squall/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity
cc3c29a81a
cutlass/include/cutlass/gemm/kernel/sm90_tile_scheduler_stream_k.hpp
Yujia Zhai cc3c29a81a
CUTLASS 3.6.0 (#1850) 
* v3.6

* update changelog

* update readme

* fix typo

* fixing typos

* hopper gemm with weight prefetch

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
2024-10-09 15:33:27 -04:00

946 lines
40 KiB
C++
Raw Blame History

/***************************************************************************************************
* Copyright (c) 2023 - 2024 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;
static constexpr bool IsDynamicPersistent = false;
// 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;
using DecompositionMode = Params::DecompositionMode;
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;
// Whether this unit of work is the final split for the given tile
bool is_separate_reduction = false;
CUTLASS_HOST_DEVICE
bool
is_valid() const {
// A work tile that computes no K tiles is invalid unless it is a separate-reduction work tile
// (which only performs reduction and epilogue)
return k_tile_count > 0 || is_separate_reduction;
}
CUTLASS_HOST_DEVICE
bool
is_reduction_unit() const {
return is_separate_reduction;
}
CUTLASS_HOST_DEVICE
int32_t
reduction_subtile_idx() const {
// For separate reduction units, the K_idx of the work tile is unused.
// Therefore, we override it to contain the subtile of that the reduction
// unit operates on.
return is_reduction_unit() ? K_idx : -1;
}
CUTLASS_HOST_DEVICE
void
setup_separate_reduction(int32_t epilogue_subtile_idx) {
// Set the epilogue subtile in the K_idx, since this is otherwise unused
// by separate reduction units.
K_idx = epilogue_subtile_idx;
is_separate_reduction = true;
k_tile_count = 0;
// Clean up remaining k tiles
k_tile_remaining = 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;
max_swizzle_size = args.max_swizzle_size;
raster_order = args.raster_order;
reduction_mode = args.reduction_mode;
decomposition_mode = args.decomposition_mode;
return *this;
}
CUTLASS_HOST_DEVICE
Arguments&
operator=(Arguments&& args) noexcept {
splits = args.splits;
max_swizzle_size = args.max_swizzle_size;
raster_order = args.raster_order;
reduction_mode = args.reduction_mode;
decomposition_mode = args.decomposition_mode;
return *this;
}
CUTLASS_HOST_DEVICE
Arguments(int splits_) : splits(splits_) {}
CUTLASS_HOST_DEVICE
Arguments(int splits_, int max_swizzle_size_, RasterOrderOptions raster_order_, DecompositionMode decomposition_mode_) :
splits(splits_),
max_swizzle_size(max_swizzle_size_),
raster_order(raster_order_),
decomposition_mode(decomposition_mode_) {}
// 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;
int max_swizzle_size = 1;
RasterOrderOptions raster_order = RasterOrderOptions::Heuristic;
ReductionMode reduction_mode = ReductionMode::Deterministic;
DecompositionMode decomposition_mode = DecompositionMode::Heuristic;
};
// Sink scheduler params as a member
Params scheduler_params;
//
// Methods
//
template <class ProblemShape>
static Params
to_underlying_arguments(
ProblemShape problem_shape,
TileShape tile_shape,
ClusterShape cluster_shape,
KernelHardwareInfo const& hw_info,
Arguments const& args,
void* workspace,
const uint32_t epilogue_subtile = 1,
[[maybe_unused]] uint32_t ktile_start_alignment_count = 1u) {
static_assert(cute::is_static<TileShape>::value);
static_assert(cute::is_static<ClusterShape>::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,
args.decomposition_mode,
workspace,
epilogue_subtile
);
return params;
}
static bool
can_implement(Arguments const& args) {
// Split count > 1 is only valid for heuristic and split-K decomposition modes
return (args.splits == 1 ||
args.decomposition_mode == DecompositionMode::Heuristic ||
args.decomposition_mode == DecompositionMode::SplitK);
}
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.divmod_splits_.divisor + params.separate_reduction_units_)) {
// 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 work_tile_info.is_valid();
}
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 <class ProblemShape>
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 <class ProblemShape>
CUTLASS_HOST_DEVICE static
dim3
get_grid_shape(
[[maybe_unused]] Params const& params,
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 invalid or data-parallel tiles
return work_tile_info.is_valid() && work_tile_info.k_tile_count != params.divmod_tiles_per_output_tile_.divisor;
}
CUTLASS_HOST_DEVICE
static bool
requires_separate_reduction(Params const& params) {
return params.requires_separate_reduction();
}
// When the work tile is not special for reduction, it's valid. Otherwise need to skip
// global loading that producer warpgroup do, also math computation that consumer warpgroup do.
