public static void MyFirstBlasEmulatorTest()
{
Console.WriteLine("MyTest()");
// Get GPU device
CudafyModes.Target = eGPUType.Emulator;
GPGPU gpu = CudafyHost.GetDevice(CudafyModes.Target);
// Create GPGPUBLAS (CUBLAS Wrapper)
using (GPGPUBLAS blas = GPGPUBLAS.Create(gpu))
{
const int N = 100;
float[] a = new float[N];
float[] b = new float[N];
float[] c = new float[N];
float alpha = -1;
float beta = 0;
float[] device_a = gpu.CopyToDevice(a);
float[] device_b = gpu.CopyToDevice(b);
float[] device_c = gpu.CopyToDevice(c);
int m = 10;
int n = 10;
int k = 10;
cublasOperation Op = cublasOperation.N;
blas.GEMM(m, k, n, alpha, device_a, device_b, beta, device_c, Op);
gpu.CopyFromDevice <float>(device_c, c);
}
}
//
// http://stackoverflow.com/questions/18628447/cudafy-throws-an-exception-while-testing
//
private static void BlasSample(int deviceId)
{
CudafyModes.Target = eGPUType.Emulator;
GPGPU gpu = CudafyHost.GetDevice(CudafyModes.Target, deviceId);
CudafyModes.DeviceId = deviceId;
eArchitecture arch = gpu.GetArchitecture();
CudafyModule km = CudafyTranslator.Cudafy(arch);
gpu.LoadModule(km);
GPGPUBLAS blas = GPGPUBLAS.Create(gpu);
const int N = 100;
float[] a = new float[N];
float[] b = new float[N];
float[] c = new float[N];
float alpha = -1;
float beta = 0;
float[] device_a = gpu.CopyToDevice(a);
float[] device_b = gpu.CopyToDevice(b);
float[] device_c = gpu.CopyToDevice(c);
int m = 10;
int n = 10;
int k = 10;
cublasOperation Op = cublasOperation.N;
blas.GEMM(m, k, n, alpha, device_a, device_b, beta, device_c, Op);
throw new NotImplementedException();
}
//http://peterwittek.com/2013/06/cublas-matrix-c-style/
//row major to col major "trick"
public void GemmRowMajor(CpuGpuArray A, CpuGpuArray B, CpuGpuArray C, float cMultiplier = 0f, bool transposeA = false, bool transposeB = false)
{
var blasA = B;
var blasB = A;
// M = rowcount A(T),C
// N = colcount B(T),C
// K = colcount A(T), rowcount B(T)
// However, we flip so
// M = colcount B(T),C
// N = rowcoun A(T), C
// K = cols A(T), rows B
var m = blasA.ColCount;
var n = blasB.RowCount;
var k = blasB.ColCount;
var lda = blasA.ColCount;
var ldb = blasB.ColCount;
var ldc = blasA.ColCount;
var transb = cublasOperation.N;
if (transposeA)
{
transb = cublasOperation.T;
n = blasB.ColCount;
k = blasB.RowCount;
}
var transa = cublasOperation.N;
if (transposeB)
{
transa = cublasOperation.T;
m = blasA.RowCount;
ldc = blasA.RowCount;
}
Blas.GEMM(m, k, n, 1f, B.GPUArray, A.GPUArray, cMultiplier, C.GPUArray, lda: lda, ldb: ldb, ldc: ldc, transb: transb, transa: transa);
}
public static void cudaTransposeAndMultiply(ref MathNet.Numerics.LinearAlgebra.Double.DenseMatrix dm)
{
Cudafy.CudafyModule km = Cudafy.Translator.CudafyTranslator.Cudafy();
km.Serialize();
GPGPU gpu = CudafyHost.GetDevice(eGPUType.Cuda);
int cols = dm.ColumnCount, rows = dm.RowCount;
dm.Storage.ToColumnMajorArray();
double[] a = dm.ToColumnWiseArray();
dm = new MathNet.Numerics.LinearAlgebra.Double.DenseMatrix(1, 1);
double[] dev_a = gpu.Allocate <double>(a.Length);
GPGPUBLAS blas = GPGPUBLAS.Create(gpu);
double[] a_d = gpu.CopyToDevice <double>(a);
double[] c_d = gpu.Allocate <double>(cols * cols);
gpu.StartTimer();
blas.GEMM(cols, rows, cols, 1, a_d, a_d, 0, c_d, Cudafy.Maths.BLAS.Types.cublasOperation.T);
a = new double[cols * cols];
gpu.CopyFromDevice <double>(c_d, a);
gpu.FreeAll();
dm = new MathNet.Numerics.LinearAlgebra.Double.DenseMatrix(cols, cols, a);
