only interrupt during an eval

* Fix typo; Fix lint warning when reuse the same name

* Add missing period

.circleci/config.yml

@@ -92,10 +92,12 @@ jobs:
       - run:
           name: Install Python package
           command: |
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF
+              CMAKE_COMPILE_WARNING_AS_ERROR=ON" \
               CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF \
+              CMAKE_COMPILE_WARNING_AS_ERROR=ON" \
               CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               python3 setup.py develop
       - run:
@@ -146,7 +148,9 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install -e . -v
+            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
+            CMAKE_ARGS="CMAKE_COMPILE_WARNING_AS_ERROR=ON" \
+              pip install -e . -v
       - run:
           name: Generate package stubs
           command: |
@@ -160,6 +164,7 @@ jobs:
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
             mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
       - run:
           name: Build example extension
           command: |

.pre-commit-config.yaml

@@ -1,16 +1,16 @@
 repos:
 -   repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v19.1.4
+    rev: v19.1.7
     hooks:
     -   id: clang-format
 # Using this mirror lets us use mypyc-compiled black, which is about 2x faster
 -   repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 24.10.0
+    rev: 25.1.0
     hooks:
     -   id: black
     
 -   repo: https://github.com/pycqa/isort
-    rev: 5.13.2
+    rev: 6.0.0
     hooks:
     -   id: isort
         args:

CMakeLists.txt

@@ -1,6 +1,24 @@
 cmake_minimum_required(VERSION 3.25)
 
-project(mlx LANGUAGES C CXX)
+if(NOT MLX_VERSION)
+  file(STRINGS "mlx/version.h" _mlx_h_version REGEX "^#define MLX_VERSION_.*$")
+  string(REGEX MATCH "#define MLX_VERSION_MAJOR ([0-9]+)" _ "${_mlx_h_version}")
+  set(_major ${CMAKE_MATCH_1})
+  string(REGEX MATCH "#define MLX_VERSION_MINOR ([0-9]+)" _ "${_mlx_h_version}")
+  set(_minor ${CMAKE_MATCH_1})
+  string(REGEX MATCH "#define MLX_VERSION_PATCH ([0-9]+)" _ "${_mlx_h_version}")
+  set(_patch ${CMAKE_MATCH_1})
+  set(MLX_PROJECT_VERSION "${_major}.${_minor}.${_patch}")
+  set(MLX_VERSION ${MLX_PROJECT_VERSION})
+else()
+  string(REGEX REPLACE "^([0-9]+\.[0-9]+\.[0-9]+).*" "\\1" MLX_PROJECT_VERSION
+                       ${MLX_VERSION})
+endif()
+
+project(
+  mlx
+  LANGUAGES C CXX
+  VERSION ${MLX_PROJECT_VERSION})
 
 # ----------------------------- Setup -----------------------------
 set(CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
@@ -24,13 +42,7 @@ option(MLX_BUILD_BLAS_FROM_SOURCE "Build OpenBLAS from source code" OFF)
 option(MLX_METAL_JIT "Use JIT compilation for Metal kernels" OFF)
 option(BUILD_SHARED_LIBS "Build mlx as a shared library" OFF)
 
-if(NOT MLX_VERSION)
-  set(MLX_VERSION 0.22.0)
-endif()
-add_compile_definitions("MLX_VERSION=${MLX_VERSION}")
-
 # --------------------- Processor tests -------------------------
-
 message(
   STATUS
     "Building MLX for ${CMAKE_SYSTEM_PROCESSOR} processor on ${CMAKE_SYSTEM_NAME}"
@@ -147,6 +159,7 @@ if(MLX_BUILD_CPU)
 
   if(MLX_BUILD_ACCELERATE)
     target_link_libraries(mlx PUBLIC ${ACCELERATE_LIBRARY})
+    add_compile_definitions(MLX_USE_ACCELERATE)
     add_compile_definitions(ACCELERATE_NEW_LAPACK)
   elseif(MLX_BUILD_BLAS_FROM_SOURCE)
     # Download and build OpenBLAS from source code.
@@ -217,6 +230,14 @@ if(MPI_FOUND)
   endif()
 endif()
 
+message(STATUS "Downloading json")
+FetchContent_Declare(
+  json
+  URL https://github.com/nlohmann/json/releases/download/v3.11.3/json.tar.xz)
+FetchContent_MakeAvailable(json)
+target_include_directories(
+  mlx PRIVATE $<BUILD_INTERFACE:${json_SOURCE_DIR}/single_include/nlohmann>)
+
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
 
 target_include_directories(

benchmarks/python/gather_bench.py

@@ -1,7 +1,6 @@
 # Copyright © 2023-2024 Apple Inc.
 
 import argparse
-from time import time
 
 import mlx.core as mx
 import torch

benchmarks/python/layer_norm_bench.py

@@ -10,7 +10,12 @@ def layer_norm(x, w, b, eps):
     x = x.astype(mx.float32)
     mu = mx.mean(x, -1, keepdims=True)
     v = mx.var(x, -1, keepdims=True)
-    return (x - mu) * mx.rsqrt(v + eps) * w + b
+    y = (x - mu) * mx.rsqrt(v + eps)
+    if w is not None:
+        y = y * w
+    if b is not None:
+        y = y + b
+    return y
 
 
 def time_layer_norm():
@@ -36,6 +41,28 @@ def time_layer_norm():
     time_fn(layer_norm_loop, mx.compile(g1), x, w, b)
     time_fn(layer_norm_loop, mx.compile(g2), x, w, b)
 
+    f1 = lambda x, y: (layer_norm(x, None, None, 1e-5) * y).sum()
+    f2 = lambda x, y: (mx.fast.layer_norm(x, None, None, 1e-5) * y).sum()
+    g1 = mx.grad(f1, argnums=(0,))
+    g2 = mx.grad(f2, argnums=(0,))
+
+    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
+    b = mx.random.uniform(shape=(4096,)).astype(mx.float16)
+    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    mx.eval(x, w, b, y)
+
+    def layer_norm_loop(g, x):
+        gx = x
+        for _ in range(32):
+            gx = g(gx, y)
+        return gx
+
+    time_fn(layer_norm_loop, g1, x)
+    time_fn(layer_norm_loop, g2, x)
+    time_fn(layer_norm_loop, mx.compile(g1), x)
+    time_fn(layer_norm_loop, mx.compile(g2), x)
+
 
 if __name__ == "__main__":
     time_layer_norm()

benchmarks/python/rms_norm_bench.py

@@ -9,7 +9,10 @@ def rms_norm(x, w, eps):
     ot = x.dtype
     x = x.astype(mx.float32)
     n = mx.rsqrt(x.square().mean(-1, keepdims=True) + eps)
-    return (x * n).astype(ot) * w
+    y = (x * n).astype(ot)
+    if w is not None:
+        y = y * w
+    return y
 
 
 def time_rms_norm():
@@ -34,6 +37,27 @@ def time_rms_norm():
     time_fn(rms_norm_loop, mx.compile(g1), x, w)
     time_fn(rms_norm_loop, mx.compile(g2), x, w)
 
+    f1 = lambda x, y: (rms_norm(x, None, 1e-5) * y).sum()
+    f2 = lambda x, y: (mx.fast.rms_norm(x, None, 1e-5) * y).sum()
+    g1 = mx.grad(f1, argnums=(0,))
+    g2 = mx.grad(f2, argnums=(0,))
+
+    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
+    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    mx.eval(x, w, y)
+
+    def rms_norm_loop(g, x):
+        gx = x
+        for _ in range(32):
+            gx = g(gx, y)
+        return gx
+
+    time_fn(rms_norm_loop, g1, x)
+    time_fn(rms_norm_loop, g2, x)
+    time_fn(rms_norm_loop, mx.compile(g1), x)
+    time_fn(rms_norm_loop, mx.compile(g2), x)
+
 
 if __name__ == "__main__":
     time_rms_norm()

benchmarks/python/sdpa_vector_bench.py

@@ -8,14 +8,23 @@ L = 16384
 H = 32
 H_k = H // 4
 D = 128
+V = 128
 dtype = mx.float16
 loops = 10
 
 
-def attention(q, k, v, mask=None):
+def upproject(x, w):
+    if w is None:
+        return x
+    else:
+        return x @ w.T
+
+
+def attention(q, k, v, mask=None, w=None):
     def _sdpa(q, k, v):
         B, Hq, L, D = q.shape
         _, Hk, S, _ = k.shape
+        _, _, _, V = v.shape
         q = q.reshape(B, Hk, Hq // Hk, L, D)
         k = k[:, :, None, :, :]
         v = v[:, :, None, :, :]
@@ -25,16 +34,18 @@ def attention(q, k, v, mask=None):
             s = mx.where(m, s, mx.finfo(s.dtype).min)
         p = mx.softmax(s.astype(mx.float32), axis=-1).astype(s.dtype)
         o = p @ v
-        return o.reshape(B, Hq, L, D)
+        return o.reshape(B, Hq, L, V)
 
     for i in range(loops):
         q = _sdpa(q, k, v)
+        q = upproject(q, w)
     return q
 
 
-def sdpa(q, k, v, mask=None):
+def sdpa(q, k, v, mask=None, w=None):
     for i in range(loops):
         q = mx.fast.scaled_dot_product_attention(q, k, v, scale=1.0, mask=mask)
+        q = upproject(q, w)
     return q
 
 
@@ -42,34 +53,37 @@ def time_self_attention_primitives():
     mx.random.seed(3)
     q = mx.random.uniform(shape=(1, H, 1, D)).astype(dtype)
     k = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
-    v = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
-    mx.eval(q, k, v)
-    time_fn(attention, q, k, v)
+    v = mx.random.uniform(shape=(1, H_k, L, V)).astype(dtype)
+    w = mx.random.uniform(shape=(D, V)).astype(dtype) if V != D else None
+    mx.eval(q, k, v, w)
+    time_fn(attention, q, k, v, w=w)
 
 
 def time_self_attention_sdpa():
     mx.random.seed(3)
     q = mx.random.uniform(shape=(1, H, 1, D)).astype(dtype)
     k = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
-    v = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
-    mx.eval(q, k, v)
-    time_fn(sdpa, q, k, v)
+    v = mx.random.uniform(shape=(1, H_k, L, V)).astype(dtype)
+    w = mx.random.uniform(shape=(D, V)).astype(dtype) if V != D else None
+    mx.eval(q, k, v, w)
+    time_fn(sdpa, q, k, v, w=w)
 
 
 def time_self_attention_sdpa_with_mask():
     mx.random.seed(3)
     q = mx.random.uniform(shape=(1, H, 1, D)).astype(dtype)
     k = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
-    v = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
+    v = mx.random.uniform(shape=(1, H_k, L, V)).astype(dtype)
+    w = mx.random.uniform(shape=(D, V)).astype(dtype) if V != D else None
     mask = mx.full((L,), True)
     mask[L // 2 :] = False
-    mx.eval(q, k, v, mask)
+    mx.eval(q, k, v, mask, w)
 
     def sdpa_mask(*args):
-        return sdpa(*args, mask=mask)
+        return sdpa(*args, mask=mask, w=w)
 
     def attention_mask(*args):
-        return attention(*args, mask=mask)
+        return attention(*args, mask=mask, w=w)
 
     time_fn(attention_mask, q, k, v)
     time_fn(sdpa_mask, q, k, v)

benchmarks/python/synchronize_bench.py

@@ -0,0 +1,55 @@
+import time
+
+import mlx.core as mx
+
+rank = mx.distributed.init().rank()
+
+
+def timeit(fn, a):
+
+    # warmup
+    for _ in range(5):
+        mx.eval(fn(a))
+
+    its = 10
+    tic = time.perf_counter()
+    for _ in range(its):
+        mx.eval(fn(a))
+    toc = time.perf_counter()
+    ms = 1000 * (toc - tic) / its
+    return ms
+
+
+def all_reduce_benchmark():
+    a = mx.ones((5, 5), mx.int32)
+
+    its_per_eval = 100
+
+    def fn(x):
+        for _ in range(its_per_eval):
+            x = mx.distributed.all_sum(x)
+            x = x - 1
+        return x
+
+    ms = timeit(fn, a) / its_per_eval
+    if rank == 0:
+        print(f"All Reduce: time per iteration {ms:.6f} (ms)")
+
+
+def all_gather_benchmark():
+    a = mx.ones((5, 5), mx.int32)
+    its_per_eval = 100
+
+    def fn(x):
+        for _ in range(its_per_eval):
+            x = mx.distributed.all_gather(x)[0]
+        return x
+
+    ms = timeit(fn, a) / its_per_eval
+    if rank == 0:
+        print(f"All gather: time per iteration {ms:.6f} (ms)")
+
+
+if __name__ == "__main__":
+    all_reduce_benchmark()
+    all_gather_benchmark()

cmake/extension.cmake

@@ -1,5 +1,7 @@
 include(CMakeParseArguments)
 
+# clang format off
+#
 # ##############################################################################
 # Build metal library
 #
@@ -11,6 +13,8 @@ include(CMakeParseArguments)
 # of source files INCLUDE_DIRS: List of include dirs DEPS: List of dependency
 # files (like headers)
 #
+# clang format on
+
 macro(mlx_build_metallib)
   # Parse args
   set(oneValueArgs TARGET TITLE OUTPUT_DIRECTORY)
@@ -21,7 +25,7 @@ macro(mlx_build_metallib)
   set(MTLLIB_BUILD_TARGET "${MTLLIB_OUTPUT_DIRECTORY}/${MTLLIB_TITLE}.metallib")
 
   # Collect compile options
-  set(MTLLIB_COMPILE_OPTIONS -Wall -Wextra -fno-fast-math)
+  set(MTLLIB_COMPILE_OPTIONS -Wall -Wextra -fno-fast-math -Wno-c++17-extensions)
 
   # Prepare metallib build command
   add_custom_command(

docs/src/dev/extensions.rst

@@ -22,12 +22,12 @@ You can do that in MLX directly:
 This function performs that operation while leaving the implementation and
 function transformations to MLX.
 
-However you may need to customize the underlying implementation, perhaps to
-make it faster or for custom differentiation. In this tutorial we will go
-through adding custom extensions. It will cover:
+However, you may want to customize the underlying implementation, perhaps to
+make it faster. In this tutorial we will go through adding custom extensions.
+It will cover:
 
 * The structure of the MLX library.
-* Implementing a CPU operation that redirects to Accelerate_ when appropriate.
+* Implementing a CPU operation.
 * Implementing a GPU operation using metal.
 * Adding the ``vjp`` and ``jvp`` function transformation.
 * Building a custom extension and binding it to python.
@@ -45,7 +45,7 @@ Operations
 Operations are the front-end functions that operate on arrays. They are defined
 in the C++ API (:ref:`cpp_ops`), and the Python API (:ref:`ops`) binds them.
 
-We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and
+We would like an operation :meth:`axpby` that takes in two arrays, ``x`` and
 ``y``, and two scalars, ``alpha`` and ``beta``. This is how to define it in
 C++:
 
@@ -55,7 +55,7 @@ C++:
     *  Scale and sum two vectors element-wise
     *  z = alpha * x + beta * y
     *
-    *  Follow numpy style broadcasting between x and y
+    *  Use NumPy-style broadcasting between x and y
     *  Inputs are upcasted to floats if needed
     **/
     array axpby(
@@ -66,7 +66,7 @@ C++:
         StreamOrDevice s = {} // Stream on which to schedule the operation
     );
 
-The simplest way to this operation is in terms of existing operations:
+The simplest way to implement this is with existing operations:
 
 .. code-block:: C++
 
@@ -153,9 +153,6 @@ more concrete:
       private:
         float alpha_;
         float beta_;
-
-        /** Fall back implementation for evaluation on CPU */
-        void eval(const std::vector<array>& inputs, array& out);
     };
 
 The :class:`Axpby` class derives from the base :class:`Primitive` class. The
@@ -188,7 +185,7 @@ Let's reimplement our operation now in terms of our :class:`Axpby` primitive.
         auto promoted_dtype = promote_types(x.dtype(), y.dtype());
 
         // Upcast to float32 for non-floating point inputs x and y
-        auto out_dtype = is_floating_point(promoted_dtype)
+        auto out_dtype = issubdtype(promoted_dtype, float32)
             ? promoted_dtype
             : promote_types(promoted_dtype, float32);
 
@@ -234,49 +231,59 @@ the execution of the computation graph, and calls :meth:`Axpby::eval_cpu` or
 Implementing the CPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Let's start by implementing a naive and generic version of
-:meth:`Axpby::eval_cpu`. We declared this as a private member function of
-:class:`Axpby` earlier called :meth:`Axpby::eval`.
+Let's start by implementing :meth:`Axpby::eval_cpu`.
 
-Our naive method will go over each element of the output array, find the
+The method will go over each element of the output array, find the
 corresponding input elements of ``x`` and ``y`` and perform the operation
 point-wise. This is captured in the templated function :meth:`axpby_impl`.
 
 .. code-block:: C++
 
-    template <typename T>
-    void axpby_impl(
-            const array& x,
-            const array& y,
-            array& out,
-            float alpha_,
-            float beta_) {
-        // We only allocate memory when we are ready to fill the output
-        // malloc_or_wait synchronously allocates available memory
-        // There may be a wait executed here if the allocation is requested
-        // under memory-pressured conditions
-        out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  template <typename T>
+  void axpby_impl(
+      const mx::array& x,
+      const mx::array& y,
+      mx::array& out,
+      float alpha_,
+      float beta_,
+      mx::Stream stream) {
+    // Allocate the output with `malloc_or_wait` which synchronously allocates
+    // memory, potentially waiting if the system is under memory pressure
+    out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
 
-        // Collect input and output data pointers
-        const T* x_ptr = x.data<T>();
-        const T* y_ptr = y.data<T>();
-        T* out_ptr = out.data<T>();
+    // Get the CPU command encoder and register input and output arrays
+    auto& encoder = mx::cpu::get_command_encoder(stream);
+    encoder.set_input_array(x);
+    encoder.set_input_array(y);
+    encoder.set_output_array(out);
 
-        // Cast alpha and beta to the relevant types
-        T alpha = static_cast<T>(alpha_);
-        T beta = static_cast<T>(beta_);
+    // Launch the CPU kernel
+    encoder.dispatch([x_ptr = x.data<T>(),
+                      y_ptr = y.data<T>(),
+                      out_ptr = out.data<T>(),
+                      size = out.size(),
+                      shape = out.shape(),
+                      x_strides = x.strides(),
+                      y_strides = y.strides(),
+                      alpha_,
+                      beta_]() {
 
-        // Do the element-wise operation for each output
-        for (size_t out_idx = 0; out_idx < out.size(); out_idx++) {
-            // Map linear indices to offsets in x and y
-            auto x_offset = elem_to_loc(out_idx, x.shape(), x.strides());
-            auto y_offset = elem_to_loc(out_idx, y.shape(), y.strides());
+      // Cast alpha and beta to the relevant types
+      T alpha = static_cast<T>(alpha_);
+      T beta = static_cast<T>(beta_);
 
-            // We allocate the output to be contiguous and regularly strided
-            // (defaults to row major) and hence it doesn't need additional mapping
-            out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
-        }
-    }
+      // Do the element-wise operation for each output
+      for (size_t out_idx = 0; out_idx < size; out_idx++) {
+        // Map linear indices to offsets in x and y
+        auto x_offset = mx::elem_to_loc(out_idx, shape, x_strides);
+        auto y_offset = mx::elem_to_loc(out_idx, shape, y_strides);
+
+        // We allocate the output to be contiguous and regularly strided
+        // (defaults to row major) and hence it doesn't need additional mapping
+        out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
+      }
+    });
+  }
 
 Our implementation should work for all incoming floating point arrays.
 Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
@@ -284,112 +291,32 @@ Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
 
 .. code-block:: C++
 
-    /** Fall back implementation for evaluation on CPU */
-    void Axpby::eval(
-      const std::vector<array>& inputs,
-      const std::vector<array>& outputs) {
-        auto& x = inputs[0];
-        auto& y = inputs[1];
-        auto& out = outputs[0];
-
-        // Dispatch to the correct dtype
-        if (out.dtype() == float32) {
-            return axpby_impl<float>(x, y, out, alpha_, beta_);
-        } else if (out.dtype() == float16) {
-            return axpby_impl<float16_t>(x, y, out, alpha_, beta_);
-        } else if (out.dtype() == bfloat16) {
-            return axpby_impl<bfloat16_t>(x, y, out, alpha_, beta_);
-        } else if (out.dtype() == complex64) {
-            return axpby_impl<complex64_t>(x, y, out, alpha_, beta_);
-        } else {
-            throw std::runtime_error(
-                "[Axpby] Only supports floating point types.");
-        }
-    }
-
-This is good as a fallback implementation. We can use the ``axpby`` routine
-provided by the Accelerate_ framework for a faster implementation in certain
-cases:
-
-#.  Accelerate does not provide implementations of ``axpby`` for half precision
-    floats. We can only use it for ``float32`` types.
-#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all
-    elements have fixed strides between them. We only direct to Accelerate
-    if both ``x`` and ``y`` are row contiguous or column contiguous.
-#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` in-place.
-    MLX expects to write the output to a new array. We must copy the elements
-    of ``y`` into the output and use that as an input to ``axpby``.
-
-Let's write an implementation that uses Accelerate in the right conditions.
-It allocates data for the output, copies ``y`` into it, and then calls the
-:func:`catlas_saxpby` from accelerate.
-
-.. code-block:: C++
-
-    template <typename T>
-    void axpby_impl_accelerate(
-            const array& x,
-            const array& y,
-            array& out,
-            float alpha_,
-            float beta_) {
-        // Accelerate library provides catlas_saxpby which does
-        // Y = (alpha * X) + (beta * Y) in place
-        // To use it, we first copy the data in y over to the output array
-        out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-        // We then copy over the elements using the contiguous vector specialization
-        copy_inplace(y, out, CopyType::Vector);
-
-        // Get x and y pointers for catlas_saxpby
-        const T* x_ptr = x.data<T>();
-        T* y_ptr = out.data<T>();
-
-        T alpha = static_cast<T>(alpha_);
-        T beta = static_cast<T>(beta_);
-
-        // Call the inplace accelerate operator
-        catlas_saxpby(
-            /* N = */ out.size(),
-            /* ALPHA = */ alpha,
-            /* X = */ x_ptr,
-            /* INCX = */ 1,
-            /* BETA = */ beta,
-            /* Y = */ y_ptr,
-            /* INCY = */ 1);
-    }
-
-For inputs that do not fit the criteria for accelerate, we fall back to
-:meth:`Axpby::eval`. With this in mind, let's finish our
-:meth:`Axpby::eval_cpu`.
-
-.. code-block:: C++
-
-    /** Evaluate primitive on CPU using accelerate specializations */
     void Axpby::eval_cpu(
-      const std::vector<array>& inputs,
-      const std::vector<array>& outputs) {
-        assert(inputs.size() == 2);
-        auto& x = inputs[0];
-        auto& y = inputs[1];
-        auto& out = outputs[0];
+        const std::vector<mx::array>& inputs,
+        std::vector<mx::array>& outputs) {
+      auto& x = inputs[0];
+      auto& y = inputs[1];
+      auto& out = outputs[0];
 
-        // Accelerate specialization for contiguous single precision float arrays
-        if (out.dtype() == float32 &&
-            ((x.flags().row_contiguous && y.flags().row_contiguous) ||
-            (x.flags().col_contiguous && y.flags().col_contiguous))) {
-            axpby_impl_accelerate<float>(x, y, out, alpha_, beta_);
-            return;
-        }
-
-        // Fall back to common back-end if specializations are not available
-        eval(inputs, outputs);
+      // Dispatch to the correct dtype
+      if (out.dtype() == mx::float32) {
+        return axpby_impl<float>(x, y, out, alpha_, beta_, stream());
+      } else if (out.dtype() == mx::float16) {
+        return axpby_impl<mx::float16_t>(x, y, out, alpha_, beta_, stream());
+      } else if (out.dtype() == mx::bfloat16) {
+        return axpby_impl<mx::bfloat16_t>(x, y, out, alpha_, beta_, stream());
+      } else if (out.dtype() == mx::complex64) {
+        return axpby_impl<mx::complex64_t>(x, y, out, alpha_, beta_, stream());
+      } else {
+        throw std::runtime_error(
+            "Axpby is only supported for floating point types.");
+      }
     }
 
 Just this much is enough to run the operation :meth:`axpby` on a CPU stream! If
 you do not plan on running the operation on the GPU or using transforms on
 computation graphs that contain :class:`Axpby`, you can stop implementing the
-primitive here and enjoy the speed-ups you get from the Accelerate library.
+primitive here.
 
 Implementing the GPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -824,7 +751,7 @@ Results
 ^^^^^^^
 
 Let's run a quick benchmark and see how our new ``axpby`` operation compares
-with the naive :meth:`simple_axpby` we first defined on the CPU.
+with the naive :meth:`simple_axpby` we first defined.
 
 .. code-block:: python
 
@@ -832,13 +759,11 @@ with the naive :meth:`simple_axpby` we first defined on the CPU.
     from mlx_sample_extensions import axpby
     import time
 
-    mx.set_default_device(mx.cpu)
-
     def simple_axpby(x: mx.array, y: mx.array, alpha: float, beta: float) -> mx.array:
         return alpha * x + beta * y
 
-    M = 256
-    N = 512
+    M = 4096
+    N = 4096
 
     x = mx.random.normal((M, N))
     y = mx.random.normal((M, N))
@@ -849,24 +774,24 @@ with the naive :meth:`simple_axpby` we first defined on the CPU.
 
     def bench(f):
         # Warm up
-        for i in range(100):
+        for i in range(5):
             z = f(x, y, alpha, beta)
             mx.eval(z)
 
         # Timed run
         s = time.time()
-        for i in range(5000):
+        for i in range(100):
             z = f(x, y, alpha, beta)
             mx.eval(z)
         e = time.time()
-        return e - s
+        return 1000 * (e - s) / 100
 
     simple_time = bench(simple_axpby)
     custom_time = bench(axpby)
 
-    print(f"Simple axpby: {simple_time:.3f} s | Custom axpby: {custom_time:.3f} s")
+    print(f"Simple axpby: {simple_time:.3f} ms | Custom axpby: {custom_time:.3f} ms")
 
-The results are ``Simple axpby: 0.114 s | Custom axpby: 0.109 s``. We see
+The results are ``Simple axpby: 1.559 ms | Custom axpby: 0.774 ms``. We see
 modest improvements right away!
 
 This operation is now good to be used to build other operations, in

docs/src/python/data_types.rst

@@ -51,11 +51,20 @@ The default floating point type is ``float32`` and the default integer type is
    * - ``float32``
      - 4 
      - 32-bit float
+   * - ``float64``
+     - 4 
+     - 64-bit double
    * - ``complex64``
      - 8 
      - 64-bit complex float
 
 
+.. note::
+
+    Arrays with type ``float64`` only work with CPU operations. Using
+    ``float64`` arrays on the GPU will result in an exception.
+
+
 Data type are aranged in a hierarchy. See the :obj:`DtypeCategory` object
 documentation for more information. Use :func:`issubdtype` to determine if one
 ``dtype`` (or category) is a subtype of another category.

docs/src/python/linalg.rst

@@ -5,8 +5,8 @@ Linear Algebra
 
 .. currentmodule:: mlx.core.linalg
 
-.. autosummary:: 
-   :toctree: _autosummary 
+.. autosummary::
+   :toctree: _autosummary
 
     inv
     tri_inv
@@ -18,3 +18,7 @@ Linear Algebra
     svd
     eigvalsh
     eigh
+    lu
+    lu_factor
+    solve
+    solve_triangular

docs/src/python/nn.rst

@@ -174,6 +174,7 @@ In detail:
 
    value_and_grad
    quantize
+   average_gradients
 
 .. toctree::
 

docs/src/python/ops.rst

@@ -32,6 +32,7 @@ Operations
    atleast_2d
    atleast_3d
    bitwise_and
+   bitwise_invert
    bitwise_or
    bitwise_xor
    block_masked_mm

docs/src/usage/distributed.rst

@@ -5,21 +5,27 @@ Distributed Communication
 
 .. currentmodule:: mlx.core.distributed
 
-MLX utilizes `MPI <https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ to
-provide distributed communication operations that allow the computational cost
-of training or inference to be shared across many physical machines. You can
-see a list of the supported operations in the :ref:`API docs<distributed>`.
+MLX supports distributed communication operations that allow the computational cost
+of training or inference to be shared across many physical machines. At the
+moment we support two different communication backends:
+
+* `MPI <https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ a
+  full-featured and mature distributed communications library
+* A **ring** backend of our own that uses native TCP sockets and should be
+  faster for thunderbolt connections.
+
+The list of all currently supported operations and their documentation can be
+seen in the :ref:`API docs<distributed>`.
 
 .. note::
-   A lot of operations may not be supported or not as fast as they should be.
+   Some operations may not be supported or not as fast as they should be.
    We are adding more and tuning the ones we have as we are figuring out the
    best way to do distributed computing on Macs using MLX.
 
 Getting Started
 ---------------
 
-MLX already comes with the ability to "talk" to MPI if it is installed on the
-machine. The minimal distributed program in MLX is as simple as:
+A distributed program in MLX is as simple as:
 
 .. code:: python
 
@@ -30,74 +36,79 @@ machine. The minimal distributed program in MLX is as simple as:
     print(world.rank(), x)
 
 The program above sums the array ``mx.ones(10)`` across all
-distributed processes. If simply run with ``python``, however, only one
-process is launched and no distributed communication takes place.
+distributed processes. However, when this script is run with ``python`` only
+one process is launched and no distributed communication takes place. Namely,
+all operations in ``mx.distributed`` are noops when the distributed group has a
+size of one. This property allows us to avoid code that checks if we are in a
+distributed setting similar to the one below:
 
-To launch the program in distributed mode we need to use ``mpirun`` or
-``mpiexec`` depending on the MPI installation. The simplest possible way is the
-following:
+.. code:: python
+
+    import mlx.core as mx
+
+    x = ...
+    world = mx.distributed.init()
+    # No need for the check we can simply do x = mx.distributed.all_sum(x)
+    if world.size() > 1:
+        x = mx.distributed.all_sum(x)
+
+Running Distributed Programs
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+MLX provides ``mlx.launch`` a helper script to launch distributed programs.
+Continuing with our initial example we can run it on localhost with 4 processes using
 
 .. code:: shell
 
-    $ mpirun -np 2 python test.py
-    1 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
-    0 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
+    $ mlx.launch -n 4 my_script.py
+    3 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
+    2 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
+    1 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
+    0 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
 
-The above launches two processes on the same (local) machine and we can see
-both standard output streams. The processes send the array of 1s to each other
-and compute the sum which is printed. Launching with ``mpirun -np 4 ...`` would
-print 4 etc.
-
-Installing MPI
----------------
-
-MPI can be installed with Homebrew, using the Anaconda package manager or
-compiled from source. Most of our testing is done using ``openmpi`` installed
-with the Anaconda package manager as follows:
+We can also run it on some remote hosts by providing their IPs (provided that
+the script exists on all hosts and they are reachable by ssh)
 
 .. code:: shell
 
-    $ conda install openmpi
+    $ mlx.launch --hosts ip1,ip2,ip3,ip4 my_script.py
+    3 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
+    2 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
+    1 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
+    0 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
 
-Installing with Homebrew may require specifying the location of ``libmpi.dyld``
-so that MLX can find it and load it at runtime. This can simply be achieved by
-passing the ``DYLD_LIBRARY_PATH`` environment variable to ``mpirun``.
+Consult the dedicated :doc:`usage guide<launching_distributed>` for more
+information on using ``mlx.launch``.
 
-.. code:: shell
+Selecting Backend
+^^^^^^^^^^^^^^^^^
 
-    $ mpirun -np 2 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python test.py
-
-Setting up Remote Hosts
------------------------
-
-MPI can automatically connect to remote hosts and set up the communication over
-the network if the remote hosts can be accessed via ssh. A good checklist to
-debug connectivity issues is the following:
-
-* ``ssh hostname`` works from all machines to all machines without asking for
-  password or host confirmation
-* ``mpirun`` is accessible on all machines. You can call ``mpirun`` using its
-  full path to force all machines to use a specific path.
-* Ensure that the ``hostname`` used by MPI is the one that you have configured
-  in the ``.ssh/config`` files on all machines.
+You can select the backend you want to use when calling :func:`init` by passing
+one of ``{'any', 'ring', 'mpi'}``. When passing ``any``, MLX will try to
+initialize the ``ring`` backend and if it fails the ``mpi`` backend. If they
+both fail then a singleton group is created.
 
 .. note::
-  For an example hostname ``foo.bar.com`` MPI can use only ``foo`` as
-  the hostname passed to ssh if the current hostname matches ``*.bar.com``.
+   After a distributed backend is successfully initialized :func:`init` will
+   return **the same backend** if called without arguments or with backend set to
+   ``any``.
 
-An easy way to pass the host names to MPI is using a host file. A host file
-looks like the following, where ``host1`` and ``host2`` should be the fully
-qualified domain names or IPs for these hosts.
+The following examples aim to clarify the backend initialization logic in MLX:
 
-.. code::
+.. code:: python
 
-    host1 slots=1
-    host2 slots=1
+    # Case 1: Initialize MPI regardless if it was possible to initialize the ring backend
+    world = mx.distributed.init(backend="mpi")
+    world2 = mx.distributed.init()  # subsequent calls return the MPI backend!
 
-When using MLX, it is very likely that you want to use 1 slot per host, ie one
-process per host.  The hostfile also needs to contain the current
-host if you want to run on the local host. Passing the host file to
-``mpirun`` is simply done using the ``--hostfile`` command line argument.
+    # Case 2: Initialize any backend
+    world = mx.distributed.init(backend="any")  # equivalent to no arguments
+    world2 = mx.distributed.init()  # same as above
+
+    # Case 3: Initialize both backends at the same time
+    world_mpi = mx.distributed.init(backend="mpi")
+    world_ring = mx.distributed.init(backend="ring")
+    world_any = mx.distributed.init()  # same as MPI because it was initialized first!
 
 Training Example
 ----------------
@@ -155,13 +166,179 @@ everything else remaining the same.
         optimizer.update(model, grads)
         return loss
 
-Tuning All Reduce
------------------
+Utilizing ``nn.average_gradients``
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-We are working on improving the performance of all reduce on MLX but for now
-the two main things one can do to extract the most out of distributed training with MLX are:
+Although the code example above works correctly; it performs one communication
+per gradient. It is significantly more efficient to aggregate several gradients
+together and perform fewer communication steps.
 
-1. Perform a few large reductions instead of many small ones to improve
-   bandwidth and latency
-2. Pass ``--mca btl_tcp_links 4`` to ``mpirun`` to configure it to use 4 tcp
-   connections between each host to improve bandwidth
+This is the purpose of :func:`mlx.nn.average_gradients`. The final code looks
+almost identical to the example above:
+
+.. code:: python
+
+    model = ...
+    optimizer = ...
+    dataset = ...
+
+    def step(model, x, y):
+        loss, grads = loss_grad_fn(model, x, y)
+        grads = mlx.nn.average_gradients(grads) # <---- This line was added
+        optimizer.update(model, grads)
+        return loss
+
+    for x, y in dataset:
+        loss = step(model, x, y)
+        mx.eval(loss, model.parameters())
+
+
+Getting Started with MPI
+------------------------
+
+MLX already comes with the ability to "talk" to MPI if it is installed on the
+machine. Launching distributed MLX programs that use MPI can be done with
+``mpirun`` as expected. However, in the following examples we will be using
+``mlx.launch --backend mpi`` which takes care of some nuisances such as setting
+absolute paths for the ``mpirun`` executable and the ``libmpi.dyld`` shared
+library.
+
+The simplest possible usage is the following which, assuming the minimal
+example in the beginning of this page, should result in:
+
+.. code:: shell
+
+    $ mlx.launch --backend mpi -n 2 test.py
+    1 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
+    0 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
+
+The above launches two processes on the same (local) machine and we can see
+both standard output streams. The processes send the array of 1s to each other
+and compute the sum which is printed. Launching with ``mlx.launch -n 4 ...`` would
+print 4 etc.
+
+Installing MPI
+^^^^^^^^^^^^^^
+
+MPI can be installed with Homebrew, using the Anaconda package manager or
+compiled from source. Most of our testing is done using ``openmpi`` installed
+with the Anaconda package manager as follows:
+
+.. code:: shell
+
+    $ conda install conda-forge::openmpi
+
+Installing with Homebrew may require specifying the location of ``libmpi.dyld``
+so that MLX can find it and load it at runtime. This can simply be achieved by
+passing the ``DYLD_LIBRARY_PATH`` environment variable to ``mpirun`` and it is
+done automatically by ``mlx.launch``.
+
+.. code:: shell
+
+    $ mpirun -np 2 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python test.py
+    $ # or simply
+    $ mlx.launch -n 2 test.py
+
+Setting up Remote Hosts
+^^^^^^^^^^^^^^^^^^^^^^^
+
+MPI can automatically connect to remote hosts and set up the communication over
+the network if the remote hosts can be accessed via ssh. A good checklist to
+debug connectivity issues is the following:
+
+* ``ssh hostname`` works from all machines to all machines without asking for
+  password or host confirmation
+* ``mpirun`` is accessible on all machines.
+* Ensure that the ``hostname`` used by MPI is the one that you have configured
+  in the ``.ssh/config`` files on all machines.
+
+Tuning MPI All Reduce
+^^^^^^^^^^^^^^^^^^^^^
+
+.. note::
+
+    For faster all reduce consider using the ring backend either with Thunderbolt
+    connections or over Ethernet.
+
+Configure MPI to use N tcp connections between each host to improve bandwidth
+by passing ``--mca btl_tcp_links N``.
+
+Force MPI to use the most performant network interface by setting ``--mca
+btl_tcp_if_include <iface>`` where ``<iface>`` should be the interface you want
+to use.
+
+Getting Started with Ring
+-------------------------
+
+The ring backend does not depend on any third party library so it is always
+available. It uses TCP sockets so the nodes need to be reachable via a network.
+As the name suggests the nodes are connected in a ring which means that rank 1
+can only communicate with rank 0 and rank 2, rank 2 only with rank 1 and rank 3
+and so on and so forth. As a result :func:`send` and :func:`recv` with
+arbitrary sender and receiver is not supported in the ring backend.
+
+Defining a Ring
+^^^^^^^^^^^^^^^
+
+The easiest way to define and use a ring is via a JSON hostfile and the
+``mlx.launch`` :doc:`helper script <launching_distributed>`. For each node one
+defines a hostname to ssh into to run commands on this node and one or more IPs
+that this node will listen to for connections.
+
+For example the hostfile below defines a 4 node ring. ``hostname1`` will be
+rank 0, ``hostname2`` rank 1 etc.
+
+.. code:: json
+
+    [
+        {"ssh": "hostname1", "ips": ["123.123.123.1"]},
+        {"ssh": "hostname2", "ips": ["123.123.123.2"]},
+        {"ssh": "hostname3", "ips": ["123.123.123.3"]},
+        {"ssh": "hostname4", "ips": ["123.123.123.4"]}
+    ]
+
+Running ``mlx.launch --hostfile ring-4.json my_script.py`` will ssh into each
+node, run the script which will listen for connections in each of the provided
+IPs. Specifically, ``hostname1`` will connect to ``123.123.123.2`` and accept a
+connection from ``123.123.123.4`` and so on and so forth.
+
+Thunderbolt Ring
+^^^^^^^^^^^^^^^^
+
+Although the ring backend can have benefits over MPI even for Ethernet, its
+main purpose is to use Thunderbolt rings for higher bandwidth communication.
+Setting up such thunderbolt rings can be done manually, but is a relatively
+tedious process. To simplify this, we provide the utility ``mlx.distributed_config``.
+
+To use ``mlx.distributed_config`` your computers need to be accessible by ssh via
+Ethernet or Wi-Fi. Subsequently, connect them via thunderbolt cables and then call the
+utility as follows:
+
+.. code:: shell
+
+   mlx.distributed_config --verbose --hosts host1,host2,host3,host4
+
+By default the script will attempt to discover the thunderbolt ring and provide
+you with the commands to configure each node as well as the ``hostfile.json``
+to use with ``mlx.launch``. If password-less ``sudo`` is available on the nodes
+then ``--auto-setup`` can be used to configure them automatically.
+
+To validate your connection without configuring anything
+``mlx.distributed_config`` can also plot the ring using DOT format.
+
+.. code:: shell
+
+   mlx.distributed_config --verbose --hosts host1,host2,host3,host4 --dot >ring.dot
+   dot -Tpng ring.dot >ring.png
+   open ring.png
+
+If you want to go through the process manually, the steps are as follows:
+
+* Disable the thunderbolt bridge interface
+* For the cable connecting rank ``i`` to rank ``i + 1`` find the interfaces
+  corresponding to that cable in nodes ``i`` and ``i + 1``.
+* Set up a unique subnetwork connecting the two nodes for the corresponding
+  interfaces. For instance if the cable corresponds to ``en2`` on node ``i``
+  and ``en2`` also on node ``i + 1`` then we may assign IPs ``192.168.0.1`` and
+  ``192.168.0.2`` respectively to the two nodes. For more details you can see
+  the commands prepared by the utility script.

docs/src/usage/launching_distributed.rst

@@ -0,0 +1,105 @@
+:orphan:
+
+.. _usage_launch_distributed:
+
+Launching Distributed Programs
+==============================
+
+.. currentmodule:: mlx.core.distributed
+
+Installing the MLX python package provides a helper script ``mlx.launch`` that
+can be used to run python scripts distributed on several nodes. It allows
+launching using either the MPI backend or the ring backend. See the
+:doc:`distributed docs <distributed>` for the different backends.
+
+Usage
+-----
+
+The minimal usage example of ``mlx.launch`` is simply
+
+.. code:: shell
+
+    mlx.launch --hosts ip1,ip2 my_script.py
+
+or for testing on localhost
+
+.. code:: shell
+
+    mlx.launch -n 2 my_script.py
+
+The ``mlx.launch`` command connects to the provided host and launches the input
+script on each host. It monitors each of the launched processes and terminates
+the rest if one of them fails unexpectedly or if ``mlx.launch`` is terminated.
+It also takes care of forwarding the output of each remote process to stdout
+and stderr respectively.
+
+Providing Hosts
+^^^^^^^^^^^^^^^^
+
+Hosts can be provided as command line arguments, like above, but the way that
+allows to fully define a list of hosts is via a JSON hostfile. The hostfile has
+a very simple schema. It is simply a list of objects that define each host via
+a hostname to ssh to and a list of IPs to utilize for the communication.
+
+.. code:: json
+
+    [
+        {"ssh": "hostname1", "ips": ["123.123.1.1", "123.123.2.1"]},
+        {"ssh": "hostname2", "ips": ["123.123.1.2", "123.123.2.2"]}
+    ]
+
+You can use ``mlx.distributed_config --over ethernet`` to create a hostfile
+with IPs corresponding to the ``en0`` interface.
+
+Setting up Remote Hosts
+^^^^^^^^^^^^^^^^^^^^^^^^
+
+In order to be able to launch the script on each host we need to be able to
+connect via ssh. Moreover the input script and python binary need to be on each
+host and on the same path. A good checklist to debug errors is the following:
+
+* ``ssh hostname`` works without asking for password or host confirmation
+* the python binary is available on all hosts at the same path. You can use
+  ``mlx.launch --print-python`` to see what that path is.
+* the script you want to run is available on all hosts at the same path
+
+.. _mpi_specifics:
+
+MPI Specifics
+-------------
+
+One can use MPI by passing ``--backend mpi`` to ``mlx.launch``. In that case,
+``mlx.launch`` is a thin wrapper over ``mpirun``. Moreover,
+
+* The IPs in the hostfile are ignored
+* The ssh connectivity requirement is stronger as every node needs to be able
+  to connect to every other node
+* ``mpirun`` needs to be available on every node at the same path
+
+Finally, one can pass arguments to ``mpirun`` using ``--mpi-arg``. For instance
+to choose a specific interface for the byte-transfer-layer of MPI we can call
+``mlx.launch`` as follows:
+
+.. code:: shell
+
+    mlx.launch --backend mpi --mpi-arg '--mca btl_tcp_if_include en0' --hostfile hosts.json my_script.py
+
+
+.. _ring_specifics:
+
+Ring Specifics
+--------------
+
+The ring backend, which is also the default backend, can be explicitly selected
+with the argument ``--backend ring``. The ring backend has some specific
+requirements and arguments that are different to MPI:
+
+* The argument ``--hosts`` only accepts IPs and not hostnames. If we need to
+  ssh to a hostname that does not correspond to the IP we want to bind to we
+  have to provide a hostfile.
+* ``--starting-port`` defines the port to bind to on the remote hosts.
+  Specifically rank 0 for the first IP will use this port and each subsequent
+  IP or rank will add 1 to this port.
+* ``--connections-per-ip`` allows us to increase the number of connections
+  between neighboring nodes. This corresponds to ``--mca btl_tcp_links 2`` for
+  ``mpirun``.

docs/src/usage/numpy.rst

@@ -21,11 +21,13 @@ Let's convert an array to NumPy and back.
 
 .. note::
 
-    Since NumPy does not support ``bfloat16`` arrays, you will need to convert to ``float16`` or ``float32`` first:
-    ``np.array(a.astype(mx.float32))``.
-    Otherwise, you will receive an error like: ``Item size 2 for PEP 3118 buffer format string does not match the dtype V item size 0.``
+    Since NumPy does not support ``bfloat16`` arrays, you will need to convert
+    to ``float16`` or ``float32`` first: ``np.array(a.astype(mx.float32))``.
+    Otherwise, you will receive an error like: ``Item size 2 for PEP 3118
+    buffer format string does not match the dtype V item size 0.``
 
-By default, NumPy copies data to a new array. This can be prevented by creating an array view:
+By default, NumPy copies data to a new array. This can be prevented by creating
+an array view:
 
 .. code-block:: python
 
@@ -35,10 +37,16 @@ By default, NumPy copies data to a new array. This can be prevented by creating
   a_view[0] = 1
   print(a[0].item()) # 1
 
-A NumPy array view is a normal NumPy array, except that it does not own its memory.
-This means writing to the view is reflected in the original array.
+.. note::
 
-While this is quite powerful to prevent copying arrays, it should be noted that external changes to the memory of arrays cannot be reflected in gradients.
+    NumPy arrays with type ``float64`` will be default converted to MLX arrays
+    with type ``float32``.
+
+A NumPy array view is a normal NumPy array, except that it does not own its
+memory. This means writing to the view is reflected in the original array.
+
+While this is quite powerful to prevent copying arrays, it should be noted that
+external changes to the memory of arrays cannot be reflected in gradients.
 
 Let's demonstrate this in an example:
 
@@ -56,11 +64,12 @@ Let's demonstrate this in an example:
 
 
 The function ``f`` indirectly modifies the array ``x`` through a memory view.
-However, this modification is not reflected in the gradient, as seen in the last line outputting ``1.0``,
-representing the gradient of the sum operation alone.
-The squaring of ``x`` occurs externally to MLX, meaning that no gradient is incorporated.
-It's important to note that a similar issue arises during array conversion and copying.
-For instance, a function defined as ``mx.array(np.array(x)**2).sum()`` would also result in an incorrect gradient,
+However, this modification is not reflected in the gradient, as seen in the
+last line outputting ``1.0``, representing the gradient of the sum operation
+alone.  The squaring of ``x`` occurs externally to MLX, meaning that no
+gradient is incorporated.  It's important to note that a similar issue arises
+during array conversion and copying.  For instance, a function defined as
+``mx.array(np.array(x)**2).sum()`` would also result in an incorrect gradient,
 even though no in-place operations on MLX memory are executed.
 
 PyTorch
@@ -71,7 +80,8 @@ PyTorch
    PyTorch Support for :obj:`memoryview` is experimental and can break for
    multi-dimensional arrays. Casting to NumPy first is advised for now.
 
-PyTorch supports the buffer protocol, but it requires an explicit :obj:`memoryview`.
+PyTorch supports the buffer protocol, but it requires an explicit
+:obj:`memoryview`.
 
 .. code-block:: python
 
@@ -82,7 +92,8 @@ PyTorch supports the buffer protocol, but it requires an explicit :obj:`memoryvi
   b = torch.tensor(memoryview(a))
   c = mx.array(b.numpy())
 
-Conversion from PyTorch tensors back to arrays must be done via intermediate NumPy arrays with ``numpy()``.
+Conversion from PyTorch tensors back to arrays must be done via intermediate
+NumPy arrays with ``numpy()``.
 
 JAX
 ---
@@ -100,7 +111,8 @@ JAX fully supports the buffer protocol.
 TensorFlow
 ----------
 
-TensorFlow supports the buffer protocol, but it requires an explicit :obj:`memoryview`.
+TensorFlow supports the buffer protocol, but it requires an explicit
+:obj:`memoryview`.
 
 .. code-block:: python
 

examples/extensions/CMakeLists.txt

@@ -10,7 +10,6 @@ set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 option(BUILD_SHARED_LIBS "Build extensions as a shared library" ON)
 
 # ----------------------------- Dependencies -----------------------------
-find_package(MLX CONFIG REQUIRED)
 find_package(
   Python 3.8
   COMPONENTS Interpreter Development.Module
@@ -21,6 +20,12 @@ execute_process(
   OUTPUT_VARIABLE nanobind_ROOT)
 find_package(nanobind CONFIG REQUIRED)
 
+execute_process(
+  COMMAND "${Python_EXECUTABLE}" -m mlx --cmake-dir
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+  OUTPUT_VARIABLE MLX_ROOT)
+find_package(MLX CONFIG REQUIRED)
+
 # ----------------------------- Extensions -----------------------------
 
 # Add library

examples/extensions/axpby/axpby.cpp

@@ -1,19 +1,14 @@
-// Copyright © 2023-2024 Apple Inc.
+// Copyright © 2023-2025 Apple Inc.
 
