Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "8826bae6555c4cdd3740e3f30397769a885c7bce"
Commit 8b2796c7 authored by rocking's avatar rocking
Browse files

Merge commit 'd8b41e1c' into gemm_layernorm_welford

parents c3107fd5 d8b41e1c
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 3454 additions and 101 deletions
Jenkinsfile
View file @ 8b2796c7
...@@ -233,6 +233,7 @@ def buildHipClangJob(Map conf=[:]){ ...@@ -233,6 +233,7 @@ def buildHipClangJob(Map conf=[:]){
def variant = env.STAGE_NAME def variant = env.STAGE_NAME
def retimage def retimage
(retimage, image) = getDockerImage(conf)
gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') { gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') { withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
...@@ -548,8 +549,9 @@ def process_results(Map conf=[:]){ ...@@ -548,8 +549,9 @@ def process_results(Map conf=[:]){
} }
} }
//launch develop branch daily at 23:00 in FULL_QA mode //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true''' : "" CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;COMPILER_VERSION=release
0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open''' : ""
pipeline { pipeline {
agent none agent none
......
example/42_groupnorm/groupnorm_sigmoid_fp16.cpp
View file @ 8b2796c7
...@@ -55,26 +55,26 @@ using DeviceInstance = ...@@ -55,26 +55,26 @@ using DeviceInstance =
YElementOp, YElementOp,
Rank, Rank,
NumReduceDim, NumReduceDim,
256, // BlockSize 1024, // BlockSize
8, // ClusterM 1, // ClusterM
32, // ClusterK 1024, // ClusterK
1, // SliceM 1, // SliceM
8, // SliceK 32, // SliceK
1, // SrcVecDim (0=M, 1=K) 1, // SrcVecDim (0=M, 1=K)
8, // SrcScalarPerVector 2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K) 1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector 2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K) 1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector 2, // BetaScalarPerVector
8>; // OutScalarPerVector 2>; // OutScalarPerVector
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
ck::index_t N = 128; ck::index_t N = 2;
ck::index_t H = 16; ck::index_t H = 32;
ck::index_t W = 16; ck::index_t W = 32;
ck::index_t G = 32; ck::index_t G = 32;
ck::index_t C = 40; ck::index_t C = 30;
if(argc == 1) if(argc == 1)
{ {
......
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt 0 → 100644
View file @ 8b2796c7
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp 0 → 100644
View file @ 8b2796c7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, 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/device_splitk_contraction_multiple_d_xdl_cshuffle.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"
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 Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using DDataType = F16;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F16;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 1;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Add;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
// clang-format off
using DeviceOpInstanceKKNN = ck::tensor_operation::device::
//############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| Gemm| A| B| DE| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle|CBlockTransferClusterLengths| CBlockTransfer|
//############################################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceSplitKContractionMultipleD_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, F16, F16, F32, F16, DsDataType, F16, AElementOp, BElementOp, CDEElementOp, GemmSpec, ABSpec, ABSpec, DESpec, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>;
// clang-format on
using DeviceOpInstance = DeviceOpInstanceKKNN;
// hardcoded for NumDimM == NumDimN == NumDimK == 2
template <ck::index_t NumDimG,
ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename EDataType,
typename AccDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
false>
struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
{
// Argument
struct Argument : public ck::tensor_operation::device::BaseArgument
{
Argument(const Tensor<ADataType>& a_gs_ms_ks,
const Tensor<BDataType>& b_gs_ns_ks,
Tensor<EDataType>& e_gs_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: a_gs_ms_ks_{a_gs_ms_ks},
b_gs_ns_ks_{b_gs_ns_ks},
e_gs_ms_ns_{e_gs_ms_ns},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
}
const Tensor<ADataType>& a_gs_ms_ks_;
const Tensor<BDataType>& b_gs_ns_ks_;
Tensor<EDataType>& e_gs_ms_ns_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
};
// Invoker
struct Invoker : public ck::tensor_operation::device::BaseInvoker
{
using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
float Run(const Argument& arg)
{
auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
AccDataType v_acc = 0;
for(int k0 = 0; k0 < K0; ++k0)
{
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(
v_a,
ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
arg.b_element_op_(
v_b,
ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
v_acc += v_a * v_b;
}
AccDataType v_c;
arg.cde_element_op_(v_c, v_acc);
arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
};
make_ParallelTensorFunctor(f_ms_ns,
arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
std::thread::hardware_concurrency());
