MapReduce Kernels

Collaboration diagram for MapReduce Kernels:

Functions
static std::string	skepu::UnaryMapReduceKernel_CL ("__kernel void UnaryMapReduceKernel_KERNELNAME(__global TYPE* input, __global TYPE* output, unsigned int n, __local TYPE* sdata, TYPE const1)\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 = FUNCTIONNAME_MAP(input[i], const1);\n"" i += gridSize;\n"" }\n"" while(i < n)\n"" {\n"" TYPE tempMap;\n"" tempMap = FUNCTIONNAME_MAP(input[i], const1);\n"" result = FUNCTIONNAME_REDUCE(result, tempMap, (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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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 std::string	skepu::BinaryMapReduceKernel_CL ("__kernel void BinaryMapReduceKernel_KERNELNAME(__global TYPE* input1, __global TYPE* input2, __global TYPE* output, unsigned int n, __local TYPE* sdata, TYPE const1)\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 = FUNCTIONNAME_MAP(input1[i], input2[i], const1);\n"" i += gridSize;\n"" }\n"" while(i < n)\n"" {\n"" TYPE tempMap;\n"" tempMap = FUNCTIONNAME_MAP(input1[i], input2[i], const1);\n"" result = FUNCTIONNAME_REDUCE(result, tempMap, (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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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 std::string	skepu::TrinaryMapReduceKernel_CL ("__kernel void TrinaryMapReduceKernel_KERNELNAME(__global TYPE* input1, __global TYPE* input2, __global TYPE* input3, __global TYPE* output, unsigned int n, __local TYPE* sdata, TYPE const1)\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 = FUNCTIONNAME_MAP(input1[i], input2[i], input3[i], const1);\n"" i += gridSize;\n"" }\n"" while(i < n)\n"" {\n"" TYPE tempMap;\n"" tempMap = FUNCTIONNAME_MAP(input1[i], input2[i], input3[i], const1);\n"" result = FUNCTIONNAME_REDUCE(result, tempMap, (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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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 UnaryFunc , typename BinaryFunc >
__global__ void	skepu::MapReduceKernel1_CU (UnaryFunc mapFunc, BinaryFunc reduceFunc, T *input, T *output, unsigned int n)
template<typename T , typename BinaryFunc1 , typename BinaryFunc2 >
__global__ void	skepu::MapReduceKernel2_CU (BinaryFunc1 mapFunc, BinaryFunc2 reduceFunc, T *input1, T *input2, T *output, unsigned int n)
template<typename T , typename TrinaryFunc , typename BinaryFunc >
__global__ void	skepu::MapReduceKernel3_CU (TrinaryFunc mapFunc, BinaryFunc reduceFunc, T *input1, T *input2, T *input3, T *output, unsigned int n)

Detailed Description

Definitions of CUDA and OpenCL kernels for the Reduce skeleton.

Function Documentation

static std::string skepu::BinaryMapReduceKernel_CL ( "__kernel void BinaryMapReduceKernel_KERNELNAME(__global TYPE* input1, __global TYPE* input2, __global TYPE* output, unsigned int n, __local TYPE* sdata, TYPE const1)\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 = FUNCTIONNAME_MAP(input1[i], input2[i], const1);\n"" i += gridSize;\n"" }\n"" while(i < n)\n"" {\n"" TYPE tempMap;\n"" tempMap = FUNCTIONNAME_MAP(input1[i], input2[i], const1);\n"" result = FUNCTIONNAME_REDUCE(result, tempMap, (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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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 MapReduce kernel, using the same pattern as reduce6 in the CUDA SDK. See whitepaper from NVIDIA on optimizing reduction for the GPU. For binary map user functions.

template<typename T , typename UnaryFunc , typename BinaryFunc >

__global__ void skepu::MapReduceKernel1_CU	(	UnaryFunc	mapFunc,
		BinaryFunc	reduceFunc,
		T *	input,
		T *	output,
		unsigned int	n
	)

CUDA MapReduce kernel, using the same pattern as reduce6 in the CUDA SDK. See whitepaper from NVIDIA on optimizing reduction for the GPU. For unary map user functions.

template<typename T , typename BinaryFunc1 , typename BinaryFunc2 >

__global__ void skepu::MapReduceKernel2_CU	(	BinaryFunc1	mapFunc,
		BinaryFunc2	reduceFunc,
		T *	input1,
		T *	input2,
		T *	output,
		unsigned int	n
	)

CUDA MapReduce kernel, using the same pattern as reduce6 in the CUDA SDK. See whitepaper from NVIDIA on optimizing reduction for the GPU. For binary map user functions.

template<typename T , typename TrinaryFunc , typename BinaryFunc >

__global__ void skepu::MapReduceKernel3_CU	(	TrinaryFunc	mapFunc,
		BinaryFunc	reduceFunc,
		T *	input1,
		T *	input2,
		T *	input3,
		T *	output,
		unsigned int	n
	)

CUDA MapReduce kernel, using the same pattern as reduce6 in the CUDA SDK. See whitepaper from NVIDIA on optimizing reduction for the GPU. For trinary map user functions.

static std::string skepu::TrinaryMapReduceKernel_CL ( "__kernel void TrinaryMapReduceKernel_KERNELNAME(__global TYPE* input1, __global TYPE* input2, __global TYPE* input3, __global TYPE* output, unsigned int n, __local TYPE* sdata, TYPE const1)\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 = FUNCTIONNAME_MAP(input1[i], input2[i], input3[i], const1);\n"" i += gridSize;\n"" }\n"" while(i < n)\n"" {\n"" TYPE tempMap;\n"" tempMap = FUNCTIONNAME_MAP(input1[i], input2[i], input3[i], const1);\n"" result = FUNCTIONNAME_REDUCE(result, tempMap, (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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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 MapReduce kernel, using the same pattern as reduce6 in the CUDA SDK. See whitepaper from NVIDIA on optimizing reduction for the GPU. For trinary map user functions.

static std::string skepu::UnaryMapReduceKernel_CL ( "__kernel void UnaryMapReduceKernel_KERNELNAME(__global TYPE* input, __global TYPE* output, unsigned int n, __local TYPE* sdata, TYPE const1)\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 = FUNCTIONNAME_MAP(input[i], const1);\n"" i += gridSize;\n"" }\n"" while(i < n)\n"" {\n"" TYPE tempMap;\n"" tempMap = FUNCTIONNAME_MAP(input[i], const1);\n"" result = FUNCTIONNAME_REDUCE(result, tempMap, (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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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_REDUCE(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 MapReduce kernel, using the same pattern as reduce6 in the CUDA SDK. See whitepaper from NVIDIA on optimizing reduction for the GPU. For unary map user functions.