CUTLASS_DEVICE
static bool
valid_warpgroup_in_work_tile(WorkTileInfo const& work_tile_info) {
return !work_tile_info.is_reduction_unit();
}
// Performs the reduction across splits for a given output tile.
template <class FrgTensorC>
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 = static_cast<int>(cutlass::arch::ReservedNamedBarriers::StreamkBarrier0);
static constexpr uint32_t MaxNumNamedBarriers = 2;
using BarrierManager = NamedBarrierManager<NumThreadsPerWarpGroup, Offset, MaxNumNamedBarriers>;
return fixup_helper<FrgTensorC, BarrierManager>(
params, work_tile_info, accumulators, num_barriers, barrier_idx);
}
// Helper for performing the reduction across splits for a given output tile.
template <class FrgTensorC, class BarrierManager>
CUTLASS_DEVICE
static void
fixup_helper(
Params const& params,
WorkTileInfo const& work_tile_info,
FrgTensorC& accumulators,
uint32_t num_barriers,
uint32_t barrier_idx,
uint32_t num_accumulator_mtxs = 1) {
using ElementAccumulator = typename FrgTensorC::value_type;
if (!requires_fixup(params, work_tile_info)) {
return;
}
uint64_t tile_idx = output_tile_index(params, work_tile_info);
// Index of the lock on which to wait
uint64_t lock_idx = (tile_idx * num_barriers) + barrier_idx;
uint64_t reduction_tile_idx = tile_idx;
uint64_t num_peers = 0;
uint64_t reduction_peer_offset = 0;
if (params.requires_separate_reduction()) {
// If separate reduction is to be performed, each stream-K unit writes its partials
// to a separate portion of the workspace. There are as many of these portions as there
// are peers for a given output tile, so we multiply the tile index by the maximum peer count.
auto [first_peer_id, my_peer_id, last_peer_id] = tile_peer_range(params, tile_idx, static_cast<uint32_t>(work_tile_info.K_idx));
num_peers = last_peer_id - first_peer_id + 1;
reduction_tile_idx *= Params::max_peers_per_tile(params.sk_units_, params.sk_tiles_);
reduction_peer_offset = my_peer_id * cute::size<0>(TileShape{}) * cute::size<1>(TileShape{});
}
// 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.
uint64_t reduction_offset =
(static_cast<uint64_t>(cute::size<0>(TileShape{})) * static_cast<uint64_t>(cute::size<1>(TileShape{})) * reduction_tile_idx * num_accumulator_mtxs) +
reduction_peer_offset +
(static_cast<uint64_t>(size(accumulators)) * barrier_idx * BarrierManager::ThreadCount);
ElementAccumulator* group_reduction_workspace = reinterpret_cast<ElementAccumulator*>(params.reduction_workspace_) + reduction_offset;
using AccumulatorArrayT = Array<typename FrgTensorC::value_type, size(FrgTensorC{})>;
using BlockStripedReduceT = BlockStripedReduce<BarrierManager::ThreadCount, AccumulatorArrayT>;
AccumulatorArrayT* reduction_workspace_array = reinterpret_cast<AccumulatorArrayT*>(group_reduction_workspace);
AccumulatorArrayT* accumulator_array = reinterpret_cast<AccumulatorArrayT*>(accumulators.data());
uint32_t barrier_group_thread_idx = threadIdx.x % BarrierManager::ThreadCount;
// 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 (potentially multiplied by peer count if using separate reduction)
// To calculate 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.
