Skip to content

GitLab

Commit aa374621 authored by Bartlomiej Kocot's avatar Bartlomiej Kocot Committed by Bartłomiej Kocot
Browse files

Add example

parent 6a9a2dc0
Show whitespace changes
Inline Side-by-side
Showing with 581 additions and 130 deletions
example/62_conv_fwd_activ/CMakeLists.txt
View file @ aa374621
...@@ -30,6 +30,15 @@ foreach(gpu IN LISTS GPU_TARGETS) ...@@ -30,6 +30,15 @@ foreach(gpu IN LISTS GPU_TARGETS)
# Elu # Elu
add_example_executable(example_convnd_fwd_xdl_elu_fp16 convnd_fwd_xdl_elu_fp16.cpp) add_example_executable(example_convnd_fwd_xdl_elu_fp16 convnd_fwd_xdl_elu_fp16.cpp)
add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_elu_fp16) add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_elu_fp16)
# ScaleAdd on A and B
add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp16 multi_AB/conv_fwd_xdl_scaleadd_ab_fp16.cpp)
add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_fp16)
add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp32 multi_AB/conv_fwd_xdl_scaleadd_ab_fp32.cpp)
add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_fp32)
add_example_executable(example_conv_fwd_xdl_scaleadd_ab_bf16 multi_AB/conv_fwd_xdl_scaleadd_ab_bf16.cpp)
add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_bf16)
add_example_executable(example_conv_fwd_xdl_scaleadd_ab_int8 multi_AB/conv_fwd_xdl_scaleadd_ab_int8.cpp)
add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_int8)
# ScaleAdd ScaleAdd Relu # ScaleAdd ScaleAdd Relu
add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp) add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp)
add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16) add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16)
......
example/62_conv_fwd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_bf16.cpp 0 → 100644
View file @ aa374621
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_multi_ab_common.hpp"
using DataType = ck::bhalf_t;
using AccDataType = float;
using InDataType = DataType;
using WeiDataType = DataType;
using OutDataType = DataType;
using ADataTypes = ck::Tuple<DataType, DataType>;
using BDataTypes = ck::Tuple<DataType, DataType>;
using InElementOp = ck::tensor_operation::element_wise::ScaleAdd;
using WeiElementOp = ck::tensor_operation::element_wise::ScaleAdd;
using DeviceGroupedConvNDFwdActivInstance = DeviceGroupedConvNDMultiABFwdInstance<DataType,
AccDataType,
ADataTypes,
BDataTypes,
InElementOp,
WeiElementOp>;
#include "../run_convnd_fwd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_fwd_example(argc, argv); }
example/62_conv_fwd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp16.cpp 0 → 100644
View file @ aa374621
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_multi_ab_common.hpp"
using DataType = ck::half_t;
using AccDataType = float;
using InDataType = DataType;
using WeiDataType = DataType;
using OutDataType = DataType;
using ADataTypes = ck::Tuple<DataType, DataType>;
using BDataTypes = ck::Tuple<DataType, DataType>;
using InElementOp = ck::tensor_operation::element_wise::ScaleAdd;
using WeiElementOp = ck::tensor_operation::element_wise::ScaleAdd;
using DeviceGroupedConvNDFwdActivInstance = DeviceGroupedConvNDMultiABFwdInstance<DataType,
AccDataType,
ADataTypes,
BDataTypes,
InElementOp,
WeiElementOp>;
#include "../run_convnd_fwd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_fwd_example(argc, argv); }
example/62_conv_fwd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp32.cpp 0 → 100644
View file @ aa374621
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_multi_ab_common.hpp"
using DataType = float;
using AccDataType = float;
using InDataType = DataType;
using WeiDataType = DataType;
using OutDataType = DataType;
using ADataTypes = ck::Tuple<DataType, DataType>;
using BDataTypes = ck::Tuple<DataType, DataType>;
using InElementOp = ck::tensor_operation::element_wise::ScaleAdd;
using WeiElementOp = ck::tensor_operation::element_wise::ScaleAdd;
using DeviceGroupedConvNDFwdActivInstance = DeviceGroupedConvNDMultiABFwdInstance<DataType,
AccDataType,
ADataTypes,
BDataTypes,
InElementOp,
WeiElementOp>;
#include "../run_convnd_fwd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_fwd_example(argc, argv); }
example/62_conv_fwd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_int8.cpp 0 → 100644
View file @ aa374621
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_multi_ab_common.hpp"
using DataType = int8_t;
using AccDataType = int32_t;
using InDataType = DataType;
using WeiDataType = DataType;
using OutDataType = DataType;
using ADataTypes = ck::Tuple<DataType, DataType>;
using BDataTypes = ck::Tuple<DataType, DataType>;
using InElementOp = ck::tensor_operation::element_wise::ScaleAdd;
using WeiElementOp = ck::tensor_operation::element_wise::ScaleAdd;
using DeviceGroupedConvNDFwdActivInstance = DeviceGroupedConvNDMultiABFwdInstance<DataType,
AccDataType,
ADataTypes,
BDataTypes,
InElementOp,
WeiElementOp>;
