ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-30 03:37:38 +00:00
Code Issues Packages Projects Releases Wiki Activity

remove unnecessary files

Browse Source
This commit is contained in:
aska-0096
2025-05-30 08:39:25 +00:00
parent 0cd2e6e782
commit bb5bdff61c
6 changed files with 0 additions and 4958 deletions
Download Patch File Download Diff File Expand all files Collapse all files

359
example/67_gemm_microscaling/gemm_mx_fp8_bpreshuffle.cpp
View File

@@ -1,359 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx_b_preshuffle.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/utility/sequence.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_mx_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/host_tensor.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F8 = ck::f8_t;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using XDataType = ck::e8m0_bexp_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = F8;
using A1DataType = XDataType;
using B0DataType = F8;
using B1DataType = XDataType;
using AccDataType = F32;
using DsDataType = ck::Tuple<>;
using CDataType = BF16;
using CShuffleDataType = CDataType;
using A0Layout = Row;
using B0Layout = Col;
using CLayout = Row;
void preShuffleBuffer(const F8* src, F8* dst, int N, int K, int NXdl)
{
int KPack = 16;
int NLane = NXdl;
int KLane = 64 / NLane;
int K0 = K / (KLane * KPack);
// K -> K0 KLane KPack
// N -> N0 NLane
// N, K -> N0 K0 KLane NLane KPack
int tempk;
for(int n = 0; n < N; ++n)
{
for(int k = 0; k < K; ++k)
{
int n0 = n / NLane;
int n1 = n % NLane;
int k0 = k / (KLane * KPack);
tempk = k % (KLane * KPack);
int k1 = tempk / KPack;
int k2 = tempk % KPack;
int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
k1 * KPack * NLane + n1 * KPack + k2;
dst[outputIndex] = src[n * K + k];
}
}
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough; // elementwise transformation for A matrix
using BElementOp = PassThrough; // elementwise transformation for B matrix
using CElementOp = PassThrough; // elementwise transformation for C matrix
constexpr ck::index_t ScaleBlockSize = 32; // scaling block size
constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMX_Xdl_CShuffleV3_BPreShuffle<
A0Layout, B0Layout, CLayout,
A0DataType, A1DataType, B0DataType, B1DataType, CDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CElementOp, GemmSpec,
ScaleBlockSize, 256,
128, 128, 128,
16, 16,
16, 16,
8, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
2, 1, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, A0DataType, B0DataType>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
bool flush_cache = true;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideC = N;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 8)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
flush_cache = std::stoi(argv[7]);
StrideA = K;
StrideB = K;
StrideC = N;
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 6: M, N, K\n");
printf("arg7: flush both I$ and L2$ (0=no, 1=yes)\n");
exit(0);
}
ck::index_t Scale_Stride_AM = (K + ScaleBlockSize - 1) / ScaleBlockSize;
ck::index_t Scale_Stride_BN = (K + ScaleBlockSize - 1) / ScaleBlockSize;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<A0DataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<A1DataType> a_m_k_scale(f_host_tensor_descriptor(
M, (K + ScaleBlockSize - 1) / ScaleBlockSize, Scale_Stride_AM, A0Layout{}));
Tensor<B0DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B0DataType> b_preshuffled(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B1DataType> b_k_n_scale(f_host_tensor_descriptor(
(K + ScaleBlockSize - 1) / ScaleBlockSize, N, Scale_Stride_BN, B0Layout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "a_m_k_scale: " << a_m_k_scale.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "b_k_n_scale: " << b_k_n_scale.mDesc << std::endl;
std::cout << "e_m_n: " << c_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b_k_n_scale.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 3:
a_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b_k_n_scale.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 4:
a_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
break;
case 5:
a_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
break;
case 6:
a_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b_k_n_scale.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
b_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
}
DeviceMem a_device_buf(sizeof(A0DataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem a_scale_device_buf(sizeof(A1DataType) * a_m_k_scale.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(B0DataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem b_scale_device_buf(sizeof(B1DataType) * b_k_n_scale.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
a_scale_device_buf.ToDevice(a_m_k_scale.mData.data());
b_scale_device_buf.ToDevice(b_k_n_scale.mData.data());
#if 1
printf("print a_m_k_scale:\n");
for(int m = 0; m < M; ++m)
{
for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
{
printf("%f ", ck::type_convert<float>(a_m_k_scale(m, k)));
}
printf("\n");
}
#endif
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CElementOp{};
// do GEMM
auto device_op = DeviceOpInstance{};
int NPerXdl = device_op.GetPreShuffleParameters();
preShuffleBuffer(b_k_n.mData.data(), b_preshuffled.mData.data(), N, K, NPerXdl);
b_device_buf.ToDevice(b_preshuffled.mData.data());
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(static_cast<A0DataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<XDataType*>(a_scale_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<XDataType*>(b_scale_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
Scale_Stride_AM,
StrideB,
Scale_Stride_BN,
StrideC,
1, // KBatch
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N * K / ScaleBlockSize;
std::size_t num_btype = sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(CDataType) * M * N +
sizeof(XDataType) * (M * K + K * N) / ScaleBlockSize;
float ave_time = .0;
if(flush_cache)
{
int rotating_buf = (512 * 1024 * 1024 + num_btype - 1) / num_btype;
ave_time = invoker.Run(argument,
StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf});
}
else
{
ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100});
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< device_op.GetTypeString() << std::endl;
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceMXGemm<A0DataType,
B0DataType,
CDataType,
AccDataType,
XDataType,
PassThrough,
PassThrough,
PassThrough,
float,
float>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k,
a_m_k_scale,
b_k_n,
b_k_n_scale,
c_m_n_host_result,
PassThrough{},
PassThrough{},
PassThrough{});
ref_invoker.Run(ref_argument);
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
return ck::utils::check_err(
c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results!", 5e-2, 5e-2)
? 0
: 1;
}
return 0;
}

94
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_mx_selector.hpp
View File

