factor/extra/cuda/demos/prefix-sum/prefix-sum.cu

104 lines
2.7 KiB
Plaintext

#include <stdio.h>
#include <stdlib.h>
#include <cuda_runtime.h>
static const int LOG_BANK_COUNT = 4;
static inline __device__ __host__ unsigned shared_offset(unsigned i)
{
return i + (i >> LOG_BANK_COUNT);
}
static inline __device__ __host__ unsigned offset_a(unsigned offset, unsigned i)
{
return shared_offset(offset * (2*i + 1) - 1);
}
static inline __device__ __host__ unsigned offset_b(unsigned offset, unsigned i)
{
return shared_offset(offset * (2*i + 2) - 1);
}
static inline __device__ __host__ unsigned lpot(unsigned x)
{
--x; x |= x>>1; x|=x>>2; x|=x>>4; x|=x>>8; x|=x>>16; return ++x;
}
template<typename T>
__global__ void prefix_sum_block(T *in, T *out, unsigned n)
{
extern __shared__ T temp[];
int idx = threadIdx.x;
int blocksize = blockDim.x;
temp[shared_offset(idx )] = (idx < n) ? in[idx ] : 0;
temp[shared_offset(idx + blocksize)] = (idx + blocksize < n) ? in[idx + blocksize] : 0;
int offset, d;
for (offset = 1, d = blocksize; d > 0; d >>= 1, offset <<= 1) {
__syncthreads();
if (idx < d) {
unsigned a = offset_a(offset, idx), b = offset_b(offset, idx);
temp[b] += temp[a];
}
}
if (idx == 0) temp[shared_offset(blocksize*2 - 1)] = 0;
for (d = 1; d <= blocksize; d <<= 1) {
offset >>= 1;
__syncthreads();
if (idx < d) {
unsigned a = offset_a(offset, idx), b = offset_b(offset, idx);
unsigned t = temp[a];
temp[a] = temp[b];
temp[b] += t;
}
}
__syncthreads();
if (idx < n) out[idx ] = temp[shared_offset(idx )];
if (idx + blocksize < n) out[idx + blocksize] = temp[shared_offset(idx + blocksize)];
}
template<typename T>
void prefix_sum(T *in, T *out, unsigned n)
{
char *device_values;
unsigned n_lpot = lpot(n);
size_t n_pitch;
cudaError_t error = cudaMallocPitch((void**)&device_values, &n_pitch, sizeof(T)*n, 2);
if (error != 0) {
printf("error %u allocating width %lu height %u\n", error, sizeof(T)*n, 2);
exit(1);
}
cudaMemcpy(device_values, in, sizeof(T)*n, cudaMemcpyHostToDevice);
prefix_sum_block<<<1, n_lpot/2, shared_offset(n_lpot)*sizeof(T)>>>
((T*)device_values, (T*)(device_values + n_pitch), n);
cudaMemcpy(out, device_values + n_pitch, sizeof(T)*n, cudaMemcpyDeviceToHost);
cudaFree(device_values);
}
int main()
{
sranddev();
static unsigned in_values[1024], out_values[1024];
for (int i = 0; i < 1024; ++i)
in_values[i] = rand() >> 21;
prefix_sum(in_values, out_values, 1024);
for (int i = 0; i < 1024; ++i)
printf("%5d => %5d\n", in_values[i], out_values[i]);
return 0;
}