#include "../run_convnd_fwd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_fwd_example(argc, argv); }
example/62_conv_fwd_activ/multi_AB/convnd_fwd_activ_multi_ab_common.hpp 0 → 100644
View file @ aa374621
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <type_traits>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
constexpr ck::index_t NDimSpatial = 3;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InLayout = ck::tensor_layout::convolution::GNDHWC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::GNDHWK;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <typename DataType,
typename AccDataType,
typename InDataTypes,
typename WeiDataTypes,
typename InElementOp,
typename WeiElementOp>
using DeviceGroupedConvNDMultiABFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataTypes,
WeiDataTypes,
AccDataType,
DataType,
ck::Tuple<>,
DataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 32, 1, 8>,
8>;
namespace {
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp,
typename DeviceConvNDFwdInstance>
bool run_grouped_conv_fwd(bool do_verification,
int init_method,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
const HostTensorDescriptor& in_g_n_c_wis_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
constexpr ck::index_t NumAs = 2;
constexpr ck::index_t NumBs = 2;
Tensor<InDataType> in(in_g_n_c_wis_desc);
Tensor<InDataType> in_bias(in_g_n_c_wis_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
Tensor<WeiDataType> wei_bias(wei_g_k_c_xs_desc);
Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
std::cout << "in: " << in.mDesc << std::endl;
std::cout << "wei: " << wei.mDesc << std::endl;
std::cout << "out: " << out_host.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
in_bias.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
wei_bias.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
in_bias.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.05, 0.05});
wei_bias.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1.0, 1.0});
}
DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
DeviceMem in_bias_device_buf(sizeof(InDataType) * in_bias.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
DeviceMem wei_bias_device_buf(sizeof(WeiDataType) * wei_bias.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
in_device_buf.ToDevice(in.mData.data());
in_bias_device_buf.ToDevice(in_bias.mData.data());
wei_device_buf.ToDevice(wei.mData.data());
wei_bias_device_buf.ToDevice(wei_bias.mData.data());
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads);
std::array<const void*, NumAs> as{in_device_buf.GetDeviceBuffer(),
in_bias_device_buf.GetDeviceBuffer()};
std::array<const void*, NumBs> bs{wei_device_buf.GetDeviceBuffer(),
wei_bias_device_buf.GetDeviceBuffer()};
std::array<const void*, 0> ds{};
// do Conv
auto conv = DeviceConvNDFwdInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(as,
bs,
ds,
out_device_buf.GetDeviceBuffer(),
a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
{},
{},
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
if(!conv.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem");
}
float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = conv_param.GetFlops() +
2 * conv_param.GetOutputByte<InDataType>() / sizeof(InDataType) +
2 * conv_param.GetOutputByte<WeiDataType>() / sizeof(WeiDataType);
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>() +
conv_param.GetInputByte<InDataType>() +
conv_param.GetWeightByte<WeiDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< conv.GetTypeString() << std::endl;
if(do_verification)
{
const std::array<Tensor<InDataType>, NumAs - 1> elementwise_a_tensors = {in_bias};
const std::array<Tensor<WeiDataType>, NumBs - 1> elementwise_b_tensors = {wei_bias};
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumAs - 1,
NumBs - 1>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in,
wei,
out_host,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op,
elementwise_a_tensors,
elementwise_b_tensors);
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(out_device.mData.data());
return ck::utils::check_err(out_device, out_host, "Error: incorrect results!");
}
return true;
}
} // namespace
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ aa374621
...@@ -3,12 +3,22 @@ ...@@ -3,12 +3,22 @@
#pragma once #pragma once
#include <iostream> #include <cmath>
#include <cstdlib>
#include <numeric>
#include <type_traits> #include <type_traits>
#include <sstream> #include <vector>
#include "ck/tensor_operation/gpu/device/device_base.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
...@@ -22,6 +32,7 @@ namespace host { ...@@ -22,6 +32,7 @@ namespace host {
// Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout // Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout
// as long as dimensions in tensor descriptor is in GNCHW order // as long as dimensions in tensor descriptor is in GNCHW order
// //
// @tparam NDimSpatial Number of spatial dimensions.
