Avoid linking to GPUTracking where possible, use GPUDataTypes / GPUCo…

…mmon / GPUO2Interface
AliceO2Group · Mar 1, 2024 · b6c1a14 · b6c1a14
1 parent 60229bc
commit b6c1a14
Show file tree

Hide file tree

Showing 27 changed files with 242 additions and 154 deletions.
diff --git a/Detectors/Base/CMakeLists.txt b/Detectors/Base/CMakeLists.txt
@@ -40,9 +40,9 @@ o2_add_library(DetectorsBase
                                      O2::SimulationDataFormat
                                      O2::SimConfig
                                      O2::CCDB
+                                     O2::GPUDataTypeHeaders
                                      MC::VMC
                                      TBB::tbb
-               PRIVATE_INCLUDE_DIRECTORIES ${CMAKE_SOURCE_DIR}/GPU/GPUTracking/Merger # Must not link to avoid cyclic dependency
                              )
 
 o2_target_root_dictionary(DetectorsBase

diff --git a/Detectors/TPC/calibration/CMakeLists.txt b/Detectors/TPC/calibration/CMakeLists.txt
@@ -60,7 +60,7 @@ o2_add_library(TPCCalibration
                                      Microsoft.GSL::GSL
                                      O2::DetectorsCalibration
                                      O2::GPUO2Interface
-                                     O2::GPUTracking
+                                     O2::GPUDataTypes
                                      O2::TPCFastTransformation
                                      O2::DataFormatsCTP
                                      O2::CCDB

diff --git a/Detectors/TPC/calibration/src/TrackDump.cxx b/Detectors/TPC/calibration/src/TrackDump.cxx
@@ -13,10 +13,7 @@
 #include <vector>
 #include <filesystem>
 #include <fmt/format.h>
-#include "GPUO2Interface.h"
-#include "GPUDefOpenCL12Templates.h"
-#include "GPUDefConstantsAndSettings.h"
-#include "SliceTracker/GPUTPCGeometry.h"
+#include "GPUTPCGeometry.h"
 
 #include "DataFormatsTPC/Defs.h"
 #include "DataFormatsTPC/ClusterNative.h"

diff --git a/Detectors/TPC/reconstruction/CMakeLists.txt b/Detectors/TPC/reconstruction/CMakeLists.txt
@@ -92,7 +92,7 @@ o2_add_test(AdcClockMonitor
 o2_add_test(GPUCATracking
             COMPONENT_NAME tpc
             LABELS tpc
-            PUBLIC_LINK_LIBRARIES O2::TPCReconstruction
+            PUBLIC_LINK_LIBRARIES O2::TPCReconstruction O2::GPUO2Interface
             SOURCES test/testGPUCATracking.cxx
             ENVIRONMENT O2_ROOT=${CMAKE_BINARY_DIR}/stage)
 

diff --git a/Detectors/TPC/reconstruction/test/testGPUCATracking.cxx b/Detectors/TPC/reconstruction/test/testGPUCATracking.cxx
@@ -27,6 +27,7 @@
 #include "CorrectionMapsHelper.h"
 #include "TPCFastTransform.h"
 #include "GPUO2Interface.h"
+#include "GPUO2InterfaceUtils.h"
 #include "GPUO2InterfaceConfiguration.h"
 #include "TPCPadGainCalib.h"
 #include "CalibdEdxContainer.h"
@@ -79,9 +80,9 @@ BOOST_AUTO_TEST_CASE(CATracking_test1)
   fastTransformHelper->setCorrMap(fastTransform.get());
   config.configCalib.fastTransform = fastTransform.get();
   config.configCalib.fastTransformHelper = fastTransformHelper.get();
-  auto dEdxCalibContainer = GPUO2Interface::getCalibdEdxContainerDefault();
+  auto dEdxCalibContainer = GPUO2InterfaceUtils::getCalibdEdxContainerDefault();
   config.configCalib.dEdxCalibContainer = dEdxCalibContainer.get();
-  std::unique_ptr<TPCPadGainCalib> gainCalib = GPUO2Interface::getPadGainCalibDefault();
+  std::unique_ptr<TPCPadGainCalib> gainCalib = GPUO2InterfaceUtils::getPadGainCalibDefault();
   config.configCalib.tpcPadGain = gainCalib.get();
 
