mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 02:02:46 +00:00
Implement the fp16xint4 scale weight only kernel for Ali (#1786)
* enable int4 scale (weight only) kernel
* format some files
* Add unit test for int4 weight only
* fixed and formatted code
* fixed
* formated
* formated
* fixed
* fixed a bug in the ckProfiler, and formatted the code
---------
Co-authored-by: mtgu0705 <mtgu@amd.com>
[ROCm/composable_kernel commit: 4f62f6e9b7]
This commit is contained in:
Mingtao Gu
@@ -30,6 +30,7 @@ add_example_executable(example_gemm_xdl_fp8_v3 gemm_xdl_fp8_v3.cpp)
|
||||
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_v3)
|
||||
add_example_executable(example_gemm_xdl_fp16_fp8_v3 gemm_xdl_fp16_fp8_v3.cpp)
|
||||
add_example_executable(example_gemm_xdl_fp16_pk_i4_v3 gemm_xdl_fp16_pk_i4_v3.cpp)
|
||||
add_example_executable(example_gemm_xdl_fp16_pk_i4_v3_b_scale gemm_xdl_fp16_pk_i4_v3_b_scale.cpp)
|
||||
add_example_executable(example_gemm_xdl_bf16_pk_i4_v3 gemm_xdl_bf16_pk_i4_v3.cpp)
|
||||
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_v3)
|
||||
add_example_executable(example_gemm_xdl_bf16_v3 gemm_xdl_bf16_v3.cpp)
|
||||
|
||||
357
example/01_gemm/gemm_xdl_fp16_pk_i4_v3_b_scale.cpp
Normal file
357
example/01_gemm/gemm_xdl_fp16_pk_i4_v3_b_scale.cpp
Normal file
@@ -0,0 +1,357 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "common.hpp"
|
||||
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_scale.hpp"
|
||||
|
||||
using ADataType = ck::half_t;
|
||||
using BDataType = ck::pk_i4_t;
|
||||
using BScaleDataType = ck::half_t;
|
||||
using AccDataType = float;
|
||||
using CShuffleDataType = ck::half_t;
|
||||
using CDataType = ck::half_t;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using CLayout = Row;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CElementOp = PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
static constexpr bool PermuteA = false;
|
||||
static constexpr bool PermuteB = true;
|
||||
|
||||
static constexpr ck::index_t Scale_Block_N = 1;
|
||||
static constexpr ck::index_t Scale_Block_K = 128;
|
||||
|
||||
static constexpr ck::index_t KPerBlock = 64;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmV2Instance =
|
||||
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
|
||||
ALayout, BLayout, CLayout,
|
||||
ADataType, BDataType, BScaleDataType, CDataType, AccDataType, CShuffleDataType,
|
||||
AElementOp, BElementOp, CElementOp, GemmDefault,
|
||||
256, Scale_Block_N, Scale_Block_K,
|
||||
128, 128,
|
||||
KPerBlock, 8, 32,
|
||||
32, 32,
|
||||
4, 1,
|
||||
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
|
||||
2, 8, 8, 0,
|
||||
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>,
|
||||
2, 32, 32, 0,
|
||||
1, 1, S<1, 32, 1, 8>, 8,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, CDataType, CDataType, PermuteA, PermuteB>;
|
||||
|
||||
// clang-format on
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
template <typename ProblemType>
|
||||
bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
|
||||
{
|
||||
using namespace ck::literals;
|
||||
|
||||
auto M = problem_size.M;
|
||||
auto N = problem_size.N;
|
||||
auto K = problem_size.K;
|
||||
auto StrideA = problem_size.StrideA;
|
||||
auto StrideB = problem_size.StrideB;
|
||||
auto StrideC = problem_size.StrideC;
|
||||
auto KBatch = problem_size.KBatch;
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
||||
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {stride, 1_uz});
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {1_uz, stride});
|
||||
}
|
||||
};
|
||||
|
||||
auto f_get_default_stride =
|
||||
[](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
|
||||
if(stride == -1)
|
||||
{
|
||||
// give a chance if stride is -1, return a default packed stride
|
||||
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
return static_cast<std::size_t>(col);
|
||||
}
|
||||
else
|
||||
{
|
||||
return static_cast<std::size_t>(row);
|
||||
}
|
||||
}
|
||||
else
|
||||
return static_cast<std::size_t>(stride);
|
||||
};
|
||||
|
||||
ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
|
||||
|
||||
StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
|
||||
StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
|
||||
StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
|
||||
|
||||
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<BDataType> b_k_n_permute(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<BScaleDataType> b1_k_n(f_host_tensor_descriptor((K + Scale_Block_K - 1) / Scale_Block_K,
|
||||
(N + Scale_Block_N - 1) / Scale_Block_N,
|
||||
Scale_Stride_BN,
|
||||
BLayout{}));
|
||||
|
||||
switch(config.init_method)
|
||||
{
|
||||
case 0:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
|
||||
break;
|
||||
case 1:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
|
||||
break;
|
||||
case 2:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
|
||||
break;
|
||||
case 3:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
|
||||
break;
|
||||
case 4:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
|
||||
break;
|
||||
case 5:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
|
||||
break;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.5, 0.5});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
|
||||
}
|
||||
|
||||
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 << "b_k_n: " << b_k_n.mDesc << std::endl;
|
||||
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
|
||||
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n_permute.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b1_scale_device_buf(sizeof(BScaleDataType) * b1_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
// weight permute
|
||||
if constexpr(PermuteB)
|
||||
{
|
||||
int K1 = KPerBlock;
|
||||
int K0 = K / KPerBlock;
|
||||
|
||||
// int K0, N, K1
|
||||
for(int j = 0; j < K0; j++)
|
||||
{
|
||||
for(int i = 0; i < N; i++)
|
||||
{
|
||||
for(int jj = 0; jj < K1; jj++)
|
||||
{
|
||||
b_k_n_permute(j * N * K1 + i * K1 + jj) = b_k_n(i * K + (j * K1 + jj));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
for(int i = 0; i < N; i++)
|
||||
{
|
||||
for(int j = 0; j < K; j++)
|
||||
{
|
||||
b_k_n_permute(i * K + j) = b_k_n(i * K + j);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// vector pk_i4x4 permute
|
||||
for(int i = 0; i < N; i++)
|
||||
{
|
||||
for(int j = 0; j < K; j += 8)
|
||||
{
|
||||
int input[8];
|
||||
|
||||
for(int k = 0; k < 4; k++)
|
||||
{
|
||||
int i4x2 = b_k_n_permute(j + k * 2, i).data;
|
||||
input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
|
||||
input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
|
||||
}
|
||||
|
||||
// permute 01234567->20643175
|
||||
{
|
||||
int hi = input[2];
|
||||
int lo = input[0];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 0, i) = i4x2;
|
||||
}
|
||||
|
||||
{
|
||||
int hi = input[6];
|
||||
int lo = input[4];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 2, i) = i4x2;
|
||||
}
|
||||
|
||||
{
|
||||
int hi = input[3];
|
||||
int lo = input[1];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 4, i) = i4x2;
|
||||
}
|
||||
|
||||
{
|
||||
int hi = input[7];
|
||||
int lo = input[5];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 6, i) = i4x2;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
|
||||
b_k_n_device_buf.ToDevice(b_k_n_permute.mData.data());
|
||||
b1_scale_device_buf.ToDevice(b1_k_n.mData.data());
|
||||
DeviceMem workspace;
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto c_element_op = CElementOp{};
|
||||
|
||||
// do GEMM
|
||||
auto gemm = DeviceGemmV2Instance{};
|
||||
auto invoker = gemm.MakeInvoker();
|
||||
float ave_time = 0;
|
||||
|
||||
auto argument =
|
||||
gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
Scale_Stride_BN,
|
||||
static_cast<BScaleDataType*>(b1_scale_device_buf.GetDeviceBuffer()),
|
||||
KBatch,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
{
|
||||
std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
bool pass = true;
|
||||
if(config.do_verification)
|
||||
{
|
||||
Tensor<float> b_k_n_dequant({K, N});
|
||||
|
||||
float v_b = 0;
|
||||
for(int n = 0; n < N; n++)
|
||||
{
|
||||
for(int k = 0; k < K; k++)
|
||||
{
|
||||
ck::pk_i4_t i4x2 = b_k_n(k, n).data;
|
||||
int8_t i4 = 0;
|
||||
if(k % 2 == 1)
|
||||
i4 = (i4x2.data >> 0) & 0xf;
|
||||
else
|
||||
i4 = (i4x2.data >> 4) & 0xf;
|
||||
i4 = i4 - 8;
|
||||
v_b = ck::type_convert<float>(i4);
|
||||
|
||||
b_k_n_dequant(k, n) =
|
||||
ck::type_convert<float>(v_b) *
|
||||
ck::type_convert<float>(b1_k_n(k / Scale_Block_K, n / Scale_Block_N));
|
||||
}
|
||||
}
|
||||
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
a_m_k, b_k_n_dequant, c_m_n_host_result, PassThrough{}, PassThrough{}, PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
ave_time = invoker.Run(argument, StreamConfig{nullptr, false, 0});
|
||||
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
||||
|
||||
pass &= ck::utils::check_err(c_m_n_device_result,
|
||||
c_m_n_host_result,
|
||||
"Error: Incorrect results!",
|
||||
get_rtol<CDataType>(),
|
||||
get_atol<CDataType>());
|
||||
}
|
||||
|
||||
if(config.time_kernel)
|
||||
{
|
||||
ave_time =
|
||||
invoker.Run(argument, StreamConfig{nullptr, config.time_kernel, 0, 20, 50, true, 50});
|
||||
|
||||
std::size_t flop = 2_uz * M * N * K;
|
||||
std::size_t num_btype =
|
||||
sizeof(ADataType) * M * K +
|
||||
sizeof(BDataType) * K * N /
|
||||
(ck::is_same_v<ck::remove_cvref_t<BDataType>, ck::pk_i4_t> ? 2 : 1) +
|
||||
sizeof(CDataType) * M * N;
|
||||
|
||||
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, " << gemm.GetTypeString() << std::endl;
|
||||
}
|
||||
return pass;
|
||||
}
|
||||
|
||||
bool run_gemm_splitk_example(int argc, char* argv[])
|
||||
{
|
||||
ProblemSizeSplitK problem_size;
|
||||
ExecutionConfig config;
|
||||
|
||||
return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
|
||||
}
|
||||
|
||||
int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }
|
||||
@@ -0,0 +1,167 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_b_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_b_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_b_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4_b_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
enum struct BlockGemmPipelineVersion
|
||||
{
|
||||
v1, // Naive
|
||||
v2, // Mem
|
||||
v3, // Comp
|
||||
v4, // Comp, double lds buffer
|
||||
v5, // Comp, double global prefetch register buffer
|
||||
};
|
||||
|
||||
template <BlockGemmPipelineVersion BlkGemmPipelineVer,
|
||||
BlockGemmPipelineScheduler BlkGemmPipeSche,
|
||||
index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
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 BlockGemmPipeline_Selector()
|
||||
{
|
||||
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v1_b_scale<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
}
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v2_b_scale<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
}
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v3_b_scale<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
}
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v4_b_scale<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
}
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v5)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v5<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cerr << "BlockGemmPipeline configuration is not available" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,403 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// Naive pipeline with lowest resource request per WGP
|
||||
// GlobalPrefetchStages: 1
|
||||
// LocalPreFillStages: 1
|
||||
// LocalPreFetchStages: 0
|
||||
// LocalSharedMemoryBuffer: 1
|
||||
|
||||
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
|
||||
index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
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 KPacks>
|
||||
struct BlockwiseGemmXdlops_pipeline_v1_b_scale
|
||||
{
|
||||
};
|
||||
|
||||
template <index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
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
|
||||
// ,bool TransposeC //disable transposec right now...
