Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "3991a1c132cd2cd51bbe074a02b0b5a4da6bc6b2"
Commit 289f15de authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into wmma_gemm

parents 9bd44685 d58b7f51
Show whitespace changes
Inline Side-by-side
Showing with 587 additions and 181 deletions
example/12_reduce/reduce_blockwise_two_call.cpp
View file @ 289f15de
...@@ -221,12 +221,12 @@ int main(int argc, char* argv[]) ...@@ -221,12 +221,12 @@ int main(int argc, char* argv[])
std::array<index_t, 3> arrOutLengths; std::array<index_t, 3> arrOutLengths;
std::array<index_t, 3> arrOutStrides; std::array<index_t, 3> arrOutStrides;
std::copy(inLengths_1.begin(), inLengths_1.end(), arrInLengths_1.begin()); ck::ranges::copy(inLengths_1, arrInLengths_1.begin());
std::copy(inStrides_1.begin(), inStrides_1.end(), arrInStrides_1.begin()); ck::ranges::copy(inStrides_1, arrInStrides_1.begin());
std::copy(inLengths_2.begin(), inLengths_2.end(), arrInLengths_2.begin()); ck::ranges::copy(inLengths_2, arrInLengths_2.begin());
std::copy(inStrides_2.begin(), inStrides_2.end(), arrInStrides_2.begin()); ck::ranges::copy(inStrides_2, arrInStrides_2.begin());
std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin()); ck::ranges::copy(outLengths, arrOutLengths.begin());
std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin()); ck::ranges::copy(outStrides, arrOutStrides.begin());
auto reduce_1 = DeviceReduceInstance_1{}; auto reduce_1 = DeviceReduceInstance_1{};
...@@ -294,7 +294,7 @@ int main(int argc, char* argv[]) ...@@ -294,7 +294,7 @@ int main(int argc, char* argv[])
if(do_verify) if(do_verify)
{ {
out_dev.FromDevice(out.mData.data()); out_dev.FromDevice(out.mData.data());
pass = pass && ck::utils::check_err(out.mData, out_ref.mData); pass = pass && ck::utils::check_err(out, out_ref);
}; };
return (pass ? 0 : 1); return (pass ? 0 : 1);
......
example/12_reduce/reduce_multiblock_atomic_add.cpp
View file @ 289f15de
...@@ -140,7 +140,7 @@ bool reduce_multiblock_atomic_add_test(bool do_verification, ...@@ -140,7 +140,7 @@ bool reduce_multiblock_atomic_add_test(bool do_verification,
std::array<int, ShapeType::NumReduceDim_> a_reduceDims; std::array<int, ShapeType::NumReduceDim_> a_reduceDims;
std::copy(reduceDims.begin(), reduceDims.end(), a_reduceDims.begin()); ck::ranges::copy(reduceDims, a_reduceDims.begin());
result = reduce_multiblock_atomic_add_impl<InOutDataType, result = reduce_multiblock_atomic_add_impl<InOutDataType,
AccDataType, AccDataType,
......
example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
View file @ 289f15de
...@@ -10,6 +10,7 @@ ...@@ -10,6 +10,7 @@
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp" #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp" #include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
...@@ -176,10 +177,10 @@ int reduce_multiblock_atomic_add_impl(bool do_verification, ...@@ -176,10 +177,10 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
std::array<index_t, NumOutDim> arrOutLengths; std::array<index_t, NumOutDim> arrOutLengths;
std::array<index_t, NumOutDim> arrOutStrides; std::array<index_t, NumOutDim> arrOutStrides;
std::copy(inLengths.begin(), inLengths.end(), arrInLengths.begin()); ck::ranges::copy(inLengths, arrInLengths.begin());
std::copy(inStrides.begin(), inStrides.end(), arrInStrides.begin()); ck::ranges::copy(inStrides, arrInStrides.begin());
std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin()); ck::ranges::copy(outLengths, arrOutLengths.begin());
std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin()); ck::ranges::copy(outStrides, arrOutStrides.begin());
auto reduce = DeviceReduceInstance{}; auto reduce = DeviceReduceInstance{};
...@@ -225,7 +226,7 @@ int reduce_multiblock_atomic_add_impl(bool do_verification, ...@@ -225,7 +226,7 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
if(do_verification) if(do_verification)
{ {
out_dev.FromDevice(out.mData.data()); out_dev.FromDevice(out.mData.data());
pass = pass && ck::utils::check_err(out.mData, out_ref.mData); pass = pass && ck::utils::check_err(out, out_ref);
}; };
return (pass ? 0 : 1); return (pass ? 0 : 1);
......
