biondizzle
a40c05f3f2
Key findings documented in docs/cuda13_tma_notes.md: - CUDA 13 globalStrides are in BYTES not elements (root cause of desc creation failures) - BFLOAT16 data type available in CUDA 13 - Driver API descriptors create OK but cp.async.bulk.tensor hangs on driver 13.0 + toolkit 13.2 - CuTeDSL tma_partition works (production path) Archived (not deleted): - fmha_tma_driver_api.cuh, fmha_6warp_tma_driver_api.cuh, test_fmha_tma_driver_api.cu - These will work once driver matches toolkit version
284 lines
12 KiB
Plaintext
284 lines
12 KiB
Plaintext
/**
|
||
* Test TMA async FMHA kernel (6-warp, multi-row, TMA loads).
|
||
* Compile with -DHD_VAL=64 etc.
|
||
*
|
||
* TMA descriptor layout:
|
||
* Q: (T, HD), tile = (16, 128) — loads (128, 16) sub-tiles, coord = (kt*16, 0)
|
||
* K: (s_k, HD), tile = (16, 128) — loads (128, 16) sub-tiles, coord = (kt*16, 0)
|
||
* V: (HD, s_k), tile = (128, 16) — loads (16, 128) sub-tiles, coord = (col_start, d_base)
|
||
*
|
||
* Note: TMA innermost dimension = columns. A tile of (16, 128) means
|
||
* 16 columns × 128 rows. The coord is (col_offset, row_offset).
|
||
*/
|
||
|
||
#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;
|
||
|
||
#include "dsv4/kernels/attention/fmha_6warp_tma.cuh"
|
||
|
||
static int compute_smem_tma() {
|
||
size_t off = 0;
|
||
off += 4; // sTmemBase
|
||
off = (off + 127) & ~(size_t)127;
|
||
off += 8; // sMbar
|
||
off += MAX_T * sizeof(float); // sRowMax
|
||
off += MAX_T * sizeof(float); // sRowSum
|
||
off = (off + 127) & ~(size_t)127;
|
||
off += 128 * MMA_K_BF16 * sizeof(bf16_t); // sQ_tma (128×16)
|
||
off = (off + 127) & ~(size_t)127;
|
||
off += 128 * MMA_K_BF16 * sizeof(bf16_t); // sK_tma (128×16)
|
||
off = (off + 127) & ~(size_t)127;
|
||
off += 128 * MMA_K_BF16 * sizeof(bf16_t); // sQ canonical
|
||
off = (off + 127) & ~(size_t)127;
|
||
off += 128 * MMA_K_BF16 * sizeof(bf16_t); // sK canonical
|
||
off = (off + 127) & ~(size_t)127;
|
||
off += 128 * MMA_K_BF16 * sizeof(bf16_t); // sPk canonical
|
||
off = (off + 127) & ~(size_t)127;
|
||
off += 16 * 128 * sizeof(bf16_t); // sV_tma (16×128)
|
||
off = (off + 127) & ~(size_t)127;
|
||
off += 128 * MMA_K_BF16 * sizeof(bf16_t); // sV canonical
|
||
return (int)off;
|
||
}
|
||
|
||
static void reference_attention_multirow(
|
||
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[512];
|
||
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;
|
||
}
|
||
}
|
||
|
||
struct TmaDescSet {
|
||
CUtensorMap tma_q, tma_k, tma_v;
|
||
CUtensorMap *d_tma_q, *d_tma_k, *d_tma_v;
|
||
|
||
bool create(bf16_t* d_q, bf16_t* d_k, bf16_t* d_v,
|
||
int T, int hd, int s_k) {
|
||
// Q: (T, HD) row-major. TMA tile = (16, 128) → innermost 16 cols, 128 rows.
|
||
// We load sub-tiles of Q at column offsets kt*16.
|
||
// For T < 128: TMA will read OOB rows beyond T, but the data is garbage.
