portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 20:57:06 +01:00
Code Releases Activity

Merge branch 'develop' into feature/dirichlet

Browse Source
This commit is contained in:
Peter Boyle
2022-08-31 19:05:57 -04:00
parent 442bfb3d42 913fbca74a
commit 1177b8f661
15 changed files with 167 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
Grid/algorithms/CoarsenedMatrix.h
View File

@ -262,7 +262,7 @@ public:
autoView( Tnp_v , (*Tnp), AcceleratorWrite);
autoView( Tnm_v , (*Tnm), AcceleratorWrite);
const int Nsimd = CComplex::Nsimd();
accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
accelerator_for(ss, FineGrid->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
});

2
Grid/algorithms/approx/Chebyshev.h
View File

@ -264,7 +264,7 @@ public:
auto Tnp_v = Tnp->View();
auto Tnm_v = Tnm->View();
constexpr int Nsimd = vector_type::Nsimd();
accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
accelerator_for(ss, in.Grid()->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
});

11
Grid/communicator/Communicator_mpi3.cc
View File

@ -395,12 +395,11 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
}
}
/* if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
* this->StencilSendToRecvFromComplete(list,dir);
* list.resize(0);
* }
*/
// if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
// this->StencilSendToRecvFromComplete(list,dir);
// }
return off_node_bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)

3
Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
View File

@ -490,6 +490,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDag); return;}
#endif
acceleratorFenceComputeStream();
} else if( interior ) {
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt); return;}
@ -497,11 +498,13 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDagInt); return;}
#endif
} else if( exterior ) {
acceleratorFenceComputeStream();
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt); return;}
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDagExt); return;}
#endif
acceleratorFenceComputeStream();
}
assert(0 && " Kernel optimisation case not covered ");
}

2
Grid/qcd/action/gauge/Photon.h
View File

@ -49,7 +49,7 @@ NAMESPACE_BEGIN(Grid);
typedef Lattice<SiteLink> LinkField;
typedef Lattice<SiteField> Field;
typedef Field ComplexField;
typedef LinkField ComplexField;
};
typedef QedGImpl<vComplex> QedGImplR;

3
Grid/threads/Accelerator.cc
View File

@ -201,12 +201,15 @@ void acceleratorInit(void)
#ifdef GRID_SYCL
cl::sycl::queue *theGridAccelerator;
cl::sycl::queue *theCopyAccelerator;
void acceleratorInit(void)
{
int nDevices = 1;
cl::sycl::gpu_selector selector;
cl::sycl::device selectedDevice { selector };
theGridAccelerator = new sycl::queue (selectedDevice);
// theCopyAccelerator = new sycl::queue (selectedDevice);
theCopyAccelerator = theGridAccelerator; // Should proceed concurrenlty anyway.
#ifdef GRID_SYCL_LEVEL_ZERO_IPC
zeInit(0);

29
Grid/threads/Accelerator.h
View File

@ -248,7 +248,6 @@ inline int acceleratorIsCommunicable(void *ptr)
//////////////////////////////////////////////
// SyCL acceleration
//////////////////////////////////////////////
#ifdef GRID_SYCL
NAMESPACE_END(Grid);
#include <CL/sycl.hpp>
@ -263,6 +262,7 @@ NAMESPACE_END(Grid);
NAMESPACE_BEGIN(Grid);
extern cl::sycl::queue *theGridAccelerator;
extern cl::sycl::queue *theCopyAccelerator;
#ifdef __SYCL_DEVICE_ONLY__
#define GRID_SIMT
@ -290,7 +290,7 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
cgh.parallel_for( \
cl::sycl::nd_range<3>(global,local), \
[=] (cl::sycl::nd_item<3> item) /*mutable*/ \
[[intel::reqd_sub_group_size(8)]] \
[[intel::reqd_sub_group_size(16)]] \
{ \
auto iter1 = item.get_global_id(0); \
auto iter2 = item.get_global_id(1); \
@ -299,19 +299,19 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
}); \
});
#define accelerator_barrier(dummy) theGridAccelerator->wait();
#define accelerator_barrier(dummy) { printf(" theGridAccelerator::wait()\n"); theGridAccelerator->wait(); }
inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) {
theGridAccelerator->memcpy(to,from,bytes);
}
inline void acceleratorCopySynchronise(void) { theGridAccelerator->wait(); std::cout<<"acceleratorCopySynchronise() wait "<<std::endl; }
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { theGridAccelerator->memset(base,value,bytes); theGridAccelerator->wait();}
inline void acceleratorCopySynchronise(void) { printf(" theCopyAccelerator::wait()\n"); theCopyAccelerator->wait(); }
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes);}
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { theCopyAccelerator->memset(base,value,bytes); theCopyAccelerator->wait();}
inline int acceleratorIsCommunicable(void *ptr)
{
#if 0
@ -514,7 +514,16 @@ inline void *acceleratorAllocCpu(size_t bytes){return memalign(GRID_ALLOC_ALIGN,
inline void acceleratorFreeCpu (void *ptr){free(ptr);};
#endif
//////////////////////////////////////////////
// Fencing needed ONLY for SYCL
//////////////////////////////////////////////
#ifdef GRID_SYCL
inline void acceleratorFenceComputeStream(void){ accelerator_barrier();};
#else
// Ordering within a stream guaranteed on Nvidia & AMD
inline void acceleratorFenceComputeStream(void){ };
#endif
///////////////////////////////////////////////////
// Synchronise across local threads for divergence resynch

7
Grid/util/Sha.h
View File

@ -27,6 +27,7 @@
/* END LEGAL */
extern "C" {
#include <openssl/sha.h>
#include <openssl/evp.h>
}
#ifdef USE_IPP
#include "ipp.h"
@ -70,10 +71,8 @@ public:
static inline std::vector<unsigned char> sha256(const void *data,size_t bytes)
{
std::vector<unsigned char> hash(SHA256_DIGEST_LENGTH);
SHA256_CTX sha256;
SHA256_Init (&sha256);
SHA256_Update(&sha256, data,bytes);
SHA256_Final (&hash[0], &sha256);
auto digest = EVP_get_digestbyname("SHA256");
EVP_Digest(data, bytes, &hash[0], NULL, digest, NULL);
return hash;
}
static inline std::vector<int> sha256_seeds(const std::string &s)