@@ -1,94 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuflle_v1_mx.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuflle_v3_mx.hpp"
namespace ck {
template <BlockGemmPipelineVersion BlkGemmPipelineVer,
BlockGemmPipelineScheduler BlkGemmPipeSche,
index_t ThreadBlockSize,
index_t ScaleBlockSize,
typename ADataType,
typename AScaleDataType,
typename BDataType,
typename BScaleDataType,
typename ComputeDataType, // TODO: remove this as in this pipeline ADataType and BDataType
// must be used for compute
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack>
constexpr auto BlockGemmMXBPreshufflePipeline_Selector()
{
// Hardware MX GEMM pipeline
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
return BlockwiseGemmXdlops_pipeline_bpreshuffle_v1_mx<BlkGemmPipeSche,
ThreadBlockSize,
ScaleBlockSize,
ADataType,
AScaleDataType,
BDataType,
BScaleDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
static_assert(MRepeat >= 4, "MRepeat should at least be 4 in BlockGemmPipelineVersion::v3");
return BlockwiseGemmXdlops_pipeline_bpreshuffle_v3_mx<BlkGemmPipeSche,
ThreadBlockSize,
ScaleBlockSize,
ADataType,
AScaleDataType,
BDataType,
BScaleDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
else
{
std::cerr << "MX GEMM Pipeline configuration is not available" << std::endl;
}
}
} // namespace ck

832
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuflle_v1_mx.hpp
View File

@@ -1,832 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/block/blockwise_gemm_mx_pipeline_xdlops_base.hpp"
namespace ck {
// Naive pipeline with lowest resource request per WGP
// GlobalPrefetchStages: 2
// LocalPreFillStages: 1
// LocalPreFetchStages: 1
// LocalSharedMemoryBuffer: 1
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
index_t ThreadBlockSize,
index_t ScaleBlockSize,
typename ADataType,
typename AScaleDataType,
typename BDataType,
typename BScaleDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat, // MXdlPerWave
index_t NRepeat, // NXdlPerWave
index_t KPack>
struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1_mx
{
};
template <index_t ThreadBlockSize,
index_t ScaleBlockSize,
typename ADataType,
typename AScaleDataType,
typename BDataType,
typename BScaleDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat, // MXdlPerWave
index_t NRepeat, // NXdlPerWave
index_t KPack>
struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1_mx<BlockGemmPipelineScheduler::Intrawave,
ThreadBlockSize,
ScaleBlockSize,
ADataType,
AScaleDataType,
BDataType,
BScaleDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
ADataType,
BDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
ADataType,
BDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::I0;
using Base::I1;
using Base::KRepeat;
using Base::MWaves;
using Base::NWaves;
using Base::WaveSize;
using Base::xdlops_gemm;
using Base::CalculateCThreadOriginDataIndex;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetWaveIdx;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
using Base::AMmaKStride;
using Base::BMmaKStride;
using Base::KThreadChunk;
using Base::APackedSize;
using Base::BPackedSize;
using Base::ComputePackedSize;
using AccType = typename Base::AccType;
using Tuple4 = typename Base::Tuple4;
using ComputeTypeA = typename Base::ComputeTypeA;
using ComputeTypeB = typename Base::ComputeTypeB;
static constexpr index_t PrefetchStages = 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 2;
template <typename TileDesc_M0_M1_M2_K>
__host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
{
constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
constexpr index_t K2 = KPack;
constexpr index_t K1 = 64 / NPerXDL;
constexpr index_t K0 = KRepeat;
return transform_tensor_descriptor(
TileDesc_M0_M1_M2_K{},
make_tuple(
make_pass_through_transform(Number<M0>{}),
make_pass_through_transform(Number<M1>{}),
make_pass_through_transform(Number<M2>{}),
make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
}
static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
static constexpr auto ScalesPerKBlockSize =
KPerBlock / ScaleBlockSize; // How many mx-vectors per K block
//> How many mx-vectors in each row/col is processed in one call to xdlops_gemm.Run()
static constexpr auto ScalesPerXdlopsRun = (KPack * xdlops_gemm.K0PerXdlops) / ScaleBlockSize;
//> How many scales a thread must read to accommodate one call to xdlops_gemm.Run()
static constexpr auto ScalesPerXdlopsRunPerThread =
ScalesPerXdlopsRun / xdlops_gemm.mfma_instr.num_input_blks;
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AScaleGridBuffer,
typename AScaleGridDesc,
typename AScaleThreadTransfer,
typename BScaleGridBuffer,
typename BScaleGridDesc,
typename BScaleThreadTransfer>
__device__ void Run(
// ABlockCopy
const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
// BBlockCopy
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
// CThread
CThreadBuffer& c_thread_buf,
// A and B scales
const AScaleGridDesc& a_scale_grid_desc,
AScaleThreadTransfer& a_scale_thread_copy,
const AScaleGridBuffer& a_scale_grid_buf,
const BScaleGridDesc& b_scale_grid_desc,
BScaleThreadTransfer& b_scale_thread_copy,
const BScaleGridBuffer& b_scale_grid_buf,
index_t num_loop) const
{
ignore = b_block_desc;
ignore = b_block_buf;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
b_thread_desc_.GetElementSpaceSize());
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
a_scale_thread_desc.GetElementSpaceSize());
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
b_scale_thread_desc.GetElementSpaceSize());
StaticallyIndexedArray<decltype(a_scale_thread_buf), Number<2>{}> a_scale_thread_bufs;
StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs;
// Global prefetch A1 B1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(I0));
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Prefetch a_scales to buf 0
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
auto a_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
a_scale_thread_desc_copy.GetElementSpaceSize());
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc_copy,
make_tuple(I0, I0),
a_scale_thread_buf_copy);
a_scale_thread_bufs(I0)(Number<a_scale_offset>{}) =
a_scale_thread_buf_copy[Number<0>{}];
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
});
// restore row id and advance to the next set of scales
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
make_multi_index(-MPerBlock, ScalesPerKBlockSize));
// Prefetch b_scales to buf 0
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
auto b_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
b_scale_thread_desc_copy.GetElementSpaceSize());
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc_copy,
make_tuple(I0, I0),
b_scale_thread_buf_copy);
b_scale_thread_bufs(I0)(Number<b_scale_offset>{}) =
b_scale_thread_buf_copy[Number<0>{}];
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
});
// restore col id and advance to the next set of scales
// NWaves * NPerXDL * NRepeat == NPerBlock
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
