void step2_gpu(int *n) {
int nprocs, procid;
MPI_Comm_rank(MPI_COMM_WORLD, &procid);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
/* Create Cartesian Communicator */
int c_dims[2];
MPI_Comm c_comm;
accfft_create_comm(MPI_COMM_WORLD, c_dims, &c_comm);
float *data, *data_cpu;
Complexf *data_hat;
double f_time = 0 * MPI_Wtime(), i_time = 0, setup_time = 0;
int alloc_max = 0;
int isize[3], osize[3], istart[3], ostart[3];
/* Get the local pencil size and the allocation size */
alloc_max = accfft_local_size_dft_r2c_gpuf(n, isize, istart, osize, ostart,
c_comm);
/* Note that both need to be allocated by alloc_max because of inplace transform*/
data_cpu = (float*) malloc(alloc_max);
cudaMalloc((void**) &data, alloc_max);
//accfft_init(nthreads);
/* Create FFT plan */
setup_time = -MPI_Wtime();
accfft_plan_gpuf * plan = accfft_plan_dft_3d_r2c_gpuf(n, data, data, c_comm,
ACCFFT_MEASURE);
setup_time += MPI_Wtime();
/* Warm Up */
accfft_execute_r2c_gpuf(plan, data, (Complexf*) data);
accfft_execute_r2c_gpuf(plan, data, (Complexf*) data);
/* Initialize data */
initialize_gpu(data, n, isize, istart);
MPI_Barrier(c_comm);
/* Perform forward FFT */
f_time -= MPI_Wtime();
accfft_execute_r2c_gpuf(plan, data, (Complexf*) data);
f_time += MPI_Wtime();
MPI_Barrier(c_comm);
/* Perform backward FFT */
i_time -= MPI_Wtime();
accfft_execute_c2r_gpuf(plan, (Complexf*) data, data);
i_time += MPI_Wtime();
/* copy back results on CPU */
cudaMemcpy(data_cpu, data, alloc_max, cudaMemcpyDeviceToHost);
/* Check Error */
check_err(data_cpu, n, c_comm);
/* Compute some timings statistics */
double g_f_time, g_i_time, g_setup_time;
MPI_Reduce(&f_time, &g_f_time, 1, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
MPI_Reduce(&i_time, &g_i_time, 1, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
MPI_Reduce(&setup_time, &g_setup_time, 1, MPI_DOUBLE, MPI_MAX, 0,
MPI_COMM_WORLD);
PCOUT << "GPU Timing for FFT of size " << n[0] << "*" << n[1] << "*" << n[2]
<< std::endl;
PCOUT << "Setup \t" << g_setup_time << std::endl;
PCOUT << "FFT \t" << g_f_time << std::endl;
PCOUT << "IFFT \t" << g_i_time << std::endl;
free(data_cpu);
cudaFree(data);
accfft_destroy_plan_gpu(plan);
accfft_cleanup_gpuf();
MPI_Comm_free(&c_comm);
return;
} // end step2_gpu
void step3_gpu(int *n) {
int nprocs, procid;
MPI_Comm_rank(MPI_COMM_WORLD, &procid);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
/* Create Cartesian Communicator */
int c_dims[2]={0};
MPI_Comm c_comm;
accfft_create_comm(MPI_COMM_WORLD,c_dims,&c_comm);
Complexf *data, *data_cpu;
Complexf *data_hat;
double f_time=0*MPI_Wtime(),i_time=0, setup_time=0;
int alloc_max=0;
int isize[3],osize[3],istart[3],ostart[3];
/* Get the local pencil size and the allocation size */
alloc_max=accfft_local_size_dft_c2c_gpuf(n,isize,istart,osize,ostart,c_comm);
#ifdef INPLACE
data_cpu=(Complexf*)malloc(alloc_max);
cudaMalloc((void**) &data, alloc_max);