|
||
// The kernel masks via the canonical transpose which zeros rows ≥ T.
|
||
// Actually, we should set the global dims to the ACTUAL tensor size (T, HD).
|
||
// TMA handles OOB by returning zeros (with FLOAT_OOB_FILL_NONE).
|
||
// But TMA requires tile ≤ global dims. So for T < 128, the tile (16, 128)
|
||
// has 128 rows > T rows. TMA REQUIRES tile ≤ global dims.
|
||
// Solution: pad the GMEM allocation to at least (128, HD) per head.
|
||
// OR: use a tile of (16, T) when T < 128. But tile dims must be power of 2.
|
||
// Simplest: always pad to (128, HD) in GMEM allocation.
|
||
// Since this is a test, we'll allocate (128, HD) for Q and (128, HD) for K,
|
||
// and (HD, 128) for V — all padded.
|
||
|
||
// Actually, let's just make the global dimensions match the tile dimensions.
|
||
// Q global dim = (HD, 128) innermost-first, which means (128 rows, HD cols).
|
||
// But the actual data is (T, HD). For T < 128, the extra rows are undefined.
|
||
// The kernel handles this via the canonical transpose zeroing.
|
||
//
|
||
// Hmm, but cuTensorMapEncodeTiled validates that global dims >= tile dims.
|
||
// If global_dim[1] = T < 128 = tile_dim[1], it fails.
|
||
//
|
||
// The real fix: allocate padded buffers. Let me do that in the test.
|
||
|
||
// Q: global (HD, 128) → 128 rows, HD cols. Tile (16, 128).
|
||
if (!create_tma_desc_2d_bf16(&tma_q, d_q, 128, (uint64_t)hd, 128, 16)) {
|
||
printf(" Failed to create Q TMA desc: rows=128, cols=%d, tile=128x16\n", hd);
|
||
return false;
|
||
}
|
||
|
||
// K: (128, HD). Tile (16, 128).
|
||
if (!create_tma_desc_2d_bf16(&tma_k, d_k, (uint64_t)s_k, (uint64_t)hd, 128, 16)) {
|
||
printf(" Failed to create K TMA desc\n");
|
||
return false;
|
||
}
|
||
|
||
// V: (HD, s_k). Tile (128, 16).
|
||
if (!create_tma_desc_2d_bf16(&tma_v, d_v, (uint64_t)hd, (uint64_t)s_k, 128, 16)) {
|
||
printf(" Failed to create V TMA desc\n");
|
||
return false;
|
||
}
|
||
|
||
cudaMalloc(&d_tma_q, sizeof(CUtensorMap));
|
||
cudaMalloc(&d_tma_k, sizeof(CUtensorMap));
|
||
cudaMalloc(&d_tma_v, sizeof(CUtensorMap));
|
||
cudaMemcpy(d_tma_q, &tma_q, sizeof(CUtensorMap), cudaMemcpyHostToDevice);
|
||
cudaMemcpy(d_tma_k, &tma_k, sizeof(CUtensorMap), cudaMemcpyHostToDevice);
|
||
cudaMemcpy(d_tma_v, &tma_v, sizeof(CUtensorMap), cudaMemcpyHostToDevice);
|
||
return true;
|
||
}
|
||
|
||
void destroy() {
|
||
if (d_tma_q) { cudaFree(d_tma_q); d_tma_q = nullptr; }
|
||
if (d_tma_k) { cudaFree(d_tma_k); d_tma_k = nullptr; }
|
||
if (d_tma_v) { cudaFree(d_tma_v); d_tma_v = nullptr; }
|
||
}
|
||
};
|
||
|
||
static int test_single(int T, int n_h = 1, int batch = 1) {
|
||
printf("\n=== TMA T=%d, n_h=%d, batch=%d, HD=%d ===\n", T, n_h, batch, HD);