make_multi_index(-NPerBlock, ScalesPerKBlockSize));
__builtin_amdgcn_sched_barrier(0);
// Local prefill A1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
// Global prefetch A2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
// Prefetch a_scales to buf 1
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
auto a_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
a_scale_thread_desc_copy.GetElementSpaceSize());
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc_copy,
make_tuple(I0, I0),
a_scale_thread_buf_copy);
a_scale_thread_bufs(I1)(Number<a_scale_offset>{}) =
a_scale_thread_buf_copy[Number<0>{}];
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
});
// restore row id and advance to the next set of scales
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
make_multi_index(-MPerBlock, ScalesPerKBlockSize));
// Prefetch b_scales to buf 1
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
auto b_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
b_scale_thread_desc_copy.GetElementSpaceSize());
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc_copy,
make_tuple(I0, I0),
b_scale_thread_buf_copy);
b_scale_thread_bufs(I1)(Number<b_scale_offset>{}) =
b_scale_thread_buf_copy[Number<0>{}];
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
});
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
make_multi_index(-NPerBlock, ScalesPerKBlockSize));
// Local prefetch A1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
constexpr auto k_step = k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
constexpr auto a_k_step_chunk =
k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
a_thread_buf);
});
});
});
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
// loop over k with the step KPerBlock
index_t i = 0;
do
{
auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(local_read_buf));
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, mfma_reg_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, local_read_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeTypeB>()(ik) =
b_thread_bufs[mfma_reg_buf]
[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
constexpr index_t a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
constexpr index_t b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
static_assert(
0 < ScalesPerXdlopsRunPerThread,
"Must have at least one scale per Xdlops per Thread.");
vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread>
a_scale_thread_vec;
vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
b_scale_thread_vec;
// Pack scale_thread_buf into scale_thread_vec
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
a_scale_thread_bufs[mfma_reg_buf]
[Number<a_scale_offset + s>{}];
b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
b_scale_thread_bufs[mfma_reg_buf]
[Number<b_scale_offset + s>{}];
});
using mfma_input_type_a =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops /
APackedSize>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB,
xdlops_gemm.K1PerXdlops /
BPackedSize>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// MFMA accumulation
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type_a>(),
a_scale_thread_vec.template AsType<AScaleDataType>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
b_scale_thread_vec.template AsType<BScaleDataType>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
block_sync_lds();
// a thread copy
static_for<0, KRepeat, 1>{}([&](auto k) {
constexpr auto k_step =
k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}(
[&](auto chunk) {
constexpr auto a_k_step_chunk =
k_step + chunk * KThreadChunk *
xdlops_gemm.mfma_instr.num_input_blks;
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
a_thread_buf);
});
});
});
// Prefetch a_scales
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
auto a_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
a_scale_thread_desc_copy.GetElementSpaceSize());
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc_copy,
make_tuple(I0, I0),
a_scale_thread_buf_copy);
a_scale_thread_bufs(mfma_reg_buf)(Number<a_scale_offset>{}) =
a_scale_thread_buf_copy[Number<0>{}];
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc,
make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
});
// restore row id and advance to the next set of scales
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, make_multi_index(-MPerBlock, ScalesPerKBlockSize));
// Prefetch b_scales
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
auto b_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
b_scale_thread_desc_copy.GetElementSpaceSize());
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc_copy,
make_tuple(I0, I0),
b_scale_thread_buf_copy);
b_scale_thread_bufs(mfma_reg_buf)(Number<b_scale_offset>{}) =
b_scale_thread_buf_copy[Number<0>{}];
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc,
make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
});
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, make_multi_index(-NPerBlock, ScalesPerKBlockSize));
};
LoopFunc(I0, I1);
LoopFunc(I1, I0);
i += 2;
} while(i < (num_loop - 2));
}
// tail
if constexpr(TailNum == TailNumber::Even)
{
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(I1));
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeTypeB>()(ik) =
b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
constexpr index_t a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
constexpr index_t b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
// Pack b_scale_thread_buf into b_scale_thread_vec
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
});
using mfma_input_type_a =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops / APackedSize>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB,
xdlops_gemm.K1PerXdlops / BPackedSize>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// MFMA accumulation
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type_a>(),
a_scale_thread_vec.template AsType<AScaleDataType>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
b_scale_thread_vec.template AsType<BScaleDataType>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
block_sync_lds();
// a thread copy
static_for<0, KRepeat, 1>{}([&](auto k) {
constexpr auto k_step =
k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
constexpr auto a_k_step_chunk =
k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
a_thread_buf);
});
});
});
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeTypeB>()(ik) =
b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
constexpr index_t a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
constexpr index_t b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
// Pack b_scale_thread_buf into b_scale_thread_vec
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
a_scale_thread_bufs[I1][Number<a_scale_offset + s>{}];
b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
b_scale_thread_bufs[I1][Number<b_scale_offset + s>{}];
});
