ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Add mxfp4 flatmm (#3080)

Browse Source 
* Squashed commit of the following:

commit 3e1a851dad834776efbe4fe365ac82c4ed312010
Author: Ding, Yi <yi.ding@amd.com>
Date:   Thu Oct 23 06:10:54 2025 +0000

    Fix & clean after rebase

commit 1edf485092f44411da9a1796a4a6b72d5cdb67c6
Author: Ding, Yi <yi.ding@amd.com>
Date:   Wed Oct 22 10:46:13 2025 +0000

    Squashed commit of the following:

    commit 0b6b9dbd1b
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Mon Sep 22 02:04:27 2025 -0500

        fix bandwidth calculation

    commit 9aebf53bb7
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Mon Sep 22 00:58:59 2025 -0500

        updates

    commit 62607de56c
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Fri Sep 19 00:39:46 2025 -0500

        fix a bug, set the A DS_read preload size to 4 for MXFP4

    commit 92ad6fcc0a
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Thu Sep 18 01:19:03 2025 -0500

        fix a_wrap preload issue for large MPerBlock.

    commit f2db44710f
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Wed Sep 17 21:34:03 2025 -0500

        optimized the VGPR repack issue for MXFP4

    commit 346a400027
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Wed Sep 17 04:19:44 2025 -0500

        fix time error

    commit 80c1743034
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Wed Sep 17 03:58:00 2025 -0500

        updated, function passed.

    commit ce26d9071e
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Tue Sep 16 22:21:39 2025 -0500

        fix, function partially passed

    commit 0a89ed13a5
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Tue Sep 16 03:01:12 2025 -0500

        fix, reference function passed, next check kernel function

    commit ec9bcef591
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Tue Sep 16 02:29:01 2025 -0500

        let pack/unpack return pk_fp4_t

    commit a333206929
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Mon Sep 15 20:50:26 2025 -0500

        fix

    commit 3893c06540
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Mon Sep 15 05:51:06 2025 -0500

        fix bug

    commit 8052bea019
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Mon Sep 15 04:02:05 2025 -0500

        fix core dump issue, function is not correct.

    commit 9ceb3fd508
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Mon Sep 15 03:03:02 2025 -0500

        updates, build pass

    commit cc94eb6045
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Mon Sep 15 00:05:18 2025 -0500

        updates

    commit 22586c3135
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Sun Sep 14 23:40:28 2025 -0500

        fix bug

    commit e92e67b8dd
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Fri Sep 12 03:28:50 2025 -0500

        fix interface

    commit 8b1dd60c08
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Fri Sep 12 02:53:50 2025 -0500

        add interface in warp_gemm_impl

    commit c6135f6abe
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Wed Sep 10 05:03:08 2025 -0500

        updates some fixes.

    commit b0d71b8d19
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Tue Sep 9 04:37:42 2025 -0500

        fix after merge ginolu/add_wgmfma_dispatcher

    commit f119c30317
    Merge: c5030e602 72c8ef856
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Mon Sep 8 22:09:15 2025 -0500

        Merge remote-tracking branch 'origin/ginolu/add_wgmfma_dispatcher' into mtgu/cktile_mxfp4_flatmm_dev

    commit c5030e602e
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Mon Sep 8 21:42:47 2025 -0500

        update mx flatmm tail pipeline

    commit 72c8ef8567
    Merge: 9661bb400 e4a772890
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Mon Sep 8 19:10:23 2025 -0500

        Merge branch 'develop' into ginolu/add_wgmfma_dispatcher

    commit 9661bb400b
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Mon Sep 8 19:09:55 2025 -0500

        fix type error

    commit 0509597f55
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Mon Sep 8 04:01:40 2025 -0500

        update hotloop pipeline

    commit 754ae0461b
    Merge: 15d44406e 83f607e2a
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Fri Sep 5 04:22:26 2025 -0500

        Merge branch 'develop' into ginolu/add_wgmfma_dispatcher

    commit 15d44406e5
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Fri Sep 5 04:21:26 2025 -0500

        fix clang format

    commit 146963d62a
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Wed Sep 3 10:00:54 2025 -0500

        some updates

    commit 12526b626a
    Merge: 47cee0471 00fd72b2d
    Author: asleepzzz <hanwen.chang@amd.com>
    Date:   Wed Sep 3 13:22:03 2025 +0800

        Merge branch 'develop' into ginolu/add_wgmfma_dispatcher

    commit 47cee04712
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Mon Sep 1 02:11:02 2025 -0500

        fix vec size error

    commit d2892925e5
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Mon Sep 1 01:23:39 2025 -0500

        fix format error

    commit 16993acd1d
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Sat Aug 30 03:19:07 2025 -0500

        update codes

    commit 9c37e55d13
    Author: mtgu0705 <mtgu@amd.com>
    Date:   Fri Aug 29 11:27:33 2025 -0500

        init ck_tile mxfp4 flatmm

    commit 5c484a5672
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Thu Aug 28 08:02:50 2025 +0000

        Add bias for f16xf4 moe_flatmm

    commit dd6539f366
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Aug 27 13:39:47 2025 +0000

        update case construction

    commit 65b702454c
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Tue Aug 26 12:32:29 2025 +0000

        support swiglu activaion and use rcpf to accelerate silu

    commit b422e41e08
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Tue Aug 26 02:33:55 2025 -0500

        first commit

    commit d05eed931d
    Author: root <root@smci355-ccs-aus-m02-25.cs-aus.dcgpu>
    Date:   Fri Aug 22 04:01:59 2025 -0500

        add line to last

    commit d69cab7f0c
    Author: root <root@smci355-ccs-aus-m02-25.cs-aus.dcgpu>
    Date:   Fri Aug 22 03:20:46 2025 -0500

        adjust A_LDS descriptor to avoid bankconflict

    commit 65989e940c
    Author: root <root@smci355-ccs-aus-m02-25.cs-aus.dcgpu>
    Date:   Thu Aug 21 09:46:52 2025 -0500

