249 lines
10 KiB
Plaintext
249 lines
10 KiB
Plaintext
/**
|
||
* FMHA HD=64 — PV via SS MMA with BLOCK_MN_B=16, 4 N-tiles per K-tile.
|
||
*
|
||
* For each PV K-tile: V is split into 4 N-sub-tiles of (16,16).
|
||
* 8 K-tiles × 4 N-tiles = 32 PV MMA calls.
|
||
* C output accumulates across all calls.
|
||
*
|
||
* This avoids the (64,16) V canonical layout that has 0.4% error.
|
||
*/
|
||
|
||
#include <cuda_runtime.h>
|
||
#include <cstdio>
|
||
#include <cmath>
|
||
#include <cstdlib>
|
||
#include <cstring>
|
||
|
||
#include "dsv4/kernels/attention/fmha_common.cuh"
|
||
#include "dsv4/kernels/attention/fmha_umma_desc.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 = 64, SK = 128, BLOCK_MN = 128;
|
||
constexpr int NKT_QK = HD / MMA_K_BF16; // 4
|
||
constexpr int NKT_PV = SK / MMA_K_BF16; // 8
|
||
constexpr int NNT_PV = HD / MMA_K_BF16; // 4 N-sub-tiles per K-tile
|
||
constexpr int TMEM_N = 128;
|
||
constexpr int TILE_SZ = BLOCK_MN * MMA_K_BF16;
|
||
|
||
__global__ void __launch_bounds__(128)
|
||
test_fmha_hd64_n16(const bf16_t* q, const bf16_t* k, const bf16_t* v,
|
||
bf16_t* o_out, float* o_scalar, float scale)
|
||
{
|
||
const int tid = threadIdx.x, wid = tid / 32, lane = tid % 32;
|
||
|
||
extern __shared__ char sbuf[];
|
||
uint32_t* sTmemBase = (uint32_t*)sbuf;
|
||
bf16_t* sQ0 = (bf16_t*)(((uintptr_t)(sbuf + 4) + 15) & ~(uintptr_t)15);
|
||
bf16_t* sK0 = sQ0 + NKT_QK * TILE_SZ;
|
||
bf16_t* sPk = (bf16_t*)(((uintptr_t)(sK0 + NKT_QK * TILE_SZ) + 127) & ~(uintptr_t)127);
|
||
// V: 8 K-tiles × 4 N-sub-tiles × (16,16) = 32 × 256 BF16
|
||
bf16_t* sV = (bf16_t*)(((uintptr_t)(sPk + TILE_SZ) + 127) & ~(uintptr_t)127);
|
||
float* s_p_vals = (float*)(sV + NKT_PV * NNT_PV * 256);
|
||
|
||
// Load Q, K
|
||
for (int kt = 0; kt < NKT_QK; kt++) {
|
||
bf16_t* sq = sQ0 + kt * TILE_SZ;
|
||
for (int i = tid; i < TILE_SZ; i += 128) sq[i] = 0;
|
||
for (int d = tid; d < MMA_K_BF16; d += 128) {
|
||
int ck = d / 8, lc = d % 8;
|
||
sq[ck * 16 * 64 + lc] = q[kt * MMA_K_BF16 + d];
|
||
}
|
||
}
|
||
for (int kt = 0; kt < NKT_QK; kt++) {
|
||
bf16_t* sk = sK0 + kt * TILE_SZ;
|
||
for (int i = tid; i < TILE_SZ; i += 128) sk[i] = 0;
|
||
for (int r = 0; r < SK; r++) {
|
||
for (int d = tid; d < MMA_K_BF16; d += 128) {
|
||
int ck = d / 8, lc = d % 8;
|
||
int tmn = r / 8, lr = r % 8;
|
||
sk[ck * 16 * 64 + tmn * 64 + lr * 8 + lc] = k[r * HD + kt * MMA_K_BF16 + d];
|
||
}
|
||
}
|
||
}
|
||
|
||
