squall/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

fix race condition when h < stride_h or w < stride_w (#562)

Browse Source 
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Haicheng Wu 2022-07-12 16:37:08 -04:00 committed by GitHub
parent fb379eaa5b
commit e7a61c761a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 20 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
include/cutlass/conv/kernel/implicit_gemm_convolution_strided_dgrad.h
View File

@ -307,6 +307,26 @@ struct ImplicitGemmConvolutionStridedDgrad {
int start_r, start_s;
params.stride_w_divmod(start_r, start_s, filter_tile_m);
int filter_r = start_r;
int filter_s = start_s;
if (params.problem_size.mode == Mode::kConvolution) {
filter_r = (params.problem_size.R - 1 - filter_r);
filter_s = (params.problem_size.S - 1 - filter_s);
}
// Starting h, w positions for filter position in gemm_k=0
int start_h, start_w;
strided_dgrad_starting_coords(
params.problem_size,
params.stride_h_divmod, params.stride_w_divmod,
filter_r, filter_s,
start_h, start_w);
if (start_h >= params.problem_size.H || start_w >= params.problem_size.W) {
return;
}
typename Mma::FragmentC accumulators;
accumulators.clear();