        enable hotloop

    commit c378e9bdf8
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Thu Aug 21 09:12:21 2025 +0000

        support atomic_pk_add_bf16 on gfx950

    commit 85976b0b87
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Thu Aug 21 06:58:55 2025 +0000

        use int64_t as expert stride to avoid overflow

    commit 9fbcc8f8a4
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Aug 20 13:53:32 2025 +0000

        use v4i32 as the storage type for B to avoid repack operation

    commit 81899bd920
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Aug 20 06:40:03 2025 +0000

        add pk_fp4_t and e8m0_t support for amd_buffer_load_impl

    commit c27eb0771a
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Aug 20 04:39:14 2025 +0000

        optimize cvt_pkf4_to_f16 implementation

    commit 3ca0bd500a
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Tue Aug 19 14:56:46 2025 +0000

        optimize A_LDS descriptor to avoid bankconflict

    commit f7f0306eea
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Aug 18 18:43:37 2025 +0000

        fix gate-up when GU_NRepeat > 1

    commit be55c0f9cb
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Aug 18 17:28:11 2025 +0000

        add fp16xf4 moe

    commit 599e1f5b32
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Sun Aug 17 17:51:18 2025 +0000

        rename example

    commit 7899fb4a8d
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Fri Aug 15 06:20:46 2025 +0000

        remove additional check when e8m0->float

    commit 714b341797
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Thu Aug 14 09:34:12 2025 +0000

        eliminate repeat dequant

    commit 53e8c0c533
    Merge: 5de620895 cc9c7b9e5
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Aug 13 16:51:49 2025 +0000

        Merge remote-tracking branch 'origin/moe_flatmm' into feat-mixed_input_flatmm

    commit 5de6208952
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Aug 13 16:16:48 2025 +0000

        update f16xMXF4

    commit 732ebdee8b
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Aug 13 10:48:53 2025 +0000

        update scale-preshuffle for MXF4

    commit edb58d0680
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Aug 11 11:24:34 2025 +0000

        update

    commit cc9c7b9e58
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Aug 11 08:38:23 2025 +0000

        optimize gemm2 atomic_add pattern

    commit 200a11afc8
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Aug 11 07:59:47 2025 +0000

        update scale for mxfp4

    commit 87aed564dc
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Aug 11 07:56:14 2025 +0000

        update case construction

    commit 8b85fa6cf2
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Aug 11 06:03:06 2025 +0000

        update granularity control

    commit 1b8c7097b8
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Aug 11 03:42:46 2025 +0000

        fix TileConfig

    commit 8ba1c708dc
    Author: Gino Lu <gino.lu@amd.com>
    Date:   Thu Aug 7 21:37:28 2025 +0800

        Add e8m0 scaled convert into CK_TILE (#2617)

        * first commit

        * remove redundent code

        * modify according to comments.

        * fix type_convert error with scaled_type_convert

    commit f788d3d629
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Fri Aug 8 20:19:16 2025 +0000

        add mixed_prec fp16xfp4

    commit 3dea10a277
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Thu Aug 7 09:22:04 2025 +0000

        debug mixed_prec flatmm

    commit 0ba513b148
    Merge: 90e910f3a c0cb4d036
    Author: lalala-sh <Jiaxing.Wen@amd.com>
    Date:   Wed Aug 6 16:49:47 2025 +0800

        Merge pull request #2626 from ROCm/felix/flatmm_fix_splitk

        fix split k

    commit 6d3cbc7c0e
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Aug 6 08:33:33 2025 +0000

        add moe_flatmm

    commit c0cb4d036d
    Author: coderfeli <coderfeli@163.com>
    Date:   Wed Aug 6 02:45:31 2025 +0000

        fix split k

    commit 90e910f3a7
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Aug 4 07:16:36 2025 +0000

        fix flatmm with scaling when WarpTileM == 32

    commit aa5e008fa5
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Fri Aug 1 11:01:23 2025 +0000

        optimize scaling epilogue

    commit ac5908c0bb
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Fri Aug 1 07:28:38 2025 +0000

        fix wrong config for fp8 scaling

    commit 3f43b841d4
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Jul 30 06:20:30 2025 +0000

        prune debug message

    commit 2e5d4c74cd
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Jul 30 04:52:08 2025 +0000

        fix compile error

    commit c117a1986a
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Tue Jul 29 15:42:58 2025 +0000

        Add persistent option on flatmm for tuning

    commit a587701117
    Author: AMD-dteng <dteng@amd.com>
    Date:   Tue Jul 29 22:48:00 2025 +0800

        update pipeline v1: add atomic IGLP schedule

    commit f9e48148d2
    Author: lalala-sh <Jiaxing.Wen@amd.com>
    Date:   Thu Jul 24 09:09:27 2025 +0000

        fix error log throwing

    commit 1b6d7cf407
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Mon Jul 28 08:24:51 2025 +0000

        crz idea

    commit 5473f06461
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Sun Jul 27 11:57:38 2025 +0000

