/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/gemm/kernel/streamk_gemm_tile_partitioner_impl.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/gemm/kernel/streamk_gemm_tile_partitioner_impl.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/gemm/kernel/streamk_gemm_tile_partitioner_impl.hpp Source File
Go to the documentation of this file.
 // Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 // SPDX-License-Identifier: MIT
  
 #pragma once
 #include "streamk_gemm_tile_partitioner.hpp"
 namespace ck_tile {
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::StreamKTilePartitionerBase(
     index_t m, index_t n, index_t k, index_t grid)
     : grid_{grid}, n_{n}
 {
     iters_per_tile_ = integer_divide_ceil(k, KPerBlock);
     num_tiles_      = integer_divide_ceil(m, MPerBlock) * integer_divide_ceil(n_, NPerBlock);
  
     bool big_enough         = num_tiles_ > grid_;
     index_t remainder_tiles = num_tiles_ % grid_;
  
     if(remainder_tiles)
     {
         sk_tiles_       = big_enough ? full_tiles_ * grid_ + (num_tiles_ % grid_) : num_tiles_;
         sk_tiles_       = min(num_tiles_, sk_tiles_);
         sk_ctas_        = grid_;
         total_sk_iters_ = sk_tiles_ * iters_per_tile_;
  
         // If there still isn't enough work to saturate all CUs, then just revert to DP only.
         if(total_sk_iters_ < grid_)
         {
             sk_tiles_       = 0;
             sk_ctas_        = 0;
             total_sk_iters_ = 0;
         }
     }
     else // Full DP (i.e., no Stream-K)
     {
         sk_tiles_       = 0;
         sk_ctas_        = 0;
         total_sk_iters_ = 0;
     }
  
     iters_per_sk_cta_ = sk_ctas_ ? total_sk_iters_ / sk_ctas_ : 0;
     extra_iters_      = sk_ctas_ ? total_sk_iters_ % sk_ctas_ : 0;
  
     dp_tiles_       = num_tiles_ - sk_tiles_;
     total_dp_iters_ = dp_tiles_ * iters_per_tile_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_partials_buffer_size(
     index_t acc_element_bytes) const noexcept
 {
     return MPerBlock * NPerBlock * acc_element_bytes * sk_ctas_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_flags_buffer_size()
     const noexcept
 {
     return sizeof(index_t) * sk_ctas_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_DEVICE void
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_iter_boundaries(
     index_t& iter, index_t& iter_end, index_t cta_idx) const noexcept
 {
     index_t extra_iters_before_me = ck_tile::min(cta_idx, extra_iters_);
     iter     = total_dp_iters_ + cta_idx * iters_per_sk_cta_ + extra_iters_before_me;
     iter_end = iter + iters_per_sk_cta_ + (cta_idx < extra_iters_);
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_tile_index(
     index_t iter) const noexcept
 {
     return iter / iters_per_tile_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_DEVICE void
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_tile_boundaries(
     index_t& tile_iter, index_t& tile_iter_end, index_t tile_idx) const noexcept
 {
     tile_iter     = tile_idx * iters_per_tile_;
     tile_iter_end = tile_iter + iters_per_tile_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_DEVICE /* static */ index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_local_iter(
     index_t iter, index_t tile_iter) noexcept
 {
     return iter - tile_iter;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_DEVICE /* static */ index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_local_iter_end(
     index_t tile_iter, index_t iter_end, index_t tile_iter_end) noexcept
 {
     return ck_tile::min(iter_end, tile_iter_end) - tile_iter;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_DEVICE auto
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_output_tile_index(
     index_t tile_idx) const noexcept -> tuple<index_t, index_t>
 {
     const index_t n_macro_tiles = integer_divide_ceil(n_, NPerBlock);
  
     const index_t im = amd_wave_read_first_lane(tile_idx / n_macro_tiles);
     const index_t in = amd_wave_read_first_lane(tile_idx - im * n_macro_tiles);
     return make_tuple(im, in);
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_workspace_size(
     index_t acc_element_bytes) const noexcept
 {
     if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
     {
  
