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

Batched GEMM Multiple D based on Universal GEMM (#1655)

Browse Source 
* Batched GEMM Multiple D based on Universal GEMM

Co-authored-by: Jing Zhang <jizhan@fb.com>

* CI fixes

Co-authored-by: Jing Zhang <jizhan@fb.com>

---------

Co-authored-by: Jing Zhang <jizhan@fb.com>
This commit is contained in:
Bartłomiej Kocot
2024-11-18 14:03:45 +01:00
committed by GitHub
parent efb34741fe
commit 754adc70e3
21 changed files with 2662 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
example/24_batched_gemm/CMakeLists.txt
View File

@@ -9,6 +9,12 @@ add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16)
add_example_executable(example_batched_gemm_xdl_bf16 batched_gemm_xdl_bf16.cpp)
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bf16)
add_example_executable(example_batched_gemm_xdl_bf16_v3 batched_gemm_xdl_bf16_v3.cpp)
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bf16_v3)
add_example_executable(example_batched_gemm_xdl_fp8_rowwise_v3 batched_gemm_xdl_fp8_rowwise_v3.cpp)
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp8_rowwise_v3)
add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int8)

99
example/24_batched_gemm/batched_gemm_xdl_bf16_v3.cpp Normal file
View File

@@ -0,0 +1,99 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_xdl_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = BF16;
using BDataType = BF16;
using AccDataType = F32;
using CShuffleDataType = BF16;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<
ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmDefault,
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXDL
32, // NPerXDL
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
S<8>, // CDEShuffleBlockTransferScalarPerVectors
ck::BlockGemmPipelineScheduler::Intrawave, // BlockGemmPipelineScheduler
ck::BlockGemmPipelineVersion::v3 // BlockGemmPipelineVersion
>;
#include "run_batched_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }

106
example/24_batched_gemm/batched_gemm_xdl_fp8_rowwise_v3.cpp Normal file
View File

@@ -0,0 +1,106 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_xdl_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F8 = ck::f8_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using MultiplyMultiply = ck::tensor_operation::element_wise::MultiplyMultiply;
using ADataType = F8;
using BDataType = F8;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = F32;
using D1DataType = F32;
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = BF16;
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Col;
using DsLayout = ck::Tuple<D0Layout, D1Layout>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MultiplyMultiply;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<
ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmDefault,
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXDL
32, // NPerXDL
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
S<8, 8, 1>, // CDEShuffleBlockTransferScalarPerVectors
ck::BlockGemmPipelineScheduler::Interwave, // BlockGemmPipelineScheduler
ck::BlockGemmPipelineVersion::v1, // BlockGemmPipelineVersion
F8 // ComputeTypeA
>;
#include "run_batched_gemm_example_rowwise.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_rowwise_example(argc, argv); }

50
example/24_batched_gemm/run_batched_gemm_example.inc
View File

@@ -210,7 +210,39 @@ bool run_batched_gemm_example(int argc, char* argv[])
problem_size.M = 256 * (dis(gen) + 1);
problem_size.N = 128 * (dis(gen) + 1);
problem_size.K = 64 * (dis(gen) + 2);
problem_size.K = 128 * (dis(gen) + 2);
problem_size.batch_count = 2;
if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else if(argc == 8)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.M = std::stoi(argv[4]);
problem_size.N = std::stoi(argv[5]);
problem_size.K = std::stoi(argv[6]);
problem_size.batch_count = std::stoi(argv[7]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("optinal\n");
printf("arg4-7: M = %d N = %d K = %d Batch = %d\n",
problem_size.M,
problem_size.N,
problem_size.K,
problem_size.batch_count);
exit(0);
}
problem_size.stride_A = problem_size.K;
problem_size.stride_B = problem_size.K;
@@ -220,21 +252,5 @@ bool run_batched_gemm_example(int argc, char* argv[])
problem_size.batch_stride_B = problem_size.K * problem_size.N;
problem_size.batch_stride_C = problem_size.M * problem_size.N;
problem_size.batch_count = 16;
if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
exit(0);
}
return run_batched_gemm(problem_size, config);
}

280
example/24_batched_gemm/run_batched_gemm_example_rowwise.inc Normal file
View File