        Add permuteN optimzization when NRepeat % 2 == 0 on flatmm

    commit bfb9f4002f
    Author: sjfeng <j514681085@icloud.com>
    Date:   Sun Jul 27 17:24:08 2025 +0800

        try to remove c_shuffle_lds

    commit 1264f4d2ab
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Fri Jul 25 07:41:48 2025 +0000

        fix loop-dim mismatch and improve c_shuffle alu parallelism

    commit 1239d8a546
    Merge: 406645448 b908f5e80
    Author: lalala-sh <Jiaxing.Wen@amd.com>
    Date:   Thu Jul 24 08:46:51 2025 +0000

        merge flatmm -scale

    commit 4066454483
    Author: lalala-sh <Jiaxing.Wen@amd.com>
    Date:   Thu Jul 24 16:19:58 2025 +0800

        revert delete of inc file

    commit 68390988c9
    Author: solin <bingzhou@amd.com>
    Date:   Thu Jul 24 04:38:16 2025 +0000

        reorg  flatmm code

    commit b908f5e803
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Jul 23 19:12:31 2025 +0000

        fix flatmm syntax error on gfx950

    commit 5a1183ebbd
    Author: Feng Shijie <Shijie.Feng@amd.com>
    Date:   Wed Jul 23 19:04:22 2025 +0000

        support flatmm scaling

    commit 89fa639207
    Author: valarLip <340077269@qq.com>
    Date:   Wed Jul 23 08:44:12 2025 +0000

        merge flatmm pipe v0 from dteng_flatmm_opt

    commit 3f7d848dd3
    Author: lalala-sh <Jiaxing.Wen@amd.com>
    Date:   Wed Jul 23 15:38:12 2025 +0800

        build pass

    commit 6dacf833da
    Author: lalala-sh <Jiaxing.Wen@amd.com>
    Date:   Wed Jul 23 07:20:26 2025 +0000

         fix bug

    commit 7e1bd4b839
    Author: lalala-sh <Jiaxing.Wen@amd.com>
    Date:   Wed Jul 23 15:01:53 2025 +0800

        sync

    commit 46a538e39e
    Author: valarLip <340077269@qq.com>
    Date:   Tue Jul 22 08:09:35 2025 +0000

        adaptive scheduler instead of Macro definition

    commit 9aa3396a79
    Author: lalala-sh <Jiaxing.Wen@amd.com>
    Date:   Thu Jul 17 08:40:35 2025 +0000

        fix tail handler bug

    commit fb76450e63
    Author: lalala-sh <Jiaxing.Wen@amd.com>
    Date:   Wed Jul 16 10:12:19 2025 +0000

        merge from dteng_flatmm_opt

---------

Co-authored-by: lalala-sh <Jiaxing.Wen@amd.com>
Co-authored-by: AMD-dteng <dteng@amd.com>
Co-authored-by: solin <bingzhou@amd.com>
Co-authored-by: sjfeng <j514681085@icloud.com>
Co-authored-by: valarLip <340077269@qq.com>
Co-authored-by: asleepzzz <hanwen.chang@amd.com>
Co-authored-by: Feng Shijie <Shijie.Feng@amd.com>
Co-authored-by: coderfeli <coderfeli@163.com>
Co-authored-by: Gino Lu <gino.lu@amd.com>
Co-authored-by: mtgu0705 <mtgu@amd.com>

* Fix crash on small M

* Apply suggestion from @Copilot

---------

Co-authored-by: lalala-sh <Jiaxing.Wen@amd.com>
Co-authored-by: AMD-dteng <dteng@amd.com>
Co-authored-by: solin <bingzhou@amd.com>
Co-authored-by: sjfeng <j514681085@icloud.com>
Co-authored-by: valarLip <340077269@qq.com>
Co-authored-by: asleepzzz <hanwen.chang@amd.com>
Co-authored-by: Feng Shijie <Shijie.Feng@amd.com>
Co-authored-by: coderfeli <coderfeli@163.com>
Co-authored-by: Gino Lu <gino.lu@amd.com>
Co-authored-by: mtgu0705 <mtgu@amd.com>
This commit is contained in:
Yi DING
2025-10-31 11:29:05 +08:00
committed by GitHub
parent b387249fd9
commit e135dd518d
13 changed files with 2953 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
include/ck_tile/ops/flatmm/kernel/flatmm_kernel.hpp
View File

@@ -902,8 +902,8 @@ struct FlatmmKernel
{
const auto [iM, iN] =
TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(partition_idx);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const SplitKBatchOffset splitk_batch_offset(kargs);
// options

