Fix a couple of slicing bugs (#1827)

* fix a few bugs

* fix conv grad

* speedup test

* comment

mlx/backend/common/slicing.cpp

@@ -35,4 +35,29 @@ void shared_buffer_slice(
   move_or_copy(in, out, 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];
+    }
+  }
+  // data_end can be -1
+  size_t data_size =
+      data_end < 0 ? (data_offset - data_end) : (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,
+    array& out,
-    size_t data_offset,
+    const Shape& start_indices,
-    size_t data_size,
+    const Shape& strides);
-    array& out);
 
 } // namespace mlx::core

mlx/backend/cpu/primitives.cpp

@@ -86,7 +86,7 @@ void NumberOfElements::eval_cpu(const std::vector<array>& inputs, array& out) {
   eval(inputs, out);
 }
 void Slice::eval_cpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
+  slice(inputs[0], out, start_indices_, strides_);
 }
 void Split::eval_cpu(
     const std::vector<array>& inputs,
@@ -262,29 +262,6 @@ void Reshape::eval_cpu(const std::vector<array>& inputs, array& out) {
   reshape(inputs[0], out);
 }
 
-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);
@@ -355,7 +332,8 @@ void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
   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_);
+  auto [data_offset, out_strides] =
+      prepare_slice(out, start_indices_, strides_);
 
   // Do copy
   copy_inplace(

mlx/backend/metal/kernels/unary.metal

@@ -11,7 +11,7 @@
   instantiate_kernel(                                                                 \
       "gn1_" #op #in_tname #out_tname, unary_g, in_type, out_type, op, 1, int)        \
   instantiate_kernel(                                                                 \
-      "gn4large" #op #in_tname #out_tname, unary_g, in_type, out_type, op, 4)
+      "gn4large_" #op #in_tname #out_tname, unary_g, in_type, out_type, op, 4)
 
 #define instantiate_unary_all_same(op, tname, type)   \
   instantiate_unary_all(op, tname, tname, type, type)

mlx/backend/metal/primitives.cpp

@@ -499,8 +499,8 @@ void SliceUpdate::eval_gpu(const std::vector<array>& inputs, array& out) {
       ? CopyType::Vector
       : CopyType::General;
   copy_gpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
-
+  auto [data_offset, out_strides] =
-  auto [data_offset, out_strides] = prepare_slice(in, start_indices_, strides_);
+      prepare_slice(out, start_indices_, strides_);
 
   // Do copy
   copy_gpu_inplace(

mlx/backend/metal/slicing.cpp

@@ -14,18 +14,7 @@ void slice_gpu(
     const Shape& start_indices,
     const Shape& strides,
     const Stream& s) {
-  // Calculate out strides and initial offset
+  slice(in, out, start_indices, strides);
-  auto [data_offset, inp_strides] = prepare_slice(in, start_indices, strides);
-
-  size_t data_end = 1;
-  for (int i = 0; i < strides.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;
-  shared_buffer_slice(in, inp_strides, data_offset, data_size, out);
 }
 
 void concatenate_gpu(

mlx/backend/metal/unary.cpp

@@ -1,5 +1,4 @@
 // Copyright © 2024 Apple Inc.
-
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
@@ -49,7 +48,7 @@ void unary_op_gpu_inplace(
   } else {
     kernel_name = "gn" + std::to_string(work_per_thread);
     if (large) {
-      kernel_name += "_large";
+      kernel_name += "large";
     }
   }
   concatenate(kernel_name, "_", op, type_to_name(in), type_to_name(out));

mlx/ops.cpp

@@ -599,7 +599,13 @@ array expand_dims(
 namespace {
 
 inline auto
-normalize_slice(const Shape& shape, Shape& start, Shape& stop, Shape& strides) {
+normalize_slice(const Shape& shape, Shape& start, Shape stop, Shape& strides) {
+  // - Start indices are normalized
+  // - End indices are unchanged as -1 means something different
+  //   pre-normalization (the end of the axis) versus post normalization (the
+  //   position left of 0).
+  // - Any strides corresponding to singleton dimension are set to 1
+
   Shape out_shape(shape.size());
   bool has_neg_strides = false;
 
@@ -624,10 +630,10 @@ normalize_slice(const Shape& shape, Shape& start, Shape& stop, Shape& strides) {
       auto ed = e > -1 ? e : -1;
 
       start[i] = st;
-      stop[i] = ed > st ? st : ed;
+      ed = ed > st ? st : ed;
 
       auto str = -strides[i];
-      out_shape[i] = (start[i] - stop[i] + str - 1) / str;
+      out_shape[i] = (start[i] - ed + str - 1) / str;
 
