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>
This commit is contained in:
Yujia Zhai
2024-10-09 15:33:27 -04:00
committed by GitHub
co-authored by yuzhai Haicheng Wu
parent 0837a2a00a
commit cc3c29a81a
354 changed files with 105937 additions and 8197 deletions
@@ -37,9 +37,12 @@
#include "cute/tensor.hpp"
#include "cute/arch/cluster_sm90.hpp"
#include "cutlass/conv/detail.hpp"
#include "cutlass/conv/convolution.h"
#include "cutlass/conv/dispatch_policy.hpp"
#include "cutlass/gemm/dispatch_policy.hpp"
#include "cutlass/pipeline/sm90_pipeline.hpp"
#include "cutlass/gemm/kernel/gemm_universal.hpp"
#include "cutlass/gemm/kernel/tile_scheduler.hpp"
///////////////////////////////////////////////////////////////////////////////
@@ -49,365 +52,25 @@ namespace cutlass::conv::kernel {
///////////////////////////////////////////////////////////////////////////////
template <
class ProblemShape_,
class CollectiveMainloop_,
class CollectiveEpilogue_,
class TileSchedulerTag_
class TileScheduler_
>
class ConvUniversal<
ProblemShape_,
CollectiveMainloop_,
CollectiveEpilogue_,
TileSchedulerTag_,
cute::enable_if_t<cute::is_base_of_v<cutlass::conv::KernelImplicitTmaWarpSpecializedSm90,
typename CollectiveMainloop_::DispatchPolicy::Schedule>>>
{
public:
//
// Type Aliases
//
// Mainloop derived types
using CollectiveMainloop = CollectiveMainloop_;
using TileShape = typename CollectiveMainloop::TileShape;
using TiledMma = typename CollectiveMainloop::TiledMma;
using ArchTag = typename CollectiveMainloop::ArchTag;
using ElementA = typename CollectiveMainloop::ElementA;
using StrideA = typename CollectiveMainloop::StrideA;
using ElementB = typename CollectiveMainloop::ElementB;
using StrideB = typename CollectiveMainloop::StrideB;
using DispatchPolicy = typename CollectiveMainloop::DispatchPolicy;
using ElementAccumulator = typename CollectiveMainloop::ElementAccumulator;
using ClusterShape = typename DispatchPolicy::ClusterShape;
using MainloopArguments = typename CollectiveMainloop::Arguments;
using MainloopParams = typename CollectiveMainloop::Params;
static constexpr int NumSpatialDimensions = CollectiveMainloop::NumSpatialDimensions;
static_assert(ArchTag::kMinComputeCapability >= 90);
// Epilogue derived types
using CollectiveEpilogue = CollectiveEpilogue_;
using ElementC = typename CollectiveEpilogue::ElementC;
using StrideC = typename CollectiveEpilogue::StrideC;
using ElementD = typename CollectiveEpilogue::ElementD;
using StrideD = typename CollectiveEpilogue::StrideD;
using EpilogueArguments = typename CollectiveEpilogue::Arguments;
using EpilogueParams = typename CollectiveEpilogue::Params;
using TileSchedulerTag = TileSchedulerTag_;
static_assert(cute::is_void_v<TileSchedulerTag>,
"TMA warp-specialized kernel does not support specializing the tile scheduler.");
using TileScheduler = typename cutlass::gemm::kernel::detail::TileSchedulerSelector<
TileSchedulerTag, ArchTag, TileShape, ClusterShape>::Scheduler;
using TileSchedulerArguments = typename TileScheduler::Arguments;
// Kernel level shared memory storage
struct SharedStorage {
union TensorStorage {
using MainloopTensorStorage = typename CollectiveMainloop::TensorStorage;
using EpilogueTensorStorage = typename CollectiveEpilogue::TensorStorage;
MainloopTensorStorage mainloop;
EpilogueTensorStorage epilogue;
} tensors;
struct PipelineStorage : cute::aligned_struct<16> {
using MainloopPipelineStorage = typename CollectiveMainloop::PipelineStorage;
using EpiLoadPipelineStorage = typename CollectiveEpilogue::PipelineStorage;
alignas(16) MainloopPipelineStorage mainloop;
alignas(16) EpiLoadPipelineStorage epi_load;
} pipelines;
};
static constexpr int SharedStorageSize = sizeof(SharedStorage);
static constexpr uint32_t NumLoadWarpGroups = 1;
static constexpr uint32_t NumMmaWarpGroups = 1;