}
public void Test_BLAS3_GEMM()
{
// A : No transpose, B : No transpose
ClearBuffer(hiMatrixAMK);
ClearBuffer(hiMatrixBKN);
ClearBuffer(hiMatrixCMN);
FillBuffer(hiMatrixAMK);
FillBuffer(hiMatrixBKN);
FillBuffer(hiMatrixCMN);
diMatrixA = _gpu.CopyToDevice(hiMatrixAMK);
diMatrixB = _gpu.CopyToDevice(hiMatrixBKN);
diMatrixC = _gpu.CopyToDevice(hiMatrixCMN);
_blas.GEMM(M, K, N, Alpha, diMatrixA, diMatrixB, Beta, diMatrixC);
_gpu.CopyFromDevice(diMatrixC, gpuResultMN);
for (int i = 0; i < M; i++)
{
for (int j = 0; j < N; j++)
{
double cpuResult = 0.0;
for (int k = 0; k < K; k++)
{
cpuResult += Alpha * hiMatrixAMK[GetIndexColumnMajor(i, k, M)] * hiMatrixBKN[GetIndexColumnMajor(k, j, K)];
}
cpuResult += Beta * hiMatrixCMN[GetIndexColumnMajor(i, j, M)];
Assert.AreEqual(cpuResult, gpuResultMN[GetIndexColumnMajor(i, j, M)]);
}
}
_gpu.FreeAll();
// A : Transpose, B : No transpose
ClearBuffer(hiMatrixAKM);
ClearBuffer(hiMatrixBKN);
ClearBuffer(hiMatrixCMN);
FillBuffer(hiMatrixAKM);
FillBuffer(hiMatrixBKN);
FillBuffer(hiMatrixCMN);
diMatrixA = _gpu.CopyToDevice(hiMatrixAKM);
diMatrixB = _gpu.CopyToDevice(hiMatrixBKN);
diMatrixC = _gpu.CopyToDevice(hiMatrixCMN);
_blas.GEMM(M, K, N, Alpha, diMatrixA, diMatrixB, Beta, diMatrixC, cublasOperation.T);
_gpu.CopyFromDevice(diMatrixC, gpuResultMN);
for (int i = 0; i < M; i++)
{
for (int j = 0; j < N; j++)
{
double cpuResult = 0.0;
for (int k = 0; k < K; k++)
{
cpuResult += Alpha * hiMatrixAKM[GetIndexColumnMajor(k, i, K)] * hiMatrixBKN[GetIndexColumnMajor(k, j, K)];
}
cpuResult += Beta * hiMatrixCMN[GetIndexColumnMajor(i, j, M)];
Assert.AreEqual(cpuResult, gpuResultMN[GetIndexColumnMajor(i, j, M)]);
}
}
_gpu.FreeAll();
// A : No transpose, B : Transpose
ClearBuffer(hiMatrixAMK);
ClearBuffer(hiMatrixBNK);
ClearBuffer(hiMatrixCMN);
FillBuffer(hiMatrixAMK);
FillBuffer(hiMatrixBNK);
FillBuffer(hiMatrixCMN);
diMatrixA = _gpu.CopyToDevice(hiMatrixAMK);
diMatrixB = _gpu.CopyToDevice(hiMatrixBNK);
diMatrixC = _gpu.CopyToDevice(hiMatrixCMN);
_blas.GEMM(M, K, N, Alpha, diMatrixA, diMatrixB, Beta, diMatrixC, cublasOperation.N, cublasOperation.T);
_gpu.CopyFromDevice(diMatrixC, gpuResultMN);
for (int i = 0; i < M; i++)
{
for (int j = 0; j < N; j++)
{
double cpuResult = 0.0;
for (int k = 0; k < K; k++)
{
cpuResult += Alpha * hiMatrixAMK[GetIndexColumnMajor(i, k, M)] * hiMatrixBNK[GetIndexColumnMajor(j, k, N)];
}
cpuResult += Beta * hiMatrixCMN[GetIndexColumnMajor(i, j, M)];
Assert.AreEqual(cpuResult, gpuResultMN[GetIndexColumnMajor(i, j, M)]);
}
}
_gpu.FreeAll();
// A : Transpose, B : Transpose
ClearBuffer(hiMatrixAKM);
ClearBuffer(hiMatrixBNK);
ClearBuffer(hiMatrixCMN);
FillBuffer(hiMatrixAKM);
FillBuffer(hiMatrixBNK);
FillBuffer(hiMatrixCMN);
diMatrixA = _gpu.CopyToDevice(hiMatrixAKM);
diMatrixB = _gpu.CopyToDevice(hiMatrixBNK);
diMatrixC = _gpu.CopyToDevice(hiMatrixCMN);
_blas.GEMM(M, K, N, Alpha, diMatrixA, diMatrixB, Beta, diMatrixC, cublasOperation.T, cublasOperation.T);
_gpu.CopyFromDevice(diMatrixC, gpuResultMN);
for (int i = 0; i < M; i++)
{
for (int j = 0; j < N; j++)
{
double cpuResult = 0.0;
for (int k = 0; k < K; k++)
{
cpuResult += Alpha * hiMatrixAKM[GetIndexColumnMajor(k, i, K)] * hiMatrixBNK[GetIndexColumnMajor(j, k, N)];
}
cpuResult += Beta * hiMatrixCMN[GetIndexColumnMajor(i, j, M)];
Assert.AreEqual(cpuResult, gpuResultMN[GetIndexColumnMajor(i, j, M)]);
}
}
_gpu.FreeAll();
}