|
||||
>
|
||||
struct BlockwiseGemmXdlops_pipeline_v1_b_scale<BlockGemmPipelineScheduler::Intrawave,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
|
||||
{
|
||||
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>;
|
||||
using Base::I0;
|
||||
using Base::KRepeat;
|
||||
using Base::xdlops_gemm;
|
||||
|
||||
using Base::CalculateCThreadOriginDataIndex;
|
||||
using Base::CalculateCThreadOriginDataIndex8D;
|
||||
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::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;
|
||||
|
||||
static constexpr index_t PrefetchStages = 1;
|
||||
static constexpr index_t PrefillStages = 1;
|
||||
static constexpr index_t GlobalBufferNum = 1;
|
||||
|
||||
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
|
||||
{
|
||||
return num_loop > PrefetchStages;
|
||||
}
|
||||
|
||||
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
|
||||
{
|
||||
ignore = num_loop;
|
||||
return TailNumber::Full;
|
||||
}
|
||||
|
||||
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,
|
||||
// BScale Thread Copy
|
||||
typename BScaleGridBuffer,
|
||||
typename BScaleGridDesc,
|
||||
typename BScaleThreadDesc,
|
||||
typename BScaleThreadTransfer,
|
||||
typename BScaleThreadTransferStep>
|
||||
__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,
|
||||
// BScaleThreadCopy
|
||||
const BScaleGridDesc& b_scale_grid_desc,
|
||||
const BScaleThreadDesc& b_scale_thread_desc,
|
||||
BScaleThreadTransfer& b_scale_thread_copy,
|
||||
const BScaleGridBuffer& b_scale_grid_buf,
|
||||
const BScaleThreadTransferStep& b_scale_thread_copy_step,
|
||||
// num_loop
|
||||
index_t num_loop,
|
||||
index_t num_loop_per_scale) const
|
||||
{
|
||||
// assume kperblock = scaleblockk
|
||||
ignore = num_loop_per_scale;
|
||||
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
a_thread_desc_.GetElementSpaceSize());
|
||||
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
b_thread_desc_.GetElementSpaceSize());
|
||||
|
||||
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
b_scale_thread_desc.GetElementSpaceSize());
|
||||
|
||||
// Global prefetch 1
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(n0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<0>{}));
|
||||
});
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<1>{}));
|
||||
|
||||
// Local prefill 1
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
// Initialize C
|
||||
c_thread_buf.Clear();
|
||||
|
||||
auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
|
||||
|
||||
// main body
|
||||
if constexpr(HasMainLoop)
|
||||
{
|
||||
index_t i = 0;
|
||||
do
|
||||
{
|
||||
// -------------------------------------------------------------------------------------------
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_buf);
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run<>(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[n0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(n0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(
|
||||
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
|
||||
});
|
||||
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<1>{}));
|
||||
|
||||
block_sync_lds();
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
i += 1;
|
||||
|
||||
} while(i < (num_loop - 1));
|
||||
}
|
||||
|
||||
// tail
|
||||
if constexpr(TailNum == TailNumber::Full)
|
||||
{
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_buf);
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
c_thread_buf_per_scale.Clear();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
xdlops_gemm.template Run<>(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf_per_scale.GetVectorTypeReference(I0));
|
||||
});
|
||||
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
|
||||
c_thread_buf(Number<c_offset>{}) +=
|
||||
c_thread_buf_per_scale[Number<t>{}] *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[n0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
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_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,530 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// Compute optimized pipeline
|
||||
// GlobalPrefetchStages: 2
|
||||
// LocalPreFillStages: 1
|
||||
// LocalPreFetchStages: 1
|
||||
// LocalSharedMemoryBuffer: 1
|
||||
|
||||
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
|
||||
index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
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 KPacks>
|
||||
struct BlockwiseGemmXdlops_pipeline_v3_b_scale
|
||||
{
|
||||
};
|
||||
|
||||
template <index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
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
|
||||
// ,bool TransposeC //disable transposec right now...
|
||||
>
|
||||
struct BlockwiseGemmXdlops_pipeline_v3_b_scale<BlockGemmPipelineScheduler::Intrawave,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
|
||||
{
|
||||
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
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::xdlops_gemm;
|
||||
using typename Base::HotLoopInstList;
|
||||
|
||||
using Base::CalculateCThreadOriginDataIndex;
|
||||
using Base::CalculateCThreadOriginDataIndex8D;
|
||||
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::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;
|
||||
|
||||
static constexpr index_t PrefetchStages = 2;
|
||||
static constexpr index_t PrefillStages = 1;
|
||||
static constexpr index_t GlobalBufferNum = 1;
|
||||
|
||||
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
|
||||
{
|
||||
return num_loop > PrefetchStages;
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
|
||||
{
|
||||
ignore = num_loop;
|
||||
return TailNumber::Full;
|
||||
}
|
||||
|
||||
__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 = NPerXDL == 16 ? 16 : 32;
|
||||
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
|
||||
});
|
||||
}
|
||||
|
||||
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 BScaleGridBuffer,
|
||||
typename BScaleGridDesc,
|
||||
typename BScaleThreadDesc,
|
||||
typename BScaleThreadTransfer,
|
||||
typename BScaleThreadTransferStep>
|
||||
__device__ void Run(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,
|
||||
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,
|
||||
CThreadBuffer& c_thread_buf,
|
||||
// BScaleThreadCopy
|
||||
const BScaleGridDesc& b_scale_grid_desc,
|
||||
const BScaleThreadDesc& b_scale_thread_desc,
|
||||
BScaleThreadTransfer& b_scale_thread_copy,
|
||||
const BScaleGridBuffer& b_scale_grid_buf,
|
||||
const BScaleThreadTransferStep& b_scale_thread_copy_step,
|
||||
// num loop
|
||||
index_t num_loop,
|
||||
index_t num_loop_per_scale) const
|
||||
{
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
|
||||
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
a_thread_desc_.GetElementSpaceSize());
|
||||
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
b_thread_desc_.GetElementSpaceSize());
|
||||
|
||||
// B scale buffer
|
||||
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
b_scale_thread_desc.GetElementSpaceSize());
|
||||
|
||||
// Global prefetch 1
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(n0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<0>{}));
|
||||
});
|
||||
|
||||
if(num_loop_per_scale == 1)
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<2>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<1>{}));
|
||||
}
|
||||
|
||||
constexpr auto num_scale_k_block = BScaleThreadDesc{}.GetLength(I1);
|
||||
constexpr auto num_scale_krepeat = KRepeat / num_scale_k_block;
|
||||
|
||||
// Local prefill 1
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
// Global prefetch 2
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
// Initialize C
|
||||
c_thread_buf.Clear();
|
||||
|
||||
// Local prefetch 1
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k0, I0),
|
||||
a_thread_buf);
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(
|
||||
b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k0 * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_scale_thread_buf[Number<n0 * num_scale_k_block + k0 / num_scale_krepeat>{}],
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
|
||||
// main body
|
||||
if constexpr(HasMainLoop)
|
||||
{
|
||||
index_t i = 0;
|
||||
do
|
||||
{
|
||||
block_sync_lds();
|
||||
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
|
||||
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(n0, I0),
|
||||
b_scale_thread_buf);
|
||||
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(
|
||||
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
|
||||
});
|
||||
|
||||
if((i + 2) % num_loop_per_scale == 0)
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(
|
||||
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<2>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(
|
||||
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<1>{}));
|
||||
}
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
block_sync_lds();
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
|
||||
a_block_buf,
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k0, I0),
|
||||
a_thread_buf);
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k0 * BMmaKStride>{}),
|
||||
b_block_buf,
|
||||
b_scale_thread_buf[Number<n0 * num_scale_k_block +
|
||||
k0 / num_scale_krepeat>{}],
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
|
||||
i += 1;
|
||||
} while(i < (num_loop - 1));
|
||||
}
|
||||
// tail
|
||||
if constexpr(TailNum == TailNumber::Full)
|
||||
{
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
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_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,686 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
|
||||
|
||||
namespace ck {
|
||||
|
||||
// Compute optimimal pipeline with highest resource request
|
||||
// GlobalPrefetchStages: 4
|
||||
// LocalPreFillStages: 2
|
||||
// LocalPreFetchStages: 1
|
||||
// LocalSharedMemoryBuffer: 2
|
||||
|
||||
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
|
||||
index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
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 KPacks>
|
||||
struct BlockwiseGemmXdlops_pipeline_v4_b_scale
|
||||
{
|
||||
};
|
||||
|
||||
template <index_t BlockSize,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename ComputeDataType,
|
||||
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
|
||||
// ,bool TransposeC //disable transposec right now...
