178 lines
7.0 KiB
Plaintext
178 lines
7.0 KiB
Plaintext
/**
|
|
* Direct comparison: load Q and K via TMA AND via direct GMEM reads,
|
|
* do MMA with both, compare TMEM output.
|
|
*/
|
|
|
|
#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 NKT = HD / MMA_K_BF16;
|
|
|
|
// Two kernels: one with TMA, one with direct loads
|
|
// Both do QK GEMM → TMEM → read out
|
|
|
|
// === DIRECT LOAD VERSION (known working pattern from fmha_6warp_multirow) ===
|
|
__global__ void __launch_bounds__(192)
|
|
test_qk_direct_kernel(float* __restrict__ out_s,
|
|
const bf16_t* __restrict__ q, const bf16_t* __restrict__ k, int T, int s_k)
|
|
{
|
|
static constexpr int TILE_SZ = 128 * MMA_K_BF16;
|
|
static constexpr int TMEM_N = (HD <= 128) ? 128 : 256;
|
|
|
|
const int tid = threadIdx.x;
|
|
const int wid = tid / 32;
|
|
const bool is_mma_warp = (wid == 4);
|
|
const bool is_softmax_warp = (wid < 4);
|
|
|
|
extern __shared__ __align__(128) char sbuf[];
|
|
size_t off = 0;
|
|
uint32_t* sTmemBase = (uint32_t*)sbuf; off = 4;
|
|
off = (off + 127) & ~(size_t)127;
|
|
bf16_t* sQ = (bf16_t*)(sbuf + off); off += 128 * HD * sizeof(bf16_t);
|
|
off = (off + 127) & ~(size_t)127;
|
|
bf16_t* sK = (bf16_t*)(sbuf + off); off += TILE_SZ * sizeof(bf16_t);
|
|
|
|
if (is_mma_warp) tmem_alloc(__cvta_generic_to_shared(sTmemBase), TMEM_N);
|
|
__syncthreads();
|
|
uint32_t tb = *sTmemBase;
|
|
|
|
// Load Q and K per K-sub-tile (same pattern as working fmha_6warp_multirow)
|
|
for (int kt = 0; kt < NKT; kt++) {
|
|
// Load Q sub-tile (128, 16) — only T rows have data
|
|
for (int i = tid; i < TILE_SZ; i += NTHREADS) sK[i] = 0; // reuse sK as temp for Q sub-tile
|
|
// Actually use a separate sQ0
|
|
// sQ0 = sK for now (they share the same buffer since we load Q first)
|
|
// Let me use the same SMEM for both Q0 and K0 since they're used sequentially
|
|
// Actually we have sQ (128*HD) and sK (128*16). Use sK for Q0 since Q0 is (128,16) too.
|
|
// But we need both Q0 and K in SMEM at the same time for MMA...
|
|
// OK let's use the first (128,16) chunk of sQ as sQ0
|
|
bf16_t* sQ0 = (bf16_t*)sbuf + 128 * 0; // overlaps with start of sQ
|
|
for (int i = tid; i < TILE_SZ; i += NTHREADS) sQ0[i] = 0;
|
|
__syncthreads();
|
|
for (int r = 0; r < T; r++) {
|
|
for (int d = threadIdx.x % 32; d < MMA_K_BF16; d += 32) {
|
|
int full_d = kt * MMA_K_BF16 + d;
|
|
if (full_d < HD) {
|
|
int ck = d/8, lc = d%8, cm = r/8, lr = r%8;
|
|
sQ0[ck*16*64 + cm*64 + lr*8 + lc] = q[r * HD + full_d];
|
|
}
|
|
}
|
|
}
|
|
__syncthreads();
|
|
|
|
// Load K sub-tile
|
|
for (int i = tid; i < TILE_SZ; i += NTHREADS) sK[i] = 0;
|
|
__syncthreads();
|
|
for (int i = tid; i < SK * MMA_K_BF16; i += NTHREADS) {
|
|
int r = i / MMA_K_BF16, c = i % MMA_K_BF16;
|
|
int gmem_c = kt * MMA_K_BF16 + c;
|
|
bf16_t val = k[r * HD + gmem_c];
|
|
int core_mn = r / 8, core_k = c / 8;
|
|
int local_r = r % 8, local_c = c % 8;
|
|
int dst_idx = core_k * 16 * 64 + core_mn * 64 + local_r * 8 + local_c;