-#include <cassert>
 #include <iostream>
 #include <sstream>
 
-#include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/utils.h"
 
 #include "axpby/axpby.h"
 
-#ifdef ACCELERATE_NEW_LAPACK
-#include <vecLib/cblas_new.h>
-#endif
-
 #ifdef _METAL_
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/utils.h"
@@ -75,136 +70,67 @@ void axpby_impl(
     const mx::array& y,
     mx::array& out,
     float alpha_,
-    float beta_) {
-  // We only allocate memory when we are ready to fill the output
-  // malloc_or_wait synchronously allocates available memory
-  // There may be a wait executed here if the allocation is requested
-  // under memory-pressured conditions
+    float beta_,
+    mx::Stream stream) {
+  // Allocate the output with `malloc_or_wait` which synchronously allocates
+  // memory, potentially waiting if the system is under memory pressure
   out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
 
-  // Collect input and output data pointers
-  const T* x_ptr = x.data<T>();
-  const T* y_ptr = y.data<T>();
-  T* out_ptr = out.data<T>();
+  // Get the CPU command encoder and register input and output arrays
+  auto& encoder = mx::cpu::get_command_encoder(stream);
+  encoder.set_input_array(x);
+  encoder.set_input_array(y);
+  encoder.set_output_array(out);
 
-  // Cast alpha and beta to the relevant types
-  T alpha = static_cast<T>(alpha_);
-  T beta = static_cast<T>(beta_);
+  // Launch the CPU kernel
+  encoder.dispatch([x_ptr = x.data<T>(),
+                    y_ptr = y.data<T>(),
+                    out_ptr = out.data<T>(),
+                    size = out.size(),
+                    shape = out.shape(),
+                    x_strides = x.strides(),
+                    y_strides = y.strides(),
+                    alpha_,
+                    beta_]() {
+    // Cast alpha and beta to the relevant types
+    T alpha = static_cast<T>(alpha_);
+    T beta = static_cast<T>(beta_);
 
-  // Do the element-wise operation for each output
-  for (size_t out_idx = 0; out_idx < out.size(); out_idx++) {
-    // Map linear indices to offsets in x and y
-    auto x_offset = mx::elem_to_loc(out_idx, x.shape(), x.strides());
-    auto y_offset = mx::elem_to_loc(out_idx, y.shape(), y.strides());
+    // Do the element-wise operation for each output
+    for (size_t out_idx = 0; out_idx < size; out_idx++) {
+      // Map linear indices to offsets in x and y
+      auto x_offset = mx::elem_to_loc(out_idx, shape, x_strides);
+      auto y_offset = mx::elem_to_loc(out_idx, shape, y_strides);
 
-    // We allocate the output to be contiguous and regularly strided
-    // (defaults to row major) and hence it doesn't need additional mapping
-    out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
-  }
+      // We allocate the output to be contiguous and regularly strided
+      // (defaults to row major) and hence it doesn't need additional mapping
+      out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
+    }
+  });
 }
 
-/** Fall back implementation for evaluation on CPU */
-void Axpby::eval(
+void Axpby::eval_cpu(
     const std::vector<mx::array>& inputs,
     std::vector<mx::array>& outputs) {
-  // Check the inputs (registered in the op while constructing the out array)
-  assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
   auto& out = outputs[0];
 
   // Dispatch to the correct dtype
   if (out.dtype() == mx::float32) {
-    return axpby_impl<float>(x, y, out, alpha_, beta_);
+    return axpby_impl<float>(x, y, out, alpha_, beta_, stream());
   } else if (out.dtype() == mx::float16) {
-    return axpby_impl<mx::float16_t>(x, y, out, alpha_, beta_);
+    return axpby_impl<mx::float16_t>(x, y, out, alpha_, beta_, stream());
   } else if (out.dtype() == mx::bfloat16) {
-    return axpby_impl<mx::bfloat16_t>(x, y, out, alpha_, beta_);
+    return axpby_impl<mx::bfloat16_t>(x, y, out, alpha_, beta_, stream());
   } else if (out.dtype() == mx::complex64) {
-    return axpby_impl<mx::complex64_t>(x, y, out, alpha_, beta_);
+    return axpby_impl<mx::complex64_t>(x, y, out, alpha_, beta_, stream());
   } else {
     throw std::runtime_error(
         "Axpby is only supported for floating point types.");
   }
 }
 
-///////////////////////////////////////////////////////////////////////////////
-// Primitive Accelerate Backend Implementation
-///////////////////////////////////////////////////////////////////////////////
-
-#ifdef ACCELERATE_NEW_LAPACK
-
-template <typename T>
-void axpby_impl_accelerate(
-    const mx::array& x,
-    const mx::array& y,
-    mx::array& out,
-    float alpha_,
-    float beta_) {
-  // Accelerate library provides catlas_saxpby which does
-  // Y = (alpha * X) + (beta * Y) in place
-  // To use it, we first copy the data in y over to the output array
-
-  // This specialization requires both x and y be contiguous in the same mode
-  // i.e: corresponding linear indices in both point to corresponding elements
-  // The data in the output array is allocated to match the strides in y
-  // such that x, y, and out are contiguous in the same mode and
-  // no transposition is needed
-  out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
-
-  // We then copy over the elements using the contiguous vector specialization
-  copy_inplace(y, out, mx::CopyType::Vector);
-
-  // Get x and y pointers for catlas_saxpby
-  const T* x_ptr = x.data<T>();
-  T* y_ptr = out.data<T>();
-
-  T alpha = static_cast<T>(alpha_);
-  T beta = static_cast<T>(beta_);
-
-  // Call the inplace accelerate operator
-  catlas_saxpby(
-      /* N = */ out.size(),
-      /* ALPHA = */ alpha,
-      /* X = */ x_ptr,
-      /* INCX = */ 1,
-      /* BETA = */ beta,
-      /* Y = */ y_ptr,
-      /* INCY = */ 1);
-}
-
-/** Evaluate primitive on CPU using accelerate specializations */
-void Axpby::eval_cpu(
-    const std::vector<mx::array>& inputs,
-    std::vector<mx::array>& outputs) {
-  assert(inputs.size() == 2);
-  auto& x = inputs[0];
-  auto& y = inputs[1];
-  auto& out = outputs[0];
-
-  // Accelerate specialization for contiguous single precision float arrays
-  if (out.dtype() == mx::float32 &&
-      ((x.flags().row_contiguous && y.flags().row_contiguous) ||
-       (x.flags().col_contiguous && y.flags().col_contiguous))) {
-    axpby_impl_accelerate<float>(x, y, out, alpha_, beta_);
-    return;
-  }
-
-  // Fall back to common backend if specializations are not available
-  eval(inputs, outputs);
-}
-
-#else // Accelerate not available
-
-/** Evaluate primitive on CPU falling back to common backend */
-void Axpby::eval_cpu(
-    const std::vector<mx::array>& inputs,
-    std::vector<mx::array>& outputs) {
-  eval(inputs, outputs);
-}
-
-#endif
-
 ///////////////////////////////////////////////////////////////////////////////
 // Primitive Metal Backend Implementation
 ///////////////////////////////////////////////////////////////////////////////
@@ -216,7 +142,6 @@ void Axpby::eval_gpu(
     const std::vector<mx::array>& inputs,
     std::vector<mx::array>& outputs) {
   // Prepare inputs
-  assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
   auto& out = outputs[0];

examples/extensions/axpby/axpby.h

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2025 Apple Inc.
 
 #pragma once
 
@@ -85,11 +85,6 @@ class Axpby : public mx::Primitive {
  private:
   float alpha_;
   float beta_;
-
-  /** Fall back implementation for evaluation on CPU */
-  void eval(
-      const std::vector<mx::array>& inputs,
-      std::vector<mx::array>& outputs);
 };
 
 } // namespace my_ext

examples/extensions/axpby/axpby.metal

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2025 Apple Inc.
 
 #include <metal_stdlib>
 

mlx/CMakeLists.txt

@@ -19,6 +19,11 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/linalg.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/backend/metal/metal.h)
 
+# Define MLX_VERSION only in the version.cpp file.
+add_library(mlx_version STATIC ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp)
+target_compile_definitions(mlx_version PRIVATE MLX_VERSION="${MLX_VERSION}")
+target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:mlx_version>)
+
 if(MSVC)
   # Disable some MSVC warnings to speed up compilation.
   target_compile_options(mlx PUBLIC /wd4068 /wd4244 /wd4267 /wd4804)
@@ -29,21 +34,16 @@ if(WIN32)
   set_target_properties(mlx PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS TRUE)
 endif()
 
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/common)
+
 if(MLX_BUILD_CPU)
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/common)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/cpu)
 else()
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_cpu)
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/distributed)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/io)
-if(MLX_BUILD_ACCELERATE)
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/accelerate)
-elseif(MLX_BUILD_CPU)
-  target_sources(
-    mlx
-    PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/backend/common/default_primitives.cpp)
-endif()
 
 if(MLX_BUILD_METAL)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/metal)

mlx/array.cpp

@@ -25,7 +25,18 @@ array::array(
           std::move(shape),
           dtype,
           std::move(primitive),
-          std::move(inputs))) {}
+          std::move(inputs))) {
+  if (has_primitive() && this->primitive().stream().device == Device::gpu) {
+    for (auto& in : this->inputs()) {
+      if (in.dtype() == float64) {
+        throw std::invalid_argument("float64 is not supported on the GPU");
+      }
+    }
+    if (this->dtype() == float64) {
+      throw std::invalid_argument("float64 is not supported on the GPU");
+    }
+  }
+}
 
 std::vector<array> array::make_arrays(
     std::vector<Shape> shapes,
@@ -45,6 +56,18 @@ std::vector<array> array::make_arrays(
   return outputs;
 }
 
+array array::unsafe_weak_copy(const array& other) {
+  auto cpy = array(other.shape(), other.dtype(), nullptr, {});
+  cpy.set_data(
+      other.buffer(),
+      other.data_size(),
+      other.strides(),
+      other.flags(),
+      [](auto) {});
+  cpy.array_desc_->data_ptr = other.array_desc_->data_ptr;
+  return cpy;
+}
+
 array::array(std::initializer_list<float> data)
     : array_desc_(std::make_shared<ArrayDesc>(
           Shape{static_cast<ShapeElem>(data.size())},
@@ -66,22 +89,26 @@ array::array(allocator::Buffer data, Shape shape, Dtype dtype, Deleter deleter)
 }
 
 void array::detach() {
+  array_desc_->primitive = nullptr;
+  for (auto& s : array_desc_->siblings) {
+    s.array_desc_->primitive = nullptr;
+  }
   for (auto& s : array_desc_->siblings) {
     s.array_desc_->inputs.clear();
     s.array_desc_->siblings.clear();
     s.array_desc_->position = 0;
-    s.array_desc_->primitive = nullptr;
   }
   array_desc_->inputs.clear();
   array_desc_->siblings.clear();
   array_desc_->position = 0;
-  array_desc_->primitive = nullptr;
 }
 
 bool array::is_available() const {
   if (status() == Status::available) {
     return true;
-  } else if (status() == Status::evaluated && event().is_signaled()) {
+  } else if (
+      status() == Status::evaluated &&
+      (!event().valid() || event().is_signaled())) {
     set_status(Status::available);
     return true;
   }
@@ -90,7 +117,10 @@ bool array::is_available() const {
 
 void array::wait() {
   if (!is_available()) {
-    event().wait();
+    if (event().valid()) {
+      event().wait();
+      detach_event();
+    }
     set_status(Status::available);
   }
 }
@@ -151,34 +181,13 @@ void array::copy_shared_buffer(const array& other) {
   copy_shared_buffer(other, other.strides(), other.flags(), other.data_size());
 }
 
-void array::move_shared_buffer(
-    array other,
-    const Strides& strides,
-    Flags flags,
-    size_t data_size,
-    size_t offset /* = 0 */) {
-  array_desc_->data = std::move(other.array_desc_->data);
-  array_desc_->strides = strides;
-  array_desc_->flags = flags;
-  array_desc_->data_size = data_size;
-  auto char_offset = sizeof(char) * itemsize() * offset;
-  auto data_ptr = other.array_desc_->data_ptr;
-  other.array_desc_->data_ptr = nullptr;
-  array_desc_->data_ptr =
-      static_cast<void*>(static_cast<char*>(data_ptr) + char_offset);
-}
-
-void array::move_shared_buffer(array other) {
-  move_shared_buffer(other, other.strides(), other.flags(), other.data_size());
-}
-
 array::~array() {
   if (array_desc_ == nullptr) {
     return;
   }
 
-  // Ignore arrays that might be detached during eval
-  if (status() == array::Status::scheduled) {
+  // Detached/detaching
+  if (array_desc_->primitive == nullptr) {
     return;
   }
 

mlx/array.h

@@ -199,6 +199,13 @@ class array {
       const std::shared_ptr<Primitive>& primitive,
       const std::vector<array>& inputs);
 
+  /**
+   * Get a new array that refers to the same data as the input but with a
+   * non-owning pointer to it. Note the array is detached from the graph and has
+   * no inputs, siblings or primitive.
+   */
+  static array unsafe_weak_copy(const array& other);
+
   /** A unique identifier for an array. */
   std::uintptr_t id() const {
     return reinterpret_cast<std::uintptr_t>(array_desc_.get());
@@ -353,11 +360,6 @@ class array {
     // For example, the status of `x` in `auto x = a + b`.
     unscheduled,
 
-    // The ouptut of a computation which has been scheduled but `eval_*` has
-    // not yet been called on the array's primitive. A possible
-    // status of `x` in `auto x = a + b; eval(x);`
-    scheduled,
-
     // The array's `eval_*` function has been run, but the computation is not
     // necessarily complete. The array will have memory allocated and if it is
     // not a tracer then it will be detached from the graph.
@@ -394,6 +396,10 @@ class array {
     array_desc_->event = std::move(e);
   }
 
+  void detach_event() const {
+    array_desc_->event = Event{};
+  }
+
   // Mark the array as a tracer array (true) or not.
   void set_tracer(bool is_tracer) {
     array_desc_->is_tracer = is_tracer;
@@ -419,15 +425,6 @@ class array {
 
   void copy_shared_buffer(const array& other);
 
-  void move_shared_buffer(
-      array other,
-      const Strides& strides,
-      Flags flags,
-      size_t data_size,
-      size_t offset = 0);
-
-  void move_shared_buffer(array other);
-
   void overwrite_descriptor(const array& other) {
     array_desc_ = other.array_desc_;
   }
@@ -594,6 +591,9 @@ void array::init(It src) {
     case float32:
       std::copy(src, src + size(), data<float>());
       break;
+    case float64:
+      std::copy(src, src + size(), data<double>());
+      break;
     case bfloat16:
       std::copy(src, src + size(), data<bfloat16_t>());
       break;

mlx/backend/accelerate/CMakeLists.txt

@@ -1,8 +0,0 @@
-target_sources(
-  mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp)

mlx/backend/accelerate/conv.cpp

@@ -1,20 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include <cassert>
-
-#include <Accelerate/Accelerate.h>
-#include <simd/vector.h>
-
-#include "mlx/backend/common/copy.h"
-#include "mlx/primitives.h"
-#include "mlx/utils.h"
-
-namespace mlx::core {
-
-void Convolution::eval_cpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-
-  // TODO: Add accelerate based optimizations for CPU conv
-}
-
-} // namespace mlx::core

mlx/backend/accelerate/matmul.cpp

@@ -1,253 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include <cassert>
-
-#include <Accelerate/Accelerate.h>
-
-#include "mlx/backend/accelerate/utils.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/primitives.h"
-#include "mlx/utils.h"
-
-namespace mlx::core {
-
-namespace {
-
-std::tuple<bool, size_t, array> check_transpose(const array& arr) {
-  auto stx = arr.strides()[arr.ndim() - 2];
-  auto sty = arr.strides()[arr.ndim() - 1];
-  if (stx == arr.shape(-1) && sty == 1) {
-    return std::make_tuple(false, stx, arr);
-  } else if (stx == 1 && sty == arr.shape(-2)) {
-    return std::make_tuple(true, sty, arr);
-  } else {
-    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-    copy(arr, arr_copy, CopyType::General);
-    size_t stx = arr.shape(-1);
-    return std::make_tuple(false, stx, arr_copy);
-  }
-}
-
-inline void matmul_cblas_general(
-    const array& a_pre,
-    const array& b_pre,
-    array& out,
-    float alpha = 1.0f,
-    float beta = 0.0f) {
-  if (out.dtype() != float32) {
-    throw std::runtime_error(
-        "[matmul_cblas] on CPU currently only supports float32");
-  }
-
-  auto [a_transposed, lda, a] = check_transpose(a_pre);
-  auto [b_transposed, ldb, b] = check_transpose(b_pre);
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
-
-  if (M == 0 || N == 0) {
-    return;
-  }
-  if (K == 0) {
-    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
-    return;
-  }
-
-  for (int i = 0; i < (a.size() / (M * K)); ++i) {
-    cblas_sgemm(
-        CblasRowMajor,
-        a_transposed ? CblasTrans : CblasNoTrans, // transA
-        b_transposed ? CblasTrans : CblasNoTrans, // transB
-        M,
-        N,
-        K,
-        alpha, // alpha
-        a.data<float>() + elem_to_loc(M * K * i, a.shape(), a.strides()),
-        lda,
-        b.data<float>() + elem_to_loc(K * N * i, b.shape(), b.strides()),
-        ldb,
-        beta, // beta
-        out.data<float>() + M * N * i,
-        out.shape(-1) // ldc
-    );
-  }
-}
-
-inline void matmul_cblas(const array& a_pre, const array& b_pre, array& out) {
-  if (out.dtype() != float32) {
-    throw std::runtime_error(
-        "[matmul_cblas] on CPU currently only supports float32");
-  }
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  return matmul_cblas_general(a_pre, b_pre, out);
-}
-
-inline void matmul_bnns_general(
-    const array& a_pre,
-    const array& b_pre,
-    array& out,
-    float alpha = 1.0f,
-    float beta = 0.0f) {
-  // TODO: Update to utilize BNNS broadcasting
-
-  auto [a_transposed, lda, a] = check_transpose(a_pre);
-  auto [b_transposed, ldb, b] = check_transpose(b_pre);
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
-
-  if (M == 0 || N == 0) {
-    return;
-  }
-  if (K == 0) {
-    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
-    return;
-  }
-
-  BNNSDataType bnns_dtype = to_bnns_dtype(out.dtype());
-
-  const BNNSLayerParametersBroadcastMatMul gemm_params{
-      /* float alpha = */ alpha,
-      /* float beta = */ beta,
-      /* bool transA = */ a_transposed,
-      /* bool transB = */ b_transposed,
-      /* bool quadratic = */ false,
-      /* bool a_is_weights = */ false,
-      /* bool b_is_weights = */ false,
-      /* BNNSNDArrayDescriptor iA_desc = */
-      BNNSNDArrayDescriptor{
-          /* BNNSNDArrayFlags flags = */ BNNSNDArrayFlagBackpropSet,
-          /* BNNSDataLayout layout = */ BNNSDataLayoutRowMajorMatrix,
-
-          /* size_t size[BNNS_MAX_TENSOR_DIMENSION] = */
-          {lda, (M * K) / lda, 0, 0, 0, 0, 0, 0},
-          /* size_t stride[BNNS_MAX_TENSOR_DIMENSION] = */
-          {1, lda, 0, 0, 0, 0, 0, 0},
-
-          /* void * _Nullable data = */ nullptr,
-          /* BNNSDataType data_type = */ bnns_dtype,
-
-          /* void * _Nullable table_data = */ nullptr,
-          /* BNNSDataType table_data_type = */ bnns_dtype,
-
-          /* float data_scale = */ 1.0,
-          /* float data_bias = */ 0.0,
-      },
-      /* BNNSNDArrayDescriptor iB_desc = */
-      BNNSNDArrayDescriptor{
-          /* BNNSNDArrayFlags flags = */ BNNSNDArrayFlagBackpropSet,
-          /* BNNSDataLayout layout = */ BNNSDataLayoutRowMajorMatrix,
-
-          /* size_t size[BNNS_MAX_TENSOR_DIMENSION] = */
-          {ldb, (K * N) / ldb, 0, 0, 0, 0, 0, 0},
-          /* size_t stride[BNNS_MAX_TENSOR_DIMENSION] = */
-          {1, ldb, 0, 0, 0, 0, 0, 0},
-
-          /* void * _Nullable data = */ nullptr,
-          /* BNNSDataType data_type = */ bnns_dtype,
-
-          /* void * _Nullable table_data = */ nullptr,
-          /* BNNSDataType table_data_type = */ bnns_dtype,
-
-          /* float data_scale = */ 1.0,
-          /* float data_bias = */ 0.0,
-      },
-      /* BNNSNDArrayDescriptor o_desc = */
-      BNNSNDArrayDescriptor{
-          /* BNNSNDArrayFlags flags = */ BNNSNDArrayFlagBackpropSet,
-          /* BNNSDataLayout layout = */ BNNSDataLayoutRowMajorMatrix,
-
-          /* size_t size[BNNS_MAX_TENSOR_DIMENSION] = */
-          {N, M, 0, 0, 0, 0, 0, 0},
-          /* size_t stride[BNNS_MAX_TENSOR_DIMENSION] = */
-          {1, N, 0, 0, 0, 0, 0, 0},
-
-          /* void * _Nullable data = */ nullptr,
-          /* BNNSDataType data_type = */ bnns_dtype,
-
-          /* void * _Nullable table_data = */ nullptr,
-          /* BNNSDataType table_data_type = */ bnns_dtype,
-
-          /* float data_scale = */ 1.0,
-          /* float data_bias = */ 0.0,
-      },
-  };
-
-  auto bnns_filter =
-      BNNSFilterCreateLayerBroadcastMatMul(&gemm_params, nullptr);
-
-  for (int i = 0; i < (a.size() / (M * K)); ++i) {
-    BNNSFilterApplyTwoInput(
-        bnns_filter,
-        a.data<uint8_t>() +
-            elem_to_loc(M * K * i, a.shape(), a.strides()) * a.itemsize(),
-        b.data<uint8_t>() +
-            elem_to_loc(K * N * i, b.shape(), b.strides()) * b.itemsize(),
-        out.data<uint8_t>() + M * N * i * out.itemsize());
-  }
-
-  BNNSFilterDestroy(bnns_filter);
-}
-
-inline void matmul_bnns(const array& a_pre, const array& b_pre, array& out) {
-  // TODO: Update to utilize BNNS broadcasting
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  return matmul_bnns_general(a_pre, b_pre, out);
-}
-
-template <typename T>
-inline void mask_matrix(
-    T* data,
-    const bool* mask,
-    int tile_size,
-    const int X,
-    const int Y,
-    const size_t X_data_str,
-    const size_t Y_data_str,
-    const size_t X_mask_str,
-    const size_t Y_mask_str) {
-  int tX = (X + tile_size - 1) / tile_size;
-  int tY = (Y + tile_size - 1) / tile_size;
-
-  for (int i = 0; i < tX; i++) {
-    for (int j = 0; j < tY; j++) {
-      bool do_mask = mask[i * X_mask_str + j * Y_mask_str];
-      if (!do_mask) {
-        int loc_x = i * tile_size;
-        int loc_y = j * tile_size;
-        T* data_block = data + loc_x * X_data_str + loc_y * Y_data_str;
-
-        int size_x = std::min(tile_size, X - loc_x);
-        int size_y = std::min(tile_size, Y - loc_y);
-        for (int ii = 0; ii < size_x; ii++) {
-          for (int jj = 0; jj < size_y; jj++) {
-            data_block[ii * X_data_str + jj * Y_data_str] = T(0.);
-          }
-        }
-      }
-    }
-  }
-}
-
-} // namespace
-
-void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {
-  if (out.dtype() == float32) {
-    return matmul_cblas(inputs[0], inputs[1], out);
-  }
-  return matmul_bnns(inputs[0], inputs[1], out);
-}
-
-void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  // Fill output with C
-  auto& c = inputs[2];
-  CopyType ctype = c.data_size() == 1 ? CopyType::Scalar : CopyType::General;
-  copy(c, out, ctype);
-
-  if (out.dtype() == float32) {
-    return matmul_cblas_general(inputs[0], inputs[1], out, alpha_, beta_);
-  }
-  return matmul_bnns_general(inputs[0], inputs[1], out, alpha_, beta_);
-}
-
-} // namespace mlx::core

mlx/backend/accelerate/primitives.cpp

@@ -1,603 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include <cassert>
-#include <cmath>
-
-#include <Accelerate/Accelerate.h>
-
-#include "mlx/allocator.h"
-#include "mlx/backend/common/binary.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/unary.h"
-#include "mlx/primitives.h"
-
-#define DEFAULT(primitive)                                                 \
-  void primitive::eval_cpu(const std::vector<array>& inputs, array& out) { \
-    primitive::eval(inputs, out);                                          \
-  }
-
-#define DEFAULT_MULTI(primitive)                                       \
-  void primitive::eval_cpu(                                            \
-      const std::vector<array>& inputs, std::vector<array>& outputs) { \
-    primitive::eval(inputs, outputs);                                  \
-  }
-
-namespace mlx::core {
-
-// Use the default implementation for the following primitives
-DEFAULT(Arange)
-DEFAULT(ArgPartition)
-DEFAULT(ArgReduce)
-DEFAULT(ArgSort)
-DEFAULT(AsStrided)
-DEFAULT(BlockMaskedMM)
-DEFAULT(Broadcast)
-DEFAULT(BroadcastAxes)
-DEFAULT(Ceil)
-DEFAULT(Concatenate)
-DEFAULT(Conjugate)
-DEFAULT(Copy)
-DEFAULT_MULTI(CustomTransforms)
-DEFAULT_MULTI(Depends)
-DEFAULT_MULTI(DivMod)
-DEFAULT(NumberOfElements)
-DEFAULT(Equal)
-DEFAULT(Erf)
-DEFAULT(ErfInv)
-DEFAULT(ExpandDims)
-DEFAULT(FFT)
-DEFAULT(Floor)
-DEFAULT(Gather)
-DEFAULT(GatherMM)
-DEFAULT(GatherQMM)
-DEFAULT(Greater)
-DEFAULT(GreaterEqual)
-DEFAULT(Hadamard)
-DEFAULT(Less)
-DEFAULT(LessEqual)
-DEFAULT(Load)
-DEFAULT(LogicalNot)
-DEFAULT(LogicalAnd)
-DEFAULT(LogicalOr)
-DEFAULT(LogAddExp)
-DEFAULT(Maximum)
-DEFAULT(Minimum)
-DEFAULT(NotEqual)
-DEFAULT(Pad)
-DEFAULT(Partition)
-DEFAULT_MULTI(QRF)
-DEFAULT(RandomBits)
-DEFAULT(Remainder)
-DEFAULT(Round)
-DEFAULT(Scatter)
-DEFAULT(Select)
-DEFAULT(Sigmoid)
-DEFAULT(Sign)
-DEFAULT(Slice)
-DEFAULT(SliceUpdate)
-DEFAULT_MULTI(Split)
-DEFAULT(Sort)
-DEFAULT(Squeeze)
-DEFAULT(StopGradient)
-DEFAULT_MULTI(SVD)
-DEFAULT(Transpose)
-DEFAULT(Inverse)
-DEFAULT(Cholesky)
-DEFAULT_MULTI(Eigh)
-
-void Abs::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (in.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    vDSP_vabs(in.data<float>(), 1, out.data<float>(), 1, in.data_size());
-  } else if (in.dtype() == int32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    vDSP_vabsi(in.data<int>(), 1, out.data<int>(), 1, in.data_size());
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-
-  if (a.dtype() == float32) {
-    binary_op<float>(
-        a,
-        b,
-        out,
-        [](auto x, auto y) { return x + y; },
-        [](const auto* s, const auto* vec, auto* o, auto n) {
-          vDSP_vsadd((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
-        },
-        [](const auto* vec, const auto* s, auto* o, auto n) {
-          vDSP_vsadd((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
-        },
-        [](const auto* a, const auto* b, auto* o, auto n) {
-          vDSP_vadd((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
-        });
-  } else if (a.dtype() == int32) {
-    binary_op<int>(
-        a,
-        b,
-        out,
-        [](auto x, auto y) { return x + y; },
-        [](const auto* s, const auto* vec, auto* o, auto n) {
-          vDSP_vsaddi((const int*)vec, 1, (const int*)s, (int*)o, 1, n);
-        },
-        [](const auto* vec, const auto* s, auto* o, auto n) {
-          vDSP_vsaddi((const int*)vec, 1, (const int*)s, (int*)o, 1, n);
-        },
-        [](const auto* a, const auto* b, auto* o, auto n) {
-          vDSP_vaddi((const int*)a, 1, (const int*)b, 1, (int*)o, 1, n);
-        });
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void ArcCos::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvacosf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void ArcCosh::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvacoshf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void ArcSin::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvasinf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void ArcSinh::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvasinhf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void ArcTan::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvatanf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void ArcTan2::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  if (out.dtype() == float32 && a.flags().row_contiguous &&
-      b.flags().row_contiguous) {
-    if (a.is_donatable()) {
-      out.copy_shared_buffer(a);
-    } else if (b.is_donatable()) {
-      out.copy_shared_buffer(b);
-    } else {
-      out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    }
-    int size = a.data_size();
-    vvatan2f(out.data<float>(), a.data<float>(), b.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void ArcTanh::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvatanhf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void AsType::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-
-  if (in.flags().contiguous) {
-    // Use accelerate functions if possible
-    if (in.dtype() == float32 && out.dtype() == uint32) {
-      set_unary_output_data(in, out);
-      vDSP_vfixu32(
-          in.data<float>(), 1, out.data<uint32_t>(), 1, in.data_size());
-      return;
-    } else if (in.dtype() == float32 && out.dtype() == int32) {
-      set_unary_output_data(in, out);
-      vDSP_vfix32(in.data<float>(), 1, out.data<int32_t>(), 1, in.data_size());
-      return;
-    } else if (in.dtype() == uint32 && out.dtype() == float32) {
-      set_unary_output_data(in, out);
-      vDSP_vfltu32(
-          in.data<uint32_t>(), 1, out.data<float>(), 1, in.data_size());
-      return;
-    } else if (in.dtype() == int32 && out.dtype() == float32) {
-      set_unary_output_data(in, out);
-      vDSP_vflt32(in.data<int32_t>(), 1, out.data<float>(), 1, in.data_size());
-      return;
-    }
-  }
-  eval(inputs, out);
-}
-
-void Cos::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvcosf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Cosh::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvcoshf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-
-  if (a.dtype() == int32) {
-    binary_op<int>(
-        a,
-        b,
-        out,
-        [](auto x, auto y) { return x / y; },
-        UseDefaultBinaryOp(),
-        [](const auto* vec, const auto* s, auto* o, auto n) {
-          vDSP_vsdivi((const int*)vec, 1, (const int*)s, (int*)o, 1, n);
-        },
-        [](const auto* a, const auto* b, auto* o, auto n) {
-          vDSP_vdivi((const int*)b, 1, (const int*)a, 1, (int*)o, 1, n);
-        });
-  } else if (a.dtype() == float32) {
-    binary_op<float>(
-        a,
-        b,
-        out,
-        [](auto x, auto y) { return x / y; },
-        [](const auto* s, const auto* vec, auto* o, auto n) {
-          vDSP_svdiv((const float*)s, (const float*)vec, 1, (float*)o, 1, n);
-        },
-        [](const auto* vec, const auto* s, auto* o, auto n) {
-          vDSP_vsdiv((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
-        },
-        [](const auto* a, const auto* b, auto* o, auto n) {
-          vDSP_vdiv((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
-        });
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Exp::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    auto size = in.data_size();
-    vvexpf(out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Expm1::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    auto size = in.data_size();
-    vvexpm1f(
-        out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  assert(in.dtype() == out.dtype());
-  if (in.data_size() == 1 && out.dtype() == float32) {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    vDSP_vfill(in.data<float>(), out.data<float>(), 1, out.size());
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Log::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    auto size = in.data_size();
-    switch (base_) {
-      case Base::e:
-        vvlogf(
-            out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-        break;
-      case Base::two:
-        vvlog2f(
-            out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-        break;
-      case Base::ten:
-        vvlog10f(
-            out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-        break;
-    }
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Log1p::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    auto size = in.data_size();
-    vvlog1pf(
-        out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-
-  if (a.dtype() == float32) {
-    binary_op<float>(
-        a,
-        b,
-        out,
-        [](auto x, auto y) { return x * y; },
-        [](const auto* s, const auto* vec, auto* o, auto n) {
-          vDSP_vsmul((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
-        },
-        [](const auto* vec, const auto* s, auto* o, auto n) {
-          vDSP_vsmul((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
-        },
-        [](const auto* a, const auto* b, auto* o, auto n) {
-          vDSP_vmul((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
-        });
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Negative::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (in.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    vDSP_vneg(in.data<float>(), 1, out.data<float>(), 1, in.data_size());
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Power::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  if (out.dtype() == float32 && a.flags().row_contiguous &&
-      b.flags().row_contiguous) {
-    int size = a.size();
-    if (a.is_donatable() && a.itemsize() == out.itemsize()) {
-      out.copy_shared_buffer(a);
-    } else if (b.is_donatable() && b.itemsize() == out.itemsize()) {
-      out.copy_shared_buffer(b);
-    } else {
-      out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    }
-    vvpowf(out.data<float>(), b.data<float>(), a.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (reduce_type_ == Scan::Sum && out.dtype() == float32 &&
-      in.flags().row_contiguous && in.strides()[axis_] == 1 && !inclusive_) {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    int stride = in.shape(axis_);
-    int count = in.size() / stride;
-    const float* input = in.data<float>();
-    float* output = out.data<float>();
-    float s = 1.0;
-    if (!reverse_) {
-      for (int i = 0; i < count; i++) {
-        vDSP_vrsum(input - 1, 1, &s, output, 1, stride);
-        input += stride;
-        output += stride;
-      }
-    } else {
-      for (int i = 0; i < count; i++) {
-        input += stride - 1;
-        output += stride - 1;
-        vDSP_vrsum(input + 1, -1, &s, output, -1, stride);
-        input++;
-        output++;
-      }
-    }
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Sin::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvsinf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Sinh::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvsinhf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Square::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (in.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    auto size = in.data_size();
-    vDSP_vsq(in.data<float>(), 1, out.data<float>(), 1, size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Sqrt::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (in.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    if (recip_) {
-      vvrsqrtf(out.data<float>(), in.data<float>(), &size);
-    } else {
-      vvsqrtf(out.data<float>(), in.data<float>(), &size);
-    }
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-
-  if (a.dtype() == float32) {
-    binary_op<float>(
-        a,
-        b,
-        out,
-        [](auto x, auto y) { return x - y; },
-        [](const auto* s, const auto* vec, auto* o, auto n) {
-          float minus_1 = -1;
-          vDSP_vsmsa(
-              (const float*)vec, 1, &minus_1, (const float*)s, (float*)o, 1, n);
-        },
-        [](const auto* vec, const auto* s, auto* o, auto n) {
-          float val = -(*s);
-          vDSP_vsadd((const float*)vec, 1, &val, (float*)o, 1, n);
-        },
-        [](const auto* a, const auto* b, auto* o, auto n) {
-          vDSP_vsub((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
-        });
-  } else if (a.dtype() == int32) {
-    binary_op<int>(
-        a,
-        b,
-        out,
-        [](auto x, auto y) { return x - y; },
-        UseDefaultBinaryOp(),
-        [](const auto* vec, const auto* s, auto* o, auto n) {
-          int val = -(*s);
-          vDSP_vsaddi((const int*)vec, 1, &val, (int*)o, 1, n);
-        },
-        UseDefaultBinaryOp());
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Tan::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvtanf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-void Tanh::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == float32 && in.flags().contiguous) {
-    set_unary_output_data(in, out);
-    int size = in.data_size();
-    vvtanhf(out.data<float>(), in.data<float>(), &size);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-} // namespace mlx::core

mlx/backend/accelerate/quantized.cpp

@@ -1,117 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cassert>
-
-#include <simd/vector.h>
-
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-namespace {
-
-void _qmm_t_4_64(
-    float* result,
-    const float* x,
-    const uint32_t* w,
-    const float* scales,
-    const float* biases,
-    int M,
-    int N,
-    int K,
-    int B,
-    bool batched_w) {
-  constexpr int bits = 4;
-  constexpr int group_size = 64;
-  constexpr int bitmask = (1 << bits) - 1;
-  constexpr int pack_factor = 32 / bits;
-  constexpr int packs_in_group = group_size / pack_factor;
-
-  int w_els = N * K / pack_factor;
-  int g_els = w_els * pack_factor / group_size;
-
-  for (int i = 0; i < B; i++) {
-    for (int m = 0; m < M; m++) {
-      const uint32_t* w_local = w;
-      const float* scales_local = scales;
-      const float* biases_local = biases;
-
-      for (int n = 0; n < N; n++) {
-        const simd_float16* x_local = (simd_float16*)x;
-        simd_float16 sum = 0;
-        for (int k = 0; k < K; k += group_size) {
-          float scale = *scales_local++;
-          float bias = *biases_local++;
-
-          for (int kw = 0; kw < packs_in_group; kw += 2) {
-            // TODO: vectorize this properly
-            simd_uint16 wi;
-            for (int e = 0; e < 2; e++) {
-              uint32_t wii = *w_local++;
-              for (int p = 0; p < 8; p++) {
-                wi[e * 8 + p] = wii & bitmask;
-                wii >>= bits;
-              }
-            }
-            simd_float16 wf = simd_float(wi);
-            wf *= scale;
-            wf += bias;
-
-            sum += (*x_local) * wf;
-            x_local++;
-          }
-        }
-
-        *result = simd_reduce_add(sum);
-        result++;
-      }
-
-      x += K;
-    }
-    if (batched_w) {
-      w += w_els;
-      scales += g_els;
-      biases += g_els;
-    }
-  }
-}
-
-} // namespace
-
-void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 4);
-
-  auto& x = inputs[0];
-  auto& w = inputs[1];
-  auto& scales = inputs[2];
-  auto& biases = inputs[3];
-
-  bool condition =
-      (transpose_ && x.flags().row_contiguous && w.flags().row_contiguous &&
-       scales.flags().row_contiguous && biases.flags().row_contiguous &&
-       x.dtype() == float32 && bits_ == 4 && group_size_ == 64);
-
-  if (condition) {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    int K = x.shape(-1);
-    int M = x.shape(-2);
-    int N = out.shape(-1);
-    int B = x.size() / K / M;
-    bool batched_w = w.ndim() > 2;
-    _qmm_t_4_64(
-        out.data<float>(),
-        x.data<float>(),
-        w.data<uint32_t>(),
-        scales.data<float>(),
-        biases.data<float>(),
-        M,
-        N,
-        K,
-        B,
-        batched_w);
-  } else {
-    eval(inputs, out);
-  }
-}
-
-} // namespace mlx::core

mlx/backend/accelerate/reduce.cpp

@@ -1,139 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cassert>
-
-#include <Accelerate/Accelerate.h>
-#include <simd/vector.h>
-
-#include "mlx/backend/common/reduce.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-namespace {
-
-template <typename T, typename VT>
-struct MinReduction {
-  T operator()(const T& a, const T& b) {
-    return std::min(a, b);
-  }
-
-  VT operator()(VT a, VT b) {
-    return simd_min(a, b);
-  }
-};
-
-template <typename T, typename VT>
-struct MaxReduction {
-  T operator()(const T& a, const T& b) {
-    return std::max(a, b);
-  }
-
-  VT operator()(VT a, VT b) {
-    return simd_max(a, b);
-  }
-};
-
-template <typename T, typename VT>
-struct SumReduction {
-  T operator()(const T& a, const T& b) {
-    return a + b;
-  }
-
-  VT operator()(VT a, VT b) {
-    return a + b;
-  }
-};
-
-template <typename T, typename VT, int N, typename Reduction>
-struct StridedReduce {
-  void operator()(const T* x, T* accum, int size, size_t stride) {
-    Reduction op;
-
-    for (int i = 0; i < size; i++) {
-      size_t s = stride;
-      T* a = accum;
-      while (s >= N) {
-        *(VT*)a = op((*(VT*)x), (*(VT*)a));
-        x += N;
-        a += N;
-        s -= N;
-      }
-      while (s-- > 0) {
-        *a = op(*a, *x);
-        a++;
-        x++;
-      }
-    }
-  }
-};
-
-} // namespace
-
-void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-
-  if (in.dtype() == float32) {
-    if (reduce_type_ == Reduce::Sum) {
-      reduction_op<float, float>(
-          in,
-          out,
-          axes_,
-          0,
-          StridedReduce<
-              float,
-              simd_float16,
-              16,
-              SumReduction<float, simd_float16>>(),
-          [](const auto* x, auto* accum, int size) {
-            float acc;
-            vDSP_sve((const float*)x, 1, &acc, size);
-            (*accum) += acc;
-          },
-          [](auto* accum, auto x) { *accum += x; });
-      return;
-    } else if (reduce_type_ == Reduce::Max) {
-      reduction_op<float, float>(
-          in,
-          out,
-          axes_,
-          -std::numeric_limits<float>::infinity(),
-          StridedReduce<
-              float,
-              simd_float16,
-              16,
-              MaxReduction<float, simd_float16>>(),
-          [](const auto* x, auto* accum, int size) {
-            float max;
-            vDSP_maxv((const float*)x, 1, &max, size);
-            (*accum) = (*accum < max) ? max : *accum;
-          },
-          [](auto* accum, auto x) { (*accum) = (*accum < x) ? x : *accum; });
-      return;
-    } else if (reduce_type_ == Reduce::Min) {
-      reduction_op<float, float>(
-          in,
-          out,
-          axes_,
-          std::numeric_limits<float>::infinity(),
-          StridedReduce<
-              float,
-              simd_float16,
-              16,
-              MinReduction<float, simd_float16>>(),
-          [](const auto* x, auto* accum, int size) {
-            float min;
-            vDSP_minv((const float*)x, 1, &min, size);
-            (*accum) = (*accum > min) ? min : *accum;
-          },
-          [](auto* accum, auto x) { (*accum) = (*accum > x) ? x : *accum; });
-      return;
-    }
-  }
-  // TODO: Add integer addition and min/max using the templates above and
-  //       simd_int16 and friends.
-  eval(inputs, out);
-}
-
-} // namespace mlx::core

mlx/backend/accelerate/softmax.cpp

@@ -1,393 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include <cassert>
-#include <limits>
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-#include <arm_neon.h>
-#endif
-
-#include <simd/math.h>
-#include <simd/vector.h>
-
-#include "mlx/backend/common/copy.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-namespace {
-
-/**
- * Compute exp(x) in an optimizer friendly way as follows:
- *
- * First change the problem to computing 2**y where y = x / ln(2).
- *
- * Now we will compute 2**y as 2**y1 * 2**y2 where y1 is the integer part
- * `ipart` and y2 is fractional part. For the integer part we perform bit
- * shifting and for the fractional part we use a polynomial approximation.
- *
- * The algorithm and constants of the polynomial taken from
- * https://github.com/akohlmey/fastermath/blob/master/src/exp.c which took them
- * from Cephes math library.
- *
- * Note: The implementation below is a general fast exp. There could be faster
- *       implementations for numbers strictly < 0.
- */
-inline simd_float16 simd_fast_exp(simd_float16 x_init) {
-  auto x = x_init * 1.442695; // multiply with log_2(e)
-  simd_float16 ipart, fpart;
-  simd_int16 epart;
-  x = simd_clamp(x, -80, 80);
-  ipart = simd::floor(x + 0.5);
-  fpart = x - ipart;
-
-  x = 1.535336188319500e-4f;
-  x = x * fpart + 1.339887440266574e-3f;
-  x = x * fpart + 9.618437357674640e-3f;
-  x = x * fpart + 5.550332471162809e-2f;
-  x = x * fpart + 2.402264791363012e-1f;
-  x = x * fpart + 6.931472028550421e-1f;
-  x = x * fpart + 1.000000000000000f;
-
-  // generate 2**ipart in the floating point representation using integer
-  // bitshifting
-  epart = (simd_int(ipart) + 127) << 23;
-
-  // Avoid supressing NaNs
-  simd_int16 eq = (x_init == x_init);
-  return simd_bitselect(x_init, (*(simd_float16*)&epart) * x, eq);
-}
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-/**
- * The ARM neon equivalent of the fast exp above.
- */
-inline float16x8_t neon_fast_exp(float16x8_t x) {
-  x = vmulq_f16(x, vdupq_n_f16(float16_t(1.442695f))); // multiply with log_2(e)
-  x = vmaxq_f16(x, vdupq_n_f16(float16_t(-14.f))); // clamp under with -14
-  x = vminq_f16(x, vdupq_n_f16(float16_t(14.f))); // clamp over with 14
-
-  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(float16_t(0.5f))));
-  float16x8_t fpart = vsubq_f16(x, ipart);
-
-  x = vdupq_n_f16(float16_t(1.535336188319500e-4f));
-  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(float16_t(9.618437357674640e-3f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(float16_t(5.550332471162809e-2f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(float16_t(2.402264791363012e-1f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(float16_t(6.931472028550421e-1f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(float16_t(1.000000000000000f)), x, fpart);
-
-  // generate 2**ipart in the floating point representation using integer
-  // bitshifting
-  int16x8_t epart = vcvtq_s16_f16(ipart);
-  epart = vaddq_s16(epart, vdupq_n_s16(15));
-  epart = vshlq_n_s16(epart, 10);
-
-  return vmulq_f16(vreinterpretq_f16_s16(epart), x);
-}
-
-/**
- * Implementation of folding maximum for ARM neon. This should possibly be
- * refactored out of softmax.cpp at some point.
- */
-inline float16_t neon_reduce_max(float16x8_t x) {
-  float16x4_t y;
-  y = vpmax_f16(vget_low_f16(x), vget_high_f16(x));
-  y = vpmax_f16(y, y);
-  y = vpmax_f16(y, y);
-  return vget_lane_f16(y, 0);
-}
-
-/**
- * Implementation of folding sum for ARM neon. This should possibly be
- * refactored out of softmax.cpp at some point.
- */
-inline float16_t neon_reduce_add(float16x8_t x) {
-  float16x4_t y;
-  float16x4_t zero = vdup_n_f16(0);
-  y = vpadd_f16(vget_low_f16(x), vget_high_f16(x));
-  y = vpadd_f16(y, zero);
-  y = vpadd_f16(y, zero);
-  return vget_lane_f16(y, 0);
-}
-
-template <typename T, typename VT>
-struct NeonFp16SimdOps {
-  VT init(T a) {
-    return vdupq_n_f16(a);
-  }
-
-  VT load(const T* a) {
-    return vld1q_f16(a);
-  }
-
-  void store(T* dst, VT x) {
-    vst1q_f16(dst, x);
-  }
-
-  VT max(VT a, VT b) {
-    return vmaxq_f16(a, b);
-  }
-
-  VT exp(VT x) {
-    return neon_fast_exp(x);
-  }
-
-  VT add(VT a, VT b) {
-    return vaddq_f16(a, b);
-  }
-
-  VT sub(VT a, T b) {
-    return vsubq_f16(a, vdupq_n_f16(b));
-  }
-
-  VT mul(VT a, VT b) {
-    return vmulq_f16(a, b);
-  }
-
-  VT mul(VT a, T b) {
-    return vmulq_f16(a, vdupq_n_f16(b));
-  }
-
-  T reduce_max(VT x) {
-    return neon_reduce_max(x);
-  }
-
-  T reduce_add(VT x) {
-    return neon_reduce_add(x);
-  }
-};
-
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-template <typename T, typename VT>
-struct AccelerateSimdOps {
-  VT init(T a) {
-    return a;
-  }
-
-  VT load(const T* a) {
-    return *(VT*)a;
-  }
-
-  void store(T* dst, VT x) {
-    *(VT*)dst = x;
-  }
-
-  VT max(VT a, VT b) {
-    return simd_max(a, b);
-  }
-
-  VT exp(VT x) {
-    return simd_fast_exp(x);
-  }
-
-  VT add(VT a, VT b) {
-    return a + b;
-  }
-
-  VT sub(VT a, T b) {
-    return a - b;
-  }
-
-  VT mul(VT a, VT b) {
-    return a * b;
-  }
-
-  VT mul(VT a, T b) {
-    return a * b;
-  }
-
-  T reduce_max(VT x) {
-    return simd_reduce_max(x);
-  }
-
-  T reduce_add(VT x) {
-    return simd_reduce_add(x);
-  }
-};
-
-template <typename T, typename AccT, typename VT, typename Ops, int N>
-void softmax(const array& in, array& out) {
-  Ops ops;
-
-  const T* in_ptr = in.data<T>();
-  T* out_ptr = out.data<T>();
-  int M = in.shape().back();
-  int L = in.data_size() / M;
-  const T* current_in_ptr;
-  T* current_out_ptr;
-
-  for (int i = 0; i < L; i++, in_ptr += M, out_ptr += M) {
-    // Find the maximum
-    current_in_ptr = in_ptr;
-    VT vmaximum = ops.init(-std::numeric_limits<float>::infinity());
-    size_t s = M;
-    while (s >= N) {
-      VT vals;
-      if constexpr (std::is_same<T, AccT>::value) {
-        vals = ops.load(current_in_ptr);
-      } else {
-        for (int i = 0; i < N; ++i) {
-          vals[i] = static_cast<AccT>(current_in_ptr[i]);
-        }
-      }
-      vmaximum = ops.max(vals, vmaximum);
-      current_in_ptr += N;
-      s -= N;
-    }
-    AccT maximum = ops.reduce_max(vmaximum);
-    while (s-- > 0) {
-      maximum = std::max(maximum, static_cast<AccT>(*current_in_ptr));
-      current_in_ptr++;
-    }
-
-    // Compute the normalizer and the exponentials
-    VT vnormalizer = ops.init(0.0);
-    current_out_ptr = out_ptr;
-    current_in_ptr = in_ptr;
-    s = M;
-    while (s >= N) {
-      VT vexp;
-      if constexpr (std::is_same<T, AccT>::value) {
-        vexp = ops.load(current_in_ptr);
-      } else {
-        for (int i = 0; i < N; ++i) {
-          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
-        }
-      }
-      vexp = ops.exp(ops.sub(vexp, maximum));
-      if constexpr (std::is_same<T, AccT>::value) {
-        ops.store(current_out_ptr, vexp);
-      }
-      vnormalizer = ops.add(vnormalizer, vexp);
-      current_in_ptr += N;
-      current_out_ptr += N;
-      s -= N;
-    }
-    AccT normalizer = ops.reduce_add(vnormalizer);
-    while (s-- > 0) {
-      AccT _exp = std::exp(*current_in_ptr - maximum);
-      if (std::is_same<T, AccT>::value) {
-        *current_out_ptr = _exp;
-      }
-      normalizer += _exp;
-      current_in_ptr++;
-      current_out_ptr++;
-    }
-    normalizer = 1 / normalizer;
-
-    // Normalize
-    current_out_ptr = out_ptr;
-    current_in_ptr = in_ptr;
-    s = M;
-    while (s >= N) {
-      if constexpr (std::is_same<T, AccT>::value) {
-        ops.store(current_out_ptr, ops.mul(*(VT*)current_out_ptr, normalizer));
-      } else {
-        VT vexp;
-        for (int i = 0; i < N; ++i) {
-          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
-        }
-        vexp = ops.mul(ops.exp(ops.sub(vexp, maximum)), normalizer);
-        for (int i = 0; i < N; ++i) {
-          current_out_ptr[i] = vexp[i];
-        }
-        current_in_ptr += N;
-      }
-      current_out_ptr += N;
-      s -= N;
-    }
-    while (s-- > 0) {
-      if constexpr (std::is_same<T, AccT>::value) {
-        *current_out_ptr *= normalizer;
-      } else {
-        AccT _exp = std::exp(*current_in_ptr - maximum);
-        *current_out_ptr = static_cast<T>(_exp * normalizer);
-        current_in_ptr++;
-      }
-      current_out_ptr++;
-    }
-  }
-}
-
-} // namespace
-
-void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-
-  // Make sure that the last dimension is contiguous
-  auto check_input = [](array x) {
-    bool no_copy = x.strides()[x.ndim() - 1] == 1;
-    if (x.ndim() > 1) {
-      auto s = x.strides()[x.ndim() - 2];
-      no_copy &= (s == 0 || s == x.shape().back());
-    }
-    if (no_copy) {
-      return x;
-    } else {
-      array x_copy(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General);
-      return x_copy;
-    }
-  };
-  array in = check_input(std::move(inputs[0]));
-  out.set_data(
-      allocator::malloc_or_wait(in.data_size() * in.itemsize()),
-      in.data_size(),
-      in.strides(),
-      in.flags());
-
-  switch (in.dtype()) {
-    case bool_:
-    case uint8:
-    case uint16:
-    case uint32:
-    case uint64:
-    case int8:
-    case int16:
-    case int32:
-    case int64:
-      throw std::invalid_argument(
-          "Softmax is defined only for floating point types");
-      break;
-    case float32:
-      softmax<
-          float,
-          float,
-          simd_float16,
-          AccelerateSimdOps<float, simd_float16>,
-          16>(in, out);
-      break;
-    case float16:
-      if (precise_) {
-        softmax<
-            float16_t,
-            float,
-            simd_float16,
-            AccelerateSimdOps<float, simd_float16>,
-            16>(in, out);
-      } else {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        softmax<
-            float16_t,
-            float16_t,
-            float16x8_t,
-            NeonFp16SimdOps<float16_t, float16x8_t>,
-            8>(in, out);
-#else // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        eval(inputs, out); // Redirect to common backend for consistency
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-      }
-      break;
-    case bfloat16:
-      eval(inputs, out);
-      break;
-    case complex64:
-      eval(inputs, out);
-      break;
-  }
-}
-
-} // namespace mlx::core

mlx/backend/accelerate/utils.h

@@ -1,28 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#pragma once
-
-#include <Accelerate/Accelerate.h>
-#include "mlx/dtype.h"
-
-namespace mlx::core {
-
-BNNSDataType to_bnns_dtype(Dtype mlx_dtype) {
-  uint32_t size_bits = size_of(mlx_dtype) * 8;
-  switch (kindof(mlx_dtype)) {
-    case Dtype::Kind::b:
-      return BNNSDataTypeBoolean;
-    case Dtype::Kind::u:
-      return BNNSDataType(BNNSDataTypeUIntBit | size_bits);
-    case Dtype::Kind::i:
-      return BNNSDataType(BNNSDataTypeIntBit | size_bits);
-    case Dtype::Kind::f:
-      return BNNSDataType(BNNSDataTypeFloatBit | size_bits);
-    case Dtype::Kind::V:
-      return BNNSDataTypeBFloat16;
-    case Dtype::Kind::c:
-      throw std::invalid_argument("BNNS does not support complex types");
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/CMakeLists.txt