|
||||
>
|
||||
struct BlockwiseGemmXdlops_pipeline_v4_b_scale<BlockGemmPipelineScheduler::Intrawave,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>
|
||||
|
||||
{
|
||||
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
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::xdlops_gemm;
|
||||
using typename Base::HotLoopInstList;
|
||||
|
||||
using Base::CalculateCThreadOriginDataIndex;
|
||||
using Base::CalculateCThreadOriginDataIndex8D;
|
||||
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::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;
|
||||
|
||||
static constexpr index_t PrefetchStages = 3;
|
||||
static constexpr index_t PrefillStages = 2;
|
||||
static constexpr index_t GlobalBufferNum = 1;
|
||||
static constexpr index_t HotloopUnroll = 2;
|
||||
|
||||
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
|
||||
{
|
||||
return num_loop > PrefetchStages;
|
||||
}
|
||||
|
||||
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
|
||||
{
|
||||
if(num_loop % HotloopUnroll == 1)
|
||||
{
|
||||
return TailNumber::Odd;
|
||||
}
|
||||
else
|
||||
{
|
||||
return TailNumber::Even;
|
||||
}
|
||||
}
|
||||
|
||||
__device__ static constexpr void HotLoopScheduler()
|
||||
{
|
||||
// TODO: Take data type into consideration as pipe ver 3
|
||||
// A-B splited schedule
|
||||
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_issue_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
|
||||
constexpr auto num_dswrite_per_issue_a =
|
||||
(HotLoopInstList::A_LDS_Write_Inst_Num + num_issue_a - 1) / num_issue_a;
|
||||
constexpr auto num_dsread_per_issue_a = num_ds_read_inst_a / num_issue_a;
|
||||
|
||||
constexpr auto num_issue_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
|
||||
constexpr auto num_dswrite_per_issue_b =
|
||||
(HotLoopInstList::B_LDS_Write_Inst_Num + num_issue_b - 1) / num_issue_b;
|
||||
constexpr auto num_dsread_per_issue_b = num_ds_read_inst_b / num_issue_b;
|
||||
|
||||
constexpr auto num_mfma_per_issue =
|
||||
HotLoopInstList::C_MFMA_Inst_Num / (num_issue_a + num_issue_b);
|
||||
|
||||
static_for<0, num_issue_a, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
static_for<0, num_dsread_per_issue_a, 1>{}([&](auto idsread) {
|
||||
ignore = idsread;
|
||||
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
|
||||
});
|
||||
|
||||
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_dsread_per_issue_a -
|
||||
num_dswrite_per_issue_a,
|
||||
0); // MFMA
|
||||
});
|
||||
|
||||
static_for<0, num_issue_b, 1>{}([&](auto i) {
|
||||
ignore = i;
|
||||
static_for<0, num_dsread_per_issue_b, 1>{}([&](auto idsread) {
|
||||
ignore = idsread;
|
||||
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
|
||||
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
|
||||
});
|
||||
|
||||
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_dsread_per_issue_a -
|
||||
num_dswrite_per_issue_b,
|
||||
0); // MFMA
|
||||
});
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
}
|
||||
|
||||
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 BScaleGridBuffer,
|
||||
typename BScaleGridDesc,
|
||||
typename BScaleThreadDesc,
|
||||
typename BScaleThreadTransfer,
|
||||
typename BScaleThreadTransferStep>
|
||||
__device__ void Run(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,
|
||||
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,
|
||||
CThreadBuffer& c_thread_buf,
|
||||
// BScaleThreadCopy
|
||||
const BScaleGridDesc& b_scale_grid_desc,
|
||||
const BScaleThreadDesc& b_scale_thread_desc,
|
||||
BScaleThreadTransfer& b_scale_thread_copy,
|
||||
const BScaleGridBuffer& b_scale_grid_buf,
|
||||
const BScaleThreadTransferStep& b_scale_thread_copy_step,
|
||||
// num loop
|
||||
index_t num_loop,
|
||||
index_t num_loop_per_scale) const
|
||||
{
|
||||
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
a_thread_desc_.GetElementSpaceSize());
|
||||
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
b_thread_desc_.GetElementSpaceSize());
|
||||
|
||||
// B scale buffer
|
||||
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
|
||||
b_scale_thread_desc.GetElementSpaceSize());
|
||||
|
||||
StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
|
||||
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
|
||||
StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs;
|
||||
|
||||
// Global prefetch 1
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(n0, I0),
|
||||
b_scale_thread_bufs(I0));
|
||||
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<0>{}));
|
||||
});
|
||||
|
||||
if(num_loop_per_scale == 1)
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<2>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<1>{}));
|
||||
}
|
||||
|
||||
// Local prefill 1
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0));
|
||||
|
||||
// Global prefetch 2
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(n0, I0),
|
||||
b_scale_thread_bufs(I1));
|
||||
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<0>{}));
|
||||
});
|
||||
|
||||
if(2 % num_loop_per_scale == 0)
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<2>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<1>{}));
|
||||
}
|
||||
|
||||
// Local prefetch 1
|
||||
block_sync_lds();
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf.At(I0),
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_bufs(I0));
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf.At(I0),
|
||||
b_scale_thread_bufs(I0)[n0],
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_bufs(I0));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
// Local prefill 2
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1));
|
||||
|
||||
// Global prefetch 3
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(n0, I0),
|
||||
b_scale_thread_bufs(I0));
|
||||
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<0>{}));
|
||||
});
|
||||
|
||||
if(3 % num_loop_per_scale == 0)
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<2>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
|
||||
b_scale_thread_copy_step.At(Number<1>{}));
|
||||
}
|
||||
|
||||
// Initialize C
|
||||
c_thread_buf.Clear();
|
||||
|
||||
// main body
|
||||
if constexpr(HasMainLoop)
|
||||
{
|
||||
index_t i = 0;
|
||||
// This hot loop has two legacy loopover, to implement the double local buffer strategy
|
||||
do
|
||||
{
|
||||
auto LoopFunc = [&](auto lds_read_buf,
|
||||
auto lds_read_reg_buf,
|
||||
auto lds_write_buf,
|
||||
auto mfma_reg_buf) {
|
||||
block_sync_lds();
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf.At(lds_read_buf),
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_bufs(lds_read_reg_buf));
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf.At(lds_read_buf),
|
||||
b_scale_thread_bufs(lds_read_buf)[n0],
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_bufs(lds_read_reg_buf));
|
||||
});
|
||||
});
|
||||
|
||||
// B scale copy
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_scale_thread_copy.Run(b_scale_grid_desc,
|
||||
b_scale_grid_buf,
|
||||
b_scale_thread_desc,
|
||||
make_tuple(n0, I0),
|
||||
b_scale_thread_bufs(lds_read_reg_buf));
|
||||
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(
|
||||
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
|
||||
});
|
||||
|
||||
if((i + 4 + mfma_reg_buf.value) % num_loop_per_scale == 0)
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(
|
||||
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<2>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
b_scale_thread_copy.MoveSrcSliceWindow(
|
||||
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<1>{}));
|
||||
}
|
||||
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf));
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf));
|
||||
|
||||
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
|
||||
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
|
||||
|
||||
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
|
||||
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_bufs[mfma_reg_buf]
|
||||
[Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_bufs[mfma_reg_buf]
|
||||
[Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType,
|
||||
xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.Run(
|
||||
a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler();
|
||||
};
|
||||
|
||||
LoopFunc(I1, I1, I0, I0);
|
||||
LoopFunc(I0, I0, I1, I1);
|
||||
|
||||
i += HotloopUnroll;
|
||||
} while(i < (num_loop - PrefetchStages));
|
||||
}
|
||||
|
||||
auto ReadWriteCompFunc = [&](auto lds_read_buf,
|
||||
auto lds_read_reg_buf,
|
||||
auto lds_write_buf,
|
||||
auto mfma_reg_buf) {
|
||||
block_sync_lds();
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf.At(lds_read_buf),
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_bufs(lds_read_reg_buf));
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf.At(lds_read_buf),
|
||||
b_scale_thread_bufs(lds_read_buf)[n0],
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_bufs(lds_read_reg_buf));
|
||||
});
|
||||
});
|
||||
|
||||
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf));
|
||||
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf));
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler();
|
||||
};
|
||||
|
||||
auto ReadCompFunc = [&](auto lds_read_buf, auto lds_read_reg_buf, auto mfma_reg_buf) {
|
||||
block_sync_lds();
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
|
||||
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
|
||||
a_block_buf.At(lds_read_buf),
|
||||
a_thread_desc_,
|
||||
make_tuple(m0, I0, k, I0),
|
||||
a_thread_bufs(lds_read_reg_buf));
|
||||
});
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
|
||||
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
|
||||
b_block_buf.At(lds_read_buf),
|
||||
b_scale_thread_bufs(lds_read_buf)[n0],
|
||||
b_thread_desc_,
|
||||
make_tuple(n0, I0, k, I0),
|
||||
b_thread_bufs(lds_read_reg_buf));
|
||||
});
|
||||
});
|
||||
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
HotLoopScheduler();
|
||||
};
|
||||
|
||||
auto CompFunc = [&](auto mfma_reg_buf) {
|
||||
