231 lines
9.0 KiB
Plaintext
231 lines
9.0 KiB
Plaintext
/**
|
|
* Test fmha_6warp_tma_multirow_multitile — multi-row + multi-tile KV + in-kernel rescale.
|
|
*/
|
|
|
|
#include <cuda_runtime.h>
|
|
#include <cuda.h>
|
|
#include <cstdio>
|
|
#include <cmath>
|
|
#include <cstdlib>
|
|
#include <cstring>
|
|
|
|
#ifndef HD_VAL
|
|
#define HD_VAL 64
|
|
#endif
|
|
|
|
#include "dsv4/kernels/attention/fmha_common.cuh"
|
|
#include "dsv4/kernels/attention/fmha_umma_desc.cuh"
|
|
#include "dsv4/kernels/attention/fmha_tma.cuh"
|
|
|
|
using namespace dsv4::kernels::attention;
|
|
|
|
static bf16_t f32_to_bf16_host(float f) { uint32_t u; memcpy(&u,&f,4); return (uint16_t)(u>>16); }
|
|
static float bf16_to_f32_host(bf16_t h) { uint32_t u=(uint32_t)h<<16; float f; memcpy(&f,&u,4); return f; }
|
|
|
|
constexpr int HD = HD_VAL;
|
|
constexpr int SK = 128;
|
|
constexpr int MAX_T = 128;
|
|
constexpr int MY_MMA_K = 16;
|
|
constexpr int TILE_SZ = 128 * MY_MMA_K;
|
|
constexpr int HD_CHUNK = (HD <= 256 ? HD : 256);
|
|
|
|
#include "dsv4/kernels/attention/fmha_6warp_tma_multirow_multitile.cuh"
|
|
|
|
static size_t compute_smem() {
|
|
size_t off = 0;
|
|
off += 4; off = (off+127)&~(size_t)127;
|
|
off += 16; off = (off+127)&~(size_t)127;
|
|
off += TILE_SZ * 2; off = (off+127)&~(size_t)127; // sTmaBuf
|
|
off += TILE_SZ * 2; off = (off+127)&~(size_t)127; // sQ0
|
|
off += TILE_SZ * 2; off = (off+127)&~(size_t)127; // sK0
|
|
off += TILE_SZ * 2; off = (off+127)&~(size_t)127; // sPk
|
|
off += 16 * MY_MMA_K * 2; // sV
|
|
off = (off+127)&~(size_t)127;
|
|
off += MAX_T * HD_CHUNK * 4; // sOacc (chunk-sized)
|
|
off += MAX_T * 4; // sRunningMax
|
|
off += MAX_T * 4; // sRunningSum
|
|
off += MAX_T * 4; // sTileRowMax
|
|
off += MAX_T * 4; // sTileRowSum
|
|
return off;
|
|
}
|
|
|
|
static void reference_attention(
|
|
const bf16_t* q, const bf16_t* k, const bf16_t* v,
|
|
float* o_ref, float* lse_ref,
|
|
int hd, int T, int s_k, float scale
|
|
) {
|
|
for (int t = 0; t < T; t++) {
|
|
float s[1024];
|
|
for (int j = 0; j < s_k; j++) {
|
|
float dot = 0.0f;
|
|
for (int d = 0; d < hd; d++) dot += bf16_to_f32_host(q[t*hd+d]) * bf16_to_f32_host(k[j*hd+d]);
|
|
s[j] = dot * scale;
|
|
}
|
|
float mx = -INFINITY;
|
|
for (int j = 0; j < s_k; j++) mx = fmaxf(mx, s[j]);
|
|
float sm = 0.0f;
|
|
for (int j = 0; j < s_k; j++) { s[j] = expf(s[j] - mx); sm += s[j]; }
|
|
for (int j = 0; j < s_k; j++) s[j] /= sm;
|
|
for (int d = 0; d < hd; d++) {
|
|
float ov = 0.0f;
|
|
for (int j = 0; j < s_k; j++) ov += s[j] * bf16_to_f32_host(v[d*s_k+j]);
|
|
o_ref[t * hd + d] = ov;
|
|
}
|
|
if (lse_ref) lse_ref[t] = logf(sm) + mx;
|
|
}
|
|
}
|
|
|
|
static int test_single(int T, int s_k, int n_h = 1, int batch = 1) {
|
|
printf("=== T=%d, s_k=%d, n_h=%d, batch=%d, HD=%d ===\n", T, s_k, n_h, batch, HD);
|
|
fflush(stdout);
|
|
const float SCALE = 1.0f / sqrtf((float)HD);
|
|
int total_heads = batch * n_h;
|
|
|
|
bf16_t* h_q = (bf16_t*)calloc(total_heads * MAX_T * HD, sizeof(bf16_t));
