static string TestcublasSgemm2()
{
string testName = "TestcublasSgemm2";
uint aw = 2;
uint bh = aw;
uint ah = 3;
uint bw = 3;
uint ch = ah;
uint cw = bw;
GpuMatrix gpuA;
GpuMatrix gpuB;
GpuMatrix gpuC;
var dataA = MatrixUtils.AA();
var dataB = MatrixUtils.BB();
var cRes = new float[ch * cw];
var cuby = new CublasClr.Cublas();
var aa = new CudaArray();
var res = aa.ResetDevice();
res = res + GpuMatrixOps.SetupGpuMatrix(
out gpuA,
new Matrix <float>(_rows: ch, _cols: cw,
host_data: ImmutableArray.Create(dataA),
matrixFormat: MatrixFormat.Column_Major));
res = res + GpuMatrixOps.SetupGpuMatrix(
out gpuB,
new Matrix <float>(_rows: bh, _cols: bw,
host_data: ImmutableArray.Create(dataB),
matrixFormat: MatrixFormat.Column_Major));
res = res + GpuMatrixOps.SetupGpuMatrix(
out gpuC,
new Matrix <float>(_rows: ch, _cols: cw,
host_data: ImmutableArray.Create(cRes),
matrixFormat: MatrixFormat.Column_Major));
IntPtr cublasHandle = new IntPtr();
res = res + cuby.MakeCublasHandle(ref cublasHandle);
GpuMatrix gpuProd;
res = res + GpuMatrixOps.Multiply(
gmOut: out gpuProd,
cublasHandle: cublasHandle,
gmA: gpuA, gmB: gpuB, gmC: gpuC);
GpuMatrix gpuSynched;
res = res + GpuMatrixOps.CopyToHost(out gpuSynched, gpuProd);
// GpuMatrixUtils.MatrixMult(C: cRes, A: dataA, B: dataB, wA: aw, hA: ah, wB: bw);
return(string.Empty);
}