mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-17 03:19:48 +00:00
update codes
This commit is contained in:
mtgu0705
@@ -71,6 +71,91 @@ CK_TILE_HOST void reference_gemm(const HostTensor<ADataType>& a_m_k,
|
||||
make_ParallelTensorFunctor(f_mn, M, N)(std::thread::hardware_concurrency());
|
||||
}
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename ScaleDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename AElementOp = ck_tile::identity,
|
||||
typename BElementOp = ck_tile::identity,
|
||||
typename ACCElementOp = ck_tile::identity>
|
||||
CK_TILE_HOST void reference_mx_gemm(const HostTensor<ADataType>& a_m_k,
|
||||
const HostTensor<BDataType>& b_k_n,
|
||||
HostTensor<CDataType>& c_m_n,
|
||||
const HostTensor<ScaleDataType>& scale_a,
|
||||
const HostTensor<ScaleDataType>& scale_b,
|
||||
const AElementOp& a_element_op = {},
|
||||
const BElementOp& b_element_op = {},
|
||||
const ACCElementOp& acc_element_op = {})
|
||||
{
|
||||
const std::size_t M = a_m_k.get_length(0);
|
||||
const std::size_t N = b_k_n.get_length(1);
|
||||
const std::size_t K = a_m_k.get_length(1);
|
||||
|
||||
const std::size_t ScaleBlockSize = K / a_m_k_scale.get_length(1);
|
||||
|
||||
HostTensor<AccDataType> a_m_k_scaled({M, K}, {K, 1});
|
||||
HostTensor<AccDataType> b_k_n_scaled({K, N}, {1, N});
|
||||
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
if constexpr(std::is_same_v<ADataType, f4x2_pk_t>)
|
||||
{
|
||||
if(k % 2 == 1)
|
||||
continue; // skip odd k
|
||||
|
||||
auto a_f4x2 = a_m_k(m, k);
|
||||
auto a_scale = a_m_k_scale(m, k / ScaleBlockSize);
|
||||
// auto f4_lo = ck_tile::type_convert<AccDataType>(f4x2)[0];
|
||||
// auto f4_hi = ck_tile::type_convert<AccDataType>(f4x2)[1];
|
||||
aut a_f4_lo =
|
||||
ck_tile::type_convert<AccDataType>(a_f4x2.template unpack<>(Number<0>{}));
|
||||
auto a_f4_hi =
|
||||
ck_tile::type_convert<AccDataType>(a_f4x2.template unpack<>(Number<1>{}));
|
||||
|
||||
a_m_k_scaled(m, k) = a_f4_lo * a_scale;
|
||||
a_m_k_scaled(m, k + 1) = a_f4_hi * a_scale;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for(int n = 0; n < N; n++)
|
||||
{
|
||||
for(int k = 0; k < K; k++)
|
||||
{
|
||||
if constexpr(std::is_same_v<BDatatype, f4x2_pk_t>)
|
||||
{
|
||||
if(k % 2 == 1)
|
||||
continue; // skip odd k
|
||||
|
||||
auto b_f4x2 = b_k_n(k, n);
|
||||
auto b_scale = b_k_n_scale(k / ScaleBlockSize, n);
|
||||
// auto f4_lo = ck_tile::type_convert<AccDataType>(f4x2)[0];
|
||||
// auto f4_hi = ck_tile::type_convert<AccDataType>(f4x2)[1];
|
||||
auto b_f4_lo =
|
||||
ck_tile::type_convert<AccDataType>(b_f4x2.template unpack<>(Number<0>{}));
|
||||
auto b_f4_hi =
|
||||
ck_tile::type_convert<AccDataType>(b_f4x2.template unpack<>(Number<1>{}));
|
||||
|
||||
b_k_n_scaled(k, n) = b_f4_lo * b_scale;
|
||||
b_k_n_scaled(k + 1, n) = b_f4_hi * b_scale;
|
||||
}
|
||||
else
|
||||
{
|
||||
b_k_n_scaled(k, n) =
|
||||
ck_tile::type_convert<AccDataType>((b_k_n(k, n))) *
|
||||
ck_tile::type_convert<AccDataType>(b_k_n_scale(k / ScaleBlockSize, n));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// call reference gemm
|
||||
reference_gemm<AccDataType, AccDataType, AccDataType, CDataType>(
|
||||
a_m_k_scaled, b_k_n_scaled, c_m_n);
|
||||
}
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename DsDataType,
|
||||
|
||||
@@ -11,10 +11,12 @@
|
||||
#include "ck_tile/ops/flatmm/block/flatmm_uk_config.hpp"
|
||||
#include "ck_tile/ops/flatmm/kernel/flatmm_kernel.hpp"
|
||||
#include "ck_tile/ops/flatmm/kernel/mixed_prec_flatmm_kernel.hpp"
|
||||
#include "ck_tile/ops/flatmm/kernel/mx_flatmm_kernel.hpp"
|
||||
#include "ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v0.hpp"
|
||||
#include "ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1.hpp"
|
||||
#include "ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1_policy.hpp"
|
||||
#include "ck_tile/ops/flatmm/pipeline/mixed_prec_flatmm_pipeline_agmem_bgmem_creg_v1.hpp"
|
||||
#include "ck_tile/ops/flatmm/pipeline/mx_flatmm_pipeline_agmem_bgmem_creg_v1.hpp"
