static int sgemv(cb_order order, cb_transpose transA, size_t M, size_t N,
float alpha, gpudata *A, size_t offA, size_t lda,
gpudata *X, size_t offX, int incX, float beta,
gpudata *Y, size_t offY, int incY) {
cl_ctx *ctx = A->ctx;
cl_uint num_ev = 0;
cl_event evl[3];
cl_event ev;
ARRAY_INIT(A);
ARRAY_INIT(X);
ARRAY_INIT(Y);
CLB_CHECK(ctx->err, clblasSgemv(convO(order), convT(transA), M, N, alpha,
A->buf, offA, lda, X->buf, offX, incX,
beta, Y->buf, offY, incY, 1, &ctx->q,
num_ev, num_ev == 0 ? NULL : evl, &ev));
ARRAY_FINI(A);
ARRAY_FINI(X);
ARRAY_FINI(Y);
clReleaseEvent(ev);
return GA_NO_ERROR;
}
void
xGemv<float>::
call_func()
{
timer.Start(timer_id);
clblasSgemv(order_, buffer_.trans_a_, buffer_.m_, buffer_.n_,
buffer_.alpha_, buffer_.buf_a_, buffer_.offA_,
buffer_.lda_, buffer_.buf_x_, 0, 1, buffer_.beta_,
buffer_.buf_y_, 0, 1, numQueues, queues_, 0, NULL, &event_);
clWaitForEvents(1, &event_);
timer.Stop(timer_id);
}
void TensorMath::GEMV(const bool is_row_major, const bool transpose_A, const int M, const int N, const datum alpha, const Conv::Tensor &A, const int smA, const int ldA, const Conv::Tensor &X, const int smX, const int incX, const datum beta, Conv::Tensor &Y, const int smY, const int incY)
{
#ifdef BUILD_CLBLAS
((Tensor&)A).MoveToGPU();
((Tensor&)X).MoveToGPU();
Y.MoveToGPU(Y.hint_ignore_content_ && beta == 0.0);
cl_event done_event = NULL;
const int offA = A.width() * A.height() * A.maps() * smA;
const int offX = X.width() * X.height() * X.maps() * smX;
const int offY = Y.width() * Y.height() * Y.maps() * smY;
cl_int err =
clblasSgemv(is_row_major ? clblasRowMajor : clblasColumnMajor,
transpose_A ? clblasTrans : clblasNoTrans,
M, N, alpha,
(cl_mem)A.cl_data_ptr_, offA, ldA,
(cl_mem)X.cl_data_ptr_, offX, incX, beta,
(cl_mem)Y.cl_data_ptr_, offY, incY,
1, &(CLHelper::queue), 0, NULL, &done_event);
if(err!=CL_SUCCESS)
FATAL("Call to clblasSgemv failed. Error: " << err);
#else
#ifdef BUILD_OPENCL
((Tensor&)A).MoveToCPU();
((Tensor&)X).MoveToCPU();
Y.MoveToCPU(Y.hint_ignore_content_ && beta == 0.0);
#endif
#ifdef BUILD_BLAS
INNERGEMV(is_row_major ? CblasRowMajor : CblasColMajor,
transpose_A ? CblasTrans : CblasNoTrans,
M, N, alpha, A.data_ptr_const(0,0,0,smA),
ldA, X.data_ptr_const(0,0,0,smX), incX, beta, Y.data_ptr(0,0,0,smY), incY);
#else
if(!is_row_major)
FATAL("Reference GEMV does not support column-major matrices!");
// ...
const datum* a_ptr = A.data_ptr_const(0, 0, 0, smA);
const datum* x_ptr = X.data_ptr_const(0, 0, 0, smX);
datum* y_ptr = Y.data_ptr(0, 0, 0, smY);
#pragma omp parallel for default(shared)
for(int i = 0; i < M; i++) {
datum sum = 0.0;
for(int j = 0; j < N; j++) {
const datum a_value = transpose_A ?
a_ptr[j * ldA + i]
:
a_ptr[i * ldA + j];
const datum x_value = x_ptr[j * incX];
sum += x_value * a_value;
}
if(beta == 0.0)
y_ptr[i * incY] = alpha * sum;
else
y_ptr[i * incY] = beta * y_ptr[i * incY] + alpha * sum;
}
#endif // BUILD_BLAS
#endif // BUILD_CLBLAS
Y.hint_ignore_content_ = false;
}
