DeepSeek/DeepGEMM
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepGEMM synced 2025-05-07 23:14:22 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Simplify

Browse Source
This commit is contained in:
Chenggang Zhao 2025-04-28 10:12:35 +08:00
parent 1e23215eb6
commit 46c7a1ef36
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
deep_gemm/include/deep_gemm/fp8_gemm.cuh
View File

@ -192,7 +192,6 @@ fp8_gemm_kernel(__nv_bfloat16* gmem_d, float* scales_b, int* grouped_layout,
if (threadIdx.x == kNumMathThreads) { if (threadIdx.x == kNumMathThreads) {
// Persistently schedule over blocks // Persistently schedule over blocks
while (scheduler.get_next_block(m_block_idx, n_block_idx)) { while (scheduler.get_next_block(m_block_idx, n_block_idx)) {
launch_k_iterations([&](uint32_t k_iter, uint32_t stage_idx, auto _) {
// Assign TMA multicast number into A and B // Assign TMA multicast number into A and B
// NOTES: there may be additional odd rows/columns or cases where multicast is not possible. // NOTES: there may be additional odd rows/columns or cases where multicast is not possible.
const bool is_tma_multicast_valid = scheduler.is_tma_multicast_valid(m_block_idx); const bool is_tma_multicast_valid = scheduler.is_tma_multicast_valid(m_block_idx);
@ -200,6 +199,7 @@ fp8_gemm_kernel(__nv_bfloat16* gmem_d, float* scales_b, int* grouped_layout,
const uint32_t num_tma_multicast_b = (not kIsTMAMulticastOnA and is_tma_multicast_valid) ? kNumTMAMulticast : 1; const uint32_t num_tma_multicast_b = (not kIsTMAMulticastOnA and is_tma_multicast_valid) ? kNumTMAMulticast : 1;
DG_STATIC_ASSERT(kNumTMAMulticast <= 2, "Scheduler does not support > 2 TMA multicast"); DG_STATIC_ASSERT(kNumTMAMulticast <= 2, "Scheduler does not support > 2 TMA multicast");
launch_k_iterations([&](uint32_t k_iter, uint32_t stage_idx, auto _) {
// Wait consumer release // Wait consumer release
auto phase_idx = ((scheduler.current_iter * kNumStagesPerBlock + k_iter) / kNumStages) & 1; auto phase_idx = ((scheduler.current_iter * kNumStagesPerBlock + k_iter) / kNumStages) & 1;
empty_barriers[stage_idx]->wait(phase_idx ^ 1); empty_barriers[stage_idx]->wait(phase_idx ^ 1);