@@ -0,0 +1,280 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <random>
#pragma once
struct ProblemSize final
{
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t stride_A = K;
ck::index_t stride_B = K;
ck::index_t stride_C = N;
ck::index_t stride_D0 = 0;
ck::index_t stride_D1 = 0;
ck::index_t batch_stride_A = M * K;
ck::index_t batch_stride_B = K * N;
ck::index_t batch_stride_C = M * N;
ck::index_t batch_stride_D0 = N;
ck::index_t batch_stride_D1 = M;
ck::index_t batch_count = 16;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
bool run_batched_gemm_rowwise(const ProblemSize& problem_size, const ExecutionConfig& config)
{
using namespace ck::literals;
auto& [M,
N,
K,
stride_A,
stride_B,
stride_C,
stride_D0,
stride_D1,
batch_stride_A,
batch_stride_B,
batch_stride_C,
batch_stride_D0,
batch_stride_D1,
batch_count] = problem_size;
// GEMM shape
auto f_host_tensor_descriptor = [](std::size_t batch_count_,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, stride, 1_uz});
}
else
{
return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, 1_uz, stride});
}
};
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
Tensor<D0DataType> d0_g_m_n(
f_host_tensor_descriptor(batch_count, M, N, stride_D0, batch_stride_D0, D0Layout{}));
Tensor<D1DataType> d1_g_m_n(
f_host_tensor_descriptor(batch_count, M, N, stride_D1, batch_stride_D1, D1Layout{}));
Tensor<EDataType> e_g_m_n_device_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "d0_g_m_n: " << d0_g_m_n.mDesc << std::endl;
std::cout << "d1_g_m_n: " << d1_g_m_n.mDesc << std::endl;
std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
switch(config.init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
d0_g_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
d1_g_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_g_m_n.mDesc.GetElementSpaceSize());
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_g_m_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
d0_device_buf.ToDevice(d0_g_m_n.mData.data());
d1_device_buf.ToDevice(d1_g_m_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
// do GEMM
auto argument =
gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
c_device_buf.GetDeviceBuffer(),
M,
N,
K,
batch_count,
stride_A,
stride_B,
{stride_D0, stride_D1},
stride_C,
batch_stride_A,
batch_stride_B,
{batch_stride_D0, batch_stride_D1},
batch_stride_C,
a_element_op,
b_element_op,
cde_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
if(config.do_verification)
{
c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
Tensor<CShuffleDataType> c_g_m_n({batch_count, M, N});
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
CShuffleDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
Tensor<EDataType> e_g_m_n_host_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, c_g_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int b = 0; b < batch_count; ++b)
{
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_g_m_n_host_result(b, m, n),
c_g_m_n(b, m, n),
d0_g_m_n(b, m, n),
d1_g_m_n(b, m, n));
}
}
}
pass = ck::utils::check_err(
e_g_m_n_device_result, e_g_m_n_host_result, "Error: Incorrect results c");
}
if(config.time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t flop = std::size_t(2) * batch_count * M * N * K;
std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
sizeof(BDataType) * batch_count * K * N +
sizeof(EDataType) * batch_count * 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 ? 0 : 1;
}
bool run_batched_gemm_rowwise_example(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
std::mt19937 gen(11939);
std::uniform_int_distribution<int> dis(0, 15);
problem_size.M = 256 * (dis(gen) + 1);
problem_size.N = 128 * (dis(gen) + 1);
problem_size.K = 128 * (dis(gen) + 2);
problem_size.batch_count = 2;
if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else if(argc == 8)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.M = std::stoi(argv[4]);
problem_size.N = std::stoi(argv[5]);
problem_size.K = std::stoi(argv[6]);
problem_size.batch_count = std::stoi(argv[7]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("optinal\n");
printf("arg4-7: M = %d N = %d K = %d Batch = %d\n",
problem_size.M,
problem_size.N,
problem_size.K,
problem_size.batch_count);
exit(0);
}
problem_size.stride_A = problem_size.K;
problem_size.stride_B = problem_size.K;
problem_size.stride_C = problem_size.N;
problem_size.stride_D0 = 0;
problem_size.stride_D1 = 0;
problem_size.batch_stride_A = problem_size.M * problem_size.K;
problem_size.batch_stride_B = problem_size.K * problem_size.N;
problem_size.batch_stride_C = problem_size.M * problem_size.N;
problem_size.batch_stride_D0 = problem_size.N;
problem_size.batch_stride_D1 = problem_size.M;
return run_batched_gemm_rowwise(problem_size, config);
}