7 #if defined(NAMD_CUDA) || defined(NAMD_HIP)
9 void cudaDie(
const char *msg, cudaError_t err) {
11 gethostname(host, 128); host[127] = 0;
12 char devstr[128] =
"";
14 if ( cudaGetDevice(&devnum) == cudaSuccess ) {
15 sprintf(devstr,
" device %d", devnum);
17 cudaDeviceProp deviceProp;
18 if ( cudaGetDeviceProperties(&deviceProp, devnum) == cudaSuccess ) {
19 sprintf(devstr,
" device %d pci %x:%x:%x", devnum,
20 deviceProp.pciDomainID, deviceProp.pciBusID, deviceProp.pciDeviceID);
23 if (err == cudaSuccess) {
24 sprintf(errmsg,
"CUDA error %s on Pe %d (%s%s)", msg, CkMyPe(), host, devstr);
26 sprintf(errmsg,
"CUDA error %s on Pe %d (%s%s): %s", msg, CkMyPe(), host, devstr, cudaGetErrorString(err));
35 cudaError_t err = cudaGetDeviceCount(&devcnt);
38 if ( err == cudaSuccess ) err = cudaSetDevice(0);
39 if ( err == cudaSuccess ) err = cudaSetDeviceFlags(cudaDeviceMapHost);
40 if ( err == cudaSuccess ) err = cudaMalloc(&dummy, 4);
42 if ( err != cudaSuccess ) {
44 gethostname(host, 128); host[127] = 0;
45 fprintf(stderr,
"CUDA initialization error on %s: %s\n", host, cudaGetErrorString(err));
52 cudaCheck(cudaMemsetAsync(data, 0, sizeofT*ndata, stream));
56 cudaCheck(cudaMemset(data, 0, sizeofT*ndata));
66 cudaCheck(cudaMallocHost(pp, sizeofT*len));
117 bool reallocate_device_T(
void **pp,
int *curlen,
const int newlen,
const float fac,
const size_t sizeofT) {
119 if (*pp != NULL && *curlen < newlen) {
126 *curlen = (int)(((
double)(newlen))*(double)fac);
130 cudaCheck(cudaMalloc(pp, sizeofT*(*curlen)));
151 const float fac,
const unsigned int flag,
const size_t sizeofT) {
153 if (*pp != NULL && *curlen < newlen) {
160 *curlen = (int)(((
double)(newlen))*(double)fac);
164 cudaCheck(cudaHostAlloc(pp, sizeofT*(*curlen), flag));
176 const size_t sizeofT) {
177 cudaCheck(cudaMemcpyAsync(d_array, h_array, sizeofT*array_len, cudaMemcpyHostToDevice, stream));
180 void copy_HtoD_T(
const void *h_array,
void *d_array,
int array_len,
181 const size_t sizeofT) {
182 cudaCheck(cudaMemcpy(d_array, h_array, sizeofT*array_len, cudaMemcpyHostToDevice));
190 const size_t sizeofT) {
191 cudaCheck(cudaMemcpyAsync(h_array, d_array, sizeofT*array_len, cudaMemcpyDeviceToHost, stream));
194 void copy_DtoH_T(
const void *d_array,
void *h_array,
const int array_len,
const size_t sizeofT) {
195 cudaCheck(cudaMemcpy(h_array, d_array, sizeofT*array_len, cudaMemcpyDeviceToHost));
203 const size_t sizeofT) {
204 cudaCheck(cudaMemcpyAsync(d_dst, d_src, sizeofT*array_len, cudaMemcpyDeviceToDevice, stream));
207 void copy_DtoD_T(
const void *d_src,
void *d_dst,
const int array_len,
const size_t sizeofT) {
208 cudaCheck(cudaMemcpy(d_dst, d_src, sizeofT*array_len, cudaMemcpyDeviceToDevice));
216 const void *d_src,
void *d_dst,
const int array_len, cudaStream_t
stream,
217 const size_t sizeofT) {
218 cudaCheck(cudaMemcpyPeerAsync(d_dst, dst_dev, d_src, src_dev, sizeofT*array_len, stream));
226 int src_x0,
int src_y0,
int src_z0,
