Skip to content

Commit

Permalink
migrate hits to SoATemplate
Browse files Browse the repository at this point in the history
  • Loading branch information
@ariostas
ariostas committed Oct 24, 2024
1 parent 534ac1b commit 018204d
Show file tree
Hide file tree
Showing 11 changed files with 298 additions and 310 deletions.
43 changes: 43 additions & 0 deletions RecoTracker/LSTCore/interface/HitsSoA.h
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
#ifndef RecoTracker_LSTCore_interface_HitsSoA_h
#define RecoTracker_LSTCore_interface_HitsSoA_h

#include "DataFormats/SoATemplate/interface/SoALayout.h"
#include "DataFormats/Portable/interface/PortableCollection.h"

namespace lst {

GENERATE_SOA_LAYOUT(HitsSoALayout,
SOA_COLUMN(float, xs),
SOA_COLUMN(float, ys),
SOA_COLUMN(float, zs),
SOA_COLUMN(uint16_t, moduleIndices),
SOA_COLUMN(unsigned int, idxs),
SOA_COLUMN(unsigned int, detid),
SOA_COLUMN(float, rts),
SOA_COLUMN(float, phis),
SOA_COLUMN(float, etas),
SOA_COLUMN(float, highEdgeXs),
SOA_COLUMN(float, highEdgeYs),
SOA_COLUMN(float, lowEdgeXs),
SOA_COLUMN(float, lowEdgeYs))

GENERATE_SOA_LAYOUT(HitsOccupancySoALayout,
SOA_COLUMN(ArrayIx2, hitRanges),
SOA_COLUMN(int, hitRangesLower),
SOA_COLUMN(int, hitRangesUpper),
SOA_COLUMN(int8_t, hitRangesnLower),
SOA_COLUMN(int8_t, hitRangesnUpper))

using HitsSoA = HitsSoALayout<>;
using HitsOccupancySoA = HitsOccupancySoALayout<>;

using Hits = HitsSoA::View;
using HitsConst = HitsSoA::ConstView;
using HitsOccupancy = HitsOccupancySoA::View;
using HitsOccupancyConst = HitsOccupancySoA::ConstView;

using HitsHostCollection = PortableHostMultiCollection<HitsSoA, HitsOccupancySoA>;

} // namespace lst

#endif
12 changes: 12 additions & 0 deletions RecoTracker/LSTCore/interface/alpaka/HitsDeviceCollection.h
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,12 @@
#ifndef RecoTracker_LSTCore_interface_alpaka_HitsDeviceCollection_h
#define RecoTracker_LSTCore_interface_alpaka_HitsDeviceCollection_h

#include "DataFormats/Portable/interface/alpaka/PortableCollection.h"

#include "RecoTracker/LSTCore/interface/HitsSoA.h"

namespace ALPAKA_ACCELERATOR_NAMESPACE::lst {
using HitsDeviceCollection = PortableCollection2<HitsSoA, HitsOccupancySoA>;
} // namespace ALPAKA_ACCELERATOR_NAMESPACE::lst

#endif
145 changes: 84 additions & 61 deletions RecoTracker/LSTCore/src/alpaka/Event.dev.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -51,8 +51,7 @@ void Event::resetEventSync() {
n_quintuplets_by_layer_endcap_[i] = 0;
}
}
hitsInGPU_.reset();
hitsBuffers_.reset();
hitsDC_.reset();
miniDoubletsDC_.reset();
rangesDC_.reset();
segmentsDC_.reset();
Expand All @@ -66,7 +65,7 @@ void Event::resetEventSync() {
pixelQuintupletsInGPU_.reset();
pixelQuintupletsBuffers_.reset();

hitsInCPU_.reset();
hitsHC_.reset();
rangesHC_.reset();
miniDoubletsHC_.reset();
segmentsHC_.reset();
Expand All @@ -87,10 +86,25 @@ void Event::addHitToEvent(std::vector<float> const& x,
unsigned int nHits = x.size();

