CudaUtils.C File Reference

#include <stdio.h>
#include "common.h"
#include "charm++.h"
#include "CudaUtils.h"

Go to the source code of this file.

Functions

void cudaDie (const char *msg, cudaError_t err)
void cudaNAMD_bug (const char *msg)
void cuda_affinity_initialize ()
void clear_device_array_async_T (void *data, const int ndata, cudaStream_t stream, const size_t sizeofT)
void clear_device_array_T (void *data, const int ndata, const size_t sizeofT)
void allocate_host_T (void **pp, const int len, const size_t sizeofT)
void allocate_device_T (void **pp, const int len, const size_t sizeofT)
void deallocate_device_T (void **pp)
void deallocate_host_T (void **pp)
bool reallocate_device_T (void **pp, int *curlen, const int newlen, const float fac, 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 copy_HtoD_async_T (const void *h_array, void *d_array, int array_len, cudaStream_t stream, const size_t sizeofT)
void copy_HtoD_T (const void *h_array, void *d_array, int array_len, const size_t sizeofT)
void copy_DtoH_async_T (const void *d_array, void *h_array, const int array_len, cudaStream_t stream, const size_t sizeofT)
void copy_DtoH_T (const void *d_array, void *h_array, const int array_len, const size_t sizeofT)
void copy_DtoD_async_T (const void *d_src, void *d_dst, const int array_len, cudaStream_t stream, 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)
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 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 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 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)

Function Documentation

void allocate_device_T ( void **  pp,
const int  len,
const size_t  sizeofT 
)

Definition at line 81 of file CudaUtils.C.

References cudaCheck.

Referenced by allocate_device().

00081                                                                        {
00082   cudaCheck(cudaMalloc(pp, sizeofT*len));
00083 }

void allocate_host_T ( void **  pp,
const int  len,
const size_t  sizeofT 
)

Definition at line 71 of file CudaUtils.C.

References cudaCheck.

Referenced by allocate_host().

00071                                                                      {
00072   cudaCheck(cudaMallocHost(pp, sizeofT*len));
00073 }

void clear_device_array_async_T ( void *  data,
const int  ndata,
cudaStream_t  stream,
const size_t  sizeofT 
)

Definition at line 57 of file CudaUtils.C.

References cudaCheck.

Referenced by clear_device_array().

00057                                                                                                         {
00058   cudaCheck(cudaMemsetAsync(data, 0, sizeofT*ndata, stream));
00059 }

void clear_device_array_T ( void *  data,
const int  ndata,
const size_t  sizeofT 
)

Definition at line 61 of file CudaUtils.C.

References cudaCheck.

Referenced by clear_device_array_sync().

00061                                                                              {
00062   cudaCheck(cudaMemset(data, 0, sizeofT*ndata));
00063 }

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 
)

Definition at line 281 of file CudaUtils.C.

References cudaCheck.

Referenced by copy3D_DtoD().

00287                                        {
00288   cudaMemcpy3DParms parms = {0};
00289 
00290   parms.srcPos = make_cudaPos(sizeofT*src_x0, src_y0, src_z0);
00291   parms.srcPtr = make_cudaPitchedPtr(src_data, sizeofT*src_xsize, src_xsize, src_ysize);
00292 
00293   parms.dstPos = make_cudaPos(sizeofT*dst_x0, dst_y0, dst_z0);
00294   parms.dstPtr = make_cudaPitchedPtr(dst_data, sizeofT*dst_xsize, dst_xsize, dst_ysize);
00295 
00296   parms.extent = make_cudaExtent(sizeofT*width, height, depth);
00297   parms.kind = cudaMemcpyDeviceToDevice;
00298 
00299   cudaCheck(cudaMemcpy3DAsync(&parms, stream));
00300 }

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 
)

Definition at line 256 of file CudaUtils.C.

References cudaCheck.

Referenced by copy3D_DtoH().

00262                                        {
00263   cudaMemcpy3DParms parms = {0};
00264 
00265   parms.srcPos = make_cudaPos(sizeofT*src_x0, src_y0, src_z0);
00266   parms.srcPtr = make_cudaPitchedPtr(src_data, sizeofT*src_xsize, src_xsize, src_ysize);
00267 
00268   parms.dstPos = make_cudaPos(sizeofT*dst_x0, dst_y0, dst_z0);
00269   parms.dstPtr = make_cudaPitchedPtr(dst_data, sizeofT*dst_xsize, dst_xsize, dst_ysize);
00270 
00271   parms.extent = make_cudaExtent(sizeofT*width, height, depth);
00272   parms.kind = cudaMemcpyDeviceToHost;
00273 
00274   cudaCheck(cudaMemcpy3DAsync(&parms, stream));
00275 }

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 
)

Definition at line 231 of file CudaUtils.C.

References cudaCheck.

Referenced by copy3D_HtoD().

