/*************************************************************************************************** * Copyright (c) 2023 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: BSD-3-Clause * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this * list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * 3. Neither the name of the copyright holder nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * **************************************************************************************************/ #pragma once #include "cutlass/barrier.h" #include "cutlass/block_striped.h" #include "cutlass/fast_math.h" #include "cutlass/gemm/kernel/sm90_tile_scheduler.hpp" #include "cutlass/kernel_hardware_info.hpp" #include "cute/layout.hpp" #include "cute/tensor.hpp" namespace cutlass::gemm::kernel::detail { // Persistent Thread Block (TB) scheduler leveraging stream-K decomposition template < class TileShape, class ClusterShape > class PersistentTileSchedulerSm90StreamK { // // Data members // private: using UnderlyingScheduler = PersistentTileSchedulerSm90; private: using UnderlyingArguments = typename UnderlyingScheduler::Arguments; using UnderlyingParams = typename UnderlyingScheduler::Params; uint64_t current_work_linear_idx_ = 0; public: using RasterOrder = UnderlyingScheduler::RasterOrder; using RasterOrderOptions = UnderlyingScheduler::RasterOrderOptions; // Use a dummy barrier manager to simply get the type used to store the barrier using BarrierType = typename NamedBarrierManager<1>::T; using Params = PersistentTileSchedulerSm90StreamKParams; using ReductionMode = Params::ReductionMode; struct WorkTileInfo { int32_t M_idx = 0; int32_t N_idx = 0; int32_t K_idx = 0; int32_t L_idx = 0; // Number of k tiles to compute for this unit of work. For stream-K, this // can indicate the number of K tiles across multiple output tiles. uint32_t k_tile_count = 0; // Number of k tiles remaining for the work unit as a whole uint32_t k_tile_remaining = 0; CUTLASS_HOST_DEVICE bool is_valid() const { // Use negative indices to denote invalid work return M_idx >= 0; } CUTLASS_HOST_DEVICE static WorkTileInfo invalid_work_tile() { return {-1, -1, -1, -1, 0}; } CUTLASS_HOST_DEVICE bool is_final_split(uint32_t k_tiles_per_output_tile) const { return (K_idx + k_tile_count) == k_tiles_per_output_tile; } }; struct Arguments { Arguments() = default; Arguments(Arguments const&) = default; Arguments(Arguments&&) = default; CUTLASS_HOST_DEVICE Arguments& operator=(Arguments const& args) { splits = args.splits; raster_order = args.raster_order; return *this; } CUTLASS_HOST_DEVICE Arguments& operator=(Arguments&& args) noexcept { splits = args.splits; raster_order = args.raster_order; return *this; } CUTLASS_HOST_DEVICE Arguments(int splits_) : splits(splits_) {} CUTLASS_HOST_DEVICE Arguments(int splits_, int max_swizzle_size_, RasterOrderOptions raster_order_) : splits(splits_), max_swizzle_size(max_swizzle_size_), raster_order(raster_order_) {} // The splitting factor to be used in a split-K decomposition of the problem. // If this is set to a value greater than 1, stream-K decomposition logic // is bypassed in favor of a split-K decomposition. int splits = 1; const int max_swizzle_size = 1; RasterOrderOptions raster_order = RasterOrderOptions::Heuristic; ReductionMode reduction_mode = ReductionMode::Deterministic; }; // Sink scheduler params as a member Params scheduler_params; // // Methods // template static Params to_underlying_arguments( ProblemShape problem_shape, TileShape tile_shape, ClusterShape cluster_shape, KernelHardwareInfo const& hw_info, Arguments const& args, void* workspace) { static_assert(cute::is_static::value); static_assert(cute::is_static::value); auto problem_shape_mnkl = cute::append<4>(problem_shape, cute::Int<1>{}); dim3 problem_blocks = get_tiled_cta_shape_mnl(problem_shape_mnkl, tile_shape, cluster_shape); uint32_t k_tile_per_output_tile = cute::size(cute::ceil_div(cute::shape<2>(problem_shape_mnkl), cute::shape<2>(TileShape{}))); Params params; params.initialize( problem_blocks, k_tile_per_output_tile, to_gemm_coord(cluster_shape), hw_info, args.splits, args.max_swizzle_size, args.raster_order, args.reduction_mode, workspace ); return params; } CUTLASS_HOST_DEVICE PersistentTileSchedulerSm90StreamK() { }; CUTLASS_HOST_DEVICE PersistentTileSchedulerSm90StreamK(Params const& params_) : scheduler_params(params_) { if (params_.raster_order_ == RasterOrder::AlongN) { current_work_linear_idx_ = uint64_t(blockIdx.x) + uint64_t(blockIdx.y) * uint64_t(gridDim.x); } else { current_work_linear_idx_ = uint64_t(blockIdx.x) * uint64_t(gridDim.y) + uint64_t(blockIdx.y); } } CUTLASS_DEVICE WorkTileInfo get_current_work() const { return get_current_work_for_linear_idx(current_work_linear_idx_, scheduler_params); } CUTLASS_DEVICE static WorkTileInfo get_current_work_for_linear_idx(uint64_t linear_idx, Params const& params) { // The maximum number of work units is units_per_problem_ * splits_. // The multiplication by splits_ is used for handling split-K, in which // units_per_problem_ is equal to the total number of output tiles. To account // for the fact that we have splits_ peers per output tile, we multiply this // value by splits_. For stream-K, this multiplication ends up being a no-op // because splits_ is set to 1 for stream-K. if (linear_idx >= params.units_per_problem_ * params.splits_) { // Invalid work. Return an empty result. return WorkTileInfo::invalid_work_tile(); } WorkTileInfo work_tile_info; assign_work(params, linear_idx, work_tile_info); return work_tile_info; } // Returns whether the current work_tile_info passed in should continue to be used. This // occurs only in the stream-K decomposition with stream-K work units, which encompass // work over multiple output tiles. If the current work_tile_info should continue to be // used, it is updated to advance to the next output tile it should cover. CUTLASS_DEVICE bool continue_current_work(WorkTileInfo& work_tile_info) const { return continue_current_work_for_linear_idx( current_work_linear_idx_, work_tile_info, scheduler_params); } CUTLASS_DEVICE static bool continue_current_work_for_linear_idx( uint64_t linear_idx, WorkTileInfo& work_tile_info, Params const& params) { work_tile_info.k_tile_remaining -= work_tile_info.k_tile_count; if (work_tile_info.k_tile_remaining == 0) { return false; } assign_work(params, linear_idx, work_tile_info); return true; } CUTLASS_DEVICE void advance_to_next_work(uint32_t advance_count = 1) { current_work_linear_idx_ += uint64_t(gridDim.x) * uint64_t(gridDim.y) * uint64_t(gridDim.z) * uint64_t(advance_count); } // Given the inputs, computes the total number of output blocks this problem will compute over // Note that this is only the logical size of our grid, not the physical grid we will actually launch. template CUTLASS_HOST_DEVICE static dim3 get_tiled_cta_shape_mnl(ProblemShape problem_shape_mnkl, TileShape cta_shape, ClusterShape cluster_shape) { return UnderlyingScheduler::get_tiled_cta_shape_mnl(problem_shape_mnkl, cta_shape, cluster_shape); } // Given the cluster shape, computes the physical grid we should launch. template CUTLASS_HOST_DEVICE static dim3 get_grid_shape( ProblemShape problem_shape, TileShape tile_shape, ClusterShape cluster_shape, KernelHardwareInfo hw_info, Arguments arguments) { auto problem_shape_mnkl = cute::append<4>(problem_shape, cute::Int<1>{}); dim3 problem_blocks = get_tiled_cta_shape_mnl(problem_shape_mnkl, tile_shape, cluster_shape); return Params::get_grid_shape( problem_blocks, to_gemm_coord(cluster_shape), hw_info, arguments.max_swizzle_size, arguments.raster_order ); } // Returns whether fixup is needed for `work_tile_info`. CUTLASS_HOST_DEVICE static bool requires_fixup(Params const& params, WorkTileInfo const& work_tile_info) { // Fixup is not needed for data-parallel tiles return work_tile_info.k_tile_count != params.divmod_tiles_per_output_tile_.divisor; } // Performs the reduction across splits for a given output tile. template CUTLASS_DEVICE static void fixup( Params const& params, WorkTileInfo const& work_tile_info, FrgTensorC& accumulators, uint32_t num_barriers, uint32_t barrier_idx) { static constexpr uint32_t Offset = 2; static constexpr uint32_t MaxNumNamedBarriers = 2; using BarrierManager = NamedBarrierManager; return fixup_helper( params, work_tile_info, accumulators, num_barriers, barrier_idx); } // Helper for performing the reduction across splits for a given output tile. template CUTLASS_DEVICE static void fixup_helper( Params const& params, WorkTileInfo const& work_tile_info, FrgTensorC& accumulators, uint32_t num_barriers, uint32_t barrier_idx) { using ElementAccumulator = typename FrgTensorC::value_type; if (!requires_fixup(params, work_tile_info)) { return; } auto tile_idx = output_tile_index(params, work_tile_info); // Index