mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-07-14 19:18:35 +00:00
Modify the kernel to 128x128x64, and use mfma_32x32x4
Add int4+scale based on Zhang, Jing pk_i4. Compile pass, function pass.
This commit is contained in:
mtgu0705
@@ -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_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_v3)
|
||||
add_example_executable(example_gemm_xdl_bf16_v3 gemm_xdl_bf16_v3.cpp)
|
||||
add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_v3)
|
||||
|
||||
@@ -23,7 +23,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
|
||||
|
||||
static constexpr bool PermuteB = true;
|
||||
|
||||
static constexpr ck::index_t KPerBlock = 128;
|
||||
static constexpr ck::index_t KPerBlock = 64;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmV2Instance =
|
||||
@@ -43,28 +43,28 @@ using DeviceGemmV2Instance =
|
||||
2, 32, 32, 0,
|
||||
1, 1, S<1, 16, 1, 8>, 4,
|
||||
#else
|
||||
128,
|
||||
16, 64,
|
||||
256,
|
||||
128, 128,
|
||||
KPerBlock, 8, 32,
|
||||
16, 16,
|
||||
1, 2,
|
||||
S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>,
|
||||
32, 32,
|
||||
2, 2,
|
||||
S<8, 32, 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>,
|
||||
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>,
|
||||
2, 32, 32, 0,
|
||||
1, 1, S<1, 16, 1, 8>, 4,
|
||||
#endif
|
||||
ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v2, CDataType, CDataType, false, PermuteB>;
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, CDataType, CDataType, false, PermuteB>;
|
||||
|
||||
// clang-format on
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
template <typename ProblemType>
|
||||
bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
|
||||
{
|
||||
|
||||
407
example/01_gemm/gemm_xdl_fp16_pk_i4_v3_b_scale.cpp
Normal file
407
example/01_gemm/gemm_xdl_fp16_pk_i4_v3_b_scale.cpp
Normal file
@@ -0,0 +1,407 @@
|
||||
// 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 PermuteB = true;
|
||||
|
||||
static constexpr ck::index_t Scale_Block_N = 1;
|
||||
static constexpr ck::index_t Scale_Block_K = 64;
|
||||
|
||||
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,
|
||||
#if 0
|
||||
128, Scale_Block_N, Scale_Block_K,
|
||||
16, 128,
|
||||
KPerBlock, 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,
|
||||
#else
|
||||
256, Scale_Block_N, Scale_Block_K,
|
||||
128, 128,
|
||||
KPerBlock, 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, 16, 1, 8>, 4,
|
||||
#endif
|
||||
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, CDataType, CDataType, false, 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, std::size_t stride, auto layout) {
|
||||
if(stride == 0)
|
||||
{
|
||||
// give a chance if stride is zero, return a default packed stride
|
||||
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
return col;
|
||||
}
|
||||
else
|
||||
{
|
||||
return row;
|
||||
}
|
||||
}
|
||||
else
|
||||
return 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_2<BScaleDataType>{0, 1});
|
||||
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;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.5, 0.5});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
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);
|
||||
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,
|
||||
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_pre({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);
|
||||
int8_t i4 = 0;
|
||||
if(k % 2 == 1)
|
||||
i4 = (i4x2 >> 0) & 0xf;
|
||||
else
|
||||
i4 = (i4x2 >> 4) & 0xf;
|
||||
i4 = i4 - 8;
|
||||
v_b = ck::type_convert<float>(i4);
|
||||
|
||||
b_k_n_pre(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_pre, 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 0
|
||||
std::cout << "a_m_k: " << std::endl;
|
||||
for(int i = 0; i < M; i++)
|
||||
{
|
||||
for(int j = 0; j < K; j++)
|
||||
{
|
||||
std::cout << ck::type_convert<float>(a_m_k(i, j)) << ",";
|
||||
}
|
||||
std::cout << std::endl;
|
||||
}
|
||||
|
||||
std::cout << "b_k_n: " << std::endl;
|
||||
for(int i = 0; i < N; i++)
|
||||
{
|
||||
for(int j = 0; j < K; j++)
|
||||
{
|
||||
ck::pk_i4_t i4x2 = b_k_n(j, i);
|
||||
int8_t i4 = 0;
|
||||
if( j % 2 == 1)
|
||||
i4 = (i4x2 >> 0) & 0xf;
|
||||
else
|
||||
i4 = (i4x2 >> 4) & 0xf;
|
||||
i4 = i4 - 8;
|
||||
std::cout << ck::type_convert<float>(i4) << ",";
|
||||
}
|
||||
std::cout << std::endl;
|
||||
}
|
||||
|
||||
std::cout << "c_m_n_device_result: " << std::endl;
|
||||
for(int i = 0; i < M; i++)
|
||||
{
|
||||
for(int j = 0; j < N; j++)
|
||||
{
|
||||
std::cout << ck::type_convert<float>(c_m_n_device_result(i, j)) << ",";
|
||||
}
|
||||
std::cout << std::endl;
|
||||
}
|
||||
|
||||
std::cout << "c_m_n_host_result: " << std::endl;
|
||||
for(int i = 0; i < M; i++)
|
||||
{
|
||||
for(int j = 0; j < N; j++)
|
||||
{
|
||||
std::cout << ck::type_convert<float>(c_m_n_host_result(i, j)) << ",";
|
||||
}
|
||||
std::cout << std::endl;
|
||||
}
|
||||
#endif
|
||||
}
|
||||
|
||||
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); }
|
||||
@@ -4,8 +4,8 @@
|
||||
#pragma once
|
||||
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.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.hpp"
|
||||
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp"
|
||||
|
||||
@@ -68,49 +68,49 @@ constexpr auto BlockGemmPipeline_Selector()
|
||||
}
|
||||
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
|
||||
{
|
||||
return BlockwiseGemmXdlops_pipeline_v2<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
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<BlkGemmPipeSche,
|
||||
BlockSize,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
ATileDesc,
|
||||
BTileDesc,
|
||||
AMmaTileDesc,
|
||||
BMmaTileDesc,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
KPerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
MRepeat,
|
||||
NRepeat,
|
||||
KPack>{};
|
||||
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)
|
||||
{
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,522 @@
|
||||
// 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);
|
||||
|
||||
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());
|
||||
|
||||
//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>{}));
|
||||
});
|
||||
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);
|
||||
|
||||
// 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();
|
||||
auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>(); // need actually?
|
||||
|
||||
// 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_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, 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;
|
||||
|
||||
// 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_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>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[n0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
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_thread_desc_,
|
||||
make_tuple(n0, I0, k0, I0),
|
||||
b_thread_buf);
|
||||
});
|
||||
});
|
||||
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>{}));
|
||||
|
||||
HotLoopScheduler();
|
||||
__builtin_amdgcn_sched_barrier(0);
|
||||
|
||||
i += 1;
|
||||
} while(i < (num_loop - 1));
|
||||
}
|
||||
// tail
|
||||
if constexpr(TailNum == TailNumber::Full)
|
||||
{
|
||||
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>(a_scale_thread_buf[I0]) *
|
||||
type_convert<AccDataType>(b_scale_thread_buf[n0]);
|
||||
});
|
||||
});
|
||||
});
|
||||
__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
|
||||
@@ -76,6 +76,42 @@ 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,
|
||||
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 +1,758 @@
|
||||
// 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/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"
|
||||
#include "ck/host_utility/flush_cache.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 ? 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,
|
||||
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, 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,
|
||||
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,
|
||||
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
|
||||
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user