return 0;
}
float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
{
return true;
}
static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
const Tensor<BDataType>& b_gs_ns_ks,
Tensor<EDataType>& e_gs_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{
a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceContraction_G2_M2_N2_K1"
<< std::endl;
// clang-format on
return str.str();
}
};
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int split_k = 1;
ck::index_t G0 = 1;
ck::index_t G1 = 2;
ck::index_t M0 = 4;
ck::index_t M1 = 256;
ck::index_t N0 = 16;
ck::index_t N1 = 128;
ck::index_t K0 = 64 * 2;
// A[G0, G1, M0, M1, K0]
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
// B[G0, G1, N0, N1, K0]
std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
// D[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
// E[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> e_gs_ms_ns_strides{
G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
if(argc == 1)
{
// use default case
}
else if(argc == 5)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
split_k = std::stoi(argv[4]);
}
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<ADataType> a_gs_ms_ks(
std::vector<std::size_t>(a_gs_ms_ks_lengths.begin(), a_gs_ms_ks_lengths.end()),
std::vector<std::size_t>(a_gs_ms_ks_strides.begin(), a_gs_ms_ks_strides.end()));
Tensor<BDataType> b_gs_ns_ks(
std::vector<std::size_t>(b_gs_ns_ks_lengths.begin(), b_gs_ns_ks_lengths.end()),
std::vector<std::size_t>(b_gs_ns_ks_strides.begin(), b_gs_ns_ks_strides.end()));
Tensor<DDataType> d_gs_ms_ns(
std::vector<std::size_t>(d_gs_ms_ns_lengths.begin(), d_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(d_gs_ms_ns_strides.begin(), d_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_host_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_device_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) *
e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
d_device_buf.ToDevice(d_gs_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{};
// 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_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b_gs_ns_ks_lengths,
b_gs_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
e_gs_ms_ns_lengths,
e_gs_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op,
split_k);
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 G = std::accumulate(e_gs_ms_ns_lengths.begin(),
e_gs_ms_ns_lengths.begin() + NumDimG,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t M = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG,
e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t N = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM + NumDimN,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t K = std::accumulate(a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM,
a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM + NumDimK,
ck::index_t{1},
std::multiplies<ck::index_t>{});
std::size_t flop = std::size_t(2) * G * M * N * K;
std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * 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_gs_ms_ns_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_ms_ns_host_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
CShuffleDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceOpInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
e_gs_ms_ns_host_result.ForEach([&](auto&, auto idx) {
cde_element_op(e_gs_ms_ns_host_result(idx), c_ms_ns_host_result(idx), d_gs_ms_ns(idx));
});
return ck::utils::check_err(e_gs_ms_ns_device_result.mData, e_gs_ms_ns_host_result.mData)
? 0
: 1;
}
return 0;
}
example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp 0 → 100644
View file @ 8b2796c7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, 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/device_splitk_contraction_multiple_d_xdl_cshuffle.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"
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 Add = ck::tensor_operation::element_wise::Add;
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F32;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F32;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 1;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Add;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
// clang-format off
using DeviceOpInstanceKKNN = ck::tensor_operation::device::
//############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| Gemm| A| B| DE| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle|CBlockTransferClusterLengths| CBlockTransfer|
//############################################| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//############################################| | | | | | | | | | | Operation| Operation| Operation| | | | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceSplitKContractionMultipleD_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, ABSpec, ABSpec, DESpec, 1, 256, 256, 128, 32, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, 1, 1, 1, S<1, 32, 1, 4>, 4>;
// clang-format on
using DeviceOpInstance = DeviceOpInstanceKKNN;
// hardcoded for NumDimM == NumDimN == NumDimK == 2
template <ck::index_t NumDimG,
ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename EDataType,