of the lock on which to wait auto lock_idx = (tile_idx * num_barriers) + barrier_idx; // Reductions use BlockStripedReduce with a width of BarrierManager::ThreadCount under the hood. // Thus, the start of the reduction space is the same across all threads in a warp group. int reduction_offset = (cute::size<0>(TileShape{}) * cute::size<1>(TileShape{}) * tile_idx) + (size(accumulators) * barrier_idx * BarrierManager::ThreadCount); ElementAccumulator* group_reduction_workspace = reinterpret_cast(params.reduction_workspace_) + reduction_offset; using AccumulatorArrayT = Array; using BlockStripedReduceT = BlockStripedReduce; // The number of tiles for which reduction is required is either: // (a) the total number of output tiles (in the case of split-K) // (b) the number of stream-K tiles // To calcualte the total number of output tiles in the split-K case, we // note that, in the split-K case, the units_per_problem_ member of Params will be // the total number of output tiles. auto reduction_tiles = params.splits_ > 1 ? params.units_per_problem_ : params.sk_tiles_; auto reduction_workspace_size = Params::get_reduction_workspace_size( reduction_tiles, to_gemm_coord(TileShape{}), sizeof_bits::value); BarrierType* lock_workspace = reinterpret_cast( reinterpret_cast(params.reduction_workspace_) + reduction_workspace_size); AccumulatorArrayT* reduction_workspace_array = reinterpret_cast(group_reduction_workspace); AccumulatorArrayT* accumulator_array = reinterpret_cast(&accumulators); int barrier_group_thread_idx = threadIdx.x % BarrierManager::ThreadCount; if (!work_tile_info.is_final_split(params.divmod_tiles_per_output_tile_.divisor)) { if (work_tile_info.K_idx == 0) { // First peer initializes the workspace partials BlockStripedReduceT::store(reduction_workspace_array, *accumulator_array, barrier_group_thread_idx); } else { if (params.reduction_mode_ == ReductionMode::Deterministic) { // Wait until the preceding split added its accumulators BarrierManager::wait_eq(barrier_idx, lock_workspace, barrier_group_thread_idx, lock_idx, work_tile_info.K_idx); } else { // Wait until the first split has stored its accumulators. Note that the first split will have // accumulated a value into the lock potentially greater than one (since the locked value is // incremented by work_tile_info.k_tile_count below for both the deterministic and non-deterministic) // cases. For non-deterministic reductions, all that non-first or last splits care about is whether // the first split has been written, so we only wait while the locked value is less than 1. This // avoids having to add logic to determine the work_tile_info.k_tile_count for the first split. BarrierManager::wait_lt(barrier_idx, lock_workspace, barrier_group_thread_idx, lock_idx, 1); } // Perform reduction in workspace BlockStripedReduceT::reduce(reduction_workspace_array, *accumulator_array, barrier_group_thread_idx); } // Signal our arrival BarrierManager::arrive_inc(barrier_idx, lock_workspace, barrier_group_thread_idx, lock_idx, work_tile_info.k_tile_count); } else { // Wait until the preceding split added its accumulators. // For both the deterministic and non-deterministic case, each preceding split will have incremented // the locked value by work_tile_info.k_tile_count. Thus, the final split konws that it can begin // loading the partially-reduced value when the locked value reaches its starting K tile index (i.e., // work_tile_info.K_idx). BarrierManager::wait_eq(barrier_idx, lock_workspace, barrier_group_thread_idx, lock_idx, work_tile_info.K_idx); // The block computing the final split for the tile adds previously-reduced partials // to its accumulators and computes the epilogue. BlockStripedReduceT::load_add(*accumulator_array, reduction_workspace_array, barrier_group_thread_idx); } } // Returns whether the block assigned this work should compute the epilogue for the corresponding // output tile. For the case of stream-K, this should only occur if the work is marked as the final split. CUTLASS_HOST_DEVICE static bool compute_epilogue(WorkTileInfo const& work_tile_info, Params const& params) { return work_tile_info.is_final_split(params.divmod_tiles_per_output_tile_.divisor); } // Returns the linearized index of the output tile corresponding to the tile with offset [L, M, K] CUTLASS_DEVICE static int output_tile_index(Params const& params, WorkTileInfo const& work_tile_info) { uint64_t