public static double2[] Zaxpy(int n, double2 alpha, double2[] x, int incx, double2[] y, int incy)
{
using (var dalpha = Worker.Malloc(new[] { alpha }))
using (var dx = Worker.Malloc(x))
using (var dy = Worker.Malloc(y))
{
Cublas.Zaxpy(n, dalpha.Ptr, dx.Ptr, incx, dy.Ptr, incy);
return(dy.Gather());
}
}
public static double2[] Zgemm(int m, int n, int k, double2 alpha, double2[] A, int lda, double2[] B, int ldb, double2 beta, double2[] C, int ldc)
{
using (var dA = Worker.Malloc(A))
using (var dB = Worker.Malloc(B))
using (var dC = Worker.Malloc(C))
using (var dAlpha = Worker.Malloc(new[] { alpha }))
using (var dBeta = Worker.Malloc(new[] { beta }))
{
Cublas.Zgemm(Transa, Transb, m, n, k, dAlpha.Ptr, dA.Ptr, lda, dB.Ptr, ldb, dBeta.Ptr, dC.Ptr, ldc);
return(dC.Gather());
}
}
//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 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);
}
public static double[][] Dgemm(int m, int n, int k, double alpha, double[][] hAs, int lda, double[][] hBs, int ldb, double beta, double[][] hCs, int ldc)
{
var batchCount = hAs.Length;
var dAs = (from hA in hAs select Worker.Malloc(hA)).ToArray();
var dBs = (from hB in hBs select Worker.Malloc(hB)).ToArray();
var dCs = (from hC in hCs select Worker.Malloc(hC)).ToArray();
try
{
using (var dAPtrs = Worker.Malloc((from dA in dAs select dA.Ptr).ToArray()))
using (var dBPtrs = Worker.Malloc((from dB in dBs select dB.Ptr).ToArray()))
using (var dCPtrs = Worker.Malloc((from dC in dCs select dC.Ptr).ToArray()))
using (var dAlpha = Worker.Malloc(new[] { alpha }))
using (var dBeta = Worker.Malloc(new[] { beta }))
{
Cublas.Dgemm(Transa, Transb, m, n, k, dAlpha.Ptr, dAPtrs.Ptr, lda, dBPtrs.Ptr, ldb, dBeta.Ptr,
dCPtrs.Ptr, ldc, batchCount);
}
return((from dC in dCs select dC.Gather()).ToArray());
}
finally
{
foreach (var dA in dAs)
{
dA.Dispose();
}
foreach (var dB in dBs)
{
dB.Dispose();
}
foreach (var dC in dCs)
{
dC.Dispose();
}
}
}
