unsigned int NuiOpenCLPrefixSum::prefixSum(cl_kernel scan1Kernel, cl_kernel scan2Kernel, cl_mem d_input, cl_mem d_output)
{
scanExclusiveLocal1(d_input, scan1Kernel);
const unsigned int cArrayLength = WORKGROUP_SIZE * 8;
unsigned int batchSize = m_numElements / cArrayLength;
scanExclusiveLocal2(batchSize, d_input, scan2Kernel);
uniformUpdate(batchSize);
generateValidIDs(d_output);
return getValidCount();
}
extern "C" size_t scanExclusiveLarge(
cl_command_queue cqCommandQueue,
cl_mem d_Dst,
cl_mem d_Src,
uint batchSize,
uint arrayLength
) {
//Check power-of-two factorization
uint log2L;
uint factorizationRemainder = factorRadix2(log2L, arrayLength);
oclCheckError( factorizationRemainder == 1, shrTRUE);
//Check supported size range
oclCheckError( (arrayLength >= MIN_LARGE_ARRAY_SIZE) && (arrayLength <= MAX_LARGE_ARRAY_SIZE), shrTRUE );
//Check total batch size limit
oclCheckError( (batchSize * arrayLength) <= MAX_BATCH_ELEMENTS, shrTRUE );
scanExclusiveLocal1(
cqCommandQueue,
d_Dst,
d_Src,
(batchSize * arrayLength) / (4 * WORKGROUP_SIZE),
4 * WORKGROUP_SIZE
);
scanExclusiveLocal2(
cqCommandQueue,
d_Buffer,
d_Dst,
d_Src,
batchSize,
arrayLength / (4 * WORKGROUP_SIZE)
);
return uniformUpdate(
cqCommandQueue,
d_Dst,
d_Buffer,
(batchSize * arrayLength) / (4 * WORKGROUP_SIZE)
);
}