// @tparam InDataType Input tensor data type. // @tparam InDataType Input tensor data type.
// @tparam WeiDataType Weights tensor data type. // @tparam WeiDataType Weights tensor data type.
// @tparam OutDataType Output tensor data type. // @tparam OutDataType Output tensor data type.
...@@ -29,7 +40,9 @@ namespace host { ...@@ -29,7 +40,9 @@ namespace host {
// operation. // operation.
// @tparam WeiElementwiseOperation Functor for weights tensor elementwise // @tparam WeiElementwiseOperation Functor for weights tensor elementwise
// operation. // operation.
// @tparam NDimSpatial Number of spatial dimensions. // @tparam NumAElementwiseTensor Number of A elementwise tensors.
// @tparam NumBElementwiseTensor Number of B elementwise tensors.
// @tparam NumDElementwiseTensor Number of D elementwise tensors.
// //
// input descriptor in [G, N, C, Do, Ho, Wo] order // input descriptor in [G, N, C, Do, Ho, Wo] order
// weight descriptor in [G, K, C, Z, Y, X] order // weight descriptor in [G, K, C, Z, Y, X] order
...@@ -42,14 +55,17 @@ template <ck::index_t NDimSpatial, ...@@ -42,14 +55,17 @@ template <ck::index_t NDimSpatial,
typename InElementwiseOperation, typename InElementwiseOperation,
typename WeiElementwiseOperation, typename WeiElementwiseOperation,
typename OutElementwiseOperation, typename OutElementwiseOperation,
ck::index_t NumDTensor = 0, ck::index_t NumAElementwiseTensor = 0,
ck::index_t NumBElementwiseTensor = 0,
ck::index_t NumDElementwiseTensor = 0,
typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false> typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
struct ReferenceConvFwd : public device::BaseOperator struct ReferenceConvFwd : public device::BaseOperator
{ {
// Argument // Argument
struct Argument : public device::BaseArgument struct Argument : public device::BaseArgument
{ {
Argument(const Tensor<InDataType>& input, Argument(
const Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight, const Tensor<WeiDataType>& weight,
Tensor<OutDataType>& output, Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::index_t> conv_filter_strides,
...@@ -59,11 +75,15 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -59,11 +75,15 @@ struct ReferenceConvFwd : public device::BaseOperator
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
const std::array<Tensor<OutDataType>, NumDTensor>& d_tensors) const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors,
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors,
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors)
: input_{input}, : input_{input},
weight_{weight}, weight_{weight},
output_{output}, output_{output},
d_tensors_{d_tensors}, elementwise_a_tensors_{elementwise_a_tensors},
elementwise_b_tensors_{elementwise_b_tensors},
elementwise_d_tensors_{elementwise_d_tensors},
conv_strides_{conv_filter_strides}, conv_strides_{conv_filter_strides},
conv_dilations_{conv_filter_dilations}, conv_dilations_{conv_filter_dilations},
in_left_pads_{input_left_pads}, in_left_pads_{input_left_pads},
...@@ -78,7 +98,9 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -78,7 +98,9 @@ struct ReferenceConvFwd : public device::BaseOperator
const Tensor<WeiDataType>& weight_; const Tensor<WeiDataType>& weight_;
Tensor<OutDataType>& output_; Tensor<OutDataType>& output_;
const std::array<Tensor<OutDataType>, NumDTensor>& d_tensors_; const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors_;
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors_;
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors_;
std::vector<index_t> conv_strides_; std::vector<index_t> conv_strides_;
std::vector<index_t> conv_dilations_; std::vector<index_t> conv_dilations_;
...@@ -119,42 +141,43 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -119,42 +141,43 @@ struct ReferenceConvFwd : public device::BaseOperator
if(wi >= 0 && if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3]) ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
