CfUtils.h

Go to the documentation of this file.

// 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.
 
 
#ifndef O2_GPU_CF_UTILS_H
#define O2_GPU_CF_UTILS_H
 
#include "clusterFinderDefs.h"
#include "GPUCommonAlgorithm.h"
#include "CfArray2D.h"
#include "CfConsts.h"
 
namespace o2::gpu
{
 
class CfUtils
{
 
 public:
  static GPUdi() bool innerAboveThreshold(uint8_t aboveThreshold, uint16_t outerIdx)
  {
    return aboveThreshold & (1 << cfconsts::OuterToInner[outerIdx]);
  }
 
  static GPUdi() bool innerAboveThresholdInv(uint8_t aboveThreshold, uint16_t outerIdx)
  {
    return aboveThreshold & (1 << cfconsts::OuterToInnerInv[outerIdx]);
  }
 
  static GPUdi() bool isPeak(uint8_t peak) { return peak & 0x01; }
 
  static GPUdi() bool isAboveThreshold(uint8_t peak) { return peak >> 1; }
 
  static GPUdi() int32_t warpPredicateScan(int32_t pred, int32_t* sum)
  {
#ifdef __HIPCC__
    int32_t iLane = hipThreadIdx_x % warpSize;
    uint64_t waveMask = __ballot(pred);
    uint64_t lowerWarpMask = (1ull << iLane) - 1ull;
    int32_t myOffset = __popcll(waveMask & lowerWarpMask);
    *sum = __popcll(waveMask);
    return myOffset;
#elif defined(__CUDACC__)
    int32_t iLane = threadIdx.x % warpSize;
    uint32_t waveMask = __ballot_sync(0xFFFFFFFF, pred);
    uint32_t lowerWarpMask = (1u << iLane) - 1u;
    int32_t myOffset = __popc(waveMask & lowerWarpMask);
    *sum = __popc(waveMask);
    return myOffset;
#else // CPU / OpenCL fallback
    int32_t myOffset = warp_scan_inclusive_add(pred ? 1 : 0);
    *sum = warp_broadcast(myOffset, GPUCA_WARP_SIZE - 1);
    myOffset--;
    return myOffset;
#endif
  }
 
  template <size_t BlockSize, typename SharedMemory>
  static GPUdi() int32_t blockPredicateScan(SharedMemory& smem, int32_t pred, int32_t* sum = nullptr)
  {
#if defined(__HIPCC__) || defined(__CUDACC__)
    int32_t iThread =
#ifdef __HIPCC__
      hipThreadIdx_x;
#else
      threadIdx.x;
#endif
 
    int32_t iWarp = iThread / warpSize;
    int32_t nWarps = BlockSize / warpSize;
 
    int32_t warpSum;
    int32_t laneOffset = warpPredicateScan(pred, &warpSum);
 
    if (iThread % warpSize == 0) {
      smem.warpPredicateSum[iWarp] = warpSum;
    }
    __syncthreads();
 
    int32_t warpOffset = 0;
 
    if (sum == nullptr) {
      for (int32_t w = 0; w < iWarp; w++) {
        int32_t s = smem.warpPredicateSum[w];
        warpOffset += s;
      }
    } else {
      *sum = 0;
      for (int32_t w = 0; w < nWarps; w++) {
        int32_t s = smem.warpPredicateSum[w];
        if (w < iWarp) {
          warpOffset += s;
        }
        *sum += s;
      }
    }
 
    return warpOffset + laneOffset;
#else // CPU / OpenCL fallback
    int32_t lpos = work_group_scan_inclusive_add(pred ? 1 : 0);
    if (sum != nullptr) {
      *sum = work_group_broadcast(lpos, BlockSize - 1);
    }
    lpos--;
    return lpos;
#endif
  }
 
  template <size_t BlockSize, typename SharedMemory>
  static GPUdi() int32_t blockPredicateSum(SharedMemory& smem, int32_t pred)
  {
#if defined(__HIPCC__) || defined(__CUDACC__)
    int32_t iThread =
#ifdef __HIPCC__
      hipThreadIdx_x;
#else
      threadIdx.x;
#endif
 
    int32_t iWarp = iThread / warpSize;
    int32_t nWarps = BlockSize / warpSize;
 