static_for<0, KRepeat, 1>{}([&](auto k0) {
|
||||
static_for<0, MRepeat, 1>{}([&](auto m0) {
|
||||
static_for<0, NRepeat, 1>{}([&](auto n0) {
|
||||
vector_type<ComputeDataType, KPack> a_thread_vec;
|
||||
vector_type<ComputeDataType, KPack> b_thread_vec;
|
||||
|
||||
static_for<0, KPack, 1>{}([&](auto ik) {
|
||||
a_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
|
||||
make_tuple(m0, I0, k0, ik))>{}];
|
||||
b_thread_vec.template AsType<ComputeDataType>()(ik) =
|
||||
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
|
||||
make_tuple(n0, I0, k0, ik))>{}];
|
||||
});
|
||||
|
||||
using mfma_input_type =
|
||||
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
|
||||
|
||||
constexpr index_t c_offset =
|
||||
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
|
||||
|
||||
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
|
||||
b_thread_vec.template AsType<mfma_input_type>(),
|
||||
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
|
||||
});
|
||||
});
|
||||
});
|
||||
};
|
||||
|
||||
// tail
|
||||
if constexpr(TailNum == TailNumber::Odd)
|
||||
{
|
||||
ReadWriteCompFunc(I1, I1, I0, I0);
|
||||
ReadCompFunc(I0, I0, I1);
|
||||
CompFunc(I0);
|
||||
}
|
||||
else if constexpr(TailNum == TailNumber::Even)
|
||||
{
|
||||
ReadCompFunc(I1, I1, I0);
|
||||
CompFunc(I1);
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
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_;
|
||||
};
|
||||
|
||||
} // namespace ck
|
||||
@@ -77,6 +77,43 @@ struct DeviceGemmV2R1 : public BaseOperator
|
||||
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
|
||||
};
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename BScaleType,
|
||||
typename CDataType,
|
||||
index_t ScaleBlockN,
|
||||
index_t ScaleBlockK,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CElementwiseOperation>
|
||||
struct DeviceGemmV2BScale : public BaseOperator
|
||||
{
|
||||
virtual std::unique_ptr<BaseArgument>
|
||||
MakeArgumentPointer(const void* p_a,
|
||||
const void* p_b,
|
||||
void* p_c,
|
||||
ck::index_t M,
|
||||
ck::index_t N,
|
||||
ck::index_t K,
|
||||
ck::index_t StrideA,
|
||||
ck::index_t StrideB,
|
||||
ck::index_t StrideC,
|
||||
ck::index_t StrideScaleB,
|
||||
const void* p_b_scale,
|
||||
ck::index_t KSplit,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CElementwiseOperation c_element_op) = 0;
|
||||
|
||||
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
|
||||
|
||||
virtual bool GetPermuteB() = 0;
|
||||
virtual ck::index_t GetKPerBlock() = 0;
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
|
||||
@@ -0,0 +1,781 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, 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_v2.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_b_scale.hpp"
|
||||
#include "ck/host_utility/device_prop.hpp"
|
||||
#include "ck/host_utility/kernel_launch.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename BScaleDataType,
|
||||
typename CDataType,
|
||||
typename GemmAccDataType,
|
||||
typename CShuffleDataType,
|
||||
typename AElementwiseOperation,
|
||||
typename BElementwiseOperation,
|
||||
typename CElementwiseOperation,
|
||||
GemmSpecialization GemmSpec,
|
||||
index_t BlockSize,
|
||||
index_t ScaleBlockN, // scale block for N
|
||||
index_t ScaleBlockK, // scale block for K
|
||||
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 = CDataType,
|
||||
typename ComputeTypeB = ComputeTypeA,
|
||||
bool PermuteA = false,
|
||||
bool PermuteB = false>
|
||||
struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2BScale<ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
BScaleDataType,
|
||||
CDataType,
|
||||
ScaleBlockN,
|
||||
ScaleBlockK,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CElementwiseOperation>
|
||||
{
|
||||
// GridwiseGemm
|
||||
using GridwiseGemm = GridwiseGemm_xdl_cshuffle_v3<
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
GemmAccDataType,
|
||||
CShuffleDataType,
|
||||
CDataType,
|
||||
AElementwiseOperation,
|
||||
BElementwiseOperation,
|
||||
CElementwiseOperation,
|
||||
GemmSpec,
|
||||
BlockSize,
|
||||
ScaleBlockN,
|
||||
ScaleBlockK,
|
||||
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,
|
||||
PermuteA,
|
||||
PermuteB>;
|
||||
|
||||
using Argument = typename GridwiseGemm::Argument;
|
||||
|
||||
// Invoker
|
||||
struct Invoker : public BaseInvoker
|
||||
{
|
||||
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
|
||||
{
|
||||
if(stream_config.log_level_ > 0)
|
||||
{
|
||||
arg.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);
|
||||
}
|
||||
};
|
||||
|
||||
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 ||
|
||||
BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
|
||||
{
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
// Tail number could be One to Seven
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
|
||||
{
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::One>;
|
||||
Run(kernel);
|
||||
}
|
||||
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Full)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Full>;
|
||||
Run(kernel);
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Two>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Three)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Three>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Four)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Four>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Five)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Five>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Six>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Seven)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Seven>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::One>;
|
||||
Run(kernel);
|
||||
}
|
||||
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Full)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Full>;
|
||||
Run(kernel);
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Two>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Three)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Three>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Four)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Four>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Five)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Five>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Six>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
|
||||
TailNumber::Seven)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Seven>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
// Tail number could be Odd or Even
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
|
||||
{
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Odd>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
|
||||
GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Even>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Odd>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Even>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Odd>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy,
|
||||
TailNumber::Even>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Odd>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
true,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy,
|
||||
TailNumber::Even>;
|
||||
Run(kernel);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
// Tail number always 1
|
||||
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
|
||||
{
|
||||
if(arg.KBatch > 1)
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
false,
|
||||
InMemoryDataOperationEnum::AtomicAdd,
|
||||
minimum_occupancy>;
|
||||
Run(kernel);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto kernel =
|
||||
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
|
||||
false,
|
||||
InMemoryDataOperationEnum::Set,
|
||||
minimum_occupancy>;
|
||||
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()
|
||||
{
|
||||
// TODO: properly implement this check
|
||||
return true;
|
||||
}
|
||||
|
||||
static bool IsSupportedArgument(const Argument& arg)
|
||||
{
|
||||
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));
|
||||
}
|
||||
|
||||
index_t GetKPerBlock() override { return KPerBlock; }
|
||||
|
||||
bool GetPermuteB() override { return PermuteB; }
|
||||
|
||||
static auto MakeArgument(const ADataType* p_a,
|
||||
const BDataType* p_b,
|
||||
CDataType* p_c,
|
||||
index_t M,
|
||||
index_t N,
|
||||
index_t K,
|
||||
index_t StrideA,
|
||||
index_t StrideB,
|
||||
index_t StrideC,
|
||||
index_t StrideScaleB,
|
||||
const BScaleDataType* p_b_scale,
|
||||
index_t KBatch,
|
||||
AElementwiseOperation a_element_op,
|
||||
BElementwiseOperation b_element_op,
|
||||
CElementwiseOperation c_element_op)
|
||||
{
|
||||
return Argument{p_a,
|
||||
p_b,
|
||||
p_c,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
StrideScaleB,
|
||||
p_b_scale,
|
||||
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_b,
|
||||
void* p_c,
|
||||
index_t M,
|
||||
index_t N,
|
||||
index_t K,
|
||||
index_t StrideA,
|
||||
index_t StrideB,
|
||||
index_t StrideC,
|
||||
index_t StrideScaleB,
|
||||
const void* p_b_scale,
|
||||
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 BDataType*>(p_b),
|
||||
static_cast<CDataType*>(p_c),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
StrideScaleB,
|
||||
static_cast<const BScaleDataType*>(p_b_scale),
|
||||
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 << "DeviceGemmXdlUniversal"
|
||||
<< "<"
|
||||
<< 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;
|
||||
// clang-format on
|
||||
|
||||
return str.str();
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -44,6 +44,40 @@ __host__ __device__ inline half4_t pki4_to_half4(int q)
|
||||
return res.template AsType<half4_t>()[Number<0>{}];
|
||||
}
|
||||
|
||||
__host__ __device__ inline half4_t pki4_to_half4_scale(int q, const ck::half2_t& scale)
|
||||
{
|
||||
const int LO = 0x000f000f;
|
||||
const int HI = 0x00f000f0;
|
||||
const int EX = 0x64006400;
|
||||
|
||||
// Extract the two int4 at low bit and create two fp16 number.