{ {
float v_in; InDataType v_in;
float v_wei; WeiDataType v_wei;
arg.in_element_op_( ExecuteElementwiseOp(arg.in_element_op_,
v_in, ck::type_convert<float>(arg.input_(g, n, c, wi))); arg.elementwise_a_tensors_,
Number<NumAElementwiseTensor>{},
arg.wei_element_op_( v_in,
v_wei, ck::type_convert<float>(arg.weight_(g, k, c, x))); arg.input_(g, n, c, wi),
g,
v_acc += v_in * v_wei; n,
c,
wi);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
arg.weight_(g, k, c, x),
g,
k,
c,
x);
v_acc +=
ck::type_convert<float>(v_in) * ck::type_convert<float>(v_wei);
} }
} }
} }
OutDataType v_out;
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc); OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
if constexpr(NumDTensor == 0) OutDataType& v_out = arg.output_(g, n, k, wo);
{ ExecuteElementwiseOp(arg.out_element_op_,
arg.out_element_op_(v_out, v_acc_converted); arg.elementwise_d_tensors_,
} Number<NumDElementwiseTensor>{},
else if constexpr(NumDTensor == 1) v_out,
{
arg.out_element_op_(v_out, v_acc_converted, arg.d_tensors_[0](g, n, k, wo));
}
else if constexpr(NumDTensor == 2)
{
arg.out_element_op_(v_out,
v_acc_converted, v_acc_converted,
arg.d_tensors_[0](g, n, k, wo), g,
arg.d_tensors_[1](g, n, k, wo)); n,
} k,
else wo);
{
throw std::runtime_error("Output ElementOp not supported in reference.");
}
arg.output_(g, n, k, wo) = v_out;
}; };
make_ParallelTensorFunctor(func, make_ParallelTensorFunctor(func,
...@@ -191,44 +214,47 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -191,44 +214,47 @@ struct ReferenceConvFwd : public device::BaseOperator
wi >= 0 && wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4]) ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
{ {
float v_in; InDataType v_in;
float v_wei; WeiDataType v_wei;
arg.in_element_op_( ExecuteElementwiseOp(arg.in_element_op_,
v_in, ck::type_convert<float>(arg.input_(g, n, c, hi, wi))); arg.elementwise_a_tensors_,
Number<NumAElementwiseTensor>{},
arg.wei_element_op_( v_in,
v_wei, ck::type_convert<float>(arg.weight_(g, k, c, y, x))); arg.input_(g, n, c, hi, wi),
g,
v_acc += v_in * v_wei; n,
c,
hi,
wi);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
arg.weight_(g, k, c, y, x),
g,
k,
c,
y,
x);
v_acc += ck::type_convert<float>(v_in) *
ck::type_convert<float>(v_wei);
} }
} }
} }
} }
OutDataType v_out;
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc); OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
if constexpr(NumDTensor == 0) OutDataType& v_out = arg.output_(g, n, k, ho, wo);
{ ExecuteElementwiseOp(arg.out_element_op_,
arg.out_element_op_(v_out, v_acc_converted); arg.elementwise_d_tensors_,
} Number<NumDElementwiseTensor>{},
else if constexpr(NumDTensor == 1) v_out,
{
arg.out_element_op_(
v_out, v_acc_converted, arg.d_tensors_[0](g, n, k, ho, wo));
}
else if constexpr(NumDTensor == 2)
{
arg.out_element_op_(v_out,
v_acc_converted, v_acc_converted,
arg.d_tensors_[0](g, n, k, ho, wo), g,
arg.d_tensors_[1](g, n, k, ho, wo)); n,
} k,
else ho,
{ wo);
throw std::runtime_error("Output ElementOp not supported in reference.");
}
arg.output_(g, n, k, ho, wo) = v_out;
}; };
make_ParallelTensorFunctor(func, make_ParallelTensorFunctor(func,
...@@ -275,47 +301,51 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -275,47 +301,51 @@ struct ReferenceConvFwd : public device::BaseOperator
ck::type_convert<std::size_t>(wi) < ck::type_convert<std::size_t>(wi) <
arg.input_.GetLengths()[5]) arg.input_.GetLengths()[5])
{ {
float v_in; InDataType v_in;
float v_wei; WeiDataType v_wei;
arg.in_element_op_(v_in, ExecuteElementwiseOp(arg.in_element_op_,
ck::type_convert<float>( arg.elementwise_a_tensors_,
arg.input_(g, n, c, di, hi, wi))); Number<NumAElementwiseTensor>{},
v_in,
arg.wei_element_op_( arg.input_(g, n, c, di, hi, wi),
g,
n,
c,
di,
hi,
wi);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei, v_wei,
ck::type_convert<float>(arg.weight_(g, k, c, z, y, x))); arg.weight_(g, k, c, z, y, x),