linear_idx_in_batch = UnderlyingScheduler::get_linear_idx_from_m_and_n( work_tile_info.M_idx, work_tile_info.N_idx, params.divmod_cluster_shape_major_, params.divmod_cluster_shape_minor_, params.divmod_cluster_blk_major_, params.log_swizzle_size_, params.raster_order_ ); uint64_t tiles_mn = params.divmod_batch_.divisor; return tiles_mn * work_tile_info.L_idx + linear_idx_in_batch; } template static int get_workspace_size( Arguments const& args, ProblemShape problem_shape, KernelHardwareInfo const& hw_info, uint32_t mma_warp_groups) { auto problem_shape_mnkl = cute::append<4>(problem_shape, 1); ClusterShape cluster_shape; TileShape tile_shape; dim3 problem_blocks = get_tiled_cta_shape_mnl(problem_shape_mnkl, tile_shape, cluster_shape); uint32_t k_tile_per_output_tile = cute::size(cute::ceil_div(cute::shape<2>(problem_shape_mnkl), cute::shape<2>(TileShape{}))); return Params::get_workspace_size( problem_blocks, k_tile_per_output_tile, to_gemm_coord(tile_shape), to_gemm_coord(cluster_shape), hw_info, args.splits, args.max_swizzle_size, args.raster_order, mma_warp_groups, sizeof_bits::value, sizeof_bits::value ); } template static cutlass::Status initialize_workspace( Arguments const& args, void* workspace, cudaStream_t stream, ProblemShape const& problem_shape, KernelHardwareInfo const& hw_info, uint32_t mma_warp_groups) { auto problem_shape_mnkl = cute::append<4>(problem_shape, 1); ClusterShape cluster_shape; TileShape tile_shape; dim3 problem_blocks = get_tiled_cta_shape_mnl(problem_shape_mnkl, tile_shape, cluster_shape); uint32_t k_tile_per_output_tile = cute::size(cute::ceil_div(cute::shape<2>(problem_shape_mnkl), cute::shape<2>(TileShape{}))); return Params::initialize_workspace( workspace, stream, problem_blocks, k_tile_per_output_tile, to_gemm_coord(tile_shape), to_gemm_coord(cluster_shape), hw_info, args.splits, args.max_swizzle_size, args.raster_order, mma_warp_groups, sizeof_bits::value, sizeof_bits::value ); } template CUTLASS_HOST_DEVICE static int get_work_k_tile_count(WorkTileInfo const& work_tile_info, ProblemShape, TileShape) { return work_tile_info.k_tile_count; } CUTLASS_HOST_DEVICE static uint32_t get_work_k_tile_start(WorkTileInfo const& work_tile_info) { return work_tile_info.K_idx; } // Sets the current stream-K work to compute within work_tile_info. If new_unit is true, work_tile_info // is populated as a new unit of work. Otherwise, state existing in work_tile_info (e.g., remaining // iterations) is used to find the next tile in the current work unit. CUTLASS_DEVICE static void assign_work( Params const& params, uint64_t linear_idx, WorkTileInfo& work_tile_info) { uint64_t true_tile_id = linear_idx; if (linear_idx >= params.sk_units_ && params.splits_ == 1) { // Data-parallel work true_tile_id = linear_idx - params.sk_units_ + params.sk_tiles_; work_tile_info.K_idx = 0; work_tile_info.k_tile_count = params.divmod_tiles_per_output_tile_.divisor; work_tile_info.k_tile_remaining = params.divmod_tiles_per_output_tile_.divisor; } else { // In the CUTLASS 2.x implementation of stream K, stream-K work is assigned to each stream-K // threadblock individually. For the most part, the set of K iterations corresponding to stream-K // work was divided amongst stream-K threadblocks, and a threadblock determined which tile // it would compute a (potentially-partial) output tile for based on the space of k iterations // assigned to it. This often results in stream-K threadblocks processing tiles with different // offsets in the K dimension from one another. This can reduce locality, but is lmitied to the // (generally few) waves of threadblocks assigned to compute stream-K work. // // With the introduction of threadblock clusters, there is additional benefit to maintaining // locality in the K dimension: shared portions of operands can be multicasted to threadblocks // within a cluster. Thus, we would like to ensure that the assignment of stream-K work to // threadblocks respects the ability to perform multicasting. // // To do so, we divide up the linearized stream-K units into clusters and share the same K // offsets for work within clusters. // Equivalent to linear_idx / cluster_size auto cluster_linear_work_idx = params.divmod_cluster_shape_minor_.divide( params.divmod_cluster_shape_major_.divide(linear_idx) ); uint64_t split; params.divmod_clusters_mnl_(split, cluster_linear_work_idx, cluster_linear_work_idx); auto