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 grad(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); 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); //data=(float*)accfft_alloc(isize[0]*isize[1]*isize[2]*sizeof(float)); float * data_cpu=(float*)accfft_alloc(alloc_max); float* data; Complexf* data_hat; cudaMalloc((void**) &data , alloc_max); cudaMalloc((void**) &data_hat, alloc_max); accfft_init(); /* Create FFT plan */ setup_time=-MPI_Wtime(); accfft_plan_gpuf * plan=accfft_plan_dft_3d_r2c_gpuf(n,data,(float*)data_hat,c_comm,ACCFFT_MEASURE); setup_time+=MPI_Wtime(); /* Initialize data */ initialize(data_cpu,n,c_comm); cudaMemcpy(data, data_cpu,alloc_max, cudaMemcpyHostToDevice); MPI_Barrier(c_comm); float * gradx,*grady, *gradz; cudaMalloc((void**) &gradx , alloc_max); cudaMalloc((void**) &grady , alloc_max); cudaMalloc((void**) &gradz , alloc_max); double timings[5]={0}; std::bitset<3> XYZ=0; XYZ[0]=1; XYZ[1]=1; XYZ[2]=1; double exec_time=-MPI_Wtime(); accfft_grad_gpuf(gradx,grady,gradz,data,plan,XYZ,timings); exec_time+=MPI_Wtime(); /* Check err*/ PCOUT<<">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>"<<std::endl; PCOUT<<">>>>>>>>Checking Gradx>>>>>>>>"<<std::endl; cudaMemcpy(data_cpu, gradx, alloc_max, cudaMemcpyDeviceToHost); check_err_grad(data_cpu,n,c_comm,0); PCOUT<<"<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<"<<std::endl; PCOUT<<"<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n"<<std::endl; PCOUT<<">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>"<<std::endl; PCOUT<<">>>>>>>>Checking Grady>>>>>>>>"<<std::endl; cudaMemcpy(data_cpu, grady, alloc_max, cudaMemcpyDeviceToHost); check_err_grad(data_cpu,n,c_comm,1); PCOUT<<"<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<"<<std::endl; PCOUT<<"<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n"<<std::endl; PCOUT<<">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>"<<std::endl; PCOUT<<">>>>>>>>Checking Gradz>>>>>>>>"<<std::endl; cudaMemcpy(data_cpu, gradz, alloc_max, cudaMemcpyDeviceToHost); check_err_grad(data_cpu,n,c_comm,2); PCOUT<<"<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<"<<std::endl; PCOUT<<"<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n"<<std::endl; /* Compute some timings statistics */ double g_setup_time,g_timings[5],g_exec_time; MPI_Reduce(timings,g_timings,5, MPI_DOUBLE, MPI_MAX,0, c_comm); MPI_Reduce(&setup_time,&g_setup_time,1, MPI_DOUBLE, MPI_MAX,0, c_comm); MPI_Reduce(&exec_time,&g_exec_time,1, MPI_DOUBLE, MPI_MAX,0, c_comm); PCOUT<<"Timing for Grad Computation for size "<<n[0]<<"*"<<n[1]<<"*"<<n[2]<<std::endl; PCOUT<<"Setup \t\t"<<g_setup_time<<std::endl; PCOUT<<"Evaluation \t"<<g_exec_time<<std::endl; accfft_free(data_cpu); cudaFree(data); cudaFree(data_hat); MPI_Barrier(c_comm); cudaFree(gradx); cudaFree(grady); cudaFree(gradz); accfft_destroy_plan(plan); accfft_cleanup_gpu(); MPI_Comm_free(&c_comm); PCOUT<<"-------------------------------------------------------"<<std::endl; PCOUT<<"-------------------------------------------------------"<<std::endl; PCOUT<<"-------------------------------------------------------\n"<<std::endl; return ; } // end grad