g,
v_acc += v_in * v_wei; k,
c,
z,
y,
x);
v_acc += ck::type_convert<float>(v_in) *
ck::type_convert<float>(v_wei);
} }
} }
} }
} }
} }
OutDataType v_out;
OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc); OutDataType v_acc_converted = ck::type_convert<OutDataType>(v_acc);
if constexpr(NumDTensor == 0) OutDataType& v_out = arg.output_(g, n, k, d_o, ho, wo);
{ ExecuteElementwiseOp(arg.out_element_op_,
arg.out_element_op_(v_out, v_acc_converted); arg.elementwise_d_tensors_,
} Number<NumDElementwiseTensor>{},
else if constexpr(NumDTensor == 1) v_out,
{
arg.out_element_op_(
v_out, v_acc_converted, arg.d_tensors_[0](g, n, k, d_o, ho, wo));
}
else if constexpr(NumDTensor == 2)
{
arg.out_element_op_(v_out,
v_acc_converted, v_acc_converted,
arg.d_tensors_[0](g, n, k, d_o, ho, wo), g,
arg.d_tensors_[1](g, n, k, d_o, ho, wo)); n,
} k,
else d_o,
{ ho,
throw std::runtime_error("Output ElementOp not supported in reference."); wo);
}
arg.output_(g, n, k, d_o, ho, wo) = v_out;
}; };
make_ParallelTensorFunctor(func, make_ParallelTensorFunctor(func,
...@@ -338,6 +368,36 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -338,6 +368,36 @@ struct ReferenceConvFwd : public device::BaseOperator
} }
}; };
template <typename... Args,
typename ElementwiseOp,
typename ElementwiseTensor,
typename NumTensor,
typename T>
static void ExecuteElementwiseOp(ElementwiseOp& elementwise_op,
ElementwiseTensor& elementwise_tensors,
NumTensor,
T& y,
const T& x,
Args... dims)
{
if constexpr(NumTensor::value == 0)
{
elementwise_op(y, x);
}
else if constexpr(NumTensor::value == 1)
{
elementwise_op(y, x, elementwise_tensors[0](dims...));
}
else if constexpr(NumTensor::value == 2)
{
elementwise_op(y, x, elementwise_tensors[0](dims...), elementwise_tensors[1](dims...));
}
else
{
throw std::runtime_error("ElementOp not supported in reference.");
}
}
static constexpr bool IsValidCompilationParameter() static constexpr bool IsValidCompilationParameter()
{ {
// TODO: properly implement this check // TODO: properly implement this check
...@@ -349,7 +409,8 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -349,7 +409,8 @@ struct ReferenceConvFwd : public device::BaseOperator
return NDimSpatial >= 1 && NDimSpatial <= 3; return NDimSpatial >= 1 && NDimSpatial <= 3;
} }
static auto MakeArgument(const Tensor<InDataType>& input, static auto MakeArgument(
const Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight, const Tensor<WeiDataType>& weight,
Tensor<OutDataType>& output, Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides, std::vector<ck::index_t> conv_filter_strides,
...@@ -359,7 +420,9 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -359,7 +420,9 @@ struct ReferenceConvFwd : public device::BaseOperator
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
const std::array<Tensor<OutDataType>, NumDTensor>& d_tensors = {}) const std::array<Tensor<OutDataType>, NumAElementwiseTensor>& elementwise_a_tensors = {},
const std::array<Tensor<OutDataType>, NumBElementwiseTensor>& elementwise_b_tensors = {},
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors = {})
{ {
return Argument{input, return Argument{input,
weight, weight,
...@@ -371,7 +434,9 @@ struct ReferenceConvFwd : public device::BaseOperator ...@@ -371,7 +434,9 @@ struct ReferenceConvFwd : public device::BaseOperator
in_element_op, in_element_op,
wei_element_op, wei_element_op,
out_element_op, out_element_op,
d_tensors}; elementwise_a_tensors,
elementwise_b_tensors,
elementwise_d_tensors};
} }
static auto MakeInvoker() { return Invoker{}; } static auto MakeInvoker() { return Invoker{}; }
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/CMakeLists.txt 0 → 100644
View file @ aa374621
set(GROUPED_CONV3D_FWD_SCALEADD_AB
xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp)
add_instance_library(device_grouped_conv3d_fwd_scaleadd_ab_instance ${GROUPED_CONV3D_FWD_SCALEADD_AB})
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