|
||||
int lo = amd_assembly_and_or_b32(q, LO, EX);
|
||||
// Extract the two int4 at hight bit and create two fp16 number.
|
||||
int hi = amd_assembly_and_or_b32(q, HI, EX);
|
||||
|
||||
const int SUB = 0xE408E408; // half2 {-1032, -1032}
|
||||
const int MUL = 0x2c002c00; // half2 {1 / 16, 1 / 16}
|
||||
const int ADD = 0xd480d480; // half2 {-72, -72}
|
||||
|
||||
vector_type<half_t, 4> res;
|
||||
|
||||
res.template AsType<half2_t>()(Number<0>{}) =
|
||||
amd_assembly_pk_add_f16(bit_cast<half2_t>(lo), bit_cast<half2_t>(SUB));
|
||||
|
||||
res.template AsType<half2_t>()(Number<1>{}) = amd_assembly_pk_fma_f16(
|
||||
bit_cast<half2_t>(hi), bit_cast<half2_t>(MUL), bit_cast<half2_t>(ADD));
|
||||
|
||||
asm volatile("v_pk_mul_f16 %0, %1, %2"
|
||||
: "=v"(res.template AsType<half2_t>()(Number<0>{}))
|
||||
: "v"(res.template AsType<half2_t>()(Number<0>{})), "v"(scale));
|
||||
|
||||
asm volatile("v_pk_mul_f16 %0, %1, %2"
|
||||
: "=v"(res.template AsType<half2_t>()(Number<1>{}))
|
||||
: "v"(res.template AsType<half2_t>()(Number<1>{})), "v"(scale));
|
||||
|
||||
return res.template AsType<half4_t>()[Number<0>{}];
|
||||
}
|
||||
|
||||
__host__ __device__ inline half2_t pki4_to_half2(pk_i4_t q)
|
||||
{
|
||||
#if 1
|
||||
@@ -171,7 +205,42 @@ struct PassThroughPack8
|
||||
dst.template AsType<bhalf2_t>()(Number<3>{}) =
|
||||
pki4_to_bhalf2(src.template AsType<pk_i4_t>()[Number<3>{}]);
|
||||
|
||||
y = dst.template AsType<bhalf8_t>()[Number<0>{}];
|
||||
y = dst.template AsType<bhalf8_t>()[Number<0>{}];
|
||||
#endif
|
||||
}
|
||||
constexpr const static bool is_pack8_invocable = true;
|
||||
};
|
||||
|
||||
struct DequantPack8
|
||||
{
|
||||
template <typename Y, typename X, typename Z>
|
||||
__host__ __device__ void operator()(Y& y, const X& x, const Z& z) const;
|
||||
|
||||
__host__ __device__ constexpr void
|
||||
operator()(ck::half8_t& y, const ck::pk_i4x4_t& x, const ck::half2_t& z) const
|
||||
{
|
||||
#if 1
|
||||
vector_type<half_t, 8> result;
|
||||
|
||||
result.template AsType<half4_t>()(Number<0>{}) = pki4_to_half4_scale(bit_cast<int>(x), z);
|
||||
result.template AsType<half4_t>()(Number<1>{}) =
|
||||
pki4_to_half4_scale(bit_cast<int>(x) >> 8, z);
|
||||
|
||||
y = result.template AsType<half8_t>()[Number<0>{}];
|
||||
#else
|
||||
vector_type<half_t, 8> dst;
|
||||
vector_type<pk_i4_t, 4> src{x};
|
||||
|
||||
dst.template AsType<half2_t>()(Number<0>{}) =
|
||||
pki4_to_half2(src.template AsType<pk_i4_t>()[Number<0>{}]);
|
||||
dst.template AsType<half2_t>()(Number<1>{}) =
|
||||
pki4_to_half2(src.template AsType<pk_i4_t>()[Number<1>{}]);
|
||||
dst.template AsType<half2_t>()(Number<2>{}) =
|
||||
pki4_to_half2(src.template AsType<pk_i4_t>()[Number<2>{}]);
|
||||
dst.template AsType<half2_t>()(Number<3>{}) =
|
||||
pki4_to_half2(src.template AsType<pk_i4_t>()[Number<3>{}]);
|
||||
|
||||
y = dst.template AsType<half8_t>()[Number<0>{}];
|
||||
#endif
|
||||
}
|
||||
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1222,6 +1222,206 @@ struct ThreadwiseTensorSliceTransfer_v4
|
||||
});
|
||||
}
|
||||
|
||||
// Fuse scale
|
||||
template <typename SrcRefToOriginDisplacement,
|
||||
typename DstOriginIdx,
|
||||
typename SrcBuffer,
|
||||
typename DstBuffer>
|
||||
__device__ void Run(const SrcDesc&,
|
||||
const SrcRefToOriginDisplacement&,
|
||||
const SrcBuffer& src_buf,
|
||||
const DstData& scale,
|
||||
const DstDesc&,
|
||||
const DstOriginIdx&,
|
||||
DstBuffer& dst_buf) const
|
||||
{
|
||||
static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
|
||||
"wrong! SrcDesc and DstDesc need to known at compile-time");
|
||||
|
||||
static_assert(
|
||||
is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
|
||||
is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
|
||||
"wrong! SrcBuffer or DstBuffer data type is wrong");
|
||||
|
||||
static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
|
||||
|
||||
static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
|
||||
is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
|
||||
"wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
|
||||
"at compile-time");
|
||||
|
||||
// SrcDesc and DstDesc are known at compile-time
|
||||
constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
|
||||
constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
|
||||
|
||||
// SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
|
||||
constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
|
||||
constexpr auto dst_origin_idx = to_multi_index(DstOriginIdx{});
|
||||
|
||||
// scalar per access of each dim
|
||||
constexpr auto src_scalar_per_access = generate_sequence_v2(
|
||||
[&](auto i) constexpr {
|
||||
if constexpr(i == SrcVectorDim)
|
||||
{
|
||||
return Number<SrcScalarPerVector>{};
|
||||
}
|
||||
else
|
||||
{
|
||||
return Number<1>{};
|
||||
}
|
||||
},
|
||||
Number<nDim>{});
|
||||
|
||||
// scalar step (if steping on SrcVectorDim) of each dim
|
||||
constexpr auto src_scalar_step_in_vector = generate_sequence_v2(
|
||||
[&](auto i) constexpr {
|
||||
if constexpr(i == SrcVectorDim)
|
||||
{
|
||||
return Number<1>{};
|
||||
}
|
||||
else
|
||||
{
|
||||
return Number<0>{};
|
||||
}
|
||||
},
|
||||
Number<nDim>{});
|
||||
|
||||
constexpr auto access_lengths = SliceLengths{} / src_scalar_per_access;
|
||||
|
||||
constexpr auto dim_access_order = DimAccessOrder{};
|
||||
|
||||
constexpr auto ordered_access_lengths =
|
||||
container_reorder_given_new2old(access_lengths, dim_access_order);
|
||||
|
||||
static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
|
||||
#if 0
|
||||
// TODO: unable to compile
|
||||
// position in slice window
|
||||
constexpr auto data_to_origin_disp_idx =
|
||||
container_reorder_given_old2new(ordered_access_idx, dim_access_order) *
|
||||
src_scalar_per_access;
|
||||
#else
|
||||
// position in slice window
|
||||
constexpr auto data_to_origin_disp_idx =
|
||||
ordered_access_idx.ReorderGivenOld2New(dim_access_order) * src_scalar_per_access;
|
||||
#endif
|
||||
// src coordinate
|
||||
constexpr auto src_ref_to_data_disp_idx =
|
||||
src_ref_to_origin_disp_idx + data_to_origin_disp_idx;
|
||||
|
||||
constexpr auto src_ref_to_data_disp_coord_step =
|
||||
make_tensor_coordinate_step(src_desc, src_ref_to_data_disp_idx);
|
||||
|
||||
auto src_data_coord = src_ref_coord_;
|
||||
|
||||
move_tensor_coordinate(src_desc, src_data_coord, src_ref_to_data_disp_coord_step);
|
||||
|
||||
vector_type_maker_t<SrcData, SrcScalarPerVector / PackedSize> src_tmp_vector;
|
||||
|
||||
using src_vector_t = typename decltype(src_tmp_vector)::type;
|
||||
|
||||
const bool is_src_valid = coordinate_has_valid_offset_assuming_visible_index_is_valid(
|
||||
src_desc, src_data_coord);
|
||||
|
||||
// copy data from src_buf into src_tmp_vector
|
||||
if constexpr(SrcBuffer::IsDynamicBuffer())
|
||||
{
|
||||
src_tmp_vector.template AsType<src_vector_t>()(Number<0>{}) =
|
||||
src_buf.template Get<src_vector_t>(src_data_coord.GetOffset() / PackedSize,
|
||||
is_src_valid);
|
||||
}
|
||||
else if constexpr(SrcBuffer::IsStaticBuffer())
|
||||
{
|
||||
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
|
||||
constexpr index_t src_offset = src_desc.CalculateOffset(
|
||||
src_ref_to_origin_disp_idx + data_to_origin_disp_idx +
|
||||
i * src_scalar_step_in_vector);
|
||||
|
||||
src_tmp_vector.template AsType<SrcData>()(i) = src_buf[Number<src_offset>{}];
|
||||
});
|
||||
}
|
||||
|
||||
if constexpr(is_same<remove_cvref_t<SrcData>, pk_i4_t>::value)
|
||||
{
|
||||
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
|
||||
// DstData)
|
||||
vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
|
||||
vector_type<DstData, 2> scale_vector;
|
||||
scale_vector.template AsType<DstData>()(Number<0>{}) = scale;
|
||||
scale_vector.template AsType<DstData>()(Number<1>{}) = scale;
|
||||
|
||||
constexpr index_t pack_size = 8;
|
||||
|
||||
static_assert(SrcScalarPerVector % pack_size == 0, "");
|
||||
|
||||
using src_v_t = typename vector_type_maker_t<SrcData, pack_size / PackedSize>::type;
|
||||
using dst_v_t = typename vector_type_maker_t<DstData, pack_size>::type;
|
||||
using scale_v_t = typename vector_type_maker_t<DstData, 2>::type;
|
||||
|
||||
static_for<0, SrcScalarPerVector / pack_size, 1>{}([&](auto i) {
|
||||
ck::tensor_operation::element_wise::DequantPack8{}(
|
||||
dst_tmp_vector.template AsType<dst_v_t>()(i),
|
||||
src_tmp_vector.template AsType<src_v_t>()[i],
|
||||
scale_vector.template AsType<scale_v_t>()[Number<0>{}]);
|
||||
});
|
||||
|
||||
// copy data from dst_tmp_vector into dst_buf
|
||||
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
|
||||
constexpr index_t dst_offset = dst_desc.CalculateOffset(
|
||||
dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
|
||||
|
||||
dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
|
||||
});
|
||||
}
|
||||
else if constexpr(is_same<remove_cvref_t<SrcData>, f8_t>::value &&
|
||||
is_same<remove_cvref_t<DstData>, half_t>::value &&
|
||||
SrcScalarPerVector % 2 == 0)
|
||||
{
|
||||
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
|
||||
