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) 2015 - 2023 Advanced Micro Devices, Inc. All rights reserved.
  
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal
 in the Software without restriction, including without limitation the rights
 to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 copies of the Software, and to permit persons to whom the Software is
 furnished to do so, subject to the following conditions:
  
 The above copyright notice and this permission notice shall be included in
 all copies or substantial portions of the Software.
  
 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 THE SOFTWARE.
 */
  
 #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__ 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__ 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(); }
 };
  
 __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 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(); }
 };
  
 __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");
     }
   }
 }
  
 #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);
   }
 }
 #endif
 }  // namespace cooperative_groups
  
 #endif  // __cplusplus
 #endif  // HIP_INCLUDE_HIP_AMD_DETAIL_HIP_COOPERATIVE_GROUPS_H