Replace TAcc with the explicit accelerator type

cms-sw · Dec 11, 2024 · 53a0974 · 53a0974
1 parent e648b61
commit 53a0974
Show file tree

Hide file tree

Showing 5 changed files with 45 additions and 56 deletions.
diff --git a/RecoLocalCalo/EcalRecProducers/plugins/alpaka/AmplitudeComputationCommonKernels.h b/RecoLocalCalo/EcalRecProducers/plugins/alpaka/AmplitudeComputationCommonKernels.h
@@ -28,8 +28,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
   ///
   class Kernel_prep_1d_and_initialize {
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   EcalDigiDeviceCollection::ConstView digisDevEB,
                                   EcalDigiDeviceCollection::ConstView digisDevEE,
                                   EcalUncalibratedRecHitDeviceCollection::View uncalibRecHitsEB,
@@ -328,8 +327,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
   ///
   class Kernel_prep_2d {
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc2D const& acc,
                                   EcalDigiDeviceCollection::ConstView digisDevEB,
                                   EcalDigiDeviceCollection::ConstView digisDevEE,
                                   EcalMultifitConditionsDevice::ConstView conditionsDev,
@@ -466,11 +464,12 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
 }  // namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit
 
 namespace alpaka::trait {
+  using namespace ALPAKA_ACCELERATOR_NAMESPACE;
   using namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit;
 
   //! The trait for getting the size of the block shared dynamic memory for Kernel_prep_1d_and_initialize.
-  template <typename TAcc>
-  struct BlockSharedMemDynSizeBytes<Kernel_prep_1d_and_initialize, TAcc> {
+  template <>
+  struct BlockSharedMemDynSizeBytes<Kernel_prep_1d_and_initialize, Acc1D> {
     //! \return The size of the shared memory allocated for a block.
     template <typename TVec, typename... TArgs>
     ALPAKA_FN_HOST_ACC static auto getBlockSharedMemDynSizeBytes(Kernel_prep_1d_and_initialize const&,

diff --git a/RecoLocalCalo/EcalRecProducers/plugins/alpaka/AmplitudeComputationKernels.dev.cc b/RecoLocalCalo/EcalRecProducers/plugins/alpaka/AmplitudeComputationKernels.dev.cc
@@ -55,8 +55,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
   ///
   class Kernel_minimize {
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   InputProduct::ConstView const& digisDevEB,
                                   InputProduct::ConstView const& digisDevEE,
                                   OutputProduct::View uncalibRecHitsEB,
@@ -293,23 +292,25 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
 }  // namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit
 
 namespace alpaka::trait {
+  using namespace ALPAKA_ACCELERATOR_NAMESPACE;
   using namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit;
 
   //! The trait for getting the size of the block shared dynamic memory for Kernel_minimize.
-  template <typename TAcc>
-  struct BlockSharedMemDynSizeBytes<Kernel_minimize, TAcc> {
+  template <>
+  struct BlockSharedMemDynSizeBytes<Kernel_minimize, Acc1D> {
     //! \return The size of the shared memory allocated for a block.
     template <typename TVec, typename... TArgs>
     ALPAKA_FN_HOST_ACC static auto getBlockSharedMemDynSizeBytes(Kernel_minimize const&,
                                                                  TVec const& threadsPerBlock,
                                                                  TVec const& elemsPerThread,
                                                                  TArgs const&...) -> std::size_t {
-      using ScalarType = ecal::multifit::SampleVector::Scalar;
+      using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
       // return the amount of dynamic shared memory needed
-      std::size_t bytes = 2 * threadsPerBlock[0u] * elemsPerThread[0u] *
-                          calo::multifit::MapSymM<ScalarType, ecal::multifit::SampleVector::RowsAtCompileTime>::total *
-                          sizeof(ScalarType);
+      std::size_t bytes =
+          2 * threadsPerBlock[0u] * elemsPerThread[0u] *
+          calo::multifit::MapSymM<ScalarType, ::ecal::multifit::SampleVector::RowsAtCompileTime>::total *
+          sizeof(ScalarType);
       return bytes;
     }
   };

diff --git a/RecoLocalCalo/EcalRecProducers/plugins/alpaka/TimeComputationKernels.h b/RecoLocalCalo/EcalRecProducers/plugins/alpaka/TimeComputationKernels.h
@@ -33,8 +33,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
     using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   ScalarType* const sample_values,
                                   ScalarType* const sample_value_errors,
                                   bool* const useless_sample_values,
@@ -120,8 +119,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
     using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   EcalDigiDeviceCollection::ConstView digisDevEB,
                                   EcalDigiDeviceCollection::ConstView digisDevEE,
                                   ScalarType* const sample_values,
@@ -523,8 +521,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
     using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   EcalDigiDeviceCollection::ConstView digisDevEB,
                                   EcalDigiDeviceCollection::ConstView digisDevEE,
                                   ScalarType* const sample_values,
@@ -730,8 +727,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
     using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   EcalDigiDeviceCollection::ConstView digisDevEB,
                                   EcalDigiDeviceCollection::ConstView digisDevEE,
                                   EcalMultifitConditionsDevice::ConstView conditionsDev,
@@ -777,8 +773,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
     using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   EcalDigiDeviceCollection::ConstView digisDevEB,
                                   EcalDigiDeviceCollection::ConstView digisDevEE,
                                   EcalMultifitConditionsDevice::ConstView conditionsDev,
@@ -960,8 +955,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
     using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   EcalDigiDeviceCollection::ConstView digisDevEB,
                                   EcalDigiDeviceCollection::ConstView digisDevEE,
                                   EcalUncalibratedRecHitDeviceCollection::View uncalibRecHitsEB,
@@ -1095,18 +1089,19 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit {
 }  // namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit
 
