Skip to content

Commit

Permalink
Merge pull request #5 from caugonnet/extended_cuda_devcontainers
Browse files Browse the repository at this point in the history 
Misc C++ fixes in stf examples using CUDA libraries
  • Loading branch information
@alliepiper
alliepiper authored Oct 31, 2024
2 parents a2d6bb4 + 4fad3b0 commit 0696ec2
Show file tree
Hide file tree
Showing 11 changed files with 222 additions and 222 deletions.
22 changes: 11 additions & 11 deletions cudax/examples/stf/linear_algebra/06-pdgemm.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -160,9 +160,9 @@ public:
{
nvtxRangePushA("FILL");
// Fill blocks by blocks
for (int colb = 0; colb < nt; colb++)
for (size_t colb = 0; colb < nt; colb++)
{
for (int rowb = 0; rowb < mt; rowb++)
for (size_t rowb = 0; rowb < mt; rowb++)
{
// Each task fills a block
auto& h = get_handle(rowb, colb);
Expand Down Expand Up @@ -251,14 +251,14 @@ void PDGEMM(stream_ctx& ctx,
double beta,
matrix<double>& C)
{
for (int m = 0; m < C.mt; m++)
for (size_t m = 0; m < C.mt; m++)
{
for (int n = 0; n < C.nt; n++)
for (size_t n = 0; n < C.nt; n++)
{
//=========================================
// alpha*A*B does not contribute; scale C
//=========================================
int inner_k = transa == CUBLAS_OP_N ? A.n : A.m;
size_t inner_k = transa == CUBLAS_OP_N ? A.n : A.m;
if (alpha == 0.0 || inner_k == 0)
{
DGEMM(ctx, transa, transb, alpha, A, 0, 0, B, 0, 0, beta, C, m, n);
Expand All @@ -271,7 +271,7 @@ void PDGEMM(stream_ctx& ctx,
if (transb == CUBLAS_OP_N)
{
assert(A.nt == B.mt);
for (int k = 0; k < A.nt; k++)
for (size_t k = 0; k < A.nt; k++)
{
double zbeta = k == 0 ? beta : 1.0;
DGEMM(ctx, transa, transb, alpha, A, m, k, B, k, n, zbeta, C, m, n);
Expand All @@ -282,7 +282,7 @@ void PDGEMM(stream_ctx& ctx,
//=====================================
else
{
for (int k = 0; k < A.nt; k++)
for (size_t k = 0; k < A.nt; k++)
{
double zbeta = k == 0 ? beta : 1.0;
DGEMM(ctx, transa, transb, alpha, A, m, k, B, n, k, zbeta, C, m, n);
Expand All @@ -296,7 +296,7 @@ void PDGEMM(stream_ctx& ctx,
//=====================================
if (transb == CUBLAS_OP_N)
{
for (int k = 0; k < A.mt; k++)
for (size_t k = 0; k < A.mt; k++)
{
double zbeta = k == 0 ? beta : 1.0;
DGEMM(ctx, transa, transb, alpha, A, k, m, B, k, n, zbeta, C, m, n);
Expand All @@ -307,7 +307,7 @@ void PDGEMM(stream_ctx& ctx,
//==========================================
else
{
for (int k = 0; k < A.mt; k++)
for (size_t k = 0; k < A.mt; k++)
{
double zbeta = k == 0 ? beta : 1.0;
DGEMM(ctx, transa, transb, alpha, A, k, m, B, n, k, zbeta, C, m, n);
Expand All @@ -328,14 +328,14 @@ void run(stream_ctx& ctx, size_t N, size_t NB)
cuda_safe_call(cudaGetDeviceCount(&ndevs));

/* Warm up allocators */
for (size_t d = 0; d < ndevs; d++)
for (int d = 0; d < ndevs; d++)
{
auto lX = ctx.logical_data(shape_of<slice<double>>(1));
ctx.parallel_for(exec_place::device(d), lX.shape(), lX.write())->*[] _CCCL_DEVICE(size_t, auto) {};
}