|
||||
#include "ck_tile/ops/flatmm/pipeline/tile_flatmm_shape.hpp"
|
||||
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
|
||||
#include "ck_tile/ops/common/tensor_layout.hpp"
|
||||
|
||||
537
include/ck_tile/ops/flatmm/kernel/mx_flatmm_kernel.hpp
Normal file
537
include/ck_tile/ops/flatmm/kernel/mx_flatmm_kernel.hpp
Normal file
@@ -0,0 +1,537 @@
|
||||
// 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>;
|
||||
|
||||
using BlockGemm = remove_cvref_t<typename GemmPipeline::BlockGemm>;
|
||||
using MThreadPerXdl = BlockGemm::WarpGemm::kM;
|
||||
using NThreadPerXdl = BlockGemm::WarpGemm::kN;
|
||||
using KThreadPerXdl = get_warp_size() / MThreadPerXdl;
|
||||
|
||||
static constexpr int APackedSize = numeric_traits<ADataType>::PackedSize;
|
||||
static constexpr int BPackedSize = numeric_traits<BDataType>::PackedSize;
|
||||
|
||||
static constexpr int MXdlPack = 2;
|
||||
static constexpr int NXdlPack = 2;
|
||||
static constexpr int KXdlPack = 2;
|
||||
|
||||
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_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 = F16xMXF4FlatmmKernel::BlockSize().x;
|
||||
int dync_smem_size = 0;
|
||||
int maxActiveBlocksPerCU = 0;
|
||||
|
||||
[[maybe_unused]] auto e = hipGetDeviceProperties(&prop, deviceId);
|
||||
|
||||
e = hipOccupancyMaxActiveBlocksPerMultiprocessor(
|
||||
&maxActiveBlocksPerCU,
|
||||
reinterpret_cast<void*>(
|
||||
kentry2<block_size,
|
||||
F16xMXF4FlatmmKernel,
|
||||
FlatmmKernelArgs<ScaleM, ScaleN, DsDataType::size()>>),
|
||||
block_size,
|
||||
dync_smem_size);
|
||||
|
||||
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;
|
||||
|
||||
assert(kargs.k_batch == 1);
|
||||
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 = decltype(scale_n)::GranularityK;
|
||||
|
||||
// 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(kargs.M / (MXdlPack * MThreadPerXdl),
|
||||
kargs.K / BlockScaleSize / (KXdlPack * KThreadPerXdl),
|
||||
KThreadPerXdl,
|
||||
MThreadPerXdl));
|
||||
const auto scale_a_desc = transform_tensor_descriptor(
|
||||
scale_a_naive_desc,
|
||||
make_tuple(
|
||||
make_merge_transform(
|
||||
make_tuple(Padded_Scale_M / (MXdlPack * MThreadPerXdl), MThreadPerXdl)),
|
||||
make_merge_transform(make_tuple(
|
||||
kargs.K / BlockScaleSize / (KXdlPack * KThreadPerXdl), 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(kargs.N / (NXdlPack * NThreadPerXdl),
|
||||
kargs.K / BlockScaleSize / (KXdlPack * KThreadPerXdl),
|
||||
KThreadPerXdl,
|
||||
NThreadPerXdl));
|
||||
const auto scale_b_desc = transform_tensor_descriptor(
|
||||
scale_b_navie_desc,
|
||||
make_tuple(
|
||||
make_merge_transform(
|
||||
make_tuple(kargs.N / (NXdlPack * NThreadPerXdl), NThreadPerXdl)),
|
||||
make_merge_transform(make_tuple(
|
||||
kargs.K / BlockScaleSize / (KXdlPack * KThreadPerXdl), 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);
|
||||
}();
|
||||
|
||||
// index_t FlatScaleK =
|
||||
// (kargs.K / decltype(scale_n)::GranularityK) * N_Pack *
|
||||
// BlockGemmShape::WarpTile::at(I1);
|
||||
// index_t FlatScaleN = kargs.N / N_Pack / BlockGemmShape::WarpTile::at(I1);
|
||||
|
||||
// const auto scale_b_flat_view = make_naive_tensor_view<address_space_enum::global>(
|
||||
// reinterpret_cast<const e8m0_t*>(scale_n.ptr),
|
||||
// make_tuple(FlatScaleN, FlatScaleK),
|
||||
// make_tuple(FlatScaleK, 1),
|
||||
// number<8>{},
|
||||
// number<1>{});
|
||||
|
||||
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});
|
||||
|
||||
// auto scale_block_window =
|
||||
// make_tile_window(views.at(I4),
|
||||
// make_tuple(number<FlatmmPipeline::flatNPerWarp>{},
|
||||
// number<FlatmmPipeline::flatKPerWarp * N_Pack * 4 /
|
||||
// 32>{}),
|
||||
// {i_n / BlockGemmShape::WarpTile::at(I1) / N_Pack, 0});
|
||||
|
||||
auto scale_a = kargs.scale_m_ptr;
|
||||
static constexpr int BlockScaleSize = decltype(scale_n)::GranularityK;
|
||||
|
||||
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_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 = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
