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

thread parallel data loading for word strided bank

Browse Source
This commit is contained in:
Richard Yan
2024-04-01 11:10:32 -07:00
parent e6db1a83af
commit 84a31f3384
1 changed files with 55 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

86
tests/kernel/gemmini_mmio/main.cpp
View File

@@ -2,23 +2,25 @@
#include <stdint.h> #include <stdint.h>
#include <vx_intrinsics.h> #include <vx_intrinsics.h>
#include <vx_print.h> #include <vx_print.h>
#include <vx_spawn.h>
#include <include/gemmini.h> #include <include/gemmini.h>
#include "gemmini_mmio.h" #include "gemmini_mmio.h"
#define rd_cycles(x) asm volatile ("csrr %0, mcycle" : "=r" (x))
int main() { int main() {
char *print_buf = (char *) PRINT_BUF; int cid;
asm volatile ("csrr %0, 0xcc2" : "=r" (cid));
if (cid > 0) return 0;
sprintf(print_buf, "\n%d\n", DIM); vx_tmc(0xff);
gemmini_config_ld(0);
gemmini_extended_config_ex(WEIGHT_STATIONARY, 0, 0, 1, 0, 0);
// load up A and B and C // load up A and B and C
uint32_t spad_A = 0x00000000; const uint32_t spad_A = 0x00000000;
uint32_t spad_B = 0x00000080; // 16B word addressed const uint32_t spad_B = 0x00000080; // 16B word addressed
uint32_t acc_C = 0x80000000; // accmem + accumulate const uint32_t acc_C = 0x80000000; // accmem + accumulate
uint32_t spad_C = 0x00000100; const uint32_t spad_C = 0x00000100;
volatile float *smem_A = (float *) SPAD_TO_SMEM(spad_A); // 0xff000000; // byte addressed volatile float *smem_A = (float *) SPAD_TO_SMEM(spad_A); // 0xff000000; // byte addressed
float *smem_B = (float *) SPAD_TO_SMEM(spad_B); // 0xff000200; float *smem_B = (float *) SPAD_TO_SMEM(spad_B); // 0xff000200;
@@ -28,33 +30,62 @@ int main() {
int J = 32 / DIM; int J = 32 / DIM;
int K = 32 / DIM; int K = 32 / DIM;
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); char *print_buf = (char *) PRINT_BUF;
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);
// int cid = vx_core_id();
int nc = vx_num_cores();
int nt = vx_num_threads();
int tid = vx_thread_id();
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(DIM * 4 * J);
sprintf(print_buf, "DIM %d\n", DIM);
sprintf(print_buf, "num cores %d\n", nc);
sprintf(print_buf, "num threads %d\n", nt);
sprintf(print_buf, "thread ids ");
vx_tmc(-1);
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 // load A with 128->1 in row-major order
for (int i = 0; i < I; i++) { for (int t = 0; t < DIM * DIM / nt; t++) {
int n = tid + t * nt;
int x = n / DIM;
int y = n % DIM;
for (int k = 0; k < K; k++) { for (int k = 0; k < K; k++) {
int tile_byte_offset = (i * K + k) * DIM * DIM; for (int i = 0; i < I; i++) {
for (int x = 0; x < DIM; x++) int tile_byte_offset = (i * K + k) * DIM * DIM;
for (int y = 0; y < DIM; y++) smem_A[tile_byte_offset + n] = (float) ((I * K * DIM * DIM - ((i * DIM + x) * DIM * K + (k * DIM + y))) % 64);
smem_A[tile_byte_offset + x * DIM + y] = (float) ((I * K * DIM * DIM - ((i * DIM + x) * DIM * K + (k * DIM + y))) % 64); // smem_A[tile_byte_offset + x * DIM + y] = (float) ((I * K * DIM * DIM - ((i * DIM + x) * DIM * K + (k * DIM + y))) % 64);
}
} }
} }
// load B with 0->191 in row-major order // load B with 0->191 in row-major order
for (int k = 0; k < K; k++) { for (int t = 0; t < DIM * DIM / nt; t++) {
for (int j = 0; j < J; j++) { int n = tid + t * nt;
int tile_byte_offset = (k * J + j) * DIM * DIM; int x = n / DIM;
for (int x = 0; x < DIM; x++) int y = n % DIM;
for (int y = 0; y < DIM; y++) for (int k = 0; k < K; k++) {
smem_B[tile_byte_offset + x * DIM + y] = (float) (((k * DIM + x) * DIM * J + (j * DIM + y)) % 64); for (int j = 0; j < J; j++) {
int tile_byte_offset = (k * J + j) * DIM * DIM;
smem_B[tile_byte_offset + n] = (float) (((k * DIM + x) * DIM * J + (j * DIM + y)) % 64);
}
// smem_B[tile_byte_offset + x * DIM + y] = (float) (((k * DIM + x) * DIM * J + (j * DIM + y)) % 64);
} }
} }
rd_cycles(end_cycles);
for (int i = 0; i < I * J * DIM * DIM; i++) smem_C[i] = 1.f; // for (int i = 0; i < I * J * DIM * DIM; i++) smem_C[i] = 1.f;
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(); fence();
@@ -65,7 +96,6 @@ int main() {
// } // }
// sprintf(print_buf, "\n"); // sprintf(print_buf, "\n");
// } // }
// sprintf(print_buf, "\nB in\n"); // sprintf(print_buf, "\nB in\n");
// for (int i = 0; i < K * DIM; i++) { // for (int i = 0; i < K * DIM; i++) {
// for (int j = 0; j < J * DIM; j++) { // for (int j = 0; j < J * DIM; j++) {
@@ -80,12 +110,6 @@ int main() {
// gemmini_extended_mvout(0xc0000000, 0xff000000, DIM, DIM); // gemmini_extended_mvout(0xc0000000, 0xff000000, DIM, DIM);
// gemmini_extended_mvout_spad(spad_C, 1, acc_C, DIM, DIM); // gemmini_extended_mvout_spad(spad_C, 1, acc_C, DIM, DIM);
uint32_t core_id;
asm volatile ("csrr %0, 0xcc2" : "=r" (core_id));
printf("core id %d\n", core_id);
if (core_id > 0) return 0;
uint32_t start_cycles, end_cycles;
asm volatile ("csrr %0, mcycle" : "=r" (start_cycles)); asm volatile ("csrr %0, mcycle" : "=r" (start_cycles));
sp_tiled_matmul_full_spad_ws(spad_A, spad_B, /*spad_D=*/0, spad_C, sp_tiled_matmul_full_spad_ws(spad_A, spad_B, /*spad_D=*/0, spad_C,
/*I=*/I, /*J=*/J, /*K=*/K, /*pad_I=*/0, /*pad_J=*/0, /*pad_K=*/0, /*I=*/I, /*J=*/J, /*K=*/K, /*pad_I=*/0, /*pad_J=*/0, /*pad_K=*/0,