         return get_partials_buffer_size(acc_element_bytes) + get_flags_buffer_size();
     }
     else // ReductionStrategy is Atomics
     {
         return 0;
     }
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_num_tiles()
     const noexcept
 {
     return num_tiles_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_grid() const noexcept
 {
     return grid_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_dp_tiles() const noexcept
 {
     return dp_tiles_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_sk_tiles() const noexcept
 {
     return sk_tiles_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_sk_ctas() const noexcept
 {
     return sk_ctas_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_total_sk_iters()
     const noexcept
 {
     return total_sk_iters_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_iters_per_tile()
     const noexcept
 {
     return iters_per_tile_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_iters_per_sk_cta()
     const noexcept
 {
     return iters_per_sk_cta_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_extra_iters()
     const noexcept
 {
     return extra_iters_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_total_dp_iters()
     const noexcept
 {
     return total_dp_iters_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::get_n() const noexcept
 {
     return n_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST index_t
 StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::estimate_num_wgs_per_tile()
     const noexcept
 {
     // In the case of non-atomic reduction or data-parallel (DP) only, there will always be 1
     // workgroup writing final results to a given macro tile in C.
     int num_wgs_per_tile = 1;
  
     // Otherwise, for atomics, multiple workgroups may be writing to the same macro tile in C.
     if(sk_ctas_ > 0 && ReductionStrategy == ck_tile::StreamKReductionStrategy::Atomic)
     {
         //  If we have DP and SK tiles, this is DP+2TSK which guarantees at most 2 workgroups per
         //  tile. We only need to check that dp_tiles is greater than zero since we know we have SK
         //  workgroups.
         if(dp_tiles_ > 0)
         {
             num_wgs_per_tile = 2;
         }
         else
         {
             ck_tile::index_t iters_per_sk_cta_non_zero = ck_tile::max(iters_per_sk_cta_, 1);
             // Estimate the number of workgroups per macro tile.
             num_wgs_per_tile = (iters_per_tile_ / iters_per_sk_cta_non_zero) +
                                ((iters_per_tile_ % iters_per_sk_cta_non_zero) != 0);
         }
     }
  
     return std::max(num_wgs_per_tile, 1);
 }
  
 template <typename BlockGemmShapeType,
           StreamKReductionStrategy ReductionStrategyType,
           bool Persistent>
 struct StreamKTilePartitioner;
  
 // child class for Persistent Tile Partitioner
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, true>::StreamKTilePartitioner(
     ck_tile::index_t m, ck_tile::index_t n, ck_tile::index_t k, ck_tile::index_t grid)
     : StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>(m, n, k, grid)
 { // inherit from base constructor
     dp_tiles_per_cta_ = this->dp_tiles_ / this->grid_;
     extra_dp_tiles_   = this->dp_tiles_ % this->grid_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST auto
 StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, true>::grid_size() const noexcept
     -> dim3
 {
     if(extra_dp_tiles_ == 0)
     {
         return dim3(this->grid_, 1, 1);
     }
     else
     {
         return dim3(this->num_tiles_, 1, 1);
     }
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, true>::get_dp_tiles_per_cta()
     const noexcept
 {
     return dp_tiles_per_cta_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, true>::get_extra_dp_tiles()
     const noexcept
 {
     return extra_dp_tiles_;
 }
  
 // child class for Non-Persistent Tile Partitioner
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, false>::StreamKTilePartitioner(
     ck_tile::index_t m, ck_tile::index_t n, ck_tile::index_t k, ck_tile::index_t grid)
     : StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>(m, n, k, grid)
 { // inherit from base constructor
     dp_ctas_            = this->dp_tiles_;
     dp_start_block_idx_ = 0;
     sk_start_block_idx_ = this->dp_tiles_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST auto
 StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, false>::grid_size() const noexcept
     -> dim3
 {
     return dim3(dp_ctas_ + this->get_sk_ctas(), 1, 1);
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, false>::get_dp_ctas()
     const noexcept
 {
     return dp_ctas_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, false>::get_dp_start_block_idx()
     const noexcept
 {
     return dp_start_block_idx_;
 }
  
 template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
 CK_TILE_HOST_DEVICE index_t
 StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, false>::get_sk_start_block_idx()
     const noexcept
 {
     return sk_start_block_idx_;
 }
  
 } // namespace ck_tile