example/13_pool2d_fwd/pool2d_fwd_common.hpp
View file @ 289f15de
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
template <typename InDataType, template <typename InDataType,
typename OutDataType, typename OutDataType,
...@@ -172,16 +173,16 @@ bool pool_test(bool do_verification, ...@@ -172,16 +173,16 @@ bool pool_test(bool do_verification,
// tensor layout // tensor layout
auto f_host_tensor_descriptor = auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) { [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
using namespace ck::literals;
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value) if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}), return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
} }
else if constexpr(ck::is_same<decltype(layout), else if constexpr(ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWC>::value) ck::tensor_layout::convolution::NHWC>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}), return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
} }
}; };
...@@ -267,14 +268,14 @@ bool pool_test(bool do_verification, ...@@ -267,14 +268,14 @@ bool pool_test(bool do_verification,
out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data()); out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
pass = pass && ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData); pass = pass && ck::utils::check_err(out_n_c_ho_wo_device, out_n_c_ho_wo_host);
if constexpr(OutputIndex) if constexpr(OutputIndex)
{ {
out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data()); out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
pass = pass && ck::utils::check_err(out_indices_n_c_ho_wo_device.mData, pass = pass &&
out_indices_n_c_ho_wo_host.mData); ck::utils::check_err(out_indices_n_c_ho_wo_device, out_indices_n_c_ho_wo_host);
}; };
} }
......
example/14_gemm_quantization/CMakeLists.txt 0 → 100644
View file @ 289f15de
add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
\ No newline at end of file
example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp 0 → 100644
View file @ 289f15de
// 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/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_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 I8 = int8_t;
using I32 = int32_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu;
using CDEElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
using ADataType = I8;
using BDataType = I8;
using AccDataType = I32;
using CShuffleDataType = I32;
using BiasDataType = I32;
using DsDataType = ck::Tuple<BiasDataType>;
using EDataType = I8;
using ALayout = Row;
using BLayout = Col;
using BiasLayout = Row;
using DsLayout = ck::Tuple<BiasLayout>;
using ELayout = Row;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
PassThrough, // AElementwiseOperation,
PassThrough, // BElementwiseOperation,
CDEElementOp, // CDEElementwiseOperation,
GemmDefault, // GemmSpecialization GemmSpec,
1, // NumGemmKPrefetchStage,
256, // BlockSize,
256, // MPerBlock,
128, // NPerBlock,
64, // KPerBlock,
16, // AK1,
16, // BK1,
32, // MPerXDL,
32, // NPerXDL,
4, // MXdlPerWave,
2, // NXdlPerWave,
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1,
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
S<1, 0, 2>, // ABlockTransferSrcAccessOrder,
2, // index_t ABlockTransferSrcVectorDim,
16, // index_t ABlockTransferSrcScalarPerVector,
16, // index_t ABlockTransferDstScalarPerVector_AK1,
1, // bool ABlockLdsExtraM,
S<4, 64, 1>, // typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
S<1, 0, 2>, // typename BBlockTransferThreadClusterArrangeOrder,
S<1, 0, 2>, // typename BBlockTransferSrcAccessOrder,
2, // index_t BBlockTransferSrcVectorDim,
8, // index_t BBlockTransferSrcScalarPerVector,
8, // index_t BBlockTransferDstScalarPerVector_BK1,
1, // bool BBlockLdsExtraN,
1, // index_t CShuffleMXdlPerWavePerShuffle,
1, // index_t CShuffleNXdlPerWavePerShuffle,
S<1, 64, 1, 4>, // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
8>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
int main()
{
bool do_verification = true;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideBias = 0;
ck::index_t StrideE = 1024;
float requant_scale = 0.03;
auto f_host_tensor_descriptor2d =
[](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(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1_uz}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1_uz, stride}));
}
};
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}),
std::vector<std::size_t>({stride}));
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "bias_n: " << bias_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
bias_n.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-128, 127});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
bias_device_buf.ToDevice(bias_n.mData.data());
auto a_element_op = PassThrough{};
auto b_element_op = PassThrough{};
auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{bias_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{StrideBias},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
