lammps/lib/cuda/comm_cuda.cu

540 lines
18 KiB
Plaintext

/* ----------------------------------------------------------------------
LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
Original Version:
http://lammps.sandia.gov, Sandia National Laboratories
Steve Plimpton, sjplimp@sandia.gov
See the README file in the top-level LAMMPS directory.
-----------------------------------------------------------------------
USER-CUDA Package and associated modifications:
https://sourceforge.net/projects/lammpscuda/
Christian Trott, christian.trott@tu-ilmenau.de
Lars Winterfeld, lars.winterfeld@tu-ilmenau.de
Theoretical Physics II, University of Technology Ilmenau, Germany
See the README file in the USER-CUDA directory.
This software is distributed under the GNU General Public License.
------------------------------------------------------------------------- */
#include <stdio.h>
#define MY_PREFIX comm_cuda
#include "cuda_shared.h"
#include "cuda_common.h"
#include "crm_cuda_utils.cu"
#include "comm_cuda_cu.h"
#include "comm_cuda_kernel.cu"
#include <ctime>
void Cuda_CommCuda_UpdateBuffer(cuda_shared_data* sdata, int n)
{
int size = n * 3 * sizeof(X_FLOAT);
if(sdata->buffersize < size) {
MYDBG(printf("Cuda_ComputeTempCuda Resizing Buffer at %p with %i kB to\n", sdata->buffer, sdata->buffersize);)
CudaWrapper_FreeCudaData(sdata->buffer, sdata->buffersize);
sdata->buffer = CudaWrapper_AllocCudaData(size);
sdata->buffersize = size;
sdata->buffer_new++;
MYDBG(printf("New buffer at %p with %i kB\n", sdata->buffer, sdata->buffersize);)
}
cudaMemcpyToSymbol(MY_AP(buffer), & sdata->buffer, sizeof(int*));
}
void Cuda_CommCuda_UpdateNmax(cuda_shared_data* sdata)
{
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
cudaMemcpyToSymbol(MY_AP(nmax) , & sdata->atom.nmax , sizeof(int));
cudaMemcpyToSymbol(MY_AP(x) , & sdata->atom.x .dev_data, sizeof(X_FLOAT*));
cudaMemcpyToSymbol(MY_AP(v) , & sdata->atom.v .dev_data, sizeof(X_FLOAT*));
cudaMemcpyToSymbol(MY_AP(f) , & sdata->atom.f .dev_data, sizeof(F_FLOAT*));
cudaMemcpyToSymbol(MY_AP(type) , & sdata->atom.type .dev_data, sizeof(int*));
}
void Cuda_CommCuda_Init(cuda_shared_data* sdata)
{
Cuda_CommCuda_UpdateNmax(sdata);
int ntypesp = sdata->atom.ntypes + 1;
cudaMemcpyToSymbol(MY_AP(cuda_ntypes) , &ntypesp, sizeof(int));
cudaMemcpyToSymbol(MY_AP(prd) , sdata->domain.prd, 3 * sizeof(X_FLOAT));
cudaMemcpyToSymbol(MY_AP(flag) , &sdata->flag, sizeof(int*));
cudaMemcpyToSymbol(MY_AP(debugdata) , &sdata->debugdata, sizeof(int*));
}
int Cuda_CommCuda_PackComm(cuda_shared_data* sdata, int n, int iswap, void* buf_send, int* pbc, int pbc_flag)
{
timespec time1, time2;
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
int size = n * 3 * sizeof(X_FLOAT);
if(sdata->buffer_new or (size > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, n);
X_FLOAT dx = 0.0;
X_FLOAT dy = 0.0;
X_FLOAT dz = 0.0;
if(pbc_flag != 0) {
if(sdata->domain.triclinic == 0) {
dx = pbc[0] * sdata->domain.prd[0];
dy = pbc[1] * sdata->domain.prd[1];
dz = pbc[2] * sdata->domain.prd[2];
} else {
dx = pbc[0] * sdata->domain.prd[0] + pbc[5] * sdata->domain.xy + pbc[4] * sdata->domain.xz;
