wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

sgemm_impl: Parameterize BM/BN/BK in single_tile

Browse Source
This commit is contained in:
Hansung Kim
2024-08-20 19:41:34 -07:00
parent dde0372769
commit 091f40c365
1 changed files with 10 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
tests/regression/sgemm_tcore/sgemm_impl.hpp
View File

@@ -606,6 +606,9 @@ load_tile_to_smem(const uint32_t dim_major, const uint32_t mn_index,
template <typename T,
MemLayout layout_a, // memory layout of `local_a`
MemLayout layout_b, // memory layout of `local_b`
uint32_t tile_dim_m,
uint32_t tile_dim_n,
uint32_t tile_dim_k,
bool load_accum = false, // if true, load the accumulation registers
// with `local_c`. used for the (C + A*B)
// operation
@@ -621,8 +624,8 @@ __attribute__((always_inline)) inline void thread_block_gemm_single_tile(
// FIXME: duplicated from thread_block_gemm
const uint32_t threads_per_warpgroup = threads_per_threadblock;
const uint32_t warp_id_in_warpgroup = tid_in_threadblock / NUM_THREADS;
const uint32_t warp_row = warp_id_in_warpgroup / (BN / WN);
const uint32_t warp_col = warp_id_in_warpgroup % (BN / WN);
const uint32_t warp_row = warp_id_in_warpgroup / (tile_dim_n / WN);
const uint32_t warp_col = warp_id_in_warpgroup % (tile_dim_n / WN);
const uint32_t tid_in_warp = tid_in_threadblock % NUM_THREADS;
const uint32_t warps_per_threadblock_per_core =
NUM_WARPS / threadblocks_per_cluster;
@@ -636,8 +639,8 @@ __attribute__((always_inline)) inline void thread_block_gemm_single_tile(
#pragma GCC unroll
for (int wn_iter = 0; wn_iter < WNITER; wn_iter++) {
// FIXME: template parameter-ize BM
wmma_load_accum(tid_in_warp, warp_col, warp_row, wn_iter, wm_iter, BN,
local_c);
wmma_load_accum(tid_in_warp, warp_col, warp_row, wn_iter, wm_iter,
tile_dim_n, local_c);
}
}
}
@@ -645,7 +648,7 @@ __attribute__((always_inline)) inline void thread_block_gemm_single_tile(
#pragma GCC unroll 1
for (int i = 0; i < BK_LOOP; i++) {
#pragma GCC unroll 4
for (uint32_t local_k = 0; local_k < BK; local_k += TCK) {
for (uint32_t local_k = 0; local_k < tile_dim_k; local_k += TCK) {
#pragma GCC unroll 2
for (int wn_iter = 0; wn_iter < WNITER; wn_iter++) {
// SMEM -> RF
@@ -682,7 +685,7 @@ __attribute__((always_inline)) inline void thread_block_gemm_single_tile(
for (int wm_iter = 0; wm_iter < WMITER; wm_iter++) {
#pragma GCC unroll
for (int wn_iter = 0; wn_iter < WNITER; wn_iter++) {
wmma_store(tid_in_warp, warp_col, warp_row, wn_iter, wm_iter, BN,
wmma_store(tid_in_warp, warp_col, warp_row, wn_iter, wm_iter, tile_dim_n,
result_addr);
}
}
@@ -918,6 +921,7 @@ inline void thread_block_gemm(const T *A, const T *B, float *C,
constexpr MemLayout layout_a =
TRANSPOSE_AT_CONSUME ? MemLayout::K_major : MemLayout::MN_major;
thread_block_gemm_single_tile<T, layout_a, MemLayout::MN_major,
BM, BN, BK,
/*load_accum=*/false,
/*write_to_mem=*/false>(
local_a_consume, local_b_consume,