typename AccDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
false>
struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
{
// Argument
struct Argument : public ck::tensor_operation::device::BaseArgument
{
Argument(const Tensor<ADataType>& a_gs_ms_ks,
const Tensor<BDataType>& b_gs_ns_ks,
Tensor<EDataType>& e_gs_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: a_gs_ms_ks_{a_gs_ms_ks},
b_gs_ns_ks_{b_gs_ns_ks},
e_gs_ms_ns_{e_gs_ms_ns},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
}
const Tensor<ADataType>& a_gs_ms_ks_;
const Tensor<BDataType>& b_gs_ns_ks_;
Tensor<EDataType>& e_gs_ms_ns_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
};
// Invoker
struct Invoker : public ck::tensor_operation::device::BaseInvoker
{
using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
float Run(const Argument& arg)
{
auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
AccDataType v_acc = 0;
for(int k0 = 0; k0 < K0; ++k0)
{
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(
v_a,
ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
arg.b_element_op_(
v_b,
ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
v_acc += v_a * v_b;
}
AccDataType v_c;
arg.cde_element_op_(v_c, v_acc);
arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
};
make_ParallelTensorFunctor(f_ms_ns,
arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
std::thread::hardware_concurrency());
return 0;
}
float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
{
return true;
}
static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
const Tensor<BDataType>& b_gs_ns_ks,
Tensor<EDataType>& e_gs_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{
a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceContraction_G2_M2_N2_K1"
<< std::endl;
// clang-format on
return str.str();
}
};
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int split_k = 1;
ck::index_t G0 = 1;
ck::index_t G1 = 2;
ck::index_t M0 = 4;
ck::index_t M1 = 256;
ck::index_t N0 = 16;
ck::index_t N1 = 128;
ck::index_t K0 = 64 * 2;
// A[G0, G1, M0, M1, K0]
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
// B[G0, G1, N0, N1, K0]
std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
// D[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
// E[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> e_gs_ms_ns_strides{
G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
if(argc == 1)
{
// use default case
}
else if(argc == 5)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
split_k = std::stoi(argv[4]);
}
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<ADataType> a_gs_ms_ks(
std::vector<std::size_t>(a_gs_ms_ks_lengths.begin(), a_gs_ms_ks_lengths.end()),
std::vector<std::size_t>(a_gs_ms_ks_strides.begin(), a_gs_ms_ks_strides.end()));
Tensor<BDataType> b_gs_ns_ks(
std::vector<std::size_t>(b_gs_ns_ks_lengths.begin(), b_gs_ns_ks_lengths.end()),
std::vector<std::size_t>(b_gs_ns_ks_strides.begin(), b_gs_ns_ks_strides.end()));
Tensor<DDataType> d_gs_ms_ns(
std::vector<std::size_t>(d_gs_ms_ns_lengths.begin(), d_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(d_gs_ms_ns_strides.begin(), d_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_host_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_device_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) *
e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
d_device_buf.ToDevice(d_gs_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{};
// 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_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b_gs_ns_ks_lengths,
b_gs_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
e_gs_ms_ns_lengths,
e_gs_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op,
split_k);
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 G = std::accumulate(e_gs_ms_ns_lengths.begin(),
e_gs_ms_ns_lengths.begin() + NumDimG,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t M = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG,
e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t N = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM + NumDimN,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t K = std::accumulate(a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM,
a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM + NumDimK,
ck::index_t{1},
std::multiplies<ck::index_t>{});
std::size_t flop = std::size_t(2) * G * M * N * K;
std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * 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_gs_ms_ns_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_ms_ns_host_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
CShuffleDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceOpInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
e_gs_ms_ns_host_result.ForEach([&](auto&, auto idx) {
cde_element_op(e_gs_ms_ns_host_result(idx), c_ms_ns_host_result(idx), d_gs_ms_ns(idx));
});
return ck::utils::check_err(e_gs_ms_ns_device_result.mData, e_gs_ms_ns_host_result.mData)
? 0
: 1;
}
return 0;
}
include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp 0 → 100644
View file @ 8b2796c7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// Tensor Contraction:
// input : A
// input : B
// input : D0, D1, ...
