public CpuGpuArray AddArray(ArrayName arrayName, int dim1, int dim2)
{
var dim = dim1 * dim2;
var res = new CpuGpuArray(this._gpuModule, dim1, dim2);
_Arrays.Add(arrayName, res);
return res;
}
public CpuGpuArray(CpuGpuArray array, int rows, int cols)
{
_gpuModule = array._gpuModule;
CPUArray = array.CPUArray;
GPUArray = array.GPUArray;
RowCount = rows;
ColCount = cols;
}
public CpuGpuArray(CpuGpuArray array, int rows, int cols)
{
_gpuModule = array._gpuModule;
CPUArray = array.CPUArray;
GPUArray = array.GPUArray;
RowCount = rows;
ColCount = cols;
}
public void FillArray(CpuGpuArray array, float value)
{
var kernelCount = array.Length;
var dim = ComputeBlocksTreads(kernelCount);
Gpu.Launch(dim.Item1, dim.Item2).FillArrayGPU(
kernelCount,
array.GPUArray,
value
);
}
void InitArrays(int dataSize, int rowCount)
{
if (NonZeroCount > dataSize)
{
throw new ArgumentException();
}
DataSize = dataSize;
RowCount = rowCount;
CPUValues = new float[DataSize];
CPUIndices = new int[DataSize];
if (_gpuModule != null)
{
GPUValues = CpuGpuArray.AllocateGPUArray(_gpuModule, DataSize);
GPUIndices = CpuGpuArrayInt.AllocateGPUArray(_gpuModule, DataSize);
}
}
public void ElementwiseMultiplication(CpuGpuArray vector, CpuGpuArray matrix)
{
if (vector.Length != matrix.Height)
{
throw new Exception("Matrix.height <> vector.length");
}
var kernelCount = matrix.Length;
var dim = ComputeBlocksTreads(kernelCount);
Gpu.Launch(dim.Item1, dim.Item2).ElementwiseMultiplicationGPU(
kernelCount,
vector.GPUArray,
matrix.GPUArray,
matrix.Width
);
}
//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 void MultiplySparse(CpuGpuMatrixSparse A, CpuGpuArray B, CpuGpuArray C, bool transposeA)
{
FillArray(C, 0f);
var kernelCount = B.ColCount * A.NonZeroCount;
var dim = ComputeBlocksTreads(kernelCount);
var colcountA = A.ColCount;
var transposeInt = transposeA ? 1 : 0;
Gpu.Launch(dim.Item1, dim.Item2).MultiplySparseGPU2(
kernelCount,
A.GPUIndices,
A.GPUValues,
A.NonZeroCount,
colcountA,
B.GPUArray,
B.ColCount,
C.GPUArray,
transposeInt
);
return;
}
//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 string SerializeArrayValues(CpuGpuArray array)
{
var res = String.Join("\n", array.CPUArray.Select(x => x.ToString(CultureInfo.InvariantCulture)));
return res;
}
public void SetSparseGpuValues(CpuGpuArray target, CpuGpuArray indices, CpuGpuArray values)
{
SetSparseGpuValues(target.GPUArray, target.RowCount, target.ColCount, indices, values);
}
public void FillArray(CpuGpuArray array, float value)
{
var kernelCount = array.Length;
var dim = ComputeBlocksTreads(kernelCount);
Gpu.Launch(dim.Item1, dim.Item2).FillArrayGPU(
kernelCount,
array.GPUArray,
value
);
}
public void MultiplySparse(CpuGpuMatrixSparse A, CpuGpuArray B, CpuGpuArray C, bool transposeA)
{
FillArray(C, 0f);
var kernelCount = B.ColCount * A.NonZeroCount;
var dim = ComputeBlocksTreads(kernelCount);
var colcountA = A.ColCount;
var transposeInt = transposeA ? 1 : 0;
Gpu.Launch(dim.Item1, dim.Item2).MultiplySparseGPU2(
kernelCount,
A.GPUIndices,
A.GPUValues,
A.NonZeroCount,
colcountA,
B.GPUArray,
B.ColCount,
C.GPUArray,
transposeInt
);
return;
}
//float[] indices, float[] values, int rowCount, int colCount
public void SetSparseGpuValues(float[] gpuTarget, int targetRowCount, int targetColCount, CpuGpuArray indices, CpuGpuArray values)
{
if (indices.Length != values.Length) throw new ArgumentException();
var kernelCount = indices.RowCount * indices.ColCount;
var dim = ComputeBlocksTreads(kernelCount);
Gpu.Launch(dim.Item1, dim.Item2).SetSparseGpuValuesGPU(
kernelCount,
gpuTarget,
targetRowCount,
targetColCount,
indices.GPUArray,
values.GPUArray,
indices.RowCount,
indices.ColCount
);
}
public void SetSparseGpuValues(CpuGpuArray target, CpuGpuArray indices, CpuGpuArray values)
{
SetSparseGpuValues(target.GPUArray, target.RowCount, target.ColCount, indices, values);
}
public void ElementwiseMultiplication(CpuGpuArray vector, CpuGpuArray matrix)
{
if (vector.Length != matrix.Height) throw new Exception("Matrix.height <> vector.length");
var kernelCount = matrix.Length;
var dim = ComputeBlocksTreads(kernelCount);
Gpu.Launch(dim.Item1, dim.Item2).ElementwiseMultiplicationGPU(
kernelCount,
vector.GPUArray,
matrix.GPUArray,
matrix.Width
);
}
public static string SerializeArrayValues(CpuGpuArray array)
{
var res = String.Join("\n", array.CPUArray.Select(x => x.ToString(CultureInfo.InvariantCulture)));
return(res);
}
//float[] indices, float[] values, int rowCount, int colCount
public void SetSparseGpuValues(float[] gpuTarget, int targetRowCount, int targetColCount, CpuGpuArray indices, CpuGpuArray values)
{
if (indices.Length != values.Length)
{
throw new ArgumentException();
}
var kernelCount = indices.RowCount * indices.ColCount;
var dim = ComputeBlocksTreads(kernelCount);
Gpu.Launch(dim.Item1, dim.Item2).SetSparseGpuValuesGPU(
kernelCount,
gpuTarget,
targetRowCount,
targetColCount,
indices.GPUArray,
values.GPUArray,
indices.RowCount,
indices.ColCount
);
}