296 lines
11 KiB
C++
296 lines
11 KiB
C++
/***************************************************************************************************
|
|
* 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/arch/arch.h"
|
|
#include "cutlass/gemm/gemm.h"
|
|
|
|
#include "cute/layout.hpp"
|
|
#include "cute/numeric/integral_constant.hpp"
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace cutlass::detail {
|
|
|
|
template <class T, template <int...> class U>
|
|
struct is_kernel_tag_of : cute::false_type {};
|
|
|
|
template <template <int...> class U, int... Args>
|
|
struct is_kernel_tag_of<U<Args...>, U> : cute::true_type {};
|
|
|
|
template <class T, template <int...> class U>
|
|
constexpr bool is_kernel_tag_of_v = is_kernel_tag_of<T, U>::value;
|
|
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace cutlass::gemm {
|
|
using namespace cute;
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace detail {
|
|
|
|
enum class KernelInputTransformType {
|
|
FastF32,
|
|
InterleavedComplexTF32
|
|
};
|
|
|
|
} // namespace detail
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace kernel::detail {
|
|
|
|
// Has_SwapAB<T>::value will be true only if:
|
|
// class T has member SwapAB and T::SwapAB is true
|
|
template <typename T, typename = void>
|
|
struct Has_SwapAB { static constexpr bool value = false; };
|
|
|
|
template <typename T>
|
|
struct Has_SwapAB <T, CUTE_STL_NAMESPACE::void_t<decltype(T::SwapAB)>>
|
|
{ static constexpr bool value = T::SwapAB; };
|
|
|
|
template <typename T>
|
|
static constexpr bool Has_SwapAB_v = Has_SwapAB<T>::value;
|
|
|
|
} // namespace kernel::detail
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
|
|
//
|
|
// Kernel schedule policies (the base class tags, one for each kernel layer file)
|
|
//
|
|
struct KernelMultistage { };
|
|
struct KernelCpAsyncWarpSpecialized { };
|
|
struct KernelCpAsyncWarpSpecializedPingpong { };
|
|
struct KernelCpAsyncWarpSpecializedCooperative { };
|
|
struct KernelTma { };
|
|
struct KernelTmaWarpSpecialized { };
|
|
struct KernelTmaWarpSpecializedPingpong { };
|
|
struct KernelTmaWarpSpecializedCooperative { };
|
|
struct KernelPtrArrayTmaWarpSpecializedCooperative { };
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
|
|
//
|
|
// Builder dispatch policies (not a part of the main CUTLASS layers, simply used to opt into
|
|
// specific collective builder dispatches)
|
|
//
|
|
|
|
// FP8 related policies (including Fast Accumulation)
|
|
struct KernelTmaWarpSpecializedFP8FastAccum : KernelTmaWarpSpecialized { };
|
|
struct KernelTmaWarpSpecializedPingpongFP8FastAccum : KernelTmaWarpSpecializedPingpong { };
|
|
struct KernelTmaWarpSpecializedCooperativeFP8FastAccum: KernelTmaWarpSpecializedCooperative { };
|
|
struct KernelPtrArrayTmaWarpSpecializedCooperativeFP8FastAccum : KernelPtrArrayTmaWarpSpecializedCooperative { };
|
|
|
|
// Policies to opt into mixed type GEMMs
|
|
struct KernelTmaWarpSpecializedMixedInput : KernelTmaWarpSpecialized { };
|
|
struct KernelTmaWarpSpecializedPingpongMixedInput : KernelTmaWarpSpecializedPingpong { };
|
|
struct KernelTmaWarpSpecializedCooperativeMixedInput: KernelTmaWarpSpecializedCooperative { };
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
|
|
// Policies for dispatch of epilogue
|
|
struct EpilogueDefault { };
|
|
struct EpilogueTransposed { };
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
|
|
//
|
|