// output : E
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Assume:
// A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
// B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
// D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
// E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
template <index_t NumDimG,
index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceSplitKContractionMultipleD : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
const std::vector<index_t>& a_gs_ms_ns_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths,
const std::vector<index_t>& b_gs_ns_ks_strides,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
const std::vector<index_t>& e_gs_ms_ns_lengths,
const std::vector<index_t>& e_gs_ms_ns_strides,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
index_t split_k) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ 8b2796c7
...@@ -209,6 +209,8 @@ struct BlockToCTileMap_KSplit_M00_N0_M01Adapt ...@@ -209,6 +209,8 @@ struct BlockToCTileMap_KSplit_M00_N0_M01Adapt
const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I0), MPerBlock); const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I0), MPerBlock);
const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I1), NPerBlock); const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I1), NPerBlock);
block_1d_id = block_1d_id % (M0 * N0 * KSplit_); // hide groups
const index_t idx_ksplit = block_1d_id / (M0 * N0); const index_t idx_ksplit = block_1d_id / (M0 * N0);
block_1d_id = block_1d_id % (M0 * N0); block_1d_id = block_1d_id % (M0 * N0);
......
include/ck/tensor_operation/gpu/grid/gridwise_layernorm_welford_variance.hpp
View file @ 8b2796c7
...@@ -57,7 +57,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -57,7 +57,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{}); make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed( using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}))); make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
using ThreadReduceDstDesc_M = using ThreadReduceDstDesc_M =
decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}))); decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
...@@ -73,8 +73,14 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -73,8 +73,14 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
static constexpr auto I1 = Number<1>{}; static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{}; static constexpr auto I2 = Number<2>{};
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize; static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize; static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
static constexpr auto XThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto GammaThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto BetaThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto YThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
__device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k, __device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
int thread_k_cluster_id) int thread_k_cluster_id)
...@@ -87,10 +93,13 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -87,10 +93,13 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
if(kPerBlockTail > 0) if(kPerBlockTail > 0)
{ {
int thread_max_len = (thread_k_cluster_id + 1) * KThreadSliceSize; static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
int delta = thread_max_len - kPerBlockTail; int thread_max_len =
delta = math::clamp(thread_max_len - kPerBlockTail, 0, KThreadSliceSize); (thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
kPerThread += KThreadSliceSize - delta; int delta = thread_max_len - kPerBlockTail;
delta = math::clamp(thread_max_len - kPerBlockTail, 0, XSrcVectorSize);
kPerThread += XSrcVectorSize - delta;
});
} }
return kPerThread; return kPerThread;
...@@ -116,19 +125,41 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -116,19 +125,41 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y_global, y_grid_desc_m_k.GetElementSpaceSize()); p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true> auto x_thread_buf = generate_tuple(
x_thread_buf; [&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true> AccDataType,
gamma_thread_buf; MThreadSliceSize * XSrcVectorSize,
true>{};
StaticBuffer<AddressSpaceEnum::Vgpr, },
AccDataType, Number<XThreadBufferNumber>{});
MThreadSliceSize * KThreadSliceSize,
true>& beta_thread_buf = gamma_thread_buf; auto gamma_thread_buf = generate_tuple(
[&](auto) {
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true> return StaticBuffer<AddressSpaceEnum::Vgpr,
y_thread_buf; AccDataType,
MThreadSliceSize * GammaSrcVectorSize,
true>{};
},
Number<GammaThreadBufferNumber>{});
