SkePU(integratedwithStarPU)  0.8.1
 All Classes Namespaces Files Functions Enumerations Friends Macros Groups Pages
Functions
Scan Kernels
Kernels
Collaboration diagram for Scan Kernels:

Functions

static std::string skepu::ScanKernel_CL ("__kernel void ScanKernel_KERNELNAME(__global TYPE* input, __global TYPE* output, __global TYPE* blockSums, unsigned int n, unsigned int numElements, __local TYPE* sdata)\n""{\n"" unsigned int threadIdx = get_local_id(0);\n"" unsigned int blockDim = get_local_size(0);\n"" unsigned int blockIdx = get_group_id(0);\n"" unsigned int gridDim = get_num_groups(0);\n"" int thid = threadIdx;\n"" int pout = 0;\n"" int pin = 1;\n"" int mem = get_global_id(0);\n"" int blockNr = blockIdx;\n"" unsigned int gridSize = blockDim*gridDim;\n"" unsigned int numBlocks = numElements/(blockDim) + (numElements%(blockDim) == 0 ? 0:1);\n"" int offset;\n"" while(blockNr < numBlocks)\n"" {\n"" sdata[pout*n+thid] = (mem < numElements) ? input[mem] : 0;\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" for(offset = 1; offset < n; offset *=2)\n"" {\n"" pout = 1-pout;\n"" pin = 1-pout;\n"" if(thid >= offset)\n"" sdata[pout*n+thid] = FUNCTIONNAME(sdata[pin*n+thid], sdata[pin*n+thid-offset], (TYPE)0);\n"" else\n"" sdata[pout*n+thid] = sdata[pin*n+thid];\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" }\n"" if(thid == blockDim - 1)\n"" blockSums[blockNr] = sdata[pout*n+blockDim-1];\n"" if(mem < numElements)\n"" output[mem] = sdata[pout*n+thid];\n"" mem += gridSize;\n"" blockNr += gridDim;\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" }\n""}\n")
 
static std::string skepu::ScanUpdate_CL ("__kernel void ScanUpdate_KERNELNAME(__global TYPE* data, __global TYPE* sums, int isInclusive, TYPE init, int n, __global TYPE* ret, __local TYPE* sdata)\n""{\n"" __local TYPE offset;\n"" __local TYPE inc_offset;\n"" unsigned int threadIdx = get_local_id(0);\n"" unsigned int blockDim = get_local_size(0);\n"" unsigned int blockIdx = get_group_id(0);\n"" unsigned int gridDim = get_num_groups(0);\n"" int thid = threadIdx;\n"" int blockNr = blockIdx;\n"" unsigned int gridSize = blockDim*gridDim;\n"" int mem = get_global_id(0);\n"" unsigned int numBlocks = n/(blockDim) + (n%(blockDim) == 0 ? 0:1);\n"" while(blockNr < numBlocks)\n"" {\n"" if(thid == 0)\n"" {\n"" if(isInclusive == 0)\n"" {\n"" offset = init;\n"" if(blockNr > 0)\n"" {\n"" offset = FUNCTIONNAME(offset, sums[blockNr-1], (TYPE)0);\n"" inc_offset = sums[blockNr-1];\n"" }\n"" }\n"" else\n"" {\n"" if(blockNr > 0)\n"" offset = sums[blockNr-1];\n"" }\n"" }\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" if(isInclusive == 1)\n"" {\n"" if(blockNr > 0)\n"" sdata[thid] = (mem < n) ? FUNCTIONNAME(offset, data[mem], (TYPE)0) : 0;\n"" else\n"" sdata[thid] = (mem < n) ? data[mem] : 0;\n"" if(mem == n-1)\n"" *ret = sdata[thid];\n"" }\n"" else\n"" {\n"" if(mem == n-1)\n"" *ret = FUNCTIONNAME(inc_offset, data[mem], (TYPE)0);\n"" if(thid == 0)\n"" sdata[thid] = offset;\n"" else\n"" sdata[thid] = (mem-1 < n) ? FUNCTIONNAME(offset, data[mem-1], (TYPE)0) : 0;\n"" }\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" if(mem < n)\n"" data[mem] = sdata[thid];\n"" mem += gridSize;\n"" blockNr += gridDim;\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" }\n""}\n")
 
static std::string skepu::ScanAdd_CL ("__kernel void ScanAdd_KERNELNAME(__global TYPE* data, TYPE sum, int n)\n""{\n"" int i = get_global_id(0);\n"" unsigned int gridSize = get_local_size(0)*get_num_groups(0);\n"" while(i < n)\n"" {\n"" data[i] = FUNCTIONNAME(data[i], sum, (TYPE)0);\n"" i += gridSize;\n"" }\n""}\n")
 
template<typename T , typename BinaryFunc >
__global__ void skepu::ScanKernel_CU (BinaryFunc scanFunc, T *input, T *output, T *blockSums, unsigned int n, unsigned int numElements)
 
template<typename T , typename BinaryFunc >
__global__ void skepu::ScanUpdate_CU (BinaryFunc scanFunc, T *data, T *sums, int isInclusive, T init, int n, T *ret)
 
template<typename T , typename BinaryFunc >
__global__ void skepu::ScanAdd_CU (BinaryFunc scanFunc, T *data, T sum, int n)
 

Detailed Description

Definitions of CUDA and OpenCL kernels for the Scan skeleton.