using mfma_input_type_a =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops / APackedSize>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB,
xdlops_gemm.K1PerXdlops / BPackedSize>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// MFMA accumulation
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type_a>(),
a_scale_thread_vec.template AsType<AScaleDataType>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
b_scale_thread_vec.template AsType<BScaleDataType>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
}
else if constexpr(TailNum == TailNumber::Odd)
{
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeTypeB>()(ik) =
b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
constexpr index_t a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
constexpr index_t b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
// Pack b_scale_thread_buf into b_scale_thread_vec
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
});
using mfma_input_type_a =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops / APackedSize>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB,
xdlops_gemm.K1PerXdlops / BPackedSize>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// MFMA accumulation
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type_a>(),
a_scale_thread_vec.template AsType<AScaleDataType>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
b_scale_thread_vec.template AsType<BScaleDataType>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
}
}
// TODO: make this field protected when a_scale_thread_copy_ is moved
// here
static constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
// Is used to copy data from a_scale_grid to a_scale_thread
static constexpr auto a_scale_thread_desc_copy =
make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
// TODO: make this field protected when b_scale_thread_copy_ is moved
// here
static constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<NRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
// Is used to copy data from b_scale_grid to b_scale_thread_buf
static constexpr auto b_scale_thread_desc_copy =
make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
protected:
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
// using Base::b_thread_desc_;
using Base::c_thread_desc_;
static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
};
} // namespace ck

932
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuflle_v3_mx.hpp
View File

@@ -1,932 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/block/blockwise_gemm_mx_pipeline_xdlops_base.hpp"
namespace ck {
// Naive pipeline with lowest resource request per WGP
// GlobalPrefetchStages: 2
// LocalPreFillStages: 1
// LocalPreFetchStages: 1
// LocalSharedMemoryBuffer: 1
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
index_t ThreadBlockSize,
index_t ScaleBlockSize,
typename ADataType,
typename AScaleDataType,
typename BDataType,
typename BScaleDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat, // MXdlPerWave
index_t NRepeat, // NXdlPerWave
index_t KPack>
struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3_mx
{
};
template <index_t ThreadBlockSize,
index_t ScaleBlockSize,
typename ADataType,
typename AScaleDataType,
typename BDataType,
typename BScaleDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat, // MXdlPerWave
index_t NRepeat, // NXdlPerWave
index_t KPack>
struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3_mx<BlockGemmPipelineScheduler::Intrawave,
ThreadBlockSize,
ScaleBlockSize,
ADataType,
AScaleDataType,
BDataType,
BScaleDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
ADataType,
BDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
ADataType,
BDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::I0;
using Base::I1;
using Base::I2;
using Base::KRepeat;
using Base::MWaves;
using Base::NWaves;
using Base::WaveSize;
using Base::xdlops_gemm;
using typename Base::HotLoopInstList;
using Base::CalculateCThreadOriginDataIndex;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetWaveIdx;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
using Base::AMmaKStride;
using Base::BMmaKStride;
using Base::KThreadChunk;
using Base::APackedSize;
using Base::BPackedSize;
using Base::ComputePackedSize;
using AccType = typename Base::AccType;
using Tuple4 = typename Base::Tuple4;
using ComputeTypeA = typename Base::ComputeTypeA;
using ComputeTypeB = typename Base::ComputeTypeB;
static constexpr index_t PrefetchStages = 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 2;
static constexpr index_t HotloopLocalBufSwitch = MRepeat % 2 == 0 ? 0 : 1;
template <typename TileDesc_M0_M1_M2_K>
__host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
{
constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
constexpr index_t K2 = KPack;
constexpr index_t K1 = 64 / NPerXDL;
constexpr index_t K0 = KRepeat;
return transform_tensor_descriptor(
TileDesc_M0_M1_M2_K{},
make_tuple(
make_pass_through_transform(Number<M0>{}),
make_pass_through_transform(Number<M1>{}),
make_pass_through_transform(Number<M2>{}),
make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
}
static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
static constexpr auto ScalesPerKBlockSize =
KPerBlock / ScaleBlockSize; // How many mx-vectors per K block
//> How many mx-vectors in each row/col is processed in one call to xdlops_gemm.Run()
static constexpr auto ScalesPerXdlopsRun = (KPack * xdlops_gemm.K0PerXdlops) / ScaleBlockSize;
//> How many scales a thread must read to accommodate one call to xdlops_gemm.Run()
static constexpr auto ScalesPerXdlopsRunPerThread =
ScalesPerXdlopsRun / xdlops_gemm.mfma_instr.num_input_blks;
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__device__ static constexpr auto HotLoopScheduler()
{
// A/B split schedule
// compiler is likely to use ds_read2 when instruction width smaller than 16bytes
constexpr auto num_ds_read_inst_a =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
? HotLoopInstList::A_LDS_Read_Inst_Num
: HotLoopInstList::A_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_read_inst_b =
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16
? HotLoopInstList::B_LDS_Read_Inst_Num
: HotLoopInstList::B_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
constexpr auto num_ds_write_inst_b = HotLoopInstList::B_LDS_Write_Inst_Num;
constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto mfma_cycle = HotLoopInstList::C_MFMA_Inst_Cycle;
constexpr auto ds_read_a_issue_cycle =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
constexpr auto ds_read_b_issue_cycle =
HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
constexpr auto ds_read_a_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
constexpr auto ds_read_b_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
constexpr auto num_dsread_a_mfma =
(num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
constexpr auto num_dsread_b_mfma =
(num_ds_read_inst_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
// stage 1
// Separate this part?
