Reduce3

1

__global__ void reduce3(float in, float out, int n) {

2

extern __shared__ float sdata[];

3

unsigned int tid = threadIdx.x;

4

unsigned int i = blockIdx.x * (blockDim.x * 2) + threadIdx.x;

5

float sum = (i < n) ? in[i] : 0;

6

if (i + blockDim.x < n) sum += in[i + blockDim.x];

7

sdata[tid] = sum;

8

__syncthreads();

9

for (unsigned int s = blockDim.x / 2; s > 0; s >>= 1) {

10

if (tid < s) {

11

sdata[tid] += sdata[tid + s];

}

12

__syncthreads();

}

13

if (tid == 0) out[blockIdx.x] = sdata[0];