dy = pbc[1] * sdata->domain.prd[1] + pbc[3] * sdata->domain.yz;
dz = pbc[2] * sdata->domain.prd[2];
}
}
int3 layout = getgrid(n);
dim3 threads(layout.z, 1, 1);
dim3 grid(layout.x, layout.y, 1);
if(sdata->atom.nlocal > 0) {
cudaMemset(sdata->flag, 0, sizeof(int));
clock_gettime(CLOCK_REALTIME, &time1);
void* buf = sdata->overlap_comm ? sdata->comm.buf_send_dev[iswap] : sdata->buffer;
Cuda_CommCuda_PackComm_Kernel <<< grid, threads, 0>>>((int*) sdata->comm.sendlist.dev_data, n
, sdata->comm.maxlistlength, iswap, dx, dy, dz, buf);
cudaThreadSynchronize();
clock_gettime(CLOCK_REALTIME, &time2);
sdata->cuda_timings.comm_forward_kernel_pack +=
time2.tv_sec - time1.tv_sec + 1.0 * (time2.tv_nsec - time1.tv_nsec) / 1000000000;
CUT_CHECK_ERROR("Cuda_CommCuda_PackComm: Kernel execution failed");
if(not sdata->overlap_comm)
cudaMemcpy(buf_send, sdata->buffer, n * 3 * sizeof(X_FLOAT), cudaMemcpyDeviceToHost);
//cudaMemcpy(buf_send, sdata->comm.buf_send_dev[iswap], n*3*sizeof(X_FLOAT), cudaMemcpyDeviceToHost);
clock_gettime(CLOCK_REALTIME, &time1);
sdata->cuda_timings.comm_forward_download +=
time1.tv_sec - time2.tv_sec + 1.0 * (time1.tv_nsec - time2.tv_nsec) / 1000000000;
int aflag;
cudaMemcpy(&aflag, sdata->flag, sizeof(int), cudaMemcpyDeviceToHost);
if(aflag != 0) printf("aflag PackComm: %i\n", aflag);
CUT_CHECK_ERROR("Cuda_CommCuda_PackComm: Kernel execution failed");
}
return 3 * n;
}
int Cuda_CommCuda_PackCommVel(cuda_shared_data* sdata, int n, int iswap, void* buf_send, int* pbc, int pbc_flag)
{
timespec time1, time2;
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
int size = n * 6 * sizeof(X_FLOAT);
if(sdata->buffer_new or (size > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, n);
X_FLOAT dx = 0.0;
X_FLOAT dy = 0.0;
X_FLOAT dz = 0.0;
if(pbc_flag != 0) {
if(sdata->domain.triclinic == 0) {
dx = pbc[0] * sdata->domain.prd[0];
dy = pbc[1] * sdata->domain.prd[1];
dz = pbc[2] * sdata->domain.prd[2];
} else {
dx = pbc[0] * sdata->domain.prd[0] + pbc[5] * sdata->domain.xy + pbc[4] * sdata->domain.xz;
dy = pbc[1] * sdata->domain.prd[1] + pbc[3] * sdata->domain.yz;
dz = pbc[2] * sdata->domain.prd[2];
}
}
int3 layout = getgrid(n);
dim3 threads(layout.z, 1, 1);
dim3 grid(layout.x, layout.y, 1);
if(sdata->atom.nlocal > 0) {
cudaMemset(sdata->flag, 0, sizeof(int));
clock_gettime(CLOCK_REALTIME, &time1);
void* buf = sdata->overlap_comm ? sdata->comm.buf_send_dev[iswap] : sdata->buffer;
Cuda_CommCuda_PackComm_Kernel <<< grid, threads, 0>>>((int*) sdata->comm.sendlist.dev_data, n
, sdata->comm.maxlistlength, iswap, dx, dy, dz, buf);
cudaThreadSynchronize();
clock_gettime(CLOCK_REALTIME, &time2);
sdata->cuda_timings.comm_forward_kernel_pack +=
time2.tv_sec - time1.tv_sec + 1.0 * (time2.tv_nsec - time1.tv_nsec) / 1000000000;
CUT_CHECK_ERROR("Cuda_CommCuda_PackComm: Kernel execution failed");
if(not sdata->overlap_comm)
cudaMemcpy(buf_send, sdata->buffer, n * 6 * sizeof(X_FLOAT), cudaMemcpyDeviceToHost);
//cudaMemcpy(buf_send, sdata->comm.buf_send_dev[iswap], n*3*sizeof(X_FLOAT), cudaMemcpyDeviceToHost);
clock_gettime(CLOCK_REALTIME, &time1);
sdata->cuda_timings.comm_forward_download +=