518
include/ck_tile/ops/flatmm/kernel/mx_flatmm_kernel.hpp Normal file
View File

@@ -0,0 +1,518 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <string>
#include "ck_tile/core.hpp"
#include "ck_tile/ops/common.hpp"
#include "ck_tile/ops/flatmm/kernel/flatmm_kernel.hpp"
namespace ck_tile {
template <typename TilePartitioner_, typename MXFlatmmPipeline_, typename EpiloguePipeline_>
struct MXFlatmmKernel : FlatmmKernel<TilePartitioner_, MXFlatmmPipeline_, EpiloguePipeline_>
{
using Underlying = FlatmmKernel<TilePartitioner_, MXFlatmmPipeline_, EpiloguePipeline_>;
using TilePartitioner = remove_cvref_t<TilePartitioner_>;
using FlatmmPipeline = remove_cvref_t<MXFlatmmPipeline_>;
using BlockGemmShape =
remove_cvref_t<typename MXFlatmmPipeline_::BlockGemmShape>; // TileFlatmmShape
using EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>;
using ALayout = remove_cvref_t<typename FlatmmPipeline::ALayout>;
using BLayout = remove_cvref_t<typename FlatmmPipeline::BLayout>;
using ELayout = remove_cvref_t<typename FlatmmPipeline::CLayout>;
using DsLayout = remove_cvref_t<typename EpiloguePipeline::DsLayout>;
using DsDataType = remove_cvref_t<typename EpiloguePipeline::DsDataType>;
static constexpr index_t KernelBlockSize = FlatmmPipeline::BlockSize;
static constexpr bool UsePersistentKernel = FlatmmPipeline::UsePersistentKernel;
using ADataType = remove_cvref_t<typename FlatmmPipeline::ADataType>;
using BDataType = remove_cvref_t<typename FlatmmPipeline::BDataType>;
// Below type is actually accumulation data type - the output of block GEMM.
using EDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>;
static constexpr int MThreadPerXdl = BlockGemmShape::WarpTile::at(number<0>{});
static constexpr int NThreadPerXdl = BlockGemmShape::WarpTile::at(number<1>{});
static constexpr int KThreadPerXdl = 64 / MThreadPerXdl;
static constexpr int APackedSize = numeric_traits<ADataType>::PackedSize;
static constexpr int BPackedSize = numeric_traits<BDataType>::PackedSize;
static constexpr int MXdlPack = FlatmmPipeline::MXdlPack;
static constexpr int NXdlPack = FlatmmPipeline::NXdlPack;
static constexpr int KXdlPack = FlatmmPipeline::KXdlPack;
static constexpr index_t NumDTensor = DsDataType::size();
static constexpr auto I0 = number<0>();
static constexpr auto I1 = number<1>();
static constexpr auto I2 = number<2>();
static constexpr auto I3 = number<3>();
static constexpr auto I4 = number<4>();
static constexpr auto I5 = number<5>();
static_assert(DsLayout::size() == DsDataType::size(),
"The size of DsLayout and DsDataType should be the same");
// using KernelArgs = FlatmmKernelArgs<DsLayout::size()>;
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "mx_flatmm_gemm", gemm_prec_str<ADataType, BDataType>, FlatmmPipeline::GetName());
// clang-format on
}
template <class ScaleM, class ScaleN>
CK_TILE_HOST static constexpr auto
GridSize(const FlatmmKernelArgs<ScaleM, ScaleN, DsDataType::size()>& kargs)
{
if constexpr(UsePersistentKernel)
{
hipDeviceProp_t prop;
int deviceId = 0; // default device
constexpr int block_size = MXFlatmmKernel::BlockSize().x;
int dync_smem_size = 0;
int maxActiveBlocksPerCU = 0;
if(hipGetDeviceProperties(&prop, deviceId) != hipSuccess)
throw std::runtime_error(std::string("hipGetDeviceProperties failed: ") +
hipGetErrorName(hipGetLastError()));
if(hipOccupancyMaxActiveBlocksPerMultiprocessor(
&maxActiveBlocksPerCU,
reinterpret_cast<void*>(
kentry<1, MXFlatmmKernel, remove_cvref_t<decltype(kargs)>>),
block_size,
dync_smem_size) != hipSuccess)
throw std::runtime_error(
std::string("hipOccupancyMaxActiveBlocksPerMultiprocessor failed: ") +
hipGetErrorName(hipGetLastError()));
const int persistent_block_size = prop.multiProcessorCount * maxActiveBlocksPerCU;
const int total_work_tile_cnt = TilePartitioner::GridSize(kargs.M, kargs.N);
// std::cout << "maxActiveBlocksPerCU: " << maxActiveBlocksPerCU
// << ", persistent_block_size: " << persistent_block_size
// << ", total_work_tile_cnt: " << total_work_tile_cnt << std::endl;
if(kargs.k_batch != 1)
throw std::runtime_error("Wrong! k_batch != 1 not supported in persistent kernel");
return dim3(min(persistent_block_size, total_work_tile_cnt), 1, kargs.k_batch);
}
else
{
return dim3(TilePartitioner::GridSize(kargs.M, kargs.N), 1, kargs.k_batch);
}
}
using SplitKBatchOffset = typename Underlying::SplitKBatchOffset;
template <memory_operation_enum DstInMemOp = memory_operation_enum::set, class KernelArgs>
CK_TILE_DEVICE static auto
MakeGemmTensorViews(const ADataType* a_ptr,
const BDataType* b_flat_ptr,
const std::array<const void*, NumDTensor>& ds_ptr,
EDataType* e_ptr,
const KernelArgs& kargs,
const SplitKBatchOffset& splitk_batch_offset)
{
const auto& a_tensor_view = [&]() {
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
{
return make_naive_tensor_view<address_space_enum::global>(
a_ptr,
make_tuple(kargs.M, splitk_batch_offset.splitted_k),
make_tuple(kargs.stride_A, 1),
number<FlatmmPipeline::GetVectorSizeA()>{},
number<1>{});
}
else
{
return make_naive_tensor_view<address_space_enum::global>(
a_ptr,
make_tuple(splitk_batch_offset.splitted_k, kargs.M),
make_tuple(kargs.stride_A, 1),
number<FlatmmPipeline::GetVectorSizeA()>{},
number<1>{});
}
}();
index_t kFlatK = kargs.K * BlockGemmShape::WarpTile::at(I1);
index_t kFlatN = kargs.N * kargs.K / kFlatK;
const auto& b_flat_tensor_view = [&]() {
return make_naive_tensor_view<address_space_enum::global>(
b_flat_ptr,
make_tuple(kFlatN, kFlatK),
make_tuple(kFlatK, 1),
number<FlatmmPipeline::GetVectorSizeB()>{},