00237                                        {
00238   cudaMemcpy3DParms parms = {0};
00239 
00240   parms.srcPos = make_cudaPos(sizeofT*src_x0, src_y0, src_z0);
00241   parms.srcPtr = make_cudaPitchedPtr(src_data, sizeofT*src_xsize, src_xsize, src_ysize);
00242 
00243   parms.dstPos = make_cudaPos(sizeofT*dst_x0, dst_y0, dst_z0);
00244   parms.dstPtr = make_cudaPitchedPtr(dst_data, sizeofT*dst_xsize, dst_xsize, dst_ysize);
00245 
00246   parms.extent = make_cudaExtent(sizeofT*width, height, depth);
00247   parms.kind = cudaMemcpyHostToDevice;
00248 
00249   cudaCheck(cudaMemcpy3DAsync(&parms, stream));
00250 }

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 
)

Definition at line 306 of file CudaUtils.C.

References cudaCheck.

Referenced by copy3D_PeerDtoD().

00313                                        {
00314   cudaMemcpy3DPeerParms parms = {0};
00315 
00316   parms.srcDevice = src_dev;
00317   parms.dstDevice = dst_dev;
00318 
00319   parms.srcPos = make_cudaPos(sizeofT*src_x0, src_y0, src_z0);
00320   parms.srcPtr = make_cudaPitchedPtr(src_data, sizeofT*src_xsize, src_xsize, src_ysize);
00321 
00322   parms.dstPos = make_cudaPos(sizeofT*dst_x0, dst_y0, dst_z0);
00323   parms.dstPtr = make_cudaPitchedPtr(dst_data, sizeofT*dst_xsize, dst_xsize, dst_ysize);
00324 
00325   parms.extent = make_cudaExtent(sizeofT*width, height, depth);
00326 
00327   cudaCheck(cudaMemcpy3DPeerAsync(&parms, stream));
00328 }

void copy_DtoD_async_T ( const void *  d_src,
void *  d_dst,
const int  array_len,
cudaStream_t  stream,
const size_t  sizeofT 
)

Definition at line 208 of file CudaUtils.C.

References cudaCheck.

Referenced by copy_DtoD().

00209                                  {
00210   cudaCheck(cudaMemcpyAsync(d_dst, d_src, sizeofT*array_len, cudaMemcpyDeviceToDevice, stream));
00211 }

void copy_DtoD_T ( const void *  d_src,
void *  d_dst,
const int  array_len,
const size_t  sizeofT 
)

Definition at line 213 of file CudaUtils.C.

References cudaCheck.

Referenced by copy_DtoD_sync().

00213                                                                                             {
00214   cudaCheck(cudaMemcpy(d_dst, d_src, sizeofT*array_len, cudaMemcpyDeviceToDevice));
00215 }

void copy_DtoH_async_T ( const void *  d_array,
void *  h_array,
const int  array_len,
cudaStream_t  stream,
const size_t  sizeofT 
)

Definition at line 195 of file CudaUtils.C.

References cudaCheck.

Referenced by copy_DtoH().

00196                                  {
00197   cudaCheck(cudaMemcpyAsync(h_array, d_array, sizeofT*array_len, cudaMemcpyDeviceToHost, stream));
00198 }

void copy_DtoH_T ( const void *  d_array,
void *  h_array,
const int  array_len,
const size_t  sizeofT 
)

Definition at line 200 of file CudaUtils.C.

References cudaCheck.

Referenced by copy_DtoH_sync().

00200                                                                                                 {
00201   cudaCheck(cudaMemcpy(h_array, d_array, sizeofT*array_len, cudaMemcpyDeviceToHost));
00202 }

void copy_HtoD_async_T ( const void *  h_array,
void *  d_array,
int  array_len,
cudaStream_t  stream,
const size_t  sizeofT 
)

Definition at line 181 of file CudaUtils.C.

References cudaCheck.

Referenced by copy_HtoD().

00182                                  {
00183   cudaCheck(cudaMemcpyAsync(d_array, h_array, sizeofT*array_len, cudaMemcpyHostToDevice, stream));
00184 }

void copy_HtoD_T ( const void *  h_array,
void *  d_array,
int  array_len,
const size_t  sizeofT 
)

Definition at line 186 of file CudaUtils.C.

References cudaCheck.

Referenced by copy_HtoD_sync().

00187                            {
00188   cudaCheck(cudaMemcpy(d_array, h_array, sizeofT*array_len, cudaMemcpyHostToDevice));
00189 }

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 
)

Definition at line 221 of file CudaUtils.C.

References cudaCheck.

Referenced by copy_PeerDtoD().

00223                         {
00224   cudaCheck(cudaMemcpyPeerAsync(d_dst, dst_dev, d_src, src_dev, sizeofT*array_len, stream));
00225 }

void cuda_affinity_initialize (  ) 

Definition at line 33 of file CudaUtils.C.

Referenced by BackEnd::init().

