clr/hipamd/include/hip/amd_detail/amd_hip_cooperative_groups.h Source File

clr/hipamd/include/hip/amd_detail/amd_hip_cooperative_groups.h Source File#

HIP Runtime API Reference: clr/hipamd/include/hip/amd_detail/amd_hip_cooperative_groups.h Source File
/*
 * Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 *
 * SPDX-License-Identifier: MIT
 */
 
#ifndef HIP_INCLUDE_HIP_AMD_DETAIL_HIP_COOPERATIVE_GROUPS_H
#define HIP_INCLUDE_HIP_AMD_DETAIL_HIP_COOPERATIVE_GROUPS_H
 
#if __cplusplus
#if !defined(__HIPCC_RTC__)
#include <hip/amd_detail/hip_cooperative_groups_helper.h>
#endif
 
namespace cooperative_groups {
 
class thread_group {
 protected:
  __hip_uint32_t _type;         
  __hip_uint32_t _num_threads;  
  __hip_uint64_t _mask;         
 
  __CG_QUALIFIER__ thread_group(internal::group_type type,
                                __hip_uint32_t num_threads = static_cast<__hip_uint64_t>(0),
                                __hip_uint64_t mask = static_cast<__hip_uint64_t>(0)) {
    _type = type;
    _num_threads = num_threads;
    _mask = mask;
  }
 
  struct _tiled_info {
    bool is_tiled;
    unsigned int num_threads;
    unsigned int meta_group_rank;
    unsigned int meta_group_size;
  };
 
  struct _coalesced_info {
    lane_mask member_mask;
    unsigned int num_threads;
    struct _tiled_info tiled_info;
  } coalesced_info;
 
  friend __CG_QUALIFIER__ thread_group this_thread();
  friend __CG_QUALIFIER__ thread_group tiled_partition(const thread_group& parent,
                                                       unsigned int tile_size);
  friend class thread_block;
 
 public:
  __CG_QUALIFIER__ __hip_uint32_t num_threads() const { return _num_threads; }
  __CG_QUALIFIER__ __hip_uint32_t size() const { return num_threads(); }
  __CG_QUALIFIER__ unsigned int cg_type() const { return _type; }
  __CG_QUALIFIER__ __hip_uint32_t thread_rank() const;
  __CG_QUALIFIER__ __hip_uint32_t block_rank() const;
  __CG_QUALIFIER__ bool is_valid() const;
 
  __CG_QUALIFIER__ void sync() const;
};
class multi_grid_group : public thread_group {
  friend __CG_QUALIFIER__ multi_grid_group this_multi_grid();
 
 protected:
  explicit __CG_QUALIFIER__ multi_grid_group(__hip_uint32_t size)
      : thread_group(internal::cg_multi_grid, size) {}
 
 public:
  __CG_QUALIFIER__ __hip_uint32_t num_grids() { return internal::multi_grid::num_grids(); }
 
  __CG_QUALIFIER__ __hip_uint32_t grid_rank() { return internal::multi_grid::grid_rank(); }
  __CG_QUALIFIER__ __hip_uint32_t thread_rank() const {
    return internal::multi_grid::thread_rank();
  }
  __CG_QUALIFIER__ bool is_valid() const { return internal::multi_grid::is_valid(); }
  __CG_QUALIFIER__ void sync() const { internal::multi_grid::sync(); }
};
 
__CG_QUALIFIER__ multi_grid_group this_multi_grid() {
  return multi_grid_group(internal::multi_grid::num_threads());
}
// Doxygen end group CooperativeGConstruct
class grid_group : public thread_group {
  friend __CG_QUALIFIER__ grid_group this_grid();
 
 protected:
  explicit __CG_QUALIFIER__ grid_group(__hip_uint32_t size)
      : thread_group(internal::cg_grid, size) {}
 
 public:
  __CG_QUALIFIER__ __hip_uint32_t thread_rank() const { return internal::grid::thread_rank(); }
  __CG_QUALIFIER__ __hip_uint32_t block_rank() const { return internal::grid::block_rank(); }
  __CG_QUALIFIER__ bool is_valid() const { return internal::grid::is_valid(); }
  __CG_QUALIFIER__ void sync() const { internal::grid::sync(); }
  __CG_QUALIFIER__ dim3 group_dim() const { return internal::grid::grid_dim(); }
  struct arrival_token {
    unsigned int signal;
  };
  __CG_QUALIFIER__ arrival_token barrier_arrive() const {
    arrival_token t;
    t.signal = internal::grid::barrier_signal();
    return t;
  }
  __CG_QUALIFIER__ void barrier_wait(arrival_token&& t) const {
    internal::grid::barrier_wait(t.signal);
  }
};
 