|
||
const float SCALE = 1.0f / sqrtf((float)HD);
|
||
int total_heads = batch * n_h;
|
||
|
||
// Allocate PADDED buffers: Q is (128, HD) per head, K is (s_k, HD), V is (HD, s_k)
|
||
// Q needs padding because TMA reads (128, 16) tiles — must have 128 rows in GMEM
|
||
constexpr int Q_PAD_ROWS = 128;
|
||
bf16_t* h_q = (bf16_t*)calloc(total_heads * Q_PAD_ROWS * HD, sizeof(bf16_t));
|
||
bf16_t* h_k = (bf16_t*)malloc(total_heads * SK * HD * sizeof(bf16_t));
|
||
bf16_t* h_v = (bf16_t*)malloc(total_heads * HD * SK * 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 + T);
|
||
// Fill Q data only for T rows
|
||
for (int i = 0; i < total_heads * T * HD; i++) h_q[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
|
||
for (int i = 0; i < total_heads * SK * HD; i++) h_k[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
|
||
for (int i = 0; i < total_heads * HD * SK; i++) h_v[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
|
||
|
||
// Device allocations — padded for TMA
|
||
bf16_t *d_q, *d_k, *d_v, *d_o; float *d_lse;
|
||
cudaMalloc(&d_q, total_heads * Q_PAD_ROWS * HD * sizeof(bf16_t)); // padded
|
||
cudaMalloc(&d_k, total_heads * SK * HD * sizeof(bf16_t));
|
||
cudaMalloc(&d_v, total_heads * HD * SK * sizeof(bf16_t));
|
||
cudaMalloc(&d_o, total_heads * MAX_T * HD * sizeof(bf16_t));
|
||
cudaMalloc(&d_lse, total_heads * MAX_T * sizeof(float));
|
||
cudaMemcpy(d_q, h_q, total_heads * Q_PAD_ROWS * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
|
||
cudaMemcpy(d_k, h_k, total_heads * SK * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
|
||
cudaMemcpy(d_v, h_v, total_heads * HD * SK * sizeof(bf16_t), cudaMemcpyHostToDevice);
|
||
|
||
int failed = 0;
|
||
float min_cos = 1.0f;
|
||
|
||
for (int b = 0; b < batch; b++) {
|
||
for (int h = 0; h < n_h; h++) {
|
||
int idx = b * n_h + h;
|
||
|
||
TmaDescSet tma;
|
||
bf16_t* d_q_h = d_q + idx * Q_PAD_ROWS * HD; // padded stride
|
||
bf16_t* d_k_h = d_k + idx * SK * HD;
|
||
bf16_t* d_v_h = d_v + idx * HD * SK;
|
||
if (!tma.create(d_q_h, d_k_h, d_v_h, T, HD, SK)) {
|
||
printf(" TMA desc creation failed for head %d batch %d\n", h, b);
|
||
failed++; continue;
|
||
}
|
||
|
||
FmhaMultiRowTmaParams params;
|
||
params.q = d_q_h; params.k = d_k_h; params.v = d_v_h;
|
||
params.o = d_o + idx * MAX_T * HD; params.lse = d_lse + idx * MAX_T;
|
||
params.s_k = SK; params.T = T; params.scale = SCALE; params.head_dim = HD;
|
||
// Strides must match PADDED allocation
|
||
params.q_head_stride = Q_PAD_ROWS * HD; params.q_batch_stride = n_h * Q_PAD_ROWS * HD;
|
||
params.k_head_stride = SK * HD; params.k_batch_stride = n_h * SK * HD;
|
||
params.v_head_stride = HD * SK; params.v_batch_stride = n_h * HD * SK;
