#include <mpi.h>
#include <cuda.h>
// kernel
__global__ void sub1(float* fx, float* fy, float* fe) {
#define BLOCK (512)
int t = threadIdx.x; // builtin
int b = blockIdx.x; // builtin
float e;
__shared__ float se[BLOCK];
__shared__ float sx[BLOCK];
__shared__ float sy[BLOCK+2];
// copy from device to processor memory
sx[t] = fx[BLOCK*b+t];
sy[t] = fy[BLOCK*b+t];
if (t<2)
sy[t+BLOCK] = fy[BLOCK*b+t+BLOCK];
__syncthreads();
// do computation
sx[t] += ( sy[t+2] + sy[t] )*.5;
e = sy[t+1] * sy[t+1];
// copy to device memory
fx[BLOCK*b+t] = sx[t];
// reduction
se[t] = e;
__syncthreads();
if (t<256) {
se[t] += se[t+256];
__syncthreads();
}
if (t<128) {
se[t] += se[t+128];
__syncthreads();
}
if (t<64) {
se[t] += se[t+64];
__syncthreads();
}
if (t<32) { // warp size
se[t] += se[t+32];
se[t] += se[t+16];
se[t] += se[t+8];
se[t] += se[t+4];
se[t] += se[t+2];
se[t] += se[t+1];
}
if (t==0)
fe[b] = se[0];
}
int main(int argc, char *argv[]) {
int n = ...;
MPI_Init(&argc, &argv);
int numproc, me;
MPI_Comm_size(MPI_COMM_WORLD, &numproc);
MPI_Comm_rank(MPI_COMM_WORLD, &me);
int p_left = -1, p_right = -1;
if (me > 0)
p_left = me-1;
if (me < numproc-1)
p_right = me+1;
int n_local0 = 1 + (me * (n-1)) / numproc;
int n_local1 = 1 + ((me+1) * (n-1)) / numproc;
// allocate only local part + ghost zone of the arrays x,y
float *x, *y;
MPI_Alloc_mem(sizeof(float) * (n_local1 - n_local0 + 2),
MPI_INFO_NULL, &x);
MPI_Alloc_mem(sizeof(float) * (n_local1 - n_local0 + 2),
MPI_INFO_NULL, &y);
x -= (n_local0 - 1);
y -= (n_local0 - 1);
MPI_Win win;
MPI_Win_create(&y[n_local0], sizeof(float) * (n_local1-n_local0+2),
sizeof(float), MPI_INFO_NULL, MPI_COMM_WORLD, &win);
... // fill x, y
// fill ghost zone
MPI_Win_fence(0, win);
if (p_left != -1)
MPI_Put(&y[n_local0], 1, MPI_FLOAT, p_left,
n_local1-n_local0+1, 1, MPI_FLOAT, win);
if (p_right != -1)
MPI_Put(&y[n_local1-1], 1, MPI_FLOAT, p_right,
0, 1, MPI_FLOAT, win);
MPI_Win_fence(0, win);
// allocate GPU memory
float *fx, *fy, *fe;
cudaMalloc((void**)&fx, (n_local1-n_local0+2) * sizeof(float));
cudaMalloc((void**)&fy, (n_local1-n_local0+2) * sizeof(float));
cudaMalloc((void**)&fe, (n_local1-n_local0+2)/BLOCK * sizeof(float));
float *de = new float[(n_local1-n_local0+2)/BLOCK];
// copy to GPU memory
cudaMemcpy(fx+1, &x[n_local0],
(n_local1-n_local0) * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(fy, &y[n_local0-1],
(n_local1-n_local0+2) * sizeof(float), cudaMemcpyHostToDevice);
dim3 dimBlock(BLOCK, 1, 1);
dim3 dimGrid((n_local1-n_local0+2)/BLOCK, 1, 1);
float e = 0;
// call GPU
sub1<<<dimGrid, dimBlock>>>(fx, fy, fe);
// copy to host memory
cudaMemcpy(fx+1, &x[n_local0], (n_local1-n_local0) * sizeof(float),
cudaMemcpyDeviceToHost);
cudaMemcpy(fe, &de[n_local0-1], (n_local1-n_local0+2)/BLOCK * sizeof(float),
cudaMemcpyDeviceToHost);
// release GPU memory
cudaFree(fe);
cudaFree(fy);
cudaFree(fx);
float e_local = 0;
for (int i=0; i<(n_local1-n_local0+2)/BLOCK; ++i)
e_local += de[i];
e += e_local;
delete[] de;
float e_local = e;
MPI_Allreduce(&e_local, &e, 1, MPI_FLOAT, MPI_SUM, MPI_COMM_WORLD);
... // output x, e
MPI_Win_free(&win);
x += (n_local0 - 1);
y += (n_local0 - 1);
MPI_Free_mem(y);
MPI_Free_mem(x);
MPI_Finalize();
return 0;
}
|