@@ -1,71 +1,8 @@
-if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
-  set(COMPILER ${CMAKE_C_COMPILER})
-  set(CLANG TRUE)
-else()
-  set(COMPILER ${CMAKE_CXX_COMPILER})
-endif()
-
-if(MSVC)
-  set(SHELL_EXT ps1)
-  set(SHELL_CMD powershell -ExecutionPolicy Bypass -File)
-else()
-  set(SHELL_EXT sh)
-  set(SHELL_CMD /bin/bash)
-endif()
-
-add_custom_command(
-  OUTPUT compiled_preamble.cpp
-  COMMAND
-    ${SHELL_CMD} ${CMAKE_CURRENT_SOURCE_DIR}/make_compiled_preamble.${SHELL_EXT}
-    ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp ${COMPILER}
-    ${PROJECT_SOURCE_DIR} ${CLANG} ${CMAKE_SYSTEM_PROCESSOR}
-  DEPENDS make_compiled_preamble.${SHELL_EXT}
-          compiled_preamble.h
-          ${PROJECT_SOURCE_DIR}/mlx/types/half_types.h
-          ${PROJECT_SOURCE_DIR}/mlx/types/fp16.h
-          ${PROJECT_SOURCE_DIR}/mlx/types/bf16.h
-          ${PROJECT_SOURCE_DIR}/mlx/types/complex.h
-          ops.h)
-
-add_custom_target(cpu_compiled_preamble DEPENDS compiled_preamble.cpp)
-
-add_dependencies(mlx cpu_compiled_preamble)
-
 target_sources(
   mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/common.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/eigh.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/erf.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/hadamard.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/masked_mm.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/reduce_utils.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/sort.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/threefry.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/qrf.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/svd.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/inverse.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/cholesky.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
-          ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp)
-
-if(IOS)
-  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/compiled_nocpu.cpp)
-else()
-  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/compiled_cpu.cpp
-                             ${CMAKE_CURRENT_SOURCE_DIR}/jit_compiler.cpp)
-endif()
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp)

mlx/backend/common/arange.h

@@ -1,74 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#pragma once
-
-#include "mlx/allocator.h"
-#include "mlx/array.h"
-
-namespace mlx::core {
-
-namespace {
-
-template <typename T>
-void arange(T start, T next, array& out, size_t size) {
-  auto ptr = out.data<T>();
-  auto step_size = next - start;
-  for (int i = 0; i < size; ++i) {
-    ptr[i] = start;
-    start += step_size;
-  }
-}
-
-} // namespace
-
-void arange(
-    const std::vector<array>& inputs,
-    array& out,
-    double start,
-    double step) {
-  assert(inputs.size() == 0);
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  switch (out.dtype()) {
-    case bool_:
-      throw std::runtime_error("Bool type unsupported for arange.");
-      break;
-    case uint8:
-      arange<uint8_t>(start, start + step, out, out.size());
-      break;
-    case uint16:
-      arange<uint16_t>(start, start + step, out, out.size());
-      break;
-    case uint32:
-      arange<uint32_t>(start, start + step, out, out.size());
-      break;
-    case uint64:
-      arange<uint64_t>(start, start + step, out, out.size());
-      break;
-    case int8:
-      arange<int8_t>(start, start + step, out, out.size());
-      break;
-    case int16:
-      arange<int16_t>(start, start + step, out, out.size());
-      break;
-    case int32:
-      arange<int32_t>(start, start + step, out, out.size());
-      break;
-    case int64:
-      arange<int64_t>(start, start + step, out, out.size());
-      break;
-    case float16:
-      arange<float16_t>(start, start + step, out, out.size());
-      break;
-    case float32:
-      arange<float>(start, start + step, out, out.size());
-      break;
-    case bfloat16:
-      arange<bfloat16_t>(start, start + step, out, out.size());
-      break;
-    case complex64:
-      arange<complex64_t>(start, start + step, out, out.size());
-      break;
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/arg_reduce.cpp

@@ -1,112 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cassert>
-
-#include "mlx/primitives.h"
-#include "utils.h"
-
-namespace mlx::core {
-
-namespace {
-
-template <typename InT, typename OpT>
-void arg_reduce(const array& in, array& out, const OpT& op, int axis) {
-  auto axis_size = in.shape()[axis];
-  auto axis_stride = in.strides()[axis];
-  Strides strides = in.strides();
-  Shape shape = in.shape();
-  strides.erase(strides.begin() + axis);
-  shape.erase(shape.begin() + axis);
-  for (uint32_t i = 0; i < out.size(); ++i) {
-    auto loc = elem_to_loc(i, shape, strides);
-    auto in_ptr = in.data<InT>() + loc;
-    uint32_t ind_v = 0;
-    InT v = (*in_ptr);
-    for (uint32_t j = 0; j < axis_size; ++j, in_ptr += axis_stride) {
-      op(j, (*in_ptr), &ind_v, &v);
-    }
-    out.data<uint32_t>()[i] = ind_v;
-  }
-}
-
-template <typename InT>
-void arg_reduce_dispatch(
-    const array& in,
-    array& out,
-    ArgReduce::ReduceType rtype,
-    int axis) {
-  switch (rtype) {
-    case ArgReduce::ArgMin: {
-      auto op = [](auto ind_x, auto x, auto ind_y, auto y) {
-        if (x < (*y)) {
-          (*y) = x;
-          (*ind_y) = ind_x;
-        }
-      };
-      arg_reduce<InT>(in, out, op, axis);
-      break;
-    }
-    case ArgReduce::ArgMax: {
-      auto op = [](auto ind_x, auto x, auto ind_y, auto y) {
-        if (x > (*y)) {
-          (*y) = x;
-          (*ind_y) = ind_x;
-        }
-      };
-      arg_reduce<InT>(in, out, op, axis);
-      break;
-    }
-  }
-}
-
-} // namespace
-
-void ArgReduce::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  switch (in.dtype()) {
-    case bool_:
-      arg_reduce_dispatch<bool>(in, out, reduce_type_, axis_);
-      break;
-    case uint8:
-      arg_reduce_dispatch<uint8_t>(in, out, reduce_type_, axis_);
-      break;
-    case uint16:
-      arg_reduce_dispatch<uint16_t>(in, out, reduce_type_, axis_);
-      break;
-    case uint32:
-      arg_reduce_dispatch<uint32_t>(in, out, reduce_type_, axis_);
-      break;
-    case uint64:
-      arg_reduce_dispatch<uint64_t>(in, out, reduce_type_, axis_);
-      break;
-    case int8:
-      arg_reduce_dispatch<int8_t>(in, out, reduce_type_, axis_);
-      break;
-    case int16:
-      arg_reduce_dispatch<int16_t>(in, out, reduce_type_, axis_);
-      break;
-    case int32:
-      arg_reduce_dispatch<int32_t>(in, out, reduce_type_, axis_);
-      break;
-    case int64:
-      arg_reduce_dispatch<int64_t>(in, out, reduce_type_, axis_);
-      break;
-    case float16:
-      arg_reduce_dispatch<float16_t>(in, out, reduce_type_, axis_);
-      break;
-    case float32:
-      arg_reduce_dispatch<float>(in, out, reduce_type_, axis_);
-      break;
-    case bfloat16:
-      arg_reduce_dispatch<bfloat16_t>(in, out, reduce_type_, axis_);
-      break;
-    case complex64:
-      arg_reduce_dispatch<complex64_t>(in, out, reduce_type_, axis_);
-      break;
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/binary.cpp

@@ -1,331 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cassert>
-#include <cmath>
-#include <sstream>
-
-#include "mlx/allocator.h"
-#include "mlx/backend/common/binary.h"
-#include "mlx/backend/common/binary_two.h"
-#include "mlx/backend/common/ops.h"
-#include "mlx/primitives.h"
-#include "mlx/utils.h"
-
-namespace mlx::core {
-
-namespace {
-
-template <typename T, typename U, typename Op>
-void comparison_op(const array& a, const array& b, array& out, Op op) {
-  DefaultScalarVector<T, U, Op> opsv(op);
-  DefaultVectorScalar<T, U, Op> opvs(op);
-  DefaultVectorVector<T, U, Op> opvv(op);
-  binary_op<T, U>(a, b, out, op, opsv, opvs, opvv);
-}
-
-template <typename Op>
-void comparison_op(const array& a, const array& b, array& out, Op op) {
-  switch (a.dtype()) {
-    case bool_:
-      comparison_op<bool, bool>(a, b, out, op);
-      break;
-    case uint8:
-      comparison_op<uint8_t, bool>(a, b, out, op);
-      break;
-    case uint16:
-      comparison_op<uint16_t, bool>(a, b, out, op);
-      break;
-    case uint32:
-      comparison_op<uint32_t, bool>(a, b, out, op);
-      break;
-    case uint64:
-      comparison_op<uint64_t, bool>(a, b, out, op);
-      break;
-    case int8:
-      comparison_op<int8_t, bool>(a, b, out, op);
-      break;
-    case int16:
-      comparison_op<int16_t, bool>(a, b, out, op);
-      break;
-    case int32:
-      comparison_op<int32_t, bool>(a, b, out, op);
-      break;
-    case int64:
-      comparison_op<int64_t, bool>(a, b, out, op);
-      break;
-    case float16:
-      comparison_op<float16_t, bool>(a, b, out, op);
-      break;
-    case float32:
-      comparison_op<float, bool>(a, b, out, op);
-      break;
-    case bfloat16:
-      comparison_op<bfloat16_t, bool>(a, b, out, op);
-      break;
-    case complex64:
-      comparison_op<complex64_t, bool>(a, b, out, op);
-      break;
-  }
-}
-
-} // namespace
-
-void Add::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  binary(a, b, out, detail::Add());
-}
-
-void DivMod::eval(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  auto integral_op = [](auto x, auto y) {
-    return std::make_pair(x / y, x % y);
-  };
-  auto float_op = [](auto x, auto y) {
-    return std::make_pair(std::trunc(x / y), std::fmod(x, y));
-  };
-  switch (outputs[0].dtype()) {
-    case bool_:
-      binary_op<bool>(a, b, outputs, integral_op);
-    case uint8:
-      binary_op<uint8_t>(a, b, outputs, integral_op);
-      break;
-    case uint16:
-      binary_op<uint16_t>(a, b, outputs, integral_op);
-      break;
-    case uint32:
-      binary_op<uint32_t>(a, b, outputs, integral_op);
-      break;
-    case uint64:
-      binary_op<uint64_t>(a, b, outputs, integral_op);
-      break;
-    case int8:
-      binary_op<int8_t>(a, b, outputs, integral_op);
-      break;
-    case int16:
-      binary_op<int16_t>(a, b, outputs, integral_op);
-      break;
-    case int32:
-      binary_op<int32_t>(a, b, outputs, integral_op);
-      break;
-    case int64:
-      binary_op<int64_t>(a, b, outputs, integral_op);
-      break;
-    case float16:
-      binary_op<float16_t>(a, b, outputs, float_op);
-      break;
-    case float32:
-      binary_op<float>(a, b, outputs, float_op);
-      break;
-    case bfloat16:
-      binary_op<bfloat16_t>(a, b, outputs, float_op);
-      break;
-    case complex64:
-      // Should never get here
-      throw std::runtime_error("[DivMod] Complex type not supported");
-      break;
-  }
-}
-
-void Divide::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  binary(a, b, out, detail::Divide());
-}
-
-void Remainder::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  binary(a, b, out, detail::Remainder());
-}
-
-void Equal::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  if (equal_nan_) {
-    comparison_op(inputs[0], inputs[1], out, detail::NaNEqual());
-  } else {
-    comparison_op(inputs[0], inputs[1], out, detail::Equal());
-  }
-}
-
-void Greater::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::Greater());
-}
-
-void GreaterEqual::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::GreaterEqual());
-}
-
-void Less::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::Less());
-}
-
-void LessEqual::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::LessEqual());
-}
-
-void LogAddExp::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  if (out.dtype() == float32) {
-    binary_op<float>(a, b, out, detail::LogAddExp());
-  } else if (out.dtype() == float16) {
-    binary_op<float16_t>(a, b, out, detail::LogAddExp());
-  } else if (out.dtype() == bfloat16) {
-    binary_op<bfloat16_t>(a, b, out, detail::LogAddExp());
-  } else if (issubdtype(out.dtype(), inexact)) {
-    std::ostringstream err;
-    err << "[logaddexp] Does not support " << out.dtype();
-    throw std::invalid_argument(err.str());
-  } else {
-    throw std::invalid_argument(
-        "[logaddexp] Cannot compute logaddexp for arrays with"
-        " non floating point type.");
-  }
-}
-
-void LogicalAnd::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2); // LogicalAnd requires two input arrays
-  auto& in1 = inputs[0];
-  auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalAnd());
-}
-
-void LogicalOr::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2); // LogicalOr requires two input arrays
-  auto& in1 = inputs[0];
-  auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalOr());
-}
-
-void Maximum::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  binary(a, b, out, detail::Maximum());
-}
-
-void Minimum::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  binary(a, b, out, detail::Minimum());
-}
-
-void Multiply::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  binary(a, b, out, detail::Multiply());
-}
-
-void NotEqual::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::NotEqual());
-}
-
-void Power::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  binary(a, b, out, detail::Power());
-}
-
-void Subtract::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  binary(a, b, out, detail::Subtract());
-}
-
-void BitwiseBinary::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  auto dispatch_type = [&a, &b, &out](auto op) {
-    switch (out.dtype()) {
-      case bool_:
-        binary_op<bool>(a, b, out, op);
-      case uint8:
-        binary_op<uint8_t>(a, b, out, op);
-        break;
-      case uint16:
-        binary_op<uint16_t>(a, b, out, op);
-        break;
-      case uint32:
-        binary_op<uint32_t>(a, b, out, op);
-        break;
-      case uint64:
-        binary_op<uint64_t>(a, b, out, op);
-        break;
-      case int8:
-        binary_op<int8_t>(a, b, out, op);
-        break;
-      case int16:
-        binary_op<int16_t>(a, b, out, op);
-        break;
-      case int32:
-        binary_op<int32_t>(a, b, out, op);
-        break;
-      case int64:
-        binary_op<int64_t>(a, b, out, op);
-        break;
-      default:
-        throw std::runtime_error(
-            "[BitwiseBinary::eval_cpu] Type not supported");
-        break;
-    }
-  };
-  switch (op_) {
-    case BitwiseBinary::And:
-      dispatch_type(detail::BitwiseAnd());
-      break;
-    case BitwiseBinary::Or:
-      dispatch_type(detail::BitwiseOr());
-      break;
-    case BitwiseBinary::Xor:
-      dispatch_type(detail::BitwiseXor());
-      break;
-    case BitwiseBinary::LeftShift:
-      dispatch_type(detail::LeftShift());
-      break;
-    case BitwiseBinary::RightShift:
-      dispatch_type(detail::RightShift());
-      break;
-  }
-}
-
-void ArcTan2::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  const auto& a = inputs[0];
-  const auto& b = inputs[1];
-  if (out.dtype() == float32) {
-    binary_op<float>(a, b, out, detail::ArcTan2());
-  } else if (out.dtype() == float16) {
-    binary_op<float16_t>(a, b, out, detail::ArcTan2());
-  } else if (out.dtype() == bfloat16) {
-    binary_op<bfloat16_t>(a, b, out, detail::ArcTan2());
-  } else if (issubdtype(out.dtype(), inexact)) {
-    std::ostringstream err;
-    err << "[arctan2] Does not support " << out.dtype();
-    throw std::invalid_argument(err.str());
-  } else {
-    throw std::invalid_argument(
-        "[arctan2] Cannot compute inverse tangent for arrays"
-        " with non floating point type.");
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/binary.h

@@ -1,7 +1,6 @@
 // Copyright © 2023 Apple Inc.
 
 #pragma once
-#include <cassert>
 
 #include "mlx/allocator.h"
 #include "mlx/array.h"
@@ -9,8 +8,6 @@
 
 namespace mlx::core {
 
-namespace {
-
 enum class BinaryOpType {
   ScalarScalar,
   ScalarVector,
@@ -19,7 +16,7 @@ enum class BinaryOpType {
   General,
 };
 
-BinaryOpType get_binary_op_type(const array& a, const array& b) {
+inline BinaryOpType get_binary_op_type(const array& a, const array& b) {
   BinaryOpType bopt;
   if (a.data_size() == 1 && b.data_size() == 1) {
     bopt = BinaryOpType::ScalarScalar;
@@ -37,12 +34,11 @@ BinaryOpType get_binary_op_type(const array& a, const array& b) {
   return bopt;
 }
 
-void set_binary_op_output_data(
+inline void set_binary_op_output_data(
     const array& a,
     const array& b,
     array& out,
-    BinaryOpType bopt,
-    bool donate_with_move = false) {
+    BinaryOpType bopt) {
   bool b_donatable = is_donatable(b, out);
   bool a_donatable = is_donatable(a, out);
   switch (bopt) {
@@ -52,11 +48,7 @@ void set_binary_op_output_data(
       break;
     case BinaryOpType::ScalarVector:
       if (b_donatable) {
-        if (donate_with_move) {
-          out.move_shared_buffer(b);
-        } else {
-          out.copy_shared_buffer(b);
-        }
+        out.copy_shared_buffer(b);
       } else {
         out.set_data(
             allocator::malloc_or_wait(b.data_size() * out.itemsize()),
@@ -67,11 +59,7 @@ void set_binary_op_output_data(
       break;
     case BinaryOpType::VectorScalar:
       if (a_donatable) {
-        if (donate_with_move) {
-          out.move_shared_buffer(a);
-        } else {
-          out.copy_shared_buffer(a);
-        }
+        out.copy_shared_buffer(a);
       } else {
         out.set_data(
             allocator::malloc_or_wait(a.data_size() * out.itemsize()),
@@ -82,17 +70,9 @@ void set_binary_op_output_data(
       break;
     case BinaryOpType::VectorVector:
       if (a_donatable) {
-        if (donate_with_move) {
-          out.move_shared_buffer(a);
-        } else {
-          out.copy_shared_buffer(a);
-        }
+        out.copy_shared_buffer(a);
       } else if (b_donatable) {
-        if (donate_with_move) {
-          out.move_shared_buffer(b);
-        } else {
-          out.copy_shared_buffer(b);
-        }
+        out.copy_shared_buffer(b);
       } else {
         out.set_data(
             allocator::malloc_or_wait(a.data_size() * out.itemsize()),
@@ -103,18 +83,10 @@ void set_binary_op_output_data(
       break;
     case BinaryOpType::General:
       if (a_donatable && a.flags().row_contiguous && a.size() == out.size()) {
-        if (donate_with_move) {
-          out.move_shared_buffer(a);
-        } else {
-          out.copy_shared_buffer(a);
-        }
+        out.copy_shared_buffer(a);
       } else if (
           b_donatable && b.flags().row_contiguous && b.size() == out.size()) {
-        if (donate_with_move) {
-          out.move_shared_buffer(b);
-        } else {
-          out.copy_shared_buffer(b);
-        }
+        out.copy_shared_buffer(b);
       } else {
         out.set_data(allocator::malloc_or_wait(out.nbytes()));
       }
@@ -122,409 +94,4 @@ void set_binary_op_output_data(
   }
 }
 
-struct UseDefaultBinaryOp {};
-
-template <typename T, typename U, typename Op>
-struct DefaultVectorScalar {
-  Op op;
-
-  DefaultVectorScalar(Op op_) : op(op_) {}
-
-  void operator()(const T* a, const T* b, U* dst, int size) {
-    T scalar = *b;
-    while (size-- > 0) {
-      *dst = op(*a, scalar);
-      dst++;
-      a++;
-    }
-  }
-};
-
-template <typename T, typename U, typename Op>
-struct DefaultScalarVector {
-  Op op;
-
-  DefaultScalarVector(Op op_) : op(op_) {}
-
-  void operator()(const T* a, const T* b, U* dst, int size) {
-    T scalar = *a;
-    while (size-- > 0) {
-      *dst = op(scalar, *b);
-      dst++;
-      b++;
-    }
-  }
-};
-
-template <typename T, typename U, typename Op>
-struct DefaultVectorVector {
-  Op op;
-
-  DefaultVectorVector(Op op_) : op(op_) {}
-
-  void operator()(const T* a, const T* b, U* dst, int size) {
-    while (size-- > 0) {
-      *dst = op(*a, *b);
-      dst++;
-      a++;
-      b++;
-    }
-  }
-};
-
-template <typename T, typename U, typename Op, int D, bool Strided>
-void binary_op_dims(
-    const T* a,
-    const T* b,
-    U* out,
-    Op op,
-    const Shape& shape,
-    const Strides& a_strides,
-    const Strides& b_strides,
-    const Strides& out_strides,
-    int axis) {
-  auto stride_a = a_strides[axis];
-  auto stride_b = b_strides[axis];
-  auto stride_out = out_strides[axis];
-  auto N = shape[axis];
-
-  for (int i = 0; i < N; i++) {
-    if constexpr (D > 1) {
-      binary_op_dims<T, U, Op, D - 1, Strided>(
-          a, b, out, op, shape, a_strides, b_strides, out_strides, axis + 1);
-    } else {
-      if constexpr (Strided) {
-        op(a, b, out, stride_out);
-      } else {
-        *out = op(*a, *b);
-      }
-    }
-    out += stride_out;
-    a += stride_a;
-    b += stride_b;
-  }
-}
-
-template <typename T, typename U, bool Strided, typename Op>
-void binary_op_dispatch_dims(
-    const array& a,
-    const array& b,
-    array& out,
-    Op op,
-    int dim,
-    const Shape& shape,
-    const Strides& a_strides,
-    const Strides& b_strides,
-    const Strides& out_strides) {
-  const T* a_ptr = a.data<T>();
-  const T* b_ptr = b.data<T>();
-  U* out_ptr = out.data<U>();
-  switch (dim) {
-    case 1:
-      binary_op_dims<T, U, Op, 1, Strided>(
-          a_ptr,
-          b_ptr,
-          out_ptr,
-          op,
-          shape,
-          a_strides,
-          b_strides,
-          out_strides,
-          0);
-      return;
-    case 2:
-      binary_op_dims<T, U, Op, 2, Strided>(
-          a_ptr,
-          b_ptr,
-          out_ptr,
-          op,
-          shape,
-          a_strides,
-          b_strides,
-          out_strides,
-          0);
-      return;
-    case 3:
-      binary_op_dims<T, U, Op, 3, Strided>(
-          a_ptr,
-          b_ptr,
-          out_ptr,
-          op,
-          shape,
-          a_strides,
-          b_strides,
-          out_strides,
-          0);
-      return;
-  }
-
-  ContiguousIterator a_it(shape, a_strides, dim - 3);
-  ContiguousIterator b_it(shape, b_strides, dim - 3);
-  auto stride = out_strides[dim - 4];
-  for (int64_t elem = 0; elem < a.size(); elem += stride) {
-    binary_op_dims<T, U, Op, 3, Strided>(
-        a_ptr + a_it.loc,
-        b_ptr + b_it.loc,
-        out_ptr + elem,
-        op,
-        shape,
-        a_strides,
-        b_strides,
-        out_strides,
-        dim - 3);
-    a_it.step();
-    b_it.step();
-  }
-}
-
-template <
-    typename T,
-    typename U,
-    typename Op,
-    typename OpSV,
-    typename OpVS,
-    typename OpVV>
-void binary_op(
-    const array& a,
-    const array& b,
-    array& out,
-    Op op,
-    OpSV opsv,
-    OpVS opvs,
-    OpVV opvv) {
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out, bopt);
-
-  // The full computation is scalar scalar so call the base op once
-  if (bopt == BinaryOpType::ScalarScalar) {
-    *(out.data<U>()) = op(*a.data<T>(), *b.data<T>());
-    return;
-  }
-
-  // The full computation is scalar vector so delegate to the op
-  if (bopt == BinaryOpType::ScalarVector) {
-    opsv(a.data<T>(), b.data<T>(), out.data<U>(), b.data_size());
-    return;
-  }
-
-  // The full computation is vector scalar so delegate to the op
-  if (bopt == BinaryOpType::VectorScalar) {
-    opvs(a.data<T>(), b.data<T>(), out.data<U>(), a.data_size());
-    return;
-  }
-
-  // The full computation is vector vector so delegate to the op
-  if (bopt == BinaryOpType::VectorVector) {
-    opvv(a.data<T>(), b.data<T>(), out.data<U>(), out.size());
-    return;
-  }
-
-  // General computation so let's try to optimize
-  auto [new_shape, new_strides] = collapse_contiguous_dims(
-      a.shape(), {a.strides(), b.strides(), out.strides()});
-  const auto& a_strides = new_strides[0];
-  const auto& b_strides = new_strides[1];
-  const auto& strides = new_strides[2];
-
-  // Get the left-most dim such that the array is row contiguous after
-  auto leftmost_rc_dim = [&strides](const auto& arr_strides) {
-    int d = arr_strides.size() - 1;
-    for (; d >= 0 && arr_strides[d] == strides[d]; d--) {
-    }
-    return d + 1;
-  };
-  auto a_rc_dim = leftmost_rc_dim(a_strides);
-  auto b_rc_dim = leftmost_rc_dim(b_strides);
-
-  // Get the left-most dim such that the array is a broadcasted "scalar" after
-  auto leftmost_s_dim = [](const auto& arr_strides) {
-    int d = arr_strides.size() - 1;
-    for (; d >= 0 && arr_strides[d] == 0; d--) {
-    }
-    return d + 1;
-  };
-  auto a_s_dim = leftmost_s_dim(a_strides);
-  auto b_s_dim = leftmost_s_dim(b_strides);
-
-  auto ndim = new_shape.size();
-
-  // Case 1: LxM and FxM where L and F are broadcastable and M is row contiguous
-  int dim = ndim;
-  if (int d = std::max(a_rc_dim, b_rc_dim); d < ndim) {
-    bopt = BinaryOpType::VectorVector;
-    dim = d;
-    // Case 2: LxM and Fx1 where L and F are broadcastable and M is row
-    // contiguous
-  } else if (int d = std::max(a_rc_dim, b_s_dim); d < ndim) {
-    bopt = BinaryOpType::VectorScalar;
-    dim = d;
-    // Case 3: Lx1 and FxM where L and F are broadcastable and M is row
-    // contiguous
-  } else if (int d = std::max(a_s_dim, b_rc_dim); d < ndim) {
-    bopt = BinaryOpType::ScalarVector;
-    dim = d;
-  }
-
-  // Can be sure dim > 0 since otherwise we would have used one of the fully
-  // contiguous methods above. Except for the case that the flags do not
-  // correspond to the underlying contiguity.
-  if (dim == 0 || strides[dim - 1] < 16) {
-    bopt = BinaryOpType::General;
-    dim = ndim;
-  }
-
-  switch (bopt) {
-    case BinaryOpType::VectorVector:
-      binary_op_dispatch_dims<T, U, true>(
-          a, b, out, opvv, dim, new_shape, a_strides, b_strides, strides);
-      break;
-    case BinaryOpType::VectorScalar:
-      binary_op_dispatch_dims<T, U, true>(
-          a, b, out, opvs, dim, new_shape, a_strides, b_strides, strides);
-      break;
-    case BinaryOpType::ScalarVector:
-      binary_op_dispatch_dims<T, U, true>(
-          a, b, out, opsv, dim, new_shape, a_strides, b_strides, strides);
-      break;
-    default:
-      binary_op_dispatch_dims<T, U, false>(
-          a, b, out, op, dim, new_shape, a_strides, b_strides, strides);
-      break;
-  }
-}
-
-template <typename T, typename Op, typename OpSV, typename OpVS, typename OpVV>
-void binary_op(
-    const array& a,
-    const array& b,
-    array& out,
-    Op op,
-    OpSV opsv,
-    OpVS opvs,
-    OpVV opvv) {
-  // TODO: The following mess of constexpr evaluations can probably be achieved
-  //       with template specializations and overloading. Would it be simpler?
-
-  if constexpr (std::is_same<decltype(opsv), UseDefaultBinaryOp>::value) {
-    if constexpr (std::is_same<decltype(opvs), UseDefaultBinaryOp>::value) {
-      if constexpr (std::is_same<decltype(opvv), UseDefaultBinaryOp>::value) {
-        // All ops are UseDefaultBinaryOp (why oh why would someone call that?)
-        binary_op<T, T>(
-            a,
-            b,
-            out,
-            op,
-            DefaultScalarVector<T, T, Op>(op),
-            DefaultVectorScalar<T, T, Op>(op),
-            DefaultVectorVector<T, T, Op>(op));
-      } else {
-        // opsv and opvs were UseDefaultBinaryOp
-        binary_op<T, T>(
-            a,
-            b,
-            out,
-            op,
-            DefaultScalarVector<T, T, Op>(op),
-            DefaultVectorScalar<T, T, Op>(op),
-            opvv);
-      }
-    } else if constexpr (std::is_same<decltype(opvv), UseDefaultBinaryOp>::
-                             value) {
-      // opsv and opvv were UseDefaultBinaryOp
-      binary_op<T, T>(
-          a,
-          b,
-          out,
-          op,
-          DefaultScalarVector<T, T, Op>(op),
-          opvs,
-          DefaultVectorVector<T, T, Op>(op));
-    } else {
-      // opsv was UseDefaultBinaryOp
-      binary_op<T, T>(
-          a, b, out, op, DefaultScalarVector<T, T, Op>(op), opvs, opvv);
-    }
-  } else if constexpr (std::is_same<decltype(opvs), UseDefaultBinaryOp>::
-                           value) {
-    if (std::is_same<decltype(opvv), UseDefaultBinaryOp>::value) {
-      // opvs and opvv were UseDefaultBinaryOp
-      binary_op<T, T>(
-          a,
-          b,
-          out,
-          op,
-          opsv,
-          DefaultVectorScalar<T, T, Op>(op),
-          DefaultVectorVector<T, T, Op>(op));
-    } else {
-      // opvs was UseDefaultBinaryOp
-      binary_op<T, T>(
-          a, b, out, op, opsv, DefaultVectorScalar<T, T, Op>(op), opvv);
-    }
-  } else if constexpr (std::is_same<decltype(opvv), UseDefaultBinaryOp>::
-                           value) {
-    // opvv was UseDefaultBinaryOp
-    binary_op<T, T>(
-        a, b, out, op, opsv, opvs, DefaultVectorVector<T, T, Op>(op));
-  } else {
-    // All ops provided
-    binary_op<T, T>(a, b, out, op, opsv, opvs, opvv);
-  }
-}
-
-template <typename T, typename Op>
-void binary_op(const array& a, const array& b, array& out, Op op) {
-  DefaultScalarVector<T, T, Op> opsv(op);
-  DefaultVectorScalar<T, T, Op> opvs(op);
-  DefaultVectorVector<T, T, Op> opvv(op);
-  binary_op<T, T>(a, b, out, op, opsv, opvs, opvv);
-}
-
-template <typename... Ops>
-void binary(const array& a, const array& b, array& out, Ops... ops) {
-  switch (out.dtype()) {
-    case bool_:
-      binary_op<bool>(a, b, out, ops...);
-      break;
-    case uint8:
-      binary_op<uint8_t>(a, b, out, ops...);
-      break;
-    case uint16:
-      binary_op<uint16_t>(a, b, out, ops...);
-      break;
-    case uint32:
-      binary_op<uint32_t>(a, b, out, ops...);
-      break;
-    case uint64:
-      binary_op<uint64_t>(a, b, out, ops...);
-      break;
-    case int8:
-      binary_op<int8_t>(a, b, out, ops...);
-      break;
-    case int16:
-      binary_op<int16_t>(a, b, out, ops...);
-      break;
-    case int32:
-      binary_op<int32_t>(a, b, out, ops...);
-      break;
-    case int64:
-      binary_op<int64_t>(a, b, out, ops...);
-      break;
-    case float16:
-      binary_op<float16_t>(a, b, out, ops...);
-      break;
-    case float32:
-      binary_op<float>(a, b, out, ops...);
-      break;
-    case bfloat16:
-      binary_op<bfloat16_t>(a, b, out, ops...);
-      break;
-    case complex64:
-      binary_op<complex64_t>(a, b, out, ops...);
-      break;
-  }
-}
-
-} // namespace
-
 } // namespace mlx::core

mlx/backend/common/cholesky.cpp

@@ -1,74 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include "mlx/allocator.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/lapack.h"
-#include "mlx/linalg.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-void cholesky_impl(const array& a, array& factor, bool upper) {
-  // Lapack uses the column-major convention. We take advantage of the fact that
-  // the matrix should be symmetric:
-  //   (A)ᵀ = A
-  // and that a column-major lower triangular matrix is a row-major upper
-  // triangular matrix, so uplo is the opposite of what we would expect from
-  // upper
-
-  char uplo = (upper) ? 'L' : 'U';
-
-  // The decomposition is computed in place, so just copy the input to the
-  // output.
-  copy(
-      a,
-      factor,
-      a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-
-  const int N = a.shape(-1);
-  const size_t num_matrices = a.size() / (N * N);
-
-  float* matrix = factor.data<float>();
-
-  for (int i = 0; i < num_matrices; i++) {
-    // Compute Cholesky factorization.
-    int info;
-    MLX_LAPACK_FUNC(spotrf)
-    (
-        /* uplo = */ &uplo,
-        /* n = */ &N,
-        /* a = */ matrix,
-        /* lda = */ &N,
-        /* info = */ &info);
-
-    // TODO: We do nothing when the matrix is not positive semi-definite
-    // because throwing an error would result in a crash. If we figure out how
-    // to catch errors from the implementation we should throw.
-    if (info < 0) {
-      std::stringstream msg;
-      msg << "[cholesky] Cholesky decomposition failed with error code "
-          << info;
-      throw std::runtime_error(msg.str());
-    }
-
-    // Zero out the upper/lower triangle while advancing the pointer to the
-    // next matrix at the same time.
-    for (int row = 0; row < N; row++) {
-      if (upper) {
-        std::fill(matrix, matrix + row, 0);
-      } else {
-        std::fill(matrix + row + 1, matrix + N, 0);
-      }
-      matrix += N;
-    }
-  }
-}
-
-void Cholesky::eval(const std::vector<array>& inputs, array& output) {
-  if (inputs[0].dtype() != float32) {
-    throw std::runtime_error("[Cholesky::eval] only supports float32.");
-  }
-  cholesky_impl(inputs[0], output, upper_);
-}
-
-} // namespace mlx::core

mlx/backend/common/common.cpp

@@ -39,7 +39,7 @@ void AsStrided::eval(const std::vector<array>& inputs, array& out) {
   // rely on data_size anyway.
   size_t data_size = out.size();
 
-  return move_or_copy(in, out, strides_, flags, data_size, offset_);
+  return out.copy_shared_buffer(in, strides_, flags, data_size, offset_);
 }
 
 void broadcast(const array& in, array& out) {
@@ -56,7 +56,7 @@ void broadcast(const array& in, array& out) {
   if (out.size() > in.size()) {
     flags.row_contiguous = flags.col_contiguous = false;
   }
-  move_or_copy(in, out, strides, flags, in.data_size());
+  out.copy_shared_buffer(in, strides, flags, in.data_size());
 }
 
 void Broadcast::eval(const std::vector<array>& inputs, array& out) {
@@ -69,7 +69,7 @@ void BroadcastAxes::eval(const std::vector<array>& inputs, array& out) {
 
 void Copy::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  move_or_copy(inputs[0], out);
+  out.copy_shared_buffer(inputs[0]);
 }
 
 void CustomTransforms::eval(
@@ -78,7 +78,7 @@ void CustomTransforms::eval(
   assert(inputs.size() > outputs.size());
   for (int i = 0, j = inputs.size() - outputs.size(); i < outputs.size();
        i++, j++) {
-    move_or_copy(inputs[j], outputs[i]);
+    outputs[i].copy_shared_buffer(inputs[j]);
   }
 }
 
@@ -87,7 +87,7 @@ void Depends::eval(
     std::vector<array>& outputs) {
   assert(inputs.size() > outputs.size());
   for (int i = 0; i < outputs.size(); i++) {
-    move_or_copy(inputs[i], outputs[i]);
+    outputs[i].copy_shared_buffer(inputs[i]);
   }
 }
 
@@ -98,7 +98,7 @@ void ExpandDims::eval(const std::vector<array>& inputs, array& out) {
   for (auto ax : axes_) {
     strides.insert(strides.begin() + ax, 1);
   }
-  move_or_copy(in, out, strides, in.flags(), in.data_size());
+  out.copy_shared_buffer(in, strides, in.flags(), in.data_size());
 }
 
 void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
@@ -151,6 +151,9 @@ void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
     case bfloat16:
       *out.data<bfloat16_t>() = static_cast<bfloat16_t>(numel);
       break;
+    case float64:
+      *out.data<double>() = static_cast<double>(numel);
+      break;
     case complex64:
       *out.data<complex64_t>() = static_cast<complex64_t>(numel);
       break;
@@ -207,7 +210,7 @@ void shared_buffer_reshape(
     auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
     flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
   }
-  move_or_copy(in, out, out_strides, flags, in.data_size());
+  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
 }
 
 void Split::eval(
@@ -273,12 +276,12 @@ void Squeeze::eval(const std::vector<array>& inputs, array& out) {
       strides.push_back(in.strides(i));
     }
   }
-  move_or_copy(in, out, strides, in.flags(), in.data_size());
+  out.copy_shared_buffer(in, strides, in.flags(), in.data_size());
 }
 
 void StopGradient::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  move_or_copy(inputs[0], out);
+  out.copy_shared_buffer(inputs[0]);
 }
 
 void Transpose::eval(const std::vector<array>& inputs, array& out) {
@@ -312,7 +315,7 @@ void Transpose::eval(const std::vector<array>& inputs, array& out) {
       b_stride *= out.shape(ri);
     }
   }
-  move_or_copy(in, out, out_strides, flags, in.data_size());
+  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
 }
 
 } // namespace mlx::core

mlx/backend/common/compiled.cpp

@@ -161,8 +161,7 @@ void compiled_allocate_outputs(
     std::vector<array>& outputs,
     const std::vector<array>& inputs_,
     const std::unordered_set<uintptr_t>& constant_ids_,
-    bool contiguous,
-    bool move_buffers /* = false */) {
+    bool contiguous) {
   if (contiguous) {
     int o = 0;
     Strides strides;
@@ -178,11 +177,7 @@ void compiled_allocate_outputs(
       if (in.itemsize() == outputs[o].itemsize() && !is_scalar(in) &&
           in.is_donatable() &&
           constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
-        if (move_buffers) {
-          outputs[o++].move_shared_buffer(in);
-        } else {
-          outputs[o++].copy_shared_buffer(in);
-        }
+        outputs[o++].copy_shared_buffer(in);
       }
       // Get representative input flags to properly set non-donated outputs
       if (strides.empty() && in.size() == outputs[0].size()) {
@@ -210,13 +205,8 @@ void compiled_allocate_outputs(
       if (in.flags().row_contiguous && in.size() == outputs[o].size() &&
           in.itemsize() == outputs[o].itemsize() && in.is_donatable() &&
           constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
-        if (move_buffers) {
-          outputs[o].move_shared_buffer(
-              in, outputs[o].strides(), in.flags(), in.data_size());
-        } else {
-          outputs[o].copy_shared_buffer(
-              in, outputs[o].strides(), in.flags(), in.data_size());
-        }
+        outputs[o].copy_shared_buffer(
+            in, outputs[o].strides(), in.flags(), in.data_size());
         o++;
       }
     }

mlx/backend/common/compiled.h

@@ -62,7 +62,6 @@ void compiled_allocate_outputs(
     std::vector<array>& outputs,
     const std::vector<array>& inputs_,
     const std::unordered_set<uintptr_t>& constant_ids_,
-    bool contiguous,
-    bool move_buffers = false);
+    bool contiguous);
 
 } // namespace mlx::core

mlx/backend/common/copy.h

@@ -3,7 +3,6 @@
 #pragma once
 
 #include "mlx/array.h"
-#include "mlx/backend/common/utils.h"
 
 namespace mlx::core {
 
@@ -23,17 +22,25 @@ enum class CopyType {
   GeneralGeneral
 };
 
-void copy(const array& src, array& dst, CopyType ctype);
-void copy_inplace(const array& src, array& dst, CopyType ctype);
-
-void copy_inplace(
-    const array& src,
-    array& dst,
-    const Shape& data_shape,
-    const Strides& i_strides,
-    const Strides& o_strides,
-    int64_t i_offset,
-    int64_t o_offset,
-    CopyType ctype);
+inline bool set_copy_output_data(const array& in, array& out, CopyType ctype) {
+  if (ctype == CopyType::Vector) {
+    // If the input is donateable, we are doing a vector copy and the types
+    // have the same size, then the input buffer can hold the output.
+    if (in.is_donatable() && in.itemsize() == out.itemsize()) {
+      out.copy_shared_buffer(in);
+      return true;
+    } else {
+      out.set_data(
+          allocator::malloc_or_wait(in.data_size() * out.itemsize()),
+          in.data_size(),
+          in.strides(),
+          in.flags());
+      return false;
+    }
+  } else {
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    return false;
+  }
+}
 
 } // namespace mlx::core

mlx/backend/common/default_primitives.cpp

@@ -1,198 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include <cstring>
-
-#include "mlx/array.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/lapack.h"
-#include "mlx/backend/common/utils.h"
-#include "mlx/primitives.h"
-
-#define DEFAULT(primitive)                                                 \
-  void primitive::eval_cpu(const std::vector<array>& inputs, array& out) { \
-    primitive::eval(inputs, out);                                          \
-  }
-
-#define DEFAULT_MULTI(primitive)                                       \
-  void primitive::eval_cpu(                                            \
-      const std::vector<array>& inputs, std::vector<array>& outputs) { \
-    primitive::eval(inputs, outputs);                                  \
-  }
-
-namespace mlx::core {
-
-DEFAULT(Abs)
-DEFAULT(Add)
-DEFAULT(Arange)
-DEFAULT(ArcCos)
-DEFAULT(ArcCosh)
-DEFAULT(ArcSin)
-DEFAULT(ArcSinh)
-DEFAULT(ArcTan)
-DEFAULT(ArcTan2)
-DEFAULT(ArcTanh)
-DEFAULT(ArgPartition)
-DEFAULT(ArgReduce)
-DEFAULT(ArgSort)
-DEFAULT(AsType)
-DEFAULT(AsStrided)
-DEFAULT(Broadcast)
-DEFAULT(BroadcastAxes)
-DEFAULT(BlockMaskedMM)
-DEFAULT(GatherMM)
-DEFAULT(GatherQMM)
-DEFAULT_MULTI(DivMod)
-DEFAULT(Ceil)
-DEFAULT(Concatenate)
-DEFAULT(Conjugate)
-DEFAULT(Convolution)
-DEFAULT(Copy)
-DEFAULT(Cos)
-DEFAULT(Cosh)
-DEFAULT_MULTI(CustomTransforms)
-DEFAULT_MULTI(Depends)
-DEFAULT(Divide)
-DEFAULT(NumberOfElements)
-DEFAULT(Remainder)
-DEFAULT(Equal)
-DEFAULT(Erf)
-DEFAULT(ErfInv)
-DEFAULT(Exp)
-DEFAULT(ExpandDims)
-DEFAULT(Expm1)
-DEFAULT(FFT)
-DEFAULT(Floor)
-DEFAULT(Full)
-DEFAULT(Gather)
-DEFAULT(Greater)
-DEFAULT(GreaterEqual)
-DEFAULT(Hadamard)
-DEFAULT(Less)
-DEFAULT(LessEqual)
-DEFAULT(Load)
-DEFAULT(Log)
-DEFAULT(Log1p)
-DEFAULT(LogicalNot)
-DEFAULT(LogicalAnd)
-DEFAULT(LogicalOr)
-DEFAULT(LogAddExp)
-DEFAULT(Maximum)
-DEFAULT(Minimum)
-DEFAULT(Multiply)
-DEFAULT(Negative)
-DEFAULT(NotEqual)
-DEFAULT(Pad)
-DEFAULT(Partition)
-DEFAULT(Power)
-DEFAULT_MULTI(QRF)
-DEFAULT(QuantizedMatmul)
-DEFAULT(RandomBits)
-DEFAULT(Reduce)
-DEFAULT(Round)
-DEFAULT(Scan)
-DEFAULT(Scatter)
-DEFAULT(Select)
-DEFAULT(Sigmoid)
-DEFAULT(Sign)
-DEFAULT(Sin)
-DEFAULT(Sinh)
-DEFAULT(Slice)
-DEFAULT(SliceUpdate)
-DEFAULT(Softmax)
-DEFAULT(Sort)
-DEFAULT_MULTI(Split)
-DEFAULT(Square)
-DEFAULT(Squeeze)
-DEFAULT(Sqrt)
-DEFAULT(StopGradient)
-DEFAULT(Subtract)
-DEFAULT_MULTI(SVD)
-DEFAULT(Tan)
-DEFAULT(Tanh)
-DEFAULT(Transpose)
-DEFAULT(Inverse)
-DEFAULT(Cholesky)
-DEFAULT_MULTI(Eigh)
-
-namespace {
-
-inline void matmul_common_general(
-    const array& a_pre,
-    const array& b_pre,
-    array& out,
-    float alpha = 1.0f,
-    float beta = 0.0f) {
-  auto check_transpose = [](const array& arr) {
-    auto stx = arr.strides()[arr.ndim() - 2];
-    auto sty = arr.strides()[arr.ndim() - 1];
-    if (stx == arr.shape(-1) && sty == 1) {
-      return std::make_tuple(false, stx, arr);
-    } else if (stx == 1 && sty == arr.shape(-2)) {
-      return std::make_tuple(true, sty, arr);
-    } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
-      stx = arr.shape(-1);
-      return std::make_tuple(false, stx, arr_copy);
-    }
-  };
-
-  auto [a_transposed, lda, a] = check_transpose(a_pre);
-  auto [b_transposed, ldb, b] = check_transpose(b_pre);
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
-  if (M == 0 || N == 0) {
-    return;
-  }
-  if (K == 0) {
-    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
-    return;
-  }
-
-  for (int i = 0; i < (a.size() / (M * K)); ++i) {
-    cblas_sgemm(
-        CblasRowMajor,
-        a_transposed ? CblasTrans : CblasNoTrans, // transA
-        b_transposed ? CblasTrans : CblasNoTrans, // transB
-        M,
-        N,
-        K,
-        alpha, // alpha
-        a.data<float>() + elem_to_loc(M * K * i, a.shape(), a.strides()),
-        lda,
-        b.data<float>() + elem_to_loc(K * N * i, b.shape(), b.strides()),
-        ldb,
-        beta, // beta
-        out.data<float>() + M * N * i,
-        out.shape(-1) // ldc
-    );
-  }
-}
-
-} // namespace
-
-void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {
-  if (out.dtype() != float32) {
-    throw std::runtime_error(
-        "[Matmul::eval_cpu] Currently only supports float32.");
-  }
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  return matmul_common_general(inputs[0], inputs[1], out);
-}
-
-void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  if (out.dtype() != float32) {
-    throw std::runtime_error(
-        "[AddMM::eval_cpu] Currently only supports float32.");
-  }
-
-  // Fill output with C
-  auto& c = inputs[2];
-  CopyType ctype = c.data_size() == 1 ? CopyType::Scalar : CopyType::General;
-  copy(c, out, ctype);
-
-  return matmul_common_general(inputs[0], inputs[1], out, alpha_, beta_);
-}
-
-} // namespace mlx::core

mlx/backend/common/eigh.cpp

@@ -1,117 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include "mlx/allocator.h"
-#include "mlx/array.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/lapack.h"
-#include "mlx/linalg.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-namespace {
-
-void ssyevd(
-    char jobz,
-    char uplo,
-    float* a,
-    int N,
-    float* w,
-    float* work,
-    int lwork,
-    int* iwork,
-    int liwork) {
-  int info;
-  MLX_LAPACK_FUNC(ssyevd)
-  (
-      /* jobz = */ &jobz,
-      /* uplo = */ &uplo,
-      /* n = */ &N,
-      /* a = */ a,
-      /* lda = */ &N,
-      /* w = */ w,
-      /* work = */ work,
-      /* lwork = */ &lwork,
-      /* iwork = */ iwork,
-      /* liwork = */ &liwork,
-      /* info = */ &info);
-  if (info != 0) {
-    std::stringstream msg;
-    msg << "[Eigh::eval_cpu] Eigenvalue decomposition failed with error code "
-        << info;
-    throw std::runtime_error(msg.str());
-  }
-}
-
-} // namespace
-
-void Eigh::eval(const std::vector<array>& inputs, std::vector<array>& outputs) {
-  const auto& a = inputs[0];
-  auto& values = outputs[0];
-
-  auto vectors = compute_eigenvectors_
-      ? outputs[1]
-      : array(a.shape(), a.dtype(), nullptr, {});
-
-  values.set_data(allocator::malloc_or_wait(values.nbytes()));
-
-  copy(
-      a,
-      vectors,
-      a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-
-  if (compute_eigenvectors_) {
-    // Set the strides and flags so the eigenvectors
-    // are in the columns of the output
-    auto flags = vectors.flags();
-    auto strides = vectors.strides();
-    auto ndim = a.ndim();
-    std::swap(strides[ndim - 1], strides[ndim - 2]);
-
-    if (a.size() > 1) {
-      flags.row_contiguous = false;
-      if (ndim > 2) {
-        flags.col_contiguous = false;
-      } else {
-        flags.col_contiguous = true;
-      }
-    }
-    vectors.move_shared_buffer(vectors, strides, flags, vectors.data_size());
-  }
-
-  auto vec_ptr = vectors.data<float>();
-  auto eig_ptr = values.data<float>();
-
-  char jobz = compute_eigenvectors_ ? 'V' : 'N';
-  auto N = a.shape(-1);
-
-  // Work query
-  int lwork;
-  int liwork;
-  {
-    float work;
-    int iwork;
-    ssyevd(jobz, uplo_[0], nullptr, N, nullptr, &work, -1, &iwork, -1);
-    lwork = static_cast<int>(work);
-    liwork = iwork;
-  }
-
-  auto work_buf = array::Data{allocator::malloc_or_wait(sizeof(float) * lwork)};
-  auto iwork_buf = array::Data{allocator::malloc_or_wait(sizeof(int) * liwork)};
-  for (size_t i = 0; i < a.size() / (N * N); ++i) {
-    ssyevd(
-        jobz,
-        uplo_[0],
-        vec_ptr,
-        N,
-        eig_ptr,
-        static_cast<float*>(work_buf.buffer.raw_ptr()),
-        lwork,
-        static_cast<int*>(iwork_buf.buffer.raw_ptr()),
-        liwork);
-    vec_ptr += N * N;
-    eig_ptr += N;
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/erf.cpp