#else
data_cpu=(Complexf*)malloc(isize[0]*isize[1]*isize[2]*2*sizeof(float));
cudaMalloc((void**) &data,isize[0]*isize[1]*isize[2]*2*sizeof(float));
cudaMalloc((void**) &data_hat, alloc_max);
#endif
//accfft_init(nthreads);
setup_time=-MPI_Wtime();
/* Create FFT plan */
#ifdef INPLACE
accfft_plan_gpuf * plan=accfft_plan_dft_3d_c2c_gpuf(n,data,data,c_comm,ACCFFT_MEASURE);
#else
accfft_plan_gpuf * plan=accfft_plan_dft_3d_c2c_gpuf(n,data,data_hat,c_comm,ACCFFT_MEASURE);
#endif
setup_time+=MPI_Wtime();
/* Warmup Runs */
#ifdef INPLACE
accfft_execute_c2c_gpuf(plan,ACCFFT_FORWARD,data,data);
accfft_execute_c2c_gpuf(plan,ACCFFT_FORWARD,data,data);
#else
accfft_execute_c2c_gpuf(plan,ACCFFT_FORWARD,data,data_hat);
accfft_execute_c2c_gpuf(plan,ACCFFT_FORWARD,data,data_hat);
#endif
/* Initialize data */
initialize(data_cpu,n,c_comm);
#ifdef INPLACE
cudaMemcpy(data, data_cpu,alloc_max, cudaMemcpyHostToDevice);
#else
cudaMemcpy(data, data_cpu,isize[0]*isize[1]*isize[2]*2*sizeof(float), cudaMemcpyHostToDevice);
#endif
MPI_Barrier(c_comm);
/* Perform forward FFT */
f_time-=MPI_Wtime();
#ifdef INPLACE
accfft_execute_c2c_gpuf(plan,ACCFFT_FORWARD,data,data);
#else
accfft_execute_c2c_gpuf(plan,ACCFFT_FORWARD,data,data_hat);
#endif
f_time+=MPI_Wtime();
MPI_Barrier(c_comm);
#ifndef INPLACE
Complexf *data2_cpu, *data2;
cudaMalloc((void**) &data2, isize[0]*isize[1]*isize[2]*2*sizeof(float));
data2_cpu=(Complexf*) malloc(isize[0]*isize[1]*isize[2]*2*sizeof(float));
#endif
/* Perform backward FFT */
i_time-=MPI_Wtime();
#ifdef INPLACE
accfft_execute_c2c_gpuf(plan,ACCFFT_BACKWARD,data,data);
#else
accfft_execute_c2c_gpuf(plan,ACCFFT_BACKWARD,data_hat,data2);
#endif
i_time+=MPI_Wtime();
/* copy back results on CPU and check error*/
#ifdef INPLACE
cudaMemcpy(data_cpu, data, alloc_max, cudaMemcpyDeviceToHost);
check_err(data_cpu,n,c_comm);
#else
cudaMemcpy(data2_cpu, data2, isize[0]*isize[1]*isize[2]*2*sizeof(float), cudaMemcpyDeviceToHost);
check_err(data2_cpu,n,c_comm);
#endif
/* Compute some timings statistics */
double g_f_time, g_i_time, g_setup_time;
MPI_Reduce(&f_time,&g_f_time,1, MPI_DOUBLE, MPI_MAX,0, MPI_COMM_WORLD);
MPI_Reduce(&i_time,&g_i_time,1, MPI_DOUBLE, MPI_MAX,0, MPI_COMM_WORLD);
MPI_Reduce(&setup_time,&g_setup_time,1, MPI_DOUBLE, MPI_MAX,0, MPI_COMM_WORLD);
#ifdef INPLACE
PCOUT<<"GPU Timing for Inplace FFT of size "<<n[0]<<"*"<<n[1]<<"*"<<n[2]<<std::endl;
#else
PCOUT<<"GPU Timing for Outplace FFT of size "<<n[0]<<"*"<<n[1]<<"*"<<n[2]<<std::endl;
#endif
PCOUT<<"Setup \t"<<g_setup_time<<std::endl;
PCOUT<<"FFT \t"<<g_f_time<<std::endl;
PCOUT<<"IFFT \t"<<g_i_time<<std::endl;
MPI_Barrier(c_comm);
cudaDeviceSynchronize();
free(data_cpu);
cudaFree(data);
#ifndef INPLACE
cudaFree(data_hat);
free(data2_cpu);
cudaFree(data2);
#endif
accfft_destroy_plan_gpu(plan);
accfft_cleanup_gpuf();
MPI_Comm_free(&c_comm);
return ;
} // end step3_gpu