// Initialize space on device/host for next event.
if (!hitsInGPU_) {
hitsInGPU_.emplace();
hitsBuffers_.emplace(nModules_, nHits, devAcc_, queue_);
hitsInGPU_->setData(*hitsBuffers_);
if (!hitsDC_) {
std::array<int, 2> const hits_sizes{{static_cast<int>(nHits), static_cast<int>(nModules_)}};
hitsDC_.emplace(hits_sizes, queue_);

auto hitsOccupancy = hitsDC_->view<HitsOccupancySoA>();
auto hitRanges_view = alpaka::createView(devAcc_, hitsOccupancy.hitRanges(), hitsOccupancy.metadata().size());
auto hitRangesLower_view =
alpaka::createView(devAcc_, hitsOccupancy.hitRangesLower(), hitsOccupancy.metadata().size());
auto hitRangesUpper_view =
alpaka::createView(devAcc_, hitsOccupancy.hitRangesUpper(), hitsOccupancy.metadata().size());
auto hitRangesnLower_view =
alpaka::createView(devAcc_, hitsOccupancy.hitRangesnLower(), hitsOccupancy.metadata().size());
auto hitRangesnUpper_view =
alpaka::createView(devAcc_, hitsOccupancy.hitRangesnUpper(), hitsOccupancy.metadata().size());
alpaka::memset(queue_, hitRanges_view, 0xff);
alpaka::memset(queue_, hitRangesLower_view, 0xff);
alpaka::memset(queue_, hitRangesUpper_view, 0xff);
alpaka::memset(queue_, hitRangesnLower_view, 0xff);
alpaka::memset(queue_, hitRangesnUpper_view, 0xff);
}

if (!rangesDC_) {
Expand All @@ -100,16 +114,18 @@ void Event::addHitToEvent(std::vector<float> const& x,
alpaka::memset(queue_, buf, 0xff);
}

// Need a view here before transferring to the device.
auto nHits_view = alpaka::createView(cms::alpakatools::host(), &nHits, (Idx)1u);

// Copy the host arrays to the GPU.
alpaka::memcpy(queue_, hitsBuffers_->xs_buf, x, nHits);
alpaka::memcpy(queue_, hitsBuffers_->ys_buf, y, nHits);
alpaka::memcpy(queue_, hitsBuffers_->zs_buf, z, nHits);
alpaka::memcpy(queue_, hitsBuffers_->detid_buf, detId, nHits);
alpaka::memcpy(queue_, hitsBuffers_->idxs_buf, idxInNtuple, nHits);
alpaka::memcpy(queue_, hitsBuffers_->nHits_buf, nHits_view);
auto hits = hitsDC_->view<HitsSoA>();
auto xs_d = alpaka::createView(devAcc_, hits.xs(), (Idx)hits.metadata().size());
auto ys_d = alpaka::createView(devAcc_, hits.ys(), (Idx)hits.metadata().size());
auto zs_d = alpaka::createView(devAcc_, hits.zs(), (Idx)hits.metadata().size());
auto detId_d = alpaka::createView(devAcc_, hits.detid(), (Idx)hits.metadata().size());
auto idxs_d = alpaka::createView(devAcc_, hits.idxs(), (Idx)hits.metadata().size());
alpaka::memcpy(queue_, xs_d, x, (Idx)nHits);
alpaka::memcpy(queue_, ys_d, y, (Idx)nHits);
alpaka::memcpy(queue_, zs_d, z, (Idx)nHits);
alpaka::memcpy(queue_, detId_d, detId, (Idx)nHits);
alpaka::memcpy(queue_, idxs_d, idxInNtuple, (Idx)nHits);
alpaka::wait(queue_); // FIXME: remove synch after inputs refactored to be in pinned memory

