zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-12-16 01:49:05 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

MLX_SWITCH macros to templates (#2320)

Browse Source
This commit is contained in:
Angelos Katharopoulos
2025-07-01 01:33:44 -07:00
committed by GitHub
parent 33bf1a244b
commit 3d5e17e507
27 changed files with 693 additions and 692 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
mlx/backend/cuda/arg_reduce.cu
View File

@@ -152,35 +152,29 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
encoder.set_input_array(in);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_REAL_TYPES_CHECKED(in.dtype(), "ArgReduce", CTYPE, {
using InType = cuda_type_t<CTYPE>;
dispatch_real_types(in.dtype(), "ArgReduce", [&](auto type_tag) {
using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
constexpr uint32_t N_READS = 4;
MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
dim3 num_blocks = get_2d_grid_dims(out.shape(), out.strides());
dim3 block_dims{BLOCK_DIM, 1, 1};
auto kernel = &cu::arg_reduce_general<
InType,
cu::ArgMax<InType>,
BLOCK_DIM,
N_READS>;
if (reduce_type_ == ArgReduce::ArgMin) {
kernel = &cu::arg_reduce_general<
InType,
cu::ArgMin<InType>,
BLOCK_DIM,
N_READS>;
}
kernel<<<num_blocks, block_dims, 0, stream>>>(
in.data<InType>(),
out.data<uint32_t>(),
out.size(),
const_param(shape),
const_param(in_strides),
const_param(out_strides),
ndim,
axis_stride,
axis_size);
});
dispatch_block_dim(
cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
dim3 num_blocks = get_2d_grid_dims(out.shape(), out.strides());
auto kernel =
cu::arg_reduce_general<T, cu::ArgMax<T>, block_dim(), N_READS>;
if (reduce_type_ == ArgReduce::ArgMin) {
kernel = cu::
arg_reduce_general<T, cu::ArgMin<T>, block_dim(), N_READS>;
}
kernel<<<num_blocks, block_dim(), 0, stream>>>(
in.data<T>(),
out.data<uint32_t>(),
out.size(),
const_param(shape),
const_param(in_strides),
const_param(out_strides),
ndim,
axis_stride,
axis_size);
});
});
});
}

94
mlx/backend/cuda/binary.cu
View File

@@ -140,54 +140,64 @@ void binary_op_gpu_inplace(
encoder.set_input_array(b);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(a.dtype(), CTYPE_IN, {
MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE_OUT, {
dispatch_all_types(a.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
using InType = cuda_type_t<CTYPE_IN>;
using OutType = cuda_type_t<CTYPE_OUT>;
auto bopt = get_binary_op_type(a, b);
if (bopt == BinaryOpType::General) {
auto [shape, strides] = collapse_contiguous_dims(a, b, out);
auto& a_strides = strides[0];
auto& b_strides = strides[1];
bool large = a.data_size() > INT32_MAX ||
b.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
MLX_SWITCH_BOOL(large, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
int ndim = shape.size();
if (ndim <= 3) {
MLX_SWITCH_1_2_3(ndim, NDIM, {
auto kernel =
&cu::binary_g_nd<Op, InType, OutType, IdxT, NDIM>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),
out.data<OutType>(),
out.size(),
const_param<NDIM>(shape),
const_param<NDIM>(a_strides),
const_param<NDIM>(b_strides));
dispatch_bool(
a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
out.data_size() > INT32_MAX,
[&](auto large) {
using IdxT = std::conditional_t<large(), int64_t, int32_t>;
Shape shape;
std::vector<Strides> strides;
std::tie(shape, strides) =
collapse_contiguous_dims(a, b, out);
auto& a_strides = strides[0];
auto& b_strides = strides[1];
int ndim = shape.size();
if (ndim <= 3) {
dispatch_1_2_3(ndim, [&](auto dims_constant) {
auto kernel = cu::binary_g_nd<
Op,
InType,
OutType,
IdxT,
dims_constant()>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),
out.data<OutType>(),
out.size(),
const_param<dims_constant()>(shape),
const_param<dims_constant()>(a_strides),
const_param<dims_constant()>(b_strides));
});
} else {
auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),
out.data<OutType>(),
out.size(),
const_param(shape),
const_param(a_strides),
const_param(b_strides),
ndim);
}
});
} else {
auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),
out.data<OutType>(),
out.size(),
const_param(shape),
const_param(a_strides),
const_param(b_strides),
ndim);
}
});
} else {
MLX_SWITCH_BOOL(out.data_size() > UINT32_MAX, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
auto kernel = cu::binary_ss<Op, InType, OutType, IdxT>;
if (bopt == BinaryOpType::ScalarVector) {
kernel = cu::binary_sv<Op, InType, OutType, IdxT>;
@@ -197,7 +207,7 @@ void binary_op_gpu_inplace(
kernel = cu::binary_vv<Op, InType, OutType, IdxT>;
}
auto [num_blocks, block_dims] = get_launch_args(
kernel, out.data_size(), out.shape(), out.strides(), LARGE);
kernel, out.data_size(), out.shape(), out.strides(), large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),

