void THClBlas_gemm2(THClState *state, char orderchar, char transa, char transb, long m, long n, long k, float alpha,
THClTensor *a, long lda, THClTensor *b, long ldb, float beta,
THClTensor *c, long ldc)
{
StatefulTimer::timeCheck("THClBlas_gemm START");
CLWrapper *aWrapper = THClTensor_wrapper(state, a);
CLWrapper *bWrapper = THClTensor_wrapper(state, b);
CLWrapper *cWrapper = THClTensor_wrapper(state, c);
long offseta = THClTensor_storageOffset(state, a);
long offsetb = THClTensor_storageOffset(state, b);
long offsetc = THClTensor_storageOffset(state, c);
// adjustLd(transa, transb, m, n, k, &lda, &ldb, &ldc);
clblasTranspose opa = convertTransToClblasOperation(transa);
clblasTranspose opb = convertTransToClblasOperation(transb);
clblasOrder order = orderchar == 'c' ? clblasColumnMajor : clblasRowMajor;
if( (m <= INT_MAX) && (n <= INT_MAX) && (k <= INT_MAX) && (lda <= INT_MAX) && (ldb <= INT_MAX) && (ldc <= INT_MAX) )
{
int i_m = (int)m;
int i_n = (int)n;
int i_k = (int)k;
int i_lda = (int)lda;
int i_ldb = (int)ldb;
int i_ldc = (int)ldc;
cl_int err;
if( !aWrapper->isOnDevice() ) {
aWrapper->createOnDevice();
}
if( !bWrapper->isOnDevice() ) {
bWrapper->createOnDevice();
}
if( !cWrapper->isOnDevice() ) {
cWrapper->createOnDevice();
}
EasyCL *cl = cWrapper->getCl();
cl_event *event = 0;
if(state->addFinish) {
event = new cl_event();
}
err = clblasSgemm(order, opa, opb, i_m, i_n, i_k,
alpha, aWrapper->getBuffer(), offseta, i_lda,
bWrapper->getBuffer(), offsetb, i_ldb, beta,
cWrapper->getBuffer(), offsetc, i_ldc,
1, cl->queue, 0, NULL, event);
if (err != CL_SUCCESS) {
THError("clblasSgemm() failed with %d", err);
}
else {
if(state->addFinish) {
err = clWaitForEvents(1, event);
if (err != CL_SUCCESS) {
// throw runtime_error("clblasSger() failed with " + easycl::toString(err));
THError("clblasSger: wait for event failed with %d", err);
}
clReleaseEvent(*event);
delete event;
}
}
cWrapper->markDeviceDirty();
StatefulTimer::timeCheck("THClBlas_gemm END");
return;
}
THError("Clblas_gemm only supports m, n, k, lda, ldb, ldc"
"with the bound [val] <= %d", INT_MAX);
}
void THClBlas_ger(THClState *state, long m, long n, float alpha,
THClTensor *x, long incx,
THClTensor *y, long incy,
THClTensor *a, long lda)
{
StatefulTimer::timeCheck("THClBlas_ger START");
if(n == 1)
lda = m;
if( (m <= INT_MAX) && (n <= INT_MAX) && (lda <= INT_MAX) && (incx <= INT_MAX) && (incy <= INT_MAX) )
{
int i_m = (int)m;
int i_n = (int)n;
int i_lda = (int)lda;
int i_incx = (int)incx;
int i_incy = (int)incy;
cl_int err;
CLWrapper *xwrap = THClTensor_wrapper(state, x);
CLWrapper *ywrap = THClTensor_wrapper(state, y);
CLWrapper *awrap = THClTensor_wrapper(state, a);
long x_offset = THClTensor_storageOffset(state, x);
long y_offset = THClTensor_storageOffset(state, y);
long a_offset = THClTensor_storageOffset(state, a);
if(!awrap->isOnDevice()) {
awrap->createOnDevice();
}
EasyCL *cl = ywrap->getCl();
cl_event *event = 0;
if(state->addFinish) {
event = new cl_event();
}
err = clblasSger(clblasColumnMajor, i_m, i_n, alpha,
xwrap->getBuffer(), x_offset, i_incx,
ywrap->getBuffer(), y_offset, i_incy,
awrap->getBuffer(), a_offset, i_lda,
1, (cl->queue), 0, NULL, event);
if (err != CL_SUCCESS) {
// throw runtime_error("clblasSger() failed with " + easycl::toString(err));
THError("clblasSger() failed with %d", err);
}
else {
if(state->addFinish) {
err = clWaitForEvents(1, event);
if (err != CL_SUCCESS) {
// throw runtime_error("clblasSger() failed with " + easycl::toString(err));
THError("clblasSger: wait for event failed with %d", err);
}
clReleaseEvent(*event);
delete event;
}
}
awrap->markDeviceDirty();
StatefulTimer::timeCheck("THClBlas_ger END");
return;
}
THError("Cublas_ger only supports m, n, lda, incx, incy"
"with the bound [val] <= %d", INT_MAX);
}