auto beta_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * BetaSrcVectorSize,
true>{};
},
Number<BetaThreadBufferNumber>{});
auto y_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * YDstVectorSize,
true>{};
},
Number<YThreadBufferNumber>{});
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf; StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf; StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
...@@ -142,9 +173,9 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -142,9 +173,9 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
const auto thread_m_cluster_id = thread_cluster_idx[I0]; const auto thread_m_cluster_id = thread_cluster_idx[I0];
const auto thread_k_cluster_id = thread_cluster_idx[I1]; const auto thread_k_cluster_id = thread_cluster_idx[I1];
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, KThreadSliceSize>; using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, XSrcVectorSize>;
constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed( constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})); make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType, auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
AccDataType, AccDataType,
...@@ -159,7 +190,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -159,7 +190,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
x_grid_desc_m_k, x_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize + make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize, thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize)); thread_k_cluster_id * XSrcVectorSize));
auto threadwise_gamma_load = auto threadwise_gamma_load =
ThreadwiseTensorSliceTransfer_v2<GammaDataType, ThreadwiseTensorSliceTransfer_v2<GammaDataType,
...@@ -175,7 +206,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -175,7 +206,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
gamma_grid_desc_m_k, gamma_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize + make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize, thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize)); thread_k_cluster_id * GammaSrcVectorSize));
auto threadwise_beta_load = auto threadwise_beta_load =
ThreadwiseTensorSliceTransfer_v2<BetaDataType, ThreadwiseTensorSliceTransfer_v2<BetaDataType,
...@@ -191,7 +222,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -191,7 +222,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
beta_grid_desc_m_k, beta_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize + make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize, thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize)); thread_k_cluster_id * BetaSrcVectorSize));
auto threadwise_y_store = auto threadwise_y_store =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType, ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
...@@ -209,13 +240,10 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -209,13 +240,10 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
y_grid_desc_m_k, y_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize + make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize, thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize), thread_k_cluster_id * YDstVectorSize),
acc_elementwise_op); acc_elementwise_op);
// Copy x from Cache constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
// one pass: fwd, second pass: bwd
constexpr auto thread_copy_fwd_step_m_k =
make_multi_index(0, SweepOnce ? 0 : K_BlockTileSize);
constexpr auto thread_copy_bwd_step_m_k = constexpr auto thread_copy_bwd_step_m_k =
make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize); make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
...@@ -238,14 +266,15 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -238,14 +266,15 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles) for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{ {
static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k, threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf, x_global_val_buf,
thread_buffer_desc_m_k, thread_buffer_desc_m_k,
make_tuple(I0, I0), make_tuple(I0, I0),
x_thread_buf); x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k); threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_welford.Run(x_thread_buf, mean_thread_buf, var_thread_buf); threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
});
} }
static_for<0, MThreadSliceSize, 1>{}([&](auto I) { static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
...@@ -256,7 +285,8 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -256,7 +285,8 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count); BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
}); });
auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k; auto thread_copy_tail_m_k =