static constexpr uint32_t MaxThreadsPerBlock = CUTE_STATIC_V(size(TiledMma{})) + (NumLoadWarpGroups * NumThreadsPerWarpGroup);
static constexpr uint32_t MinBlocksPerMultiprocessor = 1;
// Host facing host arguments
struct Arguments {
MainloopArguments mainloop{};
EpilogueArguments epilogue{};
KernelHardwareInfo hw_info{};
TileSchedulerArguments scheduler{};
};
// Kernel device entry point API
struct Params {
MainloopParams mainloop;
EpilogueParams epilogue;
};
//
// Methods
//
// Map user facing arguments to device facing params
static Params
to_underlying_arguments(Arguments const& args, void* workspace) {
(void) workspace;
auto mainloop_params = CollectiveMainloop::to_underlying_arguments(args.mainloop, workspace);
auto problem_shape_MNKL = args.mainloop.problem_shape.get_transformed_problem_shape_MNKL();
return {
mainloop_params,
CollectiveEpilogue::to_underlying_arguments(problem_shape_MNKL, args.epilogue, workspace)
};
}
// Given arguemnts, returns true if the kernel can successfully compute upon them. False otherwise.
static bool
can_implement(Arguments const& args) {
bool implementable = true;
implementable &= CollectiveMainloop::can_implement(args.mainloop.problem_shape, args.mainloop);
implementable &= CollectiveEpilogue::can_implement(args.mainloop.problem_shape.get_transformed_problem_shape_MNKL(), args.epilogue);
return implementable;
}
static size_t
get_workspace_size(Arguments const& args) {
return 0;
}
static cutlass::Status
initialize_workspace(Arguments const& args, void* workspace = nullptr, cudaStream_t stream = nullptr,
CudaHostAdapter* cuda_adapter = nullptr) {
return Status::kSuccess;
}
// Computes the kernel launch grid shape based on runtime parameters
static dim3
get_grid_shape(Params const& params) {
return cutlass::gemm::kernel::detail::PersistentTileSchedulerSm90::get_tiled_cta_shape_mnl(
params.mainloop.problem_shape, TileShape{}, ClusterShape{});
}
static dim3
get_block_shape() {
return dim3(MaxThreadsPerBlock, 1, 1);
}
CUTLASS_DEVICE
void
operator()(Params const& params, char* smem_buf) {
using namespace cute;
using X = Underscore;
// Any Tensor Op MMA Atom in the WGMMA ISA is arch conditional to sm90a.
#if ! defined(__CUDA_ARCH_FEAT_SM90_ALL)
if constexpr(size<0>(typename TiledMma::AtomShape_MNK{}) == 64) {
printf("ERROR : Arch conditional MMA instruction used without targeting sm90a compute capability. Aborting.\n");
return;
}
#endif
enum class WarpGroupRole {
Producer = 0,
Consumer = 1,
};
enum class ProducerWarpRole {
MainloopEpilogue = 0,
Warp1 = 1,
Warp2 = 2,
Warp3 = 3
};
// Kernel level shared memory storage
SharedStorage& shared_storage = *reinterpret_cast<SharedStorage*>(smem_buf);
int thread_idx = int(threadIdx.x);
int lane_idx = canonical_lane_idx();
int warp_idx = canonical_warp_idx_sync();
int warp_idx_in_warp_group = warp_idx % NumWarpsPerWarpGroup;
int warp_group_thread_idx = thread_idx % NumThreadsPerWarpGroup;
auto warp_group_role = WarpGroupRole(canonical_warp_group_idx());
auto producer_warp_role = ProducerWarpRole(warp_idx_in_warp_group);
int lane_predicate = cute::elect_one_sync();
uint32_t block_rank_in_cluster = cute::block_rank_in_cluster();
// Issue Tma Descriptor Prefetch from a single thread
if ((warp_idx == 0) && lane_predicate) {
CollectiveMainloop::prefetch_tma_descriptors(params.mainloop);
CollectiveEpilogue::prefetch_tma_descriptors(params.epilogue);
}
// Mainloop Load pipeline
using MainloopPipeline = typename CollectiveMainloop::MainloopPipeline;
typename MainloopPipeline::Params mainloop_pipeline_params;
if (warp_group_role == WarpGroupRole::Producer && producer_warp_role == ProducerWarpRole::MainloopEpilogue) {
mainloop_pipeline_params.role = MainloopPipeline::ThreadCategory::Producer;
}
if (warp_group_role == WarpGroupRole::Consumer) {