// Collective Mainloop Policies
|
|
//
|
|
|
|
// 2 stage pipeline through 1 stage in smem, 1 in rmem, WITHOUT predicated gmem loads
|
|
struct MainloopSm70TwoStageUnpredicated {
|
|
constexpr static int Stages = 2;
|
|
using ArchTag = arch::Sm70;
|
|
using Schedule = KernelMultistage;
|
|
using ClusterShape = Shape<_1,_1,_1>;
|
|
};
|
|
|
|
// 2 stage pipeline through 1 stage in smem, 1 in rmem, with predicated gmem loads
|
|
struct MainloopSm70TwoStage {
|
|
constexpr static int Stages = 2;
|
|
using ArchTag = arch::Sm70;
|
|
using Schedule = KernelMultistage;
|
|
using ClusterShape = Shape<_1,_1,_1>;
|
|
};
|
|
|
|
// n-buffer in smem (cp.async), pipelined with registers, WITHOUT predicated gmem loads
|
|
template<int Stages_>
|
|
struct MainloopSm80CpAsyncUnpredicated {
|
|
constexpr static int Stages = Stages_;
|
|
using ArchTag = arch::Sm80;
|
|
using Schedule = KernelMultistage;
|
|
using ClusterShape = Shape<_1,_1,_1>;
|
|
};
|
|
|
|
// n-buffer in smem (cp.async), pipelined with registers, with predicated gmem loads
|
|
template<
|
|
int Stages_,
|
|
class ClusterShape_ = Shape<_1,_1,_1>
|
|
>
|
|
struct MainloopSm80CpAsync {
|
|
constexpr static int Stages = Stages_;
|
|
using ArchTag = cute::conditional_t<(size(ClusterShape_{}) > 1), arch::Sm90, arch::Sm80>;
|
|
using Schedule = KernelMultistage;
|
|
using ClusterShape = ClusterShape_;
|
|
};
|
|
|
|
// n-buffer in smem (cp.async), pipelined with Hopper GMMA, with predicated gmem loads, warp specialized dynamic schedule
|
|
template<
|
|
int Stages_,
|
|
class ClusterShape_ = Shape<_1,_1,_1>,
|
|
class KernelSchedule = KernelCpAsyncWarpSpecialized
|
|
>
|
|
struct MainloopSm90CpAsyncGmmaWarpSpecialized {
|
|
constexpr static int Stages = Stages_;
|
|
using ClusterShape = ClusterShape_;
|
|
using ArchTag = arch::Sm90;
|
|
using Schedule = KernelSchedule;
|
|
};
|
|
|
|
// n-buffer in smem (cp.async), pipelined with Hopper GMMA, with predicated gmem loads, warp specialized dynamic schedule
|
|
template<
|
|
int Stages_,
|
|
class ClusterShape_ = Shape<_1,_1,_1>,
|
|
class KernelSchedule = KernelCpAsyncWarpSpecialized
|
|
>
|
|
struct MainloopSm90CpAsyncGmmaRmemAWarpSpecialized {
|
|
constexpr static int Stages = Stages_;
|
|
using ClusterShape = ClusterShape_;
|
|
using ArchTag = arch::Sm90;
|
|
using Schedule = KernelSchedule;
|
|
};
|
|
|
|
// n-buffer in smem (Hopper TMA), pipelined with Hopper GMMA and TMA, static schedule between TMA and GMMA
|
|
template<
|
|
int Stages_,
|
|
class ClusterShape_ = Shape<_1,_1,_1>,
|
|
int PipelineAsyncMmaStages_ = 1
|
|
>
|
|
struct MainloopSm90TmaGmma {
|
|
constexpr static int Stages = Stages_;
|
|
using ClusterShape = ClusterShape_;
|
|
constexpr static int PipelineAsyncMmaStages = PipelineAsyncMmaStages_;
|
|
using ArchTag = arch::Sm90;
|
|
using Schedule = KernelTma;
|
|
};
|
|
|
|
// n-buffer in smem (Hopper TMA), pipelined with Hopper GMMA and TMA, Warp specialized dynamic schedule
|
|
template<
|
|
int Stages_,
|
|
class ClusterShape_ = Shape<_1,_1,_1>,
|
|
class KernelSchedule = KernelTmaWarpSpecializedCooperative
|
|
>
|
|
struct MainloopSm90TmaGmmaWarpSpecialized {
|
|
constexpr static int Stages = Stages_;
|
|
using ClusterShape = ClusterShape_;
|
|
using ArchTag = arch::Sm90;
|
|
using Schedule = KernelSchedule;
|
|
};
|
|
|
|
// n-buffer in smem (Hopper TMA), pipelined with Hopper GMMA and TMA, Warp specialized dynamic schedule
|
|
// With GMMA's A data from registers.
