private static void IlGpuOptimisedImpl <TInt>(
CudaAccelerator gpu,
Real[] mSquaredDistances,
Real[] mCoordinates,
int c,
int n,
string name,
Action <ArrayView2D <Real>, ArrayView <Real>, TInt, int> kernelFunc,
Func <int, TInt> numCoordGetter)
where TInt : struct
{
using var cudaSquaredDistance = gpu.Allocate <Real>(n, n);
using var cudaCoordinates = gpu.Allocate(mCoordinates);
var timer = Stopwatch.StartNew();
const int blockSize = 128;
var gridSize = Util.DivUp(n, blockSize);
var lp = ((gridSize, gridSize, 1), (blockSize, 1, 1), SharedMemoryConfig.RequestDynamic <Real>(2 * c * blockSize));
gpu.Launch(kernelFunc, gpu.DefaultStream, lp, cudaSquaredDistance.View, cudaCoordinates.View, numCoordGetter(c), n);
gpu.Synchronize();
Util.PrintPerformance(timer, name, n, c, n);
cudaSquaredDistance.CopyTo(mSquaredDistances, (0, 0), 0, (n, n));
}
public static void IlGpu(
CudaAccelerator gpu,
Real[] mIntraReturn,
Real[] vClose,
Real[] vIsAlive,
Real[] vIsValidDay,
int m,
int n)
{
using (var cudaIntraReturn = gpu.Allocate(mIntraReturn))
using (var cudaClose = gpu.Allocate(vClose))
using (var cudaIsAlive = gpu.Allocate(vIsAlive))
using (var cudaIsValidDay = gpu.Allocate(vIsValidDay))
{
var timer = Stopwatch.StartNew();
var gridSizeX = Util.DivUp(n, 32);
var gridSizeY = Util.DivUp(m, 8);
var lp = ((gridSizeX, gridSizeY, 1), (32, 8));
gpu.Launch(IlGpuKernel, gpu.DefaultStream, lp, cudaIntraReturn.View, cudaClose.View, cudaIsAlive.View, cudaIsValidDay.View, m, n);
gpu.Synchronize();
Util.PrintPerformance(timer, "IntraReturn.IlGpu", 5, m, n);
cudaIntraReturn.CopyTo(mIntraReturn, 0, 0, mIntraReturn.Length);
}
}
public static void IlGpu(CudaAccelerator gpu, Real[] matrix, Real[] vector, int m, int n)
{
using (var cudaMatrix = gpu.Allocate(matrix))
using (var cudaVector = gpu.Allocate(vector))
{
var timer = Stopwatch.StartNew();
var gridSizeX = Util.DivUp(n, 32);
var gridSizeY = Util.DivUp(m, 8);
var lp = ((gridSizeX, gridSizeY, 1), (32, 8));
gpu.Launch(IlGpuKernel, gpu.DefaultStream, lp, cudaMatrix.View, cudaVector.View, m, n);
gpu.Synchronize();
Util.PrintPerformance(timer, "AddVector.IlGpu", 3, m, n);
cudaMatrix.CopyTo(matrix, 0, 0, matrix.Length);
}
}
public static void IlGpu(
CudaAccelerator gpu,
Real[] mSquaredDistances,
Real[] mCoordinates,
int c,
int n)
{
using var cudaSquaredDistance = gpu.Allocate(mSquaredDistances);
using var cudaCoordinates = gpu.Allocate(mCoordinates);
var timer = Stopwatch.StartNew();
const int blockSize = 128;
var gridSize = Util.DivUp(n * n, blockSize);
var lp = (gridSize, blockSize);
gpu.Launch(IlGpuKernel, gpu.DefaultStream, lp, cudaSquaredDistance.View, cudaCoordinates.View, c, n);
gpu.Synchronize();
Util.PrintPerformance(timer, "SquaredDistance.IlGpu", n, c, n);
cudaSquaredDistance.CopyTo(mSquaredDistances, 0, 0, mSquaredDistances.Length);
}
private static void IlGpuOptimisedImpl(
CudaAccelerator gpu,
Real[] mSquaredDistances,
Real[] mCoordinates,
int c,
int n,
string name,
Action <ArrayView2D <Real>, ArrayView <Real>, SpecializedValue <int>, SpecializedValue <int>, int> kernelFunc)
{
using var cudaSquaredDistance = gpu.Allocate <Real>(n, n);
using var cudaCoordinates = gpu.Allocate(mCoordinates);
var timer = Stopwatch.StartNew();
const int blockSize = 128;
var gridSize = Util.DivUp(n, blockSize);
var lp = ((gridSize, gridSize, 1), (blockSize, 1, 1));
gpu.Launch(kernelFunc, gpu.DefaultStream, lp, cudaSquaredDistance.View, cudaCoordinates.View, SpecializedValue.New(blockSize), SpecializedValue.New(c), n);
gpu.Synchronize();
Util.PrintPerformance(timer, name, n, c, n);
cudaSquaredDistance.CopyTo(mSquaredDistances, (0, 0), 0, (n, n));
}