__CG_QUALIFIER__ grid_group this_grid() { return grid_group(internal::grid::num_threads()); }
 
class thread_block : public thread_group {
  friend __CG_QUALIFIER__ thread_block this_thread_block();
  friend __CG_QUALIFIER__ thread_group tiled_partition(const thread_group& parent,
                                                       unsigned int tile_size);
  friend __CG_QUALIFIER__ thread_group tiled_partition(const thread_block& parent,
                                                       unsigned int tile_size);
 
 protected:
  // Construct a workgroup thread group (through the API this_thread_block())
  explicit __CG_QUALIFIER__ thread_block(__hip_uint32_t size)
      : thread_group(internal::cg_workgroup, size) {}
 
  __CG_QUALIFIER__ thread_group new_tiled_group(unsigned int tile_size) const {
    const bool pow2 = ((tile_size & (tile_size - 1)) == 0);
    // Invalid tile size, assert
    if (!tile_size || (tile_size > warpSize) || !pow2) {
      __hip_assert(false && "invalid tile size");
    }
 
    auto block_size = num_threads();
    auto rank = thread_rank();
    auto partitions = (block_size + tile_size - 1) / tile_size;
    auto tail = (partitions * tile_size) - block_size;
    auto partition_size = tile_size - tail * (rank >= (partitions - 1) * tile_size);
    thread_group tiledGroup = thread_group(internal::cg_tiled_group, partition_size);
 
    tiledGroup.coalesced_info.tiled_info.num_threads = tile_size;
    tiledGroup.coalesced_info.tiled_info.is_tiled = true;
    tiledGroup.coalesced_info.tiled_info.meta_group_rank = rank / tile_size;
    tiledGroup.coalesced_info.tiled_info.meta_group_size = partitions;
    return tiledGroup;
  }
 
 public:
  __CG_STATIC_QUALIFIER__ dim3 group_index() { return internal::workgroup::group_index(); }
  __CG_STATIC_QUALIFIER__ dim3 thread_index() { return internal::workgroup::thread_index(); }
  __CG_STATIC_QUALIFIER__ __hip_uint32_t thread_rank() {
    return internal::workgroup::thread_rank();
  }
  __CG_STATIC_QUALIFIER__ __hip_uint32_t block_rank() {
    return internal::workgroup::block_rank();
  }
  __CG_STATIC_QUALIFIER__ __hip_uint32_t num_threads() {
    return internal::workgroup::num_threads();
  }
  __CG_STATIC_QUALIFIER__ __hip_uint32_t size() { return num_threads(); }
  __CG_STATIC_QUALIFIER__ bool is_valid() { return internal::workgroup::is_valid(); }
  __CG_STATIC_QUALIFIER__ void sync() { internal::workgroup::sync(); }
  __CG_QUALIFIER__ dim3 group_dim() { return internal::workgroup::block_dim(); }
  struct arrival_token {};
  __CG_QUALIFIER__ arrival_token barrier_arrive() const {
    internal::workgroup::barrier_arrive();
    return arrival_token{};
  }
  __CG_QUALIFIER__ void barrier_wait(arrival_token&&) const { internal::workgroup::barrier_wait(); }
};
 
__CG_QUALIFIER__ thread_block this_thread_block() {
  return thread_block(internal::workgroup::num_threads());
}
 
class tiled_group : public thread_group {
 private:
  friend __CG_QUALIFIER__ thread_group tiled_partition(const thread_group& parent,
                                                       unsigned int tile_size);
  friend __CG_QUALIFIER__ tiled_group tiled_partition(const tiled_group& parent,
                                                      unsigned int tile_size);
 
  __CG_QUALIFIER__ tiled_group new_tiled_group(unsigned int tile_size) const {
    const bool pow2 = ((tile_size & (tile_size - 1)) == 0);
 
    if (!tile_size || (tile_size > warpSize) || !pow2) {
      __hip_assert(false && "invalid tile size");
    }
 
    if (num_threads() <= tile_size) {
      return *this;
    }
 
    tiled_group tiledGroup = tiled_group(tile_size);
    tiledGroup.coalesced_info.tiled_info.is_tiled = true;
    return tiledGroup;
  }
 
 protected:
  explicit __CG_QUALIFIER__ tiled_group(unsigned int tileSize)
      : thread_group(internal::cg_tiled_group, tileSize) {
    coalesced_info.tiled_info.num_threads = tileSize;
    coalesced_info.tiled_info.is_tiled = true;
  }
 
 public:
  __CG_QUALIFIER__ unsigned int num_threads() const {
    return (coalesced_info.tiled_info.num_threads);
  }
 