// Load V: 8 K-tiles × 4 N-sub-tiles of (16,16)
|
||
// V[d*SK+r] where d = nt*16+d_local, r = kt*16+lr
|
||
for (int kt = 0; kt < NKT_PV; kt++) {
|
||
for (int nt = 0; nt < NNT_PV; nt++) {
|
||
bf16_t* sv = sV + (kt * NNT_PV + nt) * 256;
|
||
for (int i = tid; i < 256; i += 128) sv[i] = 0;
|
||
for (int d_local = tid; d_local < MMA_K_BF16; d_local += 128) {
|
||
for (int lr = 0; lr < MMA_K_BF16; lr++) {
|
||
int r = kt * MMA_K_BF16 + lr;
|
||
int d = nt * MMA_K_BF16 + d_local;
|
||
int g_mn = d_local / 8, g_k = lr / 8;
|
||
int llr = d_local % 8, lc = lr % 8;
|
||
sv[g_k * 2 * 64 + g_mn * 64 + llr * 8 + lc] = v[d * SK + r];
|
||
}
|
||
}
|
||
}
|
||
}
|
||
__syncthreads();
|
||
|
||
if (wid == 1) tmem_alloc(__cvta_generic_to_shared(sTmemBase), TMEM_N);
|
||
__syncthreads();
|
||
uint32_t tb = *sTmemBase;
|
||
|
||
// QK GEMM
|
||
{
|
||
uint32_t idesc = make_idesc(BLOCK_MN, BLOCK_MN);
|
||
for (int kt = 0; kt < NKT_QK; kt++) {
|
||
bf16_t* sq = sQ0 + kt * TILE_SZ;
|
||
bf16_t* sk = sK0 + kt * TILE_SZ;
|
||
uint64_t dq = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sq), BLOCK_MN);
|
||
uint64_t dk = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sk), BLOCK_MN);
|
||
if (tid == 0) umma_ss_f16(tb, dq, dk, idesc, kt > 0);
|
||
asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
|
||
__syncthreads();
|
||
}
|
||
}
|
||
|
||
// Softmax
|
||
if (wid == 0) {
|
||
float s_vals[SK], row_max = -INFINITY;
|
||
for (int n = 0; n < SK / 8; n++) {
|
||
float tmp[8];
|
||
asm volatile("tcgen05.ld.sync.aligned.32x32b.x8.b32 {%0,%1,%2,%3,%4,%5,%6,%7},[%8];"
|
||
: "=f"(tmp[0]),"=f"(tmp[1]),"=f"(tmp[2]),"=f"(tmp[3]),
|
||
"=f"(tmp[4]),"=f"(tmp[5]),"=f"(tmp[6]),"=f"(tmp[7])
|
||
: "r"(tb + n*8));
|
||
asm volatile("tcgen05.wait::ld.sync.aligned;");
|
||
if (lane == 0) for (int c=0;c<8;c++) {
|
||
s_vals[n*8+c] = tmp[c] * scale;
|
||
row_max = fmaxf(row_max, tmp[c] * scale);
|
||
}
|
||
}
|
||
row_max = wmax(row_max);
|
||
float row_sum = 0.0f;
|
||
if (lane == 0) for (int j=0;j<SK;j++) { s_vals[j] = expf(s_vals[j] - row_max); row_sum += s_vals[j]; }
|
||
row_sum = wsum(row_sum);
|
||
if (lane == 0) for (int j=0;j<SK;j++) s_vals[j] /= row_sum;
|
||
if (lane == 0) for (int j=0;j<SK;j++) s_p_vals[j] = s_vals[j];
|
||
}
|
||
__syncthreads();
|
||
|
||
// PV GEMM: 8 K-tiles × 4 N-tiles = 32 calls
|
||
{
|
||
uint32_t idesc_pv = make_idesc(BLOCK_MN, 16); // MMA_M=8, MMA_N=1
|
||
bool first = true;
|
||
for (int kt = 0; kt < NKT_PV; kt++) {