Vec3D const threadsPerBlock1{1, 1, 256};
Expand All @@ -126,15 +142,20 @@ void Event::addHitToEvent(std::vector<float> const& x,
endcapGeometryBuffers_.geoMapDetId_buf.data(),
endcapGeometryBuffers_.geoMapPhi_buf.data(),
*modulesBuffers_.data(),
*hitsInGPU_,
hitsDC_->view<HitsSoA>(),
hitsDC_->view<HitsOccupancySoA>(),
nHits);

Vec3D const threadsPerBlock2{1, 1, 256};
Vec3D const blocksPerGrid2{1, 1, max_blocks};
WorkDiv3D const module_ranges_workdiv = createWorkDiv(blocksPerGrid2, threadsPerBlock2, elementsPerThread);

alpaka::exec<Acc3D>(
queue_, module_ranges_workdiv, ModuleRangesKernel{}, *modulesBuffers_.data(), *hitsInGPU_, nLowerModules_);
alpaka::exec<Acc3D>(queue_,
module_ranges_workdiv,
ModuleRangesKernel{},
*modulesBuffers_.data(),
hitsDC_->view<HitsOccupancySoA>(),
nLowerModules_);
}

void Event::addPixelSegmentToEvent(std::vector<unsigned int> const& hitIndices0,
Expand Down Expand Up @@ -307,7 +328,7 @@ void Event::addPixelSegmentToEvent(std::vector<unsigned int> const& hitIndices0,
AddPixelSegmentToEventKernel{},
*modulesBuffers_.data(),
rangesDC_->const_view<ObjectOccupancySoA>(),
*hitsInGPU_,
hitsDC_->view<HitsSoA>(),
miniDoubletsDC_->view<MiniDoubletsSoA>(),
segmentsDC_->view<SegmentsSoA>(),
segmentsDC_->view<SegmentsPixelSoA>(),
Expand Down Expand Up @@ -371,7 +392,8 @@ void Event::createMiniDoublets() {
createMiniDoubletsInGPUv2_workDiv,
CreateMiniDoubletsInGPUv2{},
*modulesBuffers_.data(),
*hitsInGPU_,
hitsDC_->const_view<HitsSoA>(),
hitsDC_->const_view<HitsOccupancySoA>(),
miniDoubletsDC_->view<MiniDoubletsSoA>(),
miniDoubletsDC_->view<MiniDoubletsOccupancySoA>(),
rangesDC_->const_view<ObjectOccupancySoA>());
Expand All @@ -385,7 +407,7 @@ void Event::createMiniDoublets() {
miniDoubletsDC_->view<MiniDoubletsOccupancySoA>(),
rangesDC_->view<ObjectRangesSoA>(),
rangesDC_->const_view<ObjectOccupancySoA>(),
*hitsInGPU_);
hitsDC_->const_view<HitsOccupancySoA>());

if (addObjects_) {
addMiniDoubletsToEventExplicit();
Expand Down Expand Up @@ -643,7 +665,7 @@ void Event::createTrackCandidates(bool no_pls_dupclean, bool tc_pls_triplets) {
segmentsDC_->const_view<SegmentsOccupancySoA>(),
segmentsDC_->view<SegmentsPixelSoA>(),
miniDoubletsDC_->const_view<MiniDoubletsSoA>(),
*hitsInGPU_,
hitsDC_->const_view<HitsSoA>(),
*quintupletsInGPU_);

Vec3D const threadsPerBlock_addpLSasTrackCandidateInGPU{1, 1, 384};
Expand Down Expand Up @@ -1055,8 +1077,10 @@ void Event::addMiniDoubletsToEventExplicit() {
auto module_layers_buf = allocBufWrapper<short>(cms::alpakatools::host(), nLowerModules_, queue_);
alpaka::memcpy(queue_, module_layers_buf, modulesBuffers_.layers_buf, nLowerModules_);

