Skip to content

GitLab

Commit 7b7a3978 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'amd-develop' into amd-master

parents b24d93a1 7e8230da
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2511 additions and 318 deletions
example/26_contraction/contraction_scale_xdl_fp32_compute_bf16.cpp 0 → 100644
View file @ 7b7a3978
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "common_instances.hpp"
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F32;
using ComputeDataType = BF16;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Scale;
using DeviceOpInstanceKKN = DeviceOpInstanceKK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceKNN = DeviceOpInstanceKN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMKN = DeviceOpInstanceMK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMNN = DeviceOpInstanceMN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstance = DeviceOpInstanceKKN;
#include "run_contraction_scale_example.inc"
int main(int argc, char* argv[]) { return run_contraction_scale_example(argc, argv); }
example/26_contraction/contraction_scale_xdl_fp32_compute_fp16.cpp 0 → 100644
View file @ 7b7a3978
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "common_instances.hpp"
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F32;
using ComputeDataType = F16;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Scale;
using DeviceOpInstanceKKN = DeviceOpInstanceKK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceKNN = DeviceOpInstanceKN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMKN = DeviceOpInstanceMK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMNN = DeviceOpInstanceMN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstance = DeviceOpInstanceKKN;
#include "run_contraction_scale_example.inc"
int main(int argc, char* argv[]) { return run_contraction_scale_example(argc, argv); }
example/26_contraction/contraction_scale_xdl_fp64_compute_fp32.cpp 0 → 100644
View file @ 7b7a3978
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "common_instances.hpp"
using ADataType = F64;
using BDataType = F64;
using AccDataType = F32;
using CShuffleDataType = F64;
using DsDataType = ck::Tuple<>;
using EDataType = F64;
using ComputeDataType = F32;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Scale;
using DeviceOpInstanceKKN = DeviceOpInstanceKK_FP64<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceKNN = DeviceOpInstanceKN_FP64<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMKN = DeviceOpInstanceMK_FP64<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMNN = DeviceOpInstanceMN_FP64<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstance = DeviceOpInstanceKKN;
#include "run_contraction_scale_example.inc"
int main(int argc, char* argv[]) { return run_contraction_scale_example(argc, argv); }
example/26_contraction/run_contraction_bilinear_example.inc 0 → 100644
View file @ 7b7a3978
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <iostream>
#include <string>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/numeric.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
int run_contraction_bilinear_example(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// A[M0, M1, K0, K1]
std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
// D[M0, M1, N0, N1]
std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
float alpha = 1.f;
float beta = 1.f;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 28)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t M0 = std::stoi(argv[4]);
const ck::index_t M1 = std::stoi(argv[5]);
const ck::index_t N0 = std::stoi(argv[6]);
const ck::index_t N1 = std::stoi(argv[7]);
const ck::index_t K0 = std::stoi(argv[8]);
const ck::index_t K1 = std::stoi(argv[9]);
a_ms_ks_lengths = {M0, M1, K0, K1};
a_ms_ks_strides = {
std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
b_ns_ks_lengths = {N0, N1, K0, K1};
b_ns_ks_strides = {
std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
d_ms_ns_lengths = {M0, M1, N0, N1};
d_ms_ns_strides = {
std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
e_ms_ns_lengths = {M0, M1, N0, N1};
e_ms_ns_strides = {
std::stoi(argv[22]), std::stoi(argv[23]), std::stoi(argv[24]), std::stoi(argv[25])};
alpha = std::stof(argv[26]);
beta = std::stof(argv[27]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M0, M1, N0, N1, K0, K1\n");
printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
printf("arg18 to 21: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
printf("arg22 to 25: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
printf("arg26 to 27: alpha, beta\n");
exit(0);
}
Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_ms_ks.mData.data());
b_device_buf.ToDevice(b_ns_ks.mData.data());
d_device_buf.ToDevice(d_ms_ns.mData.data());
// set zero
e_device_buf.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
// device operation
auto op = DeviceOpInstance{};
auto invoker = op.MakeInvoker();
auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument))
{
std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
ck::index_t M =
ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
ck::index_t N = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(DDataType) * M * N + sizeof(EDataType) * 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, "
<< op.GetTypeString() << std::endl;
e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
using ReferenceOpInstance =
ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
CShuffleDataType,
AccDataType,
ComputeDataType,
AElementOp,
BElementOp>;
auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_op.MakeInvoker();
auto ref_argument =
ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
ref_invoker.Run(ref_argument);
for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
c_ms_ns_host_result(m0, m1, n0, n1),
d_ms_ns(m0, m1, n0, n1));
}
}
}
}
return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
}
return 0;