98
mlx/backend/cuda/binary_two.cu
View File

@@ -138,57 +138,67 @@ void binary_op_gpu_inplace(
encoder.set_output_array(out_a);
encoder.set_output_array(out_b);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(a.dtype(), CTYPE_IN, {
MLX_SWITCH_ALL_TYPES(out_a.dtype(), CTYPE_OUT, {
dispatch_all_types(a.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out_a.dtype(), [&](auto out_type_tag) {
using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
using InType = cuda_type_t<CTYPE_IN>;
using OutType = cuda_type_t<CTYPE_OUT>;
auto bopt = get_binary_op_type(a, b);
if (bopt == BinaryOpType::General) {
auto [shape, strides] = collapse_contiguous_dims(a, b, out_a);
auto& a_strides = strides[0];
auto& b_strides = strides[1];
bool large = a.data_size() > INT32_MAX ||
b.data_size() > INT32_MAX || out_a.data_size() > INT32_MAX;
MLX_SWITCH_BOOL(large, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
int ndim = shape.size();
if (ndim <= 3) {
MLX_SWITCH_1_2_3(ndim, NDIM, {
auto kernel =
cu::binary_g_nd<Op, InType, OutType, IdxT, NDIM>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out_a, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),
out_a.data<OutType>(),
out_b.data<OutType>(),
out_a.size(),
const_param<NDIM>(shape),
const_param<NDIM>(a_strides),
const_param<NDIM>(b_strides));
dispatch_bool(
a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
out_a.data_size() > INT32_MAX,
[&](auto large) {
using IdxT = std::conditional_t<large(), int64_t, int32_t>;
Shape shape;
std::vector<Strides> strides;
std::tie(shape, strides) =
collapse_contiguous_dims(a, b, out_a);
auto& a_strides = strides[0];
auto& b_strides = strides[1];
int ndim = shape.size();
if (ndim <= 3) {
dispatch_1_2_3(ndim, [&](auto dims_constant) {
auto kernel = cu::binary_g_nd<
Op,
InType,
OutType,
IdxT,
dims_constant()>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out_a, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),
out_a.data<OutType>(),
out_b.data<OutType>(),
out_a.size(),
const_param<dims_constant()>(shape),
const_param<dims_constant()>(a_strides),
const_param<dims_constant()>(b_strides));
});
} else {
auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out_a, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),
out_a.data<OutType>(),
out_b.data<OutType>(),
out_a.size(),
const_param(shape),
const_param(a_strides),
const_param(b_strides),
ndim);
}
});
} else {
auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out_a, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),
out_a.data<OutType>(),
out_b.data<OutType>(),
out_a.size(),
const_param(shape),
const_param(a_strides),
const_param(b_strides),
ndim);
}
});
} else {
MLX_SWITCH_BOOL(out_a.data_size() > UINT32_MAX, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
dispatch_bool(out_a.data_size() > INT32_MAX, [&](auto large) {
using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
auto kernel = cu::binary_ss<Op, InType, OutType, IdxT>;
if (bopt == BinaryOpType::ScalarVector) {
kernel = cu::binary_sv<Op, InType, OutType, IdxT>;
@@ -202,7 +212,7 @@ void binary_op_gpu_inplace(
out_a.data_size(),
out_a.shape(),
out_a.strides(),
LARGE);
large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<InType>(),
b.data<InType>(),

9
mlx/backend/cuda/copy/copy.cuh
View File

@@ -10,15 +10,6 @@
namespace mlx::core {
#define MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, ...) \
MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE_IN, { \
MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE_OUT, { \
using InType = cuda_type_t<CTYPE_IN>; \
using OutType = cuda_type_t<CTYPE_OUT>; \
__VA_ARGS__; \
}); \
})
void copy_contiguous(
cu::CommandEncoder& encoder,
CopyType ctype,

30
mlx/backend/cuda/copy/copy_contiguous.cu
View File

@@ -36,19 +36,23 @@ void copy_contiguous(
int64_t in_offset,
int64_t out_offset) {
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
MLX_SWITCH_BOOL(out.data_size() > UINT32_MAX, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
auto kernel = cu::copy_s<InType, OutType, IdxT>;
if (ctype == CopyType::Vector) {
kernel = cu::copy_v<InType, OutType, IdxT>;
}
auto [num_blocks, block_dims] = get_launch_args(
kernel, out.data_size(), out.shape(), out.strides(), LARGE);
kernel<<<num_blocks, block_dims, 0, stream>>>(
in.data<InType>() + in_offset,
out.data<OutType>() + out_offset,
out.data_size());
dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
auto kernel = cu::copy_s<InType, OutType, IdxT>;
if (ctype == CopyType::Vector) {
kernel = cu::copy_v<InType, OutType, IdxT>;
}
auto [num_blocks, block_dims] = get_launch_args(
kernel, out.data_size(), out.shape(), out.strides(), large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
in.data<InType>() + in_offset,
out.data<OutType>() + out_offset,
out.data_size());
});
});
});
});