@@ -1,40 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cmath>
-
-namespace mlx::core {
-
-/* Approximation to the inverse error function.
- * Based on code from:
- *   https://stackoverflow.com/questions/27229371/inverse-error-function-in-c#answer-49743348
- */
-float erfinv(float a) {
-  auto t = std::fma(a, 0.0f - a, 1.0f);
-  t = std::log(t);
-  float p;
-  if (std::abs(t) > 6.125f) { // maximum ulp error = 2.35793
-    p = 3.03697567e-10f; //  0x1.4deb44p-32
-    p = std::fma(p, t, 2.93243101e-8f); //  0x1.f7c9aep-26
-    p = std::fma(p, t, 1.22150334e-6f); //  0x1.47e512p-20
-    p = std::fma(p, t, 2.84108955e-5f); //  0x1.dca7dep-16
-    p = std::fma(p, t, 3.93552968e-4f); //  0x1.9cab92p-12
-    p = std::fma(p, t, 3.02698812e-3f); //  0x1.8cc0dep-9
-    p = std::fma(p, t, 4.83185798e-3f); //  0x1.3ca920p-8
-    p = std::fma(p, t, -2.64646143e-1f); // -0x1.0eff66p-2
-    p = std::fma(p, t, 8.40016484e-1f); //  0x1.ae16a4p-1
-  } else { // maximum ulp error = 2.35002
-    p = 5.43877832e-9f; //  0x1.75c000p-28
-    p = std::fma(p, t, 1.43285448e-7f); //  0x1.33b402p-23
-    p = std::fma(p, t, 1.22774793e-6f); //  0x1.499232p-20
-    p = std::fma(p, t, 1.12963626e-7f); //  0x1.e52cd2p-24
-    p = std::fma(p, t, -5.61530760e-5f); // -0x1.d70bd0p-15
-    p = std::fma(p, t, -1.47697632e-4f); // -0x1.35be90p-13
-    p = std::fma(p, t, 2.31468678e-3f); //  0x1.2f6400p-9
-    p = std::fma(p, t, 1.15392581e-2f); //  0x1.7a1e50p-7
-    p = std::fma(p, t, -2.32015476e-1f); // -0x1.db2aeep-3
-    p = std::fma(p, t, 8.86226892e-1f); //  0x1.c5bf88p-1
-  }
-  return a * p;
-}
-
-} // namespace mlx::core

mlx/backend/common/fft.cpp

@@ -1,87 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <numeric>
-
-#include "mlx/3rdparty/pocketfft.h"
-#include "mlx/allocator.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-void FFT::eval(const std::vector<array>& inputs, array& out) {
-  auto& in = inputs[0];
-  std::vector<std::ptrdiff_t> strides_in(
-      in.strides().begin(), in.strides().end());
-  for (auto& s : strides_in) {
-    s *= in.itemsize();
-  }
-  std::vector<std::ptrdiff_t> strides_out(
-      out.strides().begin(), out.strides().end());
-  for (auto& s : strides_out) {
-    s *= out.itemsize();
-  }
-
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  std::vector<size_t> shape;
-  if (out.dtype() == float32) {
-    shape.insert(shape.end(), out.shape().begin(), out.shape().end());
-  } else {
-    shape.insert(shape.end(), in.shape().begin(), in.shape().end());
-  }
-
-  float scale = 1.0f;
-  if (inverse_) {
-    size_t nelem = std::accumulate(
-        axes_.begin(), axes_.end(), 1, [&shape](auto x, auto y) {
-          return x * shape[y];
-        });
-    scale /= nelem;
-  }
-  if (in.dtype() == complex64 && out.dtype() == complex64) {
-    auto in_ptr =
-        reinterpret_cast<const std::complex<float>*>(in.data<complex64_t>());
-    auto out_ptr =
-        reinterpret_cast<std::complex<float>*>(out.data<complex64_t>());
-    pocketfft::c2c(
-        shape,
-        strides_in,
-        strides_out,
-        axes_,
-        !inverse_,
-        in_ptr,
-        out_ptr,
-        scale);
-  } else if (in.dtype() == float32 && out.dtype() == complex64) {
-    auto in_ptr = in.data<float>();
-    auto out_ptr =
-        reinterpret_cast<std::complex<float>*>(out.data<complex64_t>());
-    pocketfft::r2c(
-        shape,
-        strides_in,
-        strides_out,
-        axes_,
-        !inverse_,
-        in_ptr,
-        out_ptr,
-        scale);
-  } else if (in.dtype() == complex64 && out.dtype() == float32) {
-    auto in_ptr =
-        reinterpret_cast<const std::complex<float>*>(in.data<complex64_t>());
-    auto out_ptr = out.data<float>();
-    pocketfft::c2r(
-        shape,
-        strides_in,
-        strides_out,
-        axes_,
-        !inverse_,
-        in_ptr,
-        out_ptr,
-        scale);
-  } else {
-    throw std::runtime_error(
-        "[FFT] Received unexpected input and output type combination.");
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/indexing.cpp

@@ -1,393 +0,0 @@
-// Copyright © 2023 Apple Inc.
-#include <algorithm>
-#include <cassert>
-#include <cmath>
-
-#include "mlx/allocator.h"
-#include "mlx/primitives.h"
-
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/utils.h"
-
-namespace mlx::core {
-
-template <typename IdxT>
-inline size_t offset_neg_idx(IdxT idx, size_t size) {
-  return (idx < 0) ? idx + size : idx;
-}
-
-template <>
-inline size_t offset_neg_idx(bool idx, size_t) {
-  return idx;
-}
-
-template <>
-inline size_t offset_neg_idx(uint32_t idx, size_t) {
-  return idx;
-}
-
-template <typename T, typename IdxT>
-void gather(
-    const array& src,
-    const std::vector<array>& inds,
-    array& out,
-    const std::vector<int>& axes,
-    const Shape& slice_sizes) {
-  // If the array is row contiguous then we can do a contiguous copy given
-  // two conditions on the slice size:
-  // - Any number of leading ones in the slice sizes are allowed
-  // - All other slice sizes match the corresponding dimension except the
-  //   first non-singleton slice size
-  // If the array is col contiguous then the reverse is the case:
-  // - Any number of trailing ones in the slice sizes are allowed
-  // - All other slice sizes match the corresponding dimension except the
-  //   first non-singleton slice size from the end
-
-  bool can_copy = false;
-  if (src.flags().row_contiguous) {
-    can_copy = true;
-
-    // Ignore leading 1s
-    int i = 0;
-    for (; i < slice_sizes.size() && slice_sizes[i] == 1; ++i)
-      ;
-
-    // Check the remaining
-    i++;
-    for (; i < src.ndim() && can_copy; ++i) {
-      can_copy = (src.shape(i) == slice_sizes[i]);
-    }
-  } else if (src.flags().col_contiguous) {
-    can_copy = true;
-
-    // Ignore trailing 1s
-    int i = slice_sizes.size() - 1;
-    for (; i >= 0 && slice_sizes[i] == 1; --i)
-      ;
-
-    // Skip the next slice size and check the remaining
-    i--;
-    for (; i >= 0 && can_copy; --i) {
-      can_copy = (src.shape(i) == slice_sizes[i]);
-    }
-  }
-  size_t slice_size = 1;
-  for (auto s : slice_sizes) {
-    slice_size *= s;
-  }
-  size_t ind_size = slice_size == 0 ? 0 : out.size() / slice_size;
-  const T* src_ptr = src.data<T>();
-  T* dst_ptr = out.data<T>();
-  size_t out_idx = 0;
-
-  std::vector<ContiguousIterator> its(inds.begin(), inds.end());
-  ContiguousIterator src_it;
-  if (!can_copy && src.ndim() > 0) {
-    src_it = ContiguousIterator(slice_sizes, src.strides(), src.ndim());
-  }
-  for (int idx = 0; idx < ind_size; idx++) {
-    size_t src_idx = 0;
-    for (int ii = 0; ii < inds.size(); ++ii) {
-      auto ax = axes[ii];
-      auto idx_loc = its[ii].loc;
-      its[ii].step();
-      auto idx_val =
-          offset_neg_idx(inds[ii].data<IdxT>()[idx_loc], src.shape(ax));
-      src_idx += (idx_val * src.strides()[ax]);
-    }
-
-    if (slice_size == 1) {
-      dst_ptr[out_idx++] = src_ptr[src_idx];
-    } else if (can_copy) {
-      std::copy(
-          src_ptr + src_idx, src_ptr + src_idx + slice_size, dst_ptr + out_idx);
-      out_idx += slice_size;
-    } else {
-      for (int jj = 0; jj < slice_size; jj++) {
-        dst_ptr[out_idx++] = src_ptr[src_idx + src_it.loc];
-        src_it.step();
-      }
-      src_it.reset();
-    }
-  }
-}
-
-template <typename IdxT>
-void dispatch_gather(
-    const array& src,
-    const std::vector<array>& inds,
-    array& out,
-    const std::vector<int>& axes,
-    const Shape& size) {
-  switch (out.dtype()) {
-    case bool_:
-      gather<bool, IdxT>(src, inds, out, axes, size);
-      break;
-    case uint8:
-      gather<uint8_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case uint16:
-      gather<uint16_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case uint32:
-      gather<uint32_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case uint64:
-      gather<uint64_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case int8:
-      gather<int8_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case int16:
-      gather<int16_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case int32:
-      gather<int32_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case int64:
-      gather<int64_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case float16:
-      gather<float16_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case float32:
-      gather<float, IdxT>(src, inds, out, axes, size);
-      break;
-    case bfloat16:
-      gather<bfloat16_t, IdxT>(src, inds, out, axes, size);
-      break;
-    case complex64:
-      gather<complex64_t, IdxT>(src, inds, out, axes, size);
-      break;
-  }
-}
-
-void Gather::eval(const std::vector<array>& inputs, array& out) {
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  auto& src = inputs[0];
-  std::vector<array> inds(inputs.begin() + 1, inputs.end());
-
-  if (inds.empty()) {
-    dispatch_gather<bool>(src, inds, out, axes_, slice_sizes_);
-    return;
-  }
-
-  switch (inds[0].dtype()) {
-    case bool_:
-      dispatch_gather<bool>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case uint8:
-      dispatch_gather<uint8_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case uint16:
-      dispatch_gather<uint16_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case uint32:
-      dispatch_gather<uint32_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case uint64:
-      dispatch_gather<uint64_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case int8:
-      dispatch_gather<int8_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case int16:
-      dispatch_gather<int16_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case int32:
-      dispatch_gather<int32_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case int64:
-      dispatch_gather<int64_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case float16:
-    case float32:
-    case bfloat16:
-    case complex64:
-      throw std::runtime_error(
-          "[Gather::eval] Cannot gather with floating point indices.");
-      break;
-  }
-}
-
-template <typename InT, typename IdxT, typename OpT>
-void scatter(
-    const array& updates,
-    array& out,
-    const std::vector<array>& inds,
-    const std::vector<int>& axes,
-    const OpT& op) {
-  int nind = inds.size();
-  auto inds_ndim = updates.ndim() - out.ndim();
-  size_t n_updates = nind ? inds[0].size() : 1;
-
-  Shape update_shape(
-      updates.shape().begin() + inds_ndim, updates.shape().end());
-  size_t update_size = 1;
-  for (auto us : update_shape) {
-    update_size *= us;
-  }
-
-  std::vector<ContiguousIterator> its(inds.begin(), inds.end());
-  ContiguousIterator update_it(updates);
-  ContiguousIterator out_it(update_shape, out.strides(), out.ndim());
-
-  for (int i = 0; i < n_updates; ++i) {
-    size_t out_offset = 0;
-    for (int j = 0; j < nind; ++j) {
-      auto ax = axes[j];
-      auto idx_loc = its[j].loc;
-      its[j].step();
-      auto idx_val =
-          offset_neg_idx(inds[j].data<IdxT>()[idx_loc], out.shape(ax));
-      out_offset += (idx_val * out.strides()[ax]);
-    }
-    update_it.seek(i * update_size);
-    for (int j = 0; j < update_size; ++j) {
-      op(updates.data<InT>()[update_it.loc],
-         out.data<InT>() + out_offset + out_it.loc);
-      update_it.step();
-      out_it.step();
-    }
-    out_it.reset();
-    update_it.reset();
-  }
-}
-
-template <typename InT, typename IdxT>
-void dispatch_scatter_inds(
-    array& out,
-    const std::vector<array>& indices,
-    const array& updates,
-    const std::vector<int>& axes,
-    Scatter::ReduceType rtype) {
-  switch (rtype) {
-    case Scatter::None:
-      scatter<InT, IdxT>(
-          updates, out, indices, axes, [](auto x, auto* y) { (*y) = x; });
-      break;
-    case Scatter::Sum:
-      scatter<InT, IdxT>(
-          updates, out, indices, axes, [](auto x, auto* y) { (*y) += x; });
-      break;
-    case Scatter::Prod:
-      scatter<InT, IdxT>(
-          updates, out, indices, axes, [](auto x, auto* y) { (*y) *= x; });
-      break;
-    case Scatter::Max:
-      scatter<InT, IdxT>(updates, out, indices, axes, [](auto x, auto* y) {
-        (*y) = (*y > x) ? *y : x;
-      });
-      break;
-    case Scatter::Min:
-      scatter<InT, IdxT>(updates, out, indices, axes, [](auto x, auto* y) {
-        (*y) = (*y < x) ? *y : x;
-      });
-      break;
-  }
-}
-
-template <typename InT>
-void dispatch_scatter(
-    array& out,
-    const std::vector<array>& inds,
-    const array& updates,
-    const std::vector<int>& axes,
-    Scatter::ReduceType rtype) {
-  if (inds.empty()) {
-    dispatch_scatter_inds<InT, bool>(out, inds, updates, axes, rtype);
-    return;
-  }
-
-  switch (inds[0].dtype()) {
-    case bool_:
-      dispatch_scatter_inds<InT, bool>(out, inds, updates, axes, rtype);
-      break;
-    case uint8:
-      dispatch_scatter_inds<InT, uint8_t>(out, inds, updates, axes, rtype);
-      break;
-    case uint16:
-      dispatch_scatter_inds<InT, uint16_t>(out, inds, updates, axes, rtype);
-      break;
-    case uint32:
-      dispatch_scatter_inds<InT, uint32_t>(out, inds, updates, axes, rtype);
-      break;
-    case uint64:
-      dispatch_scatter_inds<InT, uint64_t>(out, inds, updates, axes, rtype);
-      break;
-    case int8:
-      dispatch_scatter_inds<InT, int8_t>(out, inds, updates, axes, rtype);
-      break;
-    case int16:
-      dispatch_scatter_inds<InT, int16_t>(out, inds, updates, axes, rtype);
-      break;
-    case int32:
-      dispatch_scatter_inds<InT, int32_t>(out, inds, updates, axes, rtype);
-      break;
-    case int64:
-      dispatch_scatter_inds<InT, int64_t>(out, inds, updates, axes, rtype);
-      break;
-    case float16:
-    case float32:
-    case bfloat16:
-    case complex64:
-      throw std::runtime_error(
-          "[Scatter::eval_cpu] Cannot scatter with floating point indices.");
-  }
-}
-
-void Scatter::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() >= 2);
-
-  auto& src = inputs[0];
-  std::vector<array> inds(inputs.begin() + 1, inputs.end() - 1);
-  auto& updates = inputs.back();
-
-  // Copy src into out (copy allocates memory for out)
-  copy(src, out, CopyType::General);
-
-  switch (src.dtype()) {
-    case bool_:
-      dispatch_scatter<bool>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case uint8:
-      dispatch_scatter<uint8_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case uint16:
-      dispatch_scatter<uint16_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case uint32:
-      dispatch_scatter<uint32_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case uint64:
-      dispatch_scatter<uint64_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case int8:
-      dispatch_scatter<int8_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case int16:
-      dispatch_scatter<int16_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case int32:
-      dispatch_scatter<int32_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case int64:
-      dispatch_scatter<int64_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case float16:
-      dispatch_scatter<float16_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case float32:
-      dispatch_scatter<float>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case bfloat16:
-      dispatch_scatter<bfloat16_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case complex64:
-      dispatch_scatter<complex64_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/inverse.cpp

@@ -1,120 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include "mlx/allocator.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/lapack.h"
-#include "mlx/primitives.h"
-
-int strtri_wrapper(char uplo, char diag, float* matrix, int N) {
-  int info;
-  MLX_LAPACK_FUNC(strtri)
-  (
-      /* uplo = */ &uplo,
-      /* diag = */ &diag,
-      /* N = */ &N,
-      /* a = */ matrix,
-      /* lda = */ &N,
-      /* info = */ &info);
-  return info;
-}
-
-namespace mlx::core {
-
-void general_inv(array& inv, int N, int i) {
-  int info;
-  auto ipiv = array::Data{allocator::malloc_or_wait(sizeof(int) * N)};
-  // Compute LU factorization.
-  sgetrf_(
-      /* m = */ &N,
-      /* n = */ &N,
-      /* a = */ inv.data<float>() + N * N * i,
-      /* lda = */ &N,
-      /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
-      /* info = */ &info);
-
-  if (info != 0) {
-    std::stringstream ss;
-    ss << "inverse_impl: LU factorization failed with error code " << info;
-    throw std::runtime_error(ss.str());
-  }
-
-  static const int lwork_query = -1;
-  float workspace_size = 0;
-
-  // Compute workspace size.
-  sgetri_(
-      /* m = */ &N,
-      /* a = */ nullptr,
-      /* lda = */ &N,
-      /* ipiv = */ nullptr,
-      /* work = */ &workspace_size,
-      /* lwork = */ &lwork_query,
-      /* info = */ &info);
-
-  if (info != 0) {
-    std::stringstream ss;
-    ss << "inverse_impl: LU workspace calculation failed with error code "
-       << info;
-    throw std::runtime_error(ss.str());
-  }
-
-  const int lwork = workspace_size;
-  auto scratch = array::Data{allocator::malloc_or_wait(sizeof(float) * lwork)};
-
-  // Compute inverse.
-  sgetri_(
-      /* m = */ &N,
-      /* a = */ inv.data<float>() + N * N * i,
-      /* lda = */ &N,
-      /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
-      /* work = */ static_cast<float*>(scratch.buffer.raw_ptr()),
-      /* lwork = */ &lwork,
-      /* info = */ &info);
-
-  if (info != 0) {
-    std::stringstream ss;
-    ss << "inverse_impl: inversion failed with error code " << info;
-    throw std::runtime_error(ss.str());
-  }
-}
-
-void tri_inv(array& inv, int N, int i, bool upper) {
-  const char uplo = upper ? 'L' : 'U';
-  const char diag = 'N';
-  int info = strtri_wrapper(uplo, diag, inv.data<float>() + N * N * i, N);
-  if (info != 0) {
-    std::stringstream ss;
-    ss << "inverse_impl: triangular inversion failed with error code " << info;
-    throw std::runtime_error(ss.str());
-  }
-}
-
-void inverse_impl(const array& a, array& inv, bool tri, bool upper) {
-  // Lapack uses the column-major convention. We take advantage of the following
-  // identity to avoid transposing (see
-  // https://math.stackexchange.com/a/340234):
-  //   (A⁻¹)ᵀ = (Aᵀ)⁻¹
-
-  // The inverse is computed in place, so just copy the input to the output.
-  copy(a, inv, a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-
-  const int N = a.shape(-1);
-  const size_t num_matrices = a.size() / (N * N);
-
-  for (int i = 0; i < num_matrices; i++) {
-    if (tri) {
-      tri_inv(inv, N, i, upper);
-    } else {
-      general_inv(inv, N, i);
-    }
-  }
-}
-
-void Inverse::eval(const std::vector<array>& inputs, array& output) {
-  if (inputs[0].dtype() != float32) {
-    throw std::runtime_error("[Inverse::eval] only supports float32.");
-  }
-  inverse_impl(inputs[0], output, tri_, upper_);
-}
-
-} // namespace mlx::core

mlx/backend/common/load.cpp

@@ -1,12 +1,10 @@
 // Copyright © 2023 Apple Inc.
 
 #include <algorithm>
-#include <cassert>
 #include <utility>
 
-#include "mlx/allocator.h"
-#include "mlx/backend/common/load.h"
 #include "mlx/primitives.h"
+#include "mlx/scheduler.h"
 
 namespace {
 
@@ -29,33 +27,31 @@ void swap_endianness(uint8_t* data_bytes, size_t N) {
 
 namespace mlx::core {
 
-void load(
-    array& out,
-    size_t offset,
-    const std::shared_ptr<io::Reader>& reader,
-    bool swap_endianness_) {
-  reader->read(out.data<char>(), out.nbytes(), offset);
-
-  if (swap_endianness_) {
-    switch (out.itemsize()) {
-      case 2:
-        swap_endianness<2>(out.data<uint8_t>(), out.data_size());
-        break;
-      case 4:
-        swap_endianness<4>(out.data<uint8_t>(), out.data_size());
-        break;
-      case 8:
-        swap_endianness<8>(out.data<uint8_t>(), out.data_size());
-        break;
-    }
-  }
-}
-
-void Load::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 0);
+void Load::eval_cpu(const std::vector<array>& inputs, array& out) {
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  load(out, offset_, reader_, swap_endianness_);
+  auto read_task = [out_ptr = out.data<char>(),
+                    size = out.size(),
+                    itemsize = out.itemsize(),
+                    offset = offset_,
+                    reader = reader_,
+                    swap_endianness_ = swap_endianness_]() mutable {
+    reader->read(out_ptr, size * itemsize, offset);
+    if (swap_endianness_) {
+      switch (itemsize) {
+        case 2:
+          swap_endianness<2>(reinterpret_cast<uint8_t*>(out_ptr), size);
+          break;
+        case 4:
+          swap_endianness<4>(reinterpret_cast<uint8_t*>(out_ptr), size);
+          break;
+        case 8:
+          swap_endianness<8>(reinterpret_cast<uint8_t*>(out_ptr), size);
+          break;
+      }
+    }
+  };
+  auto fut = io::thread_pool().enqueue(std::move(read_task)).share();
+  scheduler::enqueue(stream(), [fut = std::move(fut)]() { fut.wait(); });
 }
 
 } // namespace mlx::core

mlx/backend/common/load.h

@@ -1,14 +0,0 @@
-// Copyright © 2024 Apple Inc.
-
-#include "mlx/array.h"
-#include "mlx/io/load.h"
-
-namespace mlx::core {
-
-void load(
-    array& out,
-    size_t offset,
-    const std::shared_ptr<io::Reader>& reader,
-    bool swap_endianess);
-
-} // namespace mlx::core

mlx/backend/common/masked_mm.cpp

@@ -1,300 +0,0 @@
-// Copyright © 2024 Apple Inc.
-
-#include <cstring>
-
-#include "mlx/array.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/lapack.h"
-#include "mlx/backend/common/utils.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-namespace {
-
-template <typename T, typename mask_t>
-inline void mask_matrix(
-    T* data,
-    const mask_t* mask,
-    int block_size,
-    const int X,
-    const int Y,
-    const int64_t X_data_str,
-    const int64_t Y_data_str,
-    const int64_t X_mask_str,
-    const int64_t Y_mask_str,
-    const size_t mask_offset) {
-  int tX = (X + block_size - 1) / block_size;
-  int tY = (Y + block_size - 1) / block_size;
-
-  for (int i = 0; i < tX; i++) {
-    for (int j = 0; j < tY; j++) {
-      mask_t do_mask = mask[mask_offset + i * X_mask_str + j * Y_mask_str];
-      if (do_mask != 1) {
-        int loc_x = i * block_size;
-        int loc_y = j * block_size;
-        T* data_block = data + loc_x * X_data_str + loc_y * Y_data_str;
-
-        int size_x = std::min(block_size, X - loc_x);
-        int size_y = std::min(block_size, Y - loc_y);
-        for (int ii = 0; ii < size_x; ii++) {
-          for (int jj = 0; jj < size_y; jj++) {
-            if constexpr (std::is_same_v<mask_t, bool>) {
-              data_block[ii * X_data_str + jj * Y_data_str] = T(0.);
-            } else {
-              data_block[ii * X_data_str + jj * Y_data_str] *= do_mask;
-            }
-          }
-        }
-      }
-    }
-  }
-}
-
-} // namespace
-
-void BlockMaskedMM::eval(const std::vector<array>& inputs, array& out) {
-  if (out.dtype() != float32) {
-    throw std::runtime_error(
-        "[BlockMaskedMM::eval] Currently only supports float32.");
-  }
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  auto& a_pre = inputs[0];
-  auto& b_pre = inputs[1];
-
-  auto check_transpose =
-      [](const array& arr, bool do_copy, bool expand_all = false) {
-        auto stx = arr.strides()[arr.ndim() - 2];
-        auto sty = arr.strides()[arr.ndim() - 1];
-        if (!expand_all && stx == arr.shape(-1) && sty == 1) {
-          if (do_copy) {
-            array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector);
-            return std::make_tuple(false, stx, arr_copy);
-          }
-          return std::make_tuple(false, stx, arr);
-        } else if (!expand_all && stx == 1 && sty == arr.shape(-2)) {
-          if (do_copy) {
-            array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector);
-            return std::make_tuple(true, sty, arr_copy);
-          }
-          return std::make_tuple(true, sty, arr);
-        } else {
-          array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-          copy(arr, arr_copy, CopyType::General);
-          int64_t stx = arr.shape(-1);
-          return std::make_tuple(false, stx, arr_copy);
-        }
-      };
-
-  bool has_op_mask = inputs.size() > 3;
-  bool has_out_mask = inputs.size() == 3 || inputs.size() == 5;
-  auto [a_transposed, lda, a] =
-      check_transpose(a_pre, has_op_mask, inputs.back().dtype() != bool_);
-  auto [b_transposed, ldb, b] =
-      check_transpose(b_pre, has_op_mask, inputs.back().dtype() != bool_);
-
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
-
-  if (M == 0 || N == 0) {
-    return;
-  }
-
-  if (K == 0) {
-    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
-    return;
-  }
-
-  auto mask_array = [](const array& mask,
-                       float* data,
-                       int block_size,
-                       int batch_idx,
-                       int X,
-                       int Y,
-                       size_t X_data_str,
-                       size_t Y_data_str) {
-    auto mask_offset = elem_to_loc(
-        mask.shape(-1) * mask.shape(-2) * batch_idx,
-        mask.shape(),
-        mask.strides());
-
-    auto X_mask_str = mask.strides()[mask.ndim() - 2];
-    auto Y_mask_str = mask.strides()[mask.ndim() - 1];
-
-    if (mask.dtype() == bool_) {
-      return mask_matrix(
-          data,
-          mask.data<bool>(),
-          block_size,
-          X,
-          Y,
-          X_data_str,
-          Y_data_str,
-          X_mask_str,
-          Y_mask_str,
-          mask_offset);
-    } else {
-      return mask_matrix(
-          data,
-          mask.data<float>(),
-          block_size,
-          X,
-          Y,
-          X_data_str,
-          Y_data_str,
-          X_mask_str,
-          Y_mask_str,
-          mask_offset);
-    }
-  };
-
-  for (int i = 0; i < (out.size() / (M * size_t(N))); ++i) {
-    // Adjust pointer
-    float* ai =
-        a.data<float>() + elem_to_loc(M * K * i, a.shape(), a.strides());
-    float* bi =
-        b.data<float>() + elem_to_loc(K * N * i, b.shape(), b.strides());
-    float* ci = out.data<float>() + M * N * i;
-
-    // Zero out blocks in a and b if needed
-    if (has_op_mask) {
-      auto& a_mask = inputs[inputs.size() - 2];
-      mask_array(
-          a_mask,
-          ai,
-          block_size_,
-          i,
-          M,
-          K,
-          a_transposed ? 1 : lda,
-          a_transposed ? lda : 1);
-
-      auto& b_mask = inputs[inputs.size() - 1];
-      mask_array(
-          b_mask,
-          bi,
-          block_size_,
-          i,
-          K,
-          N,
-          b_transposed ? 1 : ldb,
-          b_transposed ? ldb : 1);
-    }
-
-    // Do matmul
-    cblas_sgemm(
-        CblasRowMajor,
-        a_transposed ? CblasTrans : CblasNoTrans, // transA
-        b_transposed ? CblasTrans : CblasNoTrans, // transB
-        M,
-        N,
-        K,
-        1.0, // alpha
-        ai,
-        lda,
-        bi,
-        ldb,
-        0.0, // beta
-        ci,
-        out.shape(-1) // ldc
-    );
-
-    // Zero out blocks in out
-    if (has_out_mask) {
-      mask_array(inputs[2], ci, block_size_, i, M, N, N, 1);
-    }
-  }
-}
-
-void GatherMM::eval(const std::vector<array>& inputs, array& out) {
-  if (out.dtype() != float32) {
-    throw std::runtime_error(
-        "[GatherMM::eval] Currently only supports float32.");
-  }
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  auto& a_pre = inputs[0];
-  auto& b_pre = inputs[1];
-
-  auto check_transpose = [](const array& arr) {
-    auto stx = arr.strides()[arr.ndim() - 2];
-    auto sty = arr.strides()[arr.ndim() - 1];
-    if (stx == arr.shape(-1) && sty == 1) {
-      return std::make_tuple(false, stx, arr);
-    } else if (stx == 1 && sty == arr.shape(-2)) {
-      return std::make_tuple(true, sty, arr);
-    } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
-      int64_t stx = arr.shape(-1);
-      return std::make_tuple(false, stx, arr_copy);
-    }
-  };
-
-  auto [a_transposed, lda, a] = check_transpose(a_pre);
-  auto [b_transposed, ldb, b] = check_transpose(b_pre);
-
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
-
-  if (M == 0 || N == 0) {
-    return;
-  }
-
-  if (K == 0) {
-    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
-    return;
-  }
-
-  // Get batch dims
-  auto batch_size_out = out.size() / (M * N);
-  size_t matrix_stride_out = M * N;
-
-  auto get_batch_dims = [](const auto& v) {
-    return decltype(v){v.begin(), v.end() - 2};
-  };
-
-  auto& lhs_indices = inputs[2];
-  auto& rhs_indices = inputs[3];
-
-  auto batch_shape = get_batch_dims(out.shape());
-  int batch_ndim = batch_shape.size();
-
-  auto batch_shape_A = get_batch_dims(a.shape());
-  auto batch_strides_A = get_batch_dims(a.strides());
-  auto batch_shape_B = get_batch_dims(b.shape());
-  auto batch_strides_B = get_batch_dims(b.strides());
-
-  const uint32_t* lhs_indices_ptr = lhs_indices.data<uint32_t>();
-  const uint32_t* rhs_indices_ptr = rhs_indices.data<uint32_t>();
-
-  for (int i = 0; i < batch_size_out; i++) {
-    // Get index
-    uint32_t indx_A = lhs_indices_ptr[elem_to_loc(i, lhs_indices)];
-    uint32_t indx_B = rhs_indices_ptr[elem_to_loc(i, rhs_indices)];
-
-    cblas_sgemm(
-        CblasRowMajor,
-        a_transposed ? CblasTrans : CblasNoTrans, // transA
-        b_transposed ? CblasTrans : CblasNoTrans, // transB
-        M,
-        N,
-        K,
-        1.0f, // alpha
-        a.data<float>() + elem_to_loc(indx_A, batch_shape_A, batch_strides_A),
-        lda,
-        b.data<float>() + elem_to_loc(indx_B, batch_shape_B, batch_strides_B),
-        ldb,
-        0.0f, // beta
-        out.data<float>() + matrix_stride_out * i,
-        out.shape(-1) // ldc
-    );
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/ops.h

@@ -1,680 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#pragma once
-#include <stdint.h>
-#include <cmath>
-#include <complex>
-
-namespace mlx::core::detail {
-
-namespace {
-constexpr float inf = std::numeric_limits<float>::infinity();
-} // namespace
-
-typedef union {
-  int i;
-  float f;
-} IntOrFloat;
-
-inline float fast_exp(float x) {
-  if (x == -std::numeric_limits<float>::infinity()) {
-    return 0.0f;
-  } else if (x == std::numeric_limits<float>::infinity() || std::isnan(x)) {
-    return x;
-  }
-  x *= 1.442695; // multiply with log_2(e)
-  float ipart, fpart;
-  IntOrFloat epart;
-  x = std::max(-80.f, std::min(x, 80.f));
-  ipart = std::floor(x + 0.5);
-  fpart = x - ipart;
-
-  x = 1.535336188319500e-4f;
-  x = x * fpart + 1.339887440266574e-3f;
-  x = x * fpart + 9.618437357674640e-3f;
-  x = x * fpart + 5.550332471162809e-2f;
-  x = x * fpart + 2.402264791363012e-1f;
-  x = x * fpart + 6.931472028550421e-1f;
-  x = x * fpart + 1.000000000000000f;
-
-  // generate 2**ipart in the floating point representation using integer
-  // bitshifting
-  epart.i = (int(ipart) + 127) << 23;
-
-  return epart.f * x;
-}
-
-inline float fast_erf(float a) {
-  float r, s, t, u;
-  t = std::abs(a);
-  s = a * a;
-  if (t > 0.927734375f) {
-    // maximum error 0.99527 ulp
-    r = std::fma(
-        -1.72853470e-5f, t, 3.83197126e-4f); // -0x1.220000p-16,0x1.91cfb2p-12
-    u = std::fma(
-        -3.88396438e-3f, t, 2.42546219e-2f); // -0x1.fd1438p-9, 0x1.8d6342p-6
-    r = std::fma(r, s, u);
-    r = std::fma(r, t, -1.06777877e-1f); // -0x1.b55cb8p-4
-    r = std::fma(r, t, -6.34846687e-1f); // -0x1.450aa0p-1
-    r = std::fma(r, t, -1.28717512e-1f); // -0x1.079d0cp-3
-    r = std::fma(r, t, -t);
-    // TODO, replace with expm1 when implemented
-    r = 1.0f - std::exp(r);
-    r = std::copysign(r, a);
-  } else {
-    // maximum error 0.98929 ulp
-    r = -5.96761703e-4f; // -0x1.38e000p-11
-    r = std::fma(r, s, 4.99119423e-3f); //  0x1.471a58p-8
-    r = std::fma(r, s, -2.67681349e-2f); // -0x1.b691b2p-6
-    r = std::fma(r, s, 1.12819925e-1f); //  0x1.ce1c44p-4
-    r = std::fma(r, s, -3.76125336e-1f); // -0x1.812700p-2
-    r = std::fma(r, s, 1.28379166e-1f); //  0x1.06eba8p-3
-    r = std::fma(r, a, a);
-  }
-  return r;
-}
-
-inline float fast_erfinv(float a) {
-  auto t = std::fma(a, 0.0f - a, 1.0f);
-  t = std::log(t);
-  float p;
-  if (std::abs(t) > 6.125f) { // maximum ulp error = 2.35793
-    p = 3.03697567e-10f; //  0x1.4deb44p-32
-    p = std::fma(p, t, 2.93243101e-8f); //  0x1.f7c9aep-26
-    p = std::fma(p, t, 1.22150334e-6f); //  0x1.47e512p-20
-    p = std::fma(p, t, 2.84108955e-5f); //  0x1.dca7dep-16
-    p = std::fma(p, t, 3.93552968e-4f); //  0x1.9cab92p-12
-    p = std::fma(p, t, 3.02698812e-3f); //  0x1.8cc0dep-9
-    p = std::fma(p, t, 4.83185798e-3f); //  0x1.3ca920p-8
-    p = std::fma(p, t, -2.64646143e-1f); // -0x1.0eff66p-2
-    p = std::fma(p, t, 8.40016484e-1f); //  0x1.ae16a4p-1
-  } else { // maximum ulp error = 2.35002
-    p = 5.43877832e-9f; //  0x1.75c000p-28
-    p = std::fma(p, t, 1.43285448e-7f); //  0x1.33b402p-23
-    p = std::fma(p, t, 1.22774793e-6f); //  0x1.499232p-20
-    p = std::fma(p, t, 1.12963626e-7f); //  0x1.e52cd2p-24
-    p = std::fma(p, t, -5.61530760e-5f); // -0x1.d70bd0p-15
-    p = std::fma(p, t, -1.47697632e-4f); // -0x1.35be90p-13
-    p = std::fma(p, t, 2.31468678e-3f); //  0x1.2f6400p-9
-    p = std::fma(p, t, 1.15392581e-2f); //  0x1.7a1e50p-7
-    p = std::fma(p, t, -2.32015476e-1f); // -0x1.db2aeep-3
-    p = std::fma(p, t, 8.86226892e-1f); //  0x1.c5bf88p-1
-  }
-  return a * p;
-}
-
-struct Abs {
-  template <typename T>
-  T operator()(T x) {
-    return std::abs(x);
-  }
-  uint8_t operator()(uint8_t x) {
-    return x;
-  }
-  uint16_t operator()(uint16_t x) {
-    return x;
-  }
-  uint32_t operator()(uint32_t x) {
-    return x;
-  }
-  uint64_t operator()(uint64_t x) {
-    return x;
-  }
-  bool operator()(bool x) {
-    return x;
-  }
-};
-
-struct ArcCos {
-  template <typename T>
-  T operator()(T x) {
-    return std::acos(x);
-  }
-};
-
-struct ArcCosh {
-  template <typename T>
-  T operator()(T x) {
-    return std::acosh(x);
-  }
-};
-
-struct ArcSin {
-  template <typename T>
-  T operator()(T x) {
-    return std::asin(x);
-  }
-};
-
-struct ArcSinh {
-  template <typename T>
-  T operator()(T x) {
-    return std::asinh(x);
-  }
-};
-
-struct ArcTan {
-  template <typename T>
-  T operator()(T x) {
-    return std::atan(x);
-  }
-};
-
-struct ArcTan2 {
-  template <typename T>
-  T operator()(T y, T x) {
-    return std::atan2(y, x);
-  }
-};
-
-struct ArcTanh {
-  template <typename T>
-  T operator()(T x) {
-    return std::atanh(x);
-  }
-};
-
-struct Ceil {
-  template <typename T>
-  T operator()(T x) {
-    return std::ceil(x);
-  }
-  int8_t operator()(int8_t x) {
-    return x;
-  }
-  int16_t operator()(int16_t x) {
-    return x;
-  }
-  int32_t operator()(int32_t x) {
-    return x;
-  }
-  int64_t operator()(int64_t x) {
-    return x;
-  }
-  uint8_t operator()(uint8_t x) {
-    return x;
-  }
-  uint16_t operator()(uint16_t x) {
-    return x;
-  }
-  uint32_t operator()(uint32_t x) {
-    return x;
-  }
-  uint64_t operator()(uint64_t x) {
-    return x;
-  }
-  bool operator()(bool x) {
-    return x;
-  }
-};
-
-struct Conjugate {
-  complex64_t operator()(complex64_t x) {
-    return std::conj(x);
-  }
-};
-
-struct Cos {
-  template <typename T>
-  T operator()(T x) {
-    return std::cos(x);
-  }
-};
-
-struct Cosh {
-  template <typename T>
-  T operator()(T x) {
-    return std::cosh(x);
-  }
-};
-
-struct Erf {
-  template <typename T>
-  T operator()(T x) {
-    return static_cast<T>(fast_erf(static_cast<float>(x)));
-  }
-};
-
-struct ErfInv {
-  template <typename T>
-  T operator()(T x) {
-    return static_cast<T>(fast_erfinv(static_cast<float>(x)));
-  }
-};
-
-struct Exp {
-  template <typename T>
-  T operator()(T x) {
-    return fast_exp(x);
-  }
-
-  complex64_t operator()(complex64_t x) {
-    return std::exp(x);
-  }
-};
-
-struct Expm1 {
-  template <typename T>
-  T operator()(T x) {
-    return expm1(x);
-  }
-};
-
-struct Floor {
-  template <typename T>
-  T operator()(T x) {
-    return std::floor(x);
-  }
-  int8_t operator()(int8_t x) {
-    return x;
-  }
-  int16_t operator()(int16_t x) {
-    return x;
-  }
-  int32_t operator()(int32_t x) {
-    return x;
-  }
-  int64_t operator()(int64_t x) {
-    return x;
-  }
-  uint8_t operator()(uint8_t x) {
-    return x;
-  }
-  uint16_t operator()(uint16_t x) {
-    return x;
-  }
-  uint32_t operator()(uint32_t x) {
-    return x;
-  }
-  uint64_t operator()(uint64_t x) {
-    return x;
-  }
-  bool operator()(bool x) {
-    return x;
-  }
-};
-
-struct Imag {
-  template <typename T>
-  T operator()(T x) {
-    return std::imag(x);
-  }
-};
-
-struct Log {
-  template <typename T>
-  T operator()(T x) {
-    return std::log(x);
-  }
-};
-
-struct Log2 {
-  template <typename T>
-  T operator()(T x) {
-    return std::log2(x);
-  }
-};
-
-struct Log10 {
-  template <typename T>
-  T operator()(T x) {
-    return std::log10(x);
-  }
-};
-
-struct Log1p {
-  template <typename T>
-  T operator()(T x) {
-    return log1p(x);
-  }
-};
-
-struct LogicalNot {
-  template <typename T>
-  T operator()(T x) {
-    return !x;
-  }
-};
-
-struct Negative {
-  template <typename T>
-  T operator()(T x) {
-    return -x;
-  }
-};
-
-struct Real {
-  template <typename T>
-  T operator()(T x) {
-    return std::real(x);
-  }
-};
-
-struct Round {
-  template <typename T>
-  T operator()(T x) {
-    return std::rint(x);
-  }
-
-  complex64_t operator()(complex64_t x) {
-    return {std::rint(x.real()), std::rint(x.imag())};
-  }
-};
-
-struct Sigmoid {
-  template <typename T>
-  T operator()(T x) {
-    auto one = static_cast<decltype(x)>(1.0);
-    return one / (one + fast_exp(-x));
-  }
-};
-
-struct Sign {
-  template <typename T>
-  T operator()(T x) {
-    return (x > T(0)) - (x < T(0));
-  }
-  uint8_t operator()(uint8_t x) {
-    return x != 0;
-  }
-  uint16_t operator()(uint16_t x) {
-    return x != 0;
-  }
-  uint32_t operator()(uint32_t x) {
-    return x != 0;
-  }
-  uint64_t operator()(uint64_t x) {
-    return x != 0;
-  }
-
-  complex64_t operator()(complex64_t x) {
-    return x == complex64_t(0) ? x : x / std::abs(x);
-  }
-};
-
-struct Sin {
-  template <typename T>
-  T operator()(T x) {
-    return std::sin(x);
-  }
-};
-
-struct Sinh {
-  template <typename T>
-  T operator()(T x) {
-    return std::sinh(x);
-  }
-};
-
-struct Square {
-  template <typename T>
-  T operator()(T x) {
-    return x * x;
-  }
-};
-
-struct Sqrt {
-  template <typename T>
-  T operator()(T x) {
-    return std::sqrt(x);
-  }
-};
-
-struct Rsqrt {
-  template <typename T>
-  T operator()(T x) {
-    return static_cast<decltype(x)>(1.0) / std::sqrt(x);
-  }
-};
-
-struct Tan {
-  template <typename T>
-  T operator()(T x) {
-    return std::tan(x);
-  }
-};
-
-struct Tanh {
-  template <typename T>
-  T operator()(T x) {
-    return std::tanh(x);
-  }
-};
-
-struct Add {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x + y;
-  }
-};
-
-struct Divide {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x / y;
-  }
-};
-
-struct Remainder {
-  template <typename T>
-  std::enable_if_t<std::is_integral_v<T> & !std::is_signed_v<T>, T> operator()(
-      T numerator,
-      T denominator) {
-    return numerator % denominator;
-  }
-
-  template <typename T>
-  std::enable_if_t<std::is_integral_v<T> & std::is_signed_v<T>, T> operator()(
-      T numerator,
-      T denominator) {
-    auto r = numerator % denominator;
-    if (r != 0 && (r < 0 != denominator < 0))
-      r += denominator;
-    return r;
-  }
-
-  template <typename T>
-  std::enable_if_t<!std::is_integral_v<T>, T> operator()(
-      T numerator,
-      T denominator) {
-    auto r = std::fmod(numerator, denominator);
-    if (r != 0 && (r < 0 != denominator < 0)) {
-      r += denominator;
-    }
-    return r;
-  }
-
-  complex64_t operator()(complex64_t numerator, complex64_t denominator) {
-    return numerator % denominator;
-  }
-};
-
-struct Equal {
-  template <typename T>
-  bool operator()(T x, T y) {
-    return x == y;
-  }
-};
-
-struct NaNEqual {
-  template <typename T>
-  bool operator()(T x, T y) {
-    if constexpr (std::is_integral_v<T>) {
-      // isnan always returns false for integers, and MSVC refuses to compile.
-      return x == y;
-    } else {
-      return x == y || (std::isnan(x) && std::isnan(y));
-    }
-  }
-};
-
-struct Greater {
-  template <typename T>
-  bool operator()(T x, T y) {
-    return x > y;
-  }
-};
-
-struct GreaterEqual {
-  template <typename T>
-  bool operator()(T x, T y) {
-    return x >= y;
-  }
-};
-
-struct Less {
-  template <typename T>
-  bool operator()(T x, T y) {
-    return x < y;
-  }
-};
-
-struct LessEqual {
-  template <typename T>
-  bool operator()(T x, T y) {
-    return x <= y;
-  }
-};
-
-struct Maximum {
-  template <typename T>
-  std::enable_if_t<std::is_integral_v<T>, T> operator()(T x, T y) {
-    return (x > y) ? x : y;
-  }
-
-  template <typename T>
-  std::enable_if_t<!std::is_integral_v<T>, T> operator()(T x, T y) {
-    if (std::isnan(x)) {
-      return x;
-    }
-    return (x > y) ? x : y;
-  }
-};
-
-struct Minimum {
-  template <typename T>
-  std::enable_if_t<std::is_integral_v<T>, T> operator()(T x, T y) {
-    return x < y ? x : y;
-  }
-
-  template <typename T>
-  std::enable_if_t<!std::is_integral_v<T>, T> operator()(T x, T y) {
-    if (std::isnan(x)) {
-      return x;
-    }
-    return x < y ? x : y;
-  }
-};
-
-struct LogAddExp {
-  template <typename T>
-  T operator()(T x, T y) {
-    constexpr float inf = std::numeric_limits<float>::infinity();
-    auto maxval = Maximum()(x, y);
-    auto minval = Minimum()(x, y);
-    return (minval == -inf || maxval == inf)
-        ? maxval
-        : static_cast<decltype(x)>(
-              maxval + std::log1p(fast_exp(minval - maxval)));
-  }
-};
-
-struct Multiply {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x * y;
-  }
-};
-
-struct NotEqual {
-  template <typename T>
-  bool operator()(T x, T y) {
-    return x != y;
-  }
-};
-
-struct Power {
-  template <typename T>
-  std::enable_if_t<!std::is_integral_v<T>, T> operator()(T base, T exp) {
-    return std::pow(base, exp);
-  }
-
-  template <typename T>
-  std::enable_if_t<std::is_integral_v<T>, T> operator()(T base, T exp) {
-    T res = 1;
-    while (exp) {
-      if (exp & 1) {
-        res *= base;
-      }
-      exp >>= 1;
-      base *= base;
-    }
-    return res;
-  }
-};
-
-struct Subtract {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x - y;
-  }
-};
-
-struct LogicalAnd {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x && y;
-  }
-};
-
-struct LogicalOr {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x || y;
-  }
-};
-
-struct Select {
-  template <typename T>
-  T operator()(bool condition, T x, T y) {
-    return condition ? x : y;
-  }
-};
-
-struct BitwiseAnd {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x & y;
-  }
-};
-
-struct BitwiseOr {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x | y;
-  }
-};
-
-struct BitwiseXor {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x ^ y;
-  }
-};
-
-struct LeftShift {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x << y;
-  }
-};
-
-struct RightShift {
-  template <typename T>
-  T operator()(T x, T y) {
-    return x >> y;
-  }
-};
-
-} // namespace mlx::core::detail