}
example/26_contraction/run_contraction_scale_example.inc 0 → 100644
View file @ 7b7a3978
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <iostream>
#include <string>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/numeric.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
int run_contraction_scale_example(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// A[M0, M1, K0, K1]
std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
float scale = 1.f;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 23)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t M0 = std::stoi(argv[4]);
const ck::index_t M1 = std::stoi(argv[5]);
const ck::index_t N0 = std::stoi(argv[6]);
const ck::index_t N1 = std::stoi(argv[7]);
const ck::index_t K0 = std::stoi(argv[8]);
const ck::index_t K1 = std::stoi(argv[9]);
a_ms_ks_lengths = {M0, M1, K0, K1};
a_ms_ks_strides = {
std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
b_ns_ks_lengths = {N0, N1, K0, K1};
b_ns_ks_strides = {
std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
e_ms_ns_lengths = {M0, M1, N0, N1};
e_ms_ns_strides = {
std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
scale = std::stof(argv[22]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M0, M1, N0, N1, K0, K1\n");
printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
printf("arg18 to 21: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
printf("arg22: scale\n");
exit(0);
}
Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_ms_ks.mData.data());
b_device_buf.ToDevice(b_ns_ks.mData.data());
// set zero
e_device_buf.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{scale};
// device operation
auto op = DeviceOpInstance{};
auto invoker = op.MakeInvoker();
auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 0>{},
e_device_buf.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 0>{},
std::array<std::vector<ck::index_t>, 0>{},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument))
{
std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
ck::index_t M =
ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
ck::index_t N = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * 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, "
<< op.GetTypeString() << std::endl;
e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
using ReferenceOpInstance =
ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
CShuffleDataType,
AccDataType,
ComputeDataType,
AElementOp,
BElementOp>;
auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_op.MakeInvoker();
auto ref_argument =
ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
ref_invoker.Run(ref_argument);
for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
c_ms_ns_host_result(m0, m1, n0, n1));
}
}
}
}
return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
}
return 0;
}
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
View file @ 7b7a3978
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_data) add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp) add_example_executable(example_grouped_conv_bwd_data_xdl_fp16 grouped_conv_bwd_data_xdl_fp16.cpp)
if(result EQUAL 0) if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16) add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_xdl_fp16)
endif() endif()
add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp) add_example_executable(example_grouped_conv_bwd_data_bias_relu_xdl_fp16 grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp)
if(result EQUAL 0) if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16) add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_xdl_fp16)
endif()
set(target 1)
endif()
endforeach()
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_wmma_fp16 grouped_conv_bwd_data_wmma_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_wmma_fp16)
endif() endif()
set(target 1) set(target 1)
endif() endif()
......
example/38_grouped_conv_bwd_data_multiple_d/common.hpp
View file @ 7b7a3978
...@@ -10,7 +10,6 @@ ...@@ -10,7 +10,6 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp" #include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......
example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp → example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp
View file @ 7b7a3978
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
#include "common.hpp" #include "common.hpp"
using OutDataType = FP16; using OutDataType = FP16;
......
example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp 0 → 100644
View file @ 7b7a3978
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp"
#include "common.hpp"
using OutDataType = FP16;
using WeiDataType = FP16;
using AccDataType = FP32;
using CShuffleDataType = FP16;
using DsDataType = ck::Tuple<>;
using InDataType = FP16;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using DsLayout = ck::Tuple<>;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using OutElementOp = PassThrough;
using WeiElementOp = PassThrough;
using InElementOp = PassThrough;
// clang-format off
using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
//| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | WMMA| WMMA| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< 2,OutLayout, WeiLayout, DsLayout, InLayout, OutDataType, WeiDataType, AccDataType, CShuffleDataType, DsDataType, InDataType, OutElementOp, WeiElementOp, InElementOp, ConvBwdDataDefault, 128, 64, 64, 4, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>;
// clang-format on
#include "run_grouped_conv_bwd_data_example.inc"
int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_example(argc, argv); }
example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp → example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp
View file @ 7b7a3978
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
#include "common.hpp" #include "common.hpp"
using OutDataType = FP16; using OutDataType = FP16;
......