  __CG_QUALIFIER__ unsigned int size() const { return num_threads(); }
 
  __CG_QUALIFIER__ unsigned int thread_rank() const {
    return (internal::workgroup::thread_rank() & (coalesced_info.tiled_info.num_threads - 1));
  }
  __CG_QUALIFIER__ void sync() const { internal::tiled_group::sync(); }
};
 
template <unsigned int size, class ParentCGTy> class thread_block_tile;
 
class coalesced_group : public thread_group {
 private:
  friend __CG_QUALIFIER__ coalesced_group coalesced_threads();
  friend __CG_QUALIFIER__ thread_group tiled_partition(const thread_group& parent,
                                                       unsigned int tile_size);
  friend __CG_QUALIFIER__ coalesced_group tiled_partition(const coalesced_group& parent,
                                                          unsigned int tile_size);
  friend __CG_QUALIFIER__ coalesced_group binary_partition(const coalesced_group& cgrp, bool pred);
  template <unsigned int fsize, class fparent> friend __CG_QUALIFIER__ coalesced_group
  binary_partition(const thread_block_tile<fsize, fparent>& tgrp, bool pred);
 
  __CG_QUALIFIER__ coalesced_group new_tiled_group(unsigned int tile_size) const {
    const bool pow2 = ((tile_size & (tile_size - 1)) == 0);
 
    if (!tile_size || !pow2) {
      return coalesced_group(0);
    }
 
    // If a tiled group is passed to be partitioned further into a coalesced_group.
    // prepare a mask for further partitioning it so that it stays coalesced.
    if (coalesced_info.tiled_info.is_tiled) {
      unsigned int base_offset = (thread_rank() & (~(tile_size - 1)));
      unsigned int masklength =
          min(static_cast<unsigned int>(num_threads()) - base_offset, tile_size);
      lane_mask full_mask = (static_cast<int>(warpSize) == 32)
                                ? static_cast<lane_mask>((1u << 32) - 1)
                                : static_cast<lane_mask>(-1ull);
      lane_mask member_mask = full_mask >> (warpSize - masklength);
 
      member_mask <<= (__lane_id() & ~(tile_size - 1));
      coalesced_group coalesced_tile = coalesced_group(member_mask);
      coalesced_tile.coalesced_info.tiled_info.is_tiled = true;
      coalesced_tile.coalesced_info.tiled_info.meta_group_rank = thread_rank() / tile_size;
      coalesced_tile.coalesced_info.tiled_info.meta_group_size = num_threads() / tile_size;
      return coalesced_tile;
    }
    // Here the parent coalesced_group is not partitioned.
    else {
      lane_mask member_mask = 0;
      unsigned int tile_rank = 0;
      int lanes_to_skip = ((thread_rank()) / tile_size) * tile_size;
 
      for (unsigned int i = 0; i < warpSize; i++) {
        lane_mask active = coalesced_info.member_mask & (static_cast<lane_mask>(1) << i);
        // Make sure the lane is active
        if (active) {
          if (lanes_to_skip <= 0 && tile_rank < tile_size) {
            // Prepare a member_mask that is appropriate for a tile
            member_mask |= active;
            tile_rank++;
          }
          lanes_to_skip--;
        }
      }
      coalesced_group coalesced_tile = coalesced_group(member_mask);
      coalesced_tile.coalesced_info.tiled_info.meta_group_rank = thread_rank() / tile_size;
      coalesced_tile.coalesced_info.tiled_info.meta_group_size =
          (num_threads() + tile_size - 1) / tile_size;
      return coalesced_tile;
    }
    return coalesced_group(0);
  }
 
 protected:
  // Constructor
  explicit __CG_QUALIFIER__ coalesced_group(lane_mask member_mask)
      : thread_group(internal::cg_coalesced_group) {
    coalesced_info.member_mask = member_mask;  // Which threads are active
    coalesced_info.num_threads =
        __popcll(coalesced_info.member_mask);    // How many threads are active
    coalesced_info.tiled_info.is_tiled = false;  // Not a partitioned group
    coalesced_info.tiled_info.meta_group_rank = 0;
    coalesced_info.tiled_info.meta_group_size = 1;
  }
 
 public:
  __CG_QUALIFIER__ unsigned int num_threads() const { return coalesced_info.num_threads; }
 
  __CG_QUALIFIER__ unsigned int size() const { return num_threads(); }
 
  __CG_QUALIFIER__ unsigned int thread_rank() const {
    return internal::coalesced_group::masked_bit_count(coalesced_info.member_mask);
  }
 
  __CG_QUALIFIER__ void sync() const { internal::coalesced_group::sync(); }
 
  __CG_QUALIFIER__ unsigned int meta_group_rank() const {
    return coalesced_info.tiled_info.meta_group_rank;
  }
 
  __CG_QUALIFIER__ unsigned int meta_group_size() const {
    return coalesced_info.tiled_info.meta_group_size;
  }
 
  template <class T> __CG_QUALIFIER__ T shfl(T var, int srcRank) const {
    srcRank = srcRank % static_cast<int>(num_threads());
 
    int lane = (num_threads() == warpSize) ? srcRank
               : (static_cast<int>(warpSize) == 64)
                   ? __fns64(coalesced_info.member_mask, 0, (srcRank + 1))
                   : __fns32(coalesced_info.member_mask, 0, (srcRank + 1));
 
    return __shfl(var, lane, warpSize);
  }
 
  template <class T> __CG_QUALIFIER__ T shfl_down(T var, unsigned int lane_delta) const {
    // Note: The cuda implementation appears to use the remainder of lane_delta
    // and WARP_SIZE as the shift value rather than lane_delta itself.
    // This is not described in the documentation and is not done here.
 