// constexpr auto num_mfma_per_ds_read = sizeof(ComputeDataType) / sizeof(ADataType) >
// sizeof(ComputeDataType) / sizeof(BDataType)
// ? sizeof(ComputeDataType) / sizeof(ADataType)
// : sizeof(ComputeDataType) / sizeof(BDataType);
constexpr auto num_mfma_stage1 = num_mfma_inst - (num_dsread_a_mfma + num_dsread_b_mfma);
constexpr auto num_mfma_per_issue =
num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b);
constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a;
constexpr auto num_dswrite_per_issue_b = num_ds_write_inst_b / num_buffer_load_inst_b;
static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA
});
static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA
});
// stage 2
static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) {
if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >=
ds_read_a_mfma_rate)
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(0x100,
num_ds_read_inst_a - (num_dsread_a_mfma - 1) *
ds_read_a_mfma_rate,
0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
static_for<0, num_dsread_b_mfma, 1>{}([&](auto i) {
if constexpr((num_ds_read_inst_b - (i + 1) * ds_read_b_mfma_rate) >=
ds_read_b_mfma_rate)
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(0x100,
num_ds_read_inst_b - (num_dsread_b_mfma - 1) *
ds_read_b_mfma_rate,
0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
}
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AScaleGridBuffer,
typename AScaleGridDesc,
typename AScaleThreadTransfer,
typename BScaleGridBuffer,
typename BScaleGridDesc,
typename BScaleThreadTransfer>
__device__ void Run(
// ABlockCopy
const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
// BBlockCopy
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
// CThread
CThreadBuffer& c_thread_buf,
// A and B scales
const AScaleGridDesc& a_scale_grid_desc,
AScaleThreadTransfer& a_scale_thread_copy,
const AScaleGridBuffer& a_scale_grid_buf,
const BScaleGridDesc& b_scale_grid_desc,
BScaleThreadTransfer& b_scale_thread_copy,
const BScaleGridBuffer& b_scale_grid_buf,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
b_thread_desc_.GetElementSpaceSize());
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
a_scale_thread_desc.GetElementSpaceSize());
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
b_scale_thread_desc.GetElementSpaceSize());
StaticallyIndexedArray<decltype(a_scale_thread_buf), Number<2>{}> a_scale_thread_bufs;
StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs;
// Global prefetch B1
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(I0));
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Global prefetch A1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
// Prefetch a_scales 1
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
auto a_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
a_scale_thread_desc_copy.GetElementSpaceSize());
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc_copy,
make_tuple(I0, I0),
a_scale_thread_buf_copy);
a_scale_thread_bufs(I0)(Number<a_scale_offset>{}) =
a_scale_thread_buf_copy[Number<0>{}];
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
});
// restore row id and advance to the next set of scales
a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
make_multi_index(-MPerBlock, ScalesPerKBlockSize));
// Prefetch b_scales 1
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
auto b_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
b_scale_thread_desc_copy.GetElementSpaceSize());
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc_copy,
make_tuple(I0, I0),
b_scale_thread_buf_copy);
b_scale_thread_bufs(I0)(Number<b_scale_offset>{}) =
b_scale_thread_buf_copy[Number<0>{}];
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
});
// restore col id and advance to the next set of scales
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
make_multi_index(-NPerBlock, ScalesPerKBlockSize));
// Local prefill A1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0)); // vmem->vgpr-> lds0
// Global prefetch A2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// Local prefetch A1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
constexpr auto k_step = k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
constexpr auto a_k_step_chunk =
k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
a_thread_buf);
});
});
});
// main body
if constexpr(HasMainLoop)
{
// loop over k with the step KPerBlock
index_t i = 0;
do
{
auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf, auto a_buf) {
// Prefetch a_scales 2
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
auto a_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
a_scale_thread_desc_copy.GetElementSpaceSize());
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc_copy,
make_tuple(I0, I0),
a_scale_thread_buf_copy);
a_scale_thread_bufs(local_read_buf)(Number<a_scale_offset>{}) =
a_scale_thread_buf_copy[Number<0>{}];
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc,
make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
});
// restore row id and advance to the next set of scales
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, make_multi_index(-MPerBlock, ScalesPerKBlockSize));
// Prefetch b_scales 2
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
auto b_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
b_scale_thread_desc_copy.GetElementSpaceSize());
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc_copy,
make_tuple(I0, I0),
b_scale_thread_buf_copy);
b_scale_thread_bufs(local_read_buf)(Number<b_scale_offset>{}) =
b_scale_thread_buf_copy[Number<0>{}];
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc,
make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
});
// restore col id and advance to the next set of scales
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, make_multi_index(-NPerBlock, ScalesPerKBlockSize));
// Local prefill A2
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf));
// Global prefetch A1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
// Global prefetch B2
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(local_read_buf));
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// A1 * B1
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeTypeB>()(ik) =
b_thread_bufs[mfma_reg_buf]
[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
constexpr index_t a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
constexpr index_t b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread>
a_scale_thread_vec;
vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
b_scale_thread_vec;
// Pack scale_thread_buf into scale_thread_vec
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
a_scale_thread_bufs[mfma_reg_buf]
[Number<a_scale_offset + s>{}];
b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
b_scale_thread_bufs[mfma_reg_buf]
[Number<b_scale_offset + s>{}];
});
using mfma_input_type_a =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops /
APackedSize>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB,
xdlops_gemm.K1PerXdlops /
BPackedSize>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// MFMA accumulation
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type_a>(),