|
|
bf16_t* h_k = (bf16_t*)calloc(total_heads * s_k * HD, sizeof(bf16_t));
|
|
bf16_t* h_v = (bf16_t*)calloc(total_heads * HD * s_k, sizeof(bf16_t));
|
|
bf16_t* h_o = (bf16_t*)calloc(total_heads * MAX_T * HD, sizeof(bf16_t));
|
|
float* h_lse = (float*)calloc(total_heads * MAX_T, sizeof(float));
|
|
|
|
srand(42);
|
|
for (int h = 0; h < total_heads; h++) {
|
|
for (int i = 0; i < T*HD; i++) h_q[h*MAX_T*HD+i] = f32_to_bf16_host((float)(rand()%100)/100.0f-0.5f);
|
|
for (int i = 0; i < s_k*HD; i++) h_k[h*s_k*HD+i] = f32_to_bf16_host((float)(rand()%100)/100.0f-0.5f);
|
|
for (int i = 0; i < HD*s_k; i++) h_v[h*HD*s_k+i] = f32_to_bf16_host((float)(rand()%100)/100.0f-0.5f);
|
|
}
|
|
|
|
bf16_t *d_q, *d_k, *d_v, *d_o;
|
|
float *d_lse;
|
|
cudaMalloc(&d_q, total_heads * MAX_T * HD * sizeof(bf16_t));
|
|
cudaMalloc(&d_k, total_heads * s_k * HD * sizeof(bf16_t));
|
|
cudaMalloc(&d_v, total_heads * HD * s_k * sizeof(bf16_t));
|
|
cudaMalloc(&d_o, total_heads * MAX_T * HD * sizeof(bf16_t));
|
|
cudaMalloc(&d_lse, total_heads * MAX_T * sizeof(float));
|
|
printf(" cudaMalloc OK\n"); fflush(stdout);
|
|
cudaMemcpy(d_q, h_q, total_heads * MAX_T * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
|
|
cudaMemcpy(d_k, h_k, total_heads * s_k * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
|
|
cudaMemcpy(d_v, h_v, total_heads * HD * s_k * sizeof(bf16_t), cudaMemcpyHostToDevice);
|
|
cudaMemset(d_o, 0, total_heads * MAX_T * HD * sizeof(bf16_t));
|
|
cudaMemset(d_lse, 0, total_heads * MAX_T * sizeof(float));
|
|
|
|
// TMA descriptors for K: (s_k, HD) tile (128, 16)
|
|
CUtensorMap* tma_k_arr = (CUtensorMap*)malloc(total_heads * sizeof(CUtensorMap));
|
|
CUtensorMap* d_tma_k;
|
|
cudaMalloc(&d_tma_k, total_heads * sizeof(CUtensorMap));
|
|
for (int h = 0; h < total_heads; h++) {
|
|
create_tma_desc_2d_bf16(&tma_k_arr[h], d_k + h * s_k * HD, s_k, HD, 128, 16);
|
|
}
|
|
cudaMemcpy(d_tma_k, tma_k_arr, total_heads * sizeof(CUtensorMap), cudaMemcpyHostToDevice);
|
|
|
|
// TMA descriptors for V: (HD, s_k) tile (16, 16)
|
|
CUtensorMap* tma_v_arr = (CUtensorMap*)malloc(total_heads * sizeof(CUtensorMap));
|
|
CUtensorMap* d_tma_v;
|
|
cudaMalloc(&d_tma_v, total_heads * sizeof(CUtensorMap));
|
|
for (int h = 0; h < total_heads; h++) {
|
|
create_tma_desc_2d_bf16(&tma_v_arr[h], d_v + h * HD * s_k, HD, s_k, 16, 16);
|
|
}
|
|
cudaMemcpy(d_tma_v, tma_v_arr, total_heads * sizeof(CUtensorMap), cudaMemcpyHostToDevice);
|
|
|
|
FmhaTmaMultiRowMultiTileParams params;
|
|
params.q = d_q;
|
|
params.tma_k = d_tma_k;
|
|
params.tma_v = d_tma_v;
|
|
params.o = d_o;
|
|
params.lse = d_lse;
|
|
params.s_k = s_k;
|
|
params.T = T;
|
|
params.scale = SCALE;
|
|
params.n_h = n_h;
|
|
params.q_head_stride = MAX_T * HD;
|
|
params.q_batch_stride = n_h * MAX_T * HD;
|
|
params.o_head_stride = MAX_T * HD;
|
|
params.o_batch_stride = n_h * MAX_T * HD;
|
|
params.lse_head_stride = MAX_T;
|
|
params.lse_batch_stride = n_h * MAX_T;
|
|
|
|
size_t smem = compute_smem();
|