mlx/backend/common/primitives.cpp

@@ -1,714 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include <algorithm>
-#include <cassert>
-#include <cmath>
-#include <numeric>
-#include <sstream>
-
-#include "mlx/allocator.h"
-#include "mlx/backend/common/arange.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/ops.h"
-#include "mlx/backend/common/slicing.h"
-#include "mlx/backend/common/threefry.h"
-#include "mlx/backend/common/unary.h"
-#include "mlx/backend/common/utils.h"
-#include "mlx/primitives.h"
-#include "mlx/utils.h"
-
-namespace mlx::core {
-
-void reshape(const array& in, array& out) {
-  auto [copy_necessary, out_strides] = prepare_reshape(in, out);
-  if (copy_necessary) {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    copy_inplace(in, out, CopyType::General);
-  } else {
-    shared_buffer_reshape(in, out_strides, out);
-  }
-}
-
-int64_t compute_dynamic_offset(
-    const array& indices,
-    const Strides& strides,
-    const std::vector<int>& axes) {
-  auto compute_offset = [&strides, &axes](const auto* indices) {
-    int64_t offset = 0;
-    for (int i = 0; i < axes.size(); ++i) {
-      offset += indices[i] * strides[axes[i]];
-    }
-    return offset;
-  };
-  switch (indices.dtype()) {
-    case int8:
-    case uint8:
-      return compute_offset(indices.data<uint8_t>());
-    case int16:
-    case uint16:
-      return compute_offset(indices.data<uint16_t>());
-    case int32:
-    case uint32:
-      return compute_offset(indices.data<uint32_t>());
-    case int64:
-    case uint64:
-      return compute_offset(indices.data<uint64_t>());
-    default:
-      throw std::runtime_error("Invalid indices type.");
-  }
-}
-
-void Abs::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (issubdtype(in.dtype(), unsignedinteger)) {
-    // No-op for unsigned types
-    out.copy_shared_buffer(in);
-  } else {
-    unary(in, out, detail::Abs());
-  }
-}
-
-void Arange::eval(const std::vector<array>& inputs, array& out) {
-  arange(inputs, out, start_, step_);
-}
-
-void ArcCos::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::ArcCos());
-  } else {
-    throw std::invalid_argument(
-        "[arccos] Cannot compute inverse cosine of elements in array"
-        " with non floating point type.");
-  }
-}
-
-void ArcCosh::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::ArcCosh());
-  } else {
-    throw std::invalid_argument(
-        "[arccosh] Cannot compute inverse hyperbolic cosine of elements in"
-        " array with non floating point type.");
-  }
-}
-
-void ArcSin::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::ArcSin());
-  } else {
-    throw std::invalid_argument(
-        "[arcsin] Cannot compute inverse sine of elements in array"
-        " with non floating point type.");
-  }
-}
-
-void ArcSinh::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::ArcSinh());
-  } else {
-    throw std::invalid_argument(
-        "[arcsinh] Cannot compute inverse hyperbolic sine of elements in"
-        " array with non floating point type.");
-  }
-}
-
-void ArcTan::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::ArcTan());
-  } else {
-    throw std::invalid_argument(
-        "[arctan] Cannot compute inverse tangent of elements in array"
-        " with non floating point type.");
-  }
-}
-
-void ArcTanh::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::ArcTanh());
-  } else {
-    throw std::invalid_argument(
-        "[arctanh] Cannot compute inverse hyperbolic tangent of elements in"
-        " array with non floating point type.");
-  }
-}
-
-void AsType::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype);
-}
-
-void Ceil::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (issubdtype(in.dtype(), inexact)) {
-    unary_fp(in, out, detail::Ceil());
-  } else {
-    // No-op integer types
-    out.copy_shared_buffer(in);
-  }
-}
-
-void Concatenate::eval(const std::vector<array>& inputs, array& out) {
-  std::vector<int> sizes;
-  sizes.push_back(0);
-  for (auto& p : inputs) {
-    sizes.push_back(p.shape(axis_));
-  }
-  std::partial_sum(sizes.cbegin(), sizes.cend(), sizes.begin());
-
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  auto strides = out.strides();
-  auto flags = out.flags();
-  flags.row_contiguous = false;
-  flags.col_contiguous = false;
-  flags.contiguous = false;
-  for (int i = 0; i < inputs.size(); i++) {
-    array out_slice(inputs[i].shape(), out.dtype(), nullptr, {});
-    size_t data_offset = strides[axis_] * sizes[i];
-    out_slice.copy_shared_buffer(
-        out, strides, flags, out_slice.size(), data_offset);
-    copy_inplace(inputs[i], out_slice, CopyType::GeneralGeneral);
-  }
-}
-
-void Conjugate::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.dtype() == complex64) {
-    unary_fp(in, out, detail::Conjugate());
-  } else {
-    throw std::invalid_argument(
-        "[conjugate] conjugate must be called on complex input.");
-  }
-}
-
-void Contiguous::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (in.flags().row_contiguous ||
-      (allow_col_major_ && in.flags().col_contiguous)) {
-    out.copy_shared_buffer(in);
-  } else {
-    copy(in, out, CopyType::General);
-  }
-}
-
-void Cos::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Cos());
-  } else {
-    throw std::invalid_argument(
-        "[cos] Cannot compute cosine of elements in array"
-        " with non floating point type.");
-  }
-}
-
-void Cosh::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Cosh());
-  } else {
-    throw std::invalid_argument(
-        "[cosh] Cannot compute hyperbolic cosine of elements in array"
-        " with non floating point type.");
-  }
-}
-
-void Erf::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  switch (out.dtype()) {
-    case float32:
-      unary_op<float>(in, out, detail::Erf());
-      break;
-    case float16:
-      unary_op<float16_t>(in, out, detail::Erf());
-      break;
-    case bfloat16:
-      unary_op<bfloat16_t>(in, out, detail::Erf());
-      break;
-    default:
-      throw std::invalid_argument(
-          "[erf] Error function only defined for arrays"
-          " with real floating point type.");
-  }
-}
-
-void ErfInv::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  switch (out.dtype()) {
-    case float32:
-      unary_op<float>(in, out, detail::ErfInv());
-      break;
-    case float16:
-      unary_op<float16_t>(in, out, detail::ErfInv());
-      break;
-    case bfloat16:
-      unary_op<bfloat16_t>(in, out, detail::ErfInv());
-      break;
-    default:
-      throw std::invalid_argument(
-          "[erf_inv] Inverse error function only defined for arrays"
-          " with real floating point type.");
-  }
-}
-
-void Exp::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Exp());
-  } else {
-    throw std::invalid_argument(
-        "[exp] Cannot exponentiate elements in array"
-        " with non floating point type.");
-  }
-}
-
-void Expm1::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Expm1());
-  } else {
-    throw std::invalid_argument(
-        "[expm1] Cannot exponentiate elements in array"
-        " with non floating point type.");
-  }
-}
-
-void Flatten::eval_cpu(const std::vector<array>& inputs, array& out) {
-  reshape(inputs[0], out);
-}
-
-void Unflatten::eval_cpu(const std::vector<array>& inputs, array& out) {
-  reshape(inputs[0], out);
-}
-
-void Floor::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (issubdtype(in.dtype(), inexact)) {
-    unary_fp(in, out, detail::Floor());
-  } else {
-    // No-op integer types
-    out.copy_shared_buffer(in);
-  }
-}
-
-void Full::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  assert(in.dtype() == out.dtype());
-  CopyType ctype;
-  if (in.data_size() == 1) {
-    ctype = CopyType::Scalar;
-  } else if (in.flags().contiguous) {
-    ctype = CopyType::Vector;
-  } else {
-    ctype = CopyType::General;
-  }
-  copy(in, out, ctype);
-}
-
-void Imag::eval_cpu(const std::vector<array>& inputs, array& out) {
-  unary_op<complex64_t, float>(inputs[0], out, detail::Imag());
-}
-
-void Log::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    switch (base_) {
-      case Base::e:
-        unary_fp(in, out, detail::Log());
-        break;
-      case Base::two:
-        unary_fp(in, out, detail::Log2());
-        break;
-      case Base::ten:
-        unary_fp(in, out, detail::Log10());
-        break;
-    }
-  } else {
-    throw std::invalid_argument(
-        "[log] Cannot compute log of elements in array with"
-        " non floating point type.");
-  }
-}
-
-void Log1p::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Log1p());
-  } else {
-    throw std::invalid_argument(
-        "[log1p] Cannot compute log of elements in array with"
-        " non floating point type.");
-  }
-}
-
-void LogicalNot::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  unary(in, out, detail::LogicalNot());
-}
-
-void Negative::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  unary(in, out, detail::Negative());
-}
-
-void Pad::eval(const std::vector<array>& inputs, array& out) {
-  // Inputs must be base input array and scalar val array
-  assert(inputs.size() == 2);
-  auto& in = inputs[0];
-  auto& val = inputs[1];
-
-  // Padding value must be a scalar
-  assert(val.size() == 1);
-
-  // Padding value, input and output must be of the same type
-  assert(val.dtype() == in.dtype() && in.dtype() == out.dtype());
-
-  // Fill output with val
-  copy(val, out, CopyType::Scalar);
-
-  // Find offset for start of input values
-  size_t data_offset = 0;
-  for (int i = 0; i < axes_.size(); i++) {
-    auto ax = axes_[i] < 0 ? out.ndim() + axes_[i] : axes_[i];
-    data_offset += out.strides()[ax] * low_pad_size_[i];
-  }
-
-  // Extract slice from output where input will be pasted
-  array out_slice(in.shape(), out.dtype(), nullptr, {});
-  out_slice.copy_shared_buffer(
-      out, out.strides(), out.flags(), out_slice.size(), data_offset);
-
-  // Copy input values into the slice
-  copy_inplace(in, out_slice, CopyType::GeneralGeneral);
-}
-
-void RandomBits::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  // keys has shape (N1, ..., NK, 2)
-  // out has shape (N1, ..., NK, M1, M2, ...)
-  auto& keys = inputs[0];
-  size_t num_keys = keys.size() / 2;
-
-  size_t elems_per_key = out.size() / num_keys;
-  size_t bytes_per_key = out.itemsize() * elems_per_key;
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  auto kptr = inputs[0].data<uint32_t>();
-  auto cptr = out.data<char>();
-  size_t out_skip = (bytes_per_key + 4 - 1) / 4;
-  auto half_size = out_skip / 2;
-  bool even = out_skip % 2 == 0;
-  for (int i = 0; i < num_keys; ++i, cptr += bytes_per_key) {
-    auto ptr = reinterpret_cast<uint32_t*>(cptr);
-    // Get ith key
-    auto kidx = 2 * i;
-    auto k1_elem = elem_to_loc(kidx, keys.shape(), keys.strides());
-    auto k2_elem = elem_to_loc(kidx + 1, keys.shape(), keys.strides());
-    auto key = std::make_pair(kptr[k1_elem], kptr[k2_elem]);
-
-    std::pair<uintptr_t, uintptr_t> count{0, half_size + !even};
-    for (; count.first + 1 < half_size; count.first++, count.second++) {
-      std::tie(ptr[count.first], ptr[count.second]) =
-          random::threefry2x32_hash(key, count);
-    }
-    if (count.first < half_size) {
-      auto rb = random::threefry2x32_hash(key, count);
-      ptr[count.first++] = rb.first;
-      if (bytes_per_key % 4 > 0) {
-        std::copy(
-            reinterpret_cast<char*>(&rb.second),
-            reinterpret_cast<char*>(&rb.second) + bytes_per_key % 4,
-            cptr + 4 * count.second);
-      } else {
-        ptr[count.second] = rb.second;
-      }
-    }
-    if (!even) {
-      count.second = 0;
-      ptr[half_size] = random::threefry2x32_hash(key, count).first;
-    }
-  }
-}
-
-void Real::eval_cpu(const std::vector<array>& inputs, array& out) {
-  unary_op<complex64_t, float>(inputs[0], out, detail::Real());
-}
-
-void Reshape::eval_cpu(const std::vector<array>& inputs, array& out) {
-  reshape(inputs[0], out);
-}
-
-void Round::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (issubdtype(in.dtype(), inexact)) {
-    unary_fp(in, out, detail::Round());
-  } else {
-    // No-op integer types
-    out.copy_shared_buffer(in);
-  }
-}
-
-void Sigmoid::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Sigmoid());
-  } else {
-    throw std::invalid_argument(
-        "[sigmoid] Cannot sigmoid of elements in array with"
-        " non floating point type.");
-  }
-}
-
-void Sign::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (in.dtype() == bool_) {
-    out.copy_shared_buffer(in);
-  } else {
-    unary(in, out, detail::Sign());
-  }
-}
-
-void Sin::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Sin());
-  } else {
-    throw std::invalid_argument(
-        "[sin] Cannot compute sine of elements in array"
-        " with non floating point type.");
-  }
-}
-
-void Sinh::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Sinh());
-  } else {
-    throw std::invalid_argument(
-        "[sinh] Cannot compute hyperbolic sine of elements in array"
-        " with non floating point type.");
-  }
-}
-
-void Slice::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-
-  auto& in = inputs[0];
-
-  // Calculate out strides, initial offset and if copy needs to be made
-  auto [data_offset, inp_strides] = prepare_slice(in, start_indices_, strides_);
-  size_t data_end = 1;
-  for (int i = 0; i < end_indices_.size(); ++i) {
-    if (in.shape()[i] > 1) {
-      auto end_idx = start_indices_[i] + out.shape()[i] * strides_[i] - 1;
-      data_end += end_idx * in.strides()[i];
-    }
-  }
-  size_t data_size = data_end - data_offset;
-  Strides ostrides{inp_strides.begin(), inp_strides.end()};
-  shared_buffer_slice(in, ostrides, data_offset, data_size, out);
-}
-
-void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-  auto& in = inputs[0];
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  auto i_offset = compute_dynamic_offset(inputs[1], in.strides(), axes_);
-  copy_inplace(
-      /* const array& src = */ in,
-      /* array& dst = */ out,
-      /* const Shape& data_shape = */ out.shape(),
-      /* const Strides& i_strides = */ in.strides(),
-      /* const Strides& o_strides = */ out.strides(),
-      /* int64_t i_offset = */ i_offset,
-      /* int64_t o_offset = */ 0,
-      /* CopyType ctype = */ CopyType::GeneralGeneral);
-}
-
-void DynamicSliceUpdate::eval_cpu(
-    const std::vector<array>& inputs,
-    array& out) {
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-
-  auto& in = inputs[0];
-  auto& upd = inputs[1];
-
-  // Copy or move src to dst
-  auto ctype = in.flags().contiguous && in.size() == in.data_size()
-      ? CopyType::Vector
-      : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype);
-
-  auto o_offset = compute_dynamic_offset(inputs[2], out.strides(), axes_);
-  copy_inplace(
-      /* const array& src = */ upd,
-      /* array& dst = */ out,
-      /* const std::vector<int>& data_shape = */ upd.shape(),
-      /* const std::vector<stride_t>& i_strides = */ upd.strides(),
-      /* const std::vector<stride_t>& o_strides = */ out.strides(),
-      /* int64_t i_offset = */ 0,
-      /* int64_t o_offset = */ o_offset,
-      /* CopyType ctype = */ CopyType::GeneralGeneral);
-}
-
-void SliceUpdate::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2);
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-
-  auto& in = inputs[0];
-  auto& upd = inputs[1];
-
-  if (upd.size() == 0) {
-    out.copy_shared_buffer(in);
-    return;
-  }
-
-  // Check if materialization is needed
-  auto ctype = in.flags().contiguous && in.size() == in.data_size()
-      ? CopyType::Vector
-      : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype);
-
-  // Calculate out strides, initial offset and if copy needs to be made
-  auto [data_offset, out_strides] = prepare_slice(in, start_indices_, strides_);
-
-  // Do copy
-  copy_inplace(
-      /* const array& src = */ upd,
-      /* array& dst = */ out,
-      /* const std::vector<int>& data_shape = */ upd.shape(),
-      /* const std::vector<stride_t>& i_strides = */ upd.strides(),
-      /* const std::vector<stride_t>& o_strides = */ out_strides,
-      /* int64_t i_offset = */ 0,
-      /* int64_t o_offset = */ data_offset,
-      /* CopyType ctype = */ CopyType::GeneralGeneral);
-}
-
-void Square::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  unary(in, out, detail::Square());
-}
-
-void Sqrt::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  if (recip_) {
-    unary_fp(in, out, detail::Rsqrt());
-  } else {
-    unary_fp(in, out, detail::Sqrt());
-  }
-}
-
-void Tan::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Tan());
-  } else {
-    throw std::invalid_argument(
-        "[tan] Cannot compute tangent of elements in array"
-        " with non floating point type.");
-  }
-}
-
-void Tanh::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, detail::Tanh());
-  } else {
-    throw std::invalid_argument(
-        "[tanh] Cannot compute hyperbolic tangent of elements in array"
-        " with non floating point type.");
-  }
-}
-
-void View::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  auto ibytes = size_of(in.dtype());
-  auto obytes = size_of(out.dtype());
-  // Conditions for buffer copying (disjunction):
-  // - type size is the same
-  // - type size is smaller and the last axis is contiguous
-  // - the entire array is row contiguous
-  if (ibytes == obytes || (obytes < ibytes && in.strides().back() == 1) ||
-      in.flags().row_contiguous) {
-    auto strides = in.strides();
-    for (int i = 0; i < static_cast<int>(strides.size()) - 1; ++i) {
-      strides[i] *= ibytes;
-      strides[i] /= obytes;
-    }
-    out.copy_shared_buffer(
-        in, strides, in.flags(), in.data_size() * ibytes / obytes);
-  } else {
-    auto tmp = array(
-        in.shape(), in.dtype() == bool_ ? uint8 : in.dtype(), nullptr, {});
-    tmp.set_data(allocator::malloc_or_wait(tmp.nbytes()));
-    if (in.dtype() == bool_) {
-      auto in_tmp = array(in.shape(), uint8, nullptr, {});
-      in_tmp.copy_shared_buffer(in);
-      copy_inplace(in_tmp, tmp, CopyType::General);
-    } else {
-      copy_inplace(in, tmp, CopyType::General);
-    }
-
-    auto flags = out.flags();
-    flags.contiguous = true;
-    flags.row_contiguous = true;
-    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
-    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
-    out.move_shared_buffer(tmp, out.strides(), flags, out.size());
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/qrf.cpp

@@ -1,148 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include "mlx/allocator.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/lapack.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-template <typename T>
-struct lpack;
-
-template <>
-struct lpack<float> {
-  static void xgeqrf(
-      const int* m,
-      const int* n,
-      float* a,
-      const int* lda,
-      float* tau,
-      float* work,
-      const int* lwork,
-      int* info) {
-    sgeqrf_(m, n, a, lda, tau, work, lwork, info);
-  }
-  static void xorgqr(
-      const int* m,
-      const int* n,
-      const int* k,
-      float* a,
-      const int* lda,
-      const float* tau,
-      float* work,
-      const int* lwork,
-      int* info) {
-    sorgqr_(m, n, k, a, lda, tau, work, lwork, info);
-  }
-};
-
-template <typename T>
-void qrf_impl(const array& a, array& q, array& r) {
-  const int M = a.shape(-2);
-  const int N = a.shape(-1);
-  const int lda = std::max(M, N);
-  size_t num_matrices = a.size() / (M * N);
-  int num_reflectors = std::min(M, N);
-  auto tau =
-      allocator::malloc_or_wait(sizeof(T) * num_matrices * num_reflectors);
-
-  // Copy A to inplace input and make it col-contiguous
-  array in(a.shape(), float32, nullptr, {});
-  auto flags = in.flags();
-
-  // Copy the input to be column contiguous
-  flags.col_contiguous = num_matrices == 1;
-  flags.row_contiguous = false;
-  auto strides = in.strides();
-  strides[in.ndim() - 2] = 1;
-  strides[in.ndim() - 1] = M;
-  in.set_data(
-      allocator::malloc_or_wait(in.nbytes()), in.nbytes(), strides, flags);
-  copy_inplace(a, in, CopyType::GeneralGeneral);
-
-  T optimal_work;
-  int lwork = -1;
-  int info;
-
-  // Compute workspace size
-  lpack<T>::xgeqrf(
-      &M, &N, nullptr, &lda, nullptr, &optimal_work, &lwork, &info);
-
-  // Update workspace size
-  lwork = optimal_work;
-  auto work = allocator::malloc_or_wait(sizeof(T) * lwork);
-
-  // Loop over matrices
-  for (int i = 0; i < num_matrices; ++i) {
-    // Solve
-    lpack<T>::xgeqrf(
-        &M,
-        &N,
-        in.data<float>() + M * N * i,
-        &lda,
-        static_cast<T*>(tau.raw_ptr()) + num_reflectors * i,
-        static_cast<T*>(work.raw_ptr()),
-        &lwork,
-        &info);
-  }
-  allocator::free(work);
-
-  r.set_data(allocator::malloc_or_wait(r.nbytes()));
-  copy_inplace(in, r, CopyType::General);
-
-  for (int i = 0; i < num_matrices; ++i) {
-    // Zero lower triangle
-    for (int j = 0; j < r.shape(-2); ++j) {
-      for (int k = 0; k < j; ++k) {
-        r.data<T>()[i * N * M + j * N + k] = 0;
-      }
-    }
-  }
-
-  // Get work size
-  lwork = -1;
-  lpack<T>::xorgqr(
-      &M,
-      &N,
-      &num_reflectors,
-      nullptr,
-      &lda,
-      nullptr,
-      &optimal_work,
-      &lwork,
-      &info);
-  lwork = optimal_work;
-  work = allocator::malloc_or_wait(sizeof(T) * lwork);
-
-  // Loop over matrices
-  for (int i = 0; i < num_matrices; ++i) {
-    // Compute Q
-    lpack<T>::xorgqr(
-        &M,
-        &N,
-        &num_reflectors,
-        in.data<float>() + M * N * i,
-        &lda,
-        static_cast<T*>(tau.raw_ptr()) + num_reflectors * i,
-        static_cast<T*>(work.raw_ptr()),
-        &lwork,
-        &info);
-  }
-
-  q.set_data(allocator::malloc_or_wait(q.nbytes()));
-  copy_inplace(in, q, CopyType::General);
-
-  // Cleanup
-  allocator::free(work);
-  allocator::free(tau);
-}
-
-void QRF::eval(const std::vector<array>& inputs, std::vector<array>& outputs) {
-  if (!(inputs[0].dtype() == float32)) {
-    throw std::runtime_error("[QRF::eval] only supports float32.");
-  }
-  qrf_impl<float>(inputs[0], outputs[0], outputs[1]);
-}
-
-} // namespace mlx::core

mlx/backend/common/quantized.cpp

@@ -1,565 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cassert>
-
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/ops.h"
-#include "mlx/fast_primitives.h"
-#include "mlx/primitives.h"
-#include "mlx/utils.h"
-
-namespace mlx::core {
-
-namespace {
-
-template <typename T, int bits>
-void extract_bits(const uint8_t* w_in, T* w_out) {
-  assert(bits == 3 || bits == 6);
-  if (bits == 3) {
-    w_out[0] = static_cast<T>(w_in[0] & 0x7);
-    w_out[1] = static_cast<T>((w_in[0] & 0x38) >> 3);
-    w_out[2] = static_cast<T>(((w_in[0] & 0xc0) >> 6) + ((w_in[1] & 0x1) << 2));
-    w_out[3] = static_cast<T>((w_in[1] & 0xe) >> 1);
-    w_out[4] = static_cast<T>((w_in[1] & 0x70) >> 4);
-    w_out[5] = static_cast<T>(((w_in[1] & 0x80) >> 7) + ((w_in[2] & 0x3) << 1));
-    w_out[6] = static_cast<T>((w_in[2] & 0x1c) >> 2);
-    w_out[7] = static_cast<T>((w_in[2] & 0xe0) >> 5);
-  } else if (bits == 6) {
-    w_out[0] = static_cast<T>(w_in[0] & 0x3f);
-    w_out[1] =
-        static_cast<T>(((w_in[0] >> 6) & 0x03) + ((w_in[1] & 0x0f) << 2));
-    w_out[2] =
-        static_cast<T>(((w_in[1] >> 4) & 0x0f) + ((w_in[2] & 0x03) << 4));
-    w_out[3] = static_cast<T>((w_in[2] >> 2) & 0x3f);
-  }
-}
-
-template <typename T, int bits, int group_size>
-void _qmm(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const T* scales,
-    const T* biases,
-    int M,
-    int N,
-    int K) {
-  constexpr int bitmask = (1 << bits) - 1;
-  constexpr int pack_factor = bits == 3 ? 8 : bits == 6 ? 4 : 8 / bits;
-  constexpr int bytes_per_pack = (bits == 3 || bits == 6) ? 3 : 1;
-  constexpr int packs_in_group = group_size / pack_factor;
-
-  for (int m = 0; m < M; m++) {
-    const uint8_t* w_local = (const uint8_t*)w;
-    const T* scales_local = scales;
-    const T* biases_local = biases;
-
-    std::fill(result, result + N, 0);
-
-    for (int k = 0; k < K; k++) {
-      T* result_local = result;
-      T xi = *x++;
-
-      for (int n = 0; n < N; n += group_size) {
-        T scale = *scales_local++;
-        T bias = *biases_local++;
-        for (int ng = 0; ng < packs_in_group; ng++) {
-          if (bits == 3 || bits == 6) {
-            T wl[pack_factor];
-            extract_bits<T, bits>(w_local, wl);
-#pragma clang loop unroll(full)
-            for (int p = 0; p < pack_factor; p++) {
-              (*result_local++) += xi * (scale * wl[p] + bias);
-            }
-            w_local += bytes_per_pack;
-
-          } else {
-            uint8_t wi = *w_local++;
-#pragma clang loop unroll(full)
-            for (int p = 0; p < pack_factor; p++) {
-              (*result_local++) +=
-                  xi * (scale * static_cast<T>(wi & bitmask) + bias);
-              if (bits != 8) {
-                wi >>= bits;
-              }
-            }
-          }
-        }
-      }
-    }
-
-    result += N;
-  }
-}
-
-template <typename T, int bits, int group_size>
-void _qmm_t(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const T* scales,
-    const T* biases,
-    int M,
-    int N,
-    int K) {
-  constexpr int bitmask = (1 << bits) - 1;
-  constexpr int pack_factor = bits == 3 ? 8 : bits == 6 ? 4 : 8 / bits;
-  constexpr int bytes_per_pack = (bits == 3 || bits == 6) ? 3 : 1;
-  constexpr int packs_in_group = group_size / pack_factor;
-
-  for (int m = 0; m < M; m++) {
-    const uint8_t* w_local = (const uint8_t*)w;
-    const T* scales_local = scales;
-    const T* biases_local = biases;
-
-    for (int n = 0; n < N; n++) {
-      const T* x_local = x;
-      T sum = 0;
-      for (int k = 0; k < K; k += group_size) {
-        T scale = *scales_local++;
-        T bias = *biases_local++;
-
-        for (int kw = 0; kw < packs_in_group; kw++) {
-          if (bits == 3 || bits == 6) {
-            T wl[pack_factor];
-            extract_bits<T, bits>(w_local, wl);
-#pragma clang loop unroll(full)
-            for (int p = 0; p < pack_factor; p++) {
-              sum += x_local[p] * (scale * wl[p] + bias);
-            }
-            w_local += bytes_per_pack;
-            x_local += pack_factor;
-
-          } else {
-            uint8_t wi = *w_local++;
-#pragma clang loop unroll(full)
-            for (int p = 0; p < pack_factor; p++) {
-              sum +=
-                  (*x_local++) * (scale * static_cast<T>(wi & bitmask) + bias);
-              if (bits != 8) {
-                wi >>= bits;
-              }
-            }
-          }
-        }
-      }
-      *result = sum;
-      result++;
-    }
-
-    x += K;
-  }
-}
-
-template <typename T, int bits, int group_size>
-void _qmm_dispatch_transpose(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const T* scales,
-    const T* biases,
-    int M,
-    int N,
-    int K,
-    bool transposed_w) {
-  if (transposed_w) {
-    return _qmm_t<T, bits, group_size>(result, x, w, scales, biases, M, N, K);
-  } else {
-    return _qmm<T, bits, group_size>(result, x, w, scales, biases, M, N, K);
-  }
-}
-
-template <typename T, int bits>
-void _qmm_dispatch_group(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const T* scales,
-    const T* biases,
-    int M,
-    int N,
-    int K,
-    int group_size,
-    bool transposed_w) {
-  switch (group_size) {
-    case 32:
-      _qmm_dispatch_transpose<T, bits, 32>(
-          result, x, w, scales, biases, M, N, K, transposed_w);
-      break;
-    case 64:
-      _qmm_dispatch_transpose<T, bits, 64>(
-          result, x, w, scales, biases, M, N, K, transposed_w);
-      break;
-    case 128:
-      _qmm_dispatch_transpose<T, bits, 128>(
-          result, x, w, scales, biases, M, N, K, transposed_w);
-      break;
-    default:
-      throw std::invalid_argument(
-          "Quantization group size must be 32, 64 or 128.");
-  }
-}
-
-template <typename T>
-void _qmm_dispatch_typed(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const T* scales,
-    const T* biases,
-    int M,
-    int N,
-    int K,
-    int group_size,
-    int bits,
-    bool transposed_w) {
-  switch (bits) {
-    case 2:
-      _qmm_dispatch_group<T, 2>(
-          result, x, w, scales, biases, M, N, K, group_size, transposed_w);
-      break;
-    case 3:
-      _qmm_dispatch_group<T, 3>(
-          result, x, w, scales, biases, M, N, K, group_size, transposed_w);
-      break;
-    case 4:
-      _qmm_dispatch_group<T, 4>(
-          result, x, w, scales, biases, M, N, K, group_size, transposed_w);
-      break;
-    case 6:
-      _qmm_dispatch_group<T, 6>(
-          result, x, w, scales, biases, M, N, K, group_size, transposed_w);
-      break;
-    case 8:
-      _qmm_dispatch_group<T, 8>(
-          result, x, w, scales, biases, M, N, K, group_size, transposed_w);
-      break;
-    default:
-      throw std::invalid_argument("Quantization bits must be 2, 3, 4, 6 or 8.");
-  }
-}
-
-void _qmm_dispatch(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    const array& biases,
-    int bits,
-    int group_size,
-    bool transposed_w) {
-  int K = x.shape(-1);
-  int M = x.shape(-2);
-  int N = out.shape(-1);
-
-  int w_els = w.ndim() > 2 ? w.shape(-1) * w.shape(-2) : 0;
-  int g_els = w.ndim() > 2 ? scales.shape(-1) * scales.shape(-2) : 0;
-
-  int batch_size = x.size() / x.shape(-1) / x.shape(-2);
-  for (int i = 0; i < batch_size; i++) {
-    switch (x.dtype()) {
-      case float32:
-        _qmm_dispatch_typed<float>(
-            out.data<float>() + i * M * N,
-            x.data<float>() + elem_to_loc(i * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(i * w_els, w),
-            scales.data<float>() + elem_to_loc(i * g_els, scales),
-            biases.data<float>() + elem_to_loc(i * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      case float16:
-        _qmm_dispatch_typed<float16_t>(
-            out.data<float16_t>() + i * M * N,
-            x.data<float16_t>() + elem_to_loc(i * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(i * w_els, w),
-            scales.data<float16_t>() + elem_to_loc(i * g_els, scales),
-            biases.data<float16_t>() + elem_to_loc(i * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      case bfloat16:
-        _qmm_dispatch_typed<bfloat16_t>(
-            out.data<bfloat16_t>() + i * M * N,
-            x.data<bfloat16_t>() + elem_to_loc(i * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(i * w_els, w),
-            scales.data<bfloat16_t>() + elem_to_loc(i * g_els, scales),
-            biases.data<bfloat16_t>() + elem_to_loc(i * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      default:
-        throw std::invalid_argument(
-            "[quantized_matmul] only floating types are supported");
-    }
-  }
-}
-
-void _bs_qmm_dispatch(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    const array& biases,
-    const array& lhs_indices,
-    const array& rhs_indices,
-    int bits,
-    int group_size,
-    bool transposed_w) {
-  int K = x.shape(-1);
-  int M = x.shape(-2);
-  int N = out.shape(-1);
-
-  int w_els = w.shape(-1) * w.shape(-2);
-  int g_els = scales.shape(-1) * scales.shape(-2);
-
-  const uint32_t* lhs_indices_data = lhs_indices.data<uint32_t>();
-  const uint32_t* rhs_indices_data = rhs_indices.data<uint32_t>();
-
-  for (int i = 0; i < lhs_indices.size(); i++) {
-    int x_idx = lhs_indices_data[elem_to_loc(i, lhs_indices)];
-    int w_idx = rhs_indices_data[elem_to_loc(i, rhs_indices)];
-
-    switch (x.dtype()) {
-      case float32:
-        _qmm_dispatch_typed<float>(
-            out.data<float>() + i * M * N,
-            x.data<float>() + elem_to_loc(x_idx * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(w_idx * w_els, w),
-            scales.data<float>() + elem_to_loc(w_idx * g_els, scales),
-            biases.data<float>() + elem_to_loc(w_idx * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      case float16:
-        _qmm_dispatch_typed<float16_t>(
-            out.data<float16_t>() + i * M * N,
-            x.data<float16_t>() + elem_to_loc(x_idx * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(w_idx * w_els, w),
-            scales.data<float16_t>() + elem_to_loc(w_idx * g_els, scales),
-            biases.data<float16_t>() + elem_to_loc(w_idx * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      case bfloat16:
-        _qmm_dispatch_typed<bfloat16_t>(
-            out.data<bfloat16_t>() + i * M * N,
-            x.data<bfloat16_t>() + elem_to_loc(x_idx * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(w_idx * w_els, w),
-            scales.data<bfloat16_t>() + elem_to_loc(w_idx * g_els, scales),
-            biases.data<bfloat16_t>() + elem_to_loc(w_idx * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      default:
-        throw std::invalid_argument(
-            "[quantized_matmul] only floating types are supported");
-    }
-  }
-}
-
-} // namespace
-
-void QuantizedMatmul::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 4);
-
-  auto& x_pre = inputs[0];
-  auto& w_pre = inputs[1];
-  auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
-
-  auto ensure_row_contiguous = [](const array& arr) {
-    if (arr.flags().row_contiguous) {
-      return arr;
-    } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
-      return arr_copy;
-    }
-  };
-
-  auto x = ensure_row_contiguous(x_pre);
-  auto w = ensure_row_contiguous(w_pre);
-  auto scales = ensure_row_contiguous(scales_pre);
-  auto biases = ensure_row_contiguous(biases_pre);
-
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
-}
-
-void GatherQMM::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 6);
-
-  auto& x_pre = inputs[0];
-  auto& w_pre = inputs[1];
-  auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
-  auto& lhs_indices = inputs[4];
-  auto& rhs_indices = inputs[5];
-
-  auto ensure_row_contiguous_last_dims = [](const array& arr) {
-    auto stride_0 = arr.strides()[arr.ndim() - 2];
-    auto stride_1 = arr.strides()[arr.ndim() - 1];
-    if (stride_0 == arr.shape(-1) && stride_1 == 1) {
-      return arr;
-    } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
-      return arr_copy;
-    }
-  };
-
-  auto x = ensure_row_contiguous_last_dims(x_pre);
-  auto w = ensure_row_contiguous_last_dims(w_pre);
-  auto scales = ensure_row_contiguous_last_dims(scales_pre);
-  auto biases = ensure_row_contiguous_last_dims(biases_pre);
-
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  _bs_qmm_dispatch(
-      out,
-      x,
-      w,
-      scales,
-      biases,
-      lhs_indices,
-      rhs_indices,
-      group_size_,
-      bits_,
-      transpose_);
-}
-
-template <typename T, typename U>
-void quantize(
-    const array& w_,
-    array& out_,
-    array& scales_,
-    array& biases_,
-    int bits,
-    int group_size) {
-  const T* w = w_.data<T>();
-
-  auto out = out_.data<U>();
-  T* scales = scales_.data<T>();
-  T* biases = biases_.data<T>();
-
-  T n_bins = (1 << bits) - 1;
-  T eps = 1e-7;
-  bool power_of_2_bits = is_power_of_2(bits);
-  int el_per_int = bits == 3 ? 8 : bits == 6 ? 4 : 32 / bits;
-  // For 3/6 bits we read 3 uint8s at a time instead of 1 uint32
-  int bytes_per_pack = power_of_2_bits ? 1 : 3;
-  int int_per_group = group_size * bytes_per_pack / el_per_int;
-  size_t n_groups = w_.size() / group_size;
-
-  for (size_t i = 0; i < n_groups; ++i) {
-    size_t w_idx = i * group_size;
-    T w_min = std::numeric_limits<float>::infinity();
-    T w_max = -w_min;
-    for (int j = 0; j < group_size; ++j) {
-      w_max = std::max(w_max, w[w_idx + j]);
-      w_min = std::min(w_min, w[w_idx + j]);
-    }
-    bool mask = std::abs(w_min) > std::abs(w_max);
-    T scale = std::max(T((w_max - w_min) / n_bins), eps);
-    scale = mask ? scale : -scale;
-
-    auto edge = mask ? w_min : w_max;
-    auto q0 = std::rint(edge / scale);
-    if (q0 == 0) {
-      scales[i] = scale;
-      biases[i] = 0;
-    } else {
-      scales[i] = edge / q0;
-      biases[i] = edge;
-    }
-    size_t out_idx = i * int_per_group;
-    for (int j = 0; j < int_per_group / bytes_per_pack; ++j) {
-      uint32_t out_el = 0;
-      for (int k = 0; k < el_per_int; ++k) {
-        T w_el = w[w_idx + j * el_per_int + k];
-        w_el = std::rint((w_el - biases[i]) / scales[i]);
-        w_el = std::min(std::max(w_el, T(0)), n_bins);
-        out_el |= static_cast<uint32_t>(w_el) << (k * bits);
-      }
-      if (power_of_2_bits) {
-        out[out_idx + j] = out_el;
-      } else {
-        out[out_idx + bytes_per_pack * j] = out_el & 0xff;
-        out[out_idx + bytes_per_pack * j + 1] = (out_el & 0xff00) >> 8;
-        out[out_idx + bytes_per_pack * j + 2] = (out_el & 0xff0000) >> 16;
-      }
-    }
-  }
-}
-
-void fast::AffineQuantize::eval_cpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  auto ensure_row_contiguous = [](const array& arr) {
-    if (arr.flags().row_contiguous) {
-      return arr;
-    } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
-      return arr_copy;
-    }
-  };
-  auto w = ensure_row_contiguous(inputs[0]);
-
-  auto& out = outputs[0];
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  auto& scales = outputs[1];
-  auto& biases = outputs[2];
-  scales.set_data(allocator::malloc_or_wait(scales.nbytes()));
-  biases.set_data(allocator::malloc_or_wait(biases.nbytes()));
-  if (w.dtype() == float16) {
-    if (is_power_of_2(bits_)) {
-      quantize<float16_t, uint32_t>(w, out, scales, biases, bits_, group_size_);
-    } else {
-      quantize<float16_t, uint8_t>(w, out, scales, biases, bits_, group_size_);
-    }
-  } else if (w.dtype() == bfloat16) {
-    if (is_power_of_2(bits_)) {
-      quantize<bfloat16_t, uint32_t>(
-          w, out, scales, biases, bits_, group_size_);
-    } else {
-      quantize<bfloat16_t, uint8_t>(w, out, scales, biases, bits_, group_size_);
-    }
-  } else if (w.dtype() == float32) {
-    if (is_power_of_2(bits_)) {
-      quantize<float, uint32_t>(w, out, scales, biases, bits_, group_size_);
-    } else {
-      quantize<float, uint8_t>(w, out, scales, biases, bits_, group_size_);
-    }
-  } else {
-    throw std::runtime_error(
-        "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/reduce.cpp

@@ -1,312 +1,147 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cassert>
-#include <functional>
-#include <limits>
+// Copyright © 2024 Apple Inc.
 
 #include "mlx/backend/common/reduce.h"
-#include "mlx/primitives.h"
 
 namespace mlx::core {
 
-namespace {
-
-template <typename U>
-struct Limits {
-  static const U max;
-  static const U min;
-};
-
-#define instantiate_default_limit(type)                           \
-  template <>                                                     \
-  struct Limits<type> {                                           \
-    static constexpr type max = std::numeric_limits<type>::max(); \
-    static constexpr type min = std::numeric_limits<type>::min(); \
-  };
-
-instantiate_default_limit(uint8_t);
-instantiate_default_limit(uint16_t);
-instantiate_default_limit(uint32_t);
-instantiate_default_limit(uint64_t);
-instantiate_default_limit(int8_t);
-instantiate_default_limit(int16_t);
-instantiate_default_limit(int32_t);
-instantiate_default_limit(int64_t);
-
-#define instantiate_float_limit(type) \
-  template <>                         \
-  struct Limits<type> {               \
-    static const type max;            \
-    static const type min;            \
-  };
-
-instantiate_float_limit(float16_t);
-instantiate_float_limit(bfloat16_t);
-instantiate_float_limit(float);
-instantiate_float_limit(complex64_t);
-
-template <>
-struct Limits<bool> {
-  static constexpr bool max = true;
-  static constexpr bool min = false;
-};
-
-const float Limits<float>::max = std::numeric_limits<float>::infinity();
-const float Limits<float>::min = -std::numeric_limits<float>::infinity();
-const bfloat16_t Limits<bfloat16_t>::max =
-    std::numeric_limits<float>::infinity();
-const bfloat16_t Limits<bfloat16_t>::min =
-    -std::numeric_limits<float>::infinity();
-const float16_t Limits<float16_t>::max = std::numeric_limits<float>::infinity();
-const float16_t Limits<float16_t>::min =
-    -std::numeric_limits<float>::infinity();
-const complex64_t Limits<complex64_t>::max =
-    std::numeric_limits<float>::infinity();
-const complex64_t Limits<complex64_t>::min =
-    -std::numeric_limits<float>::infinity();
-
-struct AndReduce {
-  template <typename T>
-  void operator()(bool* a, T b) {
-    (*a) &= (b != 0);
-  }
-
-  void operator()(bool* y, bool x) {
-    (*y) &= x;
-  }
-};
-
-struct OrReduce {
-  template <typename T>
-  void operator()(bool* a, T b) {
-    (*a) |= (b != 0);
-  }
-
-  void operator()(bool* y, bool x) {
-    (*y) |= x;
-  }
-};
-
-struct MaxReduce {
-  template <typename T>
-  std::enable_if_t<std::is_integral_v<T>> operator()(T* y, T x) {
-    (*y) = (*y > x) ? *y : x;
-  };
-
-  template <typename T>
-  std::enable_if_t<!std::is_integral_v<T>> operator()(T* y, T x) {
-    if (std::isnan(x)) {
-      *y = x;
-    } else {
-      (*y) = (*y > x) ? *y : x;
-    }
-  };
-};
-
-struct MinReduce {
-  template <typename T>
-  std::enable_if_t<std::is_integral_v<T>> operator()(T* y, T x) {
-    (*y) = (*y < x) ? *y : x;
-  };
-
-  template <typename T>
-  std::enable_if_t<!std::is_integral_v<T>> operator()(T* y, T x) {
-    if (std::isnan(x)) {
-      *y = x;
-    } else {
-      (*y) = (*y < x) ? *y : x;
-    }
-  };
-};
-
-template <typename InT>
-void reduce_dispatch_and_or(
-    const array& in,
-    array& out,
-    Reduce::ReduceType rtype,
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    const array& x,
     const std::vector<int>& axes) {
-  if (rtype == Reduce::And) {
-    reduction_op<InT, bool>(in, out, axes, true, AndReduce());
-  } else {
-    reduction_op<InT, bool>(in, out, axes, false, OrReduce());
+  auto shape = x.shape();
+  auto strides = x.strides();
+
+  for (int i = axes.size() - 1; i >= 0; i--) {
+    int a = axes[i];
+    shape.erase(shape.begin() + a);
+    strides.erase(strides.begin() + a);
   }
+
+  return std::make_pair(shape, strides);
 }
 
-template <typename InT>
-void reduce_dispatch_sum_prod(
-    const array& in,
-    array& out,
-    Reduce::ReduceType rtype,
-    const std::vector<int>& axes) {
-  if (rtype == Reduce::Sum) {
-    auto op = [](auto y, auto x) { (*y) = (*y) + x; };
-    if constexpr (std::is_integral_v<InT> && sizeof(InT) <= 4) {
-      reduction_op<InT, int32_t>(in, out, axes, 0, op);
-    } else {
-      reduction_op<InT, InT>(in, out, axes, 0, op);
+ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
+  // The data is all there and we are reducing over everything
+  if (x.size() == x.data_size() && axes.size() == x.ndim() &&
+      x.flags().contiguous) {
+    return ContiguousAllReduce;
+  }
+
+  // Row contiguous input so the output is row contiguous
+  if (x.flags().row_contiguous) {
+    // Merge consecutive axes
+    Shape shape = {x.shape(axes[0])};
+    Strides strides = {x.strides()[axes[0]]};
+    for (int i = 1; i < axes.size(); i++) {
+      if (axes[i] - 1 == axes[i - 1] && x.shape(axes[i]) > 1) {
+        shape.back() *= x.shape(axes[i]);
+        strides.back() = x.strides()[axes[i]];
+      } else {
+        shape.push_back(x.shape(axes[i]));
+        strides.push_back(x.strides()[axes[i]]);
+      }
     }
-  } else {
-    auto op = [](auto y, auto x) { (*y) *= x; };
-    if constexpr (std::is_integral_v<InT> && sizeof(InT) <= 4) {
-      reduction_op<InT, int32_t>(in, out, axes, 1, op);
-    } else {
-      reduction_op<InT, InT>(in, out, axes, 1, op);
+
+    // Remove singleton axes from the plan
+    for (int i = shape.size() - 1; i >= 0; i--) {
+      if (shape[i] == 1) {
+        shape.erase(shape.begin() + i);
+        strides.erase(strides.begin() + i);
+      }
+    }
+
+    if (strides.back() == 1) {
+      return ReductionPlan(ContiguousReduce, shape, strides);
+    } else if (strides.back() > 1) {
+      return ReductionPlan(ContiguousStridedReduce, shape, strides);
     }
   }
-}
 
-template <typename InT>
-void reduce_dispatch_min_max(
-    const array& in,
-    array& out,
-    Reduce::ReduceType rtype,
-    const std::vector<int>& axes) {
-  if (rtype == Reduce::Max) {
-    auto init = Limits<InT>::min;
-    reduction_op<InT, InT>(in, out, axes, init, MaxReduce());
-  } else {
-    auto init = Limits<InT>::max;
-    reduction_op<InT, InT>(in, out, axes, init, MinReduce());
-  }
-}
+  // Let's check if we can optimize our access patterns
+  //
+  // 1. We have a reduction axis with stride 1. Simply call
+  //    GeneralContiguousReduce and be done with it.
+  // 2. We have transpositions and we are not reducing over the axis with
+  //    stride 1. However, we are reducing over an axis where everything is
+  //    contiguous in memory to the right of that axis. We can call strided
+  //    reduce and be done with it.
+  // 2. We have weird transpositions and expands. Copy the strides to the
+  //    output, then call strided reduce.
 