// DstData)
|
||||
vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
|
||||
|
||||
constexpr index_t pack_size = 2;
|
||||
|
||||
using dst_v_t = typename vector_type_maker_t<DstData, pack_size>::type;
|
||||
using src_v_t = typename vector_type_maker_t<SrcData, pack_size>::type;
|
||||
static_for<0, SrcScalarPerVector / pack_size, 1>{}([&](auto i) {
|
||||
ck::tensor_operation::element_wise::PassThroughPack2{}(
|
||||
dst_tmp_vector.template AsType<dst_v_t>()(i),
|
||||
src_tmp_vector.template AsType<src_v_t>()[i]);
|
||||
});
|
||||
|
||||
// copy data from dst_tmp_vector into dst_buf
|
||||
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
|
||||
constexpr index_t dst_offset = dst_desc.CalculateOffset(
|
||||
dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
|
||||
|
||||
dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
|
||||
});
|
||||
}
|
||||
else
|
||||
{
|
||||
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
|
||||
// DstData)
|
||||
vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
|
||||
|
||||
// TODO: if SrcData and DstData are vetor type, then static_cast may not compile
|
||||
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
|
||||
dst_tmp_vector.template AsType<DstData>()(i) =
|
||||
type_convert<DstData>(src_tmp_vector.template AsType<SrcData>()[i]);
|
||||
});
|
||||
|
||||
// copy data from dst_tmp_vector into dst_buf
|
||||
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
|
||||
constexpr index_t dst_offset = dst_desc.CalculateOffset(
|
||||
dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
|
||||
|
||||
dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
|
||||
});
|
||||
}
|
||||
});
|
||||
}
|
||||
|
||||
template <typename SrcSliceMoveStepIdx>
|
||||
__device__ void MoveSrcSliceWindow(const SrcDesc&,
|
||||
const SrcSliceMoveStepIdx& src_slice_move_step_idx)
|
||||
|
||||
@@ -4,8 +4,8 @@
|
||||
#ifndef CK_AMD_INLINE_ASM_HPP
|
||||
#define CK_AMD_INLINE_ASM_HPP
|
||||
|
||||
#include "data_type.hpp"
|
||||
#include "c_style_pointer_cast.hpp"
|
||||
#include "data_type.hpp"
|
||||
|
||||
// TODO: deprecate all amd_assembly_outer_product_xxx
|
||||
|
||||
@@ -21,14 +21,14 @@ inline __device__ int amd_assembly_and_or_b32(int a, int b, int d)
|
||||
inline __device__ half2_t amd_assembly_pk_fma_f16(half2_t a, half2_t b, half2_t c)
|
||||
{
|
||||
half2_t d;
|
||||
asm volatile("v_pk_fma_f16 %0, %1, %2, %3;\n" : "=v"(d) : "v"(a), "v"(b), "v"(c));
|
||||
asm volatile("v_pk_fma_f16 %0, %1, %2, %3" : "=v"(d) : "v"(a), "v"(b), "v"(c));
|
||||
return d;
|
||||
}
|
||||
|
||||
inline __device__ half2_t amd_assembly_pk_add_f16(half2_t a, half2_t b)
|
||||
{
|
||||
half2_t c;
|
||||
asm volatile("v_pk_add_f16 %0, %1, %2;\n" : "=v"(c) : "v"(a), "v"(b));
|
||||
asm volatile("v_pk_add_f16 %0, %1, %2" : "=v"(c) : "v"(a), "v"(b));
|
||||
return c;
|
||||
}
|
||||
|
||||
|
||||
@@ -19,6 +19,8 @@ struct pk_i4_t
|
||||
type data;
|
||||
__host__ __device__ constexpr pk_i4_t() : data{type{}} {}
|
||||
__host__ __device__ constexpr pk_i4_t(type init) : data{init} {}
|
||||
|
||||
__host__ __device__ constexpr operator float() const { return static_cast<int8_t>(data); }
|
||||
};
|
||||
|
||||
inline constexpr auto next_pow2(uint32_t x)
|
||||
|
||||
@@ -0,0 +1,91 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include <memory>
|
||||
#include <vector>
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
#if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_FP8))
|
||||
void add_device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmV2BScale<Row,
|
||||
Col,
|
||||
Row,
|
||||
F16,
|
||||
I4,
|
||||
F16,
|
||||
F16,
|
||||
1,
|
||||
128,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename BScaleDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout,
|
||||
index_t ScaleBlockK>
|
||||
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmV2BScale<
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
BScaleDataType,
|
||||
CDataType,
|
||||
1,
|
||||
ScaleBlockK,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough>>
|
||||
{
|
||||
using DeviceOp = DeviceGemmV2BScale<ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
BScaleDataType,
|
||||
CDataType,
|
||||
1,
|
||||
ScaleBlockK,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, pk_i4_t> &&
|
||||
is_same_v<CDataType, half_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<CLayout, Row>)
|
||||
{
|
||||
add_device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,10 @@
|
||||
# ONLY XDL_KERNELS
|
||||
set(GEMM_B_SCALE_INSTANCES)
|
||||
|
||||
list(APPEND GEMM_B_SCALE_INSTANCES
|
||||
device_gemm_b_scale_xdl_f16_i4_f16/device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instance.cpp
|
||||
)
|
||||
|
||||
set_source_files_properties(device_gemm_b_scale_xdl_f16_i4_f16/device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
|
||||
|
||||
add_instance_library(device_gemm_b_scale_instance ${GEMM_B_SCALE_INSTANCES})
|
||||
@@ -0,0 +1,105 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using I4 = pk_i4_t;
|
||||
using F16 = half_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = tensor_layout::gemm::RowMajor;
|
||||
using Col = tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
template <index_t... Is>
|
||||
using S = Sequence<Is...>;
|
||||
|
||||
using PassThrough = element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = GemmSpecialization::Default;
|
||||
static constexpr auto GemmKPadding = GemmSpecialization::KPadding;
|
||||
static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding;
|
||||
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
|
||||
|
||||
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
|
||||
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
|
||||
|
||||
#if 0
|
||||
template <GemmSpecialization GemmSpec>
|
||||
using device_gemm_xdl_b_scale_f16_i4_f16_mk_nk_mn_comp_instances = std::tuple<
|
||||
|
||||
#endif
|
||||
|
||||
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
|
||||
using device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_instances = std::tuple<
|
||||
// clang-format off
|
||||
//#########################| ALayout| BLayout| CLayout|AData| BData| BScale| CData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
|
||||
//#########################| | | | Type| Type| Data| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
|
||||
//#########################| | | | | | Type| | | | Operation| Operation| Operation| | | N| K| | | | | |Wave| Wave| | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
|
||||
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
|
||||
//Compute friendly
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 128, 8, 32, 32, 32, 2, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 8, 32, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 128, 8, 32, 32, 32, 2, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 8, 32, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 8, 32, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
|
||||
//Latency friendly
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 32, 16, 128, 8, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 16, 32, 128, 8, 32, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
|
||||
// Memory friendly v3
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 128, 8, 32, 32, 32, 2, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 128, 8, 16, 16, 16, 4, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 64, 32, 128, 8, 32, 32, 32, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 64, 16, 128, 8, 16, 16, 16, 2, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 32, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 16, 32, 128, 8, 32, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 16, 64, 128, 8, 32, 16, 16, 1, 2, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 32, 64, 128, 8, 32, 32, 32, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 16, 128, 128, 8, 32, 16, 16, 1, 4, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 32, 128, 128, 8, 32, 32, 32, 1, 2, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 16, 256, 128, 8, 32, 16, 16, 1, 4, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 32, 256, 128, 8, 32, 32, 32, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
|
||||
// Memory friendly v4
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 64, 32, 128, 8, 32, 32, 32, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 64, 16, 128, 8, 16, 16, 16, 2, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 32, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 16, 32, 128, 8, 32, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 16, 64, 128, 8, 32, 16, 16, 1, 2, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 32, 64, 128, 8, 32, 32, 32, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 16, 128, 128, 8, 32, 16, 16, 1, 4, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 32, 128, 128, 8, 32, 32, 32, 1, 2, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 16, 256, 128, 8, 32, 16, 16, 1, 4, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 32, 256, 128, 8, 32, 32, 32, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v4, half_t, half_t, false, false>,