227 size_t src_xsize,
size_t src_ysize,
228 int dst_x0,
int dst_y0,
int dst_z0,
229 size_t dst_xsize,
size_t dst_ysize,
230 size_t width,
size_t height,
size_t depth,
231 size_t sizeofT, cudaStream_t
stream) {
232 cudaMemcpy3DParms parms = {0};
234 parms.srcPos = make_cudaPos(sizeofT*src_x0, src_y0, src_z0);
235 parms.srcPtr = make_cudaPitchedPtr(src_data, sizeofT*src_xsize, src_xsize, src_ysize);
237 parms.dstPos = make_cudaPos(sizeofT*dst_x0, dst_y0, dst_z0);
238 parms.dstPtr = make_cudaPitchedPtr(dst_data, sizeofT*dst_xsize, dst_xsize, dst_ysize);
240 parms.extent = make_cudaExtent(sizeofT*width, height, depth);
242 parms.kind = cudaMemcpyHostToDevice;
244 cudaCheck(cudaMemcpy3DAsync(&parms, stream));
256 int src_x0,
int src_y0,
int src_z0,
257 size_t src_xsize,
size_t src_ysize,
258 int dst_x0,
int dst_y0,
int dst_z0,
259 size_t dst_xsize,
size_t dst_ysize,
260 size_t width,
size_t height,
size_t depth,
261 size_t sizeofT, cudaStream_t
stream) {
262 cudaMemcpy3DParms parms = {0};
264 parms.srcPos = make_cudaPos(sizeofT*src_x0, src_y0, src_z0);
265 parms.srcPtr = make_cudaPitchedPtr(src_data, sizeofT*src_xsize, src_xsize, src_ysize);
267 parms.dstPos = make_cudaPos(sizeofT*dst_x0, dst_y0, dst_z0);
268 parms.dstPtr = make_cudaPitchedPtr(dst_data, sizeofT*dst_xsize, dst_xsize, dst_ysize);
270 parms.extent = make_cudaExtent(sizeofT*width, height, depth);
271 parms.kind = cudaMemcpyDeviceToHost;
274 cudaCheck(cudaMemcpy3DAsync(&parms, stream));
286 int src_x0,
int src_y0,
int src_z0,
287 size_t src_xsize,
size_t src_ysize,
288 int dst_x0,
int dst_y0,
int dst_z0,
289 size_t dst_xsize,
size_t dst_ysize,
290 size_t width,
size_t height,
size_t depth,
291 size_t sizeofT, cudaStream_t
stream) {
292 cudaMemcpy3DParms parms = {0};
294 parms.srcPos = make_cudaPos(sizeofT*src_x0, src_y0, src_z0);
295 parms.srcPtr = make_cudaPitchedPtr(src_data, sizeofT*src_xsize, src_xsize, src_ysize);
297 parms.dstPos = make_cudaPos(sizeofT*dst_x0, dst_y0, dst_z0);
298 parms.dstPtr = make_cudaPitchedPtr(dst_data, sizeofT*dst_xsize, dst_xsize, dst_ysize);
300 parms.extent = make_cudaExtent(sizeofT*width, height, depth);
301 parms.kind = cudaMemcpyDeviceToDevice;
304 cudaCheck(cudaMemcpy3DAsync(&parms, stream));
316 void* src_data,
void* dst_data,
317 int src_x0,
int src_y0,
int src_z0,
318 size_t src_xsize,
size_t src_ysize,
319 int dst_x0,
int dst_y0,
int dst_z0,
320 size_t dst_xsize,
size_t dst_ysize,
321 size_t width,
size_t height,
size_t depth,
322 size_t sizeofT, cudaStream_t
stream) {
325 cudaDie(
"cudaMemcpy3DPeerAsync is not supported by HIP");
327 cudaMemcpy3DPeerParms parms = {0};
329 parms.srcDevice = src_dev;
330 parms.dstDevice = dst_dev;
332 parms.srcPos = make_cudaPos(sizeofT*src_x0, src_y0, src_z0);
333 parms.srcPtr = make_cudaPitchedPtr(src_data, sizeofT*src_xsize, src_xsize, src_ysize);
335 parms.dstPos = make_cudaPos(sizeofT*dst_x0, dst_y0, dst_z0);
336 parms.dstPtr = make_cudaPitchedPtr(dst_data, sizeofT*dst_xsize, dst_xsize, dst_ysize);
