POD version of TPCFastSpaceChargeCorrection

Created from the TPCFastTransform using:
std::vector<char> buff; // can be also pmr::vector from DPL make(..) const o2::gpu::TPCFastTransformPOD& pod = o2::gpu::TPCFastTransformPOD::create(buff, tpc_fast_transform);

The buff vector will be expanded during creation and can be sent over DPL. On the receiving side, it should be cast as:
const auto& podTransform = o2::gpu::TPCFastTransformPOD::get(pc.inputs().get<gsl::span<char>>(ref));

No initialization is needed, the transform methods (at the moment all methods of TPCFastSpaceChargeCorrection are implemented) can be directly queried from the object received over sh.memory.

Add perfmormance test for pod map

Author: shahor02

GPU/TPCFastTransformation/CMakeLists.txt

@@ -25,6 +25,7 @@ set(SRCS
     TPCFastSpaceChargeCorrection.cxx
     TPCFastSpaceChargeCorrectionMap.cxx
     TPCFastTransform.cxx
+    TPCFastTransformPOD.cxx
     CorrectionMapsHelper.cxx
 )
 

GPU/TPCFastTransformation/TPCFastSpaceChargeCorrection.h

@@ -37,6 +37,8 @@ namespace gpu
 ///
 class TPCFastSpaceChargeCorrection : public FlatObject
 {
+  friend class TPCFastTransformPOD;
+
  public:
   ///
   /// \brief The struct contains necessary info for TPC padrow

