portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-09-20 01:05:38 +01:00
Code Releases Activity

Merge pull request #384 from jdmaia/hip_launchbounds

Browse Source 
Changing thread block order and adding launch_bounds
This commit is contained in:
Peter Boyle 2022-02-14 11:08:28 -05:00 committed by GitHub
commit 48772f0976
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 31 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
Grid/threads/Accelerator.h
View File

@ -342,7 +342,7 @@ extern hipStream_t copyStream;
/*These routines define mapping from thread grid to loop & vector lane indexing */
accelerator_inline int acceleratorSIMTlane(int Nsimd) {
#ifdef GRID_SIMT
return hipThreadIdx_z;
return hipThreadIdx_x;
#else
return 0;
#endif
@ -356,19 +356,41 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
{ __VA_ARGS__;} \
}; \
int nt=acceleratorThreads(); \
dim3 hip_threads(nt,1,nsimd); \
dim3 hip_threads(nsimd, nt, 1); \
dim3 hip_blocks ((num1+nt-1)/nt,num2,1); \
if(hip_threads.x * hip_threads.y * hip_threads.z <= 64){ \
hipLaunchKernelGGL(LambdaApply64,hip_blocks,hip_threads, \
0,0, \
num1,num2,nsimd, lambda); \
} else { \
hipLaunchKernelGGL(LambdaApply,hip_blocks,hip_threads, \
0,0, \
num1,num2,nsimd, lambda); \
} \
}
template<typename lambda> __global__
__launch_bounds__(64,1)
void LambdaApply64(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
{
// Following the same scheme as CUDA for now
uint64_t x = threadIdx.y + blockDim.y*blockIdx.x;
uint64_t y = threadIdx.z + blockDim.z*blockIdx.y;
uint64_t z = threadIdx.x;
if ( (x < numx) && (y<numy) && (z<numz) ) {
Lambda(x,y,z);
}
}
template<typename lambda> __global__
__launch_bounds__(1024,1)
void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
{
uint64_t x = hipThreadIdx_x + hipBlockDim_x*hipBlockIdx_x;
uint64_t y = hipThreadIdx_y + hipBlockDim_y*hipBlockIdx_y;
uint64_t z = hipThreadIdx_z ;//+ hipBlockDim_z*hipBlockIdx_z;
// Following the same scheme as CUDA for now
uint64_t x = threadIdx.y + blockDim.y*blockIdx.x;
uint64_t y = threadIdx.z + blockDim.z*blockIdx.y;
uint64_t z = threadIdx.x;
if ( (x < numx) && (y<numy) && (z<numz) ) {
Lambda(x,y,z);
}