78
mlx/backend/cuda/copy/copy_general.cu
View File

@@ -56,42 +56,48 @@ void copy_general(
const Strides& strides_in,
const Strides& strides_out) {
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
const InType* in_ptr = in.data<InType>() + offset_in;
OutType* out_ptr = out.data<OutType>() + offset_out;
bool large = in.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
MLX_SWITCH_BOOL(large, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
int ndim = shape.size();
size_t data_size = 1;
for (auto& s : shape)
data_size *= s;
if (ndim <= 3) {
MLX_SWITCH_1_2_3(ndim, NDIM, {
auto kernel = cu::copy_gg_nd<InType, OutType, IdxT, NDIM>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, data_size, shape, out.strides(), large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
data_size,
const_param<NDIM>(shape),
const_param<NDIM>(strides_in),
const_param<NDIM>(strides_out));
});
} else { // ndim >= 4
auto kernel = cu::copy_gg<InType, OutType, IdxT>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, data_size, shape, out.strides(), large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
data_size,
const_param(shape),
const_param(strides_in),
const_param(strides_out),
ndim);
}
dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
dispatch_bool(
in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
[&](auto large) {
using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
using IdxT = std::conditional_t<large(), int64_t, int32_t>;
const InType* in_ptr = in.data<InType>() + offset_in;
OutType* out_ptr = out.data<OutType>() + offset_out;
int ndim = shape.size();
size_t data_size = 1;
for (auto& s : shape)
data_size *= s;
if (ndim <= 3) {
dispatch_1_2_3(ndim, [&](auto ndim_constant) {
auto kernel =
cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant()>;
auto [num_blocks, block_dims] = get_launch_args(
kernel, data_size, shape, out.strides(), large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
data_size,
const_param<ndim_constant()>(shape),
const_param<ndim_constant()>(strides_in),
const_param<ndim_constant()>(strides_out));
});
} else { // ndim >= 4
auto kernel = cu::copy_gg<InType, OutType, IdxT>;
auto [num_blocks, block_dims] = get_launch_args(
kernel, data_size, shape, out.strides(), large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
data_size,
const_param(shape),
const_param(strides_in),
const_param(strides_out),
ndim);
}
});
});
});
});

81
mlx/backend/cuda/copy/copy_general_dynamic.cu
View File

@@ -62,41 +62,52 @@ void copy_general_dynamic(
const array& dynamic_offset_in,
const array& dynamic_offset_out) {
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
const InType* in_ptr = in.data<InType>() + offset_in;
OutType* out_ptr = out.data<OutType>() + offset_out;
bool large = in.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
MLX_SWITCH_BOOL(large, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
int ndim = shape.size();
if (ndim <= 3) {
MLX_SWITCH_1_2_3(ndim, NDIM, {
auto kernel = cu::copy_gg_dynamic_nd<InType, OutType, IdxT, NDIM>;
auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
out.size(),
const_param<NDIM>(shape),
const_param<NDIM>(strides_in),
const_param<NDIM>(strides_out),
dynamic_offset_in.data<int64_t>(),
dynamic_offset_out.data<int64_t>());
});
} else { // ndim >= 4
auto kernel = cu::copy_gg_dynamic<InType, OutType, IdxT>;
auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
out.size(),
const_param(shape),
const_param(strides_in),
const_param(strides_out),
ndim,
dynamic_offset_in.data<int64_t>(),
dynamic_offset_out.data<int64_t>());
}
dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
dispatch_bool(
in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
[&](auto large) {
using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
using IdxT = std::conditional_t<large(), int64_t, int32_t>;
const InType* in_ptr = in.data<InType>() + offset_in;
OutType* out_ptr = out.data<OutType>() + offset_out;
int ndim = shape.size();
if (ndim <= 3) {
dispatch_1_2_3(ndim, [&](auto dims_constant) {
auto kernel = cu::copy_gg_dynamic_nd<
InType,
OutType,
IdxT,
dims_constant()>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
out.size(),
const_param<dims_constant()>(shape),
const_param<dims_constant()>(strides_in),
const_param<dims_constant()>(strides_out),
dynamic_offset_in.data<int64_t>(),
dynamic_offset_out.data<int64_t>());
});
} else { // ndim >= 4
auto kernel = cu::copy_gg_dynamic<InType, OutType, IdxT>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
out.size(),
const_param(shape),
const_param(strides_in),
const_param(strides_out),
ndim,
dynamic_offset_in.data<int64_t>(),
dynamic_offset_out.data<int64_t>());
}
});
});
});
});

66
mlx/backend/cuda/copy/copy_general_input.cu
View File

@@ -51,35 +51,43 @@ void copy_general_input(
const Shape& shape,
const Strides& strides_in) {
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
const InType* in_ptr = in.data<InType>() + offset_in;
OutType* out_ptr = out.data<OutType>() + offset_out;
bool large = in.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
MLX_SWITCH_BOOL(large, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
int ndim = shape.size();
if (ndim <= 3) {
MLX_SWITCH_1_2_3(ndim, NDIM, {
auto kernel = cu::copy_g_nd<InType, OutType, IdxT, NDIM>;
auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
out.size(),
const_param<NDIM>(shape),
const_param<NDIM>(strides_in));
});
} else { // ndim >= 4
auto kernel = cu::copy_g<InType, OutType, IdxT>;
auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
out.size(),
const_param(shape),
const_param(strides_in),
ndim);
}
dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
dispatch_bool(
in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
[&](auto large) {
using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
using IdxT = std::conditional_t<large(), int64_t, int32_t>;
const InType* in_ptr = in.data<InType>() + offset_in;
OutType* out_ptr = out.data<OutType>() + offset_out;
int ndim = shape.size();
if (ndim <= 3) {
dispatch_1_2_3(ndim, [&](auto dims_constant) {
auto kernel =
cu::copy_g_nd<InType, OutType, IdxT, dims_constant()>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
out.size(),
const_param<dims_constant()>(shape),
const_param<dims_constant()>(strides_in));
});
} else { // ndim >= 4
auto kernel = cu::copy_g<InType, OutType, IdxT>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
in_ptr,
out_ptr,
out.size(),
const_param(shape),
const_param(strides_in),
ndim);
}
});
});
});
});

