portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-06-18 18:03:44 +01:00
Code Releases Activity

sumD_gpu_direct: one thread per SIMD lane using extractLane

Browse Source 
Replaces one thread per outer site calling Reduce() (sequential Nsimd-wide
loop) with one thread per lane calling extractLane() — O(1) per thread.
CUB now reduces over osites*Nsimd elements. Avoids serial lane reduction
but leaves the per-lane sobjD store stride as a known remaining concern.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Peter Boyle
2026-05-18 16:21:50 -04:00
parent fca2c5dba0
commit 747c167658
1 changed files with 14 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Grid/lattice/Lattice_reduction_gpu_cub.h
+14 -9
View File
@@ -63,16 +63,21 @@ inline typename vobj::scalar_objectD sumD_gpu_direct(const vobj *lat, Integer os
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_objectD sobjD;
deviceVector<sobjD> per_site(osites);
sobjD *per_site_p = &per_site[0];
const Integer nsimd = vobj::Nsimd();
const Integer nlanes = osites * nsimd;
deviceVector<sobjD> per_lane(nlanes);
sobjD *per_lane_p = &per_lane[0];
#ifdef GRID_REDUCTION_TIMING
RealD t_for = -usecond();
#endif
accelerator_for(ss, osites, 1, {
sobj tmp = Reduce(lat[ss]);
accelerator_for(idx, nlanes, 1, {
Integer ss = idx / nsimd;
Integer lane = idx % nsimd;
sobj tmp = extractLane(lane, lat[ss]);
sobjD tmpD; tmpD = tmp;
per_site_p[ss] = tmpD;
per_lane_p[idx] = tmpD;
});
#ifdef GRID_REDUCTION_TIMING
accelerator_barrier();
@@ -85,8 +90,8 @@ inline typename vobj::scalar_objectD sumD_gpu_direct(const vobj *lat, Integer os
size_t temp_bytes = 0;
gpuError_t gpuErr;
gpuErr = gpucub::DeviceReduce::Reduce(d_temp, temp_bytes, per_site_p, d_out,
(int)osites, gpucub::Sum(), zero, computeStream);
gpuErr = gpucub::DeviceReduce::Reduce(d_temp, temp_bytes, per_lane_p, d_out,
(int)nlanes, gpucub::Sum(), zero, computeStream);
if (gpuErr != gpuSuccess) {
std::cout << GridLogError << "sumD_gpu_direct: DeviceReduce size query failed: "
<< gpuErr << std::endl;
@@ -98,8 +103,8 @@ inline typename vobj::scalar_objectD sumD_gpu_direct(const vobj *lat, Integer os
#ifdef GRID_REDUCTION_TIMING
RealD t_cub = -usecond();
#endif
gpuErr = gpucub::DeviceReduce::Reduce(d_temp, temp_bytes, per_site_p, d_out,
(int)osites, gpucub::Sum(), zero, computeStream);
gpuErr = gpucub::DeviceReduce::Reduce(d_temp, temp_bytes, per_lane_p, d_out,
(int)nlanes, gpucub::Sum(), zero, computeStream);
if (gpuErr != gpuSuccess) {
std::cout << GridLogError << "sumD_gpu_direct: DeviceReduce failed: "
<< gpuErr << std::endl;