|
printf(" SMEM: %zu bytes (%.1f KB)\n", smem, smem / 1024.0); fflush(stdout);
|
|
if (smem > 48 * 1024) {
|
|
cudaError_t sae = cudaFuncSetAttribute(fmha_6warp_tma_multirow_multitile_kernel<HD>,
|
|
cudaFuncAttributeMaxDynamicSharedMemorySize, (int)smem);
|
|
printf(" cudaFuncSetAttribute: %s\n", cudaGetErrorString(sae)); fflush(stdout);
|
|
}
|
|
|
|
dim3 grid(1, n_h, batch);
|
|
printf(" Launching kernel: grid=(%d,%d,%d) smem=%zu\n", grid.x, grid.y, grid.z, smem); fflush(stdout);
|
|
fmha_6warp_tma_multirow_multitile_kernel<HD><<<grid, 192, smem>>>(params);
|
|
printf(" Kernel launched.\n");
|
|
|
|
cudaError_t lerr = cudaGetLastError();
|
|
if (lerr != cudaSuccess) {
|
|
printf(" LAUNCH ERROR: %s\n", cudaGetErrorString(lerr));
|
|
return 1;
|
|
}
|
|
cudaError_t err = cudaDeviceSynchronize();
|
|
if (err != cudaSuccess) {
|
|
printf(" CUDA ERROR: %s\n", cudaGetErrorString(err));
|
|
return 1;
|
|
}
|
|
printf(" Kernel completed OK.\n");
|
|
|
|
cudaMemcpy(h_o, d_o, total_heads * MAX_T * HD * sizeof(bf16_t), cudaMemcpyDeviceToHost);
|
|
cudaMemcpy(h_lse, d_lse, total_heads * MAX_T * sizeof(float), cudaMemcpyDeviceToHost);
|
|
|
|
int total_bad = 0;
|
|
float min_cos = 1.0f;
|
|
for (int h = 0; h < total_heads; h++) {
|
|
float* o_ref = (float*)calloc(T * HD, sizeof(float));
|
|
reference_attention(
|
|
h_q + h * MAX_T * HD, h_k + h * s_k * HD, h_v + h * HD * s_k,
|
|
o_ref, nullptr, HD, T, s_k, SCALE);
|
|
|
|
float cs = 0, na = 0, nb = 0;
|
|
int check_hd = HD;
|
|
for (int t = 0; t < T; t++) {
|
|
for (int d = 0; d < check_hd; d++) {
|
|
float a = bf16_to_f32_host(h_o[h * MAX_T * HD + t * HD + d]);
|
|
float b = o_ref[t * HD + d];
|
|
if (fabsf(b) > 1e-4f) { cs += a * b; na += a * a; nb += b * b; }
|
|
}
|
|
}
|
|
cs /= (sqrtf(na) * sqrtf(nb) + 1e-10f);
|
|
if (cs < min_cos) min_cos = cs;
|
|
free(o_ref);
|
|
}
|
|
printf(" min_cos=%.8f %s\n", min_cos, min_cos > 0.999f ? "PASS" : "FAIL");
|
|
|
|
cudaFree(d_q); cudaFree(d_k); cudaFree(d_v); cudaFree(d_o); cudaFree(d_lse);
|
|
cudaFree(d_tma_k); cudaFree(d_tma_v);
|
|
free(h_q); free(h_k); free(h_v); free(h_o); free(h_lse);
|
|
free(tma_k_arr); free(tma_v_arr);
|
|
|
|
return min_cos > 0.999f ? 0 : 1;
|
|
}
|
|
|
|
int main() {
|
|
setbuf(stdout, NULL);
|
|
printf("START: test_fmha_6warp_tma_multirow_multitile HD=%d\n", HD);
|
|
int total_fail = 0;
|
|
|
|
printf("\n=== 6-warp TMA FMHA multi-row multi-tile HD=%d ===\n", HD);
|
|
|
|
// Single KV tile (s_k=128, baseline)
|
|
for (int T : {1, 4, 32, 128}) {
|
|
total_fail += test_single(T, 128);
|
|
}
|
|
|
|
// Multi-tile KV — the whole point of D1.5
|
|
for (int s_k : {256, 384, 512}) {
|
|
for (int T : {1, 4, 32, 128}) {
|
|
total_fail += test_single(T, s_k);
|
|
}
|
|
}
|
|
|
|
// Multi-head + batch
|
|
total_fail += test_single(4, 256, 4, 1); // n_h=4
|
|
total_fail += test_single(4, 256, 2, 2); // n_h=2, batch=2
|
|
|
|
printf("\nOverall: %s\n", total_fail == 0 ? "ALL PASSED" : "SOME FAILED");
|
|
return total_fail == 0 ? 0 : 1;
|
|
}
|