auto module_hitRanges_buf = allocBufWrapper<int>(cms::alpakatools::host(), nLowerModules_ * 2, queue_);
alpaka::memcpy(queue_, module_hitRanges_buf, hitsBuffers_->hitRanges_buf, nLowerModules_ * 2u);
auto module_hitRanges_buf = allocBufWrapper<ArrayIx2>(cms::alpakatools::host(), nLowerModules_, queue_);
auto hits = hitsDC_->view<HitsOccupancySoA>();
auto hitRanges_view = alpaka::createView(devAcc_, hits.hitRanges(), nLowerModules_); // only lower modules
alpaka::memcpy(queue_, module_hitRanges_buf, hitRanges_view, nLowerModules_);

alpaka::wait(queue_); // wait for inputs before using them

Expand All @@ -1066,7 +1090,7 @@ void Event::addMiniDoubletsToEventExplicit() {
auto const* module_hitRanges = module_hitRanges_buf.data();

for (unsigned int i = 0; i < nLowerModules_; i++) {
if (!(nMDsCPU[i] == 0 or module_hitRanges[i * 2] == -1)) {
if (!(nMDsCPU[i] == 0 or module_hitRanges[i][0] == -1)) {
if (module_subdets[i] == Barrel) {
n_minidoublets_by_layer_barrel_[module_layers[i] - 1] += nMDsCPU[i];
} else {
Expand Down Expand Up @@ -1376,47 +1400,46 @@ int Event::getNumberOfT5TrackCandidates() {
return *nTrackCandidatesT5_buf_h.data();
}

HitsBuffer<alpaka_common::DevHost>& Event::getHits(bool sync) //std::shared_ptr should take care of garbage collection
template <typename TSoA, typename TDev>
typename TSoA::ConstView Event::getHits(bool sync) //std::shared_ptr should take care of garbage collection
{
if (!hitsInCPU_) {
auto nHits_buf_h = cms::alpakatools::make_host_buffer<unsigned int[]>(queue_, 1u);
alpaka::memcpy(queue_, nHits_buf_h, hitsBuffers_->nHits_buf);
alpaka::wait(queue_); // wait for the value before using

auto const nHits = *nHits_buf_h.data();
hitsInCPU_.emplace(nModules_, nHits, cms::alpakatools::host(), queue_);
hitsInCPU_->setData(*hitsInCPU_);

alpaka::memcpy(queue_, hitsInCPU_->nHits_buf, hitsBuffers_->nHits_buf);
alpaka::memcpy(queue_, hitsInCPU_->idxs_buf, hitsBuffers_->idxs_buf, nHits);
alpaka::memcpy(queue_, hitsInCPU_->detid_buf, hitsBuffers_->detid_buf, nHits);
alpaka::memcpy(queue_, hitsInCPU_->xs_buf, hitsBuffers_->xs_buf, nHits);
alpaka::memcpy(queue_, hitsInCPU_->ys_buf, hitsBuffers_->ys_buf, nHits);
alpaka::memcpy(queue_, hitsInCPU_->zs_buf, hitsBuffers_->zs_buf, nHits);
alpaka::memcpy(queue_, hitsInCPU_->moduleIndices_buf, hitsBuffers_->moduleIndices_buf, nHits);
if (sync)
alpaka::wait(queue_); // host consumers expect filled data
if constexpr (std::is_same_v<TDev, DevHost>) {
return hitsDC_->const_view<TSoA>();
} else {
if (!hitsHC_) {
hitsHC_.emplace(cms::alpakatools::CopyToHost<PortableMultiCollection<TDev, HitsSoA, HitsOccupancySoA>>::copyAsync(
queue_, *hitsDC_));
if (sync)
alpaka::wait(queue_); // host consumers expect filled data
}
return hitsHC_->const_view<TSoA>();
}
return hitsInCPU_.value();
}
template HitsConst Event::getHits<HitsSoA>(bool);
template HitsOccupancyConst Event::getHits<HitsOccupancySoA>(bool);