 namespace alpaka::trait {
+  using namespace ALPAKA_ACCELERATOR_NAMESPACE;
   using namespace ALPAKA_ACCELERATOR_NAMESPACE::ecal::multifit;
 
   //! The trait for getting the size of the block shared dynamic memory for Kernel_time_compute_nullhypot.
-  template <typename TAcc>
-  struct BlockSharedMemDynSizeBytes<Kernel_time_compute_nullhypot, TAcc> {
+  template <>
+  struct BlockSharedMemDynSizeBytes<Kernel_time_compute_nullhypot, Acc1D> {
     //! \return The size of the shared memory allocated for a block.
     template <typename TVec, typename... TArgs>
     ALPAKA_FN_HOST_ACC static auto getBlockSharedMemDynSizeBytes(Kernel_time_compute_nullhypot const&,
                                                                  TVec const& threadsPerBlock,
                                                                  TVec const& elemsPerThread,
                                                                  TArgs const&...) -> std::size_t {
-      using ScalarType = ecal::multifit::SampleVector::Scalar;
+      using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
       // return the amount of dynamic shared memory needed
       std::size_t bytes = threadsPerBlock[0u] * elemsPerThread[0u] * 4 * sizeof(ScalarType);
@@ -1115,44 +1110,44 @@ namespace alpaka::trait {
   };
 
   //! The trait for getting the size of the block shared dynamic memory for Kernel_time_compute_makeratio.
-  template <typename TAcc>
-  struct BlockSharedMemDynSizeBytes<Kernel_time_compute_makeratio, TAcc> {
+  template <>
+  struct BlockSharedMemDynSizeBytes<Kernel_time_compute_makeratio, Acc1D> {
     template <typename TVec, typename... TArgs>
     ALPAKA_FN_HOST_ACC static auto getBlockSharedMemDynSizeBytes(Kernel_time_compute_makeratio const&,
                                                                  TVec const& threadsPerBlock,
                                                                  TVec const& elemsPerThread,
                                                                  TArgs const&...) -> std::size_t {
-      using ScalarType = ecal::multifit::SampleVector::Scalar;
+      using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
       std::size_t bytes = (8 * sizeof(ScalarType) + 3 * sizeof(bool)) * threadsPerBlock[0u] * elemsPerThread[0u];
       return bytes;
     }
   };
 
   //! The trait for getting the size of the block shared dynamic memory for Kernel_time_compute_findamplchi2_and_finish.
-  template <typename TAcc>
-  struct BlockSharedMemDynSizeBytes<Kernel_time_compute_findamplchi2_and_finish, TAcc> {
+  template <>
+  struct BlockSharedMemDynSizeBytes<Kernel_time_compute_findamplchi2_and_finish, Acc1D> {
     template <typename TVec, typename... TArgs>
     ALPAKA_FN_HOST_ACC static auto getBlockSharedMemDynSizeBytes(Kernel_time_compute_findamplchi2_and_finish const&,
                                                                  TVec const& threadsPerBlock,
                                                                  TVec const& elemsPerThread,
                                                                  TArgs const&...) -> std::size_t {
-      using ScalarType = ecal::multifit::SampleVector::Scalar;
+      using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
       std::size_t bytes = 2 * threadsPerBlock[0u] * elemsPerThread[0u] * sizeof(ScalarType);
       return bytes;
     }
   };
 
   //! The trait for getting the size of the block shared dynamic memory for Kernel_time_computation_init.
-  template <typename TAcc>
-  struct BlockSharedMemDynSizeBytes<Kernel_time_computation_init, TAcc> {
+  template <>
+  struct BlockSharedMemDynSizeBytes<Kernel_time_computation_init, Acc1D> {
     template <typename TVec, typename... TArgs>
     ALPAKA_FN_HOST_ACC static auto getBlockSharedMemDynSizeBytes(Kernel_time_computation_init const&,
                                                                  TVec const& threadsPerBlock,
                                                                  TVec const& elemsPerThread,
                                                                  TArgs const&...) -> std::size_t {
-      using ScalarType = ecal::multifit::SampleVector::Scalar;
+      using ScalarType = ::ecal::multifit::SampleVector::Scalar;
 
       std::size_t bytes = 2 * threadsPerBlock[0u] * elemsPerThread[0u] * sizeof(ScalarType);
       return bytes;

