portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-19 08:17:05 +01:00
Code Releases Activity

Merge branch 'develop' into feature/scidac-wp1

Browse Source
This commit is contained in:
Peter Boyle
2024-03-06 14:55:21 -05:00
parent 070b61f08f 7e5bd46dd3
commit cc04dc42dc
103 changed files with 5678 additions and 589 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
Grid/threads/Accelerator.cc
View File

@ -7,6 +7,8 @@ uint32_t accelerator_threads=2;
uint32_t acceleratorThreads(void) {return accelerator_threads;};
void acceleratorThreads(uint32_t t) {accelerator_threads = t;};
#define ENV_LOCAL_RANK_PALS "PALS_LOCAL_RANKID"
#define ENV_RANK_PALS "PALS_RANKID"
#define ENV_LOCAL_RANK_OMPI "OMPI_COMM_WORLD_LOCAL_RANK"
#define ENV_RANK_OMPI "OMPI_COMM_WORLD_RANK"
#define ENV_LOCAL_RANK_SLURM "SLURM_LOCALID"
@ -228,8 +230,17 @@ void acceleratorInit(void)
{
rank = atoi(localRankStr);
}
if ((localRankStr = getenv(ENV_LOCAL_RANK_PALS)) != NULL)
{
rank = atoi(localRankStr);
}
if ((localRankStr = getenv(ENV_RANK_OMPI )) != NULL) { world_rank = atoi(localRankStr);}
if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
if ((localRankStr = getenv(ENV_RANK_PALS )) != NULL) { world_rank = atoi(localRankStr);}
char hostname[HOST_NAME_MAX+1];
gethostname(hostname, HOST_NAME_MAX+1);
if ( rank==0 ) printf(" acceleratorInit world_rank %d is host %s \n",world_rank,hostname);
auto devices = cl::sycl::device::get_devices();
for(int d = 0;d<devices.size();d++){
@ -241,9 +252,10 @@ void acceleratorInit(void)
printf("AcceleratorSyclInit: " #prop ": " FMT" \n",devices[d].get_info<cl::sycl::info::device::prop>());
#define GPU_PROP(prop) GPU_PROP_FMT(prop,"%ld");
if ( world_rank == 0) {
GPU_PROP_STR(vendor);
GPU_PROP_STR(version);
GPU_PROP_STR(vendor);
GPU_PROP_STR(version);
// GPU_PROP_STR(device_type);
/*
GPU_PROP(max_compute_units);
@ -259,7 +271,8 @@ void acceleratorInit(void)
GPU_PROP(single_fp_config);
*/
// GPU_PROP(double_fp_config);
GPU_PROP(global_mem_size);
GPU_PROP(global_mem_size);
}
}
if ( world_rank == 0 ) {

54
Grid/threads/Accelerator.h
View File

@ -268,6 +268,8 @@ inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
inline void acceleratorCopyToDeviceAsync(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyHostToDevice, stream);}
inline void acceleratorCopyFromDeviceAsync(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToHost, stream);}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
{
@ -297,17 +299,13 @@ inline int acceleratorIsCommunicable(void *ptr)
#define GRID_SYCL_LEVEL_ZERO_IPC
NAMESPACE_END(Grid);
#if 0
#include <CL/sycl.hpp>
#include <CL/sycl/usm.hpp>
#include <level_zero/ze_api.h>
#include <CL/sycl/backend/level_zero.hpp>
#else
// Force deterministic reductions
#define SYCL_REDUCTION_DETERMINISTIC
#include <sycl/CL/sycl.hpp>
#include <sycl/usm.hpp>
#include <level_zero/ze_api.h>
#include <sycl/ext/oneapi/backend/level_zero.hpp>
#endif
NAMESPACE_BEGIN(Grid);
@ -336,23 +334,24 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
#define accelerator_for2dNB( iter1, num1, iter2, num2, nsimd, ... ) \
theGridAccelerator->submit([&](cl::sycl::handler &cgh) { \
unsigned long nt=acceleratorThreads(); \
unsigned long unum1 = num1; \
unsigned long unum2 = num2; \
if(nt < 8)nt=8; \
cl::sycl::range<3> local {nt,1,nsimd}; \
cl::sycl::range<3> global{unum1,unum2,nsimd}; \
cgh.parallel_for( \
cl::sycl::nd_range<3>(global,local), \
[=] (cl::sycl::nd_item<3> item) /*mutable*/ \
[[intel::reqd_sub_group_size(16)]] \
{ \
auto iter1 = item.get_global_id(0); \
auto iter2 = item.get_global_id(1); \
auto lane = item.get_global_id(2); \
{ __VA_ARGS__ }; \
}); \
});
unsigned long nt=acceleratorThreads(); \
if(nt < 8)nt=8; \
unsigned long unum1 = num1; \
unsigned long unum2 = num2; \
unsigned long unum1_divisible_by_nt = ((unum1 + nt - 1) / nt) * nt; \
cl::sycl::range<3> local {nt,1,nsimd}; \
cl::sycl::range<3> global{unum1_divisible_by_nt,unum2,nsimd}; \
cgh.parallel_for( \
cl::sycl::nd_range<3>(global,local), \
[=] (cl::sycl::nd_item<3> item) /*mutable*/ \
[[intel::reqd_sub_group_size(16)]] \
{ \
auto iter1 = item.get_global_id(0); \
auto iter2 = item.get_global_id(1); \
auto lane = item.get_global_id(2); \
{ if (iter1 < unum1){ __VA_ARGS__ } }; \
}); \
});
#define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
@ -508,6 +507,12 @@ inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes
{
auto discard=hipMemcpyDtoDAsync(to,from,bytes, copyStream);
}
inline void acceleratorCopyToDeviceAsync(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
auto r = hipMemcpyAsync(to,from,bytes, hipMemcpyHostToDevice, stream);
}
inline void acceleratorCopyFromDeviceAsync(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
auto r = hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToHost, stream);
}
inline void acceleratorCopySynchronise(void) { auto discard=hipStreamSynchronize(copyStream); };
#endif
@ -664,4 +669,5 @@ template<class T> T acceleratorGet(T& dev)
NAMESPACE_END(Grid);