Files
kernels/tests/regression/sgemm_gemmini_dma/kernel.cpp

206 lines
7.0 KiB
C++

#include <stdint.h>
#include <vx_intrinsics.h>
#include <vx_print.h>
#include <vx_spawn.h>
#include "common.h"
#include "include/gemmini.h"
#include "gemmini_mmio.h"
// fp16 16x16
#define TILE_M 128
#define TILE_N 64
#define TILE_K 128
// ampere
// #define NUM_THREADS_IN_CLUSTER 512
// hopper
#define NUM_THREADS_IN_CLUSTER 256
// fp32 8x8
// #define TILE_M 64
// #define TILE_N 64
// #define TILE_K 64
// #define SMEM_ADDR_Q0 ((float * const) 0xff000000)
// #define SMEM_ADDR_Q1 ((float * const) 0xff004000)
// #define SMEM_ADDR_Q2 ((float * const) 0xff008000)
// #define SMEM_ADDR_Q3 ((float * const) 0xff00c000)
// #define SPAD_ADDR_Q0 0x0
// #define SPAD_ADDR_Q1 0x200
// #define SPAD_ADDR_Q2 0x400
// #define SPAD_ADDR_Q3 0x600
// #define NUM_THREADS_IN_CLUSTER 256
// fp32 4x4
// #define TILE_M 32
// #define TILE_N 32
// #define TILE_K 32
// #define SMEM_ADDR_Q0 ((float * const) 0xff000000)
// #define SMEM_ADDR_Q1 ((float * const) 0xff001000)
// #define SMEM_ADDR_Q2 ((float * const) 0xff002000)
// #define SMEM_ADDR_Q3 ((float * const) 0xff003000)
// #define SPAD_ADDR_Q0 0x0
// #define SPAD_ADDR_Q1 0x80
// #define SPAD_ADDR_Q2 0x100
// #define SPAD_ADDR_Q3 0x180
// #define NUM_THREADS_IN_CLUSTER 256
#define NUM_CLUSTERS 1
// (NUM_CORES * NUM_WARPS * NUM_THREADS)
#define rd_cycles_force(x) asm volatile ("csrr %0, mcycle" : "=r" (x))
#define rd_cycles(x) rd_cycles_force(x)
#define HW_TID() ({uint32_t gtid; asm volatile ("csrr %0, mhartid" : "=r" (gtid)); gtid;})
#define PRINTF(...) sprintf(PRINT_BUF, __VA_ARGS__)
// #define PRINTF(...) vx_printf(__VA_ARGS__)
#define SWISH(beta, x) ((x) / (1 + exp(-(beta) * (x))))
// #define POWER
typedef uint16_t smem_elem_t;
// typedef float smem_elem_t;
inline void threadblock_barrier(unsigned int barrier_id, unsigned int count) __attribute__((convergent)) {
vx_fence();
vx_barrier(barrier_id, count);
}
void thread_block_matmul_gemmini(kernel_arg_t *__UNIFORM__ arg,
const uint32_t threadblock_id,
const uint32_t tid_in_threadblock) {
asm volatile ("matmul_start_%=:" :: );
const smem_elem_t * const A = (const smem_elem_t * const) arg->addr_a;
const smem_elem_t * const B = (const smem_elem_t * const) arg->addr_b;
smem_elem_t * const C = (smem_elem_t * const) arg->addr_c;
if (HW_TID() == 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);
#ifndef POWER
PRINTF("start\n");
#endif
}
vx_fence();
uint32_t marker0, marker1;
rd_cycles_force(marker0);
MARK_BEG();
const uint32_t dim_m = arg->dim_m;
const uint32_t dim_n = arg->dim_n;
const uint32_t dim_k = arg->dim_k;
const uint32_t num_tiles_m = dim_m / TILE_M;
const uint32_t num_tiles_n = dim_n / TILE_N;
const uint32_t num_tiles_k = dim_k / TILE_K;
constexpr uint32_t num_threads_in_cluster = NUM_THREADS_IN_CLUSTER;
const uint32_t num_tile_rows_per_tb = num_tiles_m / NUM_CLUSTERS;
if (HW_TID() == 0) {
gemmini_extended3_config_ld(dim_k * sizeof(elem_t), MVIN_SCALE_IDENTITY, false, 0);
gemmini_extended3_config_ld(dim_n * sizeof(elem_t), MVIN_SCALE_IDENTITY, false, 1);
gemmini_extended_config_st(dim_n * sizeof(elem_t), 0, MVIN_SCALE_IDENTITY);
// gemmini_extended_config_st(stride_C * sizeof_C, act & 3, scale);