example/60_gemm_multi_ABD/CMakeLists.txt 0 → 100644
View file @ 7b7a3978
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
set(target 1)
endif()
endforeach()
endif()
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp 0 → 100644
View file @ 7b7a3978
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
#include "ck/library/utility/check_err.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_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 = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DLayout = Row;
using ELayout = Row;
struct AddScale
{
static constexpr auto I0 = ck::Number<0>{};
static constexpr auto I1 = ck::Number<1>{};
static constexpr auto I2 = ck::Number<2>{};
static constexpr auto I3 = ck::Number<3>{};
__host__ __device__ constexpr void
operator()(ck::half4_t& a, const ck::half4_t& a0, const ck::half4_t& a1) const
{
const auto a0_v_t = ck::vector_type<ck::half_t, 4>{a0};
const auto a1_v_t = ck::vector_type<ck::half_t, 4>{a1};
auto r_v_t = ck::vector_type<ck::half_t, 4>{};
r_v_t.AsType<ck::half_t>()(I0) =
scale * (a0_v_t.AsType<ck::half_t>()[I0] + a1_v_t.AsType<ck::half_t>()[I0]);
r_v_t.AsType<ck::half_t>()(I1) =
scale * (a0_v_t.AsType<ck::half_t>()[I1] + a1_v_t.AsType<ck::half_t>()[I1]);
r_v_t.AsType<ck::half_t>()(I2) =
scale * (a0_v_t.AsType<ck::half_t>()[I2] + a1_v_t.AsType<ck::half_t>()[I2]);
r_v_t.AsType<ck::half_t>()(I3) =
scale * (a0_v_t.AsType<ck::half_t>()[I3] + a1_v_t.AsType<ck::half_t>()[I3]);
a = r_v_t.AsType<ck::half4_t>()[I0];
}
__host__ __device__ constexpr void
operator()(ck::half_t& a, const ck::half_t& a0, const ck::half_t& a1) const
{
a = scale * (a0 + a1);
}
// this attribute will force copy_function applying element_wise with vector_type
static constexpr ck::index_t vec_len = 4;
float scale = 1.0;
};
struct AlphaBetaAdd
{
AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
ck::half_t& e, const float& c, const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
};
float alpha_;
float beta_;
};
using AElementOp = AddScale;
using BElementOp = PassThrough;
using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle<
ck::Tuple<ALayout, ALayout>,
ck::Tuple<BLayout>,
ck::Tuple<DLayout>,
ELayout,
ck::Tuple<ADataType, ADataType>,
ck::Tuple<BDataType>,
AccDataType,
CShuffleDataType,
ck::Tuple<DDataType>,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
1,
256,
256,
128,
32,
8,
8,
32,
32,
4,
2,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
1,
1,
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideD = 4096;
ck::index_t StrideE = 4096;
float alpha = 1.0f;
float beta = 1.0f;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
alpha = std::stof(argv[4]);
beta = std::stof(argv[5]);
}
else if(argc == 13)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
alpha = std::stof(argv[11]);
beta = std::stof(argv[12]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
"beta\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<ADataType> a1_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
a1_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
}
DeviceMem a0_device_buf(sizeof(ADataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(ADataType) * a1_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
a1_device_buf.ToDevice(a1_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{0.2};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
a1_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{b_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
std::array<ck::index_t, 2>{StrideA, StrideA},
std::array<ck::index_t, 1>{StrideB},
std::array<ck::index_t, 1>{StrideD},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * 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"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n({M, N});
Tensor<ADataType> a_m_k({M, K});
for(int m = 0; m < M; ++m)
{
for(int k = 0; k < K; ++k)
{
a_element_op(a_m_k(m, k), a0_m_k(m, k), a1_m_k(m, k));
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CShuffleDataType,
AccDataType,
PassThrough,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
example/61_contraction_multi_ABD/CMakeLists.txt 0 → 100644
View file @ 7b7a3978
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_contraction_multi_ABD_xdl_fp16 contraction_multi_ABD_xdl_fp16.cpp)
set(target 1)
endif()
endforeach()
endif()
example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp 0 → 100644
View file @ 7b7a3978
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, 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/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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_contraction.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/numeric.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using A0DataType = F16;
using A1DataType = F32;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using EDataType = F16;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
struct AlphaBetaAdd
{
AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
ck::half_t& e, const float& c, const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
};
float alpha_;
float beta_;
};
struct Multiply
{
__host__ __device__ constexpr void
operator()(ck::half_t& a, const ck::half_t& a0, const float& a1) const
{
a = ck::type_convert<ck::half_t>(ck::type_convert<float>(a0) * a1);
}
};
using AElementOp = Multiply;
using BElementOp = PassThrough;
using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceOpInstance = ck::tensor_operation::device::DeviceContractionMultipleABD_Xdl_CShuffle<
NumDimM,
NumDimN,
NumDimK,
ck::Tuple<A0DataType, A1DataType>,
ck::Tuple<BDataType>,
AccDataType,
CShuffleDataType,
