mirror of
https://github.com/paboyle/Grid.git
synced 2024-11-09 23:45:36 +00:00
Async memcpy's and cleanup
This commit is contained in:
Dennis Bollweg
parent
79a6ed32d8
commit
b8b9dc952d
@ -4,9 +4,8 @@
|
||||
#include <cub/cub.cuh>
|
||||
#define gpucub cub
|
||||
#define gpuMalloc cudaMalloc
|
||||
#define gpuMemcpy cudaMemcpy
|
||||
#define gpuMemcpyDeviceToHost cudaMemcpyDeviceToHost
|
||||
#define gpuMemcpyHostToDevice cudaMemcpyHostToDevice
|
||||
#define gpuMemcpyDtoHAsync cudaMemcpyDtoHAsync
|
||||
#define gpuMemcpyHtoDAsync cudaMemcpyHtoDAsync
|
||||
#define gpuError_t cudaError_t
|
||||
#define gpuSuccess cudaSuccess
|
||||
|
||||
@ -15,13 +14,11 @@
|
||||
#include <hipcub/hipcub.hpp>
|
||||
#define gpucub hipcub
|
||||
#define gpuMalloc hipMalloc
|
||||
#define gpuMemcpy hipMemcpy
|
||||
#define gpuMemcpyDeviceToHost hipMemcpyDeviceToHost
|
||||
#define gpuMemcpyHostToDevice hipMemcpyHostToDevice
|
||||
#define gpuMemcpyDtoHAsync hipMemcpyDtoHAsync
|
||||
#define gpuMemcpyHtoDAsync hipMemcpyHtoDAsync
|
||||
#define gpuError_t hipError_t
|
||||
#define gpuSuccess hipSuccess
|
||||
|
||||
// extern hipStream_t computeStream;
|
||||
#endif
|
||||
|
||||
|
||||
@ -49,78 +46,101 @@ template<class vobj> inline void sliceSumGpu(const Lattice<vobj> &Data,std::vect
|
||||
int e2= grid->_slice_block [orthogdim];
|
||||
int stride=grid->_slice_stride[orthogdim];
|
||||
int ostride=grid->_ostride[orthogdim];
|
||||
size_t subvol_size = e1*e2;
|
||||
|
||||
Vector<vobj> lvSum(rd);
|
||||
Vector<sobj> lsSum(ld,Zero());
|
||||
commVector<vobj> reduction_buffer(rd*e1*e2);
|
||||
ExtractBuffer<sobj> extracted(Nsimd);
|
||||
|
||||
result.resize(fd);
|
||||
|
||||
for(int r=0;r<rd;r++){
|
||||
lvSum[r]=Zero();
|
||||
}
|
||||
vobj identity;
|
||||
zeroit(identity);
|
||||
|
||||
vobj vobj_zero; //Need to provide initial value for reduction operation
|
||||
zeroit(vobj_zero);
|
||||
|
||||
autoView( Data_v, Data, AcceleratorRead);
|
||||
auto rb_p = &reduction_buffer[0];
|
||||
|
||||
auto rb_p = &reduction_buffer[0];
|
||||
void *helperArray = NULL;
|
||||
vobj *d_out;
|
||||
size_t temp_storage_bytes = 0;
|
||||
size_t size = e1*e2;
|
||||
std::vector<int> offsets(rd+1,0);
|
||||
for (int i = 0; i < offsets.size(); i++) {
|
||||
offsets[i] = i*size;
|
||||
}
|
||||
int* d_offsets;
|
||||
|
||||
gpuError_t gpuErr = gpuMalloc(&d_out,rd*sizeof(vobj));
|
||||
if (gpuErr != gpuSuccess) {
|
||||
std::cout << "Lattice_slicesum_gpu.h: Encountered error during gpuMalloc(1) Error: " << gpuErr <<std::endl;
|
||||
}
|
||||
gpuErr = gpuMalloc(&d_offsets,sizeof(int)*(rd+1));
|
||||
if (gpuErr != gpuSuccess) {
|
||||
std::cout << "Lattice_slicesum_gpu.h: Encountered error during gpuMalloc(2) Error: " << gpuErr <<std::endl;
|
||||
}
|
||||
gpuErr = gpuMemcpy(d_offsets,&offsets[0],sizeof(int)*(rd+1),gpuMemcpyHostToDevice);
|
||||
if (gpuErr != gpuSuccess) {
|
||||
std::cout << "Lattice_slicesum_gpu.h: Encountered error during gpuMemcpy(1) Error: " << gpuErr <<std::endl;
|
||||
}
|
||||
|
||||
gpuErr = gpucub::DeviceSegmentedReduce::Reduce(helperArray, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), identity, computeStream);
|
||||
if (gpuErr!=gpuSuccess) {
|
||||
std::cout << "Lattice_slicesum_gpu.h: Encountered error during cub::DeviceReduce::Sum(1)! Error: " << gpuErr <<std::endl;
|
||||
std::vector<int> offsets(rd+1,0);
|
||||
|
||||
for (int i = 0; i < offsets.size(); i++) {
|
||||
offsets[i] = i*subvol_size;
|
||||
}
|
||||
|
||||
//Allocate memory for output and offset arrays on device
|
||||
gpuError_t gpuErr = gpuMalloc(&d_out,rd*sizeof(vobj));
|
||||
if (gpuErr != gpuSuccess) {
|
||||
std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpuMalloc (d_out)! Error: " << gpuErr <<std::endl;
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