     } else {
       // Clamp to bounds
@@ -635,9 +641,9 @@ normalize_slice(const Shape& shape, Shape& start, Shape& stop, Shape& strides) {
       auto ed = std::max(static_cast<ShapeElem>(0), std::min(e, n));
 
       start[i] = st;
-      stop[i] = ed < st ? st : ed;
+      ed = ed < st ? st : ed;
 
-      out_shape[i] = (stop[i] - start[i] + strides[i] - 1) / strides[i];
+      out_shape[i] = (ed - start[i] + strides[i] - 1) / strides[i];
     }
     // Simplify the stride if it's unused
     if (out_shape[i] == 1) {

mlx/primitives.cpp

@@ -1228,33 +1228,6 @@ std::vector<array> Convolution::vjp(
         auto grad = conv_weight_backward_patches(
             in, wt, cotan, kernel_strides_, padding_, stream());
         grads.push_back(grad);
-      } else {
-        if (flip_) {
-          auto padding = padding_;
-          for (int i = 0; i < padding.size(); i++) {
-            int wt_size = 1 + kernel_dilation_[i] * (wt.shape(1 + i) - 1);
-            padding[i] = wt_size - padding_[i] - 1;
-          }
-
-          auto cotan_trans = group_transpose(cotan, -1, 0, -1);
-          auto in_trans = swapaxes(in, 0, -1, stream());
-
-          auto grad_trans = conv_general(
-              /* const array& input = */ cotan_trans,
-              /* const array& weight = */ in_trans,
-              /* std::vector<int> stride = */ kernel_dilation_,
-              /* std::vector<int> padding_lo = */ padding,
-              /* std::vector<int> padding_hi = */ padding,
-              /* std::vector<int> kernel_dilation = */ input_dilation_,
-              /* std::vector<int> input_dilation = */ kernel_strides_,
-              /* int groups = */ groups_,
-              /* bool flip = */ false,
-              stream());
-          if (groups_ > 1) {
-            grads.push_back(group_transpose(grad_trans, -1, 0, -2));
-          } else {
-            grads.push_back(grad_trans);
-          }
       } else {
         std::vector<int> padding_lo = padding_;
         std::vector<int> padding_hi = padding_;
@@ -1279,9 +1252,23 @@ std::vector<array> Convolution::vjp(
             /* int groups = */ groups_,
             /* bool flip = */ false,
             stream());
-          grads.push_back(swapaxes(grad_trans, 0, -1, stream()));
+        if (flip_) {
+          auto start = Shape(grad_trans.ndim(), 0);
+          auto stop = Shape(grad_trans.ndim(), 0);
+          auto strides = Shape(grad_trans.ndim(), 1);
+          for (int i = 0; i < stop.size(); ++i) {
+            if (i >= 1 && i < stop.size() - 1) {
+              start[i] = grad_trans.shape(i);
+              stop[i] = -start[i] - 1;
+              strides[i] = -1;
+            } else {
+              stop[i] = grad_trans.shape(i);
             }
           }
+          grad_trans = slice(grad_trans, start, stop, strides, stream());
+        }
+        grads.push_back(swapaxes(grad_trans, 0, -1, stream()));
+      }
     }
   }
 

mlx/primitives.h

@@ -1921,7 +1921,6 @@ class Slice : public UnaryPrimitive {
   Shape start_indices_;
   Shape end_indices_;
   Shape strides_;
-  void eval(const std::vector<array>& inputs, array& out);
 };
 
 class SliceUpdate : public UnaryPrimitive {

python/tests/test_autograd.py

@@ -700,6 +700,43 @@ class TestAutograd(mlx_tests.MLXTestCase):
         expected = mx.array([0.0, 2.0, 2.0, 0.0, 2.0])[:, None]
         self.assertTrue(mx.allclose(expected, jout))
 
+    def test_slice_grads(self):
+        # Slice
+        def fun(a):
+            return a[5:-6:-1]
+
+        a = mx.ones(shape=(5,))
+        cotan = mx.random.uniform(shape=(5,))
+        _, (grad,) = mx.vjp(fun, (a,), (cotan,))
+        self.assertTrue(mx.allclose(grad, cotan[::-1]))
+
+        tan = mx.random.uniform(shape=(5,))
+        mx.eval(tan)
+        _, (grad,) = mx.jvp(fun, (a,), (tan,))
+        self.assertTrue(mx.allclose(grad, tan[::-1]))
+
+        # Slice update
+        def fun(a, b):
+            a[4:-5:-2] = b
+            return a
+
+        a = mx.ones(shape=(4,))
+        b = mx.zeros(shape=(2,))
+
+        cotan = mx.random.uniform(shape=(4,))
+        _, (grad_a, grad_b) = mx.vjp(fun, (a, b), (cotan,))
+        expected_a = mx.array(cotan)
+        expected_a[1::2] = 0.0
+        self.assertTrue(mx.allclose(grad_a, expected_a))
+        self.assertTrue(mx.allclose(grad_b, cotan[4:-5:-2]))
+
+        tan_a = mx.random.uniform(shape=(4,))
+        tan_b = mx.random.uniform(shape=(2,))
+        _, (grad,) = mx.jvp(fun, (a, b), (tan_a, tan_b))
+        expected = tan_a
+        expected[4:-5:-2] = tan_b
+        self.assertTrue(mx.allclose(grad, expected))
+
 