mainloop_pipeline_params.role = MainloopPipeline::ThreadCategory::Consumer;
}
mainloop_pipeline_params.is_leader = warp_group_thread_idx == 0;
mainloop_pipeline_params.num_consumers = NumThreadsPerWarpGroup;
mainloop_pipeline_params.transaction_bytes = params.mainloop.tma_transaction_bytes;
MainloopPipeline mainloop_pipeline(shared_storage.pipelines.mainloop, mainloop_pipeline_params, ClusterShape{});
// Epilogue Load pipeline
using EpiLoadPipeline = typename CollectiveEpilogue::LoadPipeline;
typename EpiLoadPipeline::Params epi_load_pipeline_params;
if (warp_group_role == WarpGroupRole::Producer && producer_warp_role == ProducerWarpRole::MainloopEpilogue) {
epi_load_pipeline_params.role = EpiLoadPipeline::ThreadCategory::Producer;
}
if (warp_group_role == WarpGroupRole::Consumer) {
epi_load_pipeline_params.role = EpiLoadPipeline::ThreadCategory::Consumer;
}
epi_load_pipeline_params.dst_blockid = cute::block_rank_in_cluster();
epi_load_pipeline_params.producer_arv_count = NumThreadsPerWarp;
epi_load_pipeline_params.consumer_arv_count = NumThreadsPerWarpGroup;
if constexpr (CollectiveEpilogue::RequiresTransactionBytes) {
epi_load_pipeline_params.transaction_bytes = params.epilogue.tma_transaction_bytes;
}
EpiLoadPipeline epi_load_pipeline(shared_storage.pipelines.epi_load, epi_load_pipeline_params);
// Epilogue Store pipeline
using EpiStorePipeline = typename CollectiveEpilogue::StorePipeline;
typename EpiStorePipeline::Params epi_store_pipeline_params;
epi_store_pipeline_params.always_wait = true;
EpiStorePipeline epi_store_pipeline(epi_store_pipeline_params);
// Initialize starting pipeline states for the collectives
// Epilogue store pipe is producer-only (consumer is TMA unit, waits via scoreboarding)
typename CollectiveMainloop::PipelineState mainloop_pipe_consumer_state;
typename CollectiveEpilogue::LoadPipelineState epi_load_pipe_consumer_state;
// For the DMA Load (producer) we start with an opposite phase
// i.e., we skip all waits since we know that the buffer is indeed empty
PipelineState mainloop_pipe_producer_state = cutlass::make_producer_start_state<MainloopPipeline>();
PipelineState epi_load_pipe_producer_state = cutlass::make_producer_start_state<EpiLoadPipeline>();
PipelineState epi_store_pipe_producer_state = cutlass::make_producer_start_state<EpiStorePipeline>();
auto cluster_wait_fn = [&] () {
// We need this to guarantee that the Pipeline init is visible
// To all producers and consumer thread blocks in the Cluster
if constexpr (size(ClusterShape{}) > 1) {
cute::cluster_arrive_relaxed();
return [] () { cute::cluster_wait(); };
}
else {
__syncthreads();
return [] () {}; // do nothing
}
} ();
// Separate out problem shape for convenience
auto problem_shape_MNKL = append<4>(params.mainloop.problem_shape, _1{});
auto [M, N, K, L] = problem_shape_MNKL;
// TMA requires special handling of strides to deal with coord codomain mapping
// Represent the full tensors -- get these from TMA
Tensor mA_mk = params.mainloop.tma_load_a.get_tma_tensor(make_shape(M, K));
Tensor mB_nk = params.mainloop.tma_load_b.get_tma_tensor(make_shape(N, K));
// Get the appropriate blocks for this thread block -- potential for thread block locality
auto cta_tile_shape = TileShape{}; // (BLK_M,BLK_N,BLK_K)
TiledMma tiled_mma;
// Make tiled views, defer the slice
Tensor gA_mk = local_tile(mA_mk, cta_tile_shape, make_coord(_,_,_), Step<_1, X,_1>{}); // (BLK_M,BLK_K,m,k)
Tensor gB_nk = local_tile(mB_nk, cta_tile_shape, make_coord(_,_,_), Step< X,_1,_1>{}); // (BLK_N,BLK_K,n,k)
// Compute m_coord, n_coord, and l_coord with their post-tiled shapes
auto m_coord = idx2crd(int(blockIdx.x), shape<2>(gA_mk));
auto n_coord = idx2crd(int(blockIdx.y), shape<2>(gB_nk), compact_col_major(shape<2>(gB_nk)));
// The output shape M is linearized so the output coord M here should also be linearized.