|
||
for (int nt = 0; nt < NNT_PV; nt++) {
|
||
// Fill sPk from s_p_vals[kt*16..kt*16+15]
|
||
for (int i = tid; i < TILE_SZ; i += 128) sPk[i] = 0;
|
||
if (tid < 16) {
|
||
int c = tid;
|
||
int ck = c / 8, lc = c % 8;
|
||
sPk[ck * 16 * 64 + 0 * 64 + 0 * 8 + lc] = f32_to_bf16(s_p_vals[kt * MMA_K_BF16 + c]);
|
||
}
|
||
__syncthreads();
|
||
|
||
bf16_t* sv = sV + (kt * NNT_PV + nt) * 256;
|
||
uint64_t dp = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sPk), BLOCK_MN);
|
||
uint64_t dv = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sv), 16);
|
||
if (tid == 0) umma_ss_f16(tb + nt * 16, dp, dv, idesc_pv, !first);
|
||
asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
|
||
__syncthreads();
|
||
first = false;
|
||
}
|
||
}
|
||
}
|
||
|
||
// Epilogue
|
||
if (wid == 0) {
|
||
float o_vals[HD];
|
||
for (int n = 0; n < HD / 8; n++) {
|
||
float tmp[8];
|
||
asm volatile("tcgen05.ld.sync.aligned.32x32b.x8.b32 {%0,%1,%2,%3,%4,%5,%6,%7},[%8];"
|
||
: "=f"(tmp[0]),"=f"(tmp[1]),"=f"(tmp[2]),"=f"(tmp[3]),
|
||
"=f"(tmp[4]),"=f"(tmp[5]),"=f"(tmp[6]),"=f"(tmp[7])
|
||
: "r"(tb + n*8));
|
||
asm volatile("tcgen05.wait::ld.sync.aligned;");
|
||
if (lane == 0) for (int c=0;c<8;c++) o_vals[n*8+c] = tmp[c];
|
||
}
|
||
if (lane == 0) for (int d=0;d<HD;d++) o_out[d] = f32_to_bf16(o_vals[d]);
|
||
}
|
||
__syncthreads();
|
||
|
||
// Scalar reference
|
||
if (tid == 0) {
|
||
float s[SK];
|
||
for (int j=0;j<SK;j++) {
|
||
float dot = 0.0f;
|
||
for (int d=0;d<HD;d++) dot += bf16_to_f32(q[d]) * bf16_to_f32(k[j*HD+d]);
|
||
s[j] = dot * scale;
|
||
}
|
||
float mx = -INFINITY;
|
||
for (int j=0;j<SK;j++) mx = fmaxf(mx, s[j]);
|
||
float sm = 0.0f;
|
||
for (int j=0;j<SK;j++) { s[j] = expf(s[j]-mx); sm += s[j]; }
|
||
for (int j=0;j<SK;j++) s[j] /= sm;
|
||
for (int d=0;d<HD;d++) {
|
||
float ov = 0.0f;
|
||
for (int j=0;j<SK;j++) ov += s[j] * bf16_to_f32(v[d*SK+j]);
|
||
o_scalar[d] = ov;
|
||
}
|
||
}
|
||
|
||
if (wid == 0) tmem_dealloc(tb, TMEM_N);
|
||
}
|
||
|
||
int main() {
|
||
printf("=== FMHA HD=64 BLOCK_MN_B=16, 4 N-tiles ===\n");
|
||
const float SCALE = 1.0f / sqrtf((float)HD);
|
||
|
||
bf16_t* h_q = (bf16_t*)malloc(HD*sizeof(bf16_t));
|
||
bf16_t* h_k = (bf16_t*)malloc(SK*HD*sizeof(bf16_t));
|
||
bf16_t* h_v = (bf16_t*)malloc(HD*SK*sizeof(bf16_t));
|
||
bf16_t* h_o = (bf16_t*)calloc(HD, sizeof(bf16_t));
|
||
float* h_o_scalar = (float*)calloc(HD, sizeof(float));
|
||
|
||
srand(42);