338 parms.extent = make_cudaExtent(sizeofT*width, height, depth);
340 cudaCheck(cudaMemcpy3DPeerAsync(&parms, stream));
void deallocate_device_T(void **pp)
void copy_DtoH_T(const void *d_array, void *h_array, const int array_len, const size_t sizeofT)
void copy3D_HtoD_T(void *src_data, void *dst_data, int src_x0, int src_y0, int src_z0, size_t src_xsize, size_t src_ysize, int dst_x0, int dst_y0, int dst_z0, size_t dst_xsize, size_t dst_ysize, size_t width, size_t height, size_t depth, size_t sizeofT, cudaStream_t stream)
void allocate_device_T(void **pp, const int len, const size_t sizeofT)
void copy_DtoD_T(const void *d_src, void *d_dst, const int array_len, const size_t sizeofT)
void copy_PeerDtoD_async_T(const int src_dev, const int dst_dev, const void *d_src, void *d_dst, const int array_len, cudaStream_t stream, const size_t sizeofT)
__thread cudaStream_t stream
void copy_DtoH_async_T(const void *d_array, void *h_array, const int array_len, cudaStream_t stream, const size_t sizeofT)
bool reallocate_host_T(void **pp, int *curlen, const int newlen, const float fac, const unsigned int flag, const size_t sizeofT)
void NAMD_bug(const char *err_msg)
void copy3D_DtoD_T(void *src_data, void *dst_data, int src_x0, int src_y0, int src_z0, size_t src_xsize, size_t src_ysize, int dst_x0, int dst_y0, int dst_z0, size_t dst_xsize, size_t dst_ysize, size_t width, size_t height, size_t depth, size_t sizeofT, cudaStream_t stream)
void copy_HtoD_T(const void *h_array, void *d_array, int array_len, const size_t sizeofT)
void cudaDie(const char *msg, cudaError_t err=cudaSuccess)
void NAMD_die(const char *err_msg)
void clear_device_array_async_T(void *data, const int ndata, cudaStream_t stream, const size_t sizeofT)
void cudaNAMD_bug(const char *msg)
void copy_DtoD_async_T(const void *d_src, void *d_dst, const int array_len, cudaStream_t stream, const size_t sizeofT)
bool reallocate_device_T(void **pp, int *curlen, const int newlen, const float fac, const size_t sizeofT)
void copy3D_PeerDtoD_T(int src_dev, int dst_dev, void *src_data, void *dst_data, int src_x0, int src_y0, int src_z0, size_t src_xsize, size_t src_ysize, int dst_x0, int dst_y0, int dst_z0, size_t dst_xsize, size_t dst_ysize, size_t width, size_t height, size_t depth, size_t sizeofT, cudaStream_t stream)
void copy3D_DtoH_T(void *src_data, void *dst_data, int src_x0, int src_y0, int src_z0, size_t src_xsize, size_t src_ysize, int dst_x0, int dst_y0, int dst_z0, size_t dst_xsize, size_t dst_ysize, size_t width, size_t height, size_t depth, size_t sizeofT, cudaStream_t stream)
void clear_device_array_T(void *data, const int ndata, const size_t sizeofT)
void deallocate_host_T(void **pp)
void cuda_affinity_initialize()
void allocate_host_T(void **pp, const int len, const size_t sizeofT)
void copy_HtoD_async_T(const void *h_array, void *d_array, int array_len, cudaStream_t stream, const size_t sizeofT)