auto output_tile_coord = make_coord(int(blockIdx.x), n_coord, _, Int<0>{});
// Slice with m_coord and n_coord
Tensor gA = gA_mk(_,_,m_coord,_); // (BLK_M,BLK_K,k)
Tensor gB = gB_nk(_,_,n_coord,_); // (BLK_N,BLK_K,k)
// Get pipeline iterators and increments from tensor shapes
auto k_tile_iter = cute::make_coord_iterator(shape<2>(gA));
auto k_tile_count = size<2>(gA);
// In a warp specialized kernel, collectives expose data movement and compute operations separately
CollectiveMainloop collective_mainloop;
CollectiveEpilogue collective_epilogue{params.epilogue, shared_storage.tensors.epilogue};
// Wait for all thread blocks in Cluster
cluster_wait_fn();
if (warp_group_role == WarpGroupRole::Producer) {
if (producer_warp_role == ProducerWarpRole::MainloopEpilogue) {
collective_mainloop.load(
mainloop_pipeline,
mainloop_pipe_producer_state,
gA, params.mainloop.tma_load_a,
gB, params.mainloop.tma_load_b,
k_tile_iter, k_tile_count,
lane_idx,
block_rank_in_cluster,
shared_storage.tensors.mainloop
);
// Update starting mainloop pipeline state for the pipeline drain
mainloop_pipe_producer_state.advance(k_tile_count);
// Make sure mainloop consumer has been waited upon before issuing epilogue load
collective_mainloop.load_tail(mainloop_pipeline, mainloop_pipe_producer_state);
if (collective_epilogue.is_producer_load_needed()) {
epi_load_pipe_producer_state = collective_epilogue.load(
epi_load_pipeline,
epi_load_pipe_producer_state,
problem_shape_MNKL,
cta_tile_shape,
output_tile_coord,
tiled_mma,
lane_idx,
shared_storage.tensors.epilogue
);
collective_epilogue.load_tail(epi_load_pipeline, epi_load_pipe_producer_state);
}
}
}
else if (warp_group_role == WarpGroupRole::Consumer) {
Tensor accumulators = partition_fragment_C(tiled_mma, take<0,2>(cta_tile_shape)); // (MMA,MMA_M,MMA_N)
collective_mainloop.mma(
mainloop_pipeline,
mainloop_pipe_consumer_state,
accumulators,
k_tile_count,
thread_idx,
shared_storage.tensors.mainloop,
params.mainloop
);
// Make sure the math instructions are done and free buffers before entering the epilogue
collective_mainloop.mma_tail(
mainloop_pipeline,
mainloop_pipe_consumer_state,
k_tile_count
);
// Epilogue and write to gD
auto [epi_load_pipe_consumer_state_next, epi_store_pipe_producer_state_next] =
collective_epilogue.store(
epi_load_pipeline,
epi_load_pipe_consumer_state,
epi_store_pipeline,
epi_store_pipe_producer_state,
problem_shape_MNKL,
cta_tile_shape,
output_tile_coord,
accumulators,
tiled_mma,
warp_group_thread_idx,
shared_storage.tensors.epilogue
);
collective_epilogue.store_tail(
epi_load_pipeline,
epi_load_pipe_consumer_state_next,
epi_store_pipeline,
epi_store_pipe_producer_state_next
);
}
}
};
TileScheduler_,
cute::enable_if_t<cute::is_base_of_v<KernelImplicitTmaWarpSpecializedSm90, typename CollectiveMainloop_::DispatchPolicy::Schedule>>
> : public cutlass::gemm::kernel::GemmUniversal<
ProblemShape_,
CollectiveMainloop_,
CollectiveEpilogue_,
TileScheduler_
>
{};
///////////////////////////////////////////////////////////////////////////////
} // namespace cutlass::conv::kernel