ck::Tuple<DDataType>,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
1,
256,
256,
128,
32,
8,
8,
32,
32,
4,
2,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
1,
1,
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
float alpha = 1.0f;
float beta = 1.0f;
// A0[M0, M1, K0, K1]
std::vector<ck::index_t> a0_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a0_ms_ks_strides{128 * 32 * 64, 32 * 64, 64, 1};
// A1[M1, K1] -> A1[M0, M1, K0, K1]
std::vector<ck::index_t> a1_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 0, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
// D[M0, M1, N0, N1]
std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> d_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
exit(0);
}
Tensor<A0DataType> a0_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
Tensor<A1DataType> a1_ms_ks(a1_ms_ks_lengths, a1_ms_ks_strides);
Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
std::cout << "a0_ms_ks: " << a0_ms_ks.mDesc << std::endl;
std::cout << "a1_ms_ks: " << a1_ms_ks.mDesc << std::endl;
std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_ms_ks.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
a1_ms_ks.GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a0_ms_ks.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
a1_ms_ks.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(A1DataType) * a1_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_ms_ks.mData.data());
a1_device_buf.ToDevice(a1_ms_ks.mData.data());
b_device_buf.ToDevice(b_ns_ks.mData.data());
d_device_buf.ToDevice(d_ms_ns.mData.data());
// set zero
e_device_buf.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(
std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
a1_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{b_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_lengths, a1_ms_ks_lengths},
std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_strides, a1_ms_ks_strides},
std::array<std::vector<ck::index_t>, 1>{b_ns_ks_lengths},
std::array<std::vector<ck::index_t>, 1>{b_ns_ks_strides},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_contraction with the specified compilation parameters does "
"not support this problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
if(time_kernel)
{
ck::index_t M =
ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
ck::index_t N = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a0_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * 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" << std::endl;
}
if(do_verification)
{
Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<A0DataType> a_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
for(size_t m0 = 0; m0 < a_ms_ks.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < a_ms_ks.mDesc.GetLengths()[1]; ++m1)
{
for(size_t k0 = 0; k0 < a_ms_ks.mDesc.GetLengths()[2]; ++k0)
{
for(size_t k1 = 0; k1 < a_ms_ks.mDesc.GetLengths()[3]; ++k1)
{
a_element_op(a_ms_ks(m0, m1, k0, k1),
a0_ms_ks(m0, m1, k0, k1),
a1_ms_ks(m0, m1, k0, k1));
}
}
}
}
using ReferenceOpInstance =
ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimN,
NumDimK,
A0DataType,
BDataType,
CShuffleDataType,
AccDataType,
PassThrough,
BElementOp>;
auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_op.MakeInvoker();
Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
auto ref_argument =
ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, PassThrough{}, b_element_op);
ref_invoker.Run(ref_argument);
for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
c_ms_ns_host_result(m0, m1, n0, n1),
d_ms_ns(m0, m1, n0, n1));
}
}
}
}
e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
}
return 0;
}
include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
View file @ 7b7a3978
...@@ -28,7 +28,8 @@ MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(const TileDesc_K0_MN_K1&) ...@@ -28,7 +28,8 @@ MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(const TileDesc_K0_MN_K1&)
} }
template <index_t BlockSize, template <index_t BlockSize,
typename FloatAB, typename FloatA,
typename FloatB,
typename FloatAcc, typename FloatAcc,
typename AK0MK1BlockDesc, typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc, typename BK0NK1BlockDesc,
...@@ -58,7 +59,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -58,7 +59,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2); static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2); static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack>{}; static constexpr auto xdlops_gemm = XdlopsGemm<FloatA, MPerXDL, NPerXDL, KPack, FloatB>{};
static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops; static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
...@@ -294,9 +295,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -294,9 +295,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
const BBlockBuffer& b_block_buf, const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const CThreadBuffer& c_thread_buf) const
{ {
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>( auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
a_thread_desc_.GetElementSpaceSize()); a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>( auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
b_thread_desc_.GetElementSpaceSize()); b_thread_desc_.GetElementSpaceSize());
static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, MRepeat, 1>{}([&](auto m0) {
...@@ -318,25 +319,27 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -318,25 +319,27 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
b_thread_buf); b_thread_buf);
