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);
}
static string TestCublasHandle()
{
CublasOp o = CublasClr.CublasOp.N;
string testName = "TestCublasHandle";
var cuby = new CublasClr.Cublas();
IntPtr devHandle = new IntPtr();
var aa = new CudaArray();
try
{
var res = aa.ResetDevice();
res = res + cuby.MakeCublasHandle(ref devHandle);
res = res + cuby.DestroyCublasHandle(devHandle);
if (res != String.Empty)
{
return(testName + " fail: " + res);
}
return(testName + " pass");
}
catch
{
return(testName + " fail");
}
finally
{
//aa.ReleaseDevicePtr(devData);
aa.ResetDevice();
}
}
//C = α op ( A ) op ( B ) + β C
public static string Multiply(
out GpuMatrix gmOut, IntPtr cublasHandle,
GpuMatrix gmA, GpuMatrix gmB, GpuMatrix gmC)
{
if (gmA.DevHostState == DevHostState.DeviceNotAllocated)
{
gmOut = null;
return("Device data pointer for matrix A not allocated");
}
if (gmB.DevHostState == DevHostState.DeviceNotAllocated)
{
gmOut = null;
return("Device data pointer for matrix B not allocated");
}
if (gmC.DevHostState == DevHostState.DeviceNotAllocated)
{
gmOut = null;
return("Device data pointer for matrix C not allocated");
}
var cuby = new CublasClr.Cublas();
var strRet = cuby.cublasSgemm(
cublas_handle: cublasHandle,
transa: CublasOp.N,
transb: CublasOp.N,
m: (int)gmA.Matrix.Rows,
n: (int)gmB.Matrix.Cols,
k: (int)gmA.Matrix.Cols,
alpha: 1,
dev_A: gmA.DevPtr,
lda: (int)gmA.Matrix.Rows,
dev_B: gmB.DevPtr,
ldb: (int)gmB.Matrix.Rows,
beta: 0,
dev_C: gmC.DevPtr,
ldc: (int)gmC.Matrix.Rows
);
if (!String.IsNullOrEmpty(strRet))
{
gmOut = null;
return(strRet);
}
gmOut = new GpuMatrix(
matrix: gmC.Matrix,
devPtr: gmC.DevPtr,
devHostState: DevHostState.DeviceIsNewer);
return(String.Empty);
}
static string TestcublasSgemm1()
{
string testName = "TestcublasSgemm";
uint aw = 5;
uint bh = aw;
uint ah = 5;
uint bw = 5;
uint ch = ah;
uint cw = bw;
var cuby = new CublasClr.Cublas();
var aa = new CudaArray();
var res = aa.ResetDevice();
var dataA = MatrixUtils.MakeIdentity(rows: ah, cols: aw);
GpuMatrix gpuA;
res = res + GpuMatrixOps.SetupGpuMatrix(
out gpuA,
new Matrix <float>(_rows: ah, _cols: aw,
host_data: ImmutableArray.Create(dataA),
matrixFormat: MatrixFormat.Column_Major));
var dataB = MatrixUtils.MakeIdentiPoke(rows: bh, cols: bw);
GpuMatrix gpuB;
res = res + GpuMatrixOps.SetupGpuMatrix(
out gpuB,
new Matrix <float>(_rows: bh, _cols: bw,
host_data: ImmutableArray.Create(dataB),
matrixFormat: MatrixFormat.Column_Major));
var dataC = MatrixUtils.MakeZeroes(rows: bh, cols: bw);
GpuMatrix gpuC;
res = res + GpuMatrixOps.SetupGpuMatrix(
out gpuC,
new Matrix <float>(_rows: ch, _cols: cw,
host_data: ImmutableArray.Create(dataC),
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);
var cpuRes = new float[ah * bw];
MatrixUtils.RowMajorMatrixMult(C: cpuRes, A: dataA, B: dataB, wA: aw, hA: ah, wB: bw);
var cpuRes2 = new float[bh * aw];
MatrixUtils.RowMajorMatrixMult(C: cpuRes2, A: dataB, B: dataA, wA: bw, hA: bh, wB: aw);
return(res);
}