    if (num_threads() == warpSize) {
      return __shfl_down(var, lane_delta, warpSize);
    }
 
    int lane;
    if (static_cast<int>(warpSize) == 64) {
      lane = __fns64(coalesced_info.member_mask, __lane_id(), lane_delta + 1);
    } else {
      lane = __fns32(coalesced_info.member_mask, __lane_id(), lane_delta + 1);
    }
 
    if (lane == -1) {
      lane = __lane_id();
    }
 
    return __shfl(var, lane, warpSize);
  }
 
  template <class T> __CG_QUALIFIER__ T shfl_up(T var, unsigned int lane_delta) const {
    // Note: The cuda implementation appears to use the remainder of lane_delta
    // and WARP_SIZE as the shift value rather than lane_delta itself.
    // This is not described in the documentation and is not done here.
 
    if (num_threads() == warpSize) {
      return __shfl_up(var, lane_delta, warpSize);
    }
 
    int lane;
    if (static_cast<int>(warpSize) == 64) {
      lane = __fns64(coalesced_info.member_mask, __lane_id(), -(lane_delta + 1));
    } else if (static_cast<int>(warpSize) == 32) {
      lane = __fns32(coalesced_info.member_mask, __lane_id(), -(lane_delta + 1));
    }
 
    if (lane == -1) {
      lane = __lane_id();
    }
 
    return __shfl(var, lane, warpSize);
  }
#if !defined(HIP_DISABLE_WARP_SYNC_BUILTINS)
 
  __CG_QUALIFIER__ unsigned long long ballot(int pred) const {
    return internal::helper::adjust_mask(
        coalesced_info.member_mask,
        __ballot_sync<unsigned long long>(coalesced_info.member_mask, pred));
  }
 
  __CG_QUALIFIER__ int any(int pred) const {
    return __any_sync(static_cast<unsigned long long>(coalesced_info.member_mask), pred);
  }
 
  __CG_QUALIFIER__ int all(int pred) const {
    return __all_sync(static_cast<unsigned long long>(coalesced_info.member_mask), pred);
  }
 
  template <typename T> __CG_QUALIFIER__ unsigned long long match_any(T value) const {
    return internal::helper::adjust_mask(
        coalesced_info.member_mask,
        __match_any_sync(static_cast<unsigned long long>(coalesced_info.member_mask), value));
  }
 
  template <typename T> __CG_QUALIFIER__ unsigned long long match_all(T value, int& pred) const {
    return internal::helper::adjust_mask(
        coalesced_info.member_mask,
        __match_all_sync(static_cast<unsigned long long>(coalesced_info.member_mask), value,
                         &pred));
  }
#endif  // HIP_DISABLE_WARP_SYNC_BUILTINS
};
 
__CG_QUALIFIER__ coalesced_group coalesced_threads() {
  return cooperative_groups::coalesced_group(__builtin_amdgcn_read_exec());
}
 
#ifndef DOXYGEN_SHOULD_SKIP_THIS
 
__CG_QUALIFIER__ __hip_uint32_t thread_group::thread_rank() const {
  switch (this->_type) {
    case internal::cg_multi_grid: {
      return (static_cast<const multi_grid_group*>(this)->thread_rank());
    }
    case internal::cg_grid: {
      return (static_cast<const grid_group*>(this)->thread_rank());
    }
    case internal::cg_workgroup: {
      return (static_cast<const thread_block*>(this)->thread_rank());
    }
    case internal::cg_tiled_group: {
      return (static_cast<const tiled_group*>(this)->thread_rank());
    }
    case internal::cg_coalesced_group: {
      return (static_cast<const coalesced_group*>(this)->thread_rank());
    }
    default: {
      __hip_assert(false && "invalid cooperative group type");
      return -1;
    }
  }
}
 
__CG_QUALIFIER__ bool thread_group::is_valid() const {
  switch (this->_type) {
    case internal::cg_multi_grid: {
      return (static_cast<const multi_grid_group*>(this)->is_valid());
    }
    case internal::cg_grid: {
      return (static_cast<const grid_group*>(this)->is_valid());
    }
    case internal::cg_workgroup: {
      return (static_cast<const thread_block*>(this)->is_valid());
    }
    case internal::cg_tiled_group: {
      return (static_cast<const tiled_group*>(this)->is_valid());
    }
    case internal::cg_coalesced_group: {
      return (static_cast<const coalesced_group*>(this)->is_valid());
    }
    default: {
      __hip_assert(false && "invalid cooperative group type");
      return false;
    }
  }
}
 