|
||||
|
||||
//new Compute friendly kernel
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 8, 32, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 8, 32, 32, 32, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>,
|
||||
|
||||
//new Memory friendly kernel
|
||||
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 16, 64, 256, 8, 32, 16, 16, 1, 1, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3, half_t, half_t, false, false>
|
||||
// clang-format on
|
||||
>;
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,32 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
void add_device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmV2BScale<Row,
|
||||
Col,
|
||||
Row,
|
||||
F16,
|
||||
I4,
|
||||
F16,
|
||||
F16,
|
||||
1,
|
||||
128,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_instances<Intrawave, GemmDefault>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
448
profiler/include/profiler/profile_gemm_b_scale_impl.hpp
Normal file
448
profiler/include/profiler/profile_gemm_b_scale_impl.hpp
Normal file
@@ -0,0 +1,448 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iomanip>
|
||||
#include <iostream>
|
||||
#include <typeinfo>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/gpu/gemm_b_scale.hpp"
|
||||
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
#include "ck/library/utility/device_memory.hpp"
|
||||
#include "ck/library/utility/host_tensor.hpp"
|
||||
#include "ck/library/utility/host_tensor_generator.hpp"
|
||||
#include "ck/library/utility/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace profiler {
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename BScaleDataType,
|
||||
typename ComputeDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
index_t ScaleBlockK,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
bool profile_gemm_b_scale_impl(int do_verification,
|
||||
int init_method,
|
||||
bool do_log,
|
||||
bool time_kernel,
|
||||
int M,
|
||||
int N,
|
||||
int K,
|
||||
int StrideA,
|
||||
int StrideB,
|
||||
int StrideC,
|
||||
int KBatch,
|
||||
int n_warmup,
|
||||
int n_iter,
|
||||
uint64_t rotating = 0)
|
||||
{
|
||||
bool pass = true;
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
||||
using namespace ck::literals;
|
||||
|
||||
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {stride, 1_uz});
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {1_uz, stride});
|
||||
}
|
||||
};
|
||||
|
||||
ck::index_t Scale_Stride_BN = ck::is_same_v<BLayout, ck::tensor_layout::gemm::ColumnMajor>
|
||||
? ((K + ScaleBlockK - 1) / ScaleBlockK)
|
||||
: N;
|
||||
|
||||
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<BDataType> b_k_n_permute(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<BScaleDataType> b1_k_n(f_host_tensor_descriptor(
|
||||
(K + ScaleBlockK - 1) / ScaleBlockK, // K direction group size is ScaleBlockK
|
||||
N, // N direction group size is 1
|
||||
Scale_Stride_BN,
|
||||
BLayout{}));
|
||||
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{}));
|
||||
|
||||
int total_gemm_needed = a_m_k.GetElementSpaceSizeInBytes() +
|
||||
b_k_n.GetElementSpaceSizeInBytes() +
|
||||
b1_k_n.GetElementSpaceSizeInBytes();
|
||||
|
||||
int rotating_count = std::max(
|
||||
1,
|
||||
std::min(n_iter,
|
||||
static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
|
||||
|
||||
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
|
||||
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
|
||||
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
|
||||
std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
|
||||
std::cout << "rotating count: " << rotating_count << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
|
||||
break;
|
||||
case 2:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
|
||||
break;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
|
||||
}
|
||||
|
||||
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
const auto a_element_op = AElementOp{};
|
||||
const auto b_element_op = BElementOp{};
|
||||
const auto c_element_op = CElementOp{};
|
||||
|
||||
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n_permute.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b1_device_buf(sizeof(BScaleDataType) * b1_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a_device_buf.ToDevice(a_m_k.mData.data());
|
||||
b1_device_buf.ToDevice(b1_k_n.mData.data());
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGemmV2BScale<ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
BScaleDataType,
|
||||
CDataType,
|
||||
1,
|
||||
ScaleBlockK,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CElementOp>;
|
||||
|
||||
// get device op instances
|
||||
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
DeviceOp>::GetInstances();
|
||||
|
||||
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
|
||||
|
||||
// Run reference GEMM
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<float> b_k_n_dequant({K, N});
|
||||
|
||||
float v_b = 0;
|
||||
for(int n = 0; n < N; n++)
|
||||
{
|
||||
for(int k = 0; k < K; k++)
|
||||
{
|
||||
ck::pk_i4_t i4x2 = b_k_n(k, n).data;
|
||||
int8_t i4 = 0;
|
||||
if(k % 2 == 1)
|
||||
i4 = (i4x2.data >> 0) & 0xf;
|
||||
else
|
||||
i4 = (i4x2.data >> 4) & 0xf;
|
||||
i4 = i4 - 8;
|
||||
v_b = ck::type_convert<float>(i4);
|
||||
|
||||
b_k_n_dequant(k, n) = ck::type_convert<float>(v_b) *
|
||||
ck::type_convert<float>(b1_k_n(k / ScaleBlockK, n));
|
||||
}
|
||||
}
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
AccDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CElementOp,
|
||||
ComputeDataType>;
|
||||
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
a_m_k, b_k_n_dequant, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
}
|
||||
|
||||
std::string best_op_name;
|
||||
float best_ave_time = 0;
|
||||
float best_tflops = 0;
|
||||
float best_gb_per_sec = 0;
|
||||
float best_kbatch = 0;
|
||||
|
||||
// profile device GEMM instances
|
||||
for(auto& op_ptr : op_ptrs)
|
||||
{
|
||||
const int KPerBlock = op_ptr->GetKPerBlock();
|
||||
|
||||
if(op_ptr->GetPermuteB())
|
||||
{
|
||||
int K1 = KPerBlock;
|
||||
int K0 = K / KPerBlock;
|
||||
|
||||
// int K0, N, K1
|
||||
for(int j = 0; j < K0; j++)
|
||||
{
|
||||
for(int i = 0; i < N; i++)
|
||||
{
|
||||
for(int jj = 0; jj < K1; jj++)
|
||||
{
|
||||
b_k_n_permute(j * N * K1 + i * K1 + jj) = b_k_n(i * K + (j * K1 + jj));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if(is_same_v<BDataType, pk_i4_t> && is_same_v<ADataType, half_t>)
|
||||
{
|
||||
// vector pk_i4x4 permute
|
||||
for(int i = 0; i < N; i++)
|
||||
{
|
||||
for(int j = 0; j < K; j += 8)
|
||||
{
|
||||
int input[8];
|
||||
|
||||
for(int k = 0; k < 4; k++)
|
||||
{
|
||||
int i4x2 = b_k_n_permute(j + k * 2, i).data;
|
||||
input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
|
||||
input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
|
||||
}
|
||||
|
||||
// permute 01234567->20643175
|
||||
{
|
||||
int hi = input[2];
|
||||
int lo = input[0];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 0, i) = i4x2;
|
||||
}
|
||||
|
||||
{
|
||||
int hi = input[6];
|
||||
int lo = input[4];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 2, i) = i4x2;
|
||||
}
|
||||
|
||||
{
|
||||
int hi = input[3];
|
||||
int lo = input[1];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 4, i) = i4x2;
|
||||
}
|
||||
|
||||
{
|
||||
int hi = input[7];
|
||||
int lo = input[5];
|
||||
int i4x2 = (hi << 4) | lo;
|
||||
|
||||
b_k_n_permute(j + 6, i) = i4x2;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
b_k_n_permute = b_k_n;
|
||||
}
|
||||
|
||||
b_device_buf.ToDevice(b_k_n_permute.mData.data());
|
||||
|
||||
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
|
||||
|
||||
if(KBatch > 0)
|
||||
{
|
||||
kbatch_list = {KBatch};
|
||||
}
|
||||
|
||||
for(std::size_t i = 0; i < kbatch_list.size(); i++)
|
||||
{
|
||||
auto kbatch_curr = kbatch_list[i];
|
||||
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(
|
||||
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
Scale_Stride_BN,
|
||||
static_cast<BScaleDataType*>(b1_device_buf.GetDeviceBuffer()),
|
||||
kbatch_curr,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
|
||||
// re-init C to zero before profiling next kernel
|
||||
c_device_buf.SetZero();
|
||||
|
||||
invoker_ptr->Run(argument_ptr.get(),
|
||||
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
||||
|
||||
#if defined CK_ENABLE_FP8
|
||||
// set softer tolerances for fp8
|
||||
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
|
||||
is_same_v<CDataType, f8_t>)
|
||||
{
|
||||
std::string msg = "Error: Incorrect results!";
|
||||
double rtol = 1e-1;
|
||||
double atol = 1e-1;
|
||||
pass = pass & ck::utils::check_err(
|
||||
c_m_n_device_result, c_m_n_host_result, msg, rtol, atol);
|
||||
}
|
||||
else
|
||||
{
|
||||
#endif
|
||||
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
|
||||
#if defined CK_ENABLE_FP8
|
||||
}
|
||||
#endif
|
||||
|
||||
if(do_log)
|
||||
{
|
||||
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
