public static Matrix <FieldType> GpuMultiply <FieldType, GpuStructType>(Matrix <FieldType> left, Matrix <FieldType> right) where FieldType : Field <FieldType>, IGpuCompatibleField <FieldType, GpuStructType>, new() where GpuStructType : struct
{
if (left.Width != right.Height)
{
throw new InvalidOperationException("Matrices of incompatible sizes can't be multiplied.");
}
IGpuStructManager <FieldType, GpuStructType> gpuStructManager = new FieldType().GetDefaultGpuStructManager();
GpuStructType[,] resultArr = new GpuStructType[left.Rows, right.Columns];
GpuStructType[,] leftArr = new GpuStructType[left.Rows, left.Columns];
GpuStructType[,] rightArr = new GpuStructType[right.Rows, right.Columns];
resultArr.AssignAll(gpuStructManager.GetStructDefaultValue());
leftArr.AssignAll(ind => gpuStructManager.ToStruct(left[ind[0], ind[1]]));
rightArr.AssignAll(ind => gpuStructManager.ToStruct(right[ind[0], ind[1]]));
Alea.Gpu gpu = Alea.Gpu.Default;
int threadCount = left.Rows * right.Columns;
int blockDimX = gpu.Device.Attributes.MaxThreadsPerBlock; // Threads per block
int gridDimX = (int)Math.Ceiling((double)threadCount / blockDimX); // Blocks per thread
LaunchParam lp = new LaunchParam(gridDimX, blockDimX);
gpu.Launch(multiplicationKernel, lp, leftArr, rightArr, resultArr, gpuStructManager.GetStructAddition(), gpuStructManager.GetStructMultiplication());
FieldType[,] fieldResultArr = new FieldType[resultArr.GetLength(0), resultArr.GetLength(1)];
fieldResultArr.AssignAll(ind => gpuStructManager.ToClass(resultArr[ind[0], ind[1]]));
return(new Matrix <FieldType>(fieldResultArr));
}
public Model(Alea.Gpu gpu, Reader.Reader reader)
{
GPU = gpu;
LearningReader = reader;
}