for (uint32_t tile_i = num_tile_rows_per_tb * threadblock_id;
tile_i < num_tile_rows_per_tb * (threadblock_id + 1);
tile_i += 1) {
for (uint32_t tile_j = 0; tile_j < num_tiles_n; tile_j += 1) {
for (uint32_t tile_k = 0; tile_k < num_tiles_k; tile_k += 1) {
uint32_t a_hexadecile = (tile_k & 1) << 2;
uint32_t b_hexadecile = a_hexadecile + 11;
gemmini_tile_load_ab(A, B,
a_hexadecile, b_hexadecile, tile_i, tile_j, tile_k,
dim_m, dim_n, dim_k, TILE_M, TILE_N, TILE_K);
/* DO STUFF */
gemmini_fence();
gemmini_tile_compute</*store_to_spad=*/false>(
a_hexadecile, b_hexadecile, 0 /*d_hexadecile*/, tile_k > 0);
}
/*
gemmini_fence();
gemmini_tile_store_c_spad(a_hexadecile); // then activate in spad
*/
gemmini_fence();
gemmini_tile_store_c_gmem(C, tile_i, tile_j, dim_m, dim_n, TILE_M, TILE_N);
}
}
MARK_END();
rd_cycles_force(marker1);
#ifndef POWER
if (HW_TID() == 0) {
PRINTF("\ncomplete\n");
PRINTF("total cycles: %d\n", marker1 - marker0);
for (int i = 0; i < dim_m; i += 8) {
for (int j = 0; j < dim_n; j += 8) {
PRINTF("%04x %04x ", (int) (C[i * dim_n + j]), (int) (C[i * dim_n + j + 4]));
}
PRINTF("\n");
}
}
#endif
} else {
if (HW_TID() > 8) {
asm volatile("li x1, 0xa0a0a0a0");
asm volatile("li x2, 0xa0a0a0a0");
asm volatile("li x3, 0xa0a0a0a0");
asm volatile("li x4, 0xa0a0a0a0");
asm volatile("li x5, 0xa0a0a0a0");
asm volatile("li x6, 0xa0a0a0a0");
asm volatile("li x7, 0xa0a0a0a0");
asm volatile("li x8, 0xa0a0a0a0");
asm volatile("li x9, 0xa0a0a0a0");
asm volatile("li x10, 0xa0a0a0a0");
asm volatile("li x11, 0xa0a0a0a0");
asm volatile("li x12, 0xa0a0a0a0");
asm volatile("li x13, 0xa0a0a0a0");
asm volatile("li x14, 0xa0a0a0a0");
asm volatile("li x15, 0xa0a0a0a0");
asm volatile("li x16, 0xa0a0a0a0");
asm volatile("li x17, 0xa0a0a0a0");
asm volatile("li x18, 0xa0a0a0a0");
asm volatile("li x19, 0xa0a0a0a0");
asm volatile("li x20, 0xa0a0a0a0");
asm volatile("li x21, 0xa0a0a0a0");
asm volatile("li x22, 0xa0a0a0a0");
asm volatile("li x23, 0xa0a0a0a0");
asm volatile("li x24, 0xa0a0a0a0");
asm volatile("li x25, 0xa0a0a0a0");
asm volatile("li x26, 0xa0a0a0a0");
asm volatile("li x27, 0xa0a0a0a0");
asm volatile("li x28, 0xa0a0a0a0");
asm volatile("li x29, 0xa0a0a0a0");
asm volatile("li x30, 0xa0a0a0a0");
asm volatile("li x31, 0xa0a0a0a0");
asm volatile("vx_tmc zero");
}
}
vx_fence();
// threadblock_barrier(/*barrier_id=*/0, /*count=*/NUM_WARPS);
vx_tmc(0);
}
void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
const int threadblock_id = task_id / NUM_THREADS_IN_CLUSTER;
const int tid_in_threadblock = task_id % NUM_THREADS_IN_CLUSTER;
thread_block_matmul_gemmini(arg, threadblock_id, tid_in_threadblock);
}
int main() {
kernel_arg_t *arg = (kernel_arg_t *)KERNEL_ARG_DEV_MEM_ADDR;
const uint32_t num_threads_in_cluster = NUM_THREADS_IN_CLUSTER;
const uint32_t grid_size = num_threads_in_cluster * NUM_CLUSTERS;
#ifdef RADIANCE
vx_spawn_tasks_cluster(grid_size, (vx_spawn_tasks_cb)kernel_body, arg);
#else
// NOTE: This kernel assumes contiguous thread scheduling for efficient shared
// memory allocation, and therefore does not work with original vx_spawn_tasks
vx_spawn_tasks_contiguous(grid_size, (vx_spawn_tasks_cb)kernel_body, arg);
#endif
return 0;
}