|
|
sK[dst_idx] = val;
|
|
}
|
|
__syncthreads();
|
|
|
|
if (is_mma_warp) {
|
|
uint32_t idesc = make_idesc(128, 128);
|
|
uint64_t dq = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sQ0), 128);
|
|
uint64_t dk = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sK), 128);
|
|
if (tid == 128) umma_ss_f16(tb, dq, dk, idesc, kt > 0);
|
|
asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");
|
|
}
|
|
__syncthreads();
|
|
}
|
|
|
|
asm volatile("fence.sc.gpu;" ::: "memory");
|
|
__syncthreads();
|
|
|
|
const bool my_warp_active = (wid == 0);
|
|
const int my_row = threadIdx.x % 32;
|
|
const bool my_row_active = my_row < T;
|
|
constexpr int NUM_READS = SK / 8;
|
|
|
|
if (my_warp_active) {
|
|
for (int n = 0; n < NUM_READS; 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 (my_row_active) {
|
|
for (int c = 0; c < 8; c++) {
|
|
int col = n * 8 + c;
|
|
if (col < s_k) out_s[my_row * s_k + col] = tmp[c];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
__syncthreads();
|
|
if (is_mma_warp) tmem_dealloc(tb, TMEM_N);
|
|
}
|
|
|
|
int main() {
|
|
printf("Direct QK Test (HD=%d, SK=%d)\n", HD, SK);
|
|
const int T = 1;
|
|
|
|
bf16_t* h_q = (bf16_t*)calloc(128 * HD, sizeof(bf16_t));
|
|
bf16_t* h_k = (bf16_t*)calloc(SK * HD, sizeof(bf16_t));
|
|
srand(42);
|
|
for (int i = 0; i < T * HD; i++) h_q[i] = 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);
|
|
|
|
bf16_t *d_q, *d_k; float *d_out;
|
|
cudaMalloc(&d_q, 128 * HD * sizeof(bf16_t));
|
|
cudaMalloc(&d_k, SK * HD * sizeof(bf16_t));
|
|
cudaMalloc(&d_out, 128 * SK * sizeof(float));
|
|
cudaMemcpy(d_q, h_q, 128 * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
|
|
cudaMemcpy(d_k, h_k, SK * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
|
|
|
|
int smem = 4 + 128*HD*2 + 128*16*2 + 4096;
|
|
test_qk_direct_kernel<<<1, 192, smem>>>(d_out, d_q, d_k, T, SK);
|
|
cudaError_t err = cudaDeviceSynchronize();
|
|
if (err != cudaSuccess) { printf("CUDA ERROR: %s\n", cudaGetErrorString(err)); return 1; }
|
|
|
|
float* h_out = (float*)malloc(128 * SK * sizeof(float));
|
|
cudaMemcpy(h_out, d_out, 128 * SK * sizeof(float), cudaMemcpyDeviceToHost);
|
|
|
|
float scale = 1.0f / sqrtf((float)HD);
|
|
int fail = 0; float max_rel = 0;
|
|
for (int t = 0; t < T; t++) {
|
|
for (int j = 0; j < SK; j++) {
|
|
float dot = 0;
|
|
for (int d = 0; d < HD; d++)
|
|
dot += bf16_to_f32_host(h_q[t * HD + d]) * bf16_to_f32_host(h_k[j * HD + d]);
|
|
float ref = dot * scale;
|
|
float got = h_out[t * SK + j];
|
|
float rel = fabsf(ref) > 1e-4f ? fabsf(got - ref) / fabsf(ref) : fabsf(got - ref);
|
|
if (rel > max_rel) max_rel = rel;
|
|
if (rel > 0.01f && fail < 3) printf(" t=%d j=%d: ref=%.6f got=%.6f rel=%.4f\n", t, j, ref, got, rel);
|
|
if (rel > 0.01f) fail++;
|
|
}
|
|
}
|
|
printf("Max relative error: %.6f, failures: %d\n", max_rel, fail);
|
|
printf("%s\n", fail == 0 ? "PASSED" : "FAILED");
|
|
return fail == 0 ? 0 : 1;
|
|
}
|