1
0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 04:37:05 +01:00

Async memcpy's and cleanup

This commit is contained in:
Dennis Bollweg
2024-02-01 17:55:35 -05:00
parent 79a6ed32d8
commit b8b9dc952d
2 changed files with 92 additions and 59 deletions

View File

@ -29,29 +29,42 @@ int main (int argc, char ** argv) {
sliceSumGpu(test_data,reduction_result,0);
}
int trace_id = traceStart("sliceSum benchmark");
for (int i = 0; i < Nd; i++) {
RealD t=-usecond();
sliceSum(test_data,reduction_reference,i);
t+=usecond();
std::cout << " sliceSum took "<<t<<" usecs"<<std::endl;
RealD tgpu=-usecond();
tracePush("sliceSumGpu");
sliceSumGpu(test_data,reduction_result,i);
tracePop("sliceSumGpu");
tgpu+=usecond();
std::cout <<" sliceSumGpu took "<<tgpu<<" usecs"<<std::endl;
for(int t=0;t<reduction_reference.size();t++){
auto diff = reduction_reference[t]-reduction_result[t];
// std::cout << "Difference = " << diff <<std::endl;
RealD t=-usecond();
assert(abs(TensorRemove(diff)) < 1e-8 );
}
tracePush("sliceSum");
sliceSum(test_data,reduction_reference,i);
tracePop("sliceSum");
t+=usecond();
std::cout << GridLogMessage << " sliceSum took "<<t<<" usecs"<<std::endl;
RealD tgpu=-usecond();
tracePush("sliceSumGpu");
sliceSumGpu(test_data,reduction_result,i);
tracePop("sliceSumGpu");
tgpu+=usecond();
std::cout << GridLogMessage <<" sliceSumGpu took "<<tgpu<<" usecs"<<std::endl;
for(int t=0;t<reduction_reference.size();t++) {
auto diff = reduction_reference[t]-reduction_result[t];
assert(abs(TensorRemove(diff)) < 1e-8 );
}
}
traceStop(trace_id);
Grid_finalize();
return 0;
}