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

skip move to spad

Browse Source
This commit is contained in:
Richard Yan
2024-06-19 17:45:58 -07:00
parent c06cc40e59
commit a1e165724f
1 changed files with 11 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
tests/regression/sgemm_gemmini_dma/kernel.cpp
View File

@@ -42,7 +42,7 @@
#define PRINTF(...) sprintf(PRINT_BUF, __VA_ARGS__) #define PRINTF(...) sprintf(PRINT_BUF, __VA_ARGS__)
// #define PRINTF(...) vx_printf(__VA_ARGS__) // #define PRINTF(...) vx_printf(__VA_ARGS__)
#define SWISH(beta, x) ((x) / (1 + exp(-(beta) * (x)))) #define SWISH(beta, x) ((x) / (1 + exp(-(beta) * (x))))
//#define POWER #define POWER
inline void threadblock_barrier(unsigned int barrier_id, unsigned int count) { inline void threadblock_barrier(unsigned int barrier_id, unsigned int count) {
vx_fence(); vx_fence();
@@ -60,7 +60,9 @@ void thread_block_matmul_gemmini(kernel_arg_t *__UNIFORM__ arg,
if (HW_TID() == 0) { if (HW_TID() == 0) {
gemmini_extended_config_ex(WEIGHT_STATIONARY, 0, 0, 1, 0, 0); gemmini_extended_config_ex(WEIGHT_STATIONARY, 0, 0, 1, 0, 0);
// gemmini_extended_config_ex(dataflow, act & 3, 0, 1, a_transpose, b_transpose); // gemmini_extended_config_ex(dataflow, act & 3, 0, 1, a_transpose, b_transpose);
#ifndef POWER
PRINTF("start\n"); PRINTF("start\n");
#endif
} }
vx_fence(); vx_fence();
@@ -119,15 +121,15 @@ void thread_block_matmul_gemmini(kernel_arg_t *__UNIFORM__ arg,
gemmini_fence(); gemmini_fence();
gemmini_fence(); gemmini_fence();
// mvout to scratchpad for activation // mvout to scratchpad for activation
GEMMINI_CISC_CMD_I(9); // GEMMINI_CISC_CMD_I(9);
gemmini_fence(); // gemmini_fence();
} // }
threadblock_barrier(/*barrier_id=*/0, /*count=*/NUM_WARPS); // threadblock_barrier(/*barrier_id=*/0, /*count=*/NUM_WARPS);
// activate // // activate
// move out to dram // // move out to dram
if (HW_TID() == 0) { // if (HW_TID() == 0) {
float * const dram_c_tile_start = C + tile_i * TILE_M * dim_n + tile_j * TILE_N; float * const dram_c_tile_start = C + tile_i * TILE_M * dim_n + tile_j * TILE_N;
ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, 0, BOUND_INST, k_LOOP_WS_CONFIG_BOUNDS) ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, 0, BOUND_INST, k_LOOP_WS_CONFIG_BOUNDS)
ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, 0, (uint64_t) dram_c_tile_start, k_LOOP_WS_CONFIG_ADDRS_DC) ROCC_INSTRUCTION_RS1_RS2(XCUSTOM_ACC, 0, (uint64_t) dram_c_tile_start, k_LOOP_WS_CONFIG_ADDRS_DC)
@@ -142,7 +144,7 @@ void thread_block_matmul_gemmini(kernel_arg_t *__UNIFORM__ arg,
rd_cycles_force(marker1); rd_cycles_force(marker1);
if (HW_TID() == 0) { if (HW_TID() == 0) {
#ifdef POWER #ifdef POWER
PRINTF("\nstart %d end %d\n", marker0, marker1); PRINTF("%d\n", marker1 - marker0);
#else #else
PRINTF("\ncomplete\n"); PRINTF("\ncomplete\n");
PRINTF("total cycles: %d\n", marker1 - marker0); PRINTF("total cycles: %d\n", marker1 - marker0);