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] Switch to CUDA graphs (#2317)

Browse Source 
* cuda graph prototype

fix signal bug + start to add dependencies

capture more

capture more ops

remaining ops

fix reduce and rope deps

add concurrent context

try update, but not working

cosistent topology order

use node api

use node api directly to reduce overhead

fix bug

use kernels in unary

cache graph

format

fix synchronization

format

* comment
This commit is contained in:
Awni Hannun
2025-07-02 15:59:13 -07:00
committed by GitHub
parent e76e9b87f0
commit ec0d5db67b
36 changed files with 1461 additions and 1212 deletions
Download Patch File Download Diff File Expand all files Collapse all files

82
mlx/backend/cuda/copy/copy_general.cu
View File

@@ -55,50 +55,54 @@ void copy_general(
const Shape& shape,
const Strides& strides_in,
const Strides& strides_out) {
encoder.launch_kernel([&](cudaStream_t stream) {
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>;
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>>>(
encoder.add_kernel_node(
kernel,
num_blocks,
block_dims,
in_ptr,
out_ptr,
data_size,
const_param(shape),
const_param(strides_in),
const_param(strides_out),
ndim);
}
});
});
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());
encoder.add_kernel_node(
kernel,
num_blocks,
block_dims,
in_ptr,
out_ptr,
data_size,
const_param(shape),
const_param(strides_in),
const_param(strides_out),
ndim);
}
});
});
});
}