/* Initializes CUBLAS on all devices */
for (size_t d = 0; d < ndevs; d++)
for (int d = 0; d < ndevs; d++)
{
cuda_safe_call(cudaSetDevice(d));
get_cublas_handle();
Expand Down
60 changes: 30 additions & 30 deletions cudax/examples/stf/linear_algebra/07-cholesky.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -91,10 +91,10 @@ public:

handles.resize(mt * nt);

for (int colb = 0; colb < nt; colb++)
for (size_t colb = 0; colb < nt; colb++)
{
int low_rowb = sym_matrix ? colb : 0;
for (int rowb = low_rowb; rowb < mt; rowb++)
size_t low_rowb = sym_matrix ? colb : 0;
for (size_t rowb = low_rowb; rowb < mt; rowb++)
{
T* addr_h = get_block_h(rowb, colb);
auto& h = handle(rowb, colb);
Expand Down Expand Up @@ -171,10 +171,10 @@ public:
{
nvtxRangePushA("FILL");
// Fill blocks by blocks
for (int colb = 0; colb < nt; colb++)
for (size_t colb = 0; colb < nt; colb++)
{
int low_rowb = sym_matrix ? colb : 0;
for (int rowb = low_rowb; rowb < mt; rowb++)
size_t low_rowb = sym_matrix ? colb : 0;
for (size_t rowb = low_rowb; rowb < mt; rowb++)
{
// Each task fills a block
auto& h = handle(rowb, colb);
Expand Down Expand Up @@ -363,9 +363,9 @@ void PDNRM2_HOST(matrix<double>* A, double* result)
reserved::dot::set_current_color("red");
#endif

for (int rowb = 0; rowb < A->mt; rowb++)
for (size_t rowb = 0; rowb < A->mt; rowb++)
{
for (int colb = 0; colb < A->nt; colb++)
for (size_t colb = 0; colb < A->nt; colb++)
{
ctx.host_launch(A->handle(rowb, colb).read())->*[=](auto sA) {
double res2 = 0.0;
Expand All @@ -392,24 +392,24 @@ void PDPOTRF(matrix<double>& A)
assert(A.m == A.n);
assert(A.mt == A.nt);

int NBLOCKS = A.mt;
size_t NBLOCKS = A.mt;
assert(A.mb == A.nb);

cuda_safe_call(cudaSetDevice(0));

nvtxRangePushA("SUBMIT_PDPOTRF");
for (int K = 0; K < NBLOCKS; K++)
for (size_t K = 0; K < NBLOCKS; K++)
{
int dev_akk = A.get_preferred_devid(K, K);
cuda_safe_call(cudaSetDevice(A.get_preferred_devid(K, K)));
DPOTRF(CUBLAS_FILL_MODE_LOWER, A, K, K);

for (int row = K + 1; row < NBLOCKS; row++)
for (size_t row = K + 1; row < NBLOCKS; row++)
{
cuda_safe_call(cudaSetDevice(A.get_preferred_devid(row, K)));
DTRSM(CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_LOWER, CUBLAS_OP_T, CUBLAS_DIAG_NON_UNIT, 1.0, A, K, K, A, row, K);

for (int col = K + 1; col < row; col++)
for (size_t col = K + 1; col < row; col++)
{
cuda_safe_call(cudaSetDevice(A.get_preferred_devid(row, col)));
DGEMM(CUBLAS_OP_N, CUBLAS_OP_T, -1.0, A, row, K, A, col, K, 1.0, A, row, col);
Expand Down Expand Up @@ -450,17 +450,17 @@ void PDTRSM(cublasSideMode_t side,
//===========================================
if (trans == CUBLAS_OP_N)
{
for (int k = 0; k < B.mt; k++)
for (size_t k = 0; k < B.mt; k++)
{
double lalpha = k == 0 ? alpha : 1.0;
for (int n = 0; n < B.nt; n++)
for (size_t n = 0; n < B.nt; n++)
{
cuda_safe_call(cudaSetDevice(A.get_preferred_devid(k, k)));
DTRSM(side, uplo, trans, diag, lalpha, A, k, k, B, k, n);
}
for (int m = k + 1; m < B.mt; m++)
for (size_t m = k + 1; m < B.mt; m++)
{
for (int n = 0; n < B.nt; n++)
for (size_t n = 0; n < B.nt; n++)
{
cuda_safe_call(cudaSetDevice(A.get_preferred_devid(m, k)));
DGEMM(CUBLAS_OP_N, CUBLAS_OP_N, -1.0, A, m, k, B, k, n, lalpha, B, m, n);
Expand All @@ -473,17 +473,17 @@ void PDTRSM(cublasSideMode_t side,
//================================================
else
{
for (int k = 0; k < B.mt; k++)
for (size_t k = 0; k < B.mt; k++)
{
double lalpha = k == 0 ? alpha : 1.0;
for (int n = 0; n < B.nt; n++)
for (size_t n = 0; n < B.nt; n++)
{
cuda_safe_call(cudaSetDevice(A.get_preferred_devid(B.mt - k - 1, B.mt - k - 1)));
DTRSM(side, uplo, trans, diag, lalpha, A, B.mt - k - 1, B.mt - k - 1, B, B.mt - k - 1, n);
}
for (int m = k + 1; m < B.mt; m++)
for (size_t m = k + 1; m < B.mt; m++)
{
for (int n = 0; n < B.nt; n++)
for (size_t n = 0; n < B.nt; n++)
{
cuda_safe_call(cudaSetDevice(A.get_preferred_devid(B.mt - k - 1, B.mt - 1 - m)));
DGEMM(
Expand Down Expand Up @@ -540,14 +540,14 @@ void PDGEMM(cublasOperation_t transa,
reserved::dot::set_current_color("blue");
#endif