86
mlx/backend/cuda/kernel_utils.cuh
View File

@@ -6,6 +6,8 @@
#pragma once
#include <type_traits>
#include "mlx/array.h"
#include "mlx/backend/cuda/device/utils.cuh"
@@ -17,60 +19,46 @@
namespace mlx::core {
// Convert a number between 1~3 to constexpr.
#define MLX_SWITCH_1_2_3(N, NDIM, ...) \
switch (N) { \
case 1: { \
constexpr int NDIM = 1; \
__VA_ARGS__; \
break; \
} \
case 2: { \
constexpr int NDIM = 2; \
__VA_ARGS__; \
break; \
} \
case 3: { \
constexpr int NDIM = 3; \
__VA_ARGS__; \
break; \
} \
template <typename F>
void dispatch_1_2_3(int n, F&& f) {
switch (n) {
case 1:
f(std::integral_constant<int, 1>{});
break;
case 2:
f(std::integral_constant<int, 2>{});
break;
case 3:
f(std::integral_constant<int, 3>{});
break;
}
}
// Like MLX_SWITCH_ALL_TYPES but for booleans.
#define MLX_SWITCH_BOOL(BOOL, BOOL_ALIAS, ...) \
if (BOOL) { \
constexpr bool BOOL_ALIAS = true; \
__VA_ARGS__; \
} else { \
constexpr bool BOOL_ALIAS = false; \
__VA_ARGS__; \
template <typename F>
void dispatch_bool(bool v, F&& f) {
if (v) {
f(std::true_type{});
} else {
f(std::false_type{});
}
}
// Convert a block_dim to constexpr between WARP_SIZE and WARP_SIZE ^ 2.
#define MLX_SWITCH_BLOCK_DIM(NUM_THREADS, BLOCK_DIM, ...) \
{ \
uint32_t _num_threads = NUM_THREADS; \
if (_num_threads <= WARP_SIZE) { \
constexpr uint32_t BLOCK_DIM = WARP_SIZE; \
__VA_ARGS__; \
} else if (_num_threads <= WARP_SIZE * 2) { \
constexpr uint32_t BLOCK_DIM = WARP_SIZE * 2; \
__VA_ARGS__; \
} else if (_num_threads <= WARP_SIZE * 4) { \
constexpr uint32_t BLOCK_DIM = WARP_SIZE * 4; \
__VA_ARGS__; \
} else if (_num_threads <= WARP_SIZE * 8) { \
constexpr uint32_t BLOCK_DIM = WARP_SIZE * 8; \
__VA_ARGS__; \
} else if (_num_threads <= WARP_SIZE * 16) { \
constexpr uint32_t BLOCK_DIM = WARP_SIZE * 16; \
__VA_ARGS__; \
} else { \
constexpr uint32_t BLOCK_DIM = WARP_SIZE * WARP_SIZE; \
__VA_ARGS__; \
} \
template <typename F>
void dispatch_block_dim(int threads, F&& f) {
if (threads <= WARP_SIZE) {
f(std::integral_constant<int, WARP_SIZE>{});
} else if (threads <= WARP_SIZE * 2) {
f(std::integral_constant<int, WARP_SIZE * 2>{});
} else if (threads <= WARP_SIZE * 4) {
f(std::integral_constant<int, WARP_SIZE * 4>{});
} else if (threads <= WARP_SIZE * 8) {
f(std::integral_constant<int, WARP_SIZE * 8>{});
} else if (threads <= WARP_SIZE * 16) {
f(std::integral_constant<int, WARP_SIZE * 16>{});
} else {
f(std::integral_constant<int, WARP_SIZE * 32>{});
}
}
// Maps CPU types to CUDA types.
template <typename T>

66
mlx/backend/cuda/layer_norm.cu
View File

@@ -259,21 +259,22 @@ void LayerNorm::eval_gpu(
encoder.set_input_array(b);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_FLOAT_TYPES_CHECKED(out.dtype(), "layernorm", CTYPE, {
using DataType = cuda_type_t<CTYPE>;
dispatch_float_types(out.dtype(), "layernorm", [&](auto type_tag) {
constexpr uint32_t N_READS = 4;
MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
auto kernel = cu::layer_norm<DataType, BLOCK_DIM, N_READS>;
kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
x.data<DataType>(),
w.data<DataType>(),
b.data<DataType>(),
out.data<DataType>(),
eps_,
axis_size,
w_stride,
b_stride);
});
dispatch_block_dim(
cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
auto kernel = cu::layer_norm<DataType, block_dim(), N_READS>;
kernel<<<n_rows, block_dim(), 0, stream>>>(
x.data<DataType>(),
w.data<DataType>(),
b.data<DataType>(),
out.data<DataType>(),
eps_,
axis_size,
w_stride,
b_stride);
});
});
});
}
@@ -357,22 +358,27 @@ void LayerNormVJP::eval_gpu(
encoder.set_output_array(gx);
encoder.set_output_array(gw_temp);
encoder.launch_kernel([&, x = x, g = g](cudaStream_t stream) {
MLX_SWITCH_FLOAT_TYPES_CHECKED(gx.dtype(), "layernorm_vjp", CTYPE, {
using DataType = cuda_type_t<CTYPE>;
constexpr int N_READS = 4;
MLX_SWITCH_BOOL(has_w, HAS_W, {
MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
auto kernel = cu::layer_norm_vjp<DataType, HAS_W, BLOCK_DIM, N_READS>;
kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
x.data<DataType>(),
w.data<DataType>(),
g.data<DataType>(),
gx.data<DataType>(),
gw_temp.data<DataType>(),
eps_,
axis_size,
w_stride);
});
dispatch_float_types(gx.dtype(), "layernorm_vjp", [&](auto type_tag) {
dispatch_bool(has_w, [&](auto has_w_constant) {
constexpr int N_READS = 4;
dispatch_block_dim(
cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
auto kernel = cu::layer_norm_vjp<
DataType,
has_w_constant(),
block_dim(),
N_READS>;
kernel<<<n_rows, block_dim(), 0, stream>>>(
x.data<DataType>(),
w.data<DataType>(),
g.data<DataType>(),
gx.data<DataType>(),
gw_temp.data<DataType>(),
eps_,
axis_size,
w_stride);
});
});
});
});