__CG_QUALIFIER__ void thread_group::sync() const {
  switch (this->_type) {
    case internal::cg_multi_grid: {
      static_cast<const multi_grid_group*>(this)->sync();
      break;
    }
    case internal::cg_grid: {
      static_cast<const grid_group*>(this)->sync();
      break;
    }
    case internal::cg_workgroup: {
      static_cast<const thread_block*>(this)->sync();
      break;
    }
    case internal::cg_tiled_group: {
      static_cast<const tiled_group*>(this)->sync();
      break;
    }
    case internal::cg_coalesced_group: {
      static_cast<const coalesced_group*>(this)->sync();
      break;
    }
    default: {
      __hip_assert(false && "invalid cooperative group type");
    }
  }
#if __has_builtin(__builtin_amdgcn_s_wait_asynccnt)
  __builtin_amdgcn_s_wait_asynccnt(0);
#endif
}
 
#endif
 
template <class CGTy> __CG_QUALIFIER__ __hip_uint32_t group_size(CGTy const& g) {
  return g.num_threads();
}
 
template <class CGTy> __CG_QUALIFIER__ __hip_uint32_t thread_rank(CGTy const& g) {
  return g.thread_rank();
}
 
template <class CGTy> __CG_QUALIFIER__ bool is_valid(CGTy const& g) { return g.is_valid(); }
 
template <class CGTy> __CG_QUALIFIER__ void sync(CGTy const& g) { g.sync(); }
 
// Doxygen end group CooperativeGAPI
template <unsigned int tileSize> class tile_base {
 protected:
  _CG_STATIC_CONST_DECL_ unsigned int numThreads = tileSize;
 
 public:
  _CG_STATIC_CONST_DECL_ unsigned int thread_rank() {
    return (internal::workgroup::thread_rank() & (numThreads - 1));
  }
 
  __CG_STATIC_QUALIFIER__ unsigned int num_threads() { return numThreads; }
 
  __CG_STATIC_QUALIFIER__ unsigned int size() { return num_threads(); }
};
 
template <unsigned int size> class thread_block_tile_base : public tile_base<size> {
  static_assert(is_valid_tile_size<size>::value,
                "Tile size is either not a power of 2 or greater than the wavefront size");
  using tile_base<size>::numThreads;
 
  template <unsigned int fsize, class fparent> friend __CG_QUALIFIER__ coalesced_group
  binary_partition(const thread_block_tile<fsize, fparent>& tgrp, bool pred);
 
#if !defined(HIP_DISABLE_WARP_SYNC_BUILTINS)
  __CG_QUALIFIER__ unsigned long long build_mask() const {
    unsigned long long mask = ~0ull >> (64 - numThreads);
    // thread_rank() gives thread id from 0..thread launch size.
    return mask << (((internal::workgroup::thread_rank() % warpSize) / numThreads) * numThreads);
  }
#endif  // HIP_DISABLE_WARP_SYNC_BUILTINS
 
 public:
  __CG_STATIC_QUALIFIER__ void sync() { internal::tiled_group::sync(); }
 
  template <class T> __CG_QUALIFIER__ T shfl(T var, int srcRank) const {
    return (__shfl(var, srcRank, numThreads));
  }
 
  template <class T> __CG_QUALIFIER__ T shfl_down(T var, unsigned int lane_delta) const {
    return (__shfl_down(var, lane_delta, numThreads));
  }
 
  template <class T> __CG_QUALIFIER__ T shfl_up(T var, unsigned int lane_delta) const {
    return (__shfl_up(var, lane_delta, numThreads));
  }
 
  template <class T> __CG_QUALIFIER__ T shfl_xor(T var, unsigned int laneMask) const {
    return (__shfl_xor(var, laneMask, numThreads));
  }
 
#if !defined(HIP_DISABLE_WARP_SYNC_BUILTINS)
  __CG_QUALIFIER__ unsigned long long ballot(int pred) const {
    const auto mask = build_mask();
    return internal::helper::adjust_mask(mask, __ballot_sync(mask, pred));
  }
 
  __CG_QUALIFIER__ int any(int pred) const { return __any_sync(build_mask(), pred); }
 
  __CG_QUALIFIER__ int all(int pred) const { return __all_sync(build_mask(), pred); }
 
  template <typename T> __CG_QUALIFIER__ unsigned long long match_any(T value) const {
    const auto mask = build_mask();
    return internal::helper::adjust_mask(mask, __match_any_sync(mask, value));
  }
 
  template <typename T> __CG_QUALIFIER__ unsigned long long match_all(T value, int& pred) const {
    const auto mask = build_mask();
    return internal::helper::adjust_mask(mask, __match_all_sync(mask, value, &pred));
  }
#endif  // HIP_DISABLE_WARP_SYNC_BUILTINS
};
 
template <unsigned int tileSize, typename ParentCGTy> class parent_group_info {
 public:
  __CG_STATIC_QUALIFIER__ unsigned int meta_group_rank() {
    return ParentCGTy::thread_rank() / tileSize;
  }
 