HitsBuffer<alpaka_common::DevHost>& Event::getHitsInCMSSW(bool sync) {
if (!hitsInCPU_) {
auto nHits_buf_h = cms::alpakatools::make_host_buffer<unsigned int[]>(queue_, 1u);
alpaka::memcpy(queue_, nHits_buf_h, hitsBuffers_->nHits_buf);
alpaka::wait(queue_); // wait for the value before using

auto const nHits = *nHits_buf_h.data();
hitsInCPU_.emplace(nModules_, nHits, cms::alpakatools::host(), queue_);
hitsInCPU_->setData(*hitsInCPU_);

alpaka::memcpy(queue_, hitsInCPU_->nHits_buf, hitsBuffers_->nHits_buf);
alpaka::memcpy(queue_, hitsInCPU_->idxs_buf, hitsBuffers_->idxs_buf, nHits);
if (sync)
alpaka::wait(queue_); // host consumers expect filled data
template <typename TSoA, typename TDev>
typename TSoA::ConstView Event::getHitsInCMSSW(bool sync) {
if constexpr (std::is_same_v<TDev, DevHost>) {
return hitsDC_->const_view<TSoA>();
} else {
if (!hitsHC_) {
auto hits_d = hitsDC_->view<HitsSoA>();
auto nHits = hits_d.metadata().size();
std::array<int, 2> const hits_sizes{{static_cast<int>(nHits), static_cast<int>(nModules_)}};
hitsHC_.emplace(hits_sizes, queue_);
auto hits_h = hitsHC_->view<HitsSoA>();
auto idxs_h = alpaka::createView(cms::alpakatools::host(), hits_h.idxs(), nHits);
auto idxs_d = alpaka::createView(devAcc_, hits_d.idxs(), nHits);
alpaka::memcpy(queue_, idxs_h, idxs_d);
if (sync)
alpaka::wait(queue_); // host consumers expect filled data
}
return hitsHC_->const_view<TSoA>();
}
return hitsInCPU_.value();
}
template HitsConst Event::getHitsInCMSSW<HitsSoA>(bool);
template HitsOccupancyConst Event::getHitsInCMSSW<HitsOccupancySoA>(bool);

template <typename TSoA, typename TDev>
typename TSoA::ConstView Event::getRanges(bool sync) {
Expand Down
11 changes: 6 additions & 5 deletions RecoTracker/LSTCore/src/alpaka/Event.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -45,8 +45,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::lst {

//Device stuff
std::optional<ObjectRangesDeviceCollection> rangesDC_;
std::optional<Hits> hitsInGPU_;
std::optional<HitsBuffer<Device>> hitsBuffers_;
std::optional<HitsDeviceCollection> hitsDC_;
std::optional<MiniDoubletsDeviceCollection> miniDoubletsDC_;
std::optional<SegmentsDeviceCollection> segmentsDC_;
std::optional<Triplets> tripletsInGPU_;
Expand All @@ -61,7 +60,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::lst {

//CPU interface stuff
std::optional<ObjectRangesHostCollection> rangesHC_;
std::optional<HitsBuffer<DevHost>> hitsInCPU_;
std::optional<HitsHostCollection> hitsHC_;
std::optional<MiniDoubletsHostCollection> miniDoubletsHC_;
std::optional<SegmentsHostCollection> segmentsHC_;
std::optional<TripletsBuffer<DevHost>> tripletsInCPU_;
Expand Down Expand Up @@ -178,8 +177,10 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::lst {
// (has no effect on repeated calls)
// set to false may allow faster operation with concurrent calls of get*
// HANDLE WITH CARE
HitsBuffer<DevHost>& getHits(bool sync = true);
HitsBuffer<DevHost>& getHitsInCMSSW(bool sync = true);
template <typename TSoA, typename TDev = Device>
typename TSoA::ConstView getHits(bool sync = true);
template <typename TSoA, typename TDev = Device>
typename TSoA::ConstView getHitsInCMSSW(bool sync = true);
template <typename TSoA, typename TDev = Device>
typename TSoA::ConstView getRanges(bool sync = true);
template <typename TSoA, typename TDev = Device>
Expand Down
Loading

0 comments on commit 018204d

Please sign in to comment.