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, "
<< gemm.GetTypeString() << std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
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, a_element_op, 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), bias_n(n));
}
}
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
example/14_gemm_xdl_quantization/gemm_xdl_relu_quantization_int8.cpp → example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
View file @ 289f15de
...@@ -9,62 +9,72 @@ ...@@ -9,62 +9,72 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using I8 = int8_t;
using I32 = int32_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu; using ActivationOp = PassThrough;
using CElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>; using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
using ADataType = int8_t; using ADataType = I8;
using BDataType = int8_t; using BDataType = I8;
using CDataType = int8_t; using AccDataType = I32;
using AccDataType = int32_t; using CShuffleDataType = I32;
using CShuffleDataType = float; using DsDataType = ck::Tuple<>;
using EDataType = I8;
using ALayout = ck::tensor_layout::gemm::RowMajor; using ALayout = Row;
using BLayout = ck::tensor_layout::gemm::ColumnMajor; using BLayout = Col;
using CLayout = ck::tensor_layout::gemm::RowMajor; using DsLayout = ck::Tuple<>;
using ELayout = Row;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off // clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle< using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
ALayout, // typename ALayout, ALayout,
BLayout, // typename BLayout, BLayout,
CLayout, // typename CLayout, DsLayout,
ADataType, // typename ADataType, ELayout,
BDataType, // typename BDataType, ADataType,
CDataType, // typename CDataType, BDataType,
AccDataType, // typename GemmAccDataType, AccDataType,
CShuffleDataType, // typename CShuffleDataType, CShuffleDataType,
PassThrough, // typename AElementwiseOperation, DsDataType,
PassThrough, // typename BElementwiseOperation, EDataType,
CElementOp, // typename CElementwiseOperation, PassThrough, // AElementwiseOperation,
PassThrough, // BElementwiseOperation,
CDEElementOp, // CDEElementwiseOperation,
GemmDefault, // GemmSpecialization GemmSpec, GemmDefault, // GemmSpecialization GemmSpec,
1, // index_t NumGemmKPrefetchStage, 1, // NumGemmKPrefetchStage,
256, // index_t BlockSize, 256, // BlockSize,
256, // index_t MPerBlock, 256, // MPerBlock,
128, // index_t NPerBlock, 128, // NPerBlock,
64, // index_t KPerBlock, 64, // KPerBlock,
16, // index_t AK1, 16, // AK1,
16, // index_t BK1, 16, // BK1,
32, // index_t MPerXDL, 32, // MPerXDL,
32, // index_t NPerXDL, 32, // NPerXDL,
4, // index_t MXdlPerWave, 4, // MXdlPerWave,
2, // index_t NXdlPerWave, 2, // NXdlPerWave,
S<4, 64, 1>, // typename ABlockTransferThreadClusterLengths_AK0_M_AK1, S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1,
S<1, 0, 2>, // typename ABlockTransferThreadClusterArrangeOrder, S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
S<1, 0, 2>, // typename ABlockTransferSrcAccessOrder, S<1, 0, 2>, // ABlockTransferSrcAccessOrder,
2, // index_t ABlockTransferSrcVectorDim, 2, // index_t ABlockTransferSrcVectorDim,
16, // index_t ABlockTransferSrcScalarPerVector, 16, // index_t ABlockTransferSrcScalarPerVector,
16, // index_t ABlockTransferDstScalarPerVector_AK1, 16, // index_t ABlockTransferDstScalarPerVector_AK1,
...@@ -83,115 +93,80 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle ...@@ -83,115 +93,80 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, float, PassThrough, PassThrough, CElementOp>; ReferenceGemm<ADataType, BDataType, EDataType, float, PassThrough, PassThrough, CDEElementOp>;
int main(int argc, char* argv[]) int main()
{ {
bool do_verification = true; bool do_verification = true;
int init_method = 1;
bool time_kernel = false; bool time_kernel = false;
// GEMM shape // GEMM shape
ck::index_t M = 3840; ck::index_t M = 1024;
ck::index_t N = 4096; ck::index_t N = 1024;
ck::index_t K = 4096; ck::index_t K = 1024;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
float quant_multiplier = 0.03;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]); ck::index_t StrideA = 1024;
N = std::stoi(argv[5]); ck::index_t StrideB = 1024;
K = std::stoi(argv[6]); ck::index_t StrideE = 1024;
StrideA = std::stoi(argv[7]); float requant_scale = 0.03;
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
auto f_host_tensor_descriptor = auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) { [](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) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1})); std::vector<std::size_t>({stride, 1_uz}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride})); std::vector<std::size_t>({1_uz, stride}));
} }
}; };
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl; std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method) a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