15
mlx/backend/cuda/logsumexp.cu
View File

@@ -144,14 +144,15 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
encoder.set_input_array(in);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_FLOAT_TYPES_CHECKED(out.dtype(), "logsumexp", CTYPE, {
using DataType = cuda_type_t<CTYPE>;
dispatch_float_types(out.dtype(), "logsumexp", [&](auto type_tag) {
constexpr int N_READS = 4;
MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
auto kernel = cu::logsumexp<DataType, float, BLOCK_DIM, N_READS>;
kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
in.data<DataType>(), out.data<DataType>(), axis_size);
});
dispatch_block_dim(
cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
auto kernel = cu::logsumexp<DataType, float, block_dim(), N_READS>;
kernel<<<n_rows, block_dim(), 0, stream>>>(
in.data<DataType>(), out.data<DataType>(), axis_size);
});
});
});
}

3
mlx/backend/cuda/primitives.cu
View File

@@ -28,7 +28,8 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
auto& encoder = cu::get_command_encoder(s);
encoder.set_output_array(out);
encoder.launch_kernel([&, this](cudaStream_t stream) {
MLX_SWITCH_INT_FLOAT_TYPES_CHECKED(out.dtype(), "Arange", CTYPE, {
dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
using CTYPE = MLX_GET_TYPE(type_tag);
using OutType = cuda_type_t<CTYPE>;
CTYPE step =
static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);

18
mlx/backend/cuda/reduce/all_reduce.cu
View File

@@ -111,10 +111,11 @@ void all_reduce(
encoder.add_temporary(intermediate);
encoder.set_output_array(intermediate);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(dt, CTYPE, {
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
using T = cuda_type_t<CTYPE>;
using U = cu::ReduceResult<OP, T>::type;
dispatch_all_types(dt, [&](auto type_tag) {
dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
using OP = MLX_GET_TYPE(reduce_type_tag);
using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
using U = typename cu::ReduceResult<OP, T>::type;
auto kernel = cu::all_reduce<T, U, OP, N_READS>;
kernel<<<blocks, threads, 0, stream>>>(
static_cast<T*>(indata),
@@ -135,10 +136,11 @@ void all_reduce(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(dt, CTYPE, {
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
using T = cuda_type_t<CTYPE>;
using U = cu::ReduceResult<OP, T>::type;
dispatch_all_types(dt, [&](auto type_tag) {
dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
using OP = MLX_GET_TYPE(reduce_type_tag);
using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
using U = typename cu::ReduceResult<OP, T>::type;
auto kernel = cu::all_reduce<T, U, OP, N_READS>;
kernel<<<blocks, threads, 0, stream>>>(
static_cast<T*>(indata), out.data<U>(), block_step, insize);

14
mlx/backend/cuda/reduce/col_reduce.cu
View File

@@ -215,11 +215,12 @@ void col_reduce_looped(
encoder.set_input_array(in);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
MLX_SWITCH_REDUCE_NDIM(args.reduce_ndim, NDIM, {
using T = cuda_type_t<CTYPE>;
using U = cu::ReduceResult<OP, T>::type;
dispatch_all_types(in.dtype(), [&](auto type_tag) {
dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
dispatch_reduce_ndim(args.reduce_ndim, [&](auto reduce_ndim) {
using OP = MLX_GET_TYPE(reduce_type_tag);
using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
using U = typename cu::ReduceResult<OP, T>::type;
// Cub doesn't like const pointers for vectorized loads. (sigh)
T* indata = const_cast<T*>(in.data<T>());
@@ -229,7 +230,8 @@ void col_reduce_looped(
constexpr int BN = 32;
dim3 grid = output_grid_for_col_reduce(out, args, BN);
int blocks = BM * BN / N_READS;
auto kernel = cu::col_reduce_looped<T, U, OP, NDIM, BM, BN, N_READS>;
auto kernel =
cu::col_reduce_looped<T, U, OP, reduce_ndim(), BM, BN, N_READS>;
kernel<<<grid, blocks, 0, stream>>>(indata, out.data<U>(), args);
});
});

9
mlx/backend/cuda/reduce/init_reduce.cu
View File

@@ -33,10 +33,11 @@ void init_reduce(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
using T = cuda_type_t<CTYPE>;
using U = cu::ReduceResult<OP, T>::type;
dispatch_all_types(in.dtype(), [&](auto type_tag) {
dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
using OP = MLX_GET_TYPE(reduce_type_tag);
using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
using U = typename cu::ReduceResult<OP, T>::type;
auto kernel = cu::init_reduce<T, U, OP>;
dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
dim3 block(grid.x < 1024 ? grid.x : 1024, 1, 1);