(num_k_block_tile_iteration - 1) * XThreadBufferNumber * thread_copy_fwd_step_m_k;
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k); threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k); threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
...@@ -267,62 +297,86 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk ...@@ -267,62 +297,86 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
{ {
if constexpr(!SweepOnce) if constexpr(!SweepOnce)
{ {
threadwise_x_load.Run(x_grid_desc_m_k, static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
x_global_val_buf, threadwise_x_load.Run(x_grid_desc_m_k,
thread_buffer_desc_m_k, x_global_val_buf,
make_tuple(I0, I0), thread_buffer_desc_m_k,
x_thread_buf); make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
});
} }
threadwise_gamma_load.Run(gamma_grid_desc_m_k, static_for<0, GammaThreadBufferNumber, 1>{}([&](auto i) {
gamma_global_val_buf, threadwise_gamma_load.Run(gamma_grid_desc_m_k,
thread_buffer_desc_m_k, gamma_global_val_buf,
make_tuple(I0, I0), thread_buffer_desc_m_k,
gamma_thread_buf); make_tuple(I0, I0),
gamma_thread_buf(i));
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) { static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) { auto divisor = 1 / __builtin_amdgcn_sqrtf(var_thread_buf(iM) + epsilon);
constexpr auto offset_m_k = static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK)); static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
// normalize thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
y_thread_buf(Number<offset_m_k>{}) =
(x_thread_buf(Number<offset_m_k>{}) - mean_thread_buf(iM)) / // normalize
sqrt(var_thread_buf(iM) + epsilon); y_thread_buf(iK0)(Number<offset_m_k>{}) =
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
// gamma divisor;
y_thread_buf(Number<offset_m_k>{}) =
y_thread_buf(Number<offset_m_k>{}) * gamma_thread_buf(Number<offset_m_k>{}); // gamma
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) *
gamma_thread_buf(iK0)(Number<offset_m_k>{});
});
}); });
}); });
threadwise_beta_load.Run(beta_grid_desc_m_k, static_for<0, BetaThreadBufferNumber, 1>{}([&](auto i) {
beta_global_val_buf, threadwise_beta_load.Run(beta_grid_desc_m_k,
thread_buffer_desc_m_k, beta_global_val_buf,
make_tuple(I0, I0), thread_buffer_desc_m_k,
beta_thread_buf); make_tuple(I0, I0),
beta_thread_buf(i));
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) { static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) { static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
constexpr auto offset_m_k = static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK)); constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
// beta
y_thread_buf(Number<offset_m_k>{}) = // beta
y_thread_buf(Number<offset_m_k>{}) + beta_thread_buf(Number<offset_m_k>{}); y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) +
beta_thread_buf(iK0)(Number<offset_m_k>{});
});
}); });
}); });
threadwise_y_store.Run(thread_buffer_desc_m_k, static_for<0, YThreadBufferNumber, 1>{}([&](auto i) {
make_tuple(I0, I0), threadwise_y_store.Run(thread_buffer_desc_m_k,
y_thread_buf, make_tuple(I0, I0),
y_grid_desc_m_k, y_thread_buf(i),
y_global_val_buf); y_grid_desc_m_k,
y_global_val_buf);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_fwd_step_m_k);
});
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k); threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_bwd_step_m_k); threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_bwd_step_m_k); 2 * thread_copy_bwd_step_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_bwd_step_m_k); threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
} }
} }
}; };
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f16_instance.cpp
View file @ 8b2796c7
...@@ -31,7 +31,9 @@ using device_layernorm_f16_instances = std::tuple< ...@@ -31,7 +31,9 @@ using device_layernorm_f16_instances = std::tuple<
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>, DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>, DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>, DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8> DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 2, 1, 2, 1, 2, 2>
// clang-format on // clang-format on
>; >;
......
test/layernorm/test_groupnorm_fp16.cpp
View file @ 8b2796c7
...@@ -26,6 +26,8 @@ class TestGroupnorm : public ::testing::Test ...@@ -26,6 +26,8 @@ class TestGroupnorm : public ::testing::Test
{256, 9, 9, 9, 9}, {256, 9, 9, 9, 9},
{1, 64, 64, 32, 10}, {1, 64, 64, 32, 10},
{1, 32, 32, 32, 20}, {1, 32, 32, 32, 20},
{2, 32, 32, 32, 30},
{2, 32, 32, 32, 40},
{1, 16, 16, 32, 40}}; {1, 16, 16, 32, 40}};
for(auto length : lengths) for(auto length : lengths)
......
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