Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[SYCLomatic #2114] Add more test for cublasLt API migration #734

Merged
merged 4 commits into from
Jul 19, 2024
Merged

[SYCLomatic #2114] Add more test for cublasLt API migration #734

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
cc51809
[SYCLomatic #2114] Add more test for cublasLt API migration
zhiweij1 Jul 4, 2024
b5dcbc7
Update
zhiweij1 Jul 10, 2024
fa9f8da
Remove amax
zhiweij1 Jul 10, 2024
f413d2e
Add amax
zhiweij1 Jul 11, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
145 changes: 145 additions & 0 deletions features/feature_case/cublasLt/matmul.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -722,6 +722,150 @@ bool test6() {
return !error;
}

void fgemmlt(cublasLtHandle_t ltHandle, int m, int n, int k,
const float *A, const float *B, const float *C, float *D,
float *alpha, float *beta,
int lda, int ldb, int ldc, int ldd,
cublasLtMatrixLayout_t Adesc,
cublasLtMatrixLayout_t Bdesc,
cublasLtMatrixLayout_t Cdesc,
cublasLtMatrixLayout_t Ddesc,
float *amax_d) {
cublasLtMatmulDesc_t matmulDesc = NULL;
cublasLtMatmulDescCreate(&matmulDesc, CUBLAS_COMPUTE_32F, CUDA_R_32F);

float *scale_a;
float *scale_b;
float *scale_d;
cudaMallocManaged(&scale_a, sizeof(float));
cudaMallocManaged(&scale_b, sizeof(float));
cudaMallocManaged(&scale_d, sizeof(float));
scale_a[0] = 3;
scale_b[0] = 5;
scale_d[0] = 7;

cublasLtMatmulDescSetAttribute(matmulDesc, CUBLASLT_MATMUL_DESC_A_SCALE_POINTER, &scale_a, sizeof(scale_a));
cublasLtMatmulDescSetAttribute(matmulDesc, CUBLASLT_MATMUL_DESC_B_SCALE_POINTER, &scale_b, sizeof(scale_b));
cublasLtMatmulDescSetAttribute(matmulDesc, CUBLASLT_MATMUL_DESC_D_SCALE_POINTER, &scale_d, sizeof(scale_d));
cublasLtMatmulDescSetAttribute(matmulDesc, CUBLASLT_MATMUL_DESC_AMAX_D_POINTER, &amax_d, sizeof(amax_d));

cublasLtEpilogue_t ep = CUBLASLT_EPILOGUE_RELU;
cublasLtMatmulDescSetAttribute(matmulDesc, CUBLASLT_MATMUL_DESC_EPILOGUE, &ep, sizeof(ep));

cublasLtMatmul(ltHandle, matmulDesc, alpha, A, Adesc, B, Bdesc, beta, C, Cdesc, D, Ddesc, NULL, NULL, 0, 0);

cudaStreamSynchronize(0);
cublasLtMatmulDescDestroy(matmulDesc);
}

// clang-format off
// A (4*3) B (3*2)
// 6 10 14 5 4
// 7 11 15 -3 -2
// 8 12 16 1 0
// 9 13 17 p p
//
// alpha * A * B + C = alpha * A*B + C = D
// 2*3*5 6 10 14 5 4 -10000 -5000 30 14 4 -10000 -5000 -9580 -4880
// 7 11 15 -3 -2 2000 6000 17 6 2000 6000 2510 6180
// 8 12 16 1 0 3000 7000 20 8 3000 7000 3600 7240
// 9 13 17 p p 4000 8000 23 10 4000 8000 4690 8300
// scale_d * D = D
// 7 * -9580 -4880 -67060 -34160
// 2510 6180 17570 43260
// 3600 7240 25200 50680
// 4690 8300 32830 58100
// clang-format on

bool test7() {
cublasLtHandle_t ltHandle;
cublasLtCreate(&ltHandle);
const constexpr int m = 4;
const constexpr int n = 2;
const constexpr int k = 3;
const constexpr int lda = m;
const constexpr int ldb = m;
const constexpr int ldc = m;
const constexpr int ldd = m;
void *Adev;
void *Bdev;
void *Cdev;
void *Ddev;
cudaMalloc(&Adev, lda * k * sizeof(float));
cudaMalloc(&Bdev, ldb * n * sizeof(float));
cudaMalloc(&Cdev, ldc * n * sizeof(float));
cudaMalloc(&Ddev, ldd * n * sizeof(float));

float Ahost[lda * k] = {6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17};
float Bhost[ldb * n] = {5, -3, 1, 99, 4, -2, 0, 99};
float Chost[ldc * n] = {-1000, 2000, 3000, 4000, -5000, 6000, 7000, 8000};

cudaMemcpy(Adev, Ahost, lda * k * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(Bdev, Bhost, ldb * n * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(Cdev, Chost, ldc * n * sizeof(float), cudaMemcpyHostToDevice);

cublasLtMatrixLayout_t Adesc_col_major = NULL,
Bdesc_col_major = NULL,
Cdesc_col_major = NULL,
Ddesc_col_major = NULL;
cublasLtMatrixLayoutCreate(&Adesc_col_major, CUDA_R_32F, m, k, lda);
cublasLtMatrixLayoutCreate(&Bdesc_col_major, CUDA_R_32F, k, n, ldb);
cublasLtMatrixLayoutCreate(&Cdesc_col_major, CUDA_R_32F, m, n, ldc);
cublasLtMatrixLayoutCreate(&Ddesc_col_major, CUDA_R_32F, m, n, ldd);

float alpha = 2;
float beta = 1;

// Matmul

float *amax_d;
cudaMallocManaged(&amax_d, sizeof(float));

fgemmlt(ltHandle, m, n, k, (const float *)Adev, (const float *)Bdev, (const float *)Cdev, (float *)Ddev,
&alpha, &beta, lda, ldb, ldc, ldd, Adesc_col_major, Bdesc_col_major, Cdesc_col_major, Ddesc_col_major, amax_d);
cudaStreamSynchronize(0);

// Check result
float Dhost[ldd * n];
cudaMemcpy(Dhost, Ddev, ldd * n * sizeof(float), cudaMemcpyDeviceToHost);

bool error = false;
float D_ref[ldd * n] = {0, 17570, 25200, 32830, 0, 43260, 50680, 58100};
for (int i = 0; i < ldd * n; i++) {
if (Dhost[i] != D_ref[i]) {
error = true;
break;
}
}
if (*amax_d != 8300)
error = true;

printf("d:\n");
for (int i = 0; i < ldd * n; i++)
printf("%f, ", Dhost[i]);
printf("\n");
printf("amax_d:%f\n", *amax_d);

if (error) {
printf("error\n");
} else {
printf("success\n");
}

cublasLtDestroy(ltHandle);
cublasLtMatrixLayoutDestroy(Adesc_col_major);
cublasLtMatrixLayoutDestroy(Bdesc_col_major);
cublasLtMatrixLayoutDestroy(Cdesc_col_major);
cublasLtMatrixLayoutDestroy(Ddesc_col_major);
cudaFree(Adev);
cudaFree(Bdev);
cudaFree(Ddev);
cudaFree(amax_d);

return !error;
}


// clang-format off
// A (4*3) B (2*3)
// 6 10 14 5 -3 1
Expand Down Expand Up @@ -750,5 +894,6 @@ int main() {
pass = test4() && pass;
pass = test5() && pass;
pass = test6() && pass;
pass = test7() && pass;
return pass ? 0 : 1;
}