From 85c769bdb8977dadbe4df444becebfea3c09d11b Mon Sep 17 00:00:00 2001
From: Andrea Bocci <andrea.bocci@cern.ch>
Date: Mon, 26 Feb 2024 02:07:50 +0100
Subject: [PATCH] Correct the FindClus block size

Compute the FindClus block size from the maximum number of hits per module.

Harmonise the block size used for the kernel launch and inside the kernel itself.
---
 .../plugins/alpaka/PixelClustering.h          | 20 +++++++-------
 .../alpaka/SiPixelRawToClusterKernel.dev.cc   | 26 +++++--------------
 2 files changed, 18 insertions(+), 28 deletions(-)

diff --git a/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/PixelClustering.h b/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/PixelClustering.h
index 4a2d9e72e9366..1674b56c2e97a 100644
--- a/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/PixelClustering.h
+++ b/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/PixelClustering.h
@@ -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,
@@ -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;
diff --git a/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/SiPixelRawToClusterKernel.dev.cc b/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/SiPixelRawToClusterKernel.dev.cc
index 13b971753bd75..f54c9e9af29ec 100644
--- a/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/SiPixelRawToClusterKernel.dev.cc
+++ b/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/SiPixelRawToClusterKernel.dev.cc
@@ -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
@@ -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>(