uint32_t reduction_tiles = 0;
if (params.divmod_splits_.divisor > 1) {
reduction_tiles = params.units_per_problem_;
}
else if (params.requires_separate_reduction()) {
reduction_tiles = params.sk_tiles_ * Params::max_peers_per_tile(params.sk_units_, params.sk_tiles_);
}
else {
reduction_tiles = params.sk_tiles_;
}
uint64_t reduction_workspace_size = Params::get_reduction_workspace_size(
reduction_tiles, to_gemm_coord(TileShape{}), sizeof_bits<ElementAccumulator>::value, num_accumulator_mtxs);
BarrierType* lock_workspace = reinterpret_cast<BarrierType*>(
reinterpret_cast<uint8_t*>(params.reduction_workspace_) + reduction_workspace_size);
if (work_tile_info.is_reduction_unit()) {
plus<AccumulatorArrayT> add_fragments;
uint64_t peer_offset = size(accumulators) * num_barriers * BarrierManager::ThreadCount;
// Wait until the peers collaborating on this output tile have all written
// their accumulators to workspace.
BarrierManager::wait_eq(barrier_idx, lock_workspace, barrier_group_thread_idx, lock_idx, num_peers);
// Load the first peer's data
BlockStripedReduceT::load(*accumulator_array, reduction_workspace_array, barrier_group_thread_idx);
for (uint64_t i = 1; i < num_peers; ++i) {
// Load peer fragment
AccumulatorArrayT addend_fragment;
auto peer_reduction_workspace = reinterpret_cast<AccumulatorArrayT*>(group_reduction_workspace + (i * peer_offset));
BlockStripedReduceT::load(addend_fragment, peer_reduction_workspace, barrier_group_thread_idx);
// Add peer fragment
*accumulator_array = add_fragments(*accumulator_array, addend_fragment);
}
}
else if (!compute_epilogue(work_tile_info, params)) {
if (params.requires_separate_reduction() || work_tile_info.K_idx == 0) {
// The first peer initializes the workspace partials in the non-separate-reduction case,
// and all peers write to their own location in workspace when using separate reduction
BlockStripedReduceT::store(reduction_workspace_array, *accumulator_array, barrier_group_thread_idx);
}
else {
// 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);
// Perform reduction in workspace
BlockStripedReduceT::reduce(reduction_workspace_array, *accumulator_array, barrier_group_thread_idx);
}
// If separate reduction is being performed, each participating stream-K unit increments the barrier
// by only 1. Otherwise, increment by the K tile count that this unit has processed.
uint32_t increment = params.requires_separate_reduction() ? 1 : work_tile_info.k_tile_count;
// Signal our arrival
BarrierManager::arrive_inc(barrier_idx, lock_workspace, barrier_group_thread_idx, lock_idx, increment);
}
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 unitl the first split has stored its accumulators
BarrierManager::wait_lt(barrier_idx, lock_workspace, barrier_group_thread_idx, lock_idx, 1);
}
// 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) {
// `is_final_split` will be set to `true` for the following scenarios, all of which must compute the epilogue:
// 1. The tile is computed in data-parallel mode
// 2. The tile is computed in split-/stream-K mode and this work unit represents the final split of the tile
// 3. The tile is computed in split-/stream-K mode and separate reduction is used, and this is a separate reduction unit
return work_tile_info.is_valid() &&
(work_tile_info.is_final_split(params.divmod_tiles_per_output_tile_.divisor) &&
!params.requires_separate_reduction()) || work_tile_info.is_separate_reduction;
}
// Returns the linearized index of the output tile corresponding to the tile with offset [L, M, K]
CUTLASS_DEVICE
static uint64_t
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 <class ProblemShape, class ElementAccumulator>
static size_t
get_workspace_size(
Arguments const& args,
ProblemShape problem_shape,
KernelHardwareInfo const& hw_info,
uint32_t mma_warp_groups,
const uint32_t epilogue_subtile = 1,
[[maybe_unused]] uint32_t num_accumulator_mtxs = 1) {
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,
args.decomposition_mode,
mma_warp_groups,
sizeof_bits<BarrierType>::value,
sizeof_bits<ElementAccumulator>::value,
epilogue_subtile
);
}
