static int dgemv(cb_order order, cb_transpose transA, size_t M, size_t N,
double alpha, gpudata *A, size_t offA, size_t lda,
gpudata *X, size_t offX, int incX, double 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, clblasDgemv(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<double>::
call_func()
{
timer.Start(timer_id);
clblasDgemv(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);
}