GPU/TPCFastTransformation/TPCFastTransformPOD.cxx

@@ -0,0 +1,220 @@
+// Copyright 2019-2020 CERN and copyright holders of ALICE O2.
+// See https://alice-o2.web.cern.ch/copyright for details of the copyright holders.
+// All rights not expressly granted are reserved.
+//
+// This software is distributed under the terms of the GNU General Public
+// License v3 (GPL Version 3), copied verbatim in the file "COPYING".
+//
+// In applying this license CERN does not waive the privileges and immunities
+// granted to it by virtue of its status as an Intergovernmental Organization
+// or submit itself to any jurisdiction.
+
+/// \file  TPCFastTransformPOD.cxx
+/// \brief Implementation of POD correction map
+///
+/// \author  ruben.shahoayn@cern.ch
+
+#include "TPCFastTransformPOD.h"
+#if !defined(GPUCA_GPUCODE)
+#include <TRandom.h>
+#endif
+
+namespace GPUCA_NAMESPACE
+{
+namespace gpu
+{
+#if !defined(GPUCA_GPUCODE)
+
+size_t TPCFastTransformPOD::estimateSize(const TPCFastTransform& src)
+{
+  // estimate size of own buffer
+  const auto& origCorr = src.getCorrection();
+  const size_t selfSizeFix = sizeof(TPCFastTransformPOD);
+  size_t nextDynOffs = alignOffset(selfSizeFix);
+  nextDynOffs = alignOffset(nextDynOffs + origCorr.mNumberOfScenarios * sizeof(size_t)); // spline scenarios start here
+  // space for splines
+  for (int isc = 0; isc < origCorr.mNumberOfScenarios; isc++) {
+    const auto& spline = origCorr.mScenarioPtr[isc];
+    nextDynOffs = alignOffset(nextDynOffs + sizeof(spline));
+  }
+  // space for splines data
+  for (int is = 0; is < 3; is++) {
+    for (int slice = 0; slice < origCorr.mGeo.getNumberOfSlices(); slice++) {
+      for (int row = 0; row < NROWS; row++) {
+        const auto& spline = origCorr.getSpline(slice, row);
+        int nPar = spline.getNumberOfParameters();
+        if (is == 1) {
+          nPar = nPar / 3;
+        }
+        if (is == 2) {
+          nPar = nPar * 2 / 3;
+        }
+        nextDynOffs += nPar * sizeof(float);
+      }
+    }
+  }
+  nextDynOffs = alignOffset(nextDynOffs);
+  return nextDynOffs;
+}
+
+const TPCFastTransformPOD& TPCFastTransformPOD::create(char* buff, size_t buffSize, const TPCFastTransform& src)
+{
+  // instantiate objec to already created buffer of the right size
+  assert(buffSize > sizeof(TPCFastTransformPOD));
+  const auto& origCorr = src.getCorrection();
+  auto& podMap = getNonConst(buff);
+
+  // copy fixed size data --- start
+  podMap.mNumberOfScenarios = origCorr.mNumberOfScenarios;
+  std::memcpy(&podMap.mGeo, &origCorr.mGeo, sizeof(TPCFastTransformGeo)); // copy geometry (fixed size)
+  for (int row = 0; row < NROWS; row++) {
+    podMap.mRowInfo[row] = origCorr.getRowInfo(row); // dataOffsetBytes will be modified later
+  }
+  for (int slice = 0; slice < TPCFastTransformGeo::getNumberOfSlices(); slice++) {
+    podMap.mSliceInfo[slice] = origCorr.getSliceInfo(slice);
+    for (int row = 0; row < NROWS; row++) {
+      podMap.mSliceRowInfo[NROWS * slice + row] = origCorr.getSliceRowInfo(slice, row);
+    }
+  }
+  podMap.mInterpolationSafetyMargin = origCorr.fInterpolationSafetyMargin;
+  podMap.mTimeStamp = origCorr.mTimeStamp;
+  // copy fixed size data --- end
+
+  size_t nextDynOffs = alignOffset(sizeof(TPCFastTransformPOD));
+  // copy slice scenarios
+  podMap.mOffsScenariosOffsets = nextDynOffs; // spline scenarios offsets start here
+  LOGP(debug, "Set mOffsScenariosOffsets = {}", podMap.mOffsScenariosOffsets);
+  nextDynOffs = alignOffset(nextDynOffs + podMap.mNumberOfScenarios * sizeof(size_t)); // spline scenarios start here
+
+  // copy spline objects
+  size_t* scenOffs = reinterpret_cast<size_t*>(buff + podMap.mOffsScenariosOffsets);
+  for (int isc = 0; isc < origCorr.mNumberOfScenarios; isc++) {
+    scenOffs[isc] = nextDynOffs;
+    const auto& spline = origCorr.mScenarioPtr[isc];
+    if (buffSize < nextDynOffs + sizeof(spline)) {
+      throw std::runtime_error(fmt::format("attempt to copy {} bytes for spline for scenario {} to {}, overflowing the buffer of size {}", sizeof(spline), isc, nextDynOffs + sizeof(spline), buffSize));
+    }
+    std::memcpy(buff + scenOffs[isc], &spline, sizeof(spline));
+    nextDynOffs = alignOffset(nextDynOffs + sizeof(spline));
+    LOGP(debug, "Copy {} bytes for spline scenario {} (ptr:{}) to offsset {}", sizeof(spline), isc, (void*)&spline, scenOffs[isc]);
+  }
+
+  // copy splines data
+  for (int is = 0; is < 3; is++) {
+    float* data = reinterpret_cast<float*>(buff + nextDynOffs);
+    LOGP(debug, "splinID={} start offset {} -> {}", is, nextDynOffs, (void*)data);
+    for (int slice = 0; slice < origCorr.mGeo.getNumberOfSlices(); slice++) {
+      podMap.mSplineDataOffsets[slice][is] = nextDynOffs;
+      size_t rowDataOffs = 0;
+      for (int row = 0; row < NROWS; row++) {
+        const auto& spline = origCorr.getSpline(slice, row);
+        const float* dataOr = origCorr.getSplineData(slice, row, is);