template <class ProblemShape, class ElementAccumulator>
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,
const uint32_t epilogue_subtile = 1,
[[maybe_unused]] uint32_t num_accumulator_mtxs = 1,
CudaHostAdapter* cuda_adapter = nullptr) {
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,
args.decomposition_mode,
mma_warp_groups,
sizeof_bits<BarrierType>::value,
sizeof_bits<ElementAccumulator>::value,
epilogue_subtile,
1,
cuda_adapter
);
}
template <class ProblemShape>
CUTLASS_HOST_DEVICE
static uint32_t
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;
}
// Kernel helper function to get next work tile
CUTLASS_DEVICE
auto
fetch_next_work(WorkTileInfo work_tile_info) {
if (continue_current_work(work_tile_info)) {
return cute::make_tuple(work_tile_info, true);
}
advance_to_next_work();
return cute::make_tuple(get_current_work(), true);
}
// Returns the initial work tile info that will be computed over
CUTLASS_DEVICE
WorkTileInfo
initial_work_tile_info(ClusterShape) {
return get_current_work();
}
private:
// 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) {
auto [cta_m_in_cluster_, cta_n_in_cluster_, _] = cute::block_id_in_cluster();
uint64_t cta_m_in_cluster = static_cast<uint64_t>(cta_m_in_cluster_);
uint64_t cta_n_in_cluster = static_cast<uint64_t>(cta_n_in_cluster_);
uint64_t output_tile_id = linear_idx;
if (linear_idx >= params.units_per_problem_ * params.divmod_splits_.divisor) {
// Separate-reduction work
auto cluster_size = params.get_cluster_size();
// Divide up the linearized separate reduction units into clusters
uint64_t cluster_linear_reduction_unit_idx = params.div_cluster_size((linear_idx - params.units_per_problem_));
uint64_t cluster_tile_idx, epi_subtile_idx;
params.divmod_epilogue_subtile_(cluster_tile_idx, epi_subtile_idx, cluster_linear_reduction_unit_idx);
// Bring the linearized tile ID back into the space of tiles, rather than clusters
output_tile_id = cluster_tile_idx * cluster_size;
work_tile_info.setup_separate_reduction(epi_subtile_idx);
}
else if (linear_idx >= params.sk_units_ && params.divmod_splits_.divisor == 1) {
// Data-parallel work
output_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.
uint64_t cluster_linear_work_idx = params.div_cluster_size(linear_idx);
uint64_t group_idx;
params.divmod_sk_groups_(cluster_linear_work_idx, group_idx, cluster_linear_work_idx);
// Determine whether we are in a "big group" that will process an additional
// stream-K cluster tile.
uint64_t sk_cluster_tiles = params.div_cluster_size(params.sk_tiles_);
uint64_t sk_cluster_tiles_in_group = params.divmod_sk_groups_.divide(sk_cluster_tiles);
if (group_idx < params.big_groups_) {
++sk_cluster_tiles_in_group;
}
// Determine whether we are in a "big unit" within the group, that will process
// an additional K chunk in the group.
uint64_t sk_tiles_in_group = sk_cluster_tiles_in_group * params.get_cluster_size();
uint64_t k_tiles_in_group = sk_tiles_in_group * params.divmod_tiles_per_output_tile_.divisor;
uint64_t k_tiles_per_unit_in_group = params.divmod_sk_units_per_group_.divide(k_tiles_in_group);
uint64_t big_units_in_group = params.div_cluster_size(
k_tiles_in_group - (k_tiles_per_unit_in_group * params.divmod_sk_units_per_group_.divisor));
uint64_t split;
params.divmod_clusters_mnl_(split, cluster_linear_work_idx, cluster_linear_work_idx);
bool is_split_k = params.divmod_splits_.divisor > 1;
uint64_t big_unit_cmp_lhs = is_split_k ? split : cluster_linear_work_idx;
uint64_t big_unit_cmp_rhs = is_split_k ? params.big_units_ : big_units_in_group;
uint64_t linear_idx_mult = is_split_k ? params.divmod_tiles_per_output_tile_.divisor : k_tiles_per_unit_in_group;
uint64_t k_tiles_per_split = is_split_k ? params.divmod_k_tiles_per_sk_unit_.divisor : k_tiles_per_unit_in_group;
// Determine the starting k iteration computed by this stream-K work unit
uint32_t unit_iter_start = (linear_idx_mult * cluster_linear_work_idx) +
(k_tiles_per_split * split);
// Adjust the starting position and number of k iterations for "big units," which
// compute one extra iteration. If there are any big units, they will be the first
// in the linearized ID space.