{ b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize()); DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize()); DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize()); DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data()); a_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data()); b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = PassThrough{}; auto a_element_op = PassThrough{};
auto b_element_op = PassThrough{}; auto b_element_op = PassThrough{};
auto c_element_op = CElementOp{quant_multiplier, ActivationOp{}}; auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
// do GEMM // do GEMM
auto gemm = DeviceGemmInstance{}; auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker(); auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()), auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()), b_device_buf.GetDeviceBuffer(),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()), {},
e_device_buf.GetDeviceBuffer(),
M, M,
N, N,
K, K,
StrideA, StrideA,
StrideB, StrideB,
StrideC, {},
StrideE,
a_element_op, a_element_op,
b_element_op, b_element_op,
c_element_op); cde_element_op);
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
{ {
...@@ -204,7 +179,7 @@ int main(int argc, char* argv[]) ...@@ -204,7 +179,7 @@ int main(int argc, char* argv[])
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N; sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
...@@ -213,7 +188,7 @@ int main(int argc, char* argv[]) ...@@ -213,7 +188,7 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl; << gemm.GetTypeString() << std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data()); e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification) if(do_verification)
{ {
...@@ -221,11 +196,11 @@ int main(int argc, char* argv[]) ...@@ -221,11 +196,11 @@ int main(int argc, char* argv[])
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument( auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op); a_m_k, b_k_n, e_m_n_host_result, a_element_op, b_element_op, cde_element_op);
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1; return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
} }
return 0; return 0;
......
example/14_gemm_xdl_quantization/CMakeLists.txt deleted 100644 → 0
View file @ 9bd44685
add_example_executable(example_gemm_xdl_relu_quantization_int8 gemm_xdl_relu_quantization_int8.cpp)
\ No newline at end of file
example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp
View file @ 289f15de
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.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/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
......
example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
View file @ 289f15de
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.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/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
......
example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
View file @ 289f15de
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.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/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
......
example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
View file @ 289f15de
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.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/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
......
example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
View file @ 289f15de
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.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/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
......
example/15_grouped_gemm/run_grouped_gemm_example.inc
View file @ 289f15de
...@@ -52,15 +52,15 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co ...@@ -52,15 +52,15 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
auto f_host_tensor_descriptor = auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) { [](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) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {stride, 1_uz});
std::vector<std::size_t>({stride, 1}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {1_uz, stride});
std::vector<std::size_t>({1, stride}));
} }
}; };
...@@ -208,10 +208,10 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co ...@@ -208,10 +208,10 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
#ifdef BUILD_INT4_EXAMPLE #ifdef BUILD_INT4_EXAMPLE
const Tensor<EDataType> c_device_result_converted(c_device_tensors[i]); const Tensor<EDataType> c_device_result_converted(c_device_tensors[i]);
pass &= ck::utils::check_err(c_device_result_converted.mData, c_host_tensors[i].mData); pass &= ck::utils::check_err(c_device_result_converted, c_host_tensors[i]);
#else #else
pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData); pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
#endif #endif
} }
} }
......