diff --git a/RecoLocalCalo/HGCalRecProducers/plugins/alpaka/HGCalLayerClustersSoAAlgoWrapper.dev.cc b/RecoLocalCalo/HGCalRecProducers/plugins/alpaka/HGCalLayerClustersSoAAlgoWrapper.dev.cc
@@ -13,8 +13,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
   // Set energy and number of hits in each clusters
   class HGCalLayerClustersSoAAlgoKernelEnergy {
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   const unsigned int numer_of_clusters,
                                   const HGCalSoARecHitsDeviceCollection::ConstView input_rechits_soa,
                                   const HGCalSoARecHitsExtraDeviceCollection::ConstView input_clusters_soa,
@@ -38,8 +37,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
   // Kernel to find the max for every cluster
   class HGCalLayerClustersSoAAlgoKernelPositionByHits {
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   const unsigned int numer_of_clusters,
                                   float thresholdW0,
                                   float positionDeltaRho2,
@@ -83,8 +81,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
   // Real Kernel position
   class HGCalLayerClustersSoAAlgoKernelPositionByHits2 {
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   const unsigned int numer_of_clusters,
                                   float thresholdW0,
                                   float positionDeltaRho2,
@@ -121,8 +118,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
   // Besides the final position, add also the DetId of the seed of each cluster
   class HGCalLayerClustersSoAAlgoKernelPositionByHits3 {
   public:
-    template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
+    ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                   const unsigned int numer_of_clusters,
                                   float thresholdW0,
                                   float positionDeltaRho2,

diff --git a/RecoLocalCalo/HcalRecProducers/plugins/alpaka/Mahi.dev.cc b/RecoLocalCalo/HcalRecProducers/plugins/alpaka/Mahi.dev.cc
@@ -268,8 +268,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
 
       class Kernel_prep1d_sameNumberOfSamples {
       public:
-        template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-        ALPAKA_FN_ACC void operator()(TAcc const& acc,
+        ALPAKA_FN_ACC void operator()(Acc2D const& acc,
                                       OProductType::View outputGPU,
                                       IProductTypef01::ConstView f01HEDigis,
                                       IProductTypef5::ConstView f5HBDigis,
@@ -748,8 +747,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
 
       class Kernel_prep_pulseMatrices_sameNumberOfSamples {
       public:
-        template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-        ALPAKA_FN_ACC void operator()(TAcc const& acc,
+        ALPAKA_FN_ACC void operator()(Acc3D const& acc,
                                       float* pulseMatrices,
                                       float* pulseMatricesM,
                                       float* pulseMatricesP,
@@ -968,8 +966,7 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
       template <int NSAMPLES, int NPULSES>
       class Kernel_minimize {
       public:
-        template <typename TAcc, typename = std::enable_if_t<alpaka::isAccelerator<TAcc>>>
-        ALPAKA_FN_ACC void operator()(TAcc const& acc,
+        ALPAKA_FN_ACC void operator()(Acc1D const& acc,
                                       OProductType::View outputGPU,
                                       float const* amplitudes,
                                       float* pulseMatrices,
@@ -1408,11 +1405,12 @@ namespace ALPAKA_ACCELERATOR_NAMESPACE {
 }  // namespace ALPAKA_ACCELERATOR_NAMESPACE
 
 namespace alpaka::trait {
+  using namespace ALPAKA_ACCELERATOR_NAMESPACE;
   using namespace ALPAKA_ACCELERATOR_NAMESPACE::hcal::reconstruction::mahi;
 
   //! The trait for getting the size of the block shared dynamic memory for Kernel_prep_1d_and_initialize.
-  template <typename TAcc>
-  struct BlockSharedMemDynSizeBytes<Kernel_prep1d_sameNumberOfSamples, TAcc> {
+  template <>
+  struct BlockSharedMemDynSizeBytes<Kernel_prep1d_sameNumberOfSamples, Acc2D> {
     //! \return The size of the shared memory allocated for a block.
     template <typename TVec, typename... TArgs>
     ALPAKA_FN_HOST_ACC static auto getBlockSharedMemDynSizeBytes(Kernel_prep1d_sameNumberOfSamples const&,
@@ -1431,8 +1429,8 @@ namespace alpaka::trait {
   };
 
   //! The trait for getting the size of the block shared dynamic memory for kernel_minimize.
-  template <int NSAMPLES, int NPULSES, typename TAcc>
-  struct BlockSharedMemDynSizeBytes<Kernel_minimize<NSAMPLES, NPULSES>, TAcc> {
+  template <int NSAMPLES, int NPULSES>
+  struct BlockSharedMemDynSizeBytes<Kernel_minimize<NSAMPLES, NPULSES>, Acc1D> {
     //! \return The size of the shared memory allocated for a block.
     template <typename TVec, typename... TArgs>
     ALPAKA_FN_HOST_ACC static auto getBlockSharedMemDynSizeBytes(Kernel_minimize<NSAMPLES, NPULSES> const&,