Function Documentation

static std::string skepu::ScanAdd_CL ( "__kernel void ScanAdd_KERNELNAME(__global TYPE* data, TYPE sum, int n)\n""{\n"" int i = get_global_id(0);\n"" unsigned int gridSize = get_local_size(0)*get_num_groups(0);\n"" while(i < n)\n"" {\n"" data[i] = FUNCTIONNAME(data[i], sum, (TYPE)0);\n"" i += gridSize;\n"" }\n""}\n"  )
static

Helper kernel to the OpenCL multi-GPU Scan. Applies the device sums to each devices work batch.

template<typename T , typename BinaryFunc >
__global__ void skepu::ScanAdd_CU ( BinaryFunc  scanFunc,
T *  data,
T  sum,
int  n 
)

Helper kernel to the CUDA multi-GPU Scan. Applies the device sums to each devices work batch.

static std::string skepu::ScanKernel_CL ( "__kernel void ScanKernel_KERNELNAME(__global TYPE* input, __global TYPE* output, __global TYPE* blockSums, unsigned int n, unsigned int numElements, __local TYPE* sdata)\n""{\n"" unsigned int threadIdx = get_local_id(0);\n"" unsigned int blockDim = get_local_size(0);\n"" unsigned int blockIdx = get_group_id(0);\n"" unsigned int gridDim = get_num_groups(0);\n"" int thid = threadIdx;\n"" int pout = 0;\n"" int pin = 1;\n"" int mem = get_global_id(0);\n"" int blockNr = blockIdx;\n"" unsigned int gridSize = blockDim*gridDim;\n"" unsigned int numBlocks = numElements/(blockDim) + (numElements%(blockDim) == 0 ? 0:1);\n"" int offset;\n"" while(blockNr < numBlocks)\n"" {\n"" sdata[pout*n+thid] = (mem < numElements) ? input[mem] : 0;\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" for(offset = 1; offset < n; offset *=2)\n"" {\n"" pout = 1-pout;\n"" pin = 1-pout;\n"" if(thid >= offset)\n"" sdata[pout*n+thid] = FUNCTIONNAME(sdata[pin*n+thid], sdata[pin*n+thid-offset], (TYPE)0);\n"" else\n"" sdata[pout*n+thid] = sdata[pin*n+thid];\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" }\n"" if(thid == blockDim - 1)\n"" blockSums[blockNr] = sdata[pout*n+blockDim-1];\n"" if(mem < numElements)\n"" output[mem] = sdata[pout*n+thid];\n"" mem += gridSize;\n"" blockNr += gridDim;\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" }\n""}\n"  )
static

OpenCL Scan kernel. The algorithm is a variation of the naive parallel scan algorithm. Supports exclusive and inclusive scans as well as user defined operators. Also handles vectors of any size.

template<typename T , typename BinaryFunc >
__global__ void skepu::ScanKernel_CU ( BinaryFunc  scanFunc,
T *  input,
T *  output,
T *  blockSums,
unsigned int  n,
unsigned int  numElements 
)

CUDA Scan kernel. The algorithm is a variation of the naive parallel scan algorithm. Supports exclusive and inclusive scans as well as user defined operators. Also handles vectors of any size.

static std::string skepu::ScanUpdate_CL ( "__kernel void ScanUpdate_KERNELNAME(__global TYPE* data, __global TYPE* sums, int isInclusive, TYPE init, int n, __global TYPE* ret, __local TYPE* sdata)\n""{\n"" __local TYPE offset;\n"" __local TYPE inc_offset;\n"" unsigned int threadIdx = get_local_id(0);\n"" unsigned int blockDim = get_local_size(0);\n"" unsigned int blockIdx = get_group_id(0);\n"" unsigned int gridDim = get_num_groups(0);\n"" int thid = threadIdx;\n"" int blockNr = blockIdx;\n"" unsigned int gridSize = blockDim*gridDim;\n"" int mem = get_global_id(0);\n"" unsigned int numBlocks = n/(blockDim) + (n%(blockDim) == 0 ? 0:1);\n"" while(blockNr < numBlocks)\n"" {\n"" if(thid == 0)\n"" {\n"" if(isInclusive == 0)\n"" {\n"" offset = init;\n"" if(blockNr > 0)\n"" {\n"" offset = FUNCTIONNAME(offset, sums[blockNr-1], (TYPE)0);\n"" inc_offset = sums[blockNr-1];\n"" }\n"" }\n"" else\n"" {\n"" if(blockNr > 0)\n"" offset = sums[blockNr-1];\n"" }\n"" }\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" if(isInclusive == 1)\n"" {\n"" if(blockNr > 0)\n"" sdata[thid] = (mem < n) ? FUNCTIONNAME(offset, data[mem], (TYPE)0) : 0;\n"" else\n"" sdata[thid] = (mem < n) ? data[mem] : 0;\n"" if(mem == n-1)\n"" *ret = sdata[thid];\n"" }\n"" else\n"" {\n"" if(mem == n-1)\n"" *ret = FUNCTIONNAME(inc_offset, data[mem], (TYPE)0);\n"" if(thid == 0)\n"" sdata[thid] = offset;\n"" else\n"" sdata[thid] = (mem-1 < n) ? FUNCTIONNAME(offset, data[mem-1], (TYPE)0) : 0;\n"" }\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" if(mem < n)\n"" data[mem] = sdata[thid];\n"" mem += gridSize;\n"" blockNr += gridDim;\n"" barrier(CLK_LOCAL_MEM_FENCE);\n"" }\n""}\n"  )
static

Helper kernel to the OpenCL Scan kernel. Applies the offsets for each block that were saved in main Scan kernel.

template<typename T , typename BinaryFunc >
__global__ void skepu::ScanUpdate_CU ( BinaryFunc  scanFunc,
T *  data,
T *  sums,
int  isInclusive,
T  init,
int  n,
T *  ret 
)

Helper kernel to the CUDA Scan kernel. Applies the offsets for each block that were saved in main Scan kernel.

Generated on Wed May 13 2015 11:11:54 for SkePU(integratedwithStarPU) by   doxygen 1.8.5