example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
View file @ 289f15de
...@@ -15,6 +15,7 @@ ...@@ -15,6 +15,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.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/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -109,21 +110,20 @@ void DumpPerf(float ave_time, int M, int N, int K) ...@@ -109,21 +110,20 @@ void DumpPerf(float ave_time, int M, int N, int K)
} }
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) { auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}), return HostTensorDescriptor({len}, {stride});
std::vector<std::size_t>({stride}));
}; };
auto f_host_tensor_descriptor2d = auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) { [](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) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {stride, 1_uz});
std::vector<std::size_t>({stride, 1}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {1_uz, stride});
std::vector<std::size_t>({1, stride}));
} }
}; };
...@@ -259,12 +259,9 @@ int main() ...@@ -259,12 +259,9 @@ int main()
r0_device_buf.FromDevice(r0_m.mData.data()); r0_device_buf.FromDevice(r0_m.mData.data());
r1_device_buf.FromDevice(r1_m.mData.data()); r1_device_buf.FromDevice(r1_m.mData.data());
pass = ck::utils::check_err( pass = ck::utils::check_err(e_m_n, e_m_n_host, "Error: Incorrect results c", 1e-2, 1e-2);
e_m_n.mData, e_m_n_host.mData, "Error: Incorrect results c", 1e-2, 1e-2); pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
pass &= ck::utils::check_err( pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r1_m.mData, r1_m_host.mData, "Error: Incorrect results d1", 1e-2, 1e-2);
} }
bool time_kernel = true; bool time_kernel = true;
......
example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
View file @ 289f15de
...@@ -160,14 +160,12 @@ bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M, ...@@ -160,14 +160,12 @@ bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M,
{ {
case 0: break; case 0: break;
case 1: case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k.begin(), ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
a_m_k.end()); ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n.begin(),
b_k_n.end());
break; break;
default: default:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k.begin(), a_m_k.end()); ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n.begin(), b_k_n.end()); ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
break; break;
} }
...@@ -264,15 +262,13 @@ bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M, ...@@ -264,15 +262,13 @@ bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M,
Tensor<EDataType> e_m_n_host_converted(e_m_n_host); Tensor<EDataType> e_m_n_host_converted(e_m_n_host);
pass = ck::utils::check_err( pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host_converted.mData, "Error: Incorrect results c", 1e-2, 1e-2); e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
r0_device_buf.FromDevice(r0_m.mData.data()); r0_device_buf.FromDevice(r0_m.mData.data());
r1_device_buf.FromDevice(r1_m.mData.data()); r1_device_buf.FromDevice(r1_m.mData.data());
pass &= ck::utils::check_err( pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2); pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r1_m.mData, r1_m_host.mData, "Error: Incorrect results d1", 1e-2, 1e-2);
if(pass) if(pass)
{ {
......
example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
View file @ 289f15de
...@@ -134,14 +134,12 @@ auto run_gemm_reduce_max_xdl(ck::index_t M, ...@@ -134,14 +134,12 @@ auto run_gemm_reduce_max_xdl(ck::index_t M,
{ {
case 0: break; case 0: break;
case 1: case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k.begin(), ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
a_m_k.end()); ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n.begin(),
b_k_n.end());
break; break;
default: default:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k.begin(), a_m_k.end()); ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n.begin(), b_k_n.end()); ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
break; break;
} }
...@@ -243,8 +241,8 @@ auto run_gemm_reduce_max_xdl(ck::index_t M, ...@@ -243,8 +241,8 @@ auto run_gemm_reduce_max_xdl(ck::index_t M,
if constexpr(std::is_same_v<ADataType, ck::int4_t>) if constexpr(std::is_same_v<ADataType, ck::int4_t>)
{ {
Tensor<EDataType> e_m_n_device_converted(e_m_n); Tensor<EDataType> e_m_n_device_converted(e_m_n);
pass = ck::utils::check_err(e_m_n_device_converted.mData, pass = ck::utils::check_err(e_m_n_device_converted,
e_m_n_host_converted.mData, e_m_n_host_converted,
"Error: Incorrect results c", "Error: Incorrect results c",
1e-2, 1e-2,