62
mlx/backend/cuda/reduce/reduce.cuh
View File

@@ -1,5 +1,7 @@
// Copyright © 2025 Apple Inc.
#include <type_traits>
#include "mlx/backend/common/reduce.h"
#include "mlx/backend/cuda/device/cucomplex_math.cuh"
#include "mlx/backend/cuda/kernel_utils.cuh"
@@ -9,43 +11,35 @@
namespace mlx::core {
// Dispatch dynamic ndim to constexpr.
// The behavior follows get_kernel_reduce_ndim in metal/reduce.cpp file.
#define MLX_SWITCH_REDUCE_NDIM(ndim, NDIM, ...) \
if (ndim == 1) { \
constexpr uint32_t NDIM = 1; \
__VA_ARGS__; \
} else if (ndim == 2) { \
constexpr uint32_t NDIM = 2; \
__VA_ARGS__; \
} else { \
constexpr uint32_t NDIM = 5; \
__VA_ARGS__; \
template <typename F>
void dispatch_reduce_ndim(int ndim, F&& f) {
if (ndim == 1) {
f(std::integral_constant<int, 1>{});
} else if (ndim == 2) {
f(std::integral_constant<int, 2>{});
} else {
f(std::integral_constant<int, 5>{});
}
}
// Dispatch reduce ops to constexpr.
#define MLX_SWITCH_REDUCE_OPS(REDUCE, OP, ...) \
if (REDUCE == Reduce::ReduceType::And) { \
using OP = cu::And; \
__VA_ARGS__; \
} else if (REDUCE == Reduce::ReduceType::Or) { \
using OP = cu::Or; \
__VA_ARGS__; \
} else if (REDUCE == Reduce::ReduceType::Sum) { \
using OP = cu::Sum; \
__VA_ARGS__; \
} else if (REDUCE == Reduce::ReduceType::Prod) { \
using OP = cu::Prod; \
__VA_ARGS__; \
} else if (REDUCE == Reduce::ReduceType::Max) { \
using OP = cu::Max; \
__VA_ARGS__; \
} else if (REDUCE == Reduce::ReduceType::Min) { \
using OP = cu::Min; \
__VA_ARGS__; \
} else { \
throw std::invalid_argument("Unknown reduce type."); \
template <typename F>
void dispatch_reduce_ops(Reduce::ReduceType reduce_type, F&& f) {
if (reduce_type == Reduce::ReduceType::And) {
f(type_identity<cu::And>{});
} else if (reduce_type == Reduce::ReduceType::Or) {
f(type_identity<cu::Or>{});
} else if (reduce_type == Reduce::ReduceType::Sum) {
f(type_identity<cu::Sum>{});
} else if (reduce_type == Reduce::ReduceType::Prod) {
f(type_identity<cu::Prod>{});
} else if (reduce_type == Reduce::ReduceType::Max) {
f(type_identity<cu::Max>{});
} else if (reduce_type == Reduce::ReduceType::Min) {
f(type_identity<cu::Min>{});
} else {
throw std::invalid_argument("Unknown reduce type.");
}
}
void all_reduce(
cu::CommandEncoder& encoder,

32
mlx/backend/cuda/reduce/row_reduce.cu
View File

@@ -246,10 +246,11 @@ void row_reduce_simple(
encoder.set_input_array(in);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
using T = cuda_type_t<CTYPE>;
using U = cu::ReduceResult<OP, T>::type;
dispatch_all_types(in.dtype(), [&](auto type_tag) {
dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
using OP = MLX_GET_TYPE(reduce_type_tag);
using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
using U = typename cu::ReduceResult<OP, T>::type;
// Cub doesn't like const pointers for vectorized loads. (sigh)
T* indata = const_cast<T*>(in.data<T>());
@@ -293,10 +294,11 @@ void row_reduce_looped(
encoder.set_input_array(in);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
using T = cuda_type_t<CTYPE>;
using U = cu::ReduceResult<OP, T>::type;
dispatch_all_types(in.dtype(), [&](auto type_tag) {
dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
using OP = MLX_GET_TYPE(reduce_type_tag);
using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
using U = typename cu::ReduceResult<OP, T>::type;
// Cub doesn't like const pointers for vectorized loads. (sigh)
T* indata = const_cast<T*>(in.data<T>());
@@ -311,10 +313,16 @@ void row_reduce_looped(
// Pick the kernel
auto kernel = cu::row_reduce_looped<T, U, OP, 1, 32, N_READS>;
MLX_SWITCH_REDUCE_NDIM(args.reduce_ndim, NDIM, {
MLX_SWITCH_BLOCK_DIM(threads, THREADS, {
kernel = cu::row_reduce_looped<T, U, OP, NDIM, THREADS, N_READS>;
block.x = THREADS;
dispatch_reduce_ndim(args.reduce_ndim, [&](auto reduce_ndim) {
dispatch_block_dim(threads, [&](auto threads_constant) {
kernel = cu::row_reduce_looped<
T,
U,
OP,
reduce_ndim(),
threads_constant(),
N_READS>;
block.x = threads_constant();
});
});

