Reduce Kernels

Collaboration diagram for Reduce Kernels:

Functions
static std::string	skepu::ReduceKernel_CL ("__kernel void ReduceKernel_KERNELNAME(__global TYPE* input, __global TYPE* output, unsigned int n, __local TYPE* sdata)\n""{\n"" unsigned int blockSize = get_local_size(0);\n"" unsigned int tid = get_local_id(0);\n"" unsigned int i = get_group_id(0)*blockSize + get_local_id(0);\n"" unsigned int gridSize = blockSize*get_num_groups(0);\n"" TYPE result = 0;\n"" if(i < n)\n"" {\n"" result = input[i];\n"" i += gridSize;\n"" }\n"" while(i < n)\n"" {\n"" result = FUNCTIONNAME(result, input[i], (TYPE)0);\n"" i += gridSize;\n"" }\n"" sdata[tid] = result;\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" if(blockSize >= 512) { if (tid < 256 && tid + 256 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 256], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 256) { if (tid < 128 && tid + 128 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 128], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 128) { if (tid < 64 && tid + 64 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 64], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 64) { if (tid < 32 && tid + 32 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 32], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 32) { if (tid < 16 && tid + 16 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 16], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 16) { if (tid < 8 && tid + 8 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 8], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 8) { if (tid < 4 && tid + 4 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 4], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 4) { if (tid < 2 && tid + 2 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 2], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 2) { if (tid < 1 && tid + 1 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 1], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(tid == 0)\n"" {\n"" output[get_group_id(0)] = sdata[tid];\n"" }\n""}\n")
template<typename T , typename BinaryFunc >
__global__ void	skepu::ReduceKernel_CU (BinaryFunc reduceFunc, T *input, T *output, unsigned int n)

Detailed Description

Definitions of CUDA and OpenCL kernels for the Reduce skeleton.

Function Documentation

static std::string skepu::ReduceKernel_CL ( "__kernel void ReduceKernel_KERNELNAME(__global TYPE* input, __global TYPE* output, unsigned int n, __local TYPE* sdata)\n""{\n"" unsigned int blockSize = get_local_size(0);\n"" unsigned int tid = get_local_id(0);\n"" unsigned int i = get_group_id(0)*blockSize + get_local_id(0);\n"" unsigned int gridSize = blockSize*get_num_groups(0);\n"" TYPE result = 0;\n"" if(i < n)\n"" {\n"" result = input[i];\n"" i += gridSize;\n"" }\n"" while(i < n)\n"" {\n"" result = FUNCTIONNAME(result, input[i], (TYPE)0);\n"" i += gridSize;\n"" }\n"" sdata[tid] = result;\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" if(blockSize >= 512) { if (tid < 256 && tid + 256 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 256], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 256) { if (tid < 128 && tid + 128 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 128], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 128) { if (tid < 64 && tid + 64 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 64], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 64) { if (tid < 32 && tid + 32 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 32], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 32) { if (tid < 16 && tid + 16 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 16], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 16) { if (tid < 8 && tid + 8 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 8], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 8) { if (tid < 4 && tid + 4 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 4], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 4) { if (tid < 2 && tid + 2 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 2], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(blockSize >= 2) { if (tid < 1 && tid + 1 < n) { sdata[tid] = FUNCTIONNAME(sdata[tid], sdata[tid + 1], (TYPE)0); } barrier(CLK_LOCAL_MEM_FENCE); }\n"" if(tid == 0)\n"" {\n"" output[get_group_id(0)] = sdata[tid];\n"" }\n""}\n"  )

static

OpenCL Reduce kernel, using the same pattern as reduce6 in the CUDA SDK. See whitepaper from NVIDIA on optimizing reduction for the GPU.

template<typename T , typename BinaryFunc >

__global__ void skepu::ReduceKernel_CU	(	BinaryFunc	reduceFunc,
		T *	input,
		T *	output,
		unsigned int	n
	)

CUDA Reduce kernel, using the same pattern as reduce6 in the CUDA SDK. See whitepaper from NVIDIA on optimizing reduction for the GPU.