gpuErr = gpuMalloc(&d_offsets,sizeof(int)*(rd+1));
|
||||
if (gpuErr != gpuSuccess) {
|
||||
std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpuMalloc (d_offsets)! Error: " << gpuErr <<std::endl;
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
//copy offsets to device
|
||||
gpuErr = gpuMemcpyHtoDAsync(d_offsets,&offsets[0],sizeof(int)*(rd+1),computeStream);
|
||||
if (gpuErr != gpuSuccess) {
|
||||
std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpuMemcpy (d_offsets)! Error: " << gpuErr <<std::endl;
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
//determine helperArray size
|
||||
gpuErr = gpucub::DeviceSegmentedReduce::Reduce(helperArray, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), vobj_zero, computeStream);
|
||||
if (gpuErr!=gpuSuccess) {
|
||||
std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce (setup)! Error: " << gpuErr <<std::endl;
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
//allocate memory for helperArray
|
||||
gpuErr = gpuMalloc(&helperArray,temp_storage_bytes);
|
||||
if (gpuErr!=gpuSuccess) {
|
||||
std::cout << "Lattice_slicesum_gpu.h: Encountered error during gpuMalloc Error: " << gpuErr <<std::endl;
|
||||
std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpuMalloc (helperArray)! Error: " << gpuErr <<std::endl;
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
//prepare buffer for reduction
|
||||
accelerator_for2dNB( s,e1*e2, r,rd, grid->Nsimd(),{ //use non-blocking accelerator_for to avoid syncs (ok because we submit to same computeStream)
|
||||
//use 2d accelerator_for to avoid launch latencies found when looping over rd
|
||||
//use non-blocking accelerator_for to avoid syncs (ok because we submit to same computeStream)
|
||||
//use 2d accelerator_for to avoid launch latencies found when serially looping over rd
|
||||
|
||||
accelerator_for2dNB( s,subvol_size, r,rd, grid->Nsimd(),{
|
||||
|
||||
int n = s / e2;
|
||||
int b = s % e2;
|
||||
int so=r*ostride; // base offset for start of plane
|
||||
int ss= so+n*stride+b;
|
||||
|
||||
coalescedWrite(rb_p[r*e1*e2+s], coalescedRead(Data_v[ss]));
|
||||
coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data_v[ss]));
|
||||
|
||||
});
|
||||
|
||||
//issue reductions in computeStream
|
||||
gpuErr =gpucub::DeviceSegmentedReduce::Reduce(helperArray, temp_storage_bytes, rb_p, d_out, rd, d_offsets, d_offsets+1,::gpucub::Sum(), identity, computeStream);
|
||||
//issue segmented reductions in computeStream
|
||||
gpuErr = gpucub::DeviceSegmentedReduce::Reduce(helperArray, temp_storage_bytes, rb_p, d_out, rd, d_offsets, d_offsets+1,::gpucub::Sum(), vobj_zero, computeStream);
|
||||
if (gpuErr!=gpuSuccess) {
|
||||
std::cout << "Lattice_slicesum_gpu.h: Encountered error during cub::DeviceReduce::Sum(2)! Error: " << gpuErr <<std::endl;
|
||||
std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce! Error: " << gpuErr <<std::endl;
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
//sync before copy
|
||||
accelerator_barrier();
|
||||
gpuErr = gpuMemcpy(&lvSum[0],d_out,rd*sizeof(vobj),gpuMemcpyDeviceToHost);
|
||||
gpuErr = gpuMemcpyDtoHAsync(&lvSum[0],d_out,rd*sizeof(vobj),computeStream);
|
||||
if (gpuErr!=gpuSuccess) {
|
||||
std::cout << "Lattice_slicesum_gpu.h: Encountered error during gpuMemcpy(2) Error: " << gpuErr <<std::endl;
|
||||
std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpuMemcpy (d_out)! Error: " << gpuErr <<std::endl;
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
//sync after copy
|
||||
accelerator_barrier();
|
||||
|
||||
// Sum across simd lanes in the plane, breaking out orthog dir.
|
||||
Coordinate icoor(Nd);
|
||||
|
||||
|
||||
@ -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;
|
||||
}
|
||||
Loading…
Reference in New Issue
Block a user