63
mlx/backend/cuda/rms_norm.cu
View File

@@ -225,19 +225,20 @@ void RMSNorm::eval_gpu(
encoder.set_input_array(w);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_FLOAT_TYPES_CHECKED(out.dtype(), "rms_norm", CTYPE, {
using DataType = cuda_type_t<CTYPE>;
dispatch_float_types(out.dtype(), "rms_norm", [&](auto type_tag) {
constexpr uint32_t N_READS = 4;
MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
auto kernel = cu::rms_norm<DataType, BLOCK_DIM, N_READS>;
kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
x.data<DataType>(),
w.data<DataType>(),
out.data<DataType>(),
eps_,
axis_size,
w_stride);
});
dispatch_block_dim(
cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
auto kernel = cu::rms_norm<DataType, block_dim(), N_READS>;
kernel<<<n_rows, block_dim(), 0, stream>>>(
x.data<DataType>(),
w.data<DataType>(),
out.data<DataType>(),
eps_,
axis_size,
w_stride);
});
});
});
}
@@ -311,22 +312,28 @@ void RMSNormVJP::eval_gpu(
encoder.set_output_array(gx);
encoder.set_output_array(gw_temp);
encoder.launch_kernel([&, x = x, g = g](cudaStream_t stream) {
MLX_SWITCH_FLOAT_TYPES_CHECKED(gx.dtype(), "rms_norm_vjp", CTYPE, {
using DataType = cuda_type_t<CTYPE>;
constexpr int N_READS = 4;
MLX_SWITCH_BOOL(has_w, HAS_W, {
MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
auto kernel = cu::rms_norm_vjp<DataType, HAS_W, BLOCK_DIM, N_READS>;
kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
x.data<DataType>(),
w.data<DataType>(),
g.data<DataType>(),
gx.data<DataType>(),
gw_temp.data<DataType>(),
eps_,
axis_size,
w_stride);
});
dispatch_float_types(gx.dtype(), "rms_norm_vjp", [&](auto type_tag) {
dispatch_bool(has_w, [&](auto has_w_constant) {
constexpr int N_READS = 4;
dispatch_block_dim(
cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
constexpr int N_READS = 4;
auto kernel = cu::rms_norm_vjp<
DataType,
has_w_constant(),
block_dim(),
N_READS>;
kernel<<<n_rows, block_dim(), 0, stream>>>(
x.data<DataType>(),
w.data<DataType>(),
g.data<DataType>(),
gx.data<DataType>(),
gw_temp.data<DataType>(),
eps_,
axis_size,
w_stride);
});
});
});
});

17
mlx/backend/cuda/rope.cu
View File

@@ -310,12 +310,12 @@ void RoPE::eval_gpu(
encoder.set_input_array(offset);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_FLOAT_TYPES_CHECKED(in.dtype(), "rope", CTYPE, {
using DataType = cuda_type_t<CTYPE>;
MLX_SWITCH_BOOL(traditional_, TRADITIONAL, {
MLX_SWITCH_BOOL(forward_, FORWARD, {
dispatch_float_types(out.dtype(), "rope", [&](auto type_tag) {
dispatch_bool(traditional_, [&](auto traditional) {
dispatch_bool(forward_, [&](auto forward) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
if (single && !with_freqs) {
auto kernel = cu::rope_single<DataType, TRADITIONAL, FORWARD>;
auto kernel = cu::rope_single<DataType, traditional(), forward()>;
uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
kernel<<<grid, block, 0, stream>>>(
@@ -327,7 +327,8 @@ void RoPE::eval_gpu(
mat_size,
dims);
} else if (single) {
auto kernel = cu::rope_single_freqs<DataType, TRADITIONAL, FORWARD>;
auto kernel =
cu::rope_single_freqs<DataType, traditional(), forward()>;
uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
kernel<<<grid, block, 0, stream>>>(
@@ -340,7 +341,7 @@ void RoPE::eval_gpu(
dims,
inputs[2].strides(0));
} else if (with_freqs) {
auto kernel = cu::rope_freqs<DataType, TRADITIONAL, FORWARD>;
auto kernel = cu::rope_freqs<DataType, traditional(), forward()>;
uint3 dims =
make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
dims.z = (dims.z + 3) / 4;
@@ -358,7 +359,7 @@ void RoPE::eval_gpu(
dims,
inputs[2].strides(0));
} else {
auto kernel = cu::rope<DataType, TRADITIONAL, FORWARD>;
auto kernel = cu::rope<DataType, traditional(), forward()>;
uint3 dims =
make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
dims.z = (dims.z + 3) / 4;

21
mlx/backend/cuda/softmax.cu
View File

@@ -142,17 +142,18 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
encoder.set_input_array(in);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_FLOAT_TYPES_CHECKED(out.dtype(), "softmax", CTYPE, {
using DataType = cuda_type_t<CTYPE>;
dispatch_float_types(out.dtype(), "softmax", [&](auto type_tag) {
constexpr int N_READS = 4;
MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
auto kernel = cu::softmax<DataType, DataType, BLOCK_DIM, N_READS>;
if (precise) {
kernel = cu::softmax<DataType, float, BLOCK_DIM, N_READS>;
}
kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
in.data<DataType>(), out.data<DataType>(), axis_size);
});
dispatch_block_dim(
cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
auto kernel = cu::softmax<DataType, DataType, block_dim(), N_READS>;
if (precise) {
kernel = cu::softmax<DataType, float, block_dim(), N_READS>;
}
kernel<<<n_rows, block_dim(), 0, stream>>>(
in.data<DataType>(), out.data<DataType>(), axis_size);
});
});
});
}