00033                                 {
00034   // called before Converse/Charm++ initialization so can't use Cmi/Ck utilities
00035   char *forkstr;
00036   forkstr=getenv("CmiMyForks");
00037   int forks = 0;
00038   if (forkstr) sscanf(forkstr,"%d",&forks);
00039   if ( forks ) return;  // CUDA will fail on child if initialized before fork
00040   int devcnt = 0;
00041   cudaError_t err = cudaGetDeviceCount(&devcnt);
00042   if ( devcnt == 1 ) {  // only one device so it must be ours
00043     int *dummy;
00044     if ( err == cudaSuccess ) err = cudaSetDevice(0);
00045     if ( err == cudaSuccess ) err = cudaSetDeviceFlags(cudaDeviceMapHost);
00046     if ( err == cudaSuccess ) err = cudaMalloc(&dummy, 4);
00047   }
00048   if ( err != cudaSuccess ) {
00049     char host[128];
00050     gethostname(host, 128);  host[127] = 0;
00051     fprintf(stderr,"CUDA initialization error on %s: %s\n", host, cudaGetErrorString(err));
00052   }
00053 }

void cudaDie ( const char *  msg,
cudaError_t  err 
)

Definition at line 9 of file CudaUtils.C.

References NAMD_die().

Referenced by cuda_check_local_progress(), cuda_check_pme_charges(), cuda_check_pme_forces(), cuda_check_progress(), cuda_check_remote_progress(), DeviceCUDA::initialize(), and read_CUDA_ARCH().

00009                                                {
00010   char host[128];
00011   gethostname(host, 128);  host[127] = 0;
00012   char devstr[128] = "";
00013   int devnum;
00014   if ( cudaGetDevice(&devnum) == cudaSuccess ) {
00015     sprintf(devstr, " device %d", devnum);
00016   }
00017   cudaDeviceProp deviceProp;
00018   if ( cudaGetDeviceProperties(&deviceProp, devnum) == cudaSuccess ) {
00019     sprintf(devstr, " device %d pci %x:%x:%x", devnum,
00020       deviceProp.pciDomainID, deviceProp.pciBusID, deviceProp.pciDeviceID);
00021   }
00022   char errmsg[1024];
00023   if (err == cudaSuccess) {
00024     sprintf(errmsg,"CUDA error %s on Pe %d (%s%s)", msg, CkMyPe(), host, devstr);
00025   } else {
00026     sprintf(errmsg,"CUDA error %s on Pe %d (%s%s): %s", msg, CkMyPe(), host, devstr, cudaGetErrorString(err));    
00027   }
00028   NAMD_die(errmsg);
00029 }

void cudaNAMD_bug ( const char *  msg  ) 

Definition at line 31 of file CudaUtils.C.

References NAMD_bug().

Referenced by CudaFFTCompute::backward(), CudaFFTCompute::forward(), gather_force(), and spread_charge().

00031 {NAMD_bug(msg);}

void deallocate_device_T ( void **  pp  ) 

Definition at line 90 of file CudaUtils.C.

References cudaCheck.

Referenced by deallocate_device().

00090                                     {
00091   
00092   if (*pp != NULL) {
00093     cudaCheck(cudaFree((void *)(*pp)));
00094     *pp = NULL;
00095   }
00096 
00097 }

void deallocate_host_T ( void **  pp  ) 

Definition at line 104 of file CudaUtils.C.

References cudaCheck.

Referenced by deallocate_host().

00104                                   {
00105   
00106   if (*pp != NULL) {
00107     cudaCheck(cudaFreeHost((void *)(*pp)));
00108     *pp = NULL;
00109   }
00110 
00111 }

bool reallocate_device_T ( void **  pp,
int *  curlen,
const int  newlen,
const float  fac,
const size_t  sizeofT 
)

Definition at line 123 of file CudaUtils.C.

References cudaCheck.

Referenced by reallocate_device().

00123                                                                                                           {
00124 
00125   if (*pp != NULL && *curlen < newlen) {
00126     cudaCheck(cudaFree((void *)(*pp)));
00127     *pp = NULL;
00128   }
00129 
00130   if (*pp == NULL) {
00131     if (fac > 1.0f) {
00132       *curlen = (int)(((double)(newlen))*(double)fac);
00133     } else {
00134       *curlen = newlen;
00135     }
00136     cudaCheck(cudaMalloc(pp, sizeofT*(*curlen)));
00137     return true;
00138   }
00139 
00140   return false;
00141 }

bool reallocate_host_T ( void **  pp,
int *  curlen,
const int  newlen,
const float  fac,
const unsigned int  flag,
const size_t  sizeofT 
)

Definition at line 156 of file CudaUtils.C.

References cudaCheck.

Referenced by reallocate_host().

00157                                                                                        {
00158 
00159   if (*pp != NULL && *curlen < newlen) {
00160     cudaCheck(cudaFreeHost((void *)(*pp)));
00161     *pp = NULL;
00162   }
00163 
00164   if (*pp == NULL) {
00165     if (fac > 1.0f) {
00166       *curlen = (int)(((double)(newlen))*(double)fac);
00167     } else {
00168       *curlen = newlen;
00169     }
00170     cudaCheck(cudaHostAlloc(pp, sizeofT*(*curlen), flag));
00171     return true;
00172   }
00173 
00174   return false;
00175 }


Generated on 16 Sep 2019 for NAMD by  doxygen 1.6.1