  __CG_STATIC_QUALIFIER__ unsigned int meta_group_size() {
    return (ParentCGTy::num_threads() + tileSize - 1) / tileSize;
  }
};
 
template <unsigned int tileSize, class ParentCGTy> class thread_block_tile_type
    : public thread_block_tile_base<tileSize>,
      public tiled_group,
      public parent_group_info<tileSize, ParentCGTy> {
  _CG_STATIC_CONST_DECL_ unsigned int numThreads = tileSize;
  typedef thread_block_tile_base<numThreads> tbtBase;
 
 protected:
  __CG_QUALIFIER__ thread_block_tile_type() : tiled_group(numThreads) {
    coalesced_info.tiled_info.num_threads = numThreads;
    coalesced_info.tiled_info.is_tiled = true;
  }
 
  __CG_QUALIFIER__ thread_block_tile_type(unsigned int meta_group_rank,
                                          unsigned int meta_group_size)
      : tiled_group(numThreads) {
    coalesced_info.tiled_info.num_threads = numThreads;
    coalesced_info.tiled_info.is_tiled = true;
    coalesced_info.tiled_info.meta_group_rank = meta_group_rank;
    coalesced_info.tiled_info.meta_group_size = meta_group_size;
  }
 
 public:
  using tbtBase::num_threads;
  using tbtBase::size;
  using tbtBase::sync;
  using tbtBase::thread_rank;
};
 
// Partial template specialization
template <unsigned int tileSize> class thread_block_tile_type<tileSize, void>
    : public thread_block_tile_base<tileSize>, public tiled_group {
  _CG_STATIC_CONST_DECL_ unsigned int numThreads = tileSize;
 
  typedef thread_block_tile_base<numThreads> tbtBase;
 
 protected:
  __CG_QUALIFIER__ thread_block_tile_type(unsigned int meta_group_rank,
                                          unsigned int meta_group_size)
      : tiled_group(numThreads) {
    coalesced_info.tiled_info.num_threads = numThreads;
    coalesced_info.tiled_info.is_tiled = true;
    coalesced_info.tiled_info.meta_group_rank = meta_group_rank;
    coalesced_info.tiled_info.meta_group_size = meta_group_size;
  }
 
 public:
  using tbtBase::num_threads;
  using tbtBase::size;
  using tbtBase::sync;
  using tbtBase::thread_rank;
 
  __CG_QUALIFIER__ unsigned int meta_group_rank() const {
    return coalesced_info.tiled_info.meta_group_rank;
  }
 
  __CG_QUALIFIER__ unsigned int meta_group_size() const {
    return coalesced_info.tiled_info.meta_group_size;
  }
  // Doxygen end group CooperativeG
};
 
__CG_QUALIFIER__ thread_group this_thread() {
  thread_group g(internal::group_type::cg_coalesced_group, 1, __ockl_activelane_u32());
  return g;
}
 
__CG_QUALIFIER__ thread_group tiled_partition(const thread_group& parent, unsigned int tile_size) {
  if (parent.cg_type() == internal::cg_tiled_group) {
    const tiled_group* cg = static_cast<const tiled_group*>(&parent);
    return cg->new_tiled_group(tile_size);
  } else if (parent.cg_type() == internal::cg_coalesced_group) {
    const coalesced_group* cg = static_cast<const coalesced_group*>(&parent);
    return cg->new_tiled_group(tile_size);
  } else {
    const thread_block* tb = static_cast<const thread_block*>(&parent);
    return tb->new_tiled_group(tile_size);
  }
}
 
// Thread block type overload
__CG_QUALIFIER__ thread_group tiled_partition(const thread_block& parent, unsigned int tile_size) {
  return (parent.new_tiled_group(tile_size));
}
 
__CG_QUALIFIER__ tiled_group tiled_partition(const tiled_group& parent, unsigned int tile_size) {
  return (parent.new_tiled_group(tile_size));
}
 
// If a coalesced group is passed to be partitioned, it should remain coalesced
__CG_QUALIFIER__ coalesced_group tiled_partition(const coalesced_group& parent,
                                                 unsigned int tile_size) {
  return (parent.new_tiled_group(tile_size));
}
 
namespace impl {
template <unsigned int size, class ParentCGTy> class thread_block_tile_internal;
 
template <unsigned int size, class ParentCGTy> class thread_block_tile_internal
    : public thread_block_tile_type<size, ParentCGTy> {
 protected:
  template <unsigned int tbtSize, class tbtParentT> __CG_QUALIFIER__ thread_block_tile_internal(
      const thread_block_tile_internal<tbtSize, tbtParentT>& g)
      : thread_block_tile_type<size, ParentCGTy>(g.meta_group_rank(), g.meta_group_size()) {}
 