|
||
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;
|
||
params.tma_q = tma.d_tma_q;
|
||
params.tma_k = tma.d_tma_k;
|
||
params.tma_v = tma.d_tma_v;
|
||
|
||
int smem = compute_smem_tma();
|
||
if (smem > 48 * 1024)
|
||
cudaFuncSetAttribute(fmha_6warp_tma_kernel<HD>, cudaFuncAttributeMaxDynamicSharedMemorySize, smem);
|
||
|
||
dim3 grid(1, 1, 1);
|
||
fmha_6warp_tma_kernel<HD><<<grid, 192, smem>>>(params);
|
||
|
||
cudaError_t err = cudaDeviceSynchronize();
|
||
if (err != cudaSuccess) {
|
||
printf(" CUDA ERROR b=%d h=%d: %s\n", b, h, cudaGetErrorString(err));
|
||
failed++; tma.destroy(); continue;
|
||
}
|
||
|
||
// Verify against reference using UN-padded Q data
|
||
bf16_t* h_o_head = (bf16_t*)malloc(T * HD * sizeof(bf16_t));
|
||
float* h_lse_head = (float*)malloc(T * sizeof(float));
|
||
cudaMemcpy(h_o_head, d_o + idx * MAX_T * HD, T * HD * sizeof(bf16_t), cudaMemcpyDeviceToHost);
|
||
cudaMemcpy(h_lse_head, d_lse + idx * MAX_T, T * sizeof(float), cudaMemcpyDeviceToHost);
|
||
|
||
float o_ref[MAX_T * 512]; float lse_ref[MAX_T];
|
||
reference_attention_multirow(
|
||
h_q + idx * Q_PAD_ROWS * HD, // padded but first T rows have data
|
||
h_k + idx * SK * HD, h_v + idx * HD * SK,
|
||
o_ref, lse_ref, HD, T, SK, SCALE);
|
||
|
||
for (int t = 0; t < T; t++) {
|
||
float cs=0,na=0,nb=0;
|
||
for (int d=0;d<HD;d++) {
|
||
float a=bf16_to_f32_host(h_o_head[t*HD+d]), b2=o_ref[t*HD+d];
|
||
if(fabsf(b2)>1e-4f){cs+=a*b2;na+=a*a;nb+=b2*b2;}
|
||
}
|
||
cs /= (sqrtf(na)*sqrtf(nb)+1e-10f);
|
||
if(cs<min_cos) min_cos=cs;
|
||
if(cs<0.999f) { printf(" FAIL b=%d h=%d t=%d cos=%.6f\n",b,h,t,cs); failed++; }
|
||
float lse_err = fabsf(h_lse_head[t] - lse_ref[t]);
|
||
if(lse_err > 0.01f) { printf(" FAIL LSE b=%d h=%d t=%d kernel=%.6f ref=%.6f err=%.6f\n",b,h,t,h_lse_head[t],lse_ref[t],lse_err); failed++; }
|
||
}
|
||
|
||
free(h_o_head); free(h_lse_head);
|
||
tma.destroy();
|
||
}
|
||
}
|
||
|
||
printf(" min_cos=%.8f %s\n", min_cos, failed==0?"PASSED":"FAILED");
|
||
|
||
cudaFree(d_q); cudaFree(d_k); cudaFree(d_v); cudaFree(d_o); cudaFree(d_lse);
|
||
free(h_q); free(h_k); free(h_v); free(h_o); free(h_lse);
|
||
return failed == 0;
|
||
}
|
||
|
||
int main() {
|
||
printf("TMA Async FMHA test (HD=%d)\n", HD);
|
||
|
||
int ok = 1;
|
||
|
||
printf("\n--- Single KV tile tests (TMA) ---\n");
|
||
ok &= test_single(128); // Test T=128 first (simplest — no padding issues)
|
||
ok &= test_single(64);
|
||
ok &= test_single(32);
|
||
ok &= test_single(16);
|
||
ok &= test_single(1);
|
||
|
||
printf("\n--- Multi-head tests (TMA) ---\n");
|
||
ok &= test_single(4, 4, 1);
|
||
|
||
printf("\n%s\n", ok ? "ALL PASSED" : "SOME FAILED");
|
||
return ok ? 0 : 1;
|
||
}
|