GPU TPC: Compile-time disable V/M-Shape correction during RTC

Author: davidrohr

GPU/Common/GPUCommonDef.h

@@ -97,6 +97,9 @@
 #define GPUCA_DEBUG_STREAMER_CHECK(...)
 #endif
 
+#ifndef GPUCA_RTC_SPECIAL_CODE
+#define GPUCA_RTC_SPECIAL_CODE(...)
+#endif
 
 // API Definitions for GPU Compilation
 #include "GPUCommonDefAPI.h"

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDAGenRTC.cxx

@@ -39,7 +39,7 @@ QGET_LD_BINARY_SYMBOLS(GPUReconstructionCUDArtc_command_arch);
 int GPUReconstructionCUDA::genRTC(std::string& filename, unsigned int& nCompile)
 {
 #ifndef GPUCA_ALIROOT_LIB
-  std::string rtcparam = GPUParamRTC::generateRTCCode(param(), mProcessingSettings.rtc.optConstexpr);
+  std::string rtcparam = std::string(mProcessingSettings.rtc.optSpecialCode ? "#define GPUCA_RTC_SPECIAL_CODE(...) __VA_ARGS__\n" : "#define GPUCA_RTC_SPECIAL_CODE(...)\n") + GPUParamRTC::generateRTCCode(param(), mProcessingSettings.rtc.optConstexpr);
   if (filename == "") {
     filename = "/tmp/o2cagpu_rtc_";
   }

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDArtc.cu

@@ -14,6 +14,7 @@
 
 #define GPUCA_GPUCODE_GENRTC
 #define GPUCA_GPUCODE_COMPILEKERNELS
+#define GPUCA_RTC_SPECIAL_CODE(...) GPUCA_RTC_SPECIAL_CODE(__VA_ARGS__)
 #include "GPUReconstructionCUDADef.h"
 #include "GPUReconstructionIncludesDeviceAll.h"
 

GPU/GPUTracking/Definitions/GPUSettingsList.h

@@ -198,6 +198,7 @@ EndConfig()
 BeginSubConfig(GPUSettingsProcessingRTC, rtc, configStandalone.proc, "RTC", 0, "Processing settings", proc_rtc)
 AddOption(cacheOutput, bool, false, "", 0, "Cache RTC compilation results")
 AddOption(optConstexpr, bool, true, "", 0, "Replace constant variables by static constexpr expressions")
+AddOption(optSpecialCode, signed char, -1, "", 0, "Insert GPUCA_RTC_SPECIAL_CODE special code during RTC")
 AddOption(compilePerKernel, bool, true, "", 0, "Run one RTC compilation per kernel")
 AddOption(enable, bool, false, "", 0, "Use RTC to optimize GPU code")
 AddOption(runTest, int, 0, "", 0, "Do not run the actual benchmark, but just test RTC compilation (1 full test, 2 test only compilation)")

GPU/GPUTracking/Standalone/Benchmark/standalone.cxx

@@ -434,6 +434,9 @@ int SetupReconstruction()
       procSet.eventDisplay = nullptr;
     }
   }
+  if (configStandalone.proc.rtc.optSpecialCode == -1) {
+    configStandalone.proc.rtc.optSpecialCode = configStandalone.testSyncAsync || configStandalone.testSync;
+  }
 
   rec->SetSettings(&grp, &recSet, &procSet, &steps);
   if (configStandalone.proc.doublePipeline) {
@@ -453,6 +456,7 @@ int SetupReconstruction()
     procSet.runQA = false;
     procSet.eventDisplay = eventDisplay.get();
     procSet.runCompressionStatistics = 0;
+    procSet.rtc.optSpecialCode = 0;
     if (recSet.tpc.rejectionStrategy >= GPUSettings::RejectionStrategyB) {
       procSet.tpcInputWithClusterRejection = 1;
     }

GPU/TPCFastTransformation/TPCFastTransform.h

@@ -446,6 +446,7 @@ GPUdi() void TPCFastTransform::getTOFcorrection(int slice, int /*row*/, float x,
 
 GPUdi() void TPCFastTransform::TransformInternal(int slice, int row, float& u, float& v, float& x, const TPCFastTransform* ref, const TPCFastTransform* ref2, float scale, float scale2, int scaleMode) const
 {
+  GPUCA_RTC_SPECIAL_CODE(ref2 = nullptr; scale2 = 0.f;);
   if (mApplyCorrection) {
     float dx = 0.f, du = 0.f, dv = 0.f;
     if ((scale >= 0.f) || (scaleMode == 1) || (scaleMode == 2)) {
@@ -776,6 +777,7 @@ GPUdi() float TPCFastTransform::getMaxDriftTime(int slice) const
 
 GPUdi() void TPCFastTransform::InverseTransformYZtoX(int slice, int row, float y, float z, float& x, const TPCFastTransform* ref, const TPCFastTransform* ref2, float scale, float scale2, int scaleMode) const
 {
+  GPUCA_RTC_SPECIAL_CODE(ref2 = nullptr; scale2 = 0.f;);
   /// Transformation y,z -> x
   float u = 0, v = 0;
   getGeometry().convLocalToUV(slice, y, z, u, v);
@@ -816,6 +818,7 @@ GPUdi() void TPCFastTransform::InverseTransformYZtoX(int slice, int row, float y
 
 GPUdi() void TPCFastTransform::InverseTransformYZtoNominalYZ(int slice, int row, float y, float z, float& ny, float& nz, const TPCFastTransform* ref, const TPCFastTransform* ref2, float scale, float scale2, int scaleMode) const
 {
+  GPUCA_RTC_SPECIAL_CODE(ref2 = nullptr; scale2 = 0.f;);
   /// Transformation y,z -> x
   float u = 0, v = 0, un = 0, vn = 0;
   getGeometry().convLocalToUV(slice, y, z, u, v);

GPU/Workflow/src/GPUWorkflowSpec.cxx

@@ -187,6 +187,9 @@ void GPURecoWorkflowSpec::init(InitContext& ic)
   if (mConfParam->synchronousProcessing) {
     mConfig->configReconstruction.useMatLUT = false;
   }
+  if (mConfig->configProcessing.rtc.optSpecialCode == -1) {
+    mConfig->configProcessing.rtc.optSpecialCode = mConfParam->synchronousProcessing;
+  }
 
   // Configure the "GPU workflow" i.e. which steps we run on the GPU (or CPU)
   if (mSpecConfig.outputTracks || mSpecConfig.outputCompClusters || mSpecConfig.outputCompClustersFlat) {