huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-19 11:56:45 +00:00
Code Issues Packages Projects Releases Wiki Activity

Cuda backend optimization (#1886)

Browse Source 
* Attempt at making the kernel faster.

* Also adapt the cast kernels.

* Also apply to binary ops.
This commit is contained in:
Laurent Mazare
2024-03-20 18:32:55 +01:00
committed by GitHub
parent 469635a3eb
commit b219903d0f
5 changed files with 54 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
candle-kernels/src/binary_op_macros.cuh
View File

@ -12,8 +12,8 @@ extern "C" __global__ void FN_NAME( \
const size_t *dims = dims_and_strides; \
const size_t *lhs_strides = dims_and_strides + 1 * num_dims; \
const size_t *rhs_strides = dims_and_strides + 2 * num_dims; \
bool lhs_cont = is_contiguous(num_dims, dims, lhs_strides); \
bool rhs_cont = is_contiguous(num_dims, dims, rhs_strides); \
bool lhs_cont = dims_and_strides == nullptr || is_contiguous(num_dims, dims, lhs_strides); \
bool rhs_cont = dims_and_strides == nullptr || is_contiguous(num_dims, dims, rhs_strides); \
if (lhs_cont && rhs_cont) { \
for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \
TYPENAME x = lhs[i]; \