-} // namespace
-
-void nd_loop(
-    std::function<void(int)> callback,
-    const Shape& shape,
-    const Strides& strides) {
-  std::function<void(int, int)> loop_inner;
-  loop_inner = [&](int dim, int offset) {
-    if (dim < shape.size() - 1) {
-      auto size = shape[dim];
-      auto stride = strides[dim];
-      for (int i = 0; i < size; i++) {
-        loop_inner(dim + 1, offset + i * stride);
-      }
-    } else {
-      auto size = shape[dim];
-      auto stride = strides[dim];
-      for (int i = 0; i < size; i++) {
-        callback(offset + i * stride);
-      }
-    }
-  };
-  loop_inner(0, 0);
-}
-
-void Reduce::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  switch (reduce_type_) {
-    case Reduce::And:
-    case Reduce::Or: {
-      switch (in.dtype()) {
-        case bool_:
-        case uint8:
-        case int8:
-          reduce_dispatch_and_or<int8_t>(in, out, reduce_type_, axes_);
-          break;
-        case int16:
-        case uint16:
-        case float16:
-        case bfloat16:
-          reduce_dispatch_and_or<int16_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint32:
-        case int32:
-        case float32:
-          reduce_dispatch_and_or<int32_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint64:
-        case int64:
-        case complex64:
-          reduce_dispatch_and_or<int64_t>(in, out, reduce_type_, axes_);
-          break;
-      }
-      break;
-    }
-    case Reduce::Sum:
-    case Reduce::Prod: {
-      switch (in.dtype()) {
-        case bool_:
-        case uint8:
-        case int8:
-          reduce_dispatch_sum_prod<int8_t>(in, out, reduce_type_, axes_);
-          break;
-        case int16:
-        case uint16:
-          reduce_dispatch_sum_prod<int16_t>(in, out, reduce_type_, axes_);
-          break;
-        case int32:
-        case uint32:
-          reduce_dispatch_sum_prod<int32_t>(in, out, reduce_type_, axes_);
-          break;
-        case int64:
-        case uint64:
-          reduce_dispatch_sum_prod<int64_t>(in, out, reduce_type_, axes_);
-          break;
-        case float16:
-          reduce_dispatch_sum_prod<float16_t>(in, out, reduce_type_, axes_);
-          break;
-        case bfloat16:
-          reduce_dispatch_sum_prod<bfloat16_t>(in, out, reduce_type_, axes_);
-          break;
-        case float32:
-          reduce_dispatch_sum_prod<float>(in, out, reduce_type_, axes_);
-          break;
-        case complex64:
-          reduce_dispatch_sum_prod<complex64_t>(in, out, reduce_type_, axes_);
-          break;
-      }
-      break;
-    }
-    case Reduce::Max:
-    case Reduce::Min: {
-      switch (in.dtype()) {
-        case bool_:
-          reduce_dispatch_min_max<bool>(in, out, reduce_type_, axes_);
-          break;
-        case uint8:
-          reduce_dispatch_min_max<uint8_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint16:
-          reduce_dispatch_min_max<uint16_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint32:
-          reduce_dispatch_min_max<uint32_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint64:
-          reduce_dispatch_min_max<uint64_t>(in, out, reduce_type_, axes_);
-          break;
-        case int8:
-          reduce_dispatch_min_max<uint8_t>(in, out, reduce_type_, axes_);
-          break;
-        case int16:
-          reduce_dispatch_min_max<uint16_t>(in, out, reduce_type_, axes_);
-          break;
-        case int32:
-          reduce_dispatch_min_max<int32_t>(in, out, reduce_type_, axes_);
-          break;
-        case int64:
-          reduce_dispatch_min_max<int64_t>(in, out, reduce_type_, axes_);
-          break;
-        case float16:
-          reduce_dispatch_min_max<float16_t>(in, out, reduce_type_, axes_);
-          break;
-        case float32:
-          reduce_dispatch_min_max<float>(in, out, reduce_type_, axes_);
-          break;
-        case bfloat16:
-          reduce_dispatch_min_max<bfloat16_t>(in, out, reduce_type_, axes_);
-          break;
-        case complex64:
-          reduce_dispatch_min_max<complex64_t>(in, out, reduce_type_, axes_);
-          break;
-      }
-      break;
+  // Sort reduction axes by stride in order to merge them and figure out if we
+  // have a contiguous reduction.
+  std::vector<std::pair<int, int64_t>> reductions;
+  for (auto a : axes) {
+    if (x.shape(a) > 1) {
+      reductions.push_back(std::make_pair(x.shape(a), x.strides()[a]));
     }
   }
+  std::sort(reductions.begin(), reductions.end(), [](auto a, auto b) {
+    bool a_is_zero = a.second == 0;
+    bool b_is_zero = b.second == 0;
+    return (a_is_zero != b_is_zero) ? a.second < b.second : a.second > b.second;
+  });
+  // Extract the two smallest and try to merge them in case the contiguous
+  // reduction can be bigger than just the last axis.
+  for (int i = reductions.size() - 1; i >= 1; i--) {
+    auto a = reductions[i];
+    auto b = reductions[i - 1];
+
+    // b.stride = a.shape * a.stride then a and b are contiguous
+    if (b.second == a.first * a.second) {
+      reductions.erase(reductions.begin() + i);
+      reductions[i - 1] = std::make_pair(a.first * b.first, a.second);
+    }
+  }
+
+  Shape shape;
+  Strides strides;
+  for (auto r : reductions) {
+    shape.push_back(r.first);
+    strides.push_back(r.second);
+  }
+
+  // We can call the contiguous reduction op for every weird way the input is
+  // structured in the rest of the axes.
+  if (strides.back() == 1) {
+    return ReductionPlan(GeneralContiguousReduce, shape, strides);
+  }
+
+  // Delegate to the general strided reduction op if the axes after
+  // strides.back() are contiguous.
+  if (strides.back() > 1) {
+    int64_t size = 1;
+    bool have_expand = false;
+    for (int i = x.ndim() - 1; i >= 0; i--) {
+      if (axes.back() == i) {
+        continue;
+      }
+
+      auto stride_i = x.strides()[i];
+      auto shape_i = x.shape(i);
+      if (stride_i == 0) {
+        if (shape_i == 1) {
+          continue;
+        }
+
+        have_expand = true;
+        break;
+      }
+
+      if (stride_i != size && shape_i != 1) {
+        break;
+      }
+      size *= shape_i;
+    }
+    // In the case of an expanded dimension we are being conservative and
+    // require the smallest reduction stride to be smaller than the maximum row
+    // contiguous size. The reason is that we can't easily know if the reduced
+    // axis is before or after an expanded dimension.
+    if (size > strides.back() || (size == strides.back() && !have_expand)) {
+      return ReductionPlan(GeneralStridedReduce, shape, strides);
+    }
+  }
+
+  return ReductionPlan(GeneralReduce, shape, strides);
 }
 
 } // namespace mlx::core

mlx/backend/common/reduce.h

@@ -48,186 +48,8 @@ struct ReductionPlan {
 
 ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes);
 
-// Helper for the ndimensional strided loop
-// Should this be in utils?
-void nd_loop(
-    std::function<void(int)> callback,
-    const Shape& shape,
-    const Strides& strides);
-
 std::pair<Shape, Strides> shapes_without_reduction_axes(
     const array& x,
     const std::vector<int>& axes);
 
-template <typename T, typename U, typename Op>
-struct DefaultStridedReduce {
-  Op op;
-
-  DefaultStridedReduce(Op op_) : op(op_) {}
-
-  void operator()(const T* x, U* accumulator, int size, size_t stride) {
-    for (int i = 0; i < size; i++) {
-      U* moving_accumulator = accumulator;
-      for (int j = 0; j < stride; j++) {
-        op(moving_accumulator, *x);
-        moving_accumulator++;
-        x++;
-      }
-    }
-  }
-};
-
-template <typename T, typename U, typename Op>
-struct DefaultContiguousReduce {
-  Op op;
-
-  DefaultContiguousReduce(Op op_) : op(op_) {}
-
-  void operator()(const T* x, U* accumulator, int size) {
-    while (size-- > 0) {
-      op(accumulator, *x);
-      x++;
-    }
-  }
-};
-
-template <typename T, typename U, typename OpS, typename OpC, typename Op>
-void reduction_op(
-    const array& x,
-    array& out,
-    const std::vector<int>& axes,
-    U init,
-    OpS ops,
-    OpC opc,
-    Op op) {
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  ReductionPlan plan = get_reduction_plan(x, axes);
-
-  if (plan.type == ContiguousAllReduce) {
-    U* out_ptr = out.data<U>();
-    *out_ptr = init;
-    opc(x.data<T>(), out_ptr, x.size());
-    return;
-  }
-
-  if (plan.type == ContiguousReduce && plan.shape.size() == 1) {
-    int reduction_size = plan.shape[0];
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
-    for (int i = 0; i < out.size(); i++, out_ptr++, x_ptr += reduction_size) {
-      *out_ptr = init;
-      opc(x_ptr, out_ptr, reduction_size);
-    }
-    return;
-  }
-
-  if (plan.type == GeneralContiguousReduce || plan.type == ContiguousReduce) {
-    int reduction_size = plan.shape.back();
-    plan.shape.pop_back();
-    plan.strides.pop_back();
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
-    // Unrolling the following loop (and implementing it in order for
-    // ContiguousReduce) should hold extra performance boost.
-    auto [shape, strides] = shapes_without_reduction_axes(x, axes);
-    if (plan.shape.size() == 0) {
-      for (int i = 0; i < out.size(); i++, out_ptr++) {
-        int offset = elem_to_loc(i, shape, strides);
-        *out_ptr = init;
-        opc(x_ptr + offset, out_ptr, reduction_size);
-      }
-    } else {
-      for (int i = 0; i < out.size(); i++, out_ptr++) {
-        int offset = elem_to_loc(i, shape, strides);
-        *out_ptr = init;
-        nd_loop(
-            [&](int extra_offset) {
-              opc(x_ptr + offset + extra_offset, out_ptr, reduction_size);
-            },
-            plan.shape,
-            plan.strides);
-      }
-    }
-    return;
-  }
-
-  if (plan.type == ContiguousStridedReduce && plan.shape.size() == 1) {
-    int reduction_size = plan.shape.back();
-    size_t reduction_stride = plan.strides.back();
-    plan.shape.pop_back();
-    plan.strides.pop_back();
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
-    for (int i = 0; i < out.size(); i += reduction_stride) {
-      std::fill_n(out_ptr, reduction_stride, init);
-      ops(x_ptr, out_ptr, reduction_size, reduction_stride);
-      x_ptr += reduction_stride * reduction_size;
-      out_ptr += reduction_stride;
-    }
-    return;
-  }
-
-  if (plan.type == GeneralStridedReduce ||
-      plan.type == ContiguousStridedReduce) {
-    int reduction_size = plan.shape.back();
-    size_t reduction_stride = plan.strides.back();
-    plan.shape.pop_back();
-    plan.strides.pop_back();
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
-    auto [shape, strides] = shapes_without_reduction_axes(x, axes);
-    if (plan.shape.size() == 0) {
-      for (int i = 0; i < out.size(); i += reduction_stride) {
-        int offset = elem_to_loc(i, shape, strides);
-        std::fill_n(out_ptr, reduction_stride, init);
-        ops(x_ptr + offset, out_ptr, reduction_size, reduction_stride);
-        out_ptr += reduction_stride;
-      }
-    } else {
-      for (int i = 0; i < out.size(); i += reduction_stride) {
-        int offset = elem_to_loc(i, shape, strides);
-        std::fill_n(out_ptr, reduction_stride, init);
-        nd_loop(
-            [&](int extra_offset) {
-              ops(x_ptr + offset + extra_offset,
-                  out_ptr,
-                  reduction_size,
-                  reduction_stride);
-            },
-            plan.shape,
-            plan.strides);
-        out_ptr += reduction_stride;
-      }
-    }
-    return;
-  }
-
-  if (plan.type == GeneralReduce) {
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
-    auto [shape, strides] = shapes_without_reduction_axes(x, axes);
-    for (int i = 0; i < out.size(); i++, out_ptr++) {
-      int offset = elem_to_loc(i, shape, strides);
-      U val = init;
-      nd_loop(
-          [&](int extra_offset) { op(&val, *(x_ptr + offset + extra_offset)); },
-          plan.shape,
-          plan.strides);
-      *out_ptr = val;
-    }
-  }
-}
-
-template <typename T, typename U, typename Op>
-void reduction_op(
-    const array& x,
-    array& out,
-    const std::vector<int>& axes,
-    U init,
-    Op op) {
-  DefaultStridedReduce<T, U, Op> ops(op);
-  DefaultContiguousReduce<T, U, Op> opc(op);
-  reduction_op<T, U>(x, out, axes, init, ops, opc, op);
-}
-
 } // namespace mlx::core

mlx/backend/common/reduce_utils.cpp

@@ -1,147 +0,0 @@
-// Copyright © 2024 Apple Inc.
-
-#include "mlx/backend/common/reduce.h"
-
-namespace mlx::core {
-
-std::pair<Shape, Strides> shapes_without_reduction_axes(
-    const array& x,
-    const std::vector<int>& axes) {
-  auto shape = x.shape();
-  auto strides = x.strides();
-
-  for (int i = axes.size() - 1; i >= 0; i--) {
-    int a = axes[i];
-    shape.erase(shape.begin() + a);
-    strides.erase(strides.begin() + a);
-  }
-
-  return std::make_pair(shape, strides);
-}
-
-ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
-  // The data is all there and we are reducing over everything
-  if (x.size() == x.data_size() && axes.size() == x.ndim() &&
-      x.flags().contiguous) {
-    return ContiguousAllReduce;
-  }
-
-  // Row contiguous input so the output is row contiguous
-  if (x.flags().row_contiguous) {
-    // Merge consecutive axes
-    Shape shape = {x.shape(axes[0])};
-    Strides strides = {x.strides()[axes[0]]};
-    for (int i = 1; i < axes.size(); i++) {
-      if (axes[i] - 1 == axes[i - 1] && x.shape(axes[i]) > 1) {
-        shape.back() *= x.shape(axes[i]);
-        strides.back() = x.strides()[axes[i]];
-      } else {
-        shape.push_back(x.shape(axes[i]));
-        strides.push_back(x.strides()[axes[i]]);
-      }
-    }
-
-    // Remove singleton axes from the plan
-    for (int i = shape.size() - 1; i >= 0; i--) {
-      if (shape[i] == 1) {
-        shape.erase(shape.begin() + i);
-        strides.erase(strides.begin() + i);
-      }
-    }
-
-    if (strides.back() == 1) {
-      return ReductionPlan(ContiguousReduce, shape, strides);
-    } else if (strides.back() > 1) {
-      return ReductionPlan(ContiguousStridedReduce, shape, strides);
-    }
-  }
-
-  // Let's check if we can optimize our access patterns
-  //
-  // 1. We have a reduction axis with stride 1. Simply call
-  //    GeneralContiguousReduce and be done with it.
-  // 2. We have transpositions and we are not reducing over the axis with
-  //    stride 1. However, we are reducing over an axis where everything is
-  //    contiguous in memory to the right of that axis. We can call strided
-  //    reduce and be done with it.
-  // 2. We have weird transpositions and expands. Copy the strides to the
-  //    output, then call strided reduce.
-
-  // Sort reduction axes by stride in order to merge them and figure out if we
-  // have a contiguous reduction.
-  std::vector<std::pair<int, int64_t>> reductions;
-  for (auto a : axes) {
-    if (x.shape(a) > 1) {
-      reductions.push_back(std::make_pair(x.shape(a), x.strides()[a]));
-    }
-  }
-  std::sort(reductions.begin(), reductions.end(), [](auto a, auto b) {
-    bool a_is_zero = a.second == 0;
-    bool b_is_zero = b.second == 0;
-    return (a_is_zero != b_is_zero) ? a.second < b.second : a.second > b.second;
-  });
-  // Extract the two smallest and try to merge them in case the contiguous
-  // reduction can be bigger than just the last axis.
-  for (int i = reductions.size() - 1; i >= 1; i--) {
-    auto a = reductions[i];
-    auto b = reductions[i - 1];
-
-    // b.stride = a.shape * a.stride then a and b are contiguous
-    if (b.second == a.first * a.second) {
-      reductions.erase(reductions.begin() + i);
-      reductions[i - 1] = std::make_pair(a.first * b.first, a.second);
-    }
-  }
-
-  Shape shape;
-  Strides strides;
-  for (auto r : reductions) {
-    shape.push_back(r.first);
-    strides.push_back(r.second);
-  }
-
-  // We can call the contiguous reduction op for every weird way the input is
-  // structured in the rest of the axes.
-  if (strides.back() == 1) {
-    return ReductionPlan(GeneralContiguousReduce, shape, strides);
-  }
-
-  // Delegate to the general strided reduction op if the axes after
-  // strides.back() are contiguous.
-  if (strides.back() > 1) {
-    int64_t size = 1;
-    bool have_expand = false;
-    for (int i = x.ndim() - 1; i >= 0; i--) {
-      if (axes.back() == i) {
-        continue;
-      }
-
-      auto stride_i = x.strides()[i];
-      auto shape_i = x.shape(i);
-      if (stride_i == 0) {
-        if (shape_i == 1) {
-          continue;
-        }
-
-        have_expand = true;
-        break;
-      }
-
-      if (stride_i != size && shape_i != 1) {
-        break;
-      }
-      size *= shape_i;
-    }
-    // In the case of an expanded dimension we are being conservative and
-    // require the smallest reduction stride to be smaller than the maximum row
-    // contiguous size. The reason is that we can't easily know if the reduced
-    // axis is before or after an expanded dimension.
-    if (size > strides.back() || (size == strides.back() && !have_expand)) {
-      return ReductionPlan(GeneralStridedReduce, shape, strides);
-    }
-  }
-
-  return ReductionPlan(GeneralReduce, shape, strides);
-}
-
-} // namespace mlx::core

mlx/backend/common/scan.cpp

@@ -1,325 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cassert>
-
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/utils.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-namespace {
-
-template <typename T, typename U, typename Op>
-struct DefaultContiguousScan {
-  Op op;
-  U init;
-
-  DefaultContiguousScan(Op op_, U init_) : op(op_), init(init_) {}
-
-  void operator()(
-      const T* input,
-      U* output,
-      int count,
-      int stride,
-      bool reverse,
-      bool inclusive) {
-    if (!reverse) {
-      if (inclusive) {
-        for (int i = 0; i < count; i++) {
-          *output = *input;
-          for (int j = 1; j < stride; j++) {
-            input++;
-            output++;
-            op(output, output - 1, input);
-          }
-          output++;
-          input++;
-        }
-      } else {
-        for (int i = 0; i < count; i++) {
-          *output = init;
-          for (int j = 1; j < stride; j++) {
-            op(output + 1, output, input);
-            input++;
-            output++;
-          }
-          output++;
-          input++;
-        }
-      }
-    } else {
-      if (inclusive) {
-        for (int i = 0; i < count; i++) {
-          output += stride - 1;
-          input += stride - 1;
-          *output = *input;
-          for (int j = 1; j < stride; j++) {
-            input--;
-            output--;
-            op(output, output + 1, input);
-          }
-          output += stride;
-          input += stride;
-        }
-      } else {
-        for (int i = 0; i < count; i++) {
-          output += stride - 1;
-          input += stride - 1;
-          *output = init;
-          for (int j = 1; j < stride; j++) {
-            op(output - 1, output, input);
-            input--;
-            output--;
-          }
-          output += stride;
-          input += stride;
-        }
-      }
-    }
-  }
-};
-
-template <typename T, typename U, typename Op>
-struct DefaultStridedScan {
-  Op op;
-  U init;
-
-  DefaultStridedScan(Op op_, U init_) : op(op_), init(init_) {}
-
-  void operator()(
-      const T* input,
-      U* output,
-      int count,
-      int size,
-      int stride,
-      bool reverse,
-      bool inclusive) {
-    // TODO: Vectorize the following naive implementation
-    if (!reverse) {
-      if (inclusive) {
-        for (int i = 0; i < count; i++) {
-          std::copy(input, input + stride, output);
-          output += stride;
-          input += stride;
-          for (int j = 1; j < size; j++) {
-            for (int k = 0; k < stride; k++) {
-              op(output, output - stride, input);
-              output++;
-              input++;
-            }
-          }
-        }
-      } else {
-        for (int i = 0; i < count; i++) {
-          std::fill(output, output + stride, init);
-          output += stride;
-          input += stride;
-          for (int j = 1; j < size; j++) {
-            for (int k = 0; k < stride; k++) {
-              op(output, output - stride, input - stride);
-              output++;
-              input++;
-            }
-          }
-        }
-      }
-    } else {
-      if (inclusive) {
-        for (int i = 0; i < count; i++) {
-          output += (size - 1) * stride;
-          input += (size - 1) * stride;
-          std::copy(input, input + stride, output);
-          for (int j = 1; j < size; j++) {
-            for (int k = 0; k < stride; k++) {
-              output--;
-              input--;
-              op(output, output + stride, input);
-            }
-          }
-          output += size * stride;
-          input += size * stride;
-        }
-      } else {
-        for (int i = 0; i < count; i++) {
-          output += (size - 1) * stride;
-          input += (size - 1) * stride;
-          std::fill(output, output + stride, init);
-          for (int j = 1; j < size; j++) {
-            for (int k = 0; k < stride; k++) {
-              output--;
-              input--;
-              op(output, output + stride, input + stride);
-            }
-          }
-          output += size * stride;
-          input += size * stride;
-        }
-      }
-    }
-  }
-};
-
-template <typename T, typename U, typename OpCS, typename OpSS>
-void scan_op(
-    OpCS opcs,
-    OpSS opss,
-    const array& input,
-    array& output,
-    int axis,
-    bool reverse,
-    bool inclusive) {
-  output.set_data(allocator::malloc_or_wait(output.nbytes()));
-
-  if (input.flags().row_contiguous) {
-    if (input.strides()[axis] == 1) {
-      opcs(
-          input.data<T>(),
-          output.data<U>(),
-          input.size() / input.shape(axis),
-          input.shape(axis),
-          reverse,
-          inclusive);
-    } else {
-      opss(
-          input.data<T>(),
-          output.data<U>(),
-          input.size() / input.shape(axis) / input.strides()[axis],
-          input.shape(axis),
-          input.strides()[axis],
-          reverse,
-          inclusive);
-    }
-  } else {
-    throw std::runtime_error("Scan op supports only contiguous inputs");
-  }
-}
-
-template <typename T, typename U>
-void scan_dispatch(
-    Scan::ReduceType rtype,
-    const array& input,
-    array& output,
-    int axis,
-    bool reverse,
-    bool inclusive) {
-  switch (rtype) {
-    case Scan::Sum: {
-      auto op = [](U* o, const U* y, const T* x) { *o = *y + *x; };
-      auto init = static_cast<U>(0);
-      auto opcs = DefaultContiguousScan<T, U, decltype(op)>(op, init);
-      auto opss = DefaultStridedScan<T, U, decltype(op)>(op, init);
-      scan_op<T, U>(opcs, opss, input, output, axis, reverse, inclusive);
-      break;
-    }
-    case Scan::Prod: {
-      auto op = [](U* o, const U* y, const T* x) { *o = *y * (*x); };
-      auto init = static_cast<U>(1);
-      auto opcs = DefaultContiguousScan<T, U, decltype(op)>(op, init);
-      auto opss = DefaultStridedScan<T, U, decltype(op)>(op, init);
-      scan_op<T, U>(opcs, opss, input, output, axis, reverse, inclusive);
-      break;
-    }
-    case Scan::Min: {
-      auto op = [](U* o, const U* y, const T* x) { *o = (*x < *y) ? *x : *y; };
-      auto init = (issubdtype(input.dtype(), floating))
-          ? static_cast<U>(std::numeric_limits<float>::infinity())
-          : std::numeric_limits<U>::max();
-      auto opcs = DefaultContiguousScan<T, U, decltype(op)>(op, init);
-      auto opss = DefaultStridedScan<T, U, decltype(op)>(op, init);
-      scan_op<T, U>(opcs, opss, input, output, axis, reverse, inclusive);
-      break;
-    }
-    case Scan::Max: {
-      auto op = [](U* o, const U* y, const T* x) { *o = (*x < *y) ? *y : *x; };
-      auto init = (issubdtype(input.dtype(), floating))
-          ? static_cast<U>(-std::numeric_limits<float>::infinity())
-          : std::numeric_limits<U>::min();
-      auto opcs = DefaultContiguousScan<T, U, decltype(op)>(op, init);
-      auto opss = DefaultStridedScan<T, U, decltype(op)>(op, init);
-      scan_op<T, U>(opcs, opss, input, output, axis, reverse, inclusive);
-      break;
-    }
-  }
-}
-
-} // namespace
-
-void Scan::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-
-  // Ensure contiguity
-  auto in = inputs[0];
-  if (!in.flags().row_contiguous) {
-    array arr_copy(in.shape(), in.dtype(), nullptr, {});
-    copy(in, arr_copy, CopyType::General);
-    in = arr_copy;
-  }
-
-  switch (in.dtype()) {
-    case bool_: {
-      // We could do a full dtype x dtype switch but this is the only case
-      // where we accumulate in a different type, for now.
-      //
-      // TODO: If we add the option to accumulate floats in higher precision
-      //       floats perhaps we should add the full all-to-all dispatch.
-      if (reduce_type_ == Scan::Sum && out.dtype() == int32) {
-        scan_dispatch<bool, int32_t>(
-            reduce_type_, in, out, axis_, reverse_, inclusive_);
-      } else {
-        scan_dispatch<bool, bool>(
-            reduce_type_, in, out, axis_, reverse_, inclusive_);
-      }
-      break;
-    }
-    case uint8:
-      scan_dispatch<uint8_t, uint8_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case uint16:
-      scan_dispatch<uint16_t, uint16_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case uint32:
-      scan_dispatch<uint32_t, uint32_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case uint64:
-      scan_dispatch<uint64_t, uint64_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case int8:
-      scan_dispatch<int8_t, int8_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case int16:
-      scan_dispatch<int16_t, int16_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case int32:
-      scan_dispatch<int32_t, int32_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case int64:
-      scan_dispatch<int64_t, int64_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case float16:
-      scan_dispatch<float16_t, float16_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case float32:
-      scan_dispatch<float, float>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case bfloat16:
-      scan_dispatch<bfloat16_t, bfloat16_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case complex64:
-      throw std::runtime_error("Scan ops do not support complex types yet");
-      break;
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/select.cpp

@@ -1,72 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cassert>
-
-#include "mlx/backend/common/ternary.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-namespace {
-
-template <typename Op>
-void select_op(
-    const array& a,
-    const array& b,
-    const array& c,
-    array& out,
-    Op op) {
-  switch (out.dtype()) {
-    case bool_:
-      ternary_op<bool, bool, bool, bool>(a, b, c, out, op);
-      break;
-    case uint8:
-      ternary_op<bool, uint8_t, uint8_t, uint8_t>(a, b, c, out, op);
-      break;
-    case uint16:
-      ternary_op<bool, uint16_t, uint16_t, uint16_t>(a, b, c, out, op);
-      break;
-    case uint32:
-      ternary_op<bool, uint32_t, uint32_t, uint32_t>(a, b, c, out, op);
-      break;
-    case uint64:
-      ternary_op<bool, uint64_t, uint64_t, uint64_t>(a, b, c, out, op);
-      break;
-    case int8:
-      ternary_op<bool, int8_t, int8_t, int8_t>(a, b, c, out, op);
-      break;
-    case int16:
-      ternary_op<bool, int16_t, int16_t, int16_t>(a, b, c, out, op);
-      break;
-    case int32:
-      ternary_op<bool, int32_t, int32_t, int32_t>(a, b, c, out, op);
-      break;
-    case int64:
-      ternary_op<bool, int64_t, int64_t, int64_t>(a, b, c, out, op);
-      break;
-    case float16:
-      ternary_op<bool, float16_t, float16_t, float16_t>(a, b, c, out, op);
-      break;
-    case float32:
-      ternary_op<bool, float, float, float>(a, b, c, out, op);
-      break;
-    case bfloat16:
-      ternary_op<bool, bfloat16_t, bfloat16_t, bfloat16_t>(a, b, c, out, op);
-      break;
-    case complex64:
-      ternary_op<bool, complex64_t, complex64_t, complex64_t>(a, b, c, out, op);
-      break;
-  }
-}
-
-} // namespace
-
-void Select::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 3);
-  const auto& condition = inputs[0];
-  const auto& a = inputs[1];
-  const auto& b = inputs[2];
-  select_op(condition, a, b, out, detail::Select());
-}
-
-} // namespace mlx::core

mlx/backend/common/slicing.cpp

@@ -14,6 +14,10 @@ std::tuple<int64_t, Strides> prepare_slice(
     data_offset += start_indices[i] * in.strides()[i];
     inp_strides[i] = in.strides()[i] * strides[i];
   }
+  // Normalize the offset
+  if (data_offset < 0) {
+    data_offset += in.data_size();
+  }
   return std::make_tuple(data_offset, inp_strides);
 }
 
@@ -32,7 +36,33 @@ void shared_buffer_slice(
   flags.col_contiguous = is_col_contiguous;
   flags.contiguous = (no_bsx_size == data_size);
 
-  move_or_copy(in, out, out_strides, flags, data_size, data_offset);
+  out.copy_shared_buffer(in, out_strides, flags, data_size, data_offset);
+}
+
+void slice(
+    const array& in,
+    array& out,
+    const Shape& start_indices,
+    const Shape& strides) {
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  // Calculate out strides, initial offset
+  auto [data_offset, inp_strides] = prepare_slice(in, start_indices, strides);
+  int64_t data_end = 1;
+  for (int i = 0; i < start_indices.size(); ++i) {
+    if (in.shape()[i] > 1) {
+      auto end_idx = start_indices[i] + out.shape()[i] * strides[i] - 1;
+      data_end += end_idx * in.strides()[i];
+    }
+  }
+  if (data_end < 0) {
+    data_end += in.data_size();
+  }
+  size_t data_size = (data_end - data_offset);
+  shared_buffer_slice(in, inp_strides, data_offset, data_size, out);
 }
 
 } // namespace mlx::core

mlx/backend/common/slicing.h

@@ -11,11 +11,10 @@ std::tuple<int64_t, Strides> prepare_slice(
     const Shape& start_indices,
     const Shape& strides);
 
-void shared_buffer_slice(
+void slice(
     const array& in,
-    const Strides& out_strides,
-    size_t data_offset,
-    size_t data_size,
-    array& out);
+    array& out,
+    const Shape& start_indices,
+    const Shape& strides);
 
 } // namespace mlx::core

mlx/backend/common/softmax.cpp

@@ -1,127 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include <cassert>
-#include <cmath>
-
-#include "mlx/backend/common/copy.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-namespace {
-
-template <typename T, typename AccT>
-void softmax(const array& in, array& out) {
-  const T* in_ptr = in.data<T>();
-  T* out_ptr = out.data<T>();
-  int N = in.shape().back();
-  int M = in.data_size() / N;
-  const T* current_in_ptr;
-  T* current_out_ptr;
-
-  for (int i = 0; i < M; i++, in_ptr += N, out_ptr += N) {
-    // Find the maximum
-    current_in_ptr = in_ptr;
-    AccT maximum = *current_in_ptr;
-    for (int j = 0; j < N; j++, current_in_ptr++) {
-      maximum = (maximum < *current_in_ptr) ? static_cast<AccT>(*current_in_ptr)
-                                            : maximum;
-    }
-
-    // Compute the normalizer and the exponentials
-    AccT normalizer = 0;
-    current_out_ptr = out_ptr;
-    current_in_ptr = in_ptr;
-    for (int j = 0; j < N; j++, current_out_ptr++, current_in_ptr++) {
-      AccT expv = std::exp(*current_in_ptr - maximum);
-      normalizer += expv;
-      if constexpr (std::is_same<T, AccT>::value) {
-        *current_out_ptr = expv;
-      }
-    }
-    normalizer = 1 / normalizer;
-
-    // Normalize
-    current_in_ptr = in_ptr;
-    current_out_ptr = out_ptr;
-    for (int j = 0; j < N; j++, current_out_ptr++) {
-      if constexpr (std::is_same<T, AccT>::value) {
-        *current_out_ptr *= normalizer;
-      } else {
-        auto v = std::exp(*current_in_ptr - maximum);
-        *current_out_ptr = static_cast<T>(v * normalizer);
-        current_in_ptr++;
-      }
-    }
-  }
-}
-
-} // namespace
-
-void Softmax::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-
-  // Make sure that the last dimension is contiguous
-  auto check_input = [](array x) {
-    bool no_copy = x.strides()[x.ndim() - 1] == 1;
-    if (x.ndim() > 1) {
-      auto s = x.strides()[x.ndim() - 2];
-      no_copy &= (s == 0 || s == x.shape().back());
-    }
-    if (no_copy) {
-      return x;
-    } else {
-      array x_copy(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General);
-      return x_copy;
-    }
-  };
-  array in = check_input(std::move(inputs[0]));
-  if (in.is_donatable()) {
-    out.copy_shared_buffer(in);
-  } else {
-    out.set_data(
-        allocator::malloc_or_wait(in.data_size() * in.itemsize()),
-        in.data_size(),
-        in.strides(),
-        in.flags());
-  }
-
-  switch (in.dtype()) {
-    case bool_:
-    case uint8:
-    case uint16:
-    case uint32:
-    case uint64:
-    case int8:
-    case int16:
-    case int32:
-    case int64:
-      throw std::invalid_argument(
-          "Softmax is defined only for floating point types");
-      break;
-    case float32:
-      softmax<float, float>(in, out);
-      break;
-    case float16:
-      if (precise_) {
-        softmax<float16_t, float>(in, out);
-      } else {
-        softmax<float16_t, float16_t>(in, out);
-      }
-      break;
-    case bfloat16:
-      if (precise_) {
-        softmax<bfloat16_t, float>(in, out);
-      } else {
-        softmax<bfloat16_t, bfloat16_t>(in, out);
-      }
-      break;
-    case complex64:
-      throw std::invalid_argument(
-          "[Softmax] Not yet implemented for complex64");
-      break;
-  }
-}
-
-} // namespace mlx::core

mlx/backend/common/svd.cpp

@@ -1,147 +0,0 @@
-// Copyright © 2024 Apple Inc.
-
-#include "mlx/allocator.h"
-#include "mlx/backend/common/copy.h"
-#include "mlx/backend/common/lapack.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-void svd_impl(const array& a, array& u, array& s, array& vt) {
-  // Lapack uses the column-major convention. To avoid having to transpose
-  // the input and then transpose the outputs, we swap the indices/sizes of the
-  // matrices and take advantage of the following identity (see
-  // https://math.stackexchange.com/a/30077)
-  //    A = UΣVᵀ
-  //    Aᵀ = VΣUᵀ
-  // As a result some of the indices/sizes are swapped as noted above.
-
-  // Rows and cols of the original matrix in row-major order.
-  const int M = a.shape(-2);
-  const int N = a.shape(-1);
-  const int K = std::min(M, N);
-
-  // A of shape M x N. The leading dimension is N since lapack receives Aᵀ.
-  const int lda = N;
-  // U of shape M x M. (N x N in lapack).
-  const int ldu = N;
-  // Vᵀ of shape N x N. (M x M in lapack).
-  const int ldvt = M;
-
-  size_t num_matrices = a.size() / (M * N);
-
-  // lapack clobbers the input, so we have to make a copy.
-  array in(a.shape(), float32, nullptr, {});
-  copy(a, in, a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-
-  // Allocate outputs.
-  u.set_data(allocator::malloc_or_wait(u.nbytes()));
-  s.set_data(allocator::malloc_or_wait(s.nbytes()));
-  vt.set_data(allocator::malloc_or_wait(vt.nbytes()));
-
-  static constexpr auto job_u = "V";
-  static constexpr auto job_vt = "V";
-  static constexpr auto range = "A";
-
-  // Will contain the number of singular values after the call has returned.
-  int ns = 0;
-  float workspace_dimension = 0;
-
-  // Will contain the indices of eigenvectors that failed to converge (not used
-  // here but required by lapack).
-  auto iwork = array::Data{allocator::malloc_or_wait(sizeof(int) * 12 * K)};
-
-  static const int lwork_query = -1;
-
-  static const int ignored_int = 0;
-  static const float ignored_float = 0;
-
-  int info;
-
-  // Compute workspace size.
-  MLX_LAPACK_FUNC(sgesvdx)
-  (
-      /* jobu = */ job_u,
-      /* jobvt = */ job_vt,
-      /* range = */ range,
-      // M and N are swapped since lapack expects column-major.
-      /* m = */ &N,
-      /* n = */ &M,
-      /* a = */ nullptr,
-      /* lda = */ &lda,
-      /* vl = */ &ignored_float,
-      /* vu = */ &ignored_float,
-      /* il = */ &ignored_int,
-      /* iu = */ &ignored_int,
-      /* ns = */ &ns,
-      /* s = */ nullptr,
-      /* u = */ nullptr,
-      /* ldu = */ &ldu,
-      /* vt = */ nullptr,
-      /* ldvt = */ &ldvt,
-      /* work = */ &workspace_dimension,
-      /* lwork = */ &lwork_query,
-      /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-      /* info = */ &info);
-
-  if (info != 0) {
-    std::stringstream ss;
-    ss << "svd_impl: sgesvdx_ workspace calculation failed with code " << info;
-    throw std::runtime_error(ss.str());
-  }
-
-  const int lwork = workspace_dimension;
-  auto scratch = array::Data{allocator::malloc_or_wait(sizeof(float) * lwork)};
-
-  // Loop over matrices.
-  for (int i = 0; i < num_matrices; i++) {
-    MLX_LAPACK_FUNC(sgesvdx)
-    (
-        /* jobu = */ job_u,
-        /* jobvt = */ job_vt,
-        /* range = */ range,
-        // M and N are swapped since lapack expects column-major.
-        /* m = */ &N,
-        /* n = */ &M,
-        /* a = */ in.data<float>() + M * N * i,
-        /* lda = */ &lda,
-        /* vl = */ &ignored_float,
-        /* vu = */ &ignored_float,
-        /* il = */ &ignored_int,
-        /* iu = */ &ignored_int,
-        /* ns = */ &ns,
-        /* s = */ s.data<float>() + K * i,
-        // According to the identity above, lapack will write Vᵀᵀ as U.
-        /* u = */ vt.data<float>() + N * N * i,
-        /* ldu = */ &ldu,
-        // According to the identity above, lapack will write Uᵀ as Vᵀ.
-        /* vt = */ u.data<float>() + M * M * i,
-        /* ldvt = */ &ldvt,
-        /* work = */ static_cast<float*>(scratch.buffer.raw_ptr()),
-        /* lwork = */ &lwork,
-        /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-        /* info = */ &info);
-
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "svd_impl: sgesvdx_ failed with code " << info;
-      throw std::runtime_error(ss.str());
-    }
-
-    if (ns != K) {
-      std::stringstream ss;
-      ss << "svd_impl: expected " << K << " singular values, but " << ns
-         << " were computed.";
-      throw std::runtime_error(ss.str());
-    }
-  }
-}
-
-void SVD::eval(const std::vector<array>& inputs, std::vector<array>& outputs) {
-  if (!(inputs[0].dtype() == float32)) {
-    throw std::runtime_error("[SVD::eval] only supports float32.");
-  }
-  svd_impl(inputs[0], outputs[0], outputs[1], outputs[2]);
-}
-
-} // namespace mlx::core

mlx/backend/common/ternary.h

@@ -3,11 +3,9 @@
 #pragma once
 #include "mlx/allocator.h"
 #include "mlx/array.h"
-#include "mlx/backend/common/ops.h"
 #include "mlx/backend/common/utils.h"
-namespace mlx::core {
 
-namespace {
+namespace mlx::core {
 
 // TODO: Add support for more combinations of input types.
 enum class TernaryOpType {
@@ -16,7 +14,7 @@ enum class TernaryOpType {
   General,
 };
 
-TernaryOpType
+inline TernaryOpType
 get_ternary_op_type(const array& a, const array& b, const array& c) {
   TernaryOpType topt;
   if (a.data_size() == 1 && b.data_size() == 1 && c.data_size() == 1) {
@@ -33,20 +31,15 @@ get_ternary_op_type(const array& a, const array& b, const array& c) {
   return topt;
 }
 
-void set_ternary_op_output_data(
+inline void set_ternary_op_output_data(
     const array& a,
     const array& b,
     const array& c,
     array& out,
-    TernaryOpType topt,
-    bool donate_with_move = false) {
-  auto maybe_donate = [&out, donate_with_move](const array& x) {
+    TernaryOpType topt) {
+  auto maybe_donate = [&out](const array& x) {
     if (is_donatable(x, out)) {
-      if (donate_with_move) {
-        out.move_shared_buffer(x);
-      } else {
-        out.copy_shared_buffer(x);
-      }
+      out.copy_shared_buffer(x);
       return true;
     }
     return false;
@@ -76,152 +69,5 @@ void set_ternary_op_output_data(
       break;
   }
 }
-template <typename T1, typename T2, typename T3, typename U, typename Op, int D>
-void ternary_op_dims(
-    const T1* a,
-    const T2* b,
-    const T3* c,
-    U* out,
-    Op op,
-    const Shape& shape,
-    const Strides& a_strides,
-    const Strides& b_strides,
-    const Strides& c_strides,
-    const Strides& out_strides,
-    int axis) {
-  auto stride_a = a_strides[axis];
-  auto stride_b = b_strides[axis];
-  auto stride_c = c_strides[axis];
-  auto stride_out = out_strides[axis];
-  auto N = shape[axis];
-
-  for (int i = 0; i < N; i++) {
-    if constexpr (D > 1) {
-      ternary_op_dims<T1, T2, T3, U, Op, D - 1>(
-          a,
-          b,
-          c,
-          out,
-          op,
-          shape,
-          a_strides,
-          b_strides,
-          c_strides,
-          out_strides,
-          axis + 1);
-    } else {
-      *out = op(*a, *b, *c);
-    }
-    a += stride_a;
-    b += stride_b;
-    c += stride_c;
-    out += stride_out;
-  }
-}
-
-template <typename T1, typename T2, typename T3, typename U, typename Op>
-void ternary_op_dispatch_dims(
-    const array& a,
-    const array& b,
-    const array& c,
-    array& out,
-    Op op) {
-  auto [shape, strides] = collapse_contiguous_dims(
-      a.shape(), {a.strides(), b.strides(), c.strides(), out.strides()});
-  const auto& a_strides = strides[0];
-  const auto& b_strides = strides[1];
-  const auto& c_strides = strides[2];
-  const auto& out_strides = strides[3];
-
-  const T1* a_ptr = a.data<T1>();
-  const T2* b_ptr = b.data<T2>();
-  const T3* c_ptr = c.data<T3>();
-  U* out_ptr = out.data<T3>();
-  int ndim = shape.size();
-  switch (ndim) {
-    case 1:
-      ternary_op_dims<T1, T2, T3, U, Op, 1>(
-          a_ptr,
-          b_ptr,
-          c_ptr,
-          out_ptr,
-          op,
-          shape,
-          a_strides,
-          b_strides,
-          c_strides,
-          out_strides,
-          0);
-      return;
-    case 2:
-      ternary_op_dims<T1, T2, T3, U, Op, 2>(
-          a_ptr,
-          b_ptr,
-          c_ptr,
-          out_ptr,
-          op,
-          shape,
-          a_strides,
-          b_strides,
-          c_strides,
-          out_strides,
-          0);
-      return;
-  }
-
-  ContiguousIterator a_it(shape, a_strides, ndim - 2);
-  ContiguousIterator b_it(shape, b_strides, ndim - 2);
-  ContiguousIterator c_it(shape, c_strides, ndim - 2);
-  auto stride = out_strides[ndim - 3];
-  for (size_t elem = 0; elem < a.size(); elem += stride) {
-    ternary_op_dims<T1, T2, T3, U, Op, 2>(
-        a_ptr + a_it.loc,
-        b_ptr + b_it.loc,
-        c_ptr + c_it.loc,
-        out_ptr + elem,
-        op,
-        shape,
-        a_strides,
-        b_strides,
-        c_strides,
-        out_strides,
-        ndim - 2);
-    a_it.step();
-    b_it.step();
-    c_it.step();
-  }
-}
-
-template <typename T1, typename T2, typename T3, typename U, typename Op>
-void ternary_op(
-    const array& a,
-    const array& b,
-    const array& c,
-    array& out,
-    Op op) {
-  TernaryOpType topt = get_ternary_op_type(a, b, c);
-  set_ternary_op_output_data(a, b, c, out, topt);
-
-  // The full computation is scalar-scalar-scalar so we call the base op once.
-  if (topt == TernaryOpType::ScalarScalarScalar) {
-    *(out.data<U>()) = op(*a.data<T1>(), *b.data<T2>(), *c.data<T3>());
-  } else if (topt == TernaryOpType::VectorVectorVector) {
-    const T1* a_ptr = a.data<T1>();
-    const T2* b_ptr = b.data<T2>();
-    const T3* c_ptr = c.data<T3>();
-    U* out_ptr = out.data<U>();
-    for (size_t i = 0; i < out.size(); ++i) {
-      *out_ptr = op(*a_ptr, *b_ptr, *c_ptr);
-      a_ptr++;
-      b_ptr++;
-      c_ptr++;
-      out_ptr++;
-    }
-  } else {
-    ternary_op_dispatch_dims<T1, T2, T3, U>(a, b, c, out, op);
-  }
-}
-
-} // namespace
 
 } // namespace mlx::core

mlx/backend/common/unary.h

@@ -1,130 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#pragma once
-
-#include "mlx/allocator.h"
-#include "mlx/array.h"
-#include "mlx/backend/common/utils.h"
-#include "mlx/utils.h"
-
-namespace mlx::core {
-
-namespace {
-
-void set_unary_output_data(const array& in, array& out) {
-  if (is_donatable(in, out)) {
-    out.copy_shared_buffer(in);
-  } else {
-    auto size = in.data_size();
-    out.set_data(
-        allocator::malloc_or_wait(size * out.itemsize()),
-        size,
-        in.strides(),
-        in.flags());
-  }
-}
-
-template <typename T, typename U = T, typename Op>
-void unary_op(const T* a, U* out, Op op, size_t shape, size_t stride) {
-  for (size_t i = 0; i < shape; i += 1) {
-    out[i] = op(*a);
-    a += stride;
-  }
-}
-
-template <typename T, typename U = T, typename Op>
-void unary_op(const array& a, array& out, Op op) {
-  const T* a_ptr = a.data<T>();
-  if (a.flags().contiguous) {
-    set_unary_output_data(a, out);
-    U* dst = out.data<U>();
-    for (size_t i = 0; i < a.data_size(); ++i) {
-      dst[i] = op(a_ptr[i]);
-    }
-  } else {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    U* dst = out.data<U>();
-    size_t shape = a.ndim() > 0 ? a.shape(-1) : 1;
-    size_t stride = a.ndim() > 0 ? a.strides(-1) : 1;
-    if (a.ndim() <= 1) {
-      unary_op(a_ptr, dst, op, shape, stride);
-      return;
-    }
-    ContiguousIterator it(a.shape(), a.strides(), a.ndim() - 1);
-    for (size_t elem = 0; elem < a.size(); elem += shape) {
-      unary_op(a_ptr + it.loc, dst + elem, op, shape, stride);
-      it.step();
-    }
-  }
-}
-
-template <typename Op>
-void unary(const array& a, array& out, Op op) {
-  switch (out.dtype()) {
-    case bool_:
-      unary_op<bool>(a, out, op);
-      break;
-    case uint8:
-      unary_op<uint8_t>(a, out, op);
-      break;
-    case uint16:
-      unary_op<uint16_t>(a, out, op);
-      break;
-    case uint32:
-      unary_op<uint32_t>(a, out, op);
-      break;
-    case uint64:
-      unary_op<uint64_t>(a, out, op);
-      break;
-    case int8:
-      unary_op<int8_t>(a, out, op);
-      break;
-    case int16:
-      unary_op<int16_t>(a, out, op);
-      break;
-    case int32:
-      unary_op<int32_t>(a, out, op);
-      break;
-    case int64:
-      unary_op<int64_t>(a, out, op);
-      break;
-    case float16:
-      unary_op<float16_t>(a, out, op);
-      break;
-    case float32:
-      unary_op<float>(a, out, op);
-      break;
-    case bfloat16:
-      unary_op<bfloat16_t>(a, out, op);
-      break;
-    case complex64:
-      unary_op<complex64_t>(a, out, op);
-      break;
-  }
-}
-
-template <typename Op>
-void unary_fp(const array& a, array& out, Op op) {
-  switch (out.dtype()) {
-    case bfloat16:
-      unary_op<bfloat16_t>(a, out, op);
-      break;
-    case float16:
-      unary_op<float16_t>(a, out, op);
-      break;
-    case float32:
-      unary_op<float>(a, out, op);
-      break;
-    case complex64:
-      unary_op<complex64_t>(a, out, op);
-      break;
-    default:
-      std::ostringstream err;
-      err << "[unary_fp] Does not support " << out.dtype();
-      throw std::runtime_error(err.str());
-  }
-}
-
-} // namespace
-
-} // namespace mlx::core

mlx/backend/common/utils.cpp

@@ -4,28 +4,6 @@
 
 namespace mlx::core {
 
-void move_or_copy(const array& in, array& out) {
-  if (in.is_donatable()) {
-    out.move_shared_buffer(in);
-  } else {
-    out.copy_shared_buffer(in);
-  }
-}
-
-void move_or_copy(
-    const array& in,
-    array& out,
-    const Strides& strides,
-    array::Flags flags,
-    size_t data_size,
-    size_t offset /* = 0 */) {
-  if (in.is_donatable()) {
-    out.move_shared_buffer(in, strides, flags, data_size, offset);
-  } else {
-    out.copy_shared_buffer(in, strides, flags, data_size, offset);
-  }
-}
-
 std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
     const Shape& shape,
     const std::vector<Strides>& strides,

mlx/backend/common/utils.h

@@ -159,15 +159,6 @@ inline bool is_donatable(const array& in, const array& out) {
       in.buffer_size() <= out.nbytes() + donation_extra;
 }
 