static_for<0, KPerThread, KPack>{}([&](auto k) { static_for<0, KPerThread, KPack>{}([&](auto k) {
vector_type<FloatAB, KPack> a_thread_vec; vector_type<FloatA, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec; vector_type<FloatB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto i) { static_for<0, KPack, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf a_thread_vec.template AsType<FloatA>()(i) = a_thread_buf
[Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}]; [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf b_thread_vec.template AsType<FloatB>()(i) = b_thread_buf
[Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}]; [Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
}); });
using mfma_input_type = using mfma_input_type_a =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type; typename vector_type<FloatA, xdlops_gemm.K1PerXdlops>::type;
using mfma_input_type_b =
typename vector_type<FloatB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset = constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.template Run( xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(), a_thread_vec.template AsType<mfma_input_type_a>(),
b_thread_vec.template AsType<mfma_input_type>(), b_thread_vec.template AsType<mfma_input_type_b>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{})); c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
}); });
}); });
...@@ -356,8 +359,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -356,8 +359,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed( static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops())); make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
FloatAB, FloatA,
decltype(a_block_desc_m0_m1_m2_k), decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_), decltype(a_thread_desc_),
Sequence<1, 1, 1, KPerThread>, Sequence<1, 1, 1, KPerThread>,
...@@ -366,8 +369,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -366,8 +369,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
A_K1, A_K1,
A_K1>; A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
FloatAB, FloatB,
decltype(b_block_desc_n0_n1_n2_k), decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_), decltype(b_thread_desc_),
Sequence<1, 1, 1, KPerThread>, Sequence<1, 1, 1, KPerThread>,
...@@ -385,7 +388,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -385,7 +388,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
// the latest ROCm release. For unsupported compilers, inter-wave loop scheduler falls back to the // the latest ROCm release. For unsupported compilers, inter-wave loop scheduler falls back to the
// default loop scheduler which is given by the macro CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=0 // default loop scheduler which is given by the macro CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=0
template <index_t BlockSize, template <index_t BlockSize,
typename FloatAB, typename FloatA,
typename FloatB,
typename FloatAcc, typename FloatAcc,
typename AK0MK1BlockDesc, typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc, typename BK0NK1BlockDesc,
...@@ -397,7 +401,8 @@ template <index_t BlockSize, ...@@ -397,7 +401,8 @@ template <index_t BlockSize,
index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS> index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS>
struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
: public BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize, : public BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB, FloatA,
FloatB,
FloatAcc, FloatAcc,
AK0MK1BlockDesc, AK0MK1BlockDesc,
BK0NK1BlockDesc, BK0NK1BlockDesc,
...@@ -408,7 +413,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -408,7 +413,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
KPack> KPack>
{ {
using Base = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize, using Base = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB, FloatA,
FloatB,
FloatAcc, FloatAcc,
AK0MK1BlockDesc, AK0MK1BlockDesc,
BK0NK1BlockDesc, BK0NK1BlockDesc,
...@@ -440,9 +446,9 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -440,9 +446,9 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
const BBlockBuffer& b_block_buf, const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const CThreadBuffer& c_thread_buf) const
{ {
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>( auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
a_thread_desc_.GetElementSpaceSize()); a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>( auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
b_thread_desc_.GetElementSpaceSize()); b_thread_desc_.GetElementSpaceSize());
static_for<0, KPerThread, KPerInnerLoop>{}([&](auto k) { static_for<0, KPerThread, KPerInnerLoop>{}([&](auto k) {
...@@ -479,20 +485,22 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -479,20 +485,22 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) { static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) { static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<FloatAB, KPack> a_thread_vec; vector_type<FloatA, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec; vector_type<FloatB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto i) { static_for<0, KPack, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(i) = a_thread_vec.template AsType<FloatA>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset( a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, 0, 0, k_ + i))>{}]; make_tuple(m0, 0, 0, k_ + i))>{}];
b_thread_vec.template AsType<FloatAB>()(i) = b_thread_vec.template AsType<FloatB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset( b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, 0, 0, k_ + i))>{}]; make_tuple(n0, 0, 0, k_ + i))>{}];