   tracker.Initialize(config);

diff --git a/Detectors/TPC/workflow/CMakeLists.txt b/Detectors/TPC/workflow/CMakeLists.txt
@@ -51,8 +51,8 @@ o2_add_library(TPCWorkflow
                                      O2::TPCQC O2::DetectorsCalibration
                                      O2::TPCReaderWorkflow
                                      O2::CTPWorkflowIO
-               PRIVATE_LINK_LIBRARIES O2::GPUTracking # For the Zero Suppression includes
-                                      O2::GPUWorkflow
+                                     O2::GPUWorkflow
+               PRIVATE_LINK_LIBRARIES O2::GPUTracking
            )
 
 o2_add_library(TPCWorkflowGUI
@@ -62,7 +62,6 @@ o2_add_library(TPCWorkflowGUI
                                      O2::TPCMonitor
            )
 
-
 o2_add_executable(chunkeddigit-merger
         COMPONENT_NAME tpc
         TARGETVARNAME mergertargetName
@@ -250,7 +249,7 @@ o2_add_test(workflow
 
 o2_add_executable(digits-to-rawzs
                   COMPONENT_NAME tpc
-                  PUBLIC_LINK_LIBRARIES O2::TPCBase O2::SimulationDataFormat O2::GPUO2Interface O2::GPUTracking O2::DetectorsRaw O2::TPCReconstruction
+                  PUBLIC_LINK_LIBRARIES O2::TPCBase O2::SimulationDataFormat O2::GPUO2Interface O2::DetectorsRaw O2::TPCReconstruction
                   SOURCES src/convertDigitsToRawZS.cxx)
 
 o2_add_executable(monitor-workflow

diff --git a/Detectors/TPC/workflow/src/CalibProcessingHelper.cxx b/Detectors/TPC/workflow/src/CalibProcessingHelper.cxx
@@ -17,9 +17,7 @@
 #include <fmt/format.h>
 #include <fmt/chrono.h>
 
-#include "GPUO2Interface.h"
-#include "GPUParam.h"
-#include "GPUReconstructionConvert.h"
+#include "GPUO2InterfaceUtils.h"
 #include "Framework/ConcreteDataMatcher.h"
 #include "Framework/InputRecordWalker.h"
 #include "Framework/Logger.h"
@@ -290,9 +288,8 @@ void processLinkZS(o2::framework::RawParser<>& parser, std::unique_ptr<o2::tpc::
 
     if ((decoderType == 1) && (linkID == rdh_utils::ILBZSLinkID || linkID == rdh_utils::DLBZSLinkID) && (detField == raw_data_types::Type::ZS)) {
       std::vector<Digit> digits;
-      static o2::gpu::GPUParam gpuParam;
-      static o2::gpu::GPUReconstructionZSDecoder gpuDecoder;
-      gpuDecoder.DecodePage(digits, (const void*)it.raw(), firstOrbit, gpuParam, static_cast<unsigned int>((triggerBC > 0) ? triggerBC : 0));
+      static o2::gpu::GPUO2InterfaceUtils::GPUReconstructionZSDecoder gpuDecoder;
+      gpuDecoder.DecodePage(digits, (const void*)it.raw(), firstOrbit, nullptr, static_cast<unsigned int>((triggerBC > 0) ? triggerBC : 0));
       for (const auto& digit : digits) {
         reader->getManager()->getLinkZSCallback()(digit.getCRU(), digit.getRow(), digit.getPad(), digit.getTimeStamp(), digit.getChargeFloat());
       }

diff --git a/Detectors/TPC/workflow/src/ZSSpec.cxx b/Detectors/TPC/workflow/src/ZSSpec.cxx
@@ -35,8 +35,7 @@
 #include <array>
 #include <unistd.h>
 #include <atomic>
-#include "GPUParam.h"
-#include "GPUReconstructionConvert.h"
+#include "GPUO2InterfaceUtils.h"
 #include "DetectorsRaw/RawFileWriter.h"
 #include "DetectorsRaw/HBFUtils.h"
 #include "DetectorsRaw/RDHUtils.h"
@@ -100,10 +99,7 @@ DataProcessorSpec getZSEncoderSpec(std::vector<int> const& tpcSectors, bool outR
       auto& sizes = processAttributes->sizes;
       auto& verbosity = processAttributes->verbosity;
 