number<1>{});
}();
const auto& ds_tensor_view = generate_tuple(
[&](auto i) {
using DiLayout = remove_cvref_t<std::tuple_element_t<i.value, DsLayout>>;
using DDataType_ = remove_cvref_t<std::tuple_element_t<i.value, DsDataType>>;
if constexpr(std::is_same_v<DiLayout, tensor_layout::gemm::RowMajor>)
{
return make_naive_tensor_view<address_space_enum::global>(
static_cast<const DDataType_*>(ds_ptr[i]),
make_tuple(kargs.M, kargs.N),
make_tuple(kargs.stride_Ds[i], 1),
number<EpiloguePipeline::GetVectorSizeD(i)>{},
number<1>{});
}
else
{
return make_naive_tensor_view<address_space_enum::global>(
static_cast<const DDataType_*>(ds_ptr[i]),
make_tuple(kargs.N, kargs.M),
make_tuple(kargs.stride_Ds[i], 1),
number<EpiloguePipeline::GetVectorSizeD(i)>{},
number<1>{});
}
},
number<NumDTensor>{});
// TODO: enable vector write for C in ColMajor
const auto& e_tensor_view = [&]() {
if constexpr(std::is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
e_ptr,
make_tuple(kargs.M, kargs.N),
make_tuple(kargs.stride_E, 1),
number<EpiloguePipeline::GetVectorSizeC()>{},
number<1>{});
}
else
{
return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
e_ptr,
make_tuple(kargs.N, kargs.M),
make_tuple(kargs.stride_E, 1),
number<1>{},
number<1>{});
}
}();
auto scale_a = kargs.scale_m_ptr;
auto scale_b = kargs.scale_n_ptr;
static constexpr int BlockScaleSize = 32; // decltype(scale_n)::GranularityK;
const auto&& scale_packs_m = integer_divide_ceil(kargs.M, (MXdlPack * MThreadPerXdl));
const auto&& scale_packs_n = integer_divide_ceil(kargs.N, (NXdlPack * NThreadPerXdl));
const auto&& scale_packs_k = kargs.K / BlockScaleSize / (KXdlPack * KThreadPerXdl);
// A scale tensor view
const auto& scale_a_tensor_view = [&]() {
// Pack 2x2 e8m0 over M/K dimension into 1 int32_t to trigger dword width load
const auto scale_a_naive_desc = make_naive_tensor_descriptor_packed(
make_tuple(scale_packs_m, scale_packs_k, KThreadPerXdl, MThreadPerXdl));
const auto scale_a_desc = transform_tensor_descriptor(
scale_a_naive_desc,
make_tuple(make_merge_transform(make_tuple(scale_packs_m, MThreadPerXdl)),
make_merge_transform(make_tuple(scale_packs_k, KThreadPerXdl))),
make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
return make_tensor_view<address_space_enum::global>(
reinterpret_cast<const int32_t*>(scale_a.ptr), scale_a_desc);
}();
// B scale tensor view
const auto& scale_b_tensor_view = [&]() {
const auto scale_b_navie_desc = make_naive_tensor_descriptor_packed(
make_tuple(scale_packs_n, scale_packs_k, KThreadPerXdl, NThreadPerXdl));
const auto scale_b_desc = transform_tensor_descriptor(
scale_b_navie_desc,
make_tuple(make_merge_transform(make_tuple(scale_packs_n, NThreadPerXdl)),
make_merge_transform(make_tuple(scale_packs_k, KThreadPerXdl))),
make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
return make_tensor_view<address_space_enum::global>(
reinterpret_cast<const int32_t*>(scale_b.ptr), scale_b_desc);
}();
return make_tuple(a_tensor_view,
b_flat_tensor_view,
ds_tensor_view,
e_tensor_view,
scale_a_tensor_view,
scale_b_tensor_view);
}
template <typename TensorView>
CK_TILE_DEVICE static auto MakeGemmPadViews(const TensorView& views)
{
const auto& a_pad_view = [&]() {
const auto& a_tensor_view = views.at(I0);
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
{
return pad_tensor_view(a_tensor_view,
make_tuple(number<TilePartitioner::MPerBlock>{},
number<TilePartitioner::KPerBlock>{}),
sequence<false, FlatmmPipeline::kPadK>{});
}
else
{
return pad_tensor_view(a_tensor_view,
make_tuple(number<TilePartitioner::KPerBlock>{},
number<TilePartitioner::MPerBlock>{}),
sequence<false, FlatmmPipeline::kPadM>{});
}
}();
const auto& b_flat_tensor_view = views.at(I1);
const auto& ds_pad_view = generate_tuple(
[&](auto i) {
const auto& d_tensor_view = views.at(I2);
using DiLayout = remove_cvref_t<std::tuple_element_t<i.value, DsLayout>>;
if constexpr(std::is_same_v<DiLayout, tensor_layout::gemm::RowMajor>)
{
return pad_tensor_view(d_tensor_view[i],
make_tuple(number<TilePartitioner::MPerBlock>{},
number<TilePartitioner::NPerBlock>{}),
sequence<false, FlatmmPipeline::kPadN>{});
}
else
{
return pad_tensor_view(d_tensor_view[i],
make_tuple(number<TilePartitioner::NPerBlock>{},
number<TilePartitioner::MPerBlock>{}),
sequence<false, FlatmmPipeline::kPadM>{});
}
},
number<NumDTensor>{});
// TODO vector write in for C in ColMajor
const auto& e_pad_view = [&]() {
const auto& e_tensor_view = views.at(I3);
if constexpr(std::is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
return pad_tensor_view(e_tensor_view,
make_tuple(number<TilePartitioner::MPerBlock>{},
number<TilePartitioner::NPerBlock>{}),
sequence<false, FlatmmPipeline::kPadN>{});
}
else
{
return pad_tensor_view(e_tensor_view,
make_tuple(number<TilePartitioner::MPerBlock>{},
number<TilePartitioner::NPerBlock>{}),
sequence<FlatmmPipeline::kPadM, false>{});
}
}();
return make_tuple(
a_pad_view, b_flat_tensor_view, ds_pad_view, e_pad_view, views.at(I4), views.at(I5));
}
template <typename PadView>
CK_TILE_DEVICE static auto
MakeGemmTileWindows(const PadView& views, const index_t i_m, const index_t i_n)
{
const auto& a_pad_view = views.at(I0);
const auto& b_flat_pad_view = views.at(I1);
const auto& ds_pad_view = views.at(I2);
const auto& e_pad_view = views.at(I3);
const auto& a_block_window = [&]() {
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
{
return make_tile_window(a_pad_view,
make_tuple(number<TilePartitioner::MPerBlock>{},
number<TilePartitioner::KPerBlock>{}),
{i_m, 0});
}
else
{
return make_tile_window(a_pad_view,
