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 Packages Projects Releases Wiki Activity

Move all type switching to templates

Browse Source
This commit is contained in:
Angelos Katharopoulos
2025-06-29 04:04:15 -07:00
parent 45c43dd24a
commit ef813b6d13
19 changed files with 474 additions and 431 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
mlx/backend/cuda/arg_reduce.cu
View File

@@ -155,25 +155,33 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
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<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_dims, 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);
});
dispatch_block_dim(
cuda::ceil_div(axis_size, N_READS), [&](auto block_dim_constant) {
dim3 num_blocks = get_2d_grid_dims(out.shape(), out.strides());
dim3 block_dims{block_dim_constant(), 1, 1};
auto kernel = cu::arg_reduce_general<
T,
cu::ArgMax<T>,
block_dim_constant(),
N_READS>;
if (reduce_type_ == ArgReduce::ArgMin) {
kernel = cu::arg_reduce_general<
T,
cu::ArgMin<T>,
block_dim_constant(),
N_READS>;
}
kernel<<<num_blocks, block_dims, 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);
});
});
});
}

88
mlx/backend/cuda/binary.cu
View File

@@ -149,47 +149,55 @@ void binary_op_gpu_inplace(
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>;
@@ -199,7 +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>(),

89
mlx/backend/cuda/binary_two.cu
View File

@@ -148,49 +148,54 @@ void binary_op_gpu_inplace(
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>;
auto [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>;
@@ -204,7 +209,7 @@ 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>(),

10
mlx/backend/cuda/copy/copy_contiguous.cu
View File

@@ -38,16 +38,16 @@ void copy_contiguous(
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
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 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, 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,

76
mlx/backend/cuda/copy/copy_general.cu
View File

@@ -58,44 +58,46 @@ void copy_general(
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
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));
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);
}
});
} 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);
}
});
});
});
});

80
mlx/backend/cuda/copy/copy_general_dynamic.cu
View File

@@ -64,44 +64,50 @@ void copy_general_dynamic(
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
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>());
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>());
}
});
} 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>());
}
});
});
});
});

65
mlx/backend/cuda/copy/copy_general_input.cu
View File

@@ -53,38 +53,41 @@ void copy_general_input(
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
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));
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);
}
});
} 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>

62
mlx/backend/cuda/layer_norm.cu
View File

@@ -260,20 +260,21 @@ void LayerNorm::eval_gpu(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_float_types(out.dtype(), "layernorm", [&](auto type_tag) {
using DataType = 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, {
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);
});
});
});
}
@@ -358,21 +359,26 @@ void LayerNormVJP::eval_gpu(
encoder.set_output_array(gw_temp);
encoder.launch_kernel([&, x = x, g = g](cudaStream_t stream) {
dispatch_float_types(gx.dtype(), "layernorm_vjp", [&](auto type_tag) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
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_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);
});
});
});
});

13
mlx/backend/cuda/logsumexp.cu
View File

@@ -145,13 +145,14 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_float_types(out.dtype(), "logsumexp", [&](auto type_tag) {
using DataType = cuda_type_t<MLX_GET_TYPE(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);
});
});
});
}

6
mlx/backend/cuda/reduce/all_reduce.cu
View File

@@ -112,7 +112,8 @@ void all_reduce(
encoder.set_output_array(intermediate);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(dt, [&](auto type_tag) {
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
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>;
@@ -136,7 +137,8 @@ void all_reduce(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(dt, [&](auto type_tag) {
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
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>;

11
mlx/backend/cuda/reduce/col_reduce.cu
View File

@@ -216,10 +216,10 @@ void col_reduce_looped(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(in.dtype(), [&](auto type_tag) {
using CTYPE = MLX_GET_TYPE(type_tag);
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
MLX_SWITCH_REDUCE_NDIM(args.reduce_ndim, NDIM, {
using T = cuda_type_t<CTYPE>;
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)
@@ -230,7 +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);
});
});

3
mlx/backend/cuda/reduce/init_reduce.cu
View File

@@ -34,7 +34,8 @@ void init_reduce(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(in.dtype(), [&](auto type_tag) {
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
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>;

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,

26
mlx/backend/cuda/reduce/row_reduce.cu
View File

@@ -247,9 +247,9 @@ void row_reduce_simple(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(in.dtype(), [&](auto type_tag) {
using CTYPE = MLX_GET_TYPE(type_tag);
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
using T = cuda_type_t<CTYPE>;
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)
@@ -295,9 +295,9 @@ void row_reduce_looped(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(in.dtype(), [&](auto type_tag) {
using CTYPE = MLX_GET_TYPE(type_tag);
MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
using T = cuda_type_t<CTYPE>;
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)
@@ -313,10 +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();
});
});

59
mlx/backend/cuda/rms_norm.cu
View File

@@ -226,18 +226,19 @@ void RMSNorm::eval_gpu(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_float_types(out.dtype(), "rms_norm", [&](auto type_tag) {
using DataType = 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, {
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);
});
});
});
}
@@ -312,21 +313,27 @@ void RMSNormVJP::eval_gpu(
encoder.set_output_array(gw_temp);
encoder.launch_kernel([&, x = x, g = g](cudaStream_t stream) {
dispatch_float_types(gx.dtype(), "rms_norm_vjp", [&](auto type_tag) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
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_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);
});
});
});
});

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

@@ -311,11 +311,11 @@ void RoPE::eval_gpu(
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_float_types(out.dtype(), "rope", [&](auto type_tag) {
using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
MLX_SWITCH_BOOL(traditional_, TRADITIONAL, {
MLX_SWITCH_BOOL(forward_, FORWARD, {
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;

19
mlx/backend/cuda/softmax.cu
View File

@@ -143,16 +143,17 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_float_types(out.dtype(), "softmax", [&](auto type_tag) {
using DataType = cuda_type_t<MLX_GET_TYPE(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);
});
});
});
}

89
mlx/backend/cuda/ternary.cu
View File

@@ -97,53 +97,56 @@ void ternary_op_gpu_inplace(
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>;
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];
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>(),