14
mlx/backend/cuda/sort.cu
View File

@@ -76,6 +76,14 @@ void segmented_sort(cu::CommandEncoder& encoder, Args&&... args) {
temp.data<void>(), size, args...));
}
struct OffsetTransform {
int nsort;
int __device__ operator()(int i) {
return i * nsort;
}
};
void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
array out = out_;
auto& encoder = cu::get_command_encoder(s);
@@ -106,12 +114,12 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
encoder.set_input_array(in);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
dispatch_all_types(in.dtype(), [&](auto type_tag) {
using CTYPE = MLX_GET_TYPE(type_tag);
if constexpr (!std::is_same_v<CTYPE, complex64_t>) {
using Type = cuda_type_t<CTYPE>;
auto offsets = thrust::make_transform_iterator(
thrust::make_counting_iterator(0),
[nsort] __device__(int i) { return i * nsort; });
thrust::make_counting_iterator(0), OffsetTransform{nsort});
if (argsort) {
// Indices in the sorted dimension.
array indices(

95
mlx/backend/cuda/ternary.cu
View File

@@ -92,58 +92,63 @@ void ternary_op_gpu_inplace(
encoder.set_input_array(c);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE, {
using DType = cuda_type_t<CTYPE>;
dispatch_all_types(out.dtype(), [&](auto type_tag) {
using DType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
auto topt = get_ternary_op_type(a, b, c);
if (topt == TernaryOpType::General) {
auto [shape, strides] = collapse_contiguous_dims(a, b, c, out);
auto& a_strides = strides[0];
auto& b_strides = strides[1];
auto& c_strides = strides[2];
bool large = a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
c.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
MLX_SWITCH_BOOL(large, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
int ndim = shape.size();
if (ndim <= 3) {
MLX_SWITCH_1_2_3(ndim, NDIM, {
auto kernel = cu::ternary_g_nd<Op, DType, IdxT, NDIM>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<bool>(),
b.data<DType>(),
c.data<DType>(),
out.data<DType>(),
out.size(),
const_param<NDIM>(shape),
const_param<NDIM>(a_strides),
const_param<NDIM>(b_strides),
const_param<NDIM>(c_strides));
dispatch_bool(
a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
c.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
[&](auto large) {
using IdxT = std::conditional_t<large(), int64_t, int32_t>;
Shape shape;
std::vector<Strides> strides;
std::tie(shape, strides) = collapse_contiguous_dims(a, b, c, out);
auto& a_strides = strides[0];
auto& b_strides = strides[1];
auto& c_strides = strides[2];
int ndim = shape.size();
if (ndim <= 3) {
dispatch_1_2_3(ndim, [&](auto dims_constant) {
auto kernel =
cu::ternary_g_nd<Op, DType, IdxT, dims_constant()>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<bool>(),
b.data<DType>(),
c.data<DType>(),
out.data<DType>(),
out.size(),
const_param<dims_constant()>(shape),
const_param<dims_constant()>(a_strides),
const_param<dims_constant()>(b_strides),
const_param<dims_constant()>(c_strides));
});
} else {
auto kernel = cu::ternary_g<Op, DType, IdxT>;
auto [num_blocks, block_dims] =
get_launch_args(kernel, out, large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<bool>(),
b.data<DType>(),
c.data<DType>(),
out.data<DType>(),
out.data_size(),
const_param(shape),
const_param(a_strides),
const_param(b_strides),
const_param(c_strides),
ndim);
}
});
} else {
auto kernel = cu::ternary_g<Op, DType, IdxT>;
auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<bool>(),
b.data<DType>(),
c.data<DType>(),
out.data<DType>(),
out.data_size(),
const_param(shape),
const_param(a_strides),
const_param(b_strides),
const_param(c_strides),
ndim);
}
});
} else {
MLX_SWITCH_BOOL(out.data_size() > UINT32_MAX, LARGE, {
using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
auto kernel = cu::ternary_v<Op, DType, IdxT>;
auto [num_blocks, block_dims] = get_launch_args(
kernel, out.data_size(), out.shape(), out.strides(), LARGE);
kernel, out.data_size(), out.shape(), out.strides(), large());
kernel<<<num_blocks, block_dims, 0, stream>>>(
a.data<bool>(),
b.data<DType>(),

6
mlx/backend/cuda/unary.cu
View File

@@ -79,8 +79,10 @@ void unary_op_gpu_inplace(
encoder.set_input_array(in);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE_IN, {
MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE_OUT, {
dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
if constexpr (cu::supports_unary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
using InType = cuda_type_t<CTYPE_IN>;
using OutType = cuda_type_t<CTYPE_OUT>;

46
mlx/backend/cuda/utils.cpp
View File

@@ -25,22 +25,38 @@ void check_cuda_error(const char* name, cudaError_t err) {
}
const char* dtype_to_cuda_type(const Dtype& dtype) {
if (dtype == float16) {
return "__half";
switch (dtype) {
case bool_:
return "bool";
case int8:
return "int8_t";
case int16:
return "int16_t";
case int32:
return "int32_t";
case int64:
return "int64_t";
case uint8:
return "uint8_t";
case uint16:
return "uint16_t";
case uint32:
return "uint32_t";
case uint64:
return "uint64_t";
case float16:
return "__half";
case bfloat16:
return "__nv_bfloat16";
case float32:
return "float";
case float64:
return "double";
case complex64:
return "cuComplex";
default:
return "unknown";
}
if (dtype == bfloat16) {
return "__nv_bfloat16";
}
if (dtype == complex64) {
return "cuComplex";
}
#define SPECIALIZE_DtypeToString(CPP_TYPE, DTYPE) \
if (dtype == DTYPE) { \
return #CPP_TYPE; \
}
MLX_FORALL_DTYPES(SPECIALIZE_DtypeToString)
#undef SPECIALIZE_DtypeToString
return nullptr;
}
} // namespace mlx::core