From b8d0837f4924cb88f991b367d2ffbec85e631b7f Mon Sep 17 00:00:00 2001
From: Vincenzo Innocente <vincenzo.innocente@cern.ch>
Date: Sun, 15 May 2022 15:54:23 +0200
Subject: [PATCH] use same deleter

---
 .../interface/SimplePoolAllocator.h              |  6 +++---
 .../CUDAUtilities/interface/cudaMemoryPool.h     | 16 ++++------------
 .../plugins/CAHitNtupletGeneratorKernels.cc      |  3 +--
 .../plugins/CAHitNtupletGeneratorKernels.cu      |  3 +--
 4 files changed, 9 insertions(+), 19 deletions(-)

diff --git a/HeterogeneousCore/CUDAUtilities/interface/SimplePoolAllocator.h b/HeterogeneousCore/CUDAUtilities/interface/SimplePoolAllocator.h
index e3f7dfa49bb8a..c80af3c60be98 100644
--- a/HeterogeneousCore/CUDAUtilities/interface/SimplePoolAllocator.h
+++ b/HeterogeneousCore/CUDAUtilities/interface/SimplePoolAllocator.h
@@ -126,13 +126,13 @@ template <typename T>
 struct SimplePoolAllocatorImpl final : public SimplePoolAllocator {
   using Traits = T;
 
-  SimplePoolAllocatorImpl(int maxSlots) : SimplePoolAllocator(maxSlots) {}
+  using SimplePoolAllocator::SimplePoolAllocator;
 
   ~SimplePoolAllocatorImpl() override {
     garbageCollect();
-    //#ifdef MEMORY_POOL_DEBUG
+#ifdef MEMORY_POOL_DEBUG
     dumpStat();
-    //#endif
+#endif
   }
 
   Pointer doAlloc(size_t size) override { return Traits::alloc(size); }
diff --git a/HeterogeneousCore/CUDAUtilities/interface/cudaMemoryPool.h b/HeterogeneousCore/CUDAUtilities/interface/cudaMemoryPool.h
index 9e2d284f2716f..b99d425eca25c 100644
--- a/HeterogeneousCore/CUDAUtilities/interface/cudaMemoryPool.h
+++ b/HeterogeneousCore/CUDAUtilities/interface/cudaMemoryPool.h
@@ -32,17 +32,9 @@ namespace memoryPool {
     }
 
     struct CudaDeleterBase : public DeleterBase {
-      CudaDeleterBase(cudaStream_t const &stream, Where where) : DeleterBase(getPool(where)), m_stream(stream) {
-        //         if (stream) return;
-        //         std::cout << "0 stream???" << std::endl;
-        //         throw std::bad_alloc();
-      }
+      CudaDeleterBase(cudaStream_t const &stream, Where where) : DeleterBase(getPool(where)), m_stream(stream) {}
 
-      CudaDeleterBase(cudaStream_t const &stream, SimplePoolAllocator *pool) : DeleterBase(pool), m_stream(stream) {
-        //           if (stream) return;
-        //            std::cout << "0 stream???" << std::endl;
-        //            throw std::bad_alloc();
-      }
+      CudaDeleterBase(cudaStream_t const &stream, SimplePoolAllocator *pool) : DeleterBase(pool), m_stream(stream) {}
 
       ~CudaDeleterBase() override = default;
 
@@ -57,7 +49,7 @@ namespace memoryPool {
     };
 
     struct BundleDelete final : public CudaDeleterBase {
-      using CudaDeleterBase::CudaDeleterBase;
+      BundleDelete(cudaStream_t const &stream, Where where) : CudaDeleterBase(stream, where) { m_buckets.reserve(8); }
 
       ~BundleDelete() override { free(m_stream, std::move(m_buckets), *pool()); }
 
@@ -88,7 +80,7 @@ namespace memoryPool {
 
     template <typename T>
     Buffer<T> makeBuffer(uint64_t size, cudaStream_t const &stream, Where where) {
-      return makeBuffer<T>(size, Deleter(std::make_shared<DeleteOne>(stream, getPool(where))));
+      return makeBuffer<T>(size, Deleter(std::make_shared<DeleteOne>(stream, where)));
     }
 
   }  // namespace cuda
diff --git a/RecoPixelVertexing/PixelTriplets/plugins/CAHitNtupletGeneratorKernels.cc b/RecoPixelVertexing/PixelTriplets/plugins/CAHitNtupletGeneratorKernels.cc
index f41a716d1a060..d48af27a329be 100644
--- a/RecoPixelVertexing/PixelTriplets/plugins/CAHitNtupletGeneratorKernels.cc
+++ b/RecoPixelVertexing/PixelTriplets/plugins/CAHitNtupletGeneratorKernels.cc
@@ -22,8 +22,7 @@ void CAHitNtupletGeneratorKernelsCPU::buildDoublets(HitsOnCPU const &hh, cudaStr
 #endif
 
   // use "nhits" to heuristically dimension the workspace
-  memoryPool::Deleter deleter =
-      memoryPool::Deleter(std::make_shared<memoryPool::cuda::BundleDelete>(nullptr, memoryPool::onCPU));
+  memoryPool::Deleter const &deleter = device_storage_.deleter();
   device_isOuterHitOfCell_ =
       memoryPool::cuda::makeBuffer<GPUCACell::OuterHitOfCellContainer>(std::max(1U, nhits), deleter);
   assert(device_isOuterHitOfCell_.get());
diff --git a/RecoPixelVertexing/PixelTriplets/plugins/CAHitNtupletGeneratorKernels.cu b/RecoPixelVertexing/PixelTriplets/plugins/CAHitNtupletGeneratorKernels.cu
index 5866f28f8721b..e29b15fc80e53 100644
--- a/RecoPixelVertexing/PixelTriplets/plugins/CAHitNtupletGeneratorKernels.cu
+++ b/RecoPixelVertexing/PixelTriplets/plugins/CAHitNtupletGeneratorKernels.cu
@@ -145,8 +145,7 @@ void CAHitNtupletGeneratorKernelsGPU::buildDoublets(HitsOnCPU const &hh, cudaStr
 #endif
 
   // in principle we can use "nhits" to heuristically dimension the workspace...
-  memoryPool::Deleter deleter =
-      memoryPool::Deleter(std::make_shared<memoryPool::cuda::BundleDelete>(stream, memoryPool::onDevice));
+  memoryPool::Deleter const &deleter = device_storage_.deleter();
   device_isOuterHitOfCell_ =
       memoryPool::cuda::makeBuffer<GPUCACell::OuterHitOfCellContainer>(std::max(1, nhits), deleter);
   assert(device_isOuterHitOfCell_.get());