Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Correct the FindClus block size [14.0.x] #44082

Merged
merged 1 commit into from
Feb 28, 2024
Merged

Correct the FindClus block size [14.0.x] #44082

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 11 additions & 9 deletions RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/PixelClustering.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -130,6 +130,13 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::pixelClustering {

template <typename TrackerTraits>
struct FindClus {
// assume that we can cover the whole module with up to 16 blockDimension-wide iterations
static constexpr uint32_t maxIterGPU = 16;

// this must be larger than maxPixInModule / maxIterGPU, and should be a multiple of the warp size
static constexpr uint32_t maxElementsPerBlock =
cms::alpakatools::round_up_by(TrackerTraits::maxPixInModule / maxIterGPU, 128);

template <typename TAcc>
ALPAKA_FN_ACC void operator()(const TAcc& acc,
SiPixelDigisSoAView digi_view,
Expand Down Expand Up @@ -292,17 +299,12 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::pixelClustering {
#endif

[[maybe_unused]] const uint32_t blockDimension = alpaka::getWorkDiv<alpaka::Block, alpaka::Elems>(acc)[0u];
// Assume that we can cover the whole module with up to 16 blockDimension-wide iterations
// This maxIter value was tuned for GPU, with 256 or 512 threads per block.
// Hence, also works for CPU case, with 256 or 512 elements per thread.
// Real constrainst is maxIter = hist.size() / blockDimension,
// with blockDimension = threadPerBlock * elementsPerThread.
// Hence, maxIter can be tuned accordingly to the workdiv.
constexpr unsigned int maxIterGPU = 16;
// assume that we can cover the whole module with up to maxIterGPU blockDimension-wide iterations
ALPAKA_ASSERT_ACC((hist.size() / blockDimension) < maxIterGPU);

// NB: can be tuned.
constexpr uint32_t maxElements = cms::alpakatools::requires_single_thread_per_block_v<TAcc> ? 256 : 1;
// number of elements per thread
constexpr uint32_t maxElements =
cms::alpakatools::requires_single_thread_per_block_v<TAcc> ? maxElementsPerBlock : 1;
ALPAKA_ASSERT_ACC((alpaka::getWorkDiv<alpaka::Thread, alpaka::Elems>(acc)[0u] <= maxElements));

constexpr unsigned int maxIter = maxIterGPU * maxElements;
Expand Down
26 changes: 7 additions & 19 deletions RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/SiPixelRawToClusterKernel.dev.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -641,25 +641,15 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
cms::alpakatools::make_device_view(alpaka::getDev(queue), clusters_d->view().moduleStart(), 1u);
alpaka::memcpy(queue, nModules_Clusters_h, moduleStartFirstElement);

// TODO
// - we are fixing this here since it is used at compile time also in the kernel
// - put maxIter in the Geometry traits
constexpr auto threadsOrElementsFindClus = 256;

const auto elementsPerBlockFindClus = FindClus<TrackerTraits>::maxElementsPerBlock;
const auto workDivMaxNumModules =
cms::alpakatools::make_workdiv<Acc1D>(numberOfModules, threadsOrElementsFindClus);
// NB: With present FindClus() / chargeCut() algorithm,
// threadPerBlock (GPU) or elementsPerThread (CPU) = 256 show optimal performance.
// Though, it does not have to be the same number for CPU/GPU cases.

cms::alpakatools::make_workdiv<Acc1D>(numberOfModules, elementsPerBlockFindClus);
#ifdef GPU_DEBUG
std::cout << " FindClus kernel launch with " << numberOfModules << " blocks of " << threadsOrElementsFindClus
std::cout << " FindClus kernel launch with " << numberOfModules << " blocks of " << elementsPerBlockFindClus
<< " threadsPerBlockOrElementsPerThread\n";
#endif

alpaka::exec<Acc1D>(
queue, workDivMaxNumModules, FindClus<TrackerTraits>{}, digis_d->view(), clusters_d->view(), wordCounter);

#ifdef GPU_DEBUG
alpaka::wait(queue);
#endif
Expand Down Expand Up @@ -740,14 +730,12 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
cms::alpakatools::make_device_view(alpaka::getDev(queue), clusters_d->view().moduleStart(), 1u);
alpaka::memcpy(queue, nModules_Clusters_h, moduleStartFirstElement);

/// should be larger than maxPixInModule/16 aka (maxPixInModule/maxiter in the kernel)

const auto threadsPerBlockFindClus = 256;
const auto workDivMaxNumModules = cms::alpakatools::make_workdiv<Acc1D>(numberOfModules, threadsPerBlockFindClus);

const auto elementsPerBlockFindClus = FindClus<TrackerTraits>::maxElementsPerBlock;
const auto workDivMaxNumModules =
cms::alpakatools::make_workdiv<Acc1D>(numberOfModules, elementsPerBlockFindClus);
#ifdef GPU_DEBUG
alpaka::wait(queue);
std::cout << "FindClus kernel launch with " << numberOfModules << " blocks of " << threadsPerBlockFindClus
std::cout << "FindClus kernel launch with " << numberOfModules << " blocks of " << elementsPerBlockFindClus
<< " threadsPerBlockOrElementsPerThread\n";
#endif
alpaka::exec<Acc1D>(
Expand Down