-void move_or_copy(const array& in, array& out);
-void move_or_copy(
-    const array& in,
-    array& out,
-    const Strides& strides,
-    array::Flags flags,
-    size_t data_size,
-    size_t offset = 0);
-
 std::pair<bool, Strides> prepare_reshape(const array& in, const array& out);
 
 void shared_buffer_reshape(

mlx/backend/cpu/CMakeLists.txt

@@ -0,0 +1,85 @@
+if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+  set(COMPILER ${CMAKE_C_COMPILER})
+  set(CLANG TRUE)
+else()
+  set(COMPILER ${CMAKE_CXX_COMPILER})
+endif()
+
+set(COMPILE_DEPS
+    ${PROJECT_SOURCE_DIR}/mlx/types/half_types.h
+    ${PROJECT_SOURCE_DIR}/mlx/types/fp16.h
+    ${PROJECT_SOURCE_DIR}/mlx/types/bf16.h
+    ${PROJECT_SOURCE_DIR}/mlx/types/complex.h
+    simd/simd.h
+    simd/base_simd.h
+    simd/math.h
+    simd/type.h
+    unary_ops.h
+    binary_ops.h)
+
+if(MSVC)
+  set(SHELL_EXT ps1)
+  set(SHELL_CMD powershell -ExecutionPolicy Bypass -File)
+else()
+  set(SHELL_EXT sh)
+  set(SHELL_CMD bash)
+endif()
+
+add_custom_command(
+  OUTPUT compiled_preamble.cpp
+  COMMAND
+    ${SHELL_CMD} ${CMAKE_CURRENT_SOURCE_DIR}/make_compiled_preamble.${SHELL_EXT}
+    ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp ${COMPILER}
+    ${PROJECT_SOURCE_DIR} ${CLANG} ${CMAKE_SYSTEM_PROCESSOR}
+  DEPENDS make_compiled_preamble.${SHELL_EXT} compiled_preamble.h
+          ${COMPILE_DEPS})
+
+add_custom_target(cpu_compiled_preamble DEPENDS compiled_preamble.cpp)
+
+add_dependencies(mlx cpu_compiled_preamble)
+
+target_sources(
+  mlx
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/distributed.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/eigh.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/encoder.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/hadamard.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cblas.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/masked_mm.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/sort.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/threefry.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/luf.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/qrf.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/svd.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/inverse.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/cholesky.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/unary.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
+          ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp)
+
+if(MLX_BUILD_ACCELERATE)
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/bnns.cpp)
+else()
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/no_fp16.cpp
+                             ${CMAKE_CURRENT_SOURCE_DIR}/gemms/no_bf16.cpp)
+endif()
+
+if(IOS)
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../no_cpu/compiled.cpp)
+else()
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
+                             ${CMAKE_CURRENT_SOURCE_DIR}/jit_compiler.cpp)
+endif()

mlx/backend/cpu/arange.h

@@ -0,0 +1,28 @@
+// Copyright © 2023 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/backend/cpu/encoder.h"
+
+namespace mlx::core {
+
+namespace {
+
+template <typename T>
+void arange(T start, T next, array& out, size_t size, Stream stream) {
+  auto ptr = out.data<T>();
+  auto step_size = next - start;
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(out);
+  encoder.dispatch([ptr, start, step_size, size]() mutable {
+    for (int i = 0; i < size; ++i) {
+      ptr[i] = start;
+      start += step_size;
+    }
+  });
+}
+
+} // namespace
+
+} // namespace mlx::core

mlx/backend/cpu/arg_reduce.cpp

@@ -0,0 +1,126 @@
+// Copyright © 2023 Apple Inc.
+
+#include <cassert>
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+namespace {
+
+template <typename InT, typename OpT>
+void arg_reduce(const array& in, array& out, const OpT& op, int axis) {
+  auto axis_size = in.shape()[axis];
+  auto axis_stride = in.strides()[axis];
+  Strides strides = in.strides();
+  Shape shape = in.shape();
+  strides.erase(strides.begin() + axis);
+  shape.erase(shape.begin() + axis);
+  auto in_ptr = in.data<InT>();
+  auto out_ptr = out.data<uint32_t>();
+
+  for (uint32_t i = 0; i < out.size(); ++i) {
+    auto loc = elem_to_loc(i, shape, strides);
+    auto local_in_ptr = in_ptr + loc;
+    uint32_t ind_v = 0;
+    InT v = (*local_in_ptr);
+    for (uint32_t j = 0; j < axis_size; ++j, local_in_ptr += axis_stride) {
+      op(j, (*local_in_ptr), &ind_v, &v);
+    }
+    out_ptr[i] = ind_v;
+  }
+}
+
+template <typename InT>
+void arg_reduce_dispatch(
+    const array& in,
+    array& out,
+    ArgReduce::ReduceType rtype,
+    int axis) {
+  switch (rtype) {
+    case ArgReduce::ArgMin: {
+      auto op = [](auto ind_x, auto x, auto ind_y, auto y) {
+        if (x < (*y)) {
+          (*y) = x;
+          (*ind_y) = ind_x;
+        }
+      };
+      arg_reduce<InT>(in, out, op, axis);
+      break;
+    }
+    case ArgReduce::ArgMax: {
+      auto op = [](auto ind_x, auto x, auto ind_y, auto y) {
+        if (x > (*y)) {
+          (*y) = x;
+          (*ind_y) = ind_x;
+        }
+      };
+      arg_reduce<InT>(in, out, op, axis);
+      break;
+    }
+  }
+}
+
+} // namespace
+
+void ArgReduce::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  encoder.dispatch([in = array::unsafe_weak_copy(in),
+                    out = array::unsafe_weak_copy(out),
+                    reduce_type_ = reduce_type_,
+                    axis_ = axis_]() mutable {
+    switch (in.dtype()) {
+      case bool_:
+        arg_reduce_dispatch<bool>(in, out, reduce_type_, axis_);
+        break;
+      case uint8:
+        arg_reduce_dispatch<uint8_t>(in, out, reduce_type_, axis_);
+        break;
+      case uint16:
+        arg_reduce_dispatch<uint16_t>(in, out, reduce_type_, axis_);
+        break;
+      case uint32:
+        arg_reduce_dispatch<uint32_t>(in, out, reduce_type_, axis_);
+        break;
+      case uint64:
+        arg_reduce_dispatch<uint64_t>(in, out, reduce_type_, axis_);
+        break;
+      case int8:
+        arg_reduce_dispatch<int8_t>(in, out, reduce_type_, axis_);
+        break;
+      case int16:
+        arg_reduce_dispatch<int16_t>(in, out, reduce_type_, axis_);
+        break;
+      case int32:
+        arg_reduce_dispatch<int32_t>(in, out, reduce_type_, axis_);
+        break;
+      case int64:
+        arg_reduce_dispatch<int64_t>(in, out, reduce_type_, axis_);
+        break;
+      case float16:
+        arg_reduce_dispatch<float16_t>(in, out, reduce_type_, axis_);
+        break;
+      case float32:
+        arg_reduce_dispatch<float>(in, out, reduce_type_, axis_);
+        break;
+      case bfloat16:
+        arg_reduce_dispatch<bfloat16_t>(in, out, reduce_type_, axis_);
+        break;
+      case float64:
+        arg_reduce_dispatch<double>(in, out, reduce_type_, axis_);
+        break;
+      case complex64:
+        arg_reduce_dispatch<complex64_t>(in, out, reduce_type_, axis_);
+        break;
+    }
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cpu/binary.cpp

@@ -0,0 +1,487 @@
+// Copyright © 2023 Apple Inc.
+
+#include <cassert>
+#include <cmath>
+#include <sstream>
+
+#include "mlx/allocator.h"
+#include "mlx/backend/cpu/binary.h"
+#include "mlx/backend/cpu/binary_ops.h"
+#include "mlx/backend/cpu/binary_two.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/primitives.h"
+#include "mlx/utils.h"
+
+namespace mlx::core {
+
+namespace {
+
+template <typename Op>
+void binary(const array& a, const array& b, array& out, Op op, Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        binary_op<bool, Op>(a, b, out, bopt);
+        break;
+      case uint8:
+        binary_op<uint8_t, Op>(a, b, out, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t, Op>(a, b, out, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t, Op>(a, b, out, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t, Op>(a, b, out, bopt);
+        break;
+      case int8:
+        binary_op<int8_t, Op>(a, b, out, bopt);
+        break;
+      case int16:
+        binary_op<int16_t, Op>(a, b, out, bopt);
+        break;
+      case int32:
+        binary_op<int32_t, Op>(a, b, out, bopt);
+        break;
+      case int64:
+        binary_op<int64_t, Op>(a, b, out, bopt);
+        break;
+      case float16:
+        binary_op<float16_t, Op>(a, b, out, bopt);
+        break;
+      case float32:
+        binary_op<float, Op>(a, b, out, bopt);
+        break;
+      case float64:
+        binary_op<double, Op>(a, b, out, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t, Op>(a, b, out, bopt);
+        break;
+      case complex64:
+        binary_op<complex64_t, Op>(a, b, out, bopt);
+        break;
+    }
+  });
+}
+
+template <typename Op>
+void comparison_op(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (a.dtype()) {
+      case bool_:
+        binary_op<bool, bool, Op>(a, b, out, bopt);
+        break;
+      case uint8:
+        binary_op<uint8_t, bool, Op>(a, b, out, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t, bool, Op>(a, b, out, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int8:
+        binary_op<int8_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int16:
+        binary_op<int16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int32:
+        binary_op<int32_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int64:
+        binary_op<int64_t, bool, Op>(a, b, out, bopt);
+        break;
+      case float16:
+        binary_op<float16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case float32:
+        binary_op<float, bool, Op>(a, b, out, bopt);
+        break;
+      case float64:
+        binary_op<double, bool, Op>(a, b, out, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case complex64:
+        binary_op<complex64_t, bool, Op>(a, b, out, bopt);
+        break;
+    }
+  });
+}
+
+template <typename Op>
+void binary_float(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (out.dtype()) {
+      case float16:
+        binary_op<float16_t, Op>(a, b, out, bopt);
+        break;
+      case float32:
+        binary_op<float, Op>(a, b, out, bopt);
+        break;
+      case float64:
+        binary_op<double, Op>(a, b, out, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t, Op>(a, b, out, bopt);
+        break;
+      default:
+        throw std::runtime_error(
+            "[binary_float] Only supports non-complex floating point types.");
+    }
+  });
+}
+
+template <typename Op>
+void binary_int(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        binary_op<bool, Op>(a, b, out, bopt);
+      case uint8:
+        binary_op<uint8_t, Op>(a, b, out, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t, Op>(a, b, out, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t, Op>(a, b, out, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t, Op>(a, b, out, bopt);
+        break;
+      case int8:
+        binary_op<int8_t, Op>(a, b, out, bopt);
+        break;
+      case int16:
+        binary_op<int16_t, Op>(a, b, out, bopt);
+        break;
+      case int32:
+        binary_op<int32_t, Op>(a, b, out, bopt);
+        break;
+      case int64:
+        binary_op<int64_t, Op>(a, b, out, bopt);
+        break;
+      default:
+        throw std::runtime_error("[binary_int] Type not supported");
+        break;
+    }
+  });
+}
+
+} // namespace
+
+void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Add(), stream());
+}
+
+void DivMod::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  auto& out_a = outputs[0];
+  auto& out_b = outputs[1];
+  set_binary_op_output_data(a, b, out_a, bopt);
+  set_binary_op_output_data(a, b, out_b, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out_a);
+  encoder.set_output_array(out_b);
+
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out_a = array::unsafe_weak_copy(out_a),
+                    out_b = array::unsafe_weak_copy(out_b),
+                    bopt]() mutable {
+    auto integral_op = [](auto x, auto y) {
+      return std::make_pair(x / y, x % y);
+    };
+    auto float_op = [](auto x, auto y) {
+      return std::make_pair(std::trunc(x / y), std::fmod(x, y));
+    };
+
+    switch (out_a.dtype()) {
+      case bool_:
+        binary_op<bool>(a, b, out_a, out_b, integral_op, bopt);
+      case uint8:
+        binary_op<uint8_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case int8:
+        binary_op<int8_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case int16:
+        binary_op<int16_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case int32:
+        binary_op<int32_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case int64:
+        binary_op<int64_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case float16:
+        binary_op<float16_t>(a, b, out_a, out_b, float_op, bopt);
+        break;
+      case float32:
+        binary_op<float>(a, b, out_a, out_b, float_op, bopt);
+        break;
+      case float64:
+        binary_op<double>(a, b, out_a, out_b, float_op, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t>(a, b, out_a, out_b, float_op, bopt);
+        break;
+      case complex64:
+        // Should never get here
+        throw std::runtime_error("[DivMod] Complex type not supported");
+        break;
+    }
+  });
+}
+
+void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Divide(), stream());
+}
+
+void Remainder::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Remainder(), stream());
+}
+
+void Equal::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  if (equal_nan_) {
+    auto bopt = get_binary_op_type(a, b);
+    set_binary_op_output_data(a, b, out, bopt);
+
+    auto& encoder = cpu::get_command_encoder(stream());
+    encoder.set_input_array(a);
+    encoder.set_input_array(b);
+    encoder.set_output_array(out);
+    encoder.dispatch([a = array::unsafe_weak_copy(a),
+                      b = array::unsafe_weak_copy(b),
+                      out = array::unsafe_weak_copy(out),
+                      bopt]() mutable {
+      switch (a.dtype()) {
+        case float16:
+          binary_op<float16_t, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        case float32:
+          binary_op<float, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        case float64:
+          binary_op<double, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        case bfloat16:
+          binary_op<bfloat16_t, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        case complex64:
+          binary_op<complex64_t, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        default:
+          throw std::runtime_error(
+              "[NanEqual::eval_cpu] Only for floating point types.");
+      }
+    });
+  } else {
+    comparison_op(a, b, out, detail::Equal(), stream());
+  }
+}
+
+void Greater::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::Greater(), stream());
+}
+
+void GreaterEqual::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::GreaterEqual(), stream());
+}
+
+void Less::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::Less(), stream());
+}
+
+void LessEqual::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::LessEqual(), stream());
+}
+
+void LogAddExp::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary_float(a, b, out, detail::LogAddExp(), stream());
+}
+
+void LogicalAnd::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalAnd requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalAnd(), stream());
+}
+
+void LogicalOr::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalOr requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalOr(), stream());
+}
+
+void Maximum::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Maximum(), stream());
+}
+
+void Minimum::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Minimum(), stream());
+}
+
+void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Multiply(), stream());
+}
+
+void NotEqual::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::NotEqual(), stream());
+}
+
+void Power::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Power(), stream());
+}
+
+void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Subtract(), stream());
+}
+
+void BitwiseBinary::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  switch (op_) {
+    case BitwiseBinary::And:
+      binary_int(a, b, out, detail::BitwiseAnd(), stream());
+      break;
+    case BitwiseBinary::Or:
+      binary_int(a, b, out, detail::BitwiseOr(), stream());
+      break;
+    case BitwiseBinary::Xor:
+      binary_int(a, b, out, detail::BitwiseXor(), stream());
+      break;
+    case BitwiseBinary::LeftShift:
+      binary_int(a, b, out, detail::LeftShift(), stream());
+      break;
+    case BitwiseBinary::RightShift:
+      binary_int(a, b, out, detail::RightShift(), stream());
+      break;
+  }
+}
+
+void ArcTan2::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  binary_float(a, b, out, detail::ArcTan2(), stream());
+}
+
+} // namespace mlx::core

mlx/backend/cpu/binary.h

@@ -0,0 +1,293 @@
+// Copyright © 2023 Apple Inc.
+
+#pragma once
+#include <cassert>
+
+#include "mlx/array.h"
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/common/utils.h"
+
+#include "mlx/backend/cpu/simd/simd.h"
+
+namespace mlx::core {
+
+template <typename Op>
+struct VectorScalar {
+  template <typename T, typename U>
+  void operator()(const T* a, const T* b, U* dst, int size) {
+    T scalar = *b;
+    constexpr int N = simd::max_size<T>;
+    while (size >= N) {
+      simd::store(dst, Op{}(simd::load<T, N>(a), simd::Simd<T, N>(scalar)));
+      dst += N;
+      a += N;
+      size -= N;
+    }
+    while (size-- > 0) {
+      *dst = Op{}(*a, scalar);
+      dst++;
+      a++;
+    }
+  }
+};
+
+template <typename Op>
+struct ScalarVector {
+  template <typename T, typename U>
+  void operator()(const T* a, const T* b, U* dst, int size) {
+    T scalar = *a;
+    constexpr int N = simd::max_size<T>;
+    while (size >= N) {
+      simd::store(dst, Op{}(simd::Simd<T, N>(scalar), simd::load<T, N>(b)));
+      dst += N;
+      b += N;
+      size -= N;
+    }
+    while (size-- > 0) {
+      *dst = Op{}(scalar, *b);
+      dst++;
+      b++;
+    }
+  }
+};
+
+template <typename Op>
+struct VectorVector {
+  template <typename T, typename U>
+  void operator()(const T* a, const T* b, U* dst, int size) {
+    constexpr int N = simd::max_size<T>;
+    while (size >= N) {
+      simd::store(dst, Op{}(simd::load<T, N>(a), simd::load<T, N>(b)));
+      dst += N;
+      a += N;
+      b += N;
+      size -= N;
+    }
+    while (size-- > 0) {
+      *dst = Op{}(*a, *b);
+      dst++;
+      a++;
+      b++;
+    }
+  }
+};
+
+template <typename T, typename U, typename Op, int D, bool Strided>
+void binary_op_dims(
+    const T* a,
+    const T* b,
+    U* out,
+    const Shape& shape,
+    const Strides& a_strides,
+    const Strides& b_strides,
+    const Strides& out_strides,
+    int axis) {
+  auto stride_a = a_strides[axis];
+  auto stride_b = b_strides[axis];
+  auto stride_out = out_strides[axis];
+  auto N = shape[axis];
+
+  for (int i = 0; i < N; i++) {
+    if constexpr (D > 1) {
+      binary_op_dims<T, U, Op, D - 1, Strided>(
+          a, b, out, shape, a_strides, b_strides, out_strides, axis + 1);
+    } else {
+      if constexpr (Strided) {
+        Op{}(a, b, out, stride_out);
+      } else {
+        *out = Op{}(*a, *b);
+      }
+    }
+    out += stride_out;
+    a += stride_a;
+    b += stride_b;
+  }
+}
+
+template <typename T, typename U, bool Strided, typename Op>
+void binary_op_dispatch_dims(
+    const T* a,
+    const T* b,
+    U* out,
+    int dim,
+    int size,
+    const Shape& shape,
+    const Strides& a_strides,
+    const Strides& b_strides,
+    const Strides& out_strides) {
+  switch (dim) {
+    case 1:
+      binary_op_dims<T, U, Op, 1, Strided>(
+          a, b, out, shape, a_strides, b_strides, out_strides, 0);
+      return;
+    case 2:
+      binary_op_dims<T, U, Op, 2, Strided>(
+          a, b, out, shape, a_strides, b_strides, out_strides, 0);
+      return;
+    case 3:
+      binary_op_dims<T, U, Op, 3, Strided>(
+          a, b, out, shape, a_strides, b_strides, out_strides, 0);
+      return;
+  }
+
+  ContiguousIterator a_it(shape, a_strides, dim - 3);
+  ContiguousIterator b_it(shape, b_strides, dim - 3);
+  auto stride = out_strides[dim - 4];
+  for (int64_t elem = 0; elem < size; elem += stride) {
+    binary_op_dims<T, U, Op, 3, Strided>(
+        a + a_it.loc,
+        b + b_it.loc,
+        out + elem,
+        shape,
+        a_strides,
+        b_strides,
+        out_strides,
+        dim - 3);
+    a_it.step();
+    b_it.step();
+  }
+}
+
+template <typename T, typename U, typename Op>
+void binary_op(const array& a, const array& b, array& out, BinaryOpType bopt) {
+  // The full computation is scalar scalar so call the base op once
+  auto a_ptr = a.data<T>();
+  auto b_ptr = b.data<T>();
+
+  auto out_ptr = out.data<U>();
+  if (bopt == BinaryOpType::ScalarScalar) {
+    *out_ptr = Op{}(*a_ptr, *b_ptr);
+    return;
+  }
+
+  // The full computation is scalar vector so delegate to the op
+  if (bopt == BinaryOpType::ScalarVector) {
+    ScalarVector<Op>{}(a_ptr, b_ptr, out_ptr, b.data_size());
+    return;
+  }
+
+  // The full computation is vector scalar so delegate to the op
+  if (bopt == BinaryOpType::VectorScalar) {
+    VectorScalar<Op>{}(a_ptr, b_ptr, out_ptr, a.data_size());
+    return;
+  }
+
+  // The full computation is vector vector so delegate to the op
+  if (bopt == BinaryOpType::VectorVector) {
+    VectorVector<Op>{}(a_ptr, b_ptr, out_ptr, a.size());
+    return;
+  }
+
+  // General computation so let's try to optimize
+  auto [new_shape, new_strides] = collapse_contiguous_dims(
+      a.shape(), {a.strides(), b.strides(), out.strides()});
+  auto& a_strides = new_strides[0];
+  auto& b_strides = new_strides[1];
+  auto& strides = new_strides[2];
+
+  // Get the left-most dim such that the array is row contiguous after
+  auto leftmost_rc_dim = [&strides](const auto& arr_strides) {
+    int d = arr_strides.size() - 1;
+    for (; d >= 0 && arr_strides[d] == strides[d]; d--) {
+    }
+    return d + 1;
+  };
+  auto a_rc_dim = leftmost_rc_dim(a_strides);
+  auto b_rc_dim = leftmost_rc_dim(b_strides);
+
+  // Get the left-most dim such that the array is a broadcasted "scalar" after
+  auto leftmost_s_dim = [](const auto& arr_strides) {
+    int d = arr_strides.size() - 1;
+    for (; d >= 0 && arr_strides[d] == 0; d--) {
+    }
+    return d + 1;
+  };
+  auto a_s_dim = leftmost_s_dim(a_strides);
+  auto b_s_dim = leftmost_s_dim(b_strides);
+
+  auto ndim = new_shape.size();
+
+  // Case 1: LxM and FxM where L and F are broadcastable and M is row
+  // contiguous
+  int dim = ndim;
+  if (int d = std::max(a_rc_dim, b_rc_dim); d < ndim) {
+    bopt = BinaryOpType::VectorVector;
+    dim = d;
+    // Case 2: LxM and Fx1 where L and F are broadcastable and M is row
+    // contiguous
+  } else if (int d = std::max(a_rc_dim, b_s_dim); d < ndim) {
+    bopt = BinaryOpType::VectorScalar;
+    dim = d;
+    // Case 3: Lx1 and FxM where L and F are broadcastable and M is row
+    // contiguous
+  } else if (int d = std::max(a_s_dim, b_rc_dim); d < ndim) {
+    bopt = BinaryOpType::ScalarVector;
+    dim = d;
+  }
+
+  // Can be sure dim > 0 since otherwise we would have used one of the fully
+  // contiguous methods above. Except for the case that the flags do not
+  // correspond to the underlying contiguity.
+  if (dim == 0 || strides[dim - 1] < 16) {
+    bopt = BinaryOpType::General;
+    dim = ndim;
+  }
+
+  switch (bopt) {
+    case BinaryOpType::VectorVector:
+      binary_op_dispatch_dims<T, U, true, VectorVector<Op>>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
+          dim,
+          a.size(),
+          new_shape,
+          a_strides,
+          b_strides,
+          strides);
+      break;
+    case BinaryOpType::VectorScalar:
+      binary_op_dispatch_dims<T, U, true, VectorScalar<Op>>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
+          dim,
+          a.size(),
+          new_shape,
+          a_strides,
+          b_strides,
+          strides);
+      break;
+    case BinaryOpType::ScalarVector:
+      binary_op_dispatch_dims<T, U, true, ScalarVector<Op>>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
+          dim,
+          a.size(),
+          new_shape,
+          a_strides,
+          b_strides,
+          strides);
+      break;
+    default:
+      binary_op_dispatch_dims<T, U, false, Op>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
+          dim,
+          a.size(),
+          new_shape,
+          a_strides,
+          b_strides,
+          strides);
+      break;
+  }
+}
+
+template <typename T, typename Op>
+void binary_op(const array& a, const array& b, array& out, BinaryOpType bopt) {
+  binary_op<T, T, Op>(a, b, out, bopt);
+}
+
+} // namespace mlx::core

mlx/backend/cpu/binary_ops.h

@@ -0,0 +1,98 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cpu/simd/simd.h"
+
+namespace mlx::core::detail {
+
+using namespace mlx::core::simd;
+
+#define BINARY_SINGLE()                                 \
+  template <typename T>                                 \
+  T operator()(T x, T y) {                              \
+    return (*this)(Simd<T, 1>(x), Simd<T, 1>(y)).value; \
+  }
+
+#define DEFAULT_BINARY_OP(Op, op)                       \
+  struct Op {                                           \
+    template <int N, typename T>                        \
+    Simd<T, N> operator()(Simd<T, N> x, Simd<T, N> y) { \
+      return op(x, y);                                  \
+    }                                                   \
+    BINARY_SINGLE()                                     \
+  };
+
+DEFAULT_BINARY_OP(Add, operator+)
+DEFAULT_BINARY_OP(ArcTan2, atan2)
+DEFAULT_BINARY_OP(Divide, operator/)
+DEFAULT_BINARY_OP(Multiply, operator*)
+DEFAULT_BINARY_OP(Subtract, operator-)
+DEFAULT_BINARY_OP(LogicalAnd, operator&&)
+DEFAULT_BINARY_OP(LogicalOr, operator||)
+DEFAULT_BINARY_OP(BitwiseAnd, operator&)
+DEFAULT_BINARY_OP(BitwiseOr, operator|)
+DEFAULT_BINARY_OP(BitwiseXor, operator^)
+DEFAULT_BINARY_OP(LeftShift, operator<<)
+DEFAULT_BINARY_OP(RightShift, operator>>)
+DEFAULT_BINARY_OP(Remainder, remainder)
+DEFAULT_BINARY_OP(Maximum, maximum)
+DEFAULT_BINARY_OP(Minimum, minimum)
+DEFAULT_BINARY_OP(Power, pow)
+
+#define DEFAULT_BOOL_OP(Op, op)                            \
+  struct Op {                                              \
+    template <int N, typename T>                           \
+    Simd<bool, N> operator()(Simd<T, N> x, Simd<T, N> y) { \
+      return op(x, y);                                     \
+    }                                                      \
+    template <typename T>                                  \
+    bool operator()(T x, T y) {                            \
+      return (*this)(Simd<T, 1>(x), Simd<T, 1>(y)).value;  \
+    }                                                      \
+  };
+
+DEFAULT_BOOL_OP(Equal, operator==)
+DEFAULT_BOOL_OP(Greater, operator>)
+DEFAULT_BOOL_OP(GreaterEqual, operator>=)
+DEFAULT_BOOL_OP(Less, operator<)
+DEFAULT_BOOL_OP(LessEqual, operator<=)
+DEFAULT_BOOL_OP(NotEqual, operator!=)
+
+struct NaNEqual {
+  template <int N, typename T>
+  Simd<bool, N> operator()(Simd<T, N> x, Simd<T, N> y) {
+    return x == y || (isnan(x) && isnan(y));
+  }
+  template <typename T>
+  bool operator()(T x, T y) {
+    return (*this)(Simd<T, 1>(x), Simd<T, 1>(y)).value;
+  }
+};
+
+struct LogAddExp {
+  template <int N, typename T>
+  Simd<T, N> operator()(Simd<T, N> x, Simd<T, N> y) {
+    auto maxval = maximum(x, y);
+    auto minval = minimum(x, y);
+    auto mask = minval == -inf || maxval == inf;
+    auto out = maxval + log1p(exp(minval - maxval));
+    return select(mask, Simd<T, N>(maxval), Simd<T, N>(out));
+  }
+  BINARY_SINGLE()
+};
+
+struct Select {
+  template <typename T>
+  T operator()(bool condition, T x, T y) {
+    return (*this)(Simd<bool, 1>(condition), Simd<T, 1>(x), Simd<T, 1>(y))
+        .value;
+  }
+
+  template <int N, typename T>
+  Simd<T, N> operator()(Simd<bool, N> condition, Simd<T, N> x, Simd<T, N> y) {
+    return select(condition, x, y);
+  }
+};
+
+} // namespace mlx::core::detail

mlx/backend/cpu/binary_two.h

@@ -2,8 +2,8 @@
 
 #pragma once
 
-#include "mlx/backend/common/binary.h"
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/binary.h"
 
 namespace mlx::core {
 
@@ -58,14 +58,14 @@ void binary_op_dispatch_dims(
     Op op) {
   auto [shape, strides] = collapse_contiguous_dims(
       a.shape(), {a.strides(), b.strides(), out_a.strides()});
-  const auto& a_strides = strides[0];
-  const auto& b_strides = strides[1];
-  const auto& out_strides = strides[2];
   const T* a_ptr = a.data<T>();
   const T* b_ptr = b.data<T>();
   U* out_a_ptr = out_a.data<U>();
   U* out_b_ptr = out_b.data<U>();
 
+  const auto& a_strides = strides[0];
+  const auto& b_strides = strides[1];
+  const auto& out_strides = strides[2];
   int ndim = shape.size();
   switch (ndim) {
     case 1:
@@ -120,14 +120,10 @@ template <typename T, typename U = T, typename Op>
 void binary_op(
     const array& a,
     const array& b,
-    std::vector<array>& outputs,
-    Op op) {
-  auto bopt = get_binary_op_type(a, b);
-  auto& out_a = outputs[0];
-  auto& out_b = outputs[1];
-  set_binary_op_output_data(a, b, out_a, bopt);
-  set_binary_op_output_data(a, b, out_b, bopt);
-
+    array& out_a,
+    array& out_b,
+    Op op,
+    BinaryOpType bopt) {
   // The full computation is scalar scalar so call the base op once
   if (bopt == BinaryOpType::General) {
     binary_op_dispatch_dims<T, U, Op>(a, b, out_a, out_b, op);
@@ -141,14 +137,14 @@ void binary_op(
   if (bopt == BinaryOpType::ScalarScalar) {
     std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
   } else if (bopt == BinaryOpType::ScalarVector) {
-    for (size_t i = 0; i < b.size(); ++i) {
+    for (size_t i = 0; i < b.data_size(); ++i) {
       std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
       out_a_ptr++;
       out_b_ptr++;
       b_ptr++;
     }
   } else if (bopt == BinaryOpType::VectorScalar) {
-    for (size_t i = 0; i < a.size(); ++i) {
+    for (size_t i = 0; i < a.data_size(); ++i) {
       std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
       out_a_ptr++;
       out_b_ptr++;
@@ -165,55 +161,6 @@ void binary_op(
   }
 }
 
-template <typename Op>
-void binary(
-    const array& a,
-    const array& b,
-    std::vector<array>& outputs,
-    Op op) {
-  switch (outputs[0].dtype()) {
-    case bool_:
-      binary_op<bool>(a, b, outputs, op);
-      break;
-    case uint8:
-      binary_op<uint8_t>(a, b, outputs, op);
-      break;
-    case uint16:
-      binary_op<uint16_t>(a, b, outputs, op);
-      break;
-    case uint32:
-      binary_op<uint32_t>(a, b, outputs, op);
-      break;
-    case uint64:
-      binary_op<uint64_t>(a, b, outputs, op);
-      break;
-    case int8:
-      binary_op<int8_t>(a, b, outputs, op);
-      break;
-    case int16:
-      binary_op<int16_t>(a, b, outputs, op);
-      break;
-    case int32:
-      binary_op<int32_t>(a, b, outputs, op);
-      break;
-    case int64:
-      binary_op<int64_t>(a, b, outputs, op);
-      break;
-    case float16:
-      binary_op<float16_t>(a, b, outputs, op);
-      break;
-    case float32:
-      binary_op<float>(a, b, outputs, op);
-      break;
-    case bfloat16:
-      binary_op<bfloat16_t>(a, b, outputs, op);
-      break;
-    case complex64:
-      binary_op<complex64_t>(a, b, outputs, op);
-      break;
-  }
-}
-
 } // namespace
 
 } // namespace mlx::core

mlx/backend/cpu/cholesky.cpp

@@ -0,0 +1,85 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include "mlx/allocator.h"
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/lapack.h"
+#include "mlx/linalg.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+template <typename T>
+void cholesky_impl(const array& a, array& factor, bool upper, Stream stream) {
+  // Lapack uses the column-major convention. We take advantage of the fact that
+  // the matrix should be symmetric:
+  //   (A)ᵀ = A
+  // and that a column-major lower triangular matrix is a row-major upper
+  // triangular matrix, so uplo is the opposite of what we would expect from
+  // upper
+
+  // The decomposition is computed in place, so just copy the input to the
+  // output.
+  copy(
+      a,
+      factor,
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+      stream);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(factor);
+  encoder.dispatch([matrix = factor.data<T>(),
+                    upper,
+                    N = a.shape(-1),
+                    size = a.size()]() mutable {
+    char uplo = (upper) ? 'L' : 'U';
+    size_t num_matrices = size / (N * N);
+    for (int i = 0; i < num_matrices; i++) {
+      // Compute Cholesky factorization.
+      int info;
+      potrf<T>(
+          /* uplo = */ &uplo,
+          /* n = */ &N,
+          /* a = */ matrix,
+          /* lda = */ &N,
+          /* info = */ &info);
+
+      // TODO: We do nothing when the matrix is not positive semi-definite
+      // because throwing an error would result in a crash. If we figure out how
+      // to catch errors from the implementation we should throw.
+      if (info < 0) {
+        std::stringstream msg;
+        msg << "[Cholesky::eval_cpu] Cholesky decomposition failed with error code "
+            << info;
+        throw std::runtime_error(msg.str());
+      }
+
+      // Zero out the upper/lower triangle while advancing the pointer to the
+      // next matrix at the same time.
+      for (int row = 0; row < N; row++) {
+        if (upper) {
+          std::fill(matrix, matrix + row, 0);
+        } else {
+          std::fill(matrix + row + 1, matrix + N, 0);
+        }
+        matrix += N;
+      }
+    }
+  });
+}
+
+void Cholesky::eval_cpu(const std::vector<array>& inputs, array& output) {
+  switch (inputs[0].dtype()) {
+    case float32:
+      cholesky_impl<float>(inputs[0], output, upper_, stream());
+      break;
+    case float64:
+      cholesky_impl<double>(inputs[0], output, upper_, stream());
+      break;
+    default:
+      throw std::runtime_error(
+          "[Cholesky::eval_cpu] only supports float32 or float64.");
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/compiled.cpp

@@ -10,8 +10,9 @@
 #include <fmt/format.h>
 
 #include "mlx/backend/common/compiled.h"
-#include "mlx/backend/common/compiled_preamble.h"
-#include "mlx/backend/common/jit_compiler.h"
+#include "mlx/backend/cpu/compiled_preamble.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/jit_compiler.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
 
@@ -288,6 +289,7 @@ void Compiled::eval_cpu(
   // Figure out which kernel we are using
   auto& shape = outputs[0].shape();
   auto contiguous = compiled_check_contiguity(inputs, shape);
+  auto& encoder = cpu::get_command_encoder(stream());
 
   // Handle all broadcasting and collect function input arguments
   std::vector<void*> args;
@@ -298,6 +300,7 @@ void Compiled::eval_cpu(
       continue;
     }
     auto& x = inputs[i];
+    encoder.set_input_array(x);
     args.push_back((void*)x.data<void>());
 
     if (contiguous || is_scalar(x)) {
@@ -356,18 +359,25 @@ void Compiled::eval_cpu(
   });
 
   compiled_allocate_outputs(
-      inputs, outputs, inputs_, constant_ids_, contiguous, false);
+      inputs, outputs, inputs_, constant_ids_, contiguous);
 
   for (auto& x : outputs) {
     args.push_back(x.data<void>());
+    encoder.set_output_array(x);
   }
+  Shape out_shape;
   if (!contiguous) {
-    args.push_back((void*)outputs[0].shape().data());
+    out_shape = outputs[0].shape();
+    args.push_back((void*)out_shape.data());
   } else {
     args.push_back((void*)outputs[0].data_size());
   }
   auto fun = (void (*)(void**))fn_ptr;
-  fun(args.data());
+  encoder.dispatch(
+      [fun,
+       args = std::move(args),
+       strides = std::move(strides),
+       out_shape = std::move(out_shape)]() mutable { fun(args.data()); });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/compiled_preamble.h

@@ -5,7 +5,8 @@
 // clang-format off
 #include "mlx/types/half_types.h"
 #include "mlx/types/complex.h"
-#include "mlx/backend/common/ops.h"
+#include "mlx/backend/cpu/unary_ops.h"
+#include "mlx/backend/cpu/binary_ops.h"
 // clang-format on
 
 const char* get_kernel_preamble();

mlx/backend/cpu/copy.cpp

@@ -3,8 +3,10 @@
 #include <numeric>
 
 #include "mlx/allocator.h"
-#include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/simd/simd.h"
 
 namespace mlx::core {
 
@@ -12,18 +14,19 @@ namespace {
 
 template <typename SrcT, typename DstT>
 void copy_single(const array& src, array& dst) {
-  auto val = static_cast<DstT>(src.data<SrcT>()[0]);
+  auto src_ptr = src.data<SrcT>();
   auto dst_ptr = dst.data<DstT>();
-  for (int i = 0; i < dst.size(); ++i) {
-    dst_ptr[i] = val;
-  }
+  auto size = dst.size();
+  auto val = static_cast<DstT>(src_ptr[0]);
+  std::fill_n(dst_ptr, size, val);
 }
 
 template <typename SrcT, typename DstT>
 void copy_vector(const array& src, array& dst) {
   auto src_ptr = src.data<SrcT>();
   auto dst_ptr = dst.data<DstT>();
-  std::copy(src_ptr, src_ptr + src.data_size(), dst_ptr);
+  auto size = src.data_size();
+  std::copy(src_ptr, src_ptr + size, dst_ptr);
 }
 
 template <typename SrcT, typename DstT, int D>
@@ -58,36 +61,57 @@ void copy_general_general(
     const Strides& i_strides,
     const Strides& o_strides,
     int64_t i_offset,
-    int64_t o_offset) {
+    int64_t o_offset,
+    const std::optional<array>& dynamic_i_offset,
+    const std::optional<array>& dynamic_o_offset) {
+  auto src_ptr = src.data<SrcT>() + i_offset;
+  auto dst_ptr = dst.data<DstT>() + o_offset;
+  auto i_offset_ptr =
+      dynamic_i_offset ? dynamic_i_offset->data<int64_t>() : nullptr;
+  auto o_offset_ptr =
+      dynamic_o_offset ? dynamic_o_offset->data<int64_t>() : nullptr;
+  auto size = src.size();
   if (data_shape.empty()) {
-    auto val = static_cast<DstT>(*(src.data<SrcT>() + i_offset));
-    auto dst_ptr = dst.data<DstT>() + o_offset;
+    auto val = static_cast<DstT>(*src_ptr);
     *dst_ptr = val;
     return;
   }
   auto [shape, strides] =
       collapse_contiguous_dims(data_shape, {i_strides, o_strides});
-  auto src_ptr = src.data<SrcT>() + i_offset;
-  auto dst_ptr = dst.data<DstT>() + o_offset;
+
   int ndim = shape.size();
-  if (ndim == 1) {
-    copy_dims<SrcT, DstT, 1>(
-        src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
-    return;
-  } else if (ndim == 2) {
-    copy_dims<SrcT, DstT, 2>(
-        src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
-    return;
-  } else if (ndim == 3) {
-    copy_dims<SrcT, DstT, 3>(
-        src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
+  if (ndim < 3) {
+    if (i_offset_ptr) {
+      src_ptr += i_offset_ptr[0];
+    }
+    if (o_offset_ptr) {
+      dst_ptr += o_offset_ptr[0];
+    }
+
+    if (ndim == 1) {
+      copy_dims<SrcT, DstT, 1>(
+          src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
+    } else if (ndim == 2) {
+      copy_dims<SrcT, DstT, 2>(
+          src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
+    } else if (ndim == 3) {
+      copy_dims<SrcT, DstT, 3>(
+          src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
+    }
     return;
   }
+  if (i_offset_ptr) {
+    src_ptr += i_offset_ptr[0];
+  }
+  if (o_offset_ptr) {
+    dst_ptr += o_offset_ptr[0];
+  }
+
   ContiguousIterator in(shape, strides[0], ndim - 3);
   ContiguousIterator out(shape, strides[1], ndim - 3);
   auto stride = std::accumulate(
       shape.end() - 3, shape.end(), 1, std::multiplies<int64_t>());
-  for (int64_t elem = 0; elem < src.size(); elem += stride) {
+  for (int64_t elem = 0; elem < size; elem += stride) {
     copy_dims<SrcT, DstT, 3>(
         src_ptr + in.loc,
         dst_ptr + out.loc,
@@ -103,7 +127,15 @@ void copy_general_general(
 template <typename SrcT, typename DstT>
 inline void copy_general_general(const array& src, array& dst) {
   copy_general_general<SrcT, DstT>(
-      src, dst, src.shape(), src.strides(), dst.strides(), 0, 0);
+      src,
+      dst,
+      src.shape(),
+      src.strides(),
+      dst.strides(),
+      0,
+      0,
+      std::nullopt,
+      std::nullopt);
 }
 
 template <typename SrcT, typename DstT>
@@ -114,7 +146,9 @@ void copy_general(
     const Strides& i_strides,
     const Strides&,
     int64_t i_offset,
-    int64_t o_offset) {
+    int64_t o_offset,
+    const std::optional<array>& dynamic_i_offset,
+    const std::optional<array>& dynamic_o_offset) {
   copy_general_general<SrcT, DstT>(
       src,
       dst,
@@ -122,7 +156,9 @@ void copy_general(
       i_strides,
       make_contiguous_strides(data_shape),
       i_offset,
-      o_offset);
+      o_offset,
+      dynamic_i_offset,
+      dynamic_o_offset);
 }
 
 template <typename SrcT, typename DstT>
@@ -134,7 +170,9 @@ inline void copy_general(const array& src, array& dst) {
       src.strides(),
       make_contiguous_strides(src.shape()),
       0,
-      0);
+      0,
+      std::nullopt,
+      std::nullopt);
 }
 
 template <typename SrcT, typename DstT, typename... Args>
@@ -191,6 +229,9 @@ void copy(const array& src, array& dst, CopyType ctype, Args&&... args) {
     case float32:
       copy<SrcT, float>(src, dst, ctype, std::forward<Args>(args)...);
       break;
+    case float64:
+      copy<SrcT, double>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
     case bfloat16:
       copy<SrcT, bfloat16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
@@ -240,6 +281,9 @@ inline void copy_inplace_dispatch(
     case float32:
       copy<float>(src, dst, ctype, std::forward<Args>(args)...);
       break;
+    case float64:
+      copy<double>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
     case bfloat16:
       copy<bfloat16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
@@ -251,35 +295,27 @@ inline void copy_inplace_dispatch(
 
 } // namespace
 
-void copy_inplace(const array& src, array& dst, CopyType ctype) {
-  copy_inplace_dispatch(src, dst, ctype);
+void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(src);
+  encoder.set_output_array(dst);
+  encoder.dispatch(
+      [src = array::unsafe_weak_copy(src),
+       dst = array::unsafe_weak_copy(dst),
+       ctype]() mutable { copy_inplace_dispatch(src, dst, ctype); });
 }
 
-void copy(const array& src, array& dst, CopyType ctype) {
-  // Allocate the output
-  switch (ctype) {
-    case CopyType::Vector:
-      if (src.is_donatable() && src.itemsize() == dst.itemsize()) {
-        dst.copy_shared_buffer(src);
-      } else {
-        auto size = src.data_size();
-        dst.set_data(
-            allocator::malloc_or_wait(size * dst.itemsize()),
-            size,
-            src.strides(),
-            src.flags());
-      }
-      break;
-    case CopyType::Scalar:
-    case CopyType::General:
-    case CopyType::GeneralGeneral:
-      dst.set_data(allocator::malloc_or_wait(dst.nbytes()));
-      break;
+void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
+  bool donated = set_copy_output_data(src, dst, ctype);
+  if (donated && src.dtype() == dst.dtype()) {
+    // If the output has the same type as the input then there is nothing to
+    // copy, just use the buffer.
+    return;
   }
   if (ctype == CopyType::GeneralGeneral) {
     ctype = CopyType::General;
   }
-  copy_inplace(src, dst, ctype);
+  copy_inplace(src, dst, ctype, stream);
 }
 
 void copy_inplace(
@@ -290,24 +326,51 @@ void copy_inplace(
     const Strides& o_strides,
     int64_t i_offset,
     int64_t o_offset,
-    CopyType ctype) {
-  switch (ctype) {
-    case CopyType::General:
-    case CopyType::GeneralGeneral:
-      copy_inplace_dispatch(
-          src,
-          dst,
-          ctype,
-          data_shape,
-          i_strides,
-          o_strides,
-          i_offset,
-          o_offset);
-      break;
-    case CopyType::Scalar:
-    case CopyType::Vector:
-      copy_inplace_dispatch(src, dst, ctype);
-  }
+    CopyType ctype,
+    Stream stream,
+    const std::optional<array>& dynamic_i_offset, /* = std::nullopt */
+    const std::optional<array>& dynamic_o_offset /* = std::nullopt */) {
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(src);
+  encoder.set_output_array(dst);
+  auto weak_copy_if_set = [](auto x) -> std::optional<array> {
+    if (x) {
+      return array::unsafe_weak_copy(*x);
+    } else {
+      return std::nullopt;
+    }
+  };
+  encoder.dispatch(
+      [src = array::unsafe_weak_copy(src),
+       dst = array::unsafe_weak_copy(dst),
+       data_shape,
+       i_strides,
+       o_strides,
+       i_offset,
+       o_offset,
+       ctype,
+       dynamic_i_offset = weak_copy_if_set(dynamic_i_offset),
+       dynamic_o_offset = weak_copy_if_set(dynamic_o_offset)]() mutable {
+        switch (ctype) {
+          case CopyType::General:
+          case CopyType::GeneralGeneral:
+            copy_inplace_dispatch(
+                src,
+                dst,
+                ctype,
+                data_shape,
+                i_strides,
+                o_strides,
+                i_offset,
+                o_offset,
+                dynamic_i_offset,
+                dynamic_o_offset);
+            break;
+          case CopyType::Scalar:
+          case CopyType::Vector:
+            copy_inplace_dispatch(src, dst, ctype);
+        }
+      });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/copy.h

@@ -0,0 +1,29 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#pragma once
+
+#include <optional>
+
+#include "mlx/array.h"
+#include "mlx/backend/common/copy.h"
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+void copy(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream);
+
+void copy_inplace(
+    const array& src,
+    array& dst,
+    const Shape& data_shape,
+    const Strides& i_strides,
+    const Strides& o_strides,
+    int64_t i_offset,
+    int64_t o_offset,
+    CopyType ctype,
+    Stream stream,
+    const std::optional<array>& dynamic_i_offset = std::nullopt,
+    const std::optional<array>& dynamic_o_offset = std::nullopt);
+
+} // namespace mlx::core

mlx/backend/cpu/distributed.cpp

@@ -0,0 +1,94 @@
+// Copyright © 2024 Apple Inc.
+
+#include <cassert>
+
+#include "mlx/allocator.h"
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/distributed/primitives.h"
+
+namespace mlx::core::distributed {
+
+std::pair<array, bool> ensure_row_contiguous(const array& arr, Stream stream) {
+  if (arr.flags().row_contiguous) {
+    return {arr, false};
+  } else {
+    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+    copy(arr, arr_copy, CopyType::General, stream);
+    return {arr_copy, true};
+  }
+};
+
+void AllReduce::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto donate_or_copy = [s = stream()](const array& in, array& out) {
+    if (in.flags().row_contiguous) {
+      if (in.is_donatable()) {
+        out.copy_shared_buffer(in);
+      } else {
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
+      }
+      return in;
+    } else {
+      array arr_copy(in.shape(), in.dtype(), nullptr, {});
+      copy(in, arr_copy, CopyType::General, s);
+      out.copy_shared_buffer(arr_copy);
+      return arr_copy;
+    }
+  };
+
+  auto in = donate_or_copy(inputs[0], outputs[0]);
+  switch (reduce_type_) {
+    case Sum:
+      distributed::detail::all_sum(group(), in, outputs[0], stream());
+      break;
+    default:
+      throw std::runtime_error("Only all reduce sum is supported for now");
+  }
+}
+
+void AllGather::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto [in, copied] = ensure_row_contiguous(inputs[0], stream());
+  outputs[0].set_data(allocator::malloc_or_wait(outputs[0].nbytes()));
+  distributed::detail::all_gather(group(), in, outputs[0], stream());
+  if (copied) {
+    auto& enc = cpu::get_command_encoder(stream());
+    enc.add_temporary(in);
+  }
+}
+
+void Send::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto [in, copied] = ensure_row_contiguous(inputs[0], stream());
+  distributed::detail::send(group(), in, dst_, stream());
+  outputs[0].copy_shared_buffer(inputs[0]);
+  if (copied) {
+    auto& enc = cpu::get_command_encoder(stream());
+    enc.add_temporary(in);
+  }
+}
+
+void Recv::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 0);
+  assert(outputs.size() == 1);
+
+  outputs[0].set_data(allocator::malloc_or_wait(outputs[0].nbytes()));
+  distributed::detail::recv(group(), outputs[0], src_, stream());
+}
+
+} // namespace mlx::core::distributed

mlx/backend/cpu/eigh.cpp

@@ -0,0 +1,141 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include "mlx/allocator.h"
+#include "mlx/array.h"
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/lapack.h"
+#include "mlx/linalg.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+namespace {
+
+template <typename T>
+void eigh_impl(
+    array& vectors,
+    array& values,
+    const std::string& uplo,
+    bool compute_eigenvectors,
+    Stream stream) {
+  auto vec_ptr = vectors.data<T>();
+  auto eig_ptr = values.data<T>();
+  char jobz = compute_eigenvectors ? 'V' : 'N';
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(vectors);
+  encoder.set_output_array(values);
+  encoder.dispatch([vec_ptr,
+                    eig_ptr,
+                    jobz,
+                    uplo = uplo[0],
+                    N = vectors.shape(-1),
+                    size = vectors.size()]() mutable {
+    // Work query
+    int lwork = -1;
+    int liwork = -1;
+    int info;
+    {
+      T work;
+      int iwork;
+      syevd<T>(
+          &jobz,
+          &uplo,
+          &N,
+          nullptr,
+          &N,
+          nullptr,
+          &work,
+          &lwork,
+          &iwork,
+          &liwork,
+          &info);
+      lwork = static_cast<int>(work);
+      liwork = iwork;
+    }
+
+    auto work_buf = array::Data{allocator::malloc_or_wait(sizeof(T) * lwork)};
+    auto iwork_buf =
+        array::Data{allocator::malloc_or_wait(sizeof(int) * liwork)};
+    for (size_t i = 0; i < size / (N * N); ++i) {
+      syevd<T>(
+          &jobz,
+          &uplo,
+          &N,
+          vec_ptr,
+          &N,
+          eig_ptr,
+          static_cast<T*>(work_buf.buffer.raw_ptr()),
+          &lwork,
+          static_cast<int*>(iwork_buf.buffer.raw_ptr()),
+          &liwork,
+          &info);
+      vec_ptr += N * N;
+      eig_ptr += N;
+      if (info != 0) {
+        std::stringstream msg;
+        msg << "[Eigh::eval_cpu] Eigenvalue decomposition failed with error code "
+            << info;
+        throw std::runtime_error(msg.str());
+      }
+    }
+  });
+  if (!compute_eigenvectors) {
+    encoder.add_temporary(vectors);
+  }
+}
+
+} // namespace
+
+void Eigh::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  const auto& a = inputs[0];
+  auto& values = outputs[0];
+
+  auto vectors = compute_eigenvectors_
+      ? outputs[1]
+      : array(a.shape(), a.dtype(), nullptr, {});
+
+  values.set_data(allocator::malloc_or_wait(values.nbytes()));
+
+  copy(
+      a,
+      vectors,
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+      stream());
+
+  if (compute_eigenvectors_) {
+    // Set the strides and flags so the eigenvectors
+    // are in the columns of the output
+    auto flags = vectors.flags();
+    auto strides = vectors.strides();
+    auto ndim = a.ndim();
+    std::swap(strides[ndim - 1], strides[ndim - 2]);
+
+    if (a.size() > 1) {
+      flags.row_contiguous = false;
+      if (ndim > 2) {
+        flags.col_contiguous = false;
+      } else {
+        flags.col_contiguous = true;
+      }
+    }
+    vectors.copy_shared_buffer(vectors, strides, flags, vectors.data_size());
+  }
+  switch (a.dtype()) {
+    case float32:
+      eigh_impl<float>(vectors, values, uplo_, compute_eigenvectors_, stream());
+      break;
+    case float64:
+      eigh_impl<double>(
+          vectors, values, uplo_, compute_eigenvectors_, stream());
+      break;
+    default:
+      throw std::runtime_error(
+          "[Eigh::eval_cpu] only supports float32 or float64.");
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/encoder.cpp