 if __name__ == "__main__":
     unittest.main()

python/tests/test_conv.py

@@ -911,6 +911,44 @@ class TestConv(mlx_tests.MLXTestCase):
             expected = mx.array([[dw00, dw01], [dw10, dw11]])
             self.assertTrue(mx.allclose(dw, expected, rtol=1e-5, atol=1e-5))
 
+        # Test with input dilation
+        inputs = mx.random.normal((1, 14, 14, 2))
+        kernel = mx.random.normal((2, 7, 7, 2))
+
+        def conv_flip(kernel):
+            return mx.conv_general(
+                inputs,
+                kernel,
+                stride=1,
+                padding=([6, 6], [15, 15]),
+                kernel_dilation=(1, 1),
+                input_dilation=(16, 16),
+                groups=1,
+                flip=True,
+            ).sum()
+
+        def reverse_sequence(xs, axis=0):
+            indices = mx.arange(xs.shape[axis] - 1, -1, -1)
+            return mx.take(xs, indices, axis=axis)
+
+        def conv_manual_flip(kernel):
+            for ax in range(1, kernel.ndim - 1):
+                kernel = reverse_sequence(kernel, axis=ax)
+            return mx.conv_general(
+                inputs,
+                kernel,
+                stride=1,
+                padding=([6, 6], [15, 15]),
+                kernel_dilation=(1, 1),
+                input_dilation=(16, 16),
+                groups=1,
+                flip=False,
+            ).sum()
+
+        grad = mx.grad(conv_flip)(kernel)
+        expected_grad = mx.grad(conv_manual_flip)(kernel)
+        self.assertTrue(mx.allclose(grad, expected_grad))
+
     def test_conv_groups_grad(self):
         def fn(x, w):
             num_groups = x.shape[-1] // w.shape[-1]

python/tests/test_conv_transpose.py

@@ -587,10 +587,10 @@ class TestConvTranspose(mlx_tests.MLXTestCase):
                 for idim, kdim, stride, padding, dilation in (
                     ((1, 1, 1), (1, 1, 1), (1, 1, 1), (0, 0, 0), (1, 1, 1)),
                     ((3, 3, 3), (3, 1, 1), (1, 1, 1), (0, 0, 0), (1, 1, 1)),
-                    ((15, 15, 15), (5, 5, 5), (5, 5, 5), (2, 2, 2), (1, 1, 1)),
+                    ((7, 7, 7), (5, 5, 5), (5, 5, 5), (2, 2, 2), (1, 1, 1)),
-                    ((16, 16, 16), (3, 3, 3), (2, 2, 2), (1, 1, 1), (1, 1, 1)),
+                    ((8, 8, 8), (3, 3, 3), (2, 2, 2), (1, 1, 1), (1, 1, 1)),
-                    ((15, 15, 15), (5, 5, 5), (5, 5, 5), (2, 2, 2), (3, 2, 2)),
+                    ((7, 7, 7), (5, 5, 5), (3, 3, 3), (2, 2, 2), (3, 2, 2)),
-                    ((16, 16, 16), (3, 3, 3), (2, 2, 2), (1, 1, 1), (3, 2, 2)),
+                    ((8, 8, 8), (3, 3, 3), (2, 2, 2), (1, 1, 1), (3, 2, 2)),
                 ):
                     run_conv_transpose3D_grad(
                         N, C, O, idim, kdim, stride, padding, dilation, dtype=dtype

python/tests/test_ops.py

@@ -2816,6 +2816,12 @@ class TestOps(mlx_tests.MLXTestCase):
         self.assertEqual(a.shape, (3, 4, 2))
         self.assertEqual(b.shape, (3, 4, 2))
 
+    def test_slice_update_reversed(self):
+        a = mx.array([1, 2, 3, 4])
+        b = a[::-1]
+        b[::2] = 0
+        self.assertTrue(mx.array_equal(b, mx.array([0, 3, 0, 1])))
+
 
 if __name__ == "__main__":
     unittest.main()