make_tuple(number<TilePartitioner::KPerBlock>{},
number<TilePartitioner::MPerBlock>{}),
{0, i_m});
}
}();
const auto& b_flat_block_window =
make_tile_window(b_flat_pad_view,
make_tuple(number<FlatmmPipeline::flatNPerWarp>{},
number<FlatmmPipeline::flatKPerWarp>{}),
{static_cast<int>(i_n / BlockGemmShape::WarpTile::at(I1)), 0});
const auto ds_block_window = generate_tuple(
[&](auto i) {
using DiLayout = remove_cvref_t<std::tuple_element_t<i.value, DsLayout>>;
if constexpr(std::is_same_v<DiLayout, tensor_layout::gemm::RowMajor>)
{
return make_tile_window(ds_pad_view[i],
make_tuple(number<TilePartitioner::MPerBlock>{},
number<TilePartitioner::NPerBlock>{}),
{i_m, i_n});
}
else
{
return make_tile_window(ds_pad_view[i],
make_tuple(number<TilePartitioner::NPerBlock>{},
number<TilePartitioner::MPerBlock>{}),
{i_n, i_m});
}
},
number<NumDTensor>{});
auto e_block_window = make_tile_window(
e_pad_view,
make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::NPerBlock>{}),
{i_m, i_n});
static constexpr int BlockScaleSize = 32;
auto scale_a_block_window = make_tile_window(
views.at(I4),
make_tuple(number<TilePartitioner::MPerBlock / MXdlPack>{},
number<TilePartitioner::KPerBlock / (BlockScaleSize * KXdlPack)>{}),
{i_m / MXdlPack, 0});
auto scale_b_block_window = make_tile_window(
views.at(I5),
make_tuple(number<TilePartitioner::NPerBlock / NXdlPack>{},
number<TilePartitioner::KPerBlock / (BlockScaleSize * KXdlPack)>{}),
{i_n / NXdlPack, 0});
return make_tuple(a_block_window,
b_flat_block_window,
ds_block_window,
e_block_window,
scale_a_block_window,
scale_b_block_window);
}
template <class ScaleM, class ScaleN, bool UseDefaultScheduler = true>
CK_TILE_DEVICE static void
RunFlatmm(const ADataType* a_ptr,
const BDataType* b_flat_ptr,
const std::array<const void*, NumDTensor>& ds_ptr,
EDataType* e_ptr,
void* smem_ptr_ping,
void* smem_ptr_pong,
const FlatmmKernelArgs<ScaleM, ScaleN, DsDataType::size()>& kargs,
const SplitKBatchOffset& splitk_batch_offset,
const index_t block_idx_m,
const index_t block_idx_n)
{
// Create Gemm tensor views, pad views and tile windows
const auto& gemm_tensor_views_tuple =
MakeGemmTensorViews<EpiloguePipeline::MemoryOperation>(
a_ptr, b_flat_ptr, ds_ptr, e_ptr, kargs, splitk_batch_offset);
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop = TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k);
// Run GEMM cooperatively by whole workgroup.
const auto& a_block_window = gemm_tile_windows.at(I0);
const auto& b_flat_block_window = gemm_tile_windows.at(I1);
const auto& d_block_window = gemm_tile_windows.at(I2);
const auto& scale_a_block_window = gemm_tile_windows.at(I4);
const auto& scale_b_block_window = gemm_tile_windows.at(I5);
static_assert(ScaleM::GranularityK == ScaleN::GranularityK // have the same granK
|| ScaleM::GranularityMN == -1 // or ScaleA is disable
|| ScaleN::GranularityMN == -1, // or ScaleB is disable
"ScaleM and ScaleN should have the same GranularityK");
constexpr bool DoEpiScale =
(ScaleM::GranularityMN != -1 && ScaleM::GranularityK == 0) || // per token
(ScaleN::GranularityMN != -1 && ScaleN::GranularityK == 0); // per channel
auto a_block_window_with_distr =
ck_tile::make_tile_window(a_block_window.get_bottom_tensor_view(),
a_block_window.get_window_lengths(),
a_block_window.get_window_origin(),
FlatmmPipeline::GetADramTileDistribution());
const auto& c_block_tile = FlatmmPipeline{}(a_block_window_with_distr,
b_flat_block_window,
scale_a_block_window,
scale_b_block_window,
num_loop,
smem_ptr_ping,
smem_ptr_pong);
// Run Epilogue Pipeline
if constexpr(DoEpiScale)
{
auto& c_block_window = gemm_tile_windows.at(I3);
EpiloguePipeline{}(c_block_window,
c_block_tile,
d_block_window,
smem_ptr_ping,
kargs.scale_m_ptr + block_idx_m,
kargs.scale_n_ptr + block_idx_n);
}
else if(UseDefaultScheduler || (get_warp_id() == 0))
{
// Run Epilogue Pipeline
auto& c_block_window = gemm_tile_windows.at(I3);
EpiloguePipeline{}(c_block_window, c_block_tile, d_block_window, smem_ptr_ping);
}
}
template <class ScaleM, class ScaleN>
CK_TILE_DEVICE void operator()(FlatmmKernelArgs<ScaleM, ScaleN, DsDataType::size()> kargs,
int partition_idx = blockIdx.x) const
{
int total_work_tile_cnt = TilePartitioner::GridSize(kargs.M, kargs.N);
do
{
const auto [iM, iN] =
TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(partition_idx);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const SplitKBatchOffset splitk_batch_offset(kargs);
// options
const ADataType* a_ptr = static_cast<const ADataType*>(kargs.a_ptr) +
splitk_batch_offset.a_k_split_offset / APackedSize;
const BDataType* b_flat_ptr = static_cast<const BDataType*>(kargs.b_ptr) +
splitk_batch_offset.b_k_split_offset / BPackedSize;
EDataType* e_ptr = static_cast<EDataType*>(kargs.e_ptr);
// allocate LDS
__shared__ char smem_ptr_ping[Underlying::GetSmemPingSize()];
__shared__ char smem_ptr_pong[Underlying::GetSmemPongSize()];
if constexpr(!(EpiloguePipeline::MemoryOperation == memory_operation_enum::atomic_add &&
EpiloguePipeline::GetVectorSizeC() % 2 != 0 &&
is_any_of<EDataType, fp16_t, bf16_t>::value))
{
constexpr auto scheduler_type = (FlatmmPipeline::NumWaveGroups == 1);
RunFlatmm<ScaleM, ScaleN, scheduler_type>(a_ptr,
b_flat_ptr,
kargs.ds_ptr,
e_ptr,
smem_ptr_ping,
smem_ptr_pong,
kargs,
splitk_batch_offset,
i_m,
i_n);
}
else
{
static_assert(false,
"Unimplemented: atomic_add with odd vector size for fp16/bf16");
}
partition_idx += gridDim.x;
} while(UsePersistentKernel && partition_idx < total_work_tile_cnt);
}
};
} // namespace ck_tile