@@ -0,0 +1,16 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cpu/encoder.h"
+
+namespace mlx::core::cpu {
+
+CommandEncoder& get_command_encoder(Stream stream) {
+  static std::unordered_map<int, CommandEncoder> encoder_map;
+  auto it = encoder_map.find(stream.index);
+  if (it == encoder_map.end()) {
+    it = encoder_map.emplace(stream.index, stream).first;
+  }
+  return it->second;
+}
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/encoder.h

@@ -0,0 +1,53 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <unordered_map>
+
+#include "mlx/array.h"
+#include "mlx/scheduler.h"
+
+namespace mlx::core::cpu {
+
+struct CommandEncoder {
+  CommandEncoder(Stream stream) : stream_(stream) {}
+
+  CommandEncoder(const CommandEncoder&) = delete;
+  CommandEncoder& operator=(const CommandEncoder&) = delete;
+  CommandEncoder(CommandEncoder&&) = delete;
+  CommandEncoder& operator=(CommandEncoder&&) = delete;
+
+  void set_input_array(const array& a) {}
+  void set_output_array(array& a) {}
+
+  // Hold onto a temporary until any already scheduled tasks which use it as
+  // an input are complete.
+  void add_temporary(array arr) {
+    temporaries_.push_back(std::move(arr));
+  }
+
+  void add_temporaries(std::vector<array> arrays) {
+    temporaries_.insert(
+        temporaries_.end(),
+        std::make_move_iterator(arrays.begin()),
+        std::make_move_iterator(arrays.end()));
+  }
+
+  std::vector<array>& temporaries() {
+    return temporaries_;
+  }
+
+  template <class F, class... Args>
+  void dispatch(F&& f, Args&&... args) {
+    auto task = std::bind(std::forward<F>(f), std::forward<Args>(args)...);
+    scheduler::enqueue(stream_, std::move(task));
+  }
+
+ private:
+  Stream stream_;
+  std::vector<array> temporaries_;
+};
+
+CommandEncoder& get_command_encoder(Stream stream);
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/eval.cpp

@@ -0,0 +1,44 @@
+// Copyright © 2025 Apple Inc.
+#include "mlx/backend/cpu/eval.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/primitives.h"
+#include "mlx/scheduler.h"
+#include "mlx/utils.h"
+
+namespace mlx::core::cpu {
+
+void eval(array& arr) {
+  auto s = arr.primitive().stream();
+
+  auto outputs = arr.outputs();
+  {
+    // If the array is a tracer hold a reference
+    // to its inputs so they don't get donated
+    std::vector<array> inputs;
+    if (arr.is_tracer()) {
+      inputs = arr.inputs();
+    }
+    arr.primitive().eval_cpu(arr.inputs(), outputs);
+  }
+
+  std::unordered_set<std::shared_ptr<array::Data>> buffers;
+  for (auto& in : arr.inputs()) {
+    buffers.insert(in.data_shared_ptr());
+  }
+  for (auto& s : arr.siblings()) {
+    buffers.insert(s.data_shared_ptr());
+  }
+  // Remove the output if it was donated to by an input
+  if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
+    buffers.erase(it);
+  }
+  auto& encoder = cpu::get_command_encoder(s);
+  scheduler::notify_new_task(s);
+  encoder.dispatch([s,
+                    buffers = std::move(buffers),
+                    temps = std::move(encoder.temporaries())]() {
+    scheduler::notify_task_completion(s);
+  });
+}
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/eval.h

@@ -0,0 +1,12 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/stream.h"
+
+namespace mlx::core::cpu {
+
+void eval(array& arr);
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/fft.cpp

@@ -0,0 +1,120 @@
+// Copyright © 2023 Apple Inc.
+
+#include <numeric>
+
+#include "mlx/3rdparty/pocketfft.h"
+#include "mlx/allocator.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+void FFT::eval_cpu(const std::vector<array>& inputs, array& out) {
+  auto& in = inputs[0];
+  std::vector<std::ptrdiff_t> strides_in(
+      in.strides().begin(), in.strides().end());
+  for (auto& s : strides_in) {
+    s *= in.itemsize();
+  }
+  std::vector<std::ptrdiff_t> strides_out(
+      out.strides().begin(), out.strides().end());
+  for (auto& s : strides_out) {
+    s *= out.itemsize();
+  }
+
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  std::vector<size_t> shape;
+  if (out.dtype() == float32) {
+    shape.insert(shape.end(), out.shape().begin(), out.shape().end());
+  } else {
+    shape.insert(shape.end(), in.shape().begin(), in.shape().end());
+  }
+
+  float scale = 1.0f;
+  if (inverse_) {
+    size_t nelem = std::accumulate(
+        axes_.begin(), axes_.end(), 1, [&shape](auto x, auto y) {
+          return x * shape[y];
+        });
+    scale /= nelem;
+  }
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+
+  if (in.dtype() == complex64 && out.dtype() == complex64) {
+    auto in_ptr =
+        reinterpret_cast<const std::complex<float>*>(in.data<complex64_t>());
+    auto out_ptr =
+        reinterpret_cast<std::complex<float>*>(out.data<complex64_t>());
+    encoder.dispatch([shape = std::move(shape),
+                      strides_in = std::move(strides_in),
+                      strides_out = std::move(strides_out),
+                      axes = axes_,
+                      inverse = inverse_,
+                      in_ptr,
+                      out_ptr,
+                      scale]() {
+      pocketfft::c2c(
+          shape,
+          strides_in,
+          strides_out,
+          axes,
+          !inverse,
+          in_ptr,
+          out_ptr,
+          scale);
+    });
+  } else if (in.dtype() == float32 && out.dtype() == complex64) {
+    auto in_ptr = in.data<float>();
+    auto out_ptr =
+        reinterpret_cast<std::complex<float>*>(out.data<complex64_t>());
+    encoder.dispatch([shape = std::move(shape),
+                      strides_in = std::move(strides_in),
+                      strides_out = std::move(strides_out),
+                      axes = axes_,
+                      inverse = inverse_,
+                      in_ptr,
+                      out_ptr,
+                      scale]() {
+      pocketfft::r2c(
+          shape,
+          strides_in,
+          strides_out,
+          axes,
+          !inverse,
+          in_ptr,
+          out_ptr,
+          scale);
+    });
+  } else if (in.dtype() == complex64 && out.dtype() == float32) {
+    auto in_ptr =
+        reinterpret_cast<const std::complex<float>*>(in.data<complex64_t>());
+    auto out_ptr = out.data<float>();
+    encoder.dispatch([shape = std::move(shape),
+                      strides_in = std::move(strides_in),
+                      strides_out = std::move(strides_out),
+                      axes = axes_,
+                      inverse = inverse_,
+                      in_ptr,
+                      out_ptr,
+                      scale]() {
+      pocketfft::c2r(
+          shape,
+          strides_in,
+          strides_out,
+          axes,
+          !inverse,
+          in_ptr,
+          out_ptr,
+          scale);
+    });
+  } else {
+    throw std::runtime_error(
+        "[FFT] Received unexpected input and output type combination.");
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/gemm.h

@@ -0,0 +1,26 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+#include "mlx/array.h"
+
+namespace mlx::core {
+
+template <typename T>
+void matmul(
+    const T* a,
+    const T* b,
+    T* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides);
+
+} // namespace mlx::core

mlx/backend/cpu/gemms/bnns.cpp

@@ -0,0 +1,209 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include <Accelerate/Accelerate.h>
+
+#include "mlx/array.h"
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/gemm.h"
+#include "mlx/dtype.h"
+
+namespace mlx::core {
+
+template <typename T>
+constexpr BNNSDataType to_bnns_dtype();
+
+template <>
+constexpr BNNSDataType to_bnns_dtype<float>() {
+  return BNNSDataType(BNNSDataTypeFloatBit | 32);
+}
+template <>
+constexpr BNNSDataType to_bnns_dtype<float16_t>() {
+  return BNNSDataType(BNNSDataTypeFloatBit | 16);
+}
+
+template <>
+constexpr BNNSDataType to_bnns_dtype<bfloat16_t>() {
+  return BNNSDataTypeBFloat16;
+}
+
+template <typename T>
+void matmul_bnns(
+    const T* a,
+    const T* b,
+    T* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
+
+  BNNSDataType bnns_dtype = to_bnns_dtype<T>();
+
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
+  const BNNSLayerParametersBroadcastMatMul gemm_params{
+      /* float alpha = */ alpha,
+      /* float beta = */ beta,
+      /* bool transA = */ a_transposed,
+      /* bool transB = */ b_transposed,
+      /* bool quadratic = */ false,
+      /* bool a_is_weights = */ false,
+      /* bool b_is_weights = */ false,
+      /* BNNSNDArrayDescriptor iA_desc = */
+      BNNSNDArrayDescriptor{
+          /* BNNSNDArrayFlags flags = */ BNNSNDArrayFlagBackpropSet,
+          /* BNNSDataLayout layout = */ BNNSDataLayoutRowMajorMatrix,
+
+          /* size_t size[BNNS_MAX_TENSOR_DIMENSION] = */
+          {lda, (M * K) / lda, 0, 0, 0, 0, 0, 0},
+          /* size_t stride[BNNS_MAX_TENSOR_DIMENSION] = */
+          {1, lda, 0, 0, 0, 0, 0, 0},
+
+          /* void * _Nullable data = */ nullptr,
+          /* BNNSDataType data_type = */ bnns_dtype,
+
+          /* void * _Nullable table_data = */ nullptr,
+          /* BNNSDataType table_data_type = */ bnns_dtype,
+
+          /* float data_scale = */ 1.0,
+          /* float data_bias = */ 0.0,
+      },
+      /* BNNSNDArrayDescriptor iB_desc = */
+      BNNSNDArrayDescriptor{
+          /* BNNSNDArrayFlags flags = */ BNNSNDArrayFlagBackpropSet,
+          /* BNNSDataLayout layout = */ BNNSDataLayoutRowMajorMatrix,
+
+          /* size_t size[BNNS_MAX_TENSOR_DIMENSION] = */
+          {ldb, (K * N) / ldb, 0, 0, 0, 0, 0, 0},
+          /* size_t stride[BNNS_MAX_TENSOR_DIMENSION] = */
+          {1, ldb, 0, 0, 0, 0, 0, 0},
+
+          /* void * _Nullable data = */ nullptr,
+          /* BNNSDataType data_type = */ bnns_dtype,
+
+          /* void * _Nullable table_data = */ nullptr,
+          /* BNNSDataType table_data_type = */ bnns_dtype,
+
+          /* float data_scale = */ 1.0,
+          /* float data_bias = */ 0.0,
+      },
+      /* BNNSNDArrayDescriptor o_desc = */
+      BNNSNDArrayDescriptor{
+          /* BNNSNDArrayFlags flags = */ BNNSNDArrayFlagBackpropSet,
+          /* BNNSDataLayout layout = */ BNNSDataLayoutRowMajorMatrix,
+
+          /* size_t size[BNNS_MAX_TENSOR_DIMENSION] = */
+          {N, M, 0, 0, 0, 0, 0, 0},
+          /* size_t stride[BNNS_MAX_TENSOR_DIMENSION] = */
+          {1, N, 0, 0, 0, 0, 0, 0},
+
+          /* void * _Nullable data = */ nullptr,
+          /* BNNSDataType data_type = */ bnns_dtype,
+
+          /* void * _Nullable table_data = */ nullptr,
+          /* BNNSDataType table_data_type = */ bnns_dtype,
+
+          /* float data_scale = */ 1.0,
+          /* float data_bias = */ 0.0,
+      },
+  };
+
+  auto bnns_filter =
+      BNNSFilterCreateLayerBroadcastMatMul(&gemm_params, nullptr);
+
+  for (int i = 0; i < batch_size; ++i) {
+    BNNSFilterApplyTwoInput(
+        bnns_filter,
+        reinterpret_cast<const uint8_t*>(
+            a + elem_to_loc(M * K * i, a_shape, a_strides)),
+        reinterpret_cast<const uint8_t*>(
+            b + elem_to_loc(K * N * i, b_shape, b_strides)),
+        reinterpret_cast<uint8_t*>(out + M * N * i));
+  }
+
+  BNNSFilterDestroy(bnns_filter);
+#pragma GCC diagnostic pop
+}
+
+template <>
+void matmul<float16_t>(
+    const float16_t* a,
+    const float16_t* b,
+    float16_t* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  matmul_bnns(
+      a,
+      b,
+      out,
+      a_transposed,
+      b_transposed,
+      lda,
+      ldb,
+      ldc,
+      alpha,
+      beta,
+      batch_size,
+      a_shape,
+      a_strides,
+      b_shape,
+      b_strides);
+}
+
+template <>
+void matmul<bfloat16_t>(
+    const bfloat16_t* a,
+    const bfloat16_t* b,
+    bfloat16_t* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  matmul_bnns(
+      a,
+      b,
+      out,
+      a_transposed,
+      b_transposed,
+      lda,
+      ldb,
+      ldc,
+      alpha,
+      beta,
+      batch_size,
+      a_shape,
+      a_strides,
+      b_shape,
+      b_strides);
+}
+
+} // namespace mlx::core

mlx/backend/cpu/gemms/cblas.cpp

@@ -0,0 +1,91 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/gemm.h"
+#include "mlx/backend/cpu/lapack.h"
+
+namespace mlx::core {
+
+template <>
+void matmul<float>(
+    const float* a,
+    const float* b,
+    float* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
+
+  for (int i = 0; i < batch_size; ++i) {
+    cblas_sgemm(
+        CblasRowMajor,
+        a_transposed ? CblasTrans : CblasNoTrans, // transA
+        b_transposed ? CblasTrans : CblasNoTrans, // transB
+        M,
+        N,
+        K,
+        alpha,
+        a + elem_to_loc(M * K * i, a_shape, a_strides),
+        lda,
+        b + elem_to_loc(K * N * i, b_shape, b_strides),
+        ldb,
+        beta,
+        out + M * N * i,
+        ldc);
+  }
+}
+
+template <>
+void matmul<double>(
+    const double* a,
+    const double* b,
+    double* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
+
+  for (int i = 0; i < batch_size; ++i) {
+    cblas_dgemm(
+        CblasRowMajor,
+        a_transposed ? CblasTrans : CblasNoTrans, // transA
+        b_transposed ? CblasTrans : CblasNoTrans, // transB
+        M,
+        N,
+        K,
+        alpha,
+        a + elem_to_loc(M * K * i, a_shape, a_strides),
+        lda,
+        b + elem_to_loc(K * N * i, b_shape, b_strides),
+        ldb,
+        beta,
+        out + M * N * i,
+        ldc);
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/gemms/no_bf16.cpp

@@ -0,0 +1,27 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cpu/gemm.h"
+
+namespace mlx::core {
+
+template <>
+void matmul<bfloat16_t>(
+    const bfloat16_t*,
+    const bfloat16_t*,
+    bfloat16_t*,
+    bool,
+    bool,
+    size_t,
+    size_t,
+    size_t,
+    float,
+    float,
+    size_t,
+    const Shape&,
+    const Strides&,
+    const Shape&,
+    const Strides&) {
+  throw std::runtime_error("[Matmul::eval_cpu] bfloat16 not supported.");
+}
+
+} // namespace mlx::core

mlx/backend/cpu/gemms/no_fp16.cpp

@@ -0,0 +1,27 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cpu/gemm.h"
+
+namespace mlx::core {
+
+template <>
+void matmul<float16_t>(
+    const float16_t*,
+    const float16_t*,
+    float16_t*,
+    bool,
+    bool,
+    size_t,
+    size_t,
+    size_t,
+    float,
+    float,
+    size_t,
+    const Shape&,
+    const Strides&,
+    const Shape&,
+    const Strides&) {
+  throw std::runtime_error("[Matmul::eval_cpu] float16 not supported.");
+}
+
+} // namespace mlx::core

mlx/backend/cpu/hadamard.cpp

@@ -2,18 +2,19 @@
 
 #include <cassert>
 
-#include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/hadamard.h"
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
 // n = 2^k component
 template <typename T>
-void hadamard_n(array& out, int n, int m, float scale) {
-  for (int b = 0; b < out.size() / n; b++) {
+void hadamard_n(T* out, int n, int m, float scale, size_t size) {
+  for (int b = 0; b < size / n; b++) {
     size_t loc = b * n;
-    T* data_ptr = out.data<T>() + loc;
+    T* data_ptr = out + loc;
     int h = 1;
     int n_over_2 = n / 2;
     while (h < n) {
@@ -36,7 +37,7 @@ void hadamard_n(array& out, int n, int m, float scale) {
 
 // m component
 template <typename T>
-void hadamard_m(array& out, int n, int m, float scale) {
+void hadamard_m(T* out, int n, int m, float scale, size_t size) {
   auto h_matrices = hadamard_matrices();
   auto& matrix = h_matrices[m];
   auto start = 1;
@@ -51,9 +52,9 @@ void hadamard_m(array& out, int n, int m, float scale) {
     end = matrix.find('\n', start);
   }
 
-  for (int b = 0; b < out.size() / m / n; b++) {
+  for (int b = 0; b < size / m / n; b++) {
     size_t loc = b * n * m;
-    T* data_ptr = out.data<T>() + loc;
+    T* data_ptr = out + loc;
     for (int i = 0; i < n; i++) {
       std::vector<float> out(m);
       for (int j = 0; j < m; j++) {
@@ -74,34 +75,47 @@ void hadamard_m(array& out, int n, int m, float scale) {
 }
 
 template <typename T>
-void hadamard(array& out, int n, int m, float scale) {
-  float n_scale = m > 1 ? 1.0 : scale;
-  hadamard_n<T>(out, n, m, n_scale);
-  if (m > 1) {
-    hadamard_m<T>(out, n, m, scale);
-  }
+void hadamard(array& out, int n, int m, float scale, Stream stream) {
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(out);
+  auto out_ptr = out.data<T>();
+  encoder.dispatch([out_ptr, size = out.size(), n, m, scale]() {
+    float n_scale = m > 1 ? 1.0 : scale;
+    hadamard_n<T>(out_ptr, n, m, n_scale, size);
+    if (m > 1) {
+      hadamard_m<T>(out_ptr, n, m, scale, size);
+    }
+  });
 }
 
-void Hadamard::eval(const std::vector<array>& inputs, array& out) {
+void Hadamard::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
 
   // Copy input to output
-  copy(in, out, CopyType::General);
+  if (in.flags().row_contiguous && in.is_donatable()) {
+    out.copy_shared_buffer(in);
+  } else {
+    copy(
+        in,
+        out,
+        in.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+        stream());
+  }
 
   int axis = out.ndim() - 1;
   auto [n, m] = decompose_hadamard(out.shape(axis));
 
   switch (in.dtype()) {
     case float32:
-      return hadamard<float>(out, n, m, scale_);
+      return hadamard<float>(out, n, m, scale_, stream());
     case float16:
-      return hadamard<float16_t>(out, n, m, scale_);
+      return hadamard<float16_t>(out, n, m, scale_, stream());
     case bfloat16:
-      return hadamard<bfloat16_t>(out, n, m, scale_);
+      return hadamard<bfloat16_t>(out, n, m, scale_, stream());
     default:
       throw std::invalid_argument("[hadamard] Unsupported type.");
   }
 }
 
-} // namespace mlx::core
+} // namespace mlx::core

mlx/backend/cpu/indexing.cpp

@@ -0,0 +1,758 @@
+// Copyright © 2023 Apple Inc.
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+
+#include "mlx/allocator.h"
+#include "mlx/primitives.h"
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
+
+namespace mlx::core {
+
+template <typename IdxT>
+inline size_t offset_neg_idx(IdxT idx, size_t size) {
+  return (idx < 0) ? idx + size : idx;
+}
+
+template <>
+inline size_t offset_neg_idx(uint32_t idx, size_t) {
+  return idx;
+}
+
+struct None {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) = x;
+  }
+};
+struct Sum {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) += x;
+  }
+};
+
+struct Prod {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) *= x;
+  }
+};
+
+struct Max {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) = (*y > x) ? *y : x;
+  }
+};
+
+struct Min {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) = (*y < x) ? *y : x;
+  }
+};
+
+template <typename T, typename IdxT>
+void gather(
+    const array& src,
+    const std::vector<array>& inds,
+    array& out,
+    const std::vector<int>& axes,
+    const Shape& slice_sizes) {
+  // If the array is row contiguous then we can do a contiguous copy given
+  // two conditions on the slice size:
+  // - Any number of leading ones in the slice sizes are allowed
+  // - All other slice sizes match the corresponding dimension except the
+  //   first non-singleton slice size
+  // If the array is col contiguous then the reverse is the case:
+  // - Any number of trailing ones in the slice sizes are allowed
+  // - All other slice sizes match the corresponding dimension except the
+  //   first non-singleton slice size from the end
+
+  bool can_copy = false;
+  if (src.flags().row_contiguous) {
+    can_copy = true;
+
+    // Ignore leading 1s
+    int i = 0;
+    for (; i < slice_sizes.size() && slice_sizes[i] == 1; ++i)
+      ;
+
+    // Check the remaining
+    i++;
+    for (; i < src.ndim() && can_copy; ++i) {
+      can_copy = (src.shape(i) == slice_sizes[i]);
+    }
+  } else if (src.flags().col_contiguous) {
+    can_copy = true;
+
+    // Ignore trailing 1s
+    int i = slice_sizes.size() - 1;
+    for (; i >= 0 && slice_sizes[i] == 1; --i)
+      ;
+
+    // Skip the next slice size and check the remaining
+    i--;
+    for (; i >= 0 && can_copy; --i) {
+      can_copy = (src.shape(i) == slice_sizes[i]);
+    }
+  }
+  size_t slice_size = 1;
+  for (auto s : slice_sizes) {
+    slice_size *= s;
+  }
+  size_t ind_size = slice_size == 0 ? 0 : out.size() / slice_size;
+  const T* src_ptr = src.data<T>();
+  T* dst_ptr = out.data<T>();
+
+  std::vector<ContiguousIterator> its(inds.begin(), inds.end());
+  ContiguousIterator src_it;
+  if (!can_copy && src.ndim() > 0) {
+    src_it = ContiguousIterator(slice_sizes, src.strides(), src.ndim());
+  }
+
+  size_t out_idx = 0;
+  for (int idx = 0; idx < ind_size; idx++) {
+    size_t src_idx = 0;
+    for (int ii = 0; ii < inds.size(); ++ii) {
+      auto ax = axes[ii];
+      auto idx_loc = its[ii].loc;
+      its[ii].step();
+      auto idx_val =
+          offset_neg_idx(inds[ii].data<IdxT>()[idx_loc], src.shape(ax));
+      src_idx += (idx_val * src.strides()[ax]);
+    }
+
+    if (slice_size == 1) {
+      dst_ptr[out_idx++] = src_ptr[src_idx];
+    } else if (can_copy) {
+      std::copy(
+          src_ptr + src_idx, src_ptr + src_idx + slice_size, dst_ptr + out_idx);
+      out_idx += slice_size;
+    } else {
+      for (int jj = 0; jj < slice_size; jj++) {
+        dst_ptr[out_idx++] = src_ptr[src_idx + src_it.loc];
+        src_it.step();
+      }
+      src_it.reset();
+    }
+  }
+}
+
+template <typename IdxT>
+void dispatch_gather(
+    const array& src,
+    const std::vector<array>& inds,
+    array& out,
+    const std::vector<int>& axes,
+    const Shape& size) {
+  switch (out.dtype()) {
+    case bool_:
+      gather<bool, IdxT>(src, inds, out, axes, size);
+      break;
+    case uint8:
+      gather<uint8_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case uint16:
+      gather<uint16_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case uint32:
+      gather<uint32_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case uint64:
+      gather<uint64_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case int8:
+      gather<int8_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case int16:
+      gather<int16_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case int32:
+      gather<int32_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case int64:
+      gather<int64_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case float16:
+      gather<float16_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case float32:
+      gather<float, IdxT>(src, inds, out, axes, size);
+      break;
+    case float64:
+      gather<double, IdxT>(src, inds, out, axes, size);
+      break;
+    case bfloat16:
+      gather<bfloat16_t, IdxT>(src, inds, out, axes, size);
+      break;
+    case complex64:
+      gather<complex64_t, IdxT>(src, inds, out, axes, size);
+      break;
+  }
+}
+
+void Gather::eval_cpu(const std::vector<array>& inputs, array& out) {
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  auto& src = inputs[0];
+  std::vector<array> inds;
+  for (auto it = inputs.begin() + 1; it < inputs.end(); ++it) {
+    inds.push_back(array::unsafe_weak_copy(*it));
+  }
+  auto& encoder = cpu::get_command_encoder(stream());
+  for (auto& in : inputs) {
+    encoder.set_input_array(in);
+  }
+  encoder.set_output_array(out);
+  encoder.dispatch([axes_ = axes_,
+                    slice_sizes_ = slice_sizes_,
+                    src = array::unsafe_weak_copy(src),
+                    inds = std::move(inds),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    if (inds.empty()) {
+      dispatch_gather<uint8_t>(src, inds, out, axes_, slice_sizes_);
+      return;
+    }
+
+    switch (inds[0].dtype()) {
+      case uint8:
+        dispatch_gather<uint8_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case uint16:
+        dispatch_gather<uint16_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case uint32:
+        dispatch_gather<uint32_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case uint64:
+        dispatch_gather<uint64_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case int8:
+        dispatch_gather<int8_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case int16:
+        dispatch_gather<int16_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case int32:
+        dispatch_gather<int32_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case int64:
+        dispatch_gather<int64_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      default:
+        throw std::runtime_error(
+            "[Gather::eval_cpu] Cannot gather with indices type.");
+        break;
+    }
+  });
+}
+template <typename T, typename IdxT>
+void gather_axis(
+    const array& src,
+    const array& ind,
+    array& out,
+    const int axis) {
+  auto strides = ind.strides();
+  strides.erase(strides.begin() + axis);
+  auto shape = ind.shape();
+  shape.erase(shape.begin() + axis);
+  ContiguousIterator ind_it(shape, strides, src.ndim() - 1);
+
+  strides = src.strides();
+  strides.erase(strides.begin() + axis);
+  ContiguousIterator src_it(shape, strides, src.ndim() - 1);
+
+  auto ind_ptr = ind.data<IdxT>();
+  auto src_ptr = src.data<T>();
+  auto dst_ptr = out.data<T>();
+  auto ind_ax_stride = ind.strides(axis);
+  auto src_ax_stride = src.strides(axis);
+  auto dst_ax_stride = out.strides(axis);
+  auto ind_ax_size = ind.shape(axis);
+  auto src_ax_size = src.shape(axis);
+
+  size_t size_pre = 1;
+  size_t size_post = 1;
+  for (int i = 0; i < axis; ++i) {
+    size_pre *= ind.shape(i);
+  }
+  for (int i = axis + 1; i < ind.ndim(); ++i) {
+    size_post *= ind.shape(i);
+  }
+
+  size_t stride_pre = size_post * ind_ax_size;
+  for (size_t i = 0; i < size_pre; i++) {
+    for (size_t k = 0; k < size_post; k++) {
+      for (int j = 0; j < ind_ax_size; ++j) {
+        auto ind_val = offset_neg_idx(
+            ind_ptr[ind_it.loc + j * ind_ax_stride], src_ax_size);
+        dst_ptr[k + j * dst_ax_stride] =
+            src_ptr[src_it.loc + ind_val * src_ax_stride];
+      }
+      ind_it.step();
+      src_it.step();
+    }
+    dst_ptr += stride_pre;
+  }
+}
+
+template <typename IdxT>
+void dispatch_gather_axis(
+    const array& src,
+    const array& inds,
+    array& out,
+    const int axis) {
+  switch (out.dtype()) {
+    case bool_:
+      gather_axis<bool, IdxT>(src, inds, out, axis);
+      break;
+    case uint8:
+      gather_axis<uint8_t, IdxT>(src, inds, out, axis);
+      break;
+    case uint16:
+      gather_axis<uint16_t, IdxT>(src, inds, out, axis);
+      break;
+    case uint32:
+      gather_axis<uint32_t, IdxT>(src, inds, out, axis);
+      break;
+    case uint64:
+      gather_axis<uint64_t, IdxT>(src, inds, out, axis);
+      break;
+    case int8:
+      gather_axis<int8_t, IdxT>(src, inds, out, axis);
+      break;
+    case int16:
+      gather_axis<int16_t, IdxT>(src, inds, out, axis);
+      break;
+    case int32:
+      gather_axis<int32_t, IdxT>(src, inds, out, axis);
+      break;
+    case int64:
+      gather_axis<int64_t, IdxT>(src, inds, out, axis);
+      break;
+    case float16:
+      gather_axis<float16_t, IdxT>(src, inds, out, axis);
+      break;
+    case float32:
+      gather_axis<float, IdxT>(src, inds, out, axis);
+      break;
+    case float64:
+      gather_axis<double, IdxT>(src, inds, out, axis);
+      break;
+    case bfloat16:
+      gather_axis<bfloat16_t, IdxT>(src, inds, out, axis);
+      break;
+    case complex64:
+      gather_axis<complex64_t, IdxT>(src, inds, out, axis);
+      break;
+  }
+}
+
+void GatherAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  auto& src = inputs[0];
+  auto& inds = inputs[1];
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(src);
+  encoder.set_input_array(inds);
+  encoder.set_output_array(out);
+  encoder.dispatch([axis_ = axis_,
+                    src = array::unsafe_weak_copy(src),
+                    inds = array::unsafe_weak_copy(inds),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    switch (inds.dtype()) {
+      case uint8:
+        dispatch_gather_axis<uint8_t>(src, inds, out, axis_);
+        break;
+      case uint16:
+        dispatch_gather_axis<uint16_t>(src, inds, out, axis_);
+        break;
+      case uint32:
+        dispatch_gather_axis<uint32_t>(src, inds, out, axis_);
+        break;
+      case uint64:
+        dispatch_gather_axis<uint64_t>(src, inds, out, axis_);
+        break;
+      case int8:
+        dispatch_gather_axis<int8_t>(src, inds, out, axis_);
+        break;
+      case int16:
+        dispatch_gather_axis<int16_t>(src, inds, out, axis_);
+        break;
+      case int32:
+        dispatch_gather_axis<int32_t>(src, inds, out, axis_);
+        break;
+      case int64:
+        dispatch_gather_axis<int64_t>(src, inds, out, axis_);
+        break;
+      default:
+        throw std::runtime_error(
+            "[GatherAxis::eval_cpu] Cannot gather with indices type.");
+        break;
+    }
+  });
+}
+
+template <typename InT, typename IdxT, typename OpT>
+void scatter(
+    const array& updates,
+    array& out,
+    const std::vector<array>& inds,
+    const std::vector<int>& axes) {
+  int nind = inds.size();
+  auto inds_ndim = updates.ndim() - out.ndim();
+  size_t n_updates = nind ? inds[0].size() : 1;
+
+  Shape update_shape(
+      updates.shape().begin() + inds_ndim, updates.shape().end());
+  size_t update_size = 1;
+  for (auto us : update_shape) {
+    update_size *= us;
+  }
+
+  std::vector<ContiguousIterator> its(inds.begin(), inds.end());
+  ContiguousIterator update_it(updates);
+  ContiguousIterator out_it(update_shape, out.strides(), out.ndim());
+
+  auto out_ptr = out.data<InT>();
+  auto upd_ptr = updates.data<InT>();
+  for (int i = 0; i < n_updates; ++i) {
+    size_t out_offset = 0;
+    for (int j = 0; j < inds.size(); ++j) {
+      auto ax = axes[j];
+      auto idx_loc = its[j].loc;
+      its[j].step();
+      auto idx_val =
+          offset_neg_idx(inds[j].data<IdxT>()[idx_loc], out.shape(ax));
+      out_offset += (idx_val * out.strides()[ax]);
+    }
+    update_it.seek(i * update_size);
+    for (int j = 0; j < update_size; ++j) {
+      OpT{}(upd_ptr[update_it.loc], out_ptr + out_offset + out_it.loc);
+      update_it.step();
+      out_it.step();
+    }
+    out_it.reset();
+    update_it.reset();
+  }
+}
+
+template <typename InT, typename IdxT>
+void dispatch_scatter_inds(
+    array& out,
+    const std::vector<array>& indices,
+    const array& updates,
+    const std::vector<int>& axes,
+    Scatter::ReduceType rtype) {
+  switch (rtype) {
+    case Scatter::None:
+      scatter<InT, IdxT, None>(updates, out, indices, axes);
+      break;
+    case Scatter::Sum:
+      scatter<InT, IdxT, Sum>(updates, out, indices, axes);
+      break;
+    case Scatter::Prod:
+      scatter<InT, IdxT, Prod>(updates, out, indices, axes);
+      break;
+    case Scatter::Max:
+      scatter<InT, IdxT, Max>(updates, out, indices, axes);
+      break;
+    case Scatter::Min:
+      scatter<InT, IdxT, Min>(updates, out, indices, axes);
+      break;
+  }
+}
+
+template <typename InT>
+void dispatch_scatter(
+    array& out,
+    const std::vector<array>& inds,
+    const array& updates,
+    const std::vector<int>& axes,
+    Scatter::ReduceType rtype) {
+  if (inds.empty()) {
+    dispatch_scatter_inds<InT, uint8_t>(out, inds, updates, axes, rtype);
+    return;
+  }
+
+  switch (inds[0].dtype()) {
+    case uint8:
+      dispatch_scatter_inds<InT, uint8_t>(out, inds, updates, axes, rtype);
+      break;
+    case uint16:
+      dispatch_scatter_inds<InT, uint16_t>(out, inds, updates, axes, rtype);
+      break;
+    case uint32:
+      dispatch_scatter_inds<InT, uint32_t>(out, inds, updates, axes, rtype);
+      break;
+    case uint64:
+      dispatch_scatter_inds<InT, uint64_t>(out, inds, updates, axes, rtype);
+      break;
+    case int8:
+      dispatch_scatter_inds<InT, int8_t>(out, inds, updates, axes, rtype);
+      break;
+    case int16:
+      dispatch_scatter_inds<InT, int16_t>(out, inds, updates, axes, rtype);
+      break;
+    case int32:
+      dispatch_scatter_inds<InT, int32_t>(out, inds, updates, axes, rtype);
+      break;
+    case int64:
+      dispatch_scatter_inds<InT, int64_t>(out, inds, updates, axes, rtype);
+      break;
+    default:
+      throw std::runtime_error(
+          "[Scatter::eval_cpu] Cannot scatter with indices type.");
+  }
+}
+
+void Scatter::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() >= 2);
+
+  auto& src = inputs[0];
+  auto& updates = inputs.back();
+
+  // Copy src into out (copy allocates memory for out)
+  auto ctype =
+      src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
+  copy(src, out, ctype, stream());
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  std::vector<array> inds;
+  for (auto it = inputs.begin() + 1; it < inputs.end() - 1; ++it) {
+    encoder.set_input_array(*it);
+    inds.push_back(array::unsafe_weak_copy(*it));
+  }
+  encoder.set_input_array(updates);
+  encoder.set_output_array(out);
+  encoder.dispatch([axes_ = axes_,
+                    reduce_type_ = reduce_type_,
+                    updates = array::unsafe_weak_copy(updates),
+                    inds = std::move(inds),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        dispatch_scatter<bool>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case uint8:
+        dispatch_scatter<uint8_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case uint16:
+        dispatch_scatter<uint16_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case uint32:
+        dispatch_scatter<uint32_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case uint64:
+        dispatch_scatter<uint64_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case int8:
+        dispatch_scatter<int8_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case int16:
+        dispatch_scatter<int16_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case int32:
+        dispatch_scatter<int32_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case int64:
+        dispatch_scatter<int64_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case float16:
+        dispatch_scatter<float16_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case float32:
+        dispatch_scatter<float>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case float64:
+        dispatch_scatter<double>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case bfloat16:
+        dispatch_scatter<bfloat16_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case complex64:
+        dispatch_scatter<complex64_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+    }
+  });
+}
+
+template <typename T, typename IdxT, typename OpT>
+void scatter_axis(array& out, const array idx, const array& upd, int axis) {
+  auto strides = idx.strides();
+  strides.erase(strides.begin() + axis);
+  auto shape = idx.shape();
+  shape.erase(shape.begin() + axis);
+  ContiguousIterator idx_it(shape, strides, upd.ndim() - 1);
+
+  strides = upd.strides();
+  strides.erase(strides.begin() + axis);
+  ContiguousIterator upd_it(shape, strides, upd.ndim() - 1);
+
+  auto idx_ptr = idx.data<IdxT>();
+  auto upd_ptr = upd.data<T>();
+  auto dst_ptr = out.data<T>();
+  auto idx_ax_stride = idx.strides(axis);
+  auto upd_ax_stride = upd.strides(axis);
+  auto dst_ax_stride = out.strides(axis);
+  auto idx_ax_size = idx.shape(axis);
+  auto dst_ax_size = out.shape(axis);
+
+  size_t size_pre = 1;
+  size_t size_post = 1;
+  for (int i = 0; i < axis; ++i) {
+    size_pre *= idx.shape(i);
+  }
+  for (int i = axis + 1; i < idx.ndim(); ++i) {
+    size_post *= idx.shape(i);
+  }
+  size_t stride_pre = size_post * dst_ax_size;
+  for (size_t i = 0; i < size_pre; i++) {
+    for (size_t k = 0; k < size_post; k++) {
+      for (int j = 0; j < idx_ax_size; ++j) {
+        auto ind_val = offset_neg_idx(
+            idx_ptr[idx_it.loc + j * idx_ax_stride], dst_ax_size);
+        OpT{}(
+            upd_ptr[upd_it.loc + j * upd_ax_stride],
+            dst_ptr + k + ind_val * dst_ax_stride);
+      }
+      idx_it.step();
+      upd_it.step();
+    }
+    dst_ptr += stride_pre;
+  }
+}
+
+template <typename InT, typename IdxT>
+void dispatch_scatter_axis_op(
+    array& out,
+    const array& idx,
+    const array& updates,
+    int axis,
+    ScatterAxis::ReduceType rtype) {
+  switch (rtype) {
+    case ScatterAxis::None:
+      scatter_axis<InT, IdxT, None>(out, idx, updates, axis);
+      break;
+    case ScatterAxis::Sum:
+      scatter_axis<InT, IdxT, Sum>(out, idx, updates, axis);
+      break;
+  }
+}
+
+template <typename InT>
+void dispatch_scatter_axis(
+    array& out,
+    const array& idx,
+    const array& updates,
+    int axis,
+    ScatterAxis::ReduceType rtype) {
+  switch (idx.dtype()) {
+    case uint8:
+      dispatch_scatter_axis_op<InT, uint8_t>(out, idx, updates, axis, rtype);
+      break;
+    case uint16:
+      dispatch_scatter_axis_op<InT, uint16_t>(out, idx, updates, axis, rtype);
+      break;
+    case uint32:
+      dispatch_scatter_axis_op<InT, uint32_t>(out, idx, updates, axis, rtype);
+      break;
+    case uint64:
+      dispatch_scatter_axis_op<InT, uint64_t>(out, idx, updates, axis, rtype);
+      break;
+    case int8:
+      dispatch_scatter_axis_op<InT, int8_t>(out, idx, updates, axis, rtype);
+      break;
+    case int16:
+      dispatch_scatter_axis_op<InT, int16_t>(out, idx, updates, axis, rtype);
+      break;
+    case int32:
+      dispatch_scatter_axis_op<InT, int32_t>(out, idx, updates, axis, rtype);
+      break;
+    case int64:
+      dispatch_scatter_axis_op<InT, int64_t>(out, idx, updates, axis, rtype);
+      break;
+    default:
+      throw std::runtime_error(
+          "[ScatterAxis::eval_cpu] Cannot scatter with indices type.");
+  }
+}
+
+void ScatterAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() >= 2);
+
+  auto& src = inputs[0];
+  auto& idx = inputs[1];
+  auto& updates = inputs[2];
+
+  // Copy src into out (copy allocates memory for out)
+  auto ctype =
+      src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
+  copy(src, out, ctype, stream());
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(idx);
+  encoder.set_input_array(updates);
+  encoder.set_output_array(out);
+  encoder.dispatch([axis_ = axis_,
+                    reduce_type_ = reduce_type_,
+                    idx = array::unsafe_weak_copy(idx),
+                    updates = array::unsafe_weak_copy(updates),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        dispatch_scatter_axis<bool>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case uint8:
+        dispatch_scatter_axis<uint8_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case uint16:
+        dispatch_scatter_axis<uint16_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case uint32:
+        dispatch_scatter_axis<uint32_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case uint64:
+        dispatch_scatter_axis<uint64_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case int8:
+        dispatch_scatter_axis<int8_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case int16:
+        dispatch_scatter_axis<int16_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case int32:
+        dispatch_scatter_axis<int32_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case int64:
+        dispatch_scatter_axis<int64_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case float16:
+        dispatch_scatter_axis<float16_t>(
+            out, idx, updates, axis_, reduce_type_);
+        break;
+      case float32:
+        dispatch_scatter_axis<float>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case float64:
+        dispatch_scatter_axis<double>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case bfloat16:
+        dispatch_scatter_axis<bfloat16_t>(
+            out, idx, updates, axis_, reduce_type_);
+        break;
+      case complex64:
+        dispatch_scatter_axis<complex64_t>(
+            out, idx, updates, axis_, reduce_type_);
+        break;
+    }
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cpu/inverse.cpp

@@ -0,0 +1,160 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include "mlx/allocator.h"
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/lapack.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+template <typename T>
+void general_inv(T* inv, int N) {
+  int info;
+  auto ipiv = array::Data{allocator::malloc_or_wait(sizeof(int) * N)};
+  // Compute LU factorization.
+  getrf<T>(
+      /* m = */ &N,
+      /* n = */ &N,
+      /* a = */ inv,
+      /* lda = */ &N,
+      /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
+      /* info = */ &info);
+
+  if (info != 0) {
+    std::stringstream ss;
+    ss << "[Inverse::eval_cpu] LU factorization failed with error code "
+       << info;
+    throw std::runtime_error(ss.str());
+  }
+
+  static const int lwork_query = -1;
+  T workspace_size = 0;
+
+  // Compute workspace size.
+  getri<T>(
+      /* m = */ &N,
+      /* a = */ nullptr,
+      /* lda = */ &N,
+      /* ipiv = */ nullptr,
+      /* work = */ &workspace_size,
+      /* lwork = */ &lwork_query,
+      /* info = */ &info);
+
+  if (info != 0) {
+    std::stringstream ss;
+    ss << "[Inverse::eval_cpu] LU workspace calculation failed with error code "
+       << info;
+    throw std::runtime_error(ss.str());
+  }
+
+  const int lwork = workspace_size;
+  auto scratch = array::Data{allocator::malloc_or_wait(sizeof(T) * lwork)};
+
+  // Compute inverse.
+  getri<T>(
+      /* m = */ &N,
+      /* a = */ inv,
+      /* lda = */ &N,
+      /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
+      /* work = */ static_cast<T*>(scratch.buffer.raw_ptr()),
+      /* lwork = */ &lwork,
+      /* info = */ &info);
+
+  if (info != 0) {
+    std::stringstream ss;
+    ss << "[Inverse::eval_cpu] inversion failed with error code " << info;
+    throw std::runtime_error(ss.str());
+  }
+}
+
+template <typename T>
+void tri_inv(T* inv, int N, bool upper) {
+  const char uplo = upper ? 'L' : 'U';
+  const char diag = 'N';
+  int info;
+  trtri<T>(
+      /* uplo = */ &uplo,
+      /* diag = */ &diag,
+      /* N = */ &N,
+      /* a = */ inv,
+      /* lda = */ &N,
+      /* info = */ &info);
+
+  // zero out the other triangle
+  if (upper) {
+    for (int i = 0; i < N; i++) {
+      std::fill(inv, inv + i, 0.0f);
+      inv += N;
+    }
+  } else {
+    for (int i = 0; i < N; i++) {
+      std::fill(inv + i + 1, inv + N, 0.0f);
+      inv += N;
+    }
+  }
+
+  if (info != 0) {
+    std::stringstream ss;
+    ss << "[Inverse::eval_cpu] triangular inversion failed with error code "
+       << info;
+    throw std::runtime_error(ss.str());
+  }
+}
+
+template <typename T>
+void inverse_impl(
+    const array& a,
+    array& inv,
+    bool tri,
+    bool upper,
+    Stream stream) {
+  // Lapack uses the column-major convention. We take advantage of the following
+  // identity to avoid transposing (see
+  // https://math.stackexchange.com/a/340234):
+  //   (A⁻¹)ᵀ = (Aᵀ)⁻¹
+
+  // The inverse is computed in place, so just copy the input to the output.
+  copy(
+      a,
+      inv,
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+      stream);
+
+  const int N = a.shape(-1);
+  const size_t num_matrices = a.size() / (N * N);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(inv);
+
+  auto inv_ptr = inv.data<T>();
+  if (tri) {
+    encoder.dispatch([inv_ptr, N, num_matrices, upper]() {
+      for (int i = 0; i < num_matrices; i++) {
+        tri_inv<T>(inv_ptr + N * N * i, N, upper);
+      }
+    });
+  } else {
+    encoder.dispatch([inv_ptr, N, num_matrices]() {
+      for (int i = 0; i < num_matrices; i++) {
+        general_inv<T>(inv_ptr + N * N * i, N);
+      }
+    });
+  }
+}
+
+void Inverse::eval_cpu(const std::vector<array>& inputs, array& output) {
+  switch (inputs[0].dtype()) {
+    case float32:
+      inverse_impl<float>(inputs[0], output, tri_, upper_, stream());
+      break;
+    case float64:
+      inverse_impl<double>(inputs[0], output, tri_, upper_, stream());
+      break;
+    default:
+      throw std::runtime_error(
+          "[Inverse::eval_cpu] only supports float32 or float64.");
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/jit_compiler.cpp

@@ -1,6 +1,6 @@
 // Copyright © 2024 Apple Inc.
 
-#include "mlx/backend/common/jit_compiler.h"
+#include "mlx/backend/cpu/jit_compiler.h"
 
 #include <sstream>
 #include <vector>
@@ -54,7 +54,7 @@ struct VisualStudioInfo {
       std::string value = line.substr(pos + 1);
       if (name == "LIB") {
         libpaths = str_split(value, ';');
-      } else if (name == "VCToolsInstallDir") {
+      } else if (name == "VCToolsInstallDir" || name == "VCTOOLSINSTALLDIR") {
         cl_exe = fmt::format("{0}\\bin\\Host{1}\\{1}\\cl.exe", value, arch);
       }
     }
@@ -96,7 +96,7 @@ std::string JitCompiler::build_command(
       libpaths);
 #else
   return fmt::format(
-      "g++ -std=c++17 -O3 -Wall -fPIC -shared '{0}' -o '{1}' 2>&1",
+      "g++ -std=c++17 -O3 -Wall -fPIC -shared \"{0}\" -o \"{1}\" 2>&1",
       (dir / source_file_name).string(),
       (dir / shared_lib_name).string());
 #endif
@@ -133,7 +133,7 @@ std::string JitCompiler::exec(const std::string& cmd) {
   if (status == -1) {
     throw std::runtime_error("pclose() failed.");
   }
-#ifdef _MSC_VER
+#if defined(_WIN32) || defined(__FreeBSD__)
   int code = status;
 #else
   int code = WEXITSTATUS(status);

mlx/backend/cpu/lapack.h

@@ -11,7 +11,7 @@
 #define lapack_complex_double std::complex<double>
 #endif
 
-#ifdef ACCELERATE_NEW_LAPACK
+#ifdef MLX_USE_ACCELERATE
 #include <Accelerate/Accelerate.h>
 #else
 #include <cblas.h>
@@ -31,3 +31,22 @@
 #define MLX_LAPACK_FUNC(f) f##_
 
 #endif
+
+#define INSTANTIATE_LAPACK_TYPES(FUNC)                       \
+  template <typename T, typename... Args>                    \
+  void FUNC(Args... args) {                                  \
+    if constexpr (std::is_same_v<T, float>) {                \
+      MLX_LAPACK_FUNC(s##FUNC)(std::forward<Args>(args)...); \
+    } else if constexpr (std::is_same_v<T, double>) {        \
+      MLX_LAPACK_FUNC(d##FUNC)(std::forward<Args>(args)...); \
+    }                                                        \
+  }
+
+INSTANTIATE_LAPACK_TYPES(geqrf)
+INSTANTIATE_LAPACK_TYPES(orgqr)
+INSTANTIATE_LAPACK_TYPES(syevd)
+INSTANTIATE_LAPACK_TYPES(potrf)
+INSTANTIATE_LAPACK_TYPES(gesvdx)
+INSTANTIATE_LAPACK_TYPES(getrf)
+INSTANTIATE_LAPACK_TYPES(getri)
+INSTANTIATE_LAPACK_TYPES(trtri)