auto k_tiles_in_my_split = k_tiles_per_split;
if (big_unit_cmp_lhs < big_unit_cmp_rhs) {
// 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_lhs;
++k_tiles_in_my_split;
}
else {
// Increment by one for each of the big clusters (since all big units precede this unit)
unit_iter_start += big_unit_cmp_rhs;
}
if (!is_split_k) {
// Adjust the unit starting position and number of tiles to avoid
// computing splits of size less than min_iters_per_sk_unit_
int unused, start_tile_k_tile;
params.divmod_tiles_per_output_tile_(unused, start_tile_k_tile, unit_iter_start);
if (start_tile_k_tile < Params::min_iters_per_sk_unit_) {
// Starting K tile is in range [0, Params::min_iters_per_sk_unit_), which means that another
// stream-K unit will be computing a split with fewer than Params::min_iters_per_sk_unit_ K tiles.
// Adjust our work to take over these K tiles.
unit_iter_start -= start_tile_k_tile;
k_tiles_in_my_split += start_tile_k_tile;
}
else if (start_tile_k_tile > (params.divmod_tiles_per_output_tile_.divisor - Params::min_iters_per_sk_unit_)) {
// Starting K tile is within the final Params::min_iters_per_sk_unit_ K tiles of some output tile,
// which means that this unit will compute a split with fewer than Params::min_iters_per_sk_unit_ K tiles.
// Adjust our work to shed these K tiles to a neighboring stream-K unit that will compute more consecutive K tiles.
auto adjustment_tiles = (params.divmod_tiles_per_output_tile_.divisor - start_tile_k_tile);
unit_iter_start += adjustment_tiles;
k_tiles_in_my_split -= adjustment_tiles;
}
else if (params.ktile_start_alignment_count == 2 && start_tile_k_tile % 2 != 0) {
// ktile for each SM start from even number
// If start from odd number ktile within the output tile
// now start at the ktile one before my initial ktile start (take one ktile from prev sm)
// if end on odd number ktile within the output tile
// now end at ktile that one before my ktile end (give one ktile to next sm)
unit_iter_start -= 1;
k_tiles_in_my_split += 1;
}
}
if (work_tile_info.k_tile_count == 0) {
// This is a new unit
if (!is_split_k) {
//
// Adjust the unit ending position and number of tiles to avoid
// computing splits of size less than min_iters_per_sk_unit_
//
// Begin by assuming that no adjustment is needed
auto initial_unit_iter_end = unit_iter_start + k_tiles_in_my_split;
int unused, end_tile_k_tile;
params.divmod_tiles_per_output_tile_(unused, end_tile_k_tile, initial_unit_iter_end);
if (end_tile_k_tile < Params::min_iters_per_sk_unit_) {
// Ending K tile is within the first Params::min_iters_per_sk_unit_ K tiles of some output tile,
// which means that this unit will compute a split with fewer than Params::min_iters_per_sk_unit_ K tiles.
// Adjust our work to shed these K tiles to a neighboring stream-K unit that will compute more consecutive K tiles.
k_tiles_in_my_split -= end_tile_k_tile;
}
else if (end_tile_k_tile > (params.divmod_tiles_per_output_tile_.divisor - Params::min_iters_per_sk_unit_)) {
// Ending K tile is within the final Params::min_iters_per_sk_unit_ K tiles of some output tile,
// which means that some other unit will compute a split with fewer than Params::min_iters_per_sk_unit_ K tiles.
// Adjust our work to take on these K tiles.
k_tiles_in_my_split += (params.divmod_tiles_per_output_tile_.divisor - end_tile_k_tile);
}
else if (params.ktile_start_alignment_count == 2 && end_tile_k_tile % 2 != 0) {
// ktile for each SM start from even number
// If start from odd number ktile within the output tile
// now start at the ktile one before my initial ktile start (take one ktile from prev sm)
// If end on odd number ktile within the output tile,
// now end at ktile that one before my ktile end (give one ktile to next sm)
k_tiles_in_my_split -= 1;
}
}
work_tile_info.k_tile_remaining = k_tiles_in_my_split;
}
uint32_t unit_iter_end = unit_iter_start + work_tile_info.k_tile_remaining - 1;
// Find the output tile corresponding to the final k tile 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.