  __CG_QUALIFIER__ thread_block_tile_internal(const thread_block& g)
      : thread_block_tile_type<size, ParentCGTy>() {}
};
}  // namespace impl
 
template <unsigned int size, class ParentCGTy> class thread_block_tile
    : public impl::thread_block_tile_internal<size, ParentCGTy> {
 protected:
  __CG_QUALIFIER__ thread_block_tile(const ParentCGTy& g)
      : impl::thread_block_tile_internal<size, ParentCGTy>(g) {}
 
 public:
  __CG_QUALIFIER__ operator thread_block_tile<size, void>() const {
    return thread_block_tile<size, void>(*this);
  }
 
#ifdef DOXYGEN_SHOULD_INCLUDE_THIS
 
  __CG_QUALIFIER__ unsigned int thread_rank() const;
 
  __CG_QUALIFIER__ void sync();
 
  __CG_QUALIFIER__ unsigned int meta_group_rank() const;
 
  __CG_QUALIFIER__ unsigned int meta_group_size() const;
 
  template <class T> __CG_QUALIFIER__ T shfl(T var, int srcRank) const;
 
  template <class T> __CG_QUALIFIER__ T shfl_down(T var, unsigned int lane_delta) const;
 
  template <class T> __CG_QUALIFIER__ T shfl_up(T var, unsigned int lane_delta) const;
 
  template <class T> __CG_QUALIFIER__ T shfl_xor(T var, unsigned int laneMask) const;
 
  __CG_QUALIFIER__ unsigned long long ballot(int pred) const;
 
  __CG_QUALIFIER__ int any(int pred) const;
 
  __CG_QUALIFIER__ int all(int pred) const;
 
  template <typename T> __CG_QUALIFIER__ unsigned long long match_any(T value) const;
 
  template <typename T> __CG_QUALIFIER__ unsigned long long match_all(T value, int& pred) const;
 
#endif
};
 
template <unsigned int size> class thread_block_tile<size, void>
    : public impl::thread_block_tile_internal<size, void> {
  template <unsigned int, class ParentCGTy> friend class thread_block_tile;
 
 protected:
 public:
  template <class ParentCGTy>
  __CG_QUALIFIER__ thread_block_tile(const thread_block_tile<size, ParentCGTy>& g)
      : impl::thread_block_tile_internal<size, void>(g) {}
};
 
template <unsigned int size, class ParentCGTy = void> class thread_block_tile;
 
namespace impl {
template <unsigned int size, class ParentCGTy> struct tiled_partition_internal;
 
template <unsigned int size> struct tiled_partition_internal<size, thread_block>
    : public thread_block_tile<size, thread_block> {
  __CG_QUALIFIER__ tiled_partition_internal(const thread_block& g)
      : thread_block_tile<size, thread_block>(g) {}
};
 
// ParentCGTy = thread_block_tile<ParentSize, GrandParentCGTy> specialization
template <unsigned int size, unsigned int ParentSize, class GrandParentCGTy>
struct tiled_partition_internal<size, thread_block_tile<ParentSize, GrandParentCGTy> >
    : public thread_block_tile<size, thread_block_tile<ParentSize, GrandParentCGTy> > {
  static_assert(size < ParentSize, "Sub tile size must be < parent tile size in tiled_partition");
 
  __CG_QUALIFIER__ tiled_partition_internal(const thread_block_tile<ParentSize, GrandParentCGTy>& g)
      : thread_block_tile<size, thread_block_tile<ParentSize, GrandParentCGTy> >(g) {}
};
 
}  // namespace impl
 
template <unsigned int size, class ParentCGTy>
__CG_QUALIFIER__ thread_block_tile<size, ParentCGTy> tiled_partition(const ParentCGTy& g) {
  static_assert(is_valid_tile_size<size>::value,
                "Tiled partition with size > wavefront size. Currently not supported ");
  return impl::tiled_partition_internal<size, ParentCGTy>(g);
}
 
#if !defined(HIP_DISABLE_WARP_SYNC_BUILTINS)
 