    int32_t warpSum =
#ifdef __HIPCC__
      __popcll(__ballot(pred));
#else
      __popc(__ballot_sync(0xFFFFFFFF, pred));
#endif
 
    if (iThread % warpSize == 0) {
      smem.warpPredicateSum[iWarp] = warpSum;
    }
    __syncthreads();
 
    int32_t sum = 0;
    for (int32_t w = 0; w < nWarps; w++) {
      sum += smem.warpPredicateSum[w];
    }
 
    return sum;
#else // CPU / OpenCL fallback
    return work_group_reduce_add(pred ? 1 : 0);
#endif
  }
 
  template <size_t SCRATCH_PAD_WORK_GROUP_SIZE, typename SharedMemory>
  static GPUdi() uint16_t partition(SharedMemory& smem, uint16_t ll, bool pred, uint16_t partSize, uint16_t* newPartSize)
  {
    bool participates = ll < partSize;
 
    int32_t part;
    int32_t lpos = blockPredicateScan<SCRATCH_PAD_WORK_GROUP_SIZE>(smem, int32_t(!pred && participates), &part);
 
    uint16_t pos = (participates && !pred) ? lpos : part;
 
    *newPartSize = part;
    return pos;
  }
 
  template <typename T>
  static GPUdi() void blockLoad(
    const CfArray2D<T>& map,
    uint32_t wgSize,
    uint32_t elems,
    uint16_t ll,
    uint32_t offset,
    uint32_t N,
    GPUconstexprref() const tpccf::Delta2* neighbors,
    const CfChargePos* posBcast,
    GPUgeneric() T* buf)
  {
#if defined(GPUCA_GPUCODE)
    GPUbarrier();
    uint16_t x = ll % N;
    uint16_t y = ll / N;
    tpccf::Delta2 d = neighbors[x + offset];
 
    for (uint32_t i = y; i < wgSize; i += (elems / N)) {
      CfChargePos readFrom = posBcast[i];
      uint32_t writeTo = N * i + x;
      buf[writeTo] = map[readFrom.delta(d)];
    }
    GPUbarrier();
#else
    if (ll >= wgSize) {
      return;
    }
 
    CfChargePos readFrom = posBcast[ll];
 
    GPUbarrier();
 
    for (uint32_t i = 0; i < N; i++) {
      tpccf::Delta2 d = neighbors[i + offset];
 
      uint32_t writeTo = N * ll + i;
      buf[writeTo] = map[readFrom.delta(d)];
    }
 
    GPUbarrier();
#endif
  }
 
  template <typename T, bool Inv = false>
  static GPUdi() void condBlockLoad(
    const CfArray2D<T>& map,
    uint16_t wgSize,
    uint16_t elems,
    uint16_t ll,
    uint16_t offset,
    uint16_t N,
    GPUconstexprref() const tpccf::Delta2* neighbors,
    const CfChargePos* posBcast,
    const uint8_t* aboveThreshold,
    GPUgeneric() T* buf)
  {
#if defined(GPUCA_GPUCODE)
    GPUbarrier();
    uint16_t y = ll / N;
    uint16_t x = ll % N;
    tpccf::Delta2 d = neighbors[x + offset];
    for (uint32_t i = y; i < wgSize; i += (elems / N)) {
      CfChargePos readFrom = posBcast[i];
      uint8_t above = aboveThreshold[i];
      uint32_t writeTo = N * i + x;
      T v(0);
      bool cond = (Inv) ? innerAboveThresholdInv(above, x + offset)
                        : innerAboveThreshold(above, x + offset);
      if (cond) {
        v = map[readFrom.delta(d)];
      }
      buf[writeTo] = v;
    }
    GPUbarrier();
#else
    if (ll >= wgSize) {
      return;
    }
 
    CfChargePos readFrom = posBcast[ll];
    uint8_t above = aboveThreshold[ll];
    GPUbarrier();
 
    for (uint32_t i = 0; i < N; i++) {
      tpccf::Delta2 d = neighbors[i + offset];
 
      uint32_t writeTo = N * ll + i;
      T v(0);
      bool cond = (Inv) ? innerAboveThresholdInv(above, i + offset)
                        : innerAboveThreshold(above, i + offset);
      if (cond) {
        v = map[readFrom.delta(d)];
      }
      buf[writeTo] = v;
    }
 
    GPUbarrier();
#endif
  }
};
 
} // namespace o2::gpu
 
#endif