Skip to content

tuned grid size by reducing num_warps_per_threadblock to 4 #117

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
kudomcho wants to merge 3 commits into
base: mi350_develop_0811
Choose a base branch
from
Open

tuned grid size by reducing num_warps_per_threadblock to 4 #117

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
11e0d3d
tuned grid size by reducing num_warps_per_threadblock to 4
Aug 26, 2025
91c60a6
removed print statements
kudomcho Aug 26, 2025
4a3e0cf
added rocm guard on wg size change on v3 fwd
kudomcho Aug 28, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 7 additions & 3 deletions fbgemm_gpu/codegen/training/forward/embedding_forward_split_template.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -795,15 +795,19 @@ batch_index_select_dim0_codegen_forward_cuda(
// if (!is_experimental)
} else {
// Allocate num warps per table based on max_D

const int num_warps_per_table = B * div_round_up(max_D, kWarpSize * 4);
const uint32_t num_warps_per_threadblock = kForwardMaxThreads / kWarpSize;

#ifdef USE_ROCM
const uint32_t num_warps_per_threadblock = kForwardMaxThreads / (kWarpSize*2);
#else
const uint32_t num_warps_per_threadblock = kForwardMaxThreads / kWarpSize;
#endif

const auto kernel_func =
(use_lxu_cache ? split_embedding_codegen_forward_{{ wdesc }}_v2_kernel<
emb_t, cache_t, output_t, index_t, true>
: split_embedding_codegen_forward_{{ wdesc }}_v2_kernel<
emb_t, cache_t, output_t, index_t, false>);

FBGEMM_LAUNCH_KERNEL(
kernel_func,
div_round_up(T * num_warps_per_table, num_warps_per_threadblock),
Expand Down