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

update gemmini only kernel

Browse Source
This commit is contained in:
Richard Yan
2024-04-15 10:22:00 -07:00
parent 0bb7aeb45b
commit 041d49fb58
2 changed files with 18 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
tests/kernel/gemmini_mmio/gemmini_mmio.h
View File

@@ -14,7 +14,7 @@
#define SPAD_NUM_ROWS (SMEM_SIZE / SPAD_ROW_SIZE)
#define SPAD_MASK (SPAD_NUM_ROWS - 1)
#define PRINT_BUF SMEM_ADDR_END
#define PRINT_BUF ((char *) (SMEM_ADDR_END))
#define GEMMINI_RS1_ADDR 0xff007010
#define GEMMINI_RS2_ADDR 0xff007018
#define GEMMINI_INST_ADDR 0xff007000
@@ -32,7 +32,8 @@
(((i) / DIM * (J) / DIM + (j) / DIM) * DIM * DIM + ((i) % DIM) * DIM + ((j) % DIM))
// #define fence() { for (int i = 0; i < 10; i++) *((volatile uint32_t *) (0xFFFF0000)) = 0xdeadbeef; }
#define fence() { while (*((volatile uint32_t *) GEMMINI_BUSY_ADDR)) asm volatile ("nop"); }
#undef gemmini_fence
#define gemmini_fence() { while (*((volatile uint32_t *) GEMMINI_BUSY_ADDR)) asm volatile ("nop"); }
#undef ROCC_INSTRUCTION_RS1_RS2
#define ROCC_INSTRUCTION_RS1_RS2(x, rs1, rs2, funct) { \
@@ -60,6 +61,7 @@ static void sp_tiled_matmul_full_spad_ws(const uint32_t A_sp_addr_start, const u
a_transpose, b_transpose,
full_C, low_D, false,
act, 0, 0, false);
/*
return;
@@ -72,7 +74,7 @@ static void sp_tiled_matmul_full_spad_ws(const uint32_t A_sp_addr_start, const u
const int C_blocks = 1; //full_C ? 1 : (J <= MAX_BLOCK_LEN ? J : MAX_BLOCK_LEN);
// const size_t sizeof_D = low_D ? sizeof(elem_t) : sizeof(acc_t);
const size_t sizeof_C = full_C ? sizeof(acc_t) : sizeof(elem_t);
fence();
gemmini_fence();
if (a_transpose || b_transpose || (I < 4)) {
for (size_t k = 0; k < K; k++) {
@@ -140,7 +142,7 @@ static void sp_tiled_matmul_full_spad_ws(const uint32_t A_sp_addr_start, const u
gemmini_extended_compute_accumulated(A_sp_addr + 3 * K * DIM, GARBAGE_ADDR, DIM, DIM, DIM, DIM);
}
if (k == K - 1) {
for (int x = 0; x < 3; x++) fence();
for (int x = 0; x < 3; x++) gemmini_fence();
gemmini_extended_mvout_spad((uint32_t) C_dst_sp_addr_start + (i * J + j) * DIM, 1, C_sp_addr, DIM, DIM);
gemmini_extended_mvout_spad((uint32_t) C_dst_sp_addr_start + ((i + 1) * J + j) * DIM, 1, C_sp_addr + J * DIM, DIM, DIM);
gemmini_extended_mvout_spad((uint32_t) C_dst_sp_addr_start + ((i + 2) * J + j) * DIM, 1, C_sp_addr + 2 * J * DIM, DIM, DIM);
@@ -152,7 +154,8 @@ static void sp_tiled_matmul_full_spad_ws(const uint32_t A_sp_addr_start, const u
}
}
}
fence();
gemmini_fence();
*/
}

18
tests/kernel/gemmini_mmio/main.cpp
View File

@@ -3,7 +3,7 @@
#include <vx_intrinsics.h>
#include <vx_print.h>
#include <vx_spawn.h>
#include <include/gemmini.h>
#include "include/gemmini.h"
#include "gemmini_mmio.h"
#define rd_cycles(x) asm volatile ("csrr %0, mcycle" : "=r" (x))
@@ -12,7 +12,7 @@ int main() {
int cid;
asm volatile ("csrr %0, 0xcc2" : "=r" (cid));
if (cid > 0) return 0;
if (cid > 0) vx_tmc(0);
vx_tmc(0xff);
@@ -40,7 +40,11 @@ int main() {
vx_tmc_one();
gemmini_config_ld(0);
gemmini_extended_config_ex(WEIGHT_STATIONARY, 0, 0, 1, 0, 0);
gemmini_config_st(DIM * 4 * J);
gemmini_config_st(0);
/* sprintf(print_buf, "A spad: 0x%x-0x%x, smem: 0x%x-%x\n", spad_A, spad_A + I * K * DIM, (uint32_t) smem_A, (uint32_t) smem_A + sizeof(float) * I * K * DIM * DIM);
sprintf(print_buf, "B spad: 0x%x-0x%x, smem: 0x%x-%x\n", spad_B, spad_B + K * J * DIM, (uint32_t) smem_B, (uint32_t) smem_B + sizeof(float) * K * J * DIM * DIM);
sprintf(print_buf, "C spad: 0x%x-0x%x, smem: 0x%x-%x\n", spad_C, spad_C + I * J * DIM, (uint32_t) smem_C, (uint32_t) smem_C + sizeof(float) * I * J * DIM * DIM); */
sprintf(print_buf, "DIM %d\n", DIM);
sprintf(print_buf, "num cores %d\n", nc);
sprintf(print_buf, "num threads %d\n", nt);
@@ -49,9 +53,6 @@ int main() {
sprintf(print_buf, "%d", tid);
uint32_t start_cycles, end_cycles;
/* sprintf(print_buf, "A spad: 0x%x-0x%x, smem: 0x%x-%x\n", spad_A, spad_A + I * K * DIM, (uint32_t) smem_A, (uint32_t) smem_A + sizeof(float) * I * K * DIM * DIM);
sprintf(print_buf, "B spad: 0x%x-0x%x, smem: 0x%x-%x\n", spad_B, spad_B + K * J * DIM, (uint32_t) smem_B, (uint32_t) smem_B + sizeof(float) * K * J * DIM * DIM);
sprintf(print_buf, "C spad: 0x%x-0x%x, smem: 0x%x-%x\n", spad_C, spad_C + I * J * DIM, (uint32_t) smem_C, (uint32_t) smem_C + sizeof(float) * I * J * DIM * DIM); */
rd_cycles(start_cycles);
// load A with 128->1 in row-major order
@@ -87,7 +88,7 @@ int main() {
vx_tmc_one();
sprintf(print_buf, "\ndata loading took %d cycles for %d floats\n", end_cycles - start_cycles, DIM * DIM * (I * K + J * K));
fence();
gemmini_fence();
// sprintf(print_buf, "\nA in\n");
// for (int i = 0; i < I * DIM; i++) {
@@ -113,7 +114,7 @@ int main() {
/*no_bias=*/1, /*repeating_bias=*/0, /*act=*/NO_ACTIVATION);
rd_cycles(fence_cycles);
fence();
gemmini_fence();
rd_cycles(end_cycles);
sprintf(print_buf, "gemmini cycles taken: %d, fence cycles: %d\n", end_cycles - start_cycles, end_cycles - fence_cycles);
@@ -138,5 +139,6 @@ int main() {
}
sprintf(print_buf, "TEST PASSED\n");
vx_tmc(0);
return 0;
}