time1.tv_sec - time2.tv_sec + 1.0 * (time1.tv_nsec - time2.tv_nsec) / 1000000000;
int aflag;
cudaMemcpy(&aflag, sdata->flag, sizeof(int), cudaMemcpyDeviceToHost);
if(aflag != 0) printf("aflag PackComm: %i\n", aflag);
CUT_CHECK_ERROR("Cuda_CommCuda_PackComm: Kernel execution failed");
}
return 6 * n;
}
int Cuda_CommCuda_PackComm_Self(cuda_shared_data* sdata, int n, int iswap, int first, int* pbc, int pbc_flag)
{
MYDBG(printf(" # CUDA: CommCuda_PackComm_Self\n");)
timespec time1, time2;
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
int size = n * 3 * sizeof(X_FLOAT);
if(sdata->buffer_new or (size > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, n);
static int count = -1;
count++;
X_FLOAT dx = 0.0;
X_FLOAT dy = 0.0;
X_FLOAT dz = 0.0;
if(pbc_flag != 0) {
if(sdata->domain.triclinic == 0) {
dx = pbc[0] * sdata->domain.prd[0];
dy = pbc[1] * sdata->domain.prd[1];
dz = pbc[2] * sdata->domain.prd[2];
} else {
dx = pbc[0] * sdata->domain.prd[0] + pbc[5] * sdata->domain.xy + pbc[4] * sdata->domain.xz;
dy = pbc[1] * sdata->domain.prd[1] + pbc[3] * sdata->domain.yz;
dz = pbc[2] * sdata->domain.prd[2];
}
}
int3 layout = getgrid(n);
dim3 threads(layout.z, 1, 1);
dim3 grid(layout.x, layout.y, 1);
if(sdata->atom.nlocal > 0) {
clock_gettime(CLOCK_REALTIME, &time1);
Cuda_CommCuda_PackComm_Self_Kernel <<< grid, threads, 0>>>((int*) sdata->comm.sendlist.dev_data, n, sdata->comm.maxlistlength, iswap, dx, dy, dz, first);
cudaThreadSynchronize();
clock_gettime(CLOCK_REALTIME, &time2);
sdata->cuda_timings.comm_forward_kernel_self +=
time2.tv_sec - time1.tv_sec + 1.0 * (time2.tv_nsec - time1.tv_nsec) / 1000000000;
CUT_CHECK_ERROR("Cuda_CommCuda_PackComm_Self: Kernel execution failed");
}
return 3 * n;
}
int Cuda_CommCuda_PackCommVel_Self(cuda_shared_data* sdata, int n, int iswap, int first, int* pbc, int pbc_flag)
{
MYDBG(printf(" # CUDA: CommCuda_PackComm_Self\n");)
timespec time1, time2;
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
int size = n * 6 * sizeof(X_FLOAT);
if(sdata->buffer_new or (size > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, n);
static int count = -1;
count++;
X_FLOAT dx = 0.0;
X_FLOAT dy = 0.0;
X_FLOAT dz = 0.0;
if(pbc_flag != 0) {
if(sdata->domain.triclinic == 0) {
dx = pbc[0] * sdata->domain.prd[0];
dy = pbc[1] * sdata->domain.prd[1];
dz = pbc[2] * sdata->domain.prd[2];
} else {
dx = pbc[0] * sdata->domain.prd[0] + pbc[5] * sdata->domain.xy + pbc[4] * sdata->domain.xz;
dy = pbc[1] * sdata->domain.prd[1] + pbc[3] * sdata->domain.yz;
dz = pbc[2] * sdata->domain.prd[2];
}
}
int3 layout = getgrid(n);
dim3 threads(layout.z, 1, 1);
dim3 grid(layout.x, layout.y, 1);
if(sdata->atom.nlocal > 0) {
clock_gettime(CLOCK_REALTIME, &time1);
Cuda_CommCuda_PackComm_Self_Kernel <<< grid, threads, 0>>>((int*) sdata->comm.sendlist.dev_data, n, sdata->comm.maxlistlength, iswap, dx, dy, dz, first);
cudaThreadSynchronize();
clock_gettime(CLOCK_REALTIME, &time2);
sdata->cuda_timings.comm_forward_kernel_self +=
time2.tv_sec - time1.tv_sec + 1.0 * (time2.tv_nsec - time1.tv_nsec) / 1000000000;
CUT_CHECK_ERROR("Cuda_CommCuda_PackComm_Self: Kernel execution failed");