-      GPUParam _GPUParam;
-
       processAttributes->config.configGRP.solenoidBz = 5.00668;
-      _GPUParam.SetDefaults(&processAttributes->config.configGRP, &processAttributes->config.configReconstruction, &processAttributes->config.configProcessing, nullptr);
       std::function<void(std::vector<o2::tpc::Digit>&)> digitsFilter = nullptr;
       if (processAttributes->globalConfig.zsOnTheFlyDigitsFilter) {
         digitsFilter = [processAttributes](std::vector<o2::tpc::Digit>& digits) {
@@ -119,7 +115,7 @@ DataProcessorSpec getZSEncoderSpec(std::vector<int> const& tpcSectors, bool outR
       const auto& inputs = getWorkflowTPCInput(pc, 0, false, false, tpcSectorMask, true);
       sizes.resize(NSectors * NEndpoints);
       o2::InteractionRecord ir{0, pc.services().get<o2::framework::TimingInfo>().firstTForbit};
-      o2::gpu::GPUReconstructionConvert::RunZSEncoder<DigitArray>(inputs->inputDigits, &zsoutput, sizes.data(), nullptr, &ir, _GPUParam, 4, verify, processAttributes->config.configReconstruction.tpc.zsThreshold, false, digitsFilter);
+      o2::gpu::GPUO2InterfaceUtils::RunZSEncoder<DigitArray>(inputs->inputDigits, &zsoutput, sizes.data(), nullptr, &ir, processAttributes->config, 4, verify, false, digitsFilter);
       ZeroSuppressedContainer8kb* page = reinterpret_cast<ZeroSuppressedContainer8kb*>(zsoutput.get());
       unsigned int offset = 0;
       for (unsigned int i = 0; i < NSectors; i++) {
@@ -172,7 +168,7 @@ DataProcessorSpec getZSEncoderSpec(std::vector<int> const& tpcSectors, bool outR
           writer.useCaching();
         }
         ir = o2::raw::HBFUtils::Instance().getFirstSampledTFIR();
-        o2::gpu::GPUReconstructionConvert::RunZSEncoder(inputs->inputDigits, nullptr, nullptr, &writer, &ir, _GPUParam, 4, false, processAttributes->config.configReconstruction.tpc.zsThreshold, false, digitsFilter);
+        o2::gpu::GPUO2InterfaceUtils::RunZSEncoder(inputs->inputDigits, nullptr, nullptr, &writer, &ir, processAttributes->config, 4, false, false, digitsFilter);
         writer.writeConfFile("TPC", "RAWDATA", fmt::format("{}tpcraw.cfg", outDir));
       }
       zsoutput.reset(nullptr);

diff --git a/Detectors/TPC/workflow/src/convertDigitsToRawZS.cxx b/Detectors/TPC/workflow/src/convertDigitsToRawZS.cxx
@@ -23,10 +23,8 @@
 #include "TTree.h"
 #include "TROOT.h"
 
-#include "GPUO2Interface.h"
-#include "GPUReconstructionConvert.h"
+#include "GPUO2InterfaceUtils.h"
 #include "GPUHostDataTypes.h"
-#include "GPUParam.h"
 
 #include "Framework/Logger.h"
 #include "DetectorsRaw/RawFileWriter.h"
@@ -233,13 +231,10 @@ void convertDigitsToZSfinal(std::string_view digitsFile, std::string_view output
 void convert(DigitArray& inputDigits, ProcessAttributes* processAttributes, o2::raw::RawFileWriter& writer)
 {
   const auto zsThreshold = processAttributes->zsThreshold;
-  GPUParam _GPUParam;
-  _GPUParam.SetDefaults(5.00668);
-  const GPUParam mGPUParam = _GPUParam;
 