__CG_QUALIFIER__ coalesced_group binary_partition(const coalesced_group& cgrp, bool pred) {
  auto mask = __ballot_sync<unsigned long long>(cgrp.coalesced_info.member_mask, pred);
 
  if (pred) {
    return coalesced_group(mask);
  } else {
    return coalesced_group(cgrp.coalesced_info.member_mask ^ mask);
  }
}
 
template <unsigned int size, class parent>
__CG_QUALIFIER__ coalesced_group binary_partition(const thread_block_tile<size, parent>& tgrp,
                                                  bool pred) {
  auto mask = __ballot_sync<unsigned long long>(tgrp.build_mask(), pred);
 
  if (pred) {
    return coalesced_group(mask);
  } else {
    return coalesced_group(tgrp.build_mask() ^ mask);
  }
}
 
template <class T>
struct plus {
  __CG_QUALIFIER__ T operator()(T lhs, T rhs) const
  {
    return lhs + rhs;
  }
};
 
template <class T>
struct less {
  __CG_QUALIFIER__ T operator()(T lhs, T rhs) const
  {
    return lhs < rhs? lhs : rhs;
  }
};
 
template <class T>
struct greater {
  __CG_QUALIFIER__ T operator()(T lhs, T rhs) const
  {
    return lhs < rhs? rhs : lhs;
  }
};
 
template <class T>
struct bit_and {
  __CG_QUALIFIER__ T operator()(T lhs, T rhs) const
  {
    return lhs & rhs;
  }
};
 
template <class T>
struct bit_xor {
  __CG_QUALIFIER__ T operator()(T lhs, T rhs) const
  {
    return lhs ^ rhs;
  }
};
 
template <class T>
struct bit_or {
  __CG_QUALIFIER__ T operator()(T lhs, T rhs) const
  {
    return lhs | rhs;
  }
};
#endif
 
class cluster_group {
  friend __device__ cluster_group this_cluster();
 
  // Default constructor, hidden
  __CG_QUALIFIER__ cluster_group() {}
 
 public:
  using arrival_token = struct {};
 
  // Sync the cluster, equivalent to c.barrier_wait(c.barrier_arrive());
  __CG_STATIC_QUALIFIER__ void sync() { internal::cluster::sync(); }
 
  // Arrive on a cluster barrier, returns token that needs to be passed to barrier_wait
  __CG_STATIC_QUALIFIER__ arrival_token barrier_arrive() {
    // signal user cluster barrier
    internal::cluster::barrier_arrive();
    return arrival_token();
  }
 
  // Wait on arrival_token
  __CG_STATIC_QUALIFIER__ void barrier_wait(arrival_token&&) { internal::cluster::barrier_wait(); }
 
  // TODO: implement this when compiler work is done
  // block rank to which shared memory address belongs to
  // __CG_STATIC_QUALIFIER__ unsigned int query_shared_rank(const void* addr) {}
  // Obtain the address of shared memory variable of another block in the cluster
  // template <typename T> __CG_STATIC_QUALIFIER__ T* map_shared_rank(T* addr, int rank) {}
 
  // index of the calling block within cluster
  __CG_STATIC_QUALIFIER__ dim3 block_index() { return internal::cluster::block_index(); }
 
  // Rank of calling block within [0, num_blocks)
  __CG_STATIC_QUALIFIER__ unsigned int block_rank() { return internal::cluster::block_rank(); }
 
  // index of the calling thread within cluster
  __CG_STATIC_QUALIFIER__ dim3 thread_index() { return internal::cluster::thread_index(); }
 
  // Rank of calling thread within [0, num_threads)
  __CG_STATIC_QUALIFIER__ unsigned int thread_rank() { return internal::cluster::thread_rank(); }
 
  // Dimensions of launched cluster in unit of blocks
  __CG_STATIC_QUALIFIER__ dim3 dim_blocks() { return internal::cluster::dim_blocks(); }
 
  // total number of blocks in the group
  __CG_STATIC_QUALIFIER__ unsigned int num_blocks() { return internal::cluster::num_blocks(); }
 
  // Dimensions of launched cluster in unit of threads
  __CG_STATIC_QUALIFIER__ dim3 dim_threads() { return internal::cluster::dim_threads(); }
 
  // Total number of threads in the group
  __CG_STATIC_QUALIFIER__ unsigned int num_threads() { return internal::cluster::num_threads(); }
 
  // Get address of shared memory variable in another cluster
  template <typename T> __CG_STATIC_QUALIFIER__ T* map_shared_rank(T* in, int rank) {
    return internal::cluster::map_shared_rank<T>(in, rank);
  }
 
  // Return block rank of shared memory address
  __CG_STATIC_QUALIFIER__ unsigned int query_shared_rank(const void* in) {
    return internal::cluster::query_shared_rank(in);
  }
 
  // Alias of num_threads
  __CG_STATIC_QUALIFIER__ unsigned int size() { return num_threads(); }
};
 
__device__ cluster_group this_cluster() {
  cluster_group cg;
  return cg;
}
}  // namespace cooperative_groups
 
#endif  // __cplusplus
#endif  // HIP_INCLUDE_HIP_AMD_DETAIL_HIP_COOPERATIVE_GROUPS_H