topk_softmax (#1592)

* topk_softmax

* remove some file

* fix atomix linear_offset

* address various comment, and change sfc get_index api to static(tuple)
This commit is contained in:
carlushuang
2024-10-26 23:52:49 +08:00
committed by GitHub
parent 31bf253aeb
commit b098b71b05
41 changed files with 5603 additions and 226 deletions

View File

@@ -334,7 +334,7 @@ struct BlockFmhaPipelineQRKSVSAsync
move_tile_window(k_dram_window, {0, kK0});
__builtin_amdgcn_sched_barrier(0);
buffer_load_fence(k_dram_window.get_num_access(), q.get_thread_buffer());
buffer_load_fence(k_dram_window.get_num_of_access(), q.get_thread_buffer());
(void)q_element_func; // ??? rocm-6.x if use q element func will have scratch on hdim=64/32
// auto q_tile = q; // tile_elementwise_in(q_element_func, q);
@@ -359,7 +359,7 @@ struct BlockFmhaPipelineQRKSVSAsync
if constexpr(i_k0 < k0_loops - 1)
move_tile_window(k_dram_window, {0, kK0});
async_load_fence(k_dram_window.get_num_access());
async_load_fence(k_dram_window.get_num_of_access());
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
gemm_0(s_acc,