a_scale_thread_vec.template AsType<AScaleDataType>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
b_scale_thread_vec.template AsType<BScaleDataType>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
}); // KRepeat
}); // NRepeat
}); // MRepeat
// Local prefetch A2
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
constexpr auto k_step =
k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}(
[&](auto chunk) {
constexpr auto a_k_step_chunk =
k_step + chunk * KThreadChunk *
xdlops_gemm.mfma_instr.num_input_blks;
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
a_block_buf.At(local_read_buf),
a_thread_desc_,
make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
a_thread_buf);
});
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
}; // LoopFunc
LoopFunc(I0, I1, I0);
LoopFunc(I1, I0, I1);
i += 2;
} while(i < (num_loop - 2));
}
// tail
if constexpr(TailNum == TailNumber::Even)
{
// Prefetch a_scales 2
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
auto a_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
a_scale_thread_desc_copy.GetElementSpaceSize());
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc_copy,
make_tuple(I0, I0),
a_scale_thread_buf_copy);
a_scale_thread_bufs(I1)(Number<a_scale_offset>{}) =
a_scale_thread_buf_copy[Number<0>{}];
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
a_scale_thread_copy.MoveSrcSliceWindow(
a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
});
// Prefetch b_scales 2
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
constexpr auto b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
auto b_scale_thread_buf_copy =
make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
b_scale_thread_desc_copy.GetElementSpaceSize());
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc_copy,
make_tuple(I0, I0),
b_scale_thread_buf_copy);
b_scale_thread_bufs(I1)(Number<b_scale_offset>{}) =
b_scale_thread_buf_copy[Number<0>{}];
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc,
make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
});
});
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
});
// Local prefill A2
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
// Global prefetch B2
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(I1));
// A1 * B1
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeTypeB>()(ik) =
b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
constexpr index_t a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
constexpr index_t b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
// Pack b_scale_thread_buf into b_scale_thread_vec
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
});
using mfma_input_type_a =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops / APackedSize>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB,
xdlops_gemm.K1PerXdlops / BPackedSize>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// MFMA accumulation
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type_a>(),
a_scale_thread_vec.template AsType<AScaleDataType>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
b_scale_thread_vec.template AsType<BScaleDataType>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
}); // KRepeat
}); // NRepeat
}); // MRepeat
// Local prefetch A2
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
constexpr auto k_step =
k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
constexpr auto a_k_step_chunk =
k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
a_block_buf.At(I1),
a_thread_desc_,
make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
a_thread_buf);
});
});
});
// A2 * B2
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeTypeB>()(ik) =
b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
constexpr index_t a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
constexpr index_t b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
// Pack b_scale_thread_buf into b_scale_thread_vec
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
a_scale_thread_bufs[I1][Number<a_scale_offset + s>{}];
b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
b_scale_thread_bufs[I1][Number<b_scale_offset + s>{}];
});
using mfma_input_type_a =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops / APackedSize>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB,
xdlops_gemm.K1PerXdlops / BPackedSize>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// MFMA accumulation
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type_a>(),
a_scale_thread_vec.template AsType<AScaleDataType>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
b_scale_thread_vec.template AsType<BScaleDataType>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
}); // KRepeat
}); // NRepeat
}); // MRepeat
}
else if constexpr(TailNum == TailNumber::Odd)
{
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeTypeB>()(ik) =
b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
constexpr index_t a_scale_offset =
a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
constexpr index_t b_scale_offset =
b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
// Pack b_scale_thread_buf into b_scale_thread_vec
static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
});
using mfma_input_type_a =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops / APackedSize>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB,
xdlops_gemm.K1PerXdlops / BPackedSize>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// MFMA accumulation
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type_a>(),
a_scale_thread_vec.template AsType<AScaleDataType>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
b_scale_thread_vec.template AsType<BScaleDataType>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
}); // KRepeat
}); // NRepeat
}); // MRepeat
}
}
// TODO: make this field protected when a_scale_thread_copy_ is moved
// here
static constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
// Is used to copy data from a_scale_grid to a_scale_thread
static constexpr auto a_scale_thread_desc_copy =
make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
// TODO: make this field protected when b_scale_thread_copy_ is moved
// here
static constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<NRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
// Is used to copy data from b_scale_grid to b_scale_thread_buf
static constexpr auto b_scale_thread_desc_copy =
make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
protected:
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
// using Base::b_thread_desc_;
using Base::c_thread_desc_;
static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
};
} // namespace ck

605
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx_b_preshuffle.hpp
View File

@@ -1,605 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/host_utility/flush_cache.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_mx.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_mx_b_preshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// clang-format off
/**
* \brief WIP: Implements XDL CShuffle V3 GEMM for microscale-compliant data types
*
* This class is a work-in-progress implementation of the XDL CShuffle V3 GEMM for
* microscale-compliant data types.