|
|
template<
|
|
int Stages_,
|
|
class ClusterShape_ = Shape<_1,_1,_1>,
|
|
class KernelSchedule = KernelTmaWarpSpecialized
|
|
>
|
|
struct MainloopSm90TmaGmmaRmemAWarpSpecialized {
|
|
constexpr static int Stages = Stages_;
|
|
using ClusterShape = ClusterShape_;
|
|
using ArchTag = arch::Sm90;
|
|
using Schedule = KernelSchedule;
|
|
static_assert(
|
|
cute::is_same_v<Schedule, KernelTmaWarpSpecialized> ||
|
|
cute::is_same_v<Schedule, KernelTmaWarpSpecializedPingpong> ||
|
|
cute::is_same_v<Schedule, KernelTmaWarpSpecializedCooperative>,
|
|
"KernelSchedule must be one of the warp specialized policies");
|
|
};
|
|
|
|
template<
|
|
int Stages_,
|
|
class ClusterShape_ = Shape<_1,_1,_1>,
|
|
class KernelSchedule = KernelTmaWarpSpecialized
|
|
>
|
|
struct MainloopSm90TmaGmmaRmemAWarpSpecializedMixedInput {
|
|
constexpr static int Stages = Stages_;
|
|
using ClusterShape = ClusterShape_;
|
|
using ArchTag = arch::Sm90;
|
|
using Schedule = KernelSchedule;
|
|
static_assert(
|
|
cute::is_same_v<Schedule, KernelTmaWarpSpecialized> ||
|
|
cute::is_same_v<Schedule, KernelTmaWarpSpecializedMixedInput> ||
|
|
cute::is_same_v<Schedule, KernelTmaWarpSpecializedPingpong> ||
|
|
cute::is_same_v<Schedule, KernelTmaWarpSpecializedPingpongMixedInput> ||
|
|
cute::is_same_v<Schedule, KernelTmaWarpSpecializedCooperative> ||
|
|
cute::is_same_v<Schedule, KernelTmaWarpSpecializedCooperativeMixedInput>,
|
|
"KernelSchedule must be one of the warp specialized policies");
|
|
};
|
|
|
|
// n-buffer in smem (Hopper TMA), pipelined with Hopper GMMA and TMA, Warp specialized dynamic schedule
|
|
// For FP8 kernels
|
|
template<
|
|
int Stages_,
|
|
class ClusterShape_ = Shape<_1,_1,_1>,
|
|
class KernelSchedule = KernelTmaWarpSpecialized
|
|
>
|
|
struct MainloopSm90TmaGmmaWarpSpecializedFP8
|
|
: MainloopSm90TmaGmmaWarpSpecialized<Stages_, ClusterShape_, KernelSchedule> {
|
|
static_assert(
|
|
cute::is_same_v<KernelSchedule, KernelTmaWarpSpecialized> ||
|
|
cute::is_same_v<KernelSchedule, KernelTmaWarpSpecializedPingpong> ||
|
|
cute::is_same_v<KernelSchedule, KernelTmaWarpSpecializedCooperative>,
|
|
"KernelSchedule must be one of the warp specialized policies");
|
|
};
|
|
|
|
// n-buffer in smem (Hopper TMA), pipelined with Hopper GMMA and TMA, Warp specialized dynamic schedule for Ptr-Array and Grouped Gemm
|
|
template<
|
|
int Stages_,
|
|
class ClusterShape_ = Shape<_1,_1,_1>,
|
|
class KernelSchedule = KernelPtrArrayTmaWarpSpecializedCooperative
|
|
>
|
|
struct MainloopSm90ArrayTmaGmmaWarpSpecialized {
|
|
constexpr static int Stages = Stages_;
|
|
using ClusterShape = ClusterShape_;
|
|
using ArchTag = arch::Sm90;
|
|
using Schedule = KernelSchedule;
|
|
static_assert(
|
|
cute::is_base_of_v<KernelPtrArrayTmaWarpSpecializedCooperative, KernelSchedule>,
|
|
"KernelSchedule must be one of the Ptr-Array or Grouped Gemm TMA Warp Specialized Cooperative policies");
|
|
};
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
|
|
} // namespace cutlass::gemm
|