public static Tensor Invoke(CudaReduceAllKernels reduceAllKernels, float init, ReduceInitType initType, string kernelName, Tensor result, Tensor src, object extraArg = null)
{
int deviceId = CudaHelpers.GetDeviceId(src);
TSCudaContext context = CudaHelpers.TSContextForTensor(src);
CudaContext cudaContext = context.CudaContextForDevice(deviceId);
if (src.DimensionCount > TSCudaContext.MaxDims)
{
throw new InvalidOperationException("Tensors with dimension count > " + TSCudaContext.MaxDims + " are not supported");
}
Tensor writeTarget = TensorResultBuilder.GetWriteTarget(result, src, false, 1);
if (src.DimensionCount == 0)
{
return(result);
}
long totalElements = src.ElementCount();
ApplySpecialization config = new ApplySpecialization(src);
object totalElementsTyped = config.Use32BitIndices ? (uint)totalElements : (ulong)totalElements;
object initValueTyped = ReduceInitConverter.GetInitValue(init, initType, src.ElementType);
dim3 grid;
dim3 block;
byte[] ptx = reduceAllKernels.GetPtx(context.Compiler);
string fullKernelName = PermutationGenerator.GetMangledName(kernelName, config);
ManagedCuda.BasicTypes.CUdeviceptr outputDevicePtr = CudaHelpers.GetBufferStart(writeTarget);
if (isTwoPassReductionSize(totalElements))
{
getPass1ReduceBlockGrid(context, deviceId, totalElements, out grid, out block);
uint smemSize = block.x * sizeof(float);
ManagedCuda.BasicTypes.CUdeviceptr scratchSpace = context.ScratchSpaceForDevice(deviceId).buffer;
if (extraArg == null)
{
InvokeReduceAll(context, cudaContext, ptx, "twoPassA_" + fullKernelName, grid, block, smemSize, config, src, totalElementsTyped, initValueTyped, scratchSpace);
}
else
{
InvokeReduceAll(context, cudaContext, ptx, "twoPassA_" + fullKernelName, grid, block, smemSize, config, src, totalElementsTyped, initValueTyped, scratchSpace, extraArg);
}
uint numPass1Blocks = grid.x;
getPass2ReduceBlockGrid(context, deviceId, totalElements, out grid, out block);
smemSize = block.x * sizeof(float);
InvokeReduceAllPass2(context, cudaContext, ptx, "twoPassB_" + fullKernelName, grid, block, smemSize, config.Use32BitIndices, numPass1Blocks, initValueTyped, scratchSpace, outputDevicePtr);
}
else
{
getSinglePassReduceBlockGrid(totalElements, out grid, out block);
uint smemSize = block.x * sizeof(float);
if (extraArg == null)
{
InvokeReduceAll(context, cudaContext, ptx, "onePass_" + fullKernelName, grid, block, smemSize, config, src, totalElementsTyped, initValueTyped, outputDevicePtr);
}
else
{
InvokeReduceAll(context, cudaContext, ptx, "onePass_" + fullKernelName, grid, block, smemSize, config, src, totalElementsTyped, initValueTyped, outputDevicePtr, extraArg);
}
}
return(writeTarget);
}
public static Tensor Invoke(CudaReduceKernels reduceKernels, string kernelName, float init, ReduceInitType initType, Tensor result, Tensor src, int dim, object extraArg = null)
{
if (src.DimensionCount == 0)
{
return(result);
}
var context = CudaHelpers.TSContextForTensor(src);
var cudaContext = context.CudaContextForTensor(src);
var requiredOutputSize = (long[])src.Sizes.Clone();
requiredOutputSize[dim] = 1;
var writeTarget = TensorResultBuilder.GetWriteTarget(result, src, false, requiredOutputSize);
ThrowIfAnyTensorInvalid(writeTarget, src);
var inElements = src.ElementCount();
var reductionSize = src.Sizes[dim];
var reductionStride = src.Strides[dim];
var outElements = inElements / reductionSize;
var contigReduction = reductionStride == 1;
// We must make sure that when the tensor is passed to the kernel, src.Sizes[dim] is set to 1
// This includes for the purposes of determining which tensor specializations to use (changing
// the dimension size to 1 may make the tensor non-contiguous
var newSizes = (long[])src.Sizes.Clone();
newSizes[dim] = 1;
var srcSlim = new Tensor(newSizes, src.Strides, src.Storage, src.StorageOffset);
var config = new ApplySpecialization(writeTarget, srcSlim);
object totalSlices = config.Use32BitIndices ? (uint)outElements : (ulong)outElements;
object reductionSizeTyped = config.Use32BitIndices ? (uint)reductionSize : (ulong)reductionSize;
object reductionStrideTyped = config.Use32BitIndices ? (uint)reductionStride : (ulong)reductionStride;
var initValueTyped = ReduceInitConverter.GetInitValue(init, initType, src.ElementType);
var ptx = reduceKernels.GetPtx(context.Compiler);
if (contigReduction)
{
var block = GetContigReduceBlock(cudaContext, outElements, reductionSize);
var grid = GetContigReduceGrid(outElements);
var smemSize = (uint)src.ElementType.Size() * block.x;
var fullName = "contig_" + PermutationGenerator.GetMangledName(kernelName, config);
if (extraArg == null)
{
InvokeReduce(context, cudaContext, ptx, fullName, grid, block, smemSize, config, writeTarget, srcSlim, reductionSizeTyped, totalSlices, initValueTyped);
}
else
{
InvokeReduce(context, cudaContext, ptx, fullName, grid, block, smemSize, config, writeTarget, srcSlim, reductionSizeTyped, totalSlices, initValueTyped, extraArg);
}
}
else
{
var deviceProps = context.DeviceInfoForContext(cudaContext);
var block = GetNonContigReduceBlock(deviceProps);
var grid = GetNoncontigReduceGrid(deviceProps, outElements);
uint smemSize = 0;
var fullName = "noncontig_" + PermutationGenerator.GetMangledName(kernelName, config);
if (extraArg == null)
{
InvokeReduce(context, cudaContext, ptx, fullName, grid, block, smemSize, config, writeTarget, srcSlim, reductionStrideTyped, reductionSizeTyped, totalSlices, initValueTyped);
}
else
{
InvokeReduce(context, cudaContext, ptx, fullName, grid, block, smemSize, config, writeTarget, srcSlim, reductionStrideTyped, reductionSizeTyped, totalSlices, initValueTyped, extraArg);
}
}
return(writeTarget);
}