|
||||
const index_t i_n = __builtin_amdgcn_readfirstlane(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);
|
||||
}
|
||||
partition_idx += gridDim.x;
|
||||
} while(UsePersistentKernel && partition_idx < total_work_tile_cnt);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,188 @@
|
||||
// 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 index_t N_Pack = 2; // it's fixed for fp4
|
||||
static constexpr index_t K_Pack = 2; // it's fixed for fp4
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeF16xF4_ALdsBlockDescriptor()
|
||||
{
|
||||
using namespace ck_tile;
|
||||
|
||||
constexpr index_t MPerXdl = Problem::BlockGemmShape::WarpTile::at(I0);
|
||||
constexpr index_t NPerXdl = Problem::BlockGemmShape::WarpTile::at(I1);
|
||||
|
||||
static_assert(MPerXdl == 16 && NPerXdl == 16);
|
||||
|
||||
/*reduce transform layers,compare with old ck*/
|
||||
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
|
||||
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
|
||||
constexpr index_t KPack = GetSmemPackA<Problem>();
|
||||
|
||||
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 int ContiguousThreadsCntInDS_READ_16B = 4;
|
||||
|
||||
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<ContiguousThreadsCntInDS_READ_16B>{})),
|
||||
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;
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeFp16xF4_ADramTileDistribution()
|
||||
{
|
||||
using ADataType = remove_cvref_t<typename Problem::ADataType>;
|
||||
using ALayout = remove_cvref_t<typename Problem::ALayout>;
|
||||
|
||||
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 MakeF16xF4_ALDS_TileDistribution()
|
||||
{
|
||||
using TileShape = typename Problem::BlockGemmShape;
|
||||
using ADataType = remove_cvref_t<typename Problem::ADataType>;
|
||||
using ALayout = remove_cvref_t<typename Problem::ALayout>;
|
||||
|
||||
static_assert(TileShape::WarpTile::at(I1) == 16, "requires XDL_N == 16");
|
||||
static_assert(TileShape::BlockWarps::at(I0) == 1, "requires Wave_M == 1");
|
||||
|
||||
constexpr int Repeat = TileShape::BlockWarps::at(number<1>{});
|
||||
constexpr int M0 = TileShape::WarpTile::at(I0);
|
||||
|
||||
constexpr int K_Lane = 64 / TileShape::WarpTile::at(I1); // 4
|
||||
|
||||
constexpr int K2 = TileShape::WarpTile::at(I2) / K_Lane; // 8
|
||||
constexpr int XDL_PerThreadK = KBPerLoad / K2; // 4
|
||||
constexpr int K0 = K_Lane; // 4
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<sequence<Repeat>,
|
||||
tuple<sequence<M0>, sequence<K0, XDL_PerThreadK, K2>>,
|
||||
tuple<sequence<0>, sequence<2, 1>>,
|
||||
tuple<sequence<0>, sequence<0, 0>>,
|
||||
sequence<2>,
|
||||
sequence<2>>{});
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeFp4BFlatDramTileDistribution()
|
||||
{
|
||||
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, N_Pack>, // second
|
||||
// direction
|
||||
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 MakeFp4ScaleBFlatDramTileDistribution()
|
||||
{
|
||||
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 N_Warp = TileShape::BlockWarps::at(number<1>{});
|
||||
|
||||
constexpr index_t XDLPerBlock = TileShape::kK / TileShape::WarpTile::at(I2);
|
||||
constexpr index_t K_Lane = 64 / TileShape::WarpTile::at(I1);
|
||||
constexpr index_t N_Lane = TileShape::WarpTile::at(I1);
|
||||
|
||||
constexpr index_t NWavePerBlk = N_Warp;
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<
|
||||
sequence<>, // ?
|
||||
tuple<sequence<NWavePerBlk>, // second direction
|
||||
sequence<K_Lane, N_Lane, N_Pack * K_Pack>>, // first
|
||||
// direction
|
||||
// wave in blk, // thd in wave
|
||||
// <M, K> // <M, K>
|
||||
tuple<sequence<1>, sequence<2, 2>>, // which direction
|
||||
tuple<sequence<0>, sequence<0, 1>>, // which index
|
||||
// <repeat, vec_load>
|
||||
sequence<2>,
|
||||
sequence<2>>{});
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
Reference in New Issue
Block a user