}
return 6 * n;
}
void Cuda_CommCuda_UnpackComm(cuda_shared_data* sdata, int n, int first, void* buf_recv, int iswap)
{
timespec time1, time2;
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
int size = n * 3 * sizeof(X_FLOAT);
if(sdata->buffer_new or (size > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, n);
int3 layout = getgrid(n);
dim3 threads(layout.z, 1, 1);
dim3 grid(layout.x, layout.y, 1);
if(sdata->atom.nlocal > 0) {
clock_gettime(CLOCK_REALTIME, &time1);
if(not sdata->overlap_comm || iswap < 0)
cudaMemcpy(sdata->buffer, (void*)buf_recv, n * 3 * sizeof(X_FLOAT), cudaMemcpyHostToDevice);
clock_gettime(CLOCK_REALTIME, &time2);
sdata->cuda_timings.comm_forward_upload +=
time2.tv_sec - time1.tv_sec + 1.0 * (time2.tv_nsec - time1.tv_nsec) / 1000000000;
void* buf = (sdata->overlap_comm && iswap >= 0) ? sdata->comm.buf_recv_dev[iswap] : sdata->buffer;
Cuda_CommCuda_UnpackComm_Kernel <<< grid, threads, 0>>>(n, first, buf);
cudaThreadSynchronize();
clock_gettime(CLOCK_REALTIME, &time1);
sdata->cuda_timings.comm_forward_kernel_unpack +=
time1.tv_sec - time2.tv_sec + 1.0 * (time1.tv_nsec - time2.tv_nsec) / 1000000000;
CUT_CHECK_ERROR("Cuda_CommCuda_UnpackComm: Kernel execution failed");
}
}
void Cuda_CommCuda_UnpackCommVel(cuda_shared_data* sdata, int n, int first, void* buf_recv, int iswap)
{
timespec time1, time2;
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
int size = n * 6 * sizeof(X_FLOAT);
if(sdata->buffer_new or (size > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, n);
int3 layout = getgrid(n);
dim3 threads(layout.z, 1, 1);
dim3 grid(layout.x, layout.y, 1);
if(sdata->atom.nlocal > 0) {
clock_gettime(CLOCK_REALTIME, &time1);
if(not sdata->overlap_comm || iswap < 0)
cudaMemcpy(sdata->buffer, (void*)buf_recv, n * 6 * sizeof(X_FLOAT), cudaMemcpyHostToDevice);
clock_gettime(CLOCK_REALTIME, &time2);
sdata->cuda_timings.comm_forward_upload +=
time2.tv_sec - time1.tv_sec + 1.0 * (time2.tv_nsec - time1.tv_nsec) / 1000000000;
void* buf = (sdata->overlap_comm && iswap >= 0) ? sdata->comm.buf_recv_dev[iswap] : sdata->buffer;
Cuda_CommCuda_UnpackComm_Kernel <<< grid, threads, 0>>>(n, first, buf);
cudaThreadSynchronize();
clock_gettime(CLOCK_REALTIME, &time1);
sdata->cuda_timings.comm_forward_kernel_unpack +=
time1.tv_sec - time2.tv_sec + 1.0 * (time1.tv_nsec - time2.tv_nsec) / 1000000000;
CUT_CHECK_ERROR("Cuda_CommCuda_UnpackComm: Kernel execution failed");
}
}
int Cuda_CommCuda_PackReverse(cuda_shared_data* sdata, int n, int first, void* buf_send)
{
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
int size = n * 3 * sizeof(F_FLOAT);
if(sdata->buffer_new or (size > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, n);
F_FLOAT* buf = (F_FLOAT*)buf_send;
F_FLOAT* f_dev = (F_FLOAT*)sdata->atom.f.dev_data;
f_dev += first;
cudaMemcpy(buf, f_dev, n * sizeof(F_FLOAT), cudaMemcpyDeviceToHost);
buf += n;
f_dev += sdata->atom.nmax;
cudaMemcpy(buf, f_dev, n * sizeof(F_FLOAT), cudaMemcpyDeviceToHost);
buf += n;
f_dev += sdata->atom.nmax;
cudaMemcpy(buf, f_dev, n * sizeof(F_FLOAT), cudaMemcpyDeviceToHost);
return n * 3;
}
