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

sgemm_tcore: Split K-dim loop between consumer/producer

Browse Source 
... so that you don't have to run (warpgroup_id == 0) condition at every
loop iteration which is expensive due to vx_split/join.
This commit is contained in:
Hansung Kim
2024-06-05 17:14:39 -07:00
parent 150f14af25
commit 95b5719847
1 changed files with 75 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
tests/regression/sgemm_tcore/kernel.cpp
View File

@@ -557,14 +557,11 @@ void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
threadblock_barrier(threadblock_id_in_cluster, threadblock_dim_y); threadblock_barrier(threadblock_id_in_cluster, threadblock_dim_y);
} }
if (warpgroup_id == 0) {
// TODO: bring initiation pipeline here
uint32_t k_index = 0; uint32_t k_index = 0;
#pragma GCC unroll 1 #pragma GCC unroll 1
for (uint32_t k = 0; k < dim_k; k += BK) { for (uint32_t k = 0; k < dim_k - BK; k += BK) {
// register volatile float *local_a_produce asm("t0");
// register volatile float *local_b_produce asm("t1");
// register volatile float *local_a_consume asm("t2");
// register volatile float *local_b_consume asm("t3");
volatile float *local_a_produce; volatile float *local_a_produce;
volatile float *local_b_produce; volatile float *local_b_produce;
volatile float *local_a_consume; volatile float *local_a_consume;
@@ -574,8 +571,6 @@ void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
local_b_produce = (k_index % 2) ? local_b : local_b_buf; local_b_produce = (k_index % 2) ? local_b : local_b_buf;
local_a_consume = (k_index % 2) ? local_a_buf : local_a; local_a_consume = (k_index % 2) ? local_a_buf : local_a;
local_b_consume = (k_index % 2) ? local_b_buf : local_b; local_b_consume = (k_index % 2) ? local_b_buf : local_b;
// local_a_consume = local_a_produce;
// local_b_consume = local_b_produce;
} else { } else {
local_a_produce = local_a; local_a_produce = local_a;
local_b_produce = local_b; local_b_produce = local_b;
@@ -584,17 +579,35 @@ void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
} }
k_index++; k_index++;
if (warpgroup_id == 0) { global_dmem_load(dim_n, dim_k, k + BK /*runahead*/, A, B, local_a_produce,
if (k != (dim_k - BK)) { local_b_produce, tid_in_warpgroup, threadblock_id_x,
global_dmem_load(dim_n, dim_k, k + BK /*runahead*/, A, B, threadblock_id_y);
local_a_produce, local_b_produce, tid_in_warpgroup,
threadblock_id_x, threadblock_id_y);
}
threadblock_barrier(threadblock_id_in_cluster, threadblock_dim_y); threadblock_barrier(threadblock_id_in_cluster, threadblock_dim_y);
} }
else { threadblock_barrier(threadblock_id_in_cluster, threadblock_dim_y);
} else {
uint32_t k_index = 0;
#pragma GCC unroll 1
for (uint32_t k = 0; k < dim_k; k += BK) {
volatile float *local_a_produce;
volatile float *local_b_produce;
volatile float *local_a_consume;
volatile float *local_b_consume;
if constexpr (DOUBLE_BUFFER) {
local_a_produce = (k_index % 2) ? local_a : local_a_buf;
local_b_produce = (k_index % 2) ? local_b : local_b_buf;
local_a_consume = (k_index % 2) ? local_a_buf : local_a;
local_b_consume = (k_index % 2) ? local_b_buf : local_b;
} else {
local_a_produce = local_a;
local_b_produce = local_b;
local_a_consume = local_a;
local_b_consume = local_b;
}
k_index++;
#if USE_TENSOR_CORE #if USE_TENSOR_CORE
// @perf: this loop spills to stack a lot because of all the flws in // @perf: this loop spills to stack a lot because of all the flws in
// vx_wmma_load // vx_wmma_load
@@ -603,12 +616,14 @@ void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
#pragma GCC unroll 1 #pragma GCC unroll 1
for (uint32_t local_k = 0; local_k < BK; local_k += TCK) { for (uint32_t local_k = 0; local_k < BK; local_k += TCK) {
// perform wmma // perform wmma
// vx_wmma_load(local_a_consume, local_b_consume, warp_x, warp_y, tid_in_warp); // vx_wmma_load(local_a_consume, local_b_consume, warp_x, warp_y,
// tid_in_warp);
// FIXME: this is wrong!! need separate accumulation register for // FIXME: this is wrong!! need separate accumulation register for
// WM/WN_ITERS // WM/WN_ITERS
#pragma GCC unroll 2 #pragma GCC unroll 2
for (int wn_iter = 0; wn_iter < WNITER; wn_iter++) { for (int wn_iter = 0; wn_iter < WNITER; wn_iter++) {
vx_wmma_load_b(local_b_consume, local_k, warp_col, wn_iter, tid_in_warp); vx_wmma_load_b(local_b_consume, local_k, warp_col, wn_iter,
tid_in_warp);
// vx_wmma_load_b(local_b_consume, 0, 0, 0, tid_in_warp); // vx_wmma_load_b(local_b_consume, 0, 0, 0, tid_in_warp);
#pragma GCC unroll 1 #pragma GCC unroll 1
for (int wm_iter = 0; wm_iter < WMITER; wm_iter++) { for (int wm_iter = 0; wm_iter < WMITER; wm_iter++) {
@@ -641,7 +656,6 @@ void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
} }
threadblock_barrier(threadblock_id_in_cluster, threadblock_dim_y); threadblock_barrier(threadblock_id_in_cluster, threadblock_dim_y);
}
#else #else
@@ -660,7 +674,8 @@ void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
local_b[BN * local_k + (TN * local_c_col + res_idx_n)]; local_b[BN * local_k + (TN * local_c_col + res_idx_n)];
} }
// Next, compute multiple result elements (TM*TN) by reusing data in RF // Next, compute multiple result elements (TM*TN) by reusing data in
// RF
#pragma GCC unroll TM #pragma GCC unroll TM
for (uint32_t res_idx_m = 0; res_idx_m < TM; res_idx_m++) { for (uint32_t res_idx_m = 0; res_idx_m < TM; res_idx_m++) {
#pragma GCC unroll TN #pragma GCC unroll TN
@@ -679,6 +694,7 @@ void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
threadblock_dim_y); threadblock_dim_y);
#endif #endif
} }
}
#if USE_TENSOR_CORE #if USE_TENSOR_CORE
#pragma GCC unroll 1 #pragma GCC unroll 1