   o2::InteractionRecord ir = o2::raw::HBFUtils::Instance().getFirstSampledTFIR();
   ir.bc = 0; // By convention the TF starts at BC = 0
-  o2::gpu::GPUReconstructionConvert::RunZSEncoder(inputDigits, nullptr, nullptr, &writer, &ir, mGPUParam, processAttributes->version, false, zsThreshold, processAttributes->padding, processAttributes->digitsFilter);
+  o2::gpu::GPUO2InterfaceUtils::RunZSEncoder(inputDigits, nullptr, nullptr, &writer, &ir, processAttributes->version, false, zsThreshold, processAttributes->padding, processAttributes->digitsFilter);
 }
 
 int main(int argc, char** argv)

diff --git a/Detectors/TRD/calibration/CMakeLists.txt b/Detectors/TRD/calibration/CMakeLists.txt
@@ -28,7 +28,6 @@ o2_add_library(TRDCalibration
                                      O2::DetectorsBase
                                      O2::DetectorsCalibration
                                      O2::MathUtils
-                                     O2::GPUTracking
                                      O2::DetectorsDCS)
 
  o2_target_root_dictionary(TRDCalibration

diff --git a/Detectors/TRD/pid/CMakeLists.txt b/Detectors/TRD/pid/CMakeLists.txt
@@ -17,7 +17,6 @@ o2_add_library(TRDPID
                                       O2::DataFormatsGlobalTracking
                                       O2::DetectorsBase
                                       O2::MathUtils
-                                      O2::GPUTracking
                                       O2::Framework
                                       O2::ReconstructionDataFormats
                                       fmt::fmt

diff --git a/Detectors/TRD/qc/CMakeLists.txt b/Detectors/TRD/qc/CMakeLists.txt
@@ -21,8 +21,7 @@ o2_add_library(TRDQC
                                      O2::DataFormatsTRD
                                      O2::DataFormatsGlobalTracking
                                      O2::DetectorsBase
-                                     O2::MathUtils
-                                     O2::GPUTracking)
+                                     O2::MathUtils)
 
  o2_target_root_dictionary(TRDQC
                            HEADERS include/TRDQC/Tracking.h

diff --git a/GPU/Common/GPUCommonDef.h b/GPU/Common/GPUCommonDef.h
@@ -54,6 +54,7 @@
 #ifdef GPUCA_NOCOMPAT
   #define CON_DELETE = delete
   #define CON_DEFAULT = default
+  #define GPUCA_CPP11_INIT(...) __VA_ARGS__
   #if defined(__cplusplus) && __cplusplus >= 201703L
     #define CONSTEXPR constexpr
   #else
@@ -63,6 +64,7 @@
   #define CON_DELETE
   #define CON_DEFAULT
   #define CONSTEXPR
+  #define GPUCA_CPP11_INIT(...)
 #endif
 #if defined(__ROOT__) && !defined(GPUCA_NOCOMPAT)
   #define VOLATILE // ROOT5 has a problem with volatile in CINT

diff --git a/GPU/GPUTracking/Base/GPUReconstructionConvert.cxx b/GPU/GPUTracking/Base/GPUReconstructionConvert.cxx
@@ -1464,10 +1464,14 @@ void GPUReconstructionConvert::RunZSFilter(std::unique_ptr<o2::tpc::Digit[]>* bu
 
 #ifdef GPUCA_O2_LIB
 template <class T>
-static inline auto GetDecoder_internal(const GPUParam& param, int version)
+static inline auto GetDecoder_internal(const GPUParam* param, int version)
 {
   std::shared_ptr<T> enc = std::make_shared<T>();
-  enc->param = &param;
+  if (param == nullptr) {
+    static GPUParam dummyParam;
+    param = &dummyParam;
+  }
+  enc->param = param;
   enc->zsVersion = version;
   enc->init();
   return [enc](std::vector<o2::tpc::Digit>& outBuffer, const void* page, unsigned int firstTfOrbit, unsigned int triggerBC = 0) {
@@ -1485,33 +1489,16 @@ static inline auto GetDecoder_internal(const GPUParam& param, int version)
   };
 }
 