auto output_tile_id_in_group = params.divmod_tiles_per_output_tile_.divide(unit_iter_end);
uint32_t output_tile_iter_start = output_tile_id_in_group * params.divmod_tiles_per_output_tile_.divisor;
uint32_t output_tile_iter_end = output_tile_iter_start + params.divmod_tiles_per_output_tile_.divisor;
// Convert the output tile from the linearized space within each group to the
// overall linearized space.
output_tile_id = (output_tile_id_in_group * params.divmod_sk_groups_.divisor) + group_idx;
// Bring the linearized tile ID back into the space of tiles, rather than clusters
output_tile_id *= params.get_cluster_size();
// The final linearized tile ID is in units of the cluster dimension over which we rasterize.
if (params.raster_order_ == RasterOrder::AlongN) {
output_tile_id += cta_n_in_cluster * params.divmod_cluster_shape_minor_.divisor;
}
else {
output_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(output_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(output_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<int32_t>(tile_iter_start - output_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, output_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<int32_t>(work_idx_l);
}
// Returns the starting and ending peer ID of this tile
CUTLASS_HOST_DEVICE
static auto
tile_peer_range(Params const& params, uint32_t tile_idx, uint32_t cur_k_tile) {
uint32_t tile_idx_in_cluster_path = params.div_cluster_size(tile_idx);
uint32_t start_k_tile = params.divmod_tiles_per_output_tile_.divisor * tile_idx_in_cluster_path;
uint32_t end_k_tile = start_k_tile + params.divmod_tiles_per_output_tile_.divisor - 1;
uint32_t big_unit_k_tiles = params.big_units_ * (params.divmod_k_tiles_per_sk_unit_.divisor + 1);
auto adjust_unit = [&](uint32_t k_tile, uint32_t unit_idx, uint32_t k_tiles_per_unit) {
uint32_t unit_k_start = unit_idx * k_tiles_per_unit;
uint32_t unit_k_end = unit_k_start + k_tiles_per_unit;
if (k_tile - start_k_tile < Params::min_iters_per_sk_unit_ &&
unit_k_end - start_k_tile < Params::min_iters_per_sk_unit_) {
// k_tile is within the first min_iters_per_sk_unit_ K tiles of this output tile,
// and the stream-K unit computes fewer than min_iters_per_sk_unit_ K tiles for this
// output tile. This work will thus be subsumed by the next stream-K unit.
++unit_idx;
}
if (end_k_tile + 1 - k_tile < Params::min_iters_per_sk_unit_ &&
end_k_tile + 1 - unit_k_start < Params::min_iters_per_sk_unit_) {
// k_tile is within the last min_iters_per_sk_unit_ K tiles of this output tile,
// and the stream-K unit computes fewer than min_iters_per_sk_unit_ K tiles for this
// output tile. This work will thus be subsumed by the previous stream-K unit.
--unit_idx;
}
return unit_idx;
};
// Lambda to find the ID of the stream-K unit that computes this K tile
auto find_unit = [&](uint32_t k_tile) {
if (k_tile < big_unit_k_tiles) {
// The tile is within the "big unit range"
uint32_t unit_idx = params.divmod_k_tiles_per_sk_big_unit_.divide(k_tile);
return static_cast<uint64_t>(adjust_unit(k_tile, unit_idx, params.divmod_k_tiles_per_sk_big_unit_.divisor));
}
else {
// The tile is after the "big unit range." Account for this by finding the "normal unit"
// that it belongs to, and then offsetting by the number of big units
uint32_t unit_idx = params.divmod_k_tiles_per_sk_unit_.divide(k_tile - big_unit_k_tiles) + params.big_units_;
return static_cast<uint64_t>(adjust_unit(k_tile, unit_idx, params.divmod_k_tiles_per_sk_unit_.divisor));
}
};
return cute::make_tuple(find_unit(start_k_tile), find_unit(cur_k_tile), find_unit(end_k_tile));
}
};
} // namespace cutlass::gemm::kernel::detail
Reference in New Issue View Git Blame Copy Permalink