1e-2); 1e-2);
...@@ -253,12 +251,11 @@ auto run_gemm_reduce_max_xdl(ck::index_t M, ...@@ -253,12 +251,11 @@ auto run_gemm_reduce_max_xdl(ck::index_t M,
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4 #endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
{ {
pass = ck::utils::check_err( pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host_converted.mData, "Error: Incorrect results c", 1e-2, 1e-2); e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
} }
r0_device_buf.FromDevice(r0_m.mData.data()); r0_device_buf.FromDevice(r0_m.mData.data());
pass &= ck::utils::check_err( pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2);
if(pass) if(pass)
{ {
...@@ -339,14 +336,12 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M, ...@@ -339,14 +336,12 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M,
{ {
case 0: break; case 0: break;
case 1: case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k.begin(), ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
a_m_k.end()); ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n.begin(),
b_k_n.end());
break; break;
default: default:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k.begin(), a_m_k.end()); ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n.begin(), b_k_n.end()); ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
break; break;
} }
...@@ -460,8 +455,8 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M, ...@@ -460,8 +455,8 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M,
if constexpr(std::is_same_v<ADataType, ck::int4_t>) if constexpr(std::is_same_v<ADataType, ck::int4_t>)
{ {
Tensor<EDataType> e_m_n_device_converted(e_m_n); Tensor<EDataType> e_m_n_device_converted(e_m_n);
pass = ck::utils::check_err(e_m_n_device_converted.mData, pass = ck::utils::check_err(e_m_n_device_converted,
e_m_n_host_converted.mData, e_m_n_host_converted,
"Error: Incorrect results c", "Error: Incorrect results c",
1e-2, 1e-2,
1e-2); 1e-2);
...@@ -470,16 +465,14 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M, ...@@ -470,16 +465,14 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M,
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4 #endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
{ {
pass = ck::utils::check_err( pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host_converted.mData, "Error: Incorrect results c", 1e-2, 1e-2); e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
} }
r0_device_buf.FromDevice(r0_m.mData.data()); r0_device_buf.FromDevice(r0_m.mData.data());
r1_device_buf.FromDevice(r1_m.mData.data()); r1_device_buf.FromDevice(r1_m.mData.data());
pass &= ck::utils::check_err( pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2); pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r1_m.mData, r1_m_host.mData, "Error: Incorrect results d1", 1e-2, 1e-2);
if(pass) if(pass)
{ {
......
example/17_convnd_bwd_data/CMakeLists.txt
View file @ 289f15de
add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp) add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility) target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
View file @ 289f15de
...@@ -61,9 +61,13 @@ int run_conv_bwd_data(bool do_verification, ...@@ -61,9 +61,13 @@ int run_conv_bwd_data(bool do_verification,
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}); out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5}); wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break; break;
default: case 2:
out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0}); out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5}); wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
break;
default:
out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
} }
DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize()); DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
...@@ -98,9 +102,8 @@ int run_conv_bwd_data(bool do_verification, ...@@ -98,9 +102,8 @@ int run_conv_bwd_data(bool do_verification,
if(!conv.IsSupportedArgument(argument)) if(!conv.IsSupportedArgument(argument))
{ {
throw std::runtime_error( std::cout << "Not support,please check parameters or device";
"wrong! device_conv with the specified compilation parameters does " return 0;
"not support this Conv problem");
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
...@@ -142,7 +145,7 @@ int run_conv_bwd_data(bool do_verification, ...@@ -142,7 +145,7 @@ int run_conv_bwd_data(bool do_verification,
in_device_buf.FromDevice(in_device.mData.data()); in_device_buf.FromDevice(in_device.mData.data());
return ck::utils::check_err(in_device.mData, in_host.mData) ? 0 : 1; return ck::utils::check_err(in_device, in_host) ? 0 : 1;
} }
return 0; return 0;
......