-std::function<void(std::vector<o2::tpc::Digit>&, const void*, unsigned int, unsigned int)> GPUReconstructionConvert::GetDecoder(int version, const GPUParam& param)
+std::function<void(std::vector<o2::tpc::Digit>&, const void*, unsigned int, unsigned int)> GPUReconstructionConvert::GetDecoder(int version, const GPUParam* param)
 {
-  if (version >= ZSVersion::ZSVersionRowBased10BitADC && version <= ZSVersion::ZSVersionRowBased12BitADC) {
+  if (version >= o2::tpc::ZSVersion::ZSVersionRowBased10BitADC && version <= o2::tpc::ZSVersion::ZSVersionRowBased12BitADC) {
     return GetDecoder_internal<zsEncoderRow>(param, version);
-  } else if (version == ZSVersion::ZSVersionLinkBasedWithMeta) {
+  } else if (version == o2::tpc::ZSVersion::ZSVersionLinkBasedWithMeta) {
     return GetDecoder_internal<zsEncoderImprovedLinkBased>(param, version);
-  } else if (version >= ZSVersion::ZSVersionDenseLinkBased && version <= ZSVersion::ZSVersionDenseLinkBasedV2) {
+  } else if (version >= o2::tpc::ZSVersion::ZSVersionDenseLinkBased && version <= o2::tpc::ZSVersion::ZSVersionDenseLinkBasedV2) {
     return GetDecoder_internal<zsEncoderDenseLinkBased>(param, version);
   } else {
     throw std::runtime_error("Invalid ZS version "s + std::to_string(version) + ", cannot create decoder"s);
   }
 }
-
-void GPUReconstructionZSDecoder::DecodePage(std::vector<o2::tpc::Digit>& outputBuffer, const void* page, unsigned int tfFirstOrbit, const GPUParam& param, unsigned int triggerBC)
-{
-  const o2::header::RAWDataHeader* rdh = (const o2::header::RAWDataHeader*)page;
-  if (o2::raw::RDHUtils::getMemorySize(*rdh) == sizeof(o2::header::RAWDataHeader)) {
-    return;
-  }
-  TPCZSHDR* const hdr = (TPCZSHDR*)(rdh_utils::getLink(o2::raw::RDHUtils::getFEEID(*rdh)) == rdh_utils::DLBZSLinkID ? ((const char*)page + o2::raw::RDHUtils::getMemorySize(*rdh) - sizeof(TPCZSHDRV2)) : ((const char*)page + sizeof(o2::header::RAWDataHeader)));
-
-  if (mDecoders.size() < hdr->version + 1) {
-    mDecoders.resize(hdr->version + 1);
-  }
-  if (mDecoders[hdr->version] == nullptr) {
-    mDecoders[hdr->version] = GPUReconstructionConvert::GetDecoder(hdr->version, param);
-  }
-  mDecoders[hdr->version](outputBuffer, page, tfFirstOrbit, triggerBC);
-}
 #endif
diff --git a/GPU/GPUTracking/Base/GPUReconstructionConvert.h b/GPU/GPUTracking/Base/GPUReconstructionConvert.h
@@ -60,16 +60,7 @@ class GPUReconstructionConvert
   static int GetMaxTimeBin(const o2::tpc::ClusterNativeAccess& native);
   static int GetMaxTimeBin(const GPUTrackingInOutDigits& digits);
   static int GetMaxTimeBin(const GPUTrackingInOutZS& zspages);
-  static std::function<void(std::vector<o2::tpc::Digit>&, const void*, unsigned int, unsigned int)> GetDecoder(int version, const GPUParam& param);
-};
-
-class GPUReconstructionZSDecoder
-{
- public:
-  void DecodePage(std::vector<o2::tpc::Digit>& outputBuffer, const void* page, unsigned int tfFirstOrbit, const GPUParam& param, unsigned int triggerBC = 0);
-
- private:
-  std::vector<std::function<void(std::vector<o2::tpc::Digit>&, const void*, unsigned int, unsigned int)>> mDecoders;
+  static std::function<void(std::vector<o2::tpc::Digit>&, const void*, unsigned int, unsigned int)> GetDecoder(int version, const GPUParam* param);
 };
 
 } // namespace gpu