8
include/ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1_policy.hpp
View File

@@ -291,10 +291,12 @@ struct UniversalFlatmmPipelineAgBgCrPolicy
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t APackedSize = numeric_traits<ADataType>::PackedSize;
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
constexpr index_t M1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t M0 = MPerBlock / M1;
constexpr index_t M1 = Problem::VectorLoadSize / sizeof(ADataType) * APackedSize;
constexpr index_t M0 = MPerBlock / M1;
constexpr index_t total_pixels = MPerBlock * KPerBlock / BlockSize;
static_assert(total_pixels % M1 == 0);
constexpr index_t K3 = total_pixels / M1;
@@ -331,7 +333,7 @@ struct UniversalFlatmmPipelineAgBgCrPolicy
}
else
{
constexpr index_t K1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t K1 = Problem::VectorLoadSize / sizeof(ADataType) * APackedSize;
constexpr index_t K0 = KPerBlock / K1;
// coalesce reading for each blocks
if constexpr(get_warp_size() % K0 == 0)

1330
include/ck_tile/ops/flatmm/pipeline/mx_flatmm_pipeline_agmem_bgmem_creg_v1.hpp Normal file
View File

File diff suppressed because it is too large Load Diff

275
include/ck_tile/ops/flatmm/pipeline/mx_flatmm_pipeline_agmem_bgmem_creg_v1_policy.hpp Normal file
View File