*
* Assumptions:
* - A and B data types are compliant with the OCP Microscaling Formats (MX) Specification
* - Each scale applies to ScaleBlockSize elements in K direction
* - A scale matrix is a row-major
* - B scale matrix is a column-major
* - Scale data types must have get_exponent_value() specialization, whereas lowest 8 bits of the
* exponent will be interpreted as conventional biased Float32 exponent (E8M0)
*
* Tunable parameters.
* The CK instance includes a series of tunable template parameters to control the parallel
* granularity of the workload to achieve load balancing on different hardware platforms. These
* parameters include Block Size, M/N/K Per Block, M/N per XDL, AK1, BK1, etc.
* - Block Size determines the number of threads in the thread block.
* - M/N/K Per Block determines the size of tile that each thread block is responsible for
* calculating.
* - M/N Per XDL refers to M/N size for Instinct accelerator Matrix Fused Multiply Add (MFMA)
* instructions operating on a per-wavefront basis.
* - A/B K1 is related to the data type. It can be any value ranging from 1 to K Per Block. To
* achieve the optimal load/store performance, 128bit per load is suggested. In addition, the A/B
* loading parameters must be changed accordingly to match the A/B K1 value; otherwise, it will
* result in compilation errors.
*
* Conditions for achieving computational load balancing on different hardware platforms can vary.
*
* Serialized version of the algorithm:
* \code
* // E = A * B + C
* // Loop over E[MPerBlock,NPerBlock] tiles
* for(int mb = 0; mb < M; mb += MPerBlock){
* for(int nb = 0; nb < N; nb += NPerBlock){
* // initialize E[MPerBlock,NPerBlock] tile
* for(int mt = mb; mt < mb + MPerBlock; mt++){
* for(int nt = nb; nt < nb + NPerBlock; nt++){
* E[mt,nt] = C[mt,nt];
* }
* }
*
* // multiply-accumulate per tile
* for(int kb = 0; kb < K; kb += KPerBlock){
* for(int m0 = mb; m0 < mb + MPerBlock; m0 += MWaves * MPerXDL){
* for(int n0 = nb; n0 < nb + NPerBlock; n0 += NWaves * NPerXDL){
* for(int mw = m0; mw < m0 + MWaves * MPerXDL; mw += MPerXDL){
* for(int nw = n0; nw < n0 + NWaves * NPerXDL; nw += NPerXDL){
* for(int k0 = kb; k0 < kb + KPerBlock; k0 += mfma.num_input_blks*KPack){
* // MFMA accumulation
* for(int k_pack = k0; k_pack < k0 + mfma.num_input_blks*KPack; k_pack += KPerXdlops){
* // MFMA instruction
* for(int k_mfma = k_pack; k_mfma < k_pack + KPerXdlops; k_mfma += mfma.k_per_blk){
* for(int m = mw; m < mw + MPerXDL; m++){
* for(int n = nw; n < nw + NPerXDL; n++){
* for(int k = k_mfma; k < k_mfma + mfma.k_per_blk; k++){
* E[m,n] += A[m,k] * B[k,n];
* }
* }
* }
* }
* }
* }
* }
* }
* }
* }
* }
* }
* }
* \endcode
*
*/
// clang-format on
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename AScaleDataType,
typename BDataType,
typename BScaleDataType,
typename CDataType,
typename GemmAccDataType, // TODO: always float
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
GemmSpecialization GemmSpec,
index_t ScaleBlockSize, // Scaling block size
index_t BlockSize, // Thread block size
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1,
index_t BK1,
index_t MPerXDL,
index_t NPerXDL,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
typename ComputeTypeA =
ADataType, // XXX: These should always be the same as ADataType and BDataType
typename ComputeTypeB =
BDataType // TODO: Hardcode them and remove from the list of template parameters
>
struct DeviceGemmMX_Xdl_CShuffleV3_BPreShuffle
: public DeviceGemmMX_BPreshuffle<ALayout,
BLayout,
CLayout,
ADataType,
AScaleDataType,
BDataType,
BScaleDataType,
CDataType,
ScaleBlockSize,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
{
// GridwiseGemm
using GridwiseGemm = GridwiseGemmMX_xdl_cshuffle_v3_b_preshuffle<
ALayout,
BLayout,
CLayout,
ADataType,
AScaleDataType,
BDataType,
BScaleDataType,
GemmAccDataType,
CShuffleDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
GemmSpec,
ScaleBlockSize,
BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
AK1,
BK1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
false,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
false,
BBlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock,
BlkGemmPipeSched,
BlkGemmPipelineVer,
ComputeTypeA,
ComputeTypeB>;
using Argument = typename GridwiseGemm::Argument;
int GetPreShuffleParameters() override { return NPerXDL; }
// Invoker
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
if(stream_config.log_level_ > 0)
{
arg.Print();
GridwiseGemm::BlockwiseGemmPipe::HotLoopInstList::Print();
}
if(!GridwiseGemm::CheckValidity(arg))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
index_t gdx, gdy, gdz;
std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch);
float ave_time = 0;
index_t k_grain = arg.KBatch * KPerBlock;
index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
const auto Run = [&](const auto& kernel) {
if(stream_config.flush_cache)
{
Argument arg_ = arg;
const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
auto size_a_buffer =