|
||||
LogRangeAsType<int8_t>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
|
||||
LogRangeAsType<float>(
|
||||
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
|
||||
<< std::endl;
|
||||
LogRangeAsType<float>(
|
||||
std::cout << "c_device: ", c_m_n_device_result.mData, ",")
|
||||
<< std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
float ave_time = invoker_ptr->Run(argument_ptr.get(),
|
||||
StreamConfig{nullptr,
|
||||
time_kernel,
|
||||
0,
|
||||
n_warmup,
|
||||
n_iter,
|
||||
rotating_count > 1,
|
||||
rotating_count});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
|
||||
static constexpr index_t BPackedSize = []() {
|
||||
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
|
||||
return 2;
|
||||
else
|
||||
return 1;
|
||||
}();
|
||||
|
||||
std::size_t num_btype = sizeof(ADataType) * M * K +
|
||||
sizeof(BDataType) * K * N / BPackedSize +
|
||||
sizeof(CDataType) * M * N;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
|
||||
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
|
||||
<< kbatch_curr << std::endl;
|
||||
|
||||
if(tflops > best_tflops && ave_time > 1e-10)
|
||||
{
|
||||
best_op_name = op_name;
|
||||
best_tflops = tflops;
|
||||
best_ave_time = ave_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
best_kbatch = kbatch_curr;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << op_ptr->GetTypeString() << " does not support this problem"
|
||||
<< std::endl;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if constexpr(is_same<CDataType, float>::value)
|
||||
{
|
||||
std::cout << "Best Perf for datatype = f32";
|
||||
}
|
||||
else if constexpr(is_same<CDataType, half_t>::value)
|
||||
{
|
||||
std::cout << "Best Perf for datatype = f16";
|
||||
}
|
||||
else if constexpr(is_same<CDataType, bhalf_t>::value)
|
||||
{
|
||||
std::cout << "Best Perf for datatype = bf16";
|
||||
}
|
||||
else if constexpr(is_same<CDataType, int8_t>::value)
|
||||
{
|
||||
std::cout << "Best Perf for datatype = int8";
|
||||
}
|
||||
|
||||
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
std::cout << " ALayout = RowMajor";
|
||||
}
|
||||
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
|
||||
{
|
||||
std::cout << " ALayout = ColumnMajor";
|
||||
}
|
||||
|
||||
if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
std::cout << " BLayout = RowMajor";
|
||||
}
|
||||
else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
|
||||
{
|
||||
std::cout << " BLayout = ColumnMajor";
|
||||
}
|
||||
|
||||
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
|
||||
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << best_kbatch
|
||||
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
|
||||
<< " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
return pass;
|
||||
}
|
||||
|
||||
} // namespace profiler
|
||||
} // namespace ck
|
||||
@@ -58,6 +58,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
|
||||
list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
|
||||
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
|
||||
list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
|
||||
list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp)
|
||||
list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
|
||||
list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
|
||||
list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
|
||||
@@ -141,6 +142,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
|
||||
endif()
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
|
||||
|
||||
181
profiler/src/profile_gemm_b_scale.cpp
Normal file
181
profiler/src/profile_gemm_b_scale.cpp
Normal file
@@ -0,0 +1,181 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
#include <initializer_list>
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
|
||||
#include "profiler/profile_gemm_b_scale_impl.hpp"
|
||||
#include "profiler_operation_registry.hpp"
|
||||
|
||||
enum struct GemmMatrixLayout
|
||||
{
|
||||
MK_KN_MN, // 0
|
||||
MK_NK_MN, // 1
|
||||
KM_KN_MN, // 2
|
||||
KM_NK_MN, // 3
|
||||
};
|
||||
|
||||
enum struct GemmDataType
|
||||
{
|
||||
F32_F32_F32, // 0
|
||||
F16_F16_F16, // 1
|
||||
BF16_BF16_BF16, // 2
|
||||
INT8_INT8_INT8, // 3
|
||||
F8_F16_F16, // 4
|
||||
F16_F8_F16, // 5
|
||||
F16_F16_F16_F8, // 6
|
||||
F8_F8_BF16, // 7
|
||||
F16_I4_F16, // 8
|
||||
};
|
||||
|
||||
enum struct BScaleBlockTile
|
||||
{
|
||||
K_64, // 0
|
||||
K_128, // 1
|
||||
};
|
||||
|
||||
#define OP_NAME "gemm_b_scale"
|
||||
#define OP_DESC "Int4-dequant GEMM"
|
||||
|
||||
int profile_gemm_b_scale(int argc, char* argv[])
|
||||
{
|
||||
if(argc != 16 && argc != 19)
|
||||
{
|
||||
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
|
||||
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
|
||||
"f16->f8; 7: f8->bf16, "
|
||||
"comp f8; 8: f16@i4)\n");
|
||||
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
|
||||
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
|
||||
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
|
||||
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
|
||||
printf("arg4: B scale block tile (0: 64, 1: 128):\n");
|
||||
printf("arg5: verification (0: no; 1: yes)\n");
|
||||
printf("arg6: initialization (0: no init; 1: integer value; 2: decimal value)\n");
|
||||
printf("arg7: print tensor value (0: no; 1: yes)\n");
|
||||
printf("arg8: time kernel (0=no, 1=yes)\n");
|
||||
printf("arg9 to 14: M, N, K, StrideA, StrideB, StrideC\n");
|
||||
printf("arg15: split k into mulitiple batch\n");
|
||||
printf("optional:\n");
|
||||
printf("arg16: number of warm-up cycles (default 1)\n");
|
||||
printf("arg17: number of iterations (default 10)\n");
|
||||
printf("arg18: memory for rotating buffer (default 0, size in MB)\n");
|
||||
exit(1);
|
||||
}
|
||||
|
||||
printf("Start profiling\n");
|
||||
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
|
||||
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
|
||||
const auto B_scale_block = static_cast<BScaleBlockTile>(std::stoi(argv[4]));
|
||||
const bool do_verification = std::stoi(argv[5]);
|
||||
const int init_method = std::stoi(argv[6]);
|
||||
const bool do_log = std::stoi(argv[7]);
|
||||
const bool time_kernel = std::stoi(argv[8]);
|
||||
|
||||
const int M = std::stoi(argv[9]);
|
||||
const int N = std::stoi(argv[10]);
|
||||
const int K = std::stoi(argv[11]);
|
||||
|
||||
const int StrideA = std::stoi(argv[12]);
|
||||
const int StrideB = std::stoi(argv[13]);
|
||||
const int StrideC = std::stoi(argv[14]);
|
||||
const int KBatch = std::stoi(argv[15]);
|
||||
printf("M:%d, N:%d, K:%d, StrideA:%d, StrideB:%d, StrideC:%d, KBatch:%d\n",
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
KBatch);
|
||||
|
||||
int n_warmup = 1;
|
||||
int n_iter = 10;
|
||||
uint64_t rotating = 0;
|
||||
if(argc == 19)
|
||||
{
|
||||
n_warmup = std::stoi(argv[16]);
|
||||
n_iter = std::stoi(argv[17]);
|
||||
rotating = std::stoull(argv[18]) * 1024 * 1024;
|
||||
|
||||
printf("n_warmup:%d, n_iter:%d, rotating:%lu\n", n_warmup, n_iter, rotating);
|
||||
}
|
||||
|
||||
using F32 = float;
|
||||
using F16 = ck::half_t;
|
||||
using I4 = ck::pk_i4_t;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
auto profile = [&](auto a_type,
|
||||
auto b_type,
|
||||
auto b_scale_type,
|
||||
auto comp_type,
|
||||
auto acc_type,
|
||||
auto c_type,
|
||||
auto scale_block_k,
|
||||
auto a_layout,
|
||||
auto b_layout,
|
||||
auto c_layout) {
|
||||
using ADataType = decltype(a_type);
|
||||
using BDataType = decltype(b_type);
|
||||
using BScaleDataType = decltype(b_scale_type);
|
||||
using ComputeDataType = decltype(comp_type);
|
||||
using AccDataType = decltype(acc_type);
|
||||
using CDataType = decltype(c_type);
|
||||
|
||||
using ALayout = decltype(a_layout);
|
||||
using BLayout = decltype(b_layout);
|
||||
using CLayout = decltype(c_layout);
|
||||
|
||||
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
|
||||
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
|
||||
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
|
||||
|
||||
bool pass = ck::profiler::profile_gemm_b_scale_impl<ADataType,
|
||||
BDataType,
|
||||
BScaleDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
scale_block_k,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
(StrideA < 0) ? DefaultStrideA : StrideA,
|
||||
(StrideB < 0) ? DefaultStrideB : StrideB,
|
||||
(StrideC < 0) ? DefaultStrideC : StrideC,
|
||||
KBatch,
|
||||
n_warmup,
|
||||
n_iter,
|
||||
rotating);
|
||||
|
||||
return pass ? 0 : 1;
|
||||
};
|
||||
|
||||
if(data_type == GemmDataType::F16_I4_F16 && layout == GemmMatrixLayout::MK_NK_MN &&
|
||||
B_scale_block == BScaleBlockTile::K_128)
|
||||
{
|
||||
printf("F16_I4_F16 MK_NK_MN K_128\n");
|
||||
return profile(
|
||||
F16{}, I4{}, F16{}, F16{}, F32{}, F16{}, ck::Number<128>{}, Row{}, Col{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "this data_type & layout is not implemented" << std::endl;
|
||||
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
|
||||
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_b_scale);
|
||||
Reference in New Issue
Block a user