big_unit_cmp = params.splits_ > 1 ? split : cluster_linear_work_idx; auto linear_idx_mult = params.splits_ > 1 ? params.divmod_tiles_per_output_tile_.divisor : params.k_tiles_per_sk_unit_; // Determine the starting k iteration computed by this stream-K work unit uint32_t unit_iter_start = (linear_idx_mult * cluster_linear_work_idx) + (params.k_tiles_per_sk_unit_ * split); // Adjust the starting position and number of k iterations for "big units," which // compute one extra iteration. These are the first big_units_ units in the // linearized ID space. bool is_big_unit = big_unit_cmp < params.big_units_; if (is_big_unit) { // Since the "big units" are the first units in the linearized ID space, each // of the units preceding this big unit computed one extra iteration. Thus, // we must offset our start iteration by the number of units that precede // the current unit in the linearized ID space. unit_iter_start += big_unit_cmp; } else { // Increment by one for each of the big clusters (since all big units precede this unit) unit_iter_start += params.big_units_; } if (work_tile_info.k_tile_count == 0) { // This is a new unit work_tile_info.k_tile_remaining = params.k_tiles_per_sk_unit_; // Only adjust iteration count for big unit if we are initializing this // work unit. For existing work units, the extra iteration for big units // has already been accounted for in k_tiles_reamaining if (is_big_unit) { ++work_tile_info.k_tile_remaining; } } // Find the output tile corresponding to the final k iteration covered by this // work unit. Stream-K work units will work backwards in terms of the tiles they // are responsible computing. This is beneficial because the final (partial) // tile computed by a stream-K block is typically the beginning of the output // tile, while the beginning (partial) tile is typically the ending of another // output tile. Since ending portions of an output tile must reduce across // other work units computing portions of that output tile, it is preferable // for them to be computed later, so as to reduce the likelihood of blocking // on other work. uint32_t unit_iter_end = unit_iter_start + work_tile_info.k_tile_remaining - 1; true_tile_id = params.divmod_tiles_per_output_tile_.divide(unit_iter_end); uint32_t true_tile_iter_start = true_tile_id * params.divmod_tiles_per_output_tile_.divisor; uint32_t true_tile_iter_end = true_tile_iter_start + params.divmod_tiles_per_output_tile_.divisor; // Bring the linearized tile ID back into the space of tiles, rather than clusters true_tile_id *= params.divmod_cluster_shape_major_.divisor * params.divmod_cluster_shape_minor_.divisor; auto [cta_m_in_cluster, cta_n_in_cluster, _] = cute::block_id_in_cluster(); // The final linearized tile ID is in units of the cluster dimension over which we rasterize. if (params.raster_order_ == RasterOrder::AlongN) { true_tile_id += cta_n_in_cluster * params.divmod_cluster_shape_minor_.divisor; } else { true_tile_id += cta_m_in_cluster * params.divmod_cluster_shape_minor_.divisor; } // The unit's starting k iteration in the current tile is either the starting // iteration for the tile as a whole, or the starting k iteration for the unit // as a whole (if the latter is greater than the former). uint32_t tile_iter_start = max(true_tile_iter_start, unit_iter_start); // Similarly, the unit's ending k iteration (exclusive) is either the end of // the current tile it is assigned, or the ending iteration of the unit as a whole // (if the latter is less than the former). uint32_t tile_iter_end = min(true_tile_iter_end, unit_iter_end + 1); // Set the k offset to be the starting k tile for this output tile work_tile_info.K_idx = static_cast(tile_iter_start - true_tile_iter_start); work_tile_info.k_tile_count = tile_iter_end - tile_iter_start; } uint64_t work_idx_l, remainder; params.divmod_batch_(work_idx_l, remainder, true_tile_id); uint64_t cta_per_grid_dim = params.divmod_cluster_shape_minor_.divide(remainder); auto [work_idx_m, work_idx_n] = UnderlyingScheduler::get_work_idx_m_and_n( cta_per_grid_dim, params.divmod_cluster_shape_major_, params.divmod_cluster_shape_minor_, params.divmod_cluster_blk_major_, params.log_swizzle_size_, params.raster_order_); // Set the M, N, and L block offsets work_tile_info.M_idx = work_idx_m; work_tile_info.N_idx = work_idx_n; work_tile_info.L_idx = static_cast(work_idx_l); } }; } // namespace cutlass::gemm::kernel::detail