|
||
for (int d=0;d<HD;d++) h_q[d] = f32_to_bf16_host((float)(rand()%100)/100.0f-0.5f);
|
||
for (int i=0;i<SK*HD;i++) h_k[i] = f32_to_bf16_host((float)(rand()%100)/100.0f-0.5f);
|
||
for (int i=0;i<HD*SK;i++) h_v[i] = f32_to_bf16_host((float)(rand()%100)/100.0f-0.5f);
|
||
|
||
bf16_t *d_q,*d_k,*d_v,*d_o; float *d_o_scalar;
|
||
cudaMalloc(&d_q, HD*sizeof(bf16_t));
|
||
cudaMalloc(&d_k, SK*HD*sizeof(bf16_t));
|
||
cudaMalloc(&d_v, HD*SK*sizeof(bf16_t));
|
||
cudaMalloc(&d_o, HD*sizeof(bf16_t));
|
||
cudaMalloc(&d_o_scalar, HD*sizeof(float));
|
||
cudaMemcpy(d_q, h_q, HD*sizeof(bf16_t), cudaMemcpyHostToDevice);
|
||
cudaMemcpy(d_k, h_k, SK*HD*sizeof(bf16_t), cudaMemcpyHostToDevice);
|
||
cudaMemcpy(d_v, h_v, HD*SK*sizeof(bf16_t), cudaMemcpyHostToDevice);
|
||
|
||
int smem = (4+16 + NKT_QK*TILE_SZ*2 + NKT_QK*TILE_SZ*2 + TILE_SZ*2 + NKT_PV*NNT_PV*256*2 + SK*4 + 256 + 127) & ~127;
|
||
printf("SMEM: %d bytes (%.1f KB)\n", smem, smem/1024.0f);
|
||
cudaFuncSetAttribute(test_fmha_hd64_n16, cudaFuncAttributeMaxDynamicSharedMemorySize, smem);
|
||
test_fmha_hd64_n16<<<1, 128, smem>>>(d_q, d_k, d_v, d_o, d_o_scalar, SCALE);
|
||
|
||
cudaError_t err = cudaDeviceSynchronize();
|
||
if (err != cudaSuccess) { printf("CUDA ERROR: %s\n", cudaGetErrorString(err)); return 1; }
|
||
|
||
cudaMemcpy(h_o, d_o, HD*sizeof(bf16_t), cudaMemcpyDeviceToHost);
|
||
cudaMemcpy(h_o_scalar, d_o_scalar, HD*sizeof(float), cudaMemcpyDeviceToHost);
|
||
|
||
printf("O[0..7] MMA: "); for(int d=0;d<8;d++) printf("%.6f ",bf16_to_f32_host(h_o[d])); printf("\n");
|
||
printf("O[0..7] ref: "); for(int d=0;d<8;d++) printf("%.6f ",h_o_scalar[d]); printf("\n");
|
||
|
||
float max_diff=0, max_val=0;
|
||
for (int d=0;d<HD;d++) {
|
||
max_diff = fmaxf(max_diff, fabsf(bf16_to_f32_host(h_o[d]) - h_o_scalar[d]));
|
||
max_val = fmaxf(max_val, fabsf(h_o_scalar[d]));
|
||
}
|
||
float rel_err = max_val>0 ? max_diff/max_val : max_diff;
|
||
float cos_sim=0,na=0,nb=0;
|
||
for (int d=0;d<HD;d++) { float a=bf16_to_f32_host(h_o[d]),b=h_o_scalar[d]; cos_sim+=a*b; na+=a*a; nb+=b*b; }
|
||
cos_sim /= (sqrtf(na)*sqrtf(nb)+1e-10f);
|
||
printf("Max rel err: %.8f | cosine: %.8f\n", rel_err, cos_sim);
|
||
printf("Test %s\n", cos_sim > 0.999f ? "PASSED" : "FAILED");
|
||
|
||
cudaFree(d_q); cudaFree(d_k); cudaFree(d_v); cudaFree(d_o); cudaFree(d_o_scalar);
|
||
free(h_q); free(h_k); free(h_v); free(h_o); free(h_o_scalar);
|
||
return cos_sim > 0.999f ? 0 : 1;
|
||
}
|