portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-18 07:47:06 +01:00
Code Releases Activity

Merge pull request #17 from paboyle/develop

Browse Source 
sync upstream
This commit is contained in:
Christoph Lehner
2020-10-06 09:01:39 +02:00
committed by GitHub
parent 5534921bee ace9cd64bb
commit 80fd6ab407
20 changed files with 170 additions and 132 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
Grid/lattice/Lattice_arith.h
View File

@ -60,9 +60,9 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
autoView( lhs_v , lhs, AcceleratorRead);
autoView( rhs_v , rhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
auto rhs_t=rhs_v(ss);
auto tmp =ret_v(ss);
mac(&tmp,&lhs_t,&rhs_t);
coalescedWrite(ret_v[ss],tmp);
});
@ -124,7 +124,7 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto tmp =ret_v(ss);
auto lhs_t=lhs_v(ss);
mac(&tmp,&lhs_t,&rhs);
coalescedWrite(ret_v[ss],tmp);
@ -182,7 +182,7 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
autoView( ret_v , ret, AcceleratorWrite);
autoView( rhs_v , lhs, AcceleratorRead);
accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto tmp =ret_v(ss);
auto rhs_t=rhs_v(ss);
mac(&tmp,&lhs,&rhs_t);
coalescedWrite(ret_v[ss],tmp);

11
Grid/lattice/Lattice_reduction_gpu.h
View File

@ -2,12 +2,13 @@ NAMESPACE_BEGIN(Grid);
#ifdef GRID_HIP
extern hipDeviceProp_t *gpu_props;
#define WARP_SIZE 64
#endif
#ifdef GRID_CUDA
extern cudaDeviceProp *gpu_props;
#define WARP_SIZE 32
#endif
#define WARP_SIZE 32
__device__ unsigned int retirementCount = 0;
template <class Iterator>
@ -64,7 +65,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
// cannot use overloaded operators for sobj as they are not volatile-qualified
memcpy((void *)&sdata[tid], (void *)&mySum, sizeof(sobj));
__syncwarp();
acceleratorSynchronise();
const Iterator VEC = WARP_SIZE;
const Iterator vid = tid & (VEC-1);
@ -78,9 +79,9 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
beta += temp;
memcpy((void *)&sdata[tid], (void *)&beta, sizeof(sobj));
}
__syncwarp();
acceleratorSynchronise();
}
__syncthreads();
acceleratorSynchroniseAll();
if (threadIdx.x == 0) {
beta = Zero();
@ -90,7 +91,7 @@ __device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid
}
memcpy((void *)&sdata[0], (void *)&beta, sizeof(sobj));
}
__syncthreads();
acceleratorSynchroniseAll();
}