for (int m = 0; m < C.mt; m++)
for (size_t m = 0; m < C.mt; m++)
{
for (int n = 0; n < C.nt; n++)
for (size_t n = 0; n < C.nt; n++)
{
//=========================================
// alpha*A*B does not contribute; scale C
//=========================================
int inner_k = transa == CUBLAS_OP_N ? A.n : A.m;
size_t inner_k = transa == CUBLAS_OP_N ? A.n : A.m;
if (alpha == 0.0 || inner_k == 0)
{
DGEMM(transa, transb, alpha, A, 0, 0, B, 0, 0, beta, C, m, n);
Expand All @@ -559,7 +559,7 @@ void PDGEMM(cublasOperation_t transa,
//================================
if (transb == CUBLAS_OP_N)
{
for (int k = 0; k < A.nt; k++)
for (size_t k = 0; k < A.nt; k++)
{
double zbeta = k == 0 ? beta : 1.0;
DGEMM(transa, transb, alpha, A, m, k, B, k, n, zbeta, C, m, n);
Expand All @@ -570,7 +570,7 @@ void PDGEMM(cublasOperation_t transa,
//=====================================
else
{
for (int k = 0; k < A.nt; k++)
for (size_t k = 0; k < A.nt; k++)
{
double zbeta = k == 0 ? beta : 1.0;
DGEMM(transa, transb, alpha, A, m, k, B, n, k, zbeta, C, m, n);
Expand All @@ -584,7 +584,7 @@ void PDGEMM(cublasOperation_t transa,
//=====================================
if (transb == CUBLAS_OP_N)
{
for (int k = 0; k < A.mt; k++)
for (size_t k = 0; k < A.mt; k++)
{
double zbeta = k == 0 ? beta : 1.0;
DGEMM(transa, transb, alpha, A, k, m, B, k, n, zbeta, C, m, n);
Expand All @@ -595,7 +595,7 @@ void PDGEMM(cublasOperation_t transa,
//==========================================
else
{
for (int k = 0; k < A.mt; k++)
for (size_t k = 0; k < A.mt; k++)
{
double zbeta = k == 0 ? beta : 1.0;
DGEMM(transa, transb, alpha, A, k, m, B, n, k, zbeta, C, m, n);
Expand Down Expand Up @@ -637,7 +637,7 @@ int main(int argc, char** argv)
int ndevs;
cuda_safe_call(cudaGetDeviceCount(&ndevs));

for (size_t d = 0; d < ndevs; d++)
for (int d = 0; d < ndevs; d++)
{
auto lX = ctx.logical_data(shape_of<slice<double>>(1));
ctx.parallel_for(exec_place::device(d), lX.shape(), lX.write())->*[] _CCCL_DEVICE(size_t, auto) {};
Expand Down Expand Up @@ -688,9 +688,9 @@ int main(int argc, char** argv)
cudaEvent_t startEvent_pdpotrf, stopEvent_pdpotrf;
float milliseconds_pdpotrf = 0;

// for (int row = 0; row < A.mt; row++)
// for (size_t row = 0; row < A.mt; row++)
// {
// for (int col = 0; col <= row; col++)
// for (size_t col = 0; col <= row; col++)
// {
// cuda_safe_call(cudaSetDevice(A.get_preferred_devid(row, col)));
// NOOP(A, row, col);
Expand Down
Loading

0 comments on commit 0696ec2

Please sign in to comment.