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

[CUDA] Do vectorized store/load in contiguous elementwise ops (#2342)

Browse Source 
* Do vectorized store/load in unary ops

* Do vectorized store/load in binary_two ops

* Do vectorized store/load in copy ops

* Do vectorized store/load in ternary ops

* Use int32_t for IdxT

* binary => binary_two in binary_two.cu

* Fix tests on large arrays

* Use uint as index type

* Contig uses uint as index and non-contig uses int
This commit is contained in:
Cheng
2025-07-10 10:48:43 +09:00
committed by GitHub
parent e14ee12491
commit 85873cb162
5 changed files with 223 additions and 96 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
mlx/backend/cuda/copy/copy_contiguous.cu
View File

@@ -10,19 +10,43 @@ namespace cu {
namespace cg = cooperative_groups;
template <typename In, typename Out, typename IdxT>
template <typename In, typename Out, typename IdxT, int N_READS>
__global__ void copy_s(const In* in, Out* out, IdxT size) {
IdxT index = cg::this_grid().thread_rank();
if (index < size) {
out[index] = CastOp<In, Out>{}(in[0]);
if ((index + 1) * N_READS > size) {
for (IdxT i = index * N_READS; i < size; ++i) {
out[i] = cast_to<Out>(in[0]);
}
} else {
AlignedVector<Out, N_READS> out_vec;
#pragma unroll
for (int i = 0; i < N_READS; ++i) {
out_vec.val[i] = cast_to<Out>(in[0]);
}
store_vector<N_READS>(out, index, out_vec);
}
}
template <typename In, typename Out, typename IdxT>
template <typename In, typename Out, typename IdxT, int N_READS>
__global__ void copy_v(const In* in, Out* out, IdxT size) {
IdxT index = cg::this_grid().thread_rank();
if (index < size) {
out[index] = CastOp<In, Out>{}(in[index]);
if ((index + 1) * N_READS > size) {
for (IdxT i = index * N_READS; i < size; ++i) {
out[i] = cast_to<Out>(in[i]);
}
} else {
auto in_vec = load_vector<N_READS>(in, index);
AlignedVector<Out, N_READS> out_vec;
#pragma unroll
for (int i = 0; i < N_READS; ++i) {
out_vec.val[i] = cast_to<Out>(in_vec.val[i]);
}
store_vector<N_READS>(out, index, out_vec);
}
}
@@ -41,12 +65,19 @@ void copy_contiguous(
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>;
// TODO: Choose optimized value based on type size.
constexpr int N_READS = 4;
auto kernel = cu::copy_s<InType, OutType, IdxT, N_READS>;
if (ctype == CopyType::Vector) {
kernel = cu::copy_v<InType, OutType, IdxT>;
kernel = cu::copy_v<InType, OutType, IdxT, N_READS>;
}
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(),
N_READS);
encoder.add_kernel_node(
kernel,
num_blocks,