a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
auto size_b_buffer =
b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
ck::utility::RotatingMemWrapper<Argument> rotating_mem(
arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck::utility::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(arg_.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
0,
arg_.M * arg_.N * sizeof(CDataType),
stream_config.stream_id_));
};
ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
stream_config,
run_flush_cache,
kernel,
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg_);
}
else
{
if(arg.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
0,
arg.M * arg.N * sizeof(CDataType),
stream_config.stream_id_));
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
}
};
// TODO: Check if this is the right algorithm for minimum_occupancy
constexpr index_t minimum_occupancy =
BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave
? (BlkGemmPipelineVer == BlockGemmPipelineVersion::v3 &&
MPerBlock * NPerBlock * KPerBlock * sizeof(ADataType) <= 128 * 128 * 64 * 2)
? 2
: 1
: 2;
if(has_main_k_block_loop)
{
// Tail number always full
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_b_preshuffle<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_b_preshuffle<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
else
{
// Tail number always 1
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_b_preshuffle<GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_b_preshuffle<GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
static_assert(is_scale_mfma_data_type<ADataType>() && is_scale_mfma_data_type<BDataType>(),
"Only microscaling formats are supported for ADataType and BDataType");
static_assert(ScaleBlockSize == 32, "Only ScaleBlockSize 32 is supported");
static_assert(is_same_v<ComputeTypeA, ADataType> && is_same_v<ComputeTypeB, BDataType>,
"ComputeTypeA and ComputeTypeB must be the same as ADataType and BDataType");
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if constexpr(!IsValidCompilationParameter())
{
return false;
}
if(!ck::is_xdl_supported())
{
return false;
}
if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
{
return false;
}
if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding ||
GemmSpec == GemmSpecialization::KPadding))
{
return false;
}
return GridwiseGemm::CheckValidity(arg);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const ADataType* p_a,
const AScaleDataType* p_a_scale,
const BDataType* p_b,
const BScaleDataType* p_b_scale,
CDataType* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideScaleA,
index_t StrideB,
index_t StrideScaleB,
index_t StrideC,
index_t KBatch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return Argument{p_a,
p_a_scale,
p_b,
p_b_scale,
p_c,
M,
N,
K,
StrideA,
StrideScaleA,
StrideB,
StrideScaleB,
StrideC,
KBatch,
a_element_op,
b_element_op,
c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_a_scale,
const void* p_b,
const void* p_b_scale,
void* p_c,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideScaleA,
ck::index_t StrideB,
ck::index_t StrideScaleB,
ck::index_t StrideC,
ck::index_t KBatch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const AScaleDataType*>(p_a_scale),
static_cast<const BDataType*>(p_b),
static_cast<const BScaleDataType*>(p_b_scale),
static_cast<CDataType*>(p_c),
M,
N,
K,
StrideA,
StrideScaleA,
StrideB,
StrideScaleB,
StrideC,
KBatch,
a_element_op,
b_element_op,
c_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
{BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
{BlockGemmPipelineScheduler::Interwave, "Interwave"}};
std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
{BlockGemmPipelineVersion::v1, "v1"},
{BlockGemmPipelineVersion::v2, "v2"},
{BlockGemmPipelineVersion::v3, "v3"},
{BlockGemmPipelineVersion::v4, "v4"},
{BlockGemmPipelineVersion::v5, "v5"}};
// clang-format off
str << "DeviceGemmMX_Xdl_CShuffleV3"
<< "<"
<< getGemmSpecializationString(GemmSpec) << ", "
<< std::string(ALayout::name)[0]
<< std::string(BLayout::name)[0]
<< std::string(CLayout::name)[0]
<< ">"
<< " BlkSize: "
<< BlockSize << ", "
<< "BlkTile: "
<< MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
<< "WaveTile: "
<< MPerXDL<<"x"<<NPerXDL << ", "
<< "WaveMap: "
<< MXdlPerWave<<"x" << NXdlPerWave<<", "
<< "VmemReadVec: "
<< ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
<< "BlkGemmPipelineScheduler: "
<< BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
<< "BlkGemmPipelineVersion: "
<< BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
<< "BlkGemmPipelinePrefetchStages: "
<< GridwiseGemm::BlockwiseGemmPipe::PrefetchStages << ", "
<< "Kpack: "
<< GridwiseGemm::BlockwiseGemmPipe::AMmaKStride << ", "
<< "ScaleBlockSize: "
<< ScaleBlockSize;
// clang-format on
return str.str();
}
REGISTER_EXTRA_PRINTING_METHODS
};
} // namespace device
} // namespace tensor_operation
} // namespace ck

2136
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_mx_b_preshuffle.hpp
View File

File diff suppressed because it is too large Load Diff