@@ -0,0 +1,275 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1_policy.hpp"
namespace ck_tile {
struct MXF4FlatmmPipelineAgBgCrPolicy : UniversalFlatmmPipelineAgBgCrPolicy
{
static constexpr auto I0 = number<0>{};
static constexpr auto I1 = number<1>{};
static constexpr auto I2 = number<2>{};
static constexpr index_t KBPerLoad = 32;
static constexpr int MXdlPack = 2;
static constexpr int NXdlPack = 2;
static constexpr int KXdlPack = 2;
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeMXFP4_ALdsBlockDescriptor()
{
using namespace ck_tile;
using ADataType = remove_cvref_t<typename Problem::ADataType>;
using ALayout = remove_cvref_t<typename Problem::ALayout>;
constexpr index_t MPerXdl = Problem::BlockGemmShape::WarpTile::at(I0);
constexpr index_t NPerXdl = Problem::BlockGemmShape::WarpTile::at(I1);
static_assert(MPerXdl == 16 && NPerXdl == 16);
static_assert(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>);
/*reduce transform layers,compare with old ck*/
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t APackedSize = numeric_traits<ADataType>::PackedSize;
constexpr index_t KPack = GetSmemPackA<Problem>() * APackedSize;
constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<KPerBlock / KPack>{}, number<MPerBlock>{}, number<KPack>{}),
make_tuple(number<KPack>{}, number<KPerBlock>{}, number<1>{}),
number<KPack>{},
number<1>{});
constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
a_lds_block_desc_0,
make_tuple(
make_xor_transform(make_tuple(number<MPerBlock>{}, number<KPerBlock / KPack>{})),
make_pass_through_transform(number<KPack>{})),
make_tuple(sequence<1, 0>{}, sequence<2>{}),
make_tuple(sequence<1, 0>{}, sequence<2>{}));
constexpr auto a_lds_block_desc = transform_tensor_descriptor(
a_lds_block_desc_permuted,
make_tuple(make_pass_through_transform(number<MPerBlock>{}),
make_merge_transform_v3_division_mod(
make_tuple(number<KPerBlock / KPack>{}, number<KPack>{}))),
make_tuple(sequence<1>{}, sequence<0, 2>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
// return a_lds_block_desc_permuted;
return a_lds_block_desc;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeMXFP4_ADramTileDistribution()
{
using ADataType = remove_cvref_t<typename Problem::ADataType>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t K1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t K0 = KPerBlock / K1;
constexpr index_t M2 = get_warp_size() / K0;
constexpr index_t M1 = BlockSize / get_warp_size();
static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
constexpr index_t M0 = MPerBlock / (M2 * M1);
static_assert(M0 * M1 * M2 == MPerBlock,
"Incorrect M0, M2, M1 configuration! "
"M0, M1, M2 must cover whole MPerBlock!");
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeMXF4_ALDS_TileDistribution()
{
using TileShape = typename Problem::BlockGemmShape;
static_assert(TileShape::WarpTile::at(I1) == 16, "requires XDL_N == 16");
static_assert(TileShape::BlockWarps::at(I0) == 1, "requires Wave_M == 1");
constexpr int M_warps = TileShape::BlockWarps::at(number<0>{});
constexpr int N_warps = TileShape::BlockWarps::at(number<1>{});
constexpr int M_Lane = TileShape::WarpTile::at(I0);
constexpr int K_Lane = 64 / TileShape::WarpTile::at(I0); // 4
constexpr int K1 = TileShape::WarpTile::at(I2) / K_Lane; // 32
return make_static_tile_distribution(
tile_distribution_encoding<
sequence<N_warps>,
tuple<sequence<M_warps, MXdlPack, M_Lane>, sequence<K_Lane, K1>>,
tuple<sequence<1, 0>, sequence<2, 1>>,
tuple<sequence<0, 0>, sequence<0, 2>>,
sequence<2>,
sequence<1>>{});
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeMXFP4_BFlatDramTileDistribution()
{
using TileShape = typename Problem::BlockGemmShape;
static_assert(TileShape::WarpTile::at(I1) == 16, "only for XDL_N == 16");
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t WaveSize = get_warp_size();
constexpr index_t WaveNum = BlockSize / WaveSize;
constexpr index_t KThdPerWave = WaveSize; // threads cnt in K dim
constexpr index_t KWavePerBlk = 1;
constexpr index_t NWavePerBlk = TileShape::BlockWarps::at(number<1>{}); // N_Warp
constexpr index_t WaveRepeat = WaveNum / TileShape::flatNPerWarp;
return make_static_tile_distribution(
tile_distribution_encoding<
sequence<WaveRepeat>,
tuple<sequence<NWavePerBlk, NXdlPack>,
sequence<KWavePerBlk, KThdPerWave, KBPerLoad>>, // first direction
// wave in blk, // thd in wave
// <M, K> // <M, K>
tuple<sequence<0, 1, 2>, sequence<2>>, // which direction
tuple<sequence<0, 0, 0>, sequence<1>>, // which index
// <repeat, vec_load>
sequence<2>,
sequence<2>>{});
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeMXFP4_ScaleA_DramTileDistribution()
{
using TileShape = typename Problem::BlockGemmShape; // ck_tile::TileFlatmmShape
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t WaveSize = get_warp_size();
constexpr index_t WaveNum = BlockSize / WaveSize;
constexpr index_t kMPerBlock = TileShape::BlockTile::at(I0);
constexpr index_t M_Warps = TileShape::BlockWarps::at(I0);
constexpr index_t N_Warps = TileShape::BlockWarps::at(I1);
static_assert(WaveNum == M_Warps * N_Warps, "Block warps do not match block size");
constexpr index_t M_Lanes = TileShape::WarpTile::at(I0);
constexpr index_t K_Lanes = 64 / M_Lanes;
// Y dimension (M) decomposition
constexpr index_t Y2 = M_Lanes;
constexpr index_t Y1 = M_Warps;
constexpr index_t Y0 = kMPerBlock / (MXdlPack * Y1 * Y2);
// X dimension (K) decomposition
constexpr index_t X0 = K_Lanes;
constexpr index_t X1 = 1; // packed 2x2 E8M0 data into 1 int32_t for load
return make_static_tile_distribution(
tile_distribution_encoding<sequence<N_Warps>, // repeat N_warps
tuple<sequence<Y0, Y1, Y2>, sequence<X0, X1>>,
tuple<sequence<1, 0>, sequence<2, 1>>,
tuple<sequence<1, 0>, sequence<0, 2>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeMXFP4_ScaleB_DramTileDistribution()
{
using TileShape = typename Problem::BlockGemmShape; // ck_tile::TileFlatmmShape
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t WaveSize = get_warp_size();
constexpr index_t WaveNum = BlockSize / WaveSize;
constexpr index_t kNPerBlock = TileShape::BlockTile::at(I1);
constexpr index_t M_Warps = TileShape::BlockWarps::at(I0);
constexpr index_t N_Warps = TileShape::BlockWarps::at(I1);
static_assert(WaveNum == M_Warps * N_Warps, "Block warps do not match block size");
constexpr index_t N_Lanes = TileShape::WarpTile::at(I1);
constexpr index_t K_Lanes = 64 / N_Lanes;
// Y dimension (M) decomposition
constexpr index_t Y2 = N_Lanes;
constexpr index_t Y1 = N_Warps;
constexpr index_t Y0 = kNPerBlock / (NXdlPack * Y1 * Y2);
// X dimension (K) decomposition
constexpr index_t X0 = K_Lanes;
constexpr index_t X1 = 1; // packed 2x2 E8M0 data into 1 int32_t for load
return make_static_tile_distribution(
tile_distribution_encoding<sequence<M_Warps>, // ?
tuple<sequence<Y0, Y1, Y2>, sequence<X0, X1>>,
tuple<sequence<0, 1>, sequence<2, 1>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeMXFP4_ScaleA_FlatDramTileDistribution()
{
using TileShape = typename Problem::BlockGemmShape;
constexpr index_t M_Warp = TileShape::BlockWarps::at(number<0>{});
constexpr index_t K_Lane = 64 / TileShape::WarpTile::at(I0);
constexpr index_t M_Lane = TileShape::WarpTile::at(I0);
constexpr index_t N_Wrap = TileShape::BlockWarps::at(number<1>{});
constexpr index_t MWavePerBlk = M_Warp;
return make_static_tile_distribution(
tile_distribution_encoding<sequence<N_Wrap>, // ?
tuple<sequence<MWavePerBlk, M_Lane>, // second direction
sequence<K_Lane, 1>>, // first direction
tuple<sequence<1, 0>, sequence<2, 1>>, // which direction
tuple<sequence<0, 0>, sequence<0, 1>>, // which index
// <repeat, vec_load>
sequence<2>,
sequence<1>>{});
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeMXFP4_ScaleB_FlatDramTileDistribution()
{
using TileShape = typename Problem::BlockGemmShape;
constexpr index_t N_Warp = TileShape::BlockWarps::at(number<1>{});
constexpr index_t K_Lane = 64 / TileShape::WarpTile::at(I1);
constexpr index_t N_Lane = TileShape::WarpTile::at(I1);
constexpr index_t M_Wrap = TileShape::BlockWarps::at(number<0>{});
constexpr index_t NWavePerBlk = N_Warp;
return make_static_tile_distribution(
tile_distribution_encoding<sequence<M_Wrap>, // ?
tuple<sequence<NWavePerBlk, N_Lane>, // second direction
sequence<K_Lane, 1>>, // first direction
tuple<sequence<0, 1>, sequence<2, 1>>, // which direction
tuple<sequence<0, 0>, sequence<0, 1>>, // which index
// <repeat, vec_load>
sequence<2>,
sequence<1>>{});
}
};
} // namespace ck_tile