void Cuda_CommCuda_UnpackReverse(cuda_shared_data* sdata, int n, int iswap, void* buf_recv)
{
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
int size = n * 3 * sizeof(F_FLOAT);
if(sdata->buffer_new or (size > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, n);
int3 layout = getgrid(n);
dim3 threads(layout.z, 1, 1);
dim3 grid(layout.x, layout.y, 1);
if(sdata->atom.nlocal > 0) {
cudaMemcpy(sdata->buffer, buf_recv, size, cudaMemcpyHostToDevice);
Cuda_CommCuda_UnpackReverse_Kernel <<< grid, threads, 0>>>((int*) sdata->comm.sendlist.dev_data, n, sdata->comm.maxlistlength, iswap);
cudaThreadSynchronize();
CUT_CHECK_ERROR("Cuda_CommCuda_UnpackReverse: Kernel execution failed");
}
}
void Cuda_CommCuda_UnpackReverse_Self(cuda_shared_data* sdata, int n, int iswap, int first)
{
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
int size = n * 3 * sizeof(X_FLOAT);
if(sdata->buffer_new or (size > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, n);
int3 layout = getgrid(n);
dim3 threads(layout.z, 1, 1);
dim3 grid(layout.x, layout.y, 1);
if(sdata->atom.nlocal > 0) {
Cuda_CommCuda_UnpackReverse_Self_Kernel <<< grid, threads, 0>>>((int*) sdata->comm.sendlist.dev_data, n, sdata->comm.maxlistlength, iswap, first);
cudaThreadSynchronize();
CUT_CHECK_ERROR("Cuda_CommCuda_PackReverse_Self: Kernel execution failed");
}
}
int Cuda_CommCuda_BuildSendlist(cuda_shared_data* sdata, int bordergroup, int ineed, int style, int atom_nfirst, int nfirst, int nlast, int dim, int iswap)
{
MYDBG(printf(" # CUDA: CommCuda_BuildSendlist\n");)
timespec time1, time2;
if(sdata->atom.update_nmax)
Cuda_CommCuda_UpdateNmax(sdata);
if(sdata->atom.update_nlocal)
cudaMemcpyToSymbol(MY_AP(nlocal) , & sdata->atom.nlocal , sizeof(int));
if(sdata->buffer_new or (80 > sdata->buffersize))
Cuda_CommCuda_UpdateBuffer(sdata, 10);
int n;
if(!bordergroup || ineed >= 2)
n = nlast - nfirst + 1;
else {
n = atom_nfirst;
if(nlast - sdata->atom.nlocal + 1 > n) n = nlast - sdata->atom.nlocal + 1;
}
int3 layout = getgrid(n, 0, 512, true);
dim3 threads(layout.z, 1, 1);
dim3 grid(layout.x + 1, layout.y, 1);
cudaMemset((int*)(sdata->buffer), 0, sizeof(int));
clock_gettime(CLOCK_REALTIME, &time1);
if(style == 1)
Cuda_CommCuda_BuildSendlist_Single <<< grid, threads, (threads.x + 1)*sizeof(int) >>> (bordergroup, ineed, atom_nfirst, nfirst, nlast, dim, iswap, (X_FLOAT*) sdata->comm.slablo.dev_data, (X_FLOAT*) sdata->comm.slabhi.dev_data, (int*) sdata->comm.sendlist.dev_data, sdata->comm.maxlistlength);
else
Cuda_CommCuda_BuildSendlist_Multi <<< grid, threads, (threads.x + 1)*sizeof(int) >>> (bordergroup, ineed, atom_nfirst, nfirst, nlast, dim, iswap, (X_FLOAT*) sdata->comm.multilo.dev_data, (X_FLOAT*) sdata->comm.multihi.dev_data, (int*) sdata->comm.sendlist.dev_data, sdata->comm.maxlistlength);
cudaThreadSynchronize();
clock_gettime(CLOCK_REALTIME, &time2);
sdata->cuda_timings.comm_border_kernel_buildlist +=
time2.tv_sec - time1.tv_sec + 1.0 * (time2.tv_nsec - time1.tv_nsec) / 1000000000;
CUT_CHECK_ERROR("Cuda_CommCuda_BuildSendlist: Kernel execution failed");
int nsend;
cudaMemcpy(&nsend, sdata->buffer, sizeof(int), cudaMemcpyDeviceToHost);
return nsend;
}