+        int nPar = spline.getNumberOfParameters();
+        if (is == 1) {
+          nPar = nPar / 3;
+        }
+        if (is == 2) {
+          nPar = nPar * 2 / 3;
+        }
+        LOGP(debug, "Copying {} floats for spline{} of slice:{} row:{} to offset {}", nPar, is, slice, row, nextDynOffs);
+        size_t nbcopy = nPar * sizeof(float);
+        if (buffSize < nextDynOffs + nbcopy) {
+          throw std::runtime_error(fmt::format("attempt to copy {} bytes of data for spline{} of slice{}/row{} to {}, overflowing the buffer of size {}", nbcopy, is, slice, row, nextDynOffs, buffSize));
+        }
+        std::memcpy(data, dataOr, nbcopy);
+        podMap.getRowInfo(row).dataOffsetBytes[is] = rowDataOffs;
+        rowDataOffs += nbcopy;
+        data += nPar;
+        nextDynOffs += nbcopy;
+      }
+    }
+  }
+  podMap.mTotalSize = alignOffset(nextDynOffs);
+  if (buffSize != podMap.mTotalSize) {
+    throw std::runtime_error(fmt::format("Estimated buffer size {} differs from filled one {}", buffSize, podMap.mTotalSize));
+  }
+  return podMap;
+}
+
+bool TPCFastTransformPOD::test(const TPCFastTransform& src, int npoints) const
+{
+  if (npoints < 1) {
+    return false;
+  }
+  std::vector<unsigned char> slice, row;
+  std::vector<float> u, v, dxO, duO, dvO, dxP, duP, dvP, corrXO, corrXP, nomUO, nomVO, nomUP, nomVP;
+  slice.reserve(npoints);
+  row.reserve(npoints);
+  u.reserve(npoints);
+  v.reserve(npoints);
+  dxO.resize(npoints);
+  duO.resize(npoints);
+  dvO.resize(npoints);
+  corrXO.resize(npoints);
+  nomUO.resize(npoints);
+  nomVO.resize(npoints);
+  dxP.resize(npoints);
+  duP.resize(npoints);
+  dvP.resize(npoints);
+  corrXP.resize(npoints);
+  nomUP.resize(npoints);
+  nomVP.resize(npoints);
+
+  for (int i = 0; i < npoints; i++) {
+    slice.push_back(gRandom->Integer(NSLICES));
+    row.push_back(gRandom->Integer(NROWS));
+    u.push_back(gRandom->Rndm() * 15);
+    v.push_back(gRandom->Rndm() * 200);
+  }
+  const auto& origCorr = src.getCorrection();
+  long origStart[3], origEnd[3], thisStart[3], thisEnd[3];
+  origStart[0] = std::chrono::time_point_cast<std::chrono::microseconds>(std::chrono::system_clock::now()).time_since_epoch().count();
+  for (int i = 0; i < npoints; i++) {
+    origCorr.getCorrection(slice[i], row[i], u[i], v[i], dxO[i], duO[i], dvO[i]);
+  }
+  origEnd[0] = origStart[1] = std::chrono::time_point_cast<std::chrono::microseconds>(std::chrono::system_clock::now()).time_since_epoch().count();
+  for (int i = 0; i < npoints; i++) {
+    origCorr.getCorrectionInvCorrectedX(slice[i], row[i], u[i], v[i], corrXO[i]);
+  }
+  origEnd[1] = origStart[2] = std::chrono::time_point_cast<std::chrono::microseconds>(std::chrono::system_clock::now()).time_since_epoch().count();
+  for (int i = 0; i < npoints; i++) {
+    origCorr.getCorrectionInvUV(slice[i], row[i], u[i], v[i], nomUO[i], nomVO[i]);
+  }
+  origEnd[2] = thisStart[0] = std::chrono::time_point_cast<std::chrono::microseconds>(std::chrono::system_clock::now()).time_since_epoch().count();
+  for (int i = 0; i < npoints; i++) {
+    this->getCorrection(slice[i], row[i], u[i], v[i], dxP[i], duP[i], dvP[i]);
+  }
+  thisEnd[0] = thisStart[1] = std::chrono::time_point_cast<std::chrono::microseconds>(std::chrono::system_clock::now()).time_since_epoch().count();
+  for (int i = 0; i < npoints; i++) {
+    this->getCorrectionInvCorrectedX(slice[i], row[i], u[i], v[i], corrXP[i]);
+  }
+  thisEnd[1] = thisStart[2] = std::chrono::time_point_cast<std::chrono::microseconds>(std::chrono::system_clock::now()).time_since_epoch().count();
+  for (int i = 0; i < npoints; i++) {
+    this->getCorrectionInvUV(slice[i], row[i], u[i], v[i], nomUP[i], nomVP[i]);
+  }
+  thisEnd[2] = std::chrono::time_point_cast<std::chrono::microseconds>(std::chrono::system_clock::now()).time_since_epoch().count();
+  //
+  size_t ndiff[3] = {};
+  for (int i = 0; i < npoints; i++) {
+    if (dxO[i] != dxP[i] || duO[i] != duP[i] || dvO[i] != dvP[i]) {
+      ndiff[0]++;
+    }
+    if (corrXO[i] != corrXP[i]) {
+      ndiff[1]++;
+    }
+    if (nomUO[i] != nomUP[i] || nomVO[i] != nomVP[i]) {
+      ndiff[2]++;
+    }
+  }
+  //
+  LOGP(info, " (ns per call)              original       this      Nmissmatch");
+  LOGP(info, "getCorrection               {:.3e}    {:.3e}   {}", double(origEnd[0] - origStart[0]) / npoints, double(thisEnd[0] - thisStart[0]) / npoints, ndiff[0]);
+  LOGP(info, "getCorrectionInvCorrectedX  {:.3e}    {:.3e}   {}", double(origEnd[1] - origStart[1]) / npoints, double(thisEnd[1] - thisStart[1]) / npoints, ndiff[1]);
+  LOGP(info, "getCorrectionInvUV          {:.3e}    {:.3e}   {}", double(origEnd[2] - origStart[2]) / npoints, double(thisEnd[2] - thisStart[2]) / npoints, ndiff[2]);
+  return ndiff[0] == 0 && ndiff[1] == 0 && ndiff[2] == 0;
+}
+
+#endif
+
+} // namespace gpu
+} // namespace GPUCA_NAMESPACE