}); });
using mfma_input_type = using mfma_input_type_a =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type; typename vector_type<FloatA, xdlops_gemm.K1PerXdlops>::type;
using mfma_input_type_b =
typename vector_type<FloatB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset = constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0)); c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
...@@ -514,8 +522,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -514,8 +522,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
// TODO: insert setprio in more precise manner since we // TODO: insert setprio in more precise manner since we
// could have more than >1 MFMA instructions in single call // could have more than >1 MFMA instructions in single call
xdlops_gemm.template Run( xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(), a_thread_vec.template AsType<mfma_input_type_a>(),
b_thread_vec.template AsType<mfma_input_type>(), b_thread_vec.template AsType<mfma_input_type_b>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{})); c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0) if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{ {
...@@ -541,8 +549,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -541,8 +549,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed( static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<NRepeat>{}, I1, I1, Number<KPerInnerLoop>{})); make_tuple(Number<NRepeat>{}, I1, I1, Number<KPerInnerLoop>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
FloatAB, FloatA,
decltype(a_block_desc_m0_m1_m2_k), decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_), decltype(a_thread_desc_),
Sequence<1, 1, 1, KPerInnerLoop>, Sequence<1, 1, 1, KPerInnerLoop>,
...@@ -551,8 +559,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -551,8 +559,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
A_K1, A_K1,
A_K1>; A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
FloatAB, FloatB,
decltype(b_block_desc_n0_n1_n2_k), decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_), decltype(b_thread_desc_),
Sequence<1, 1, 1, KPerInnerLoop>, Sequence<1, 1, 1, KPerInnerLoop>,
...@@ -568,7 +576,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -568,7 +576,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
}; };
template <index_t BlockSize, template <index_t BlockSize,
typename FloatAB, typename FloatA,
typename FloatB,
typename FloatAcc, typename FloatAcc,
typename AK0MK1BlockDesc, typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc, typename BK0NK1BlockDesc,
...@@ -583,7 +592,8 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector() ...@@ -583,7 +592,8 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
if constexpr(LoopSched == LoopScheduler::Default) if constexpr(LoopSched == LoopScheduler::Default)
{ {
return BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize, return BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB, FloatA,
FloatB,
FloatAcc, FloatAcc,
AK0MK1BlockDesc, AK0MK1BlockDesc,
BK0NK1BlockDesc, BK0NK1BlockDesc,
...@@ -596,7 +606,8 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector() ...@@ -596,7 +606,8 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
else if constexpr(LoopSched == LoopScheduler::Interwave) else if constexpr(LoopSched == LoopScheduler::Interwave)
{ {
return BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize, return BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB, FloatA,
FloatB,
FloatAcc, FloatAcc,
AK0MK1BlockDesc, AK0MK1BlockDesc,
BK0NK1BlockDesc, BK0NK1BlockDesc,
...@@ -618,26 +629,27 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector() ...@@ -618,26 +629,27 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
* 3. configurable k index starting position and step size after each FMA/XDL instruction * 3. configurable k index starting position and step size after each FMA/XDL instruction
*/ */
template <index_t BlockSize, template <
typename FloatAB, index_t BlockSize,
typename FloatAcc, typename FloatAB,
typename ATileDesc, typename FloatAcc,
typename BTileDesc, typename ATileDesc,
typename AMmaTileDesc, typename BTileDesc,
typename BMmaTileDesc, typename AMmaTileDesc,
index_t MPerBlock, typename BMmaTileDesc,
index_t NPerBlock, index_t MPerBlock,
index_t KPerBlock, index_t NPerBlock,
index_t MPerXDL, index_t KPerBlock,
index_t NPerXDL, index_t MPerXDL,
index_t MRepeat, index_t NPerXDL,
index_t NRepeat, index_t MRepeat,
index_t KPack, index_t NRepeat,
bool TransposeC = false, index_t KPack,
index_t AMmaKStride = bool TransposeC = false,
KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{}.K0PerXdlops, index_t AMmaKStride =
index_t BMmaKStride = KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops,
KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{}.K0PerXdlops> index_t BMmaKStride =
KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops>
struct BlockwiseGemmXdlops_v2 struct BlockwiseGemmXdlops_v2
{ {
static constexpr auto I0 = Number<0>{}; static constexpr auto I0 = Number<0>{};
...@@ -654,7 +666,8 @@ struct BlockwiseGemmXdlops_v2 ...@@ -654,7 +666,8 @@ struct BlockwiseGemmXdlops_v2
static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2); static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2); static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{}; static constexpr auto xdlops_gemm =
XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{};
static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops; static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
......