example/20_convnd_bwd_weight/convnd_bwd_weight_xdl_fp16.cpp → example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
View file @ 289f15de
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_bwd_weight_common.hpp" #include "convnd_bwd_data_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_weight_nwc_kxc_nwk_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
using InDataType = ck::half_t; using InDataType = ck::half_t;
using WeiDataType = ck::half_t; using WeiDataType = ck::half_t;
...@@ -17,61 +17,31 @@ using InElementOp = ck::tensor_operation::element_wise::PassThrough; ...@@ -17,61 +17,31 @@ using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough; using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault = static constexpr auto ConvBwdDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default; ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
template <ck::index_t NDimSpatial> template <ck::index_t NDimSpatial>
using DeviceConvndBwdWeightInstance = // clang-format off
ck::tensor_operation::device::DeviceConvNdBwdWeightNwcKxcNwk_Xdl_CShuffle< using DeviceConvNdBwdDataInstance = ck::tensor_operation::device::DeviceConvNdBwdDataNwcKxcNwk_Dl<
NDimSpatial, // NDimSpatial // ######| NDim| InData| WeiData| OutData| AccData| In| Wei| Out| Convolution| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
InDataType, // InDataType // ######| Spatial| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Forward| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
WeiDataType, // WeiDataType // ######| | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
OutDataType, // OutDataType // ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
AccDataType, // AccDataType NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InElementOp, WeiElementOp, OutElementOp, ConvBwdDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<1, 1, 8, 2>, S<16, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
InElementOp, // InElementwiseOperation // clang-format on
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
namespace ctc = ck::tensor_layout::convolution; namespace ctc = ck::tensor_layout::convolution;
print_helper_msg();
bool do_verification = true; bool do_verification = true;
int init_method = 1; int init_method = 1;
bool time_kernel = false; bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{ ck::utils::conv::ConvParam conv_param{
2, 1, 32, 256, 1024, {3, 3}, {14, 14}, {2, 2}, {1, 1}, {1, 1}, {1, 1}}; 2, 1, 128, 256, 256, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
ck::index_t split_k = 4;
if(argc == 1) if(argc == 1)
{ {
...@@ -91,9 +61,6 @@ int main(int argc, char* argv[]) ...@@ -91,9 +61,6 @@ int main(int argc, char* argv[])
const ck::index_t num_dim_spatial = std::stoi(argv[4]); const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv); conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
split_k = std::stoi(argv[5 + 3 + 6 * num_dim_spatial - 1]);
split_k = std::max(1, split_k);
} }
const auto in_element_op = InElementOp{}; const auto in_element_op = InElementOp{};
...@@ -118,14 +85,14 @@ int main(int argc, char* argv[]) ...@@ -118,14 +85,14 @@ int main(int argc, char* argv[])
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>( ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param); conv_param);
return run_conv_bwd_weight<1, return run_conv_bwd_data<1,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvndBwdWeightInstance<1>>(do_verification, DeviceConvNdBwdDataInstance<1>>(do_verification,
init_method, init_method,
time_kernel, time_kernel,
conv_param, conv_param,
...@@ -134,8 +101,7 @@ int main(int argc, char* argv[]) ...@@ -134,8 +101,7 @@ int main(int argc, char* argv[])
out_g_n_k_wos_desc, out_g_n_k_wos_desc,
in_element_op, in_element_op,
wei_element_op, wei_element_op,
out_element_op, out_element_op);
split_k);
} }
else if(conv_param.num_dim_spatial_ == 2) else if(conv_param.num_dim_spatial_ == 2)
{ {
...@@ -155,14 +121,14 @@ int main(int argc, char* argv[]) ...@@ -155,14 +121,14 @@ int main(int argc, char* argv[])
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>( ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param); conv_param);
return run_conv_bwd_weight<2, return run_conv_bwd_data<2,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvndBwdWeightInstance<2>>(do_verification, DeviceConvNdBwdDataInstance<2>>(do_verification,
init_method, init_method,
time_kernel, time_kernel,
conv_param, conv_param,
...@@ -171,8 +137,7 @@ int main(int argc, char* argv[]) ...@@ -171,8 +137,7 @@ int main(int argc, char* argv[])
out_g_n_k_wos_desc, out_g_n_k_wos_desc,
in_element_op, in_element_op,
wei_element_op, wei_element_op,
out_element_op, out_element_op);
split_k);
} }
else if(conv_param.num_dim_spatial_ == 3) else if(conv_param.num_dim_spatial_ == 3)
{ {
...@@ -192,14 +157,14 @@ int main(int argc, char* argv[]) ...@@ -192,14 +157,14 @@ int main(int argc, char* argv[])
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>( ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param); conv_param);
return run_conv_bwd_weight<3, return run_conv_bwd_data<3,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvndBwdWeightInstance<3>>(do_verification, DeviceConvNdBwdDataInstance<3>>(do_verification,
init_method, init_method,
time_kernel, time_kernel,
conv_param, conv_param,
...@@ -208,8 +173,7 @@ int main(int argc, char* argv[]) ...@@ -208,8 +173,7 @@ int main(int argc, char* argv[])
out_g_n_k_wos_desc, out_g_n_k_wos_desc,
in_element_op, in_element_op,
wei_element_op, wei_element_op,
out_element_op, out_element_op);
split_k);
} }
return 0; return 0;
......
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