include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp 0 → 100644
View file @ 7b7a3978
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <array>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// GEMM:
// input : A0[M0, M1, ... K0, K1, ...], ...
// input : B0[N0, N1, ... K0, K1, ...], ...
// input : D0[M0, M1, ... N0, N1, ...], D1[M0, M1, ... N0, N1, ...], ...
// output : E[M0, M1, ... N0, N1, ...]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Assume:
// D0, D1, ... and E have the same layout
template <index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
typename AsDataType,
typename BsDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceContractionMultipleABD : public BaseOperator
{
static constexpr index_t NumATensor = AsDataType::Size();
static constexpr index_t NumBTensor = BsDataType::Size();
static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::array<const void*, NumATensor> p_as,
std::array<const void*, NumBTensor> p_bs,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_lengths,
const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_strides,
const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_lengths,
const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_strides,
const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_lengths,
const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_strides,
const std::vector<index_t>& e_ms_ns_length,
const std::vector<index_t>& e_ms_ns_stride,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
View file @ 7b7a3978
...@@ -33,7 +33,8 @@ template <index_t NumDimM, ...@@ -33,7 +33,8 @@ template <index_t NumDimM,
typename EDataType, typename EDataType,
typename AElementwiseOperation, typename AElementwiseOperation,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CDEElementwiseOperation> typename CDEElementwiseOperation,
typename ComputeDataType = ADataType>
struct DeviceContractionMultipleD : public BaseOperator struct DeviceContractionMultipleD : public BaseOperator
{ {
static constexpr index_t NumDTensor = DsDataType::Size(); static constexpr index_t NumDTensor = DsDataType::Size();
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 7b7a3978
...@@ -112,6 +112,7 @@ template <index_t NumDimM, ...@@ -112,6 +112,7 @@ template <index_t NumDimM,
typename CShuffleDataType, typename CShuffleDataType,
typename DsDataType, typename DsDataType,
typename EDataType, typename EDataType,
typename ComputeDataType,
typename AElementwiseOperation, typename AElementwiseOperation,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CDEElementwiseOperation, typename CDEElementwiseOperation,
...@@ -156,7 +157,8 @@ struct DeviceContractionMultipleD_Xdl_CShuffle ...@@ -156,7 +157,8 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
EDataType, EDataType,
AElementwiseOperation, AElementwiseOperation,
BElementwiseOperation, BElementwiseOperation,
CDEElementwiseOperation> CDEElementwiseOperation,
ComputeDataType>
{ {
using DeviceOp = DeviceContractionMultipleD_Xdl_CShuffle; using DeviceOp = DeviceContractionMultipleD_Xdl_CShuffle;
...@@ -310,8 +312,6 @@ struct DeviceContractionMultipleD_Xdl_CShuffle ...@@ -310,8 +312,6 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>; using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N({}, {})); using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N({}, {}));
using ComputeDataType = ADataType;
// GridwiseGemm // GridwiseGemm
using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle< using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
ADataType, // TODO: distinguish A/B datatype ADataType, // TODO: distinguish A/B datatype
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment