[Bug Fix] Support casting lightweight float formats to complex types

[youge325]

PR Category

Environment Adaptation

PR Types

Bug fixes

Description

• 加入了 CastDataTypeFunctor 对 float8_e4m3fn、float8_e5m2、bfloat16、float16 到 complex64/complex128 的专门实现；这些组合之前没有合法的 static_cast，会在模版实例化时报错或导致转换缺失。编译时的错误日志如下：

[283/3505] Building CXX object paddle\phi\CMakeFiles\phi.dir\core\framework\data_type_transform.cc.obj
FAILED: paddle/phi/CMakeFiles/phi.dir/core/framework/data_type_transform.cc.obj
"D:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\bin\Hostx64\x64\cl.exe"  /nologo /TP -DCUDA_TOOLKIT_ROOT_DIR="\"C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.0\"" -DCUDA_VERSION_MAJOR=\"13\" -DCUDA_VERSION_MINOR=\"0\" -DCUDNN_MAJOR_VERSION=\"9\" -DEIGEN_STRONG_INLINE=inline -DEIGEN_USE_GPU -DGLOG_NO_ABBREVIATED_SEVERITIES -DGOOGLE_GLOG_DLL_DECL="" -DLAPACK_FOUND -DNOMINMAX -DPADDLE_DISABLE_PROFILER -DPADDLE_DLL_EXPORT -DPADDLE_DLL_INFERENCE -DPADDLE_ON_INFERENCE -DPADDLE_VERSION=0.0.0 -DPADDLE_VERSION_INTEGER=0 -DPADDLE_WITH_AVX -DPADDLE_WITH_CCCL -DPADDLE_WITH_CRYPTO -DPADDLE_WITH_CUDA -DPADDLE_WITH_DNNL -DPADDLE_WITH_MKLML -DPADDLE_WITH_PIP_CUDA_LIBRARIES -DPADDLE_WITH_POCKETFFT -DPADDLE_WITH_SSE3 -DPADDLE_WITH_TENSORRT -DPHI_INNER -DPHI_SHARED -DSTATIC_IR -DTRT_PLUGIN_FP16_AVAILABLE -DUTF8PROC_STATIC -DWIN32_LEAN_AND_MEAN -DYAML_CPP_STATIC_DEFINE -D_XKEYCHECK_H -Dphi_EXPORTS -ID:\Paddle\third_party\cccl\thrust -ID:\Paddle\third_party\cccl\libcudacxx\include -ID:\Paddle\third_party\cccl\cub -ID:\Paddle\build -ID:\Paddle\paddle\fluid\framework\io -ID:\Paddle\patches\thrust -ID:\TensorRT-10.13.3.9\include -ID:\Paddle\build\third_party\install\zlib\include -ID:\Paddle\build\third_party\install -ID:\Paddle\build\third_party\install\gflags\include -ID:\Paddle\build\third_party\install\glog\include -ID:\Paddle\third_party\eigen3 -ID:\Paddle\third_party\threadpool -ID:\Paddle\third_party\dlpack\include -ID:\Paddle\build\third_party\install\xxhash\include -ID:\Paddle\build\third_party\install\warpctc\include -ID:\Paddle\build\third_party\install\warprnnt\include -ID:\Paddle\build\third_party\install\utf8proc\include -ID:\Paddle\build\third_party\install\mklml\include -ID:\Paddle\build\third_party\install\onednn\include -ID:\Paddle\build\third_party\install\protobuf\include -ID:\Paddle\third_party\nlohmann_json\include -ID:\Paddle\build\third_party\install\yaml-cpp\include -ID:\Users\Lenovo\AppData\Local\Programs\Python\Python310\include -ID:\Users\Lenovo\AppData\Local\Programs\Python\Python310\Lib\site-packages\numpy\core\include -ID:\Paddle\build\third_party\pybind\src\extern_pybind\include -ID:\Paddle\build\third_party\install\libuv\include -ID:\Paddle\third_party\cccl -ID:\Paddle\build\third_party\install\cryptopp\include -ID:\Paddle\build\third_party\pocketfft\src -ID:\Paddle\build\third_party\dirent\src\extern_dirent\include -I"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0\include" -ID:\Paddle -ID:\Paddle\paddle\phi\api\include\compat -ID:\Paddle\paddle\phi\api\include\compat\torch\csrc\api\include -ID:\Paddle\build\..\paddle\fluid\framework\io /DWIN32 /D_WINDOWS  /W0  /GR /EHsc /w /wd4068 /wd4129 /wd4244 /wd4267 /wd4297 /wd4530 /wd4577 /wd4819 /wd4838  /arch:AVX /MT /O2 /Ob2 /DNDEBUG /bigobj /Zc:inline -std:c++17 /showIncludes /Fopaddle\phi\CMakeFiles\phi.dir\core\framework\data_type_transform.cc.obj /Fdpaddle\phi\CMakeFiles\phi.dir\ /FS -c D:\Paddle\paddle\phi\core\framework\data_type_transform.cc
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(33): error C2440: 'static_cast': cannot convert from 'InType' to 'OutType'
        with
        [
            InType=phi::dtype::float8_e5m2
        ]
        and
        [
            OutType=phi::dtype::complex<double>
        ]
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(33): note: 'phi::dtype::complex<double>::complex': no overloaded function could convert all the argument types
D:\Paddle\paddle/phi/common/complex.h(127): note: could be 'phi::dtype::complex<double>::complex<T>(const phi::dtype::complex<float> &)'
        with
        [
            T=double
        ]
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(33): note: 'phi::dtype::complex<double>::complex<T>(const phi::dtype::complex<float> &)': cannot convert argument 1 from 'InType' to 'const phi::dtype::complex<float> &'
        with
        [
            T=double
        ]
        and
        [
            InType=phi::dtype::float8_e5m2
        ]
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(33): note: Reason: cannot convert from 'InType' to 'const phi::dtype::complex<float>'
        with
        [
            InType=phi::dtype::float8_e5m2
        ]
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(33): note: No user-defined-conversion operator available that can perform this conversion, or the operator cannot be called
D:\Paddle\paddle/phi/common/complex.h(140): note: or       'phi::dtype::complex<double>::complex(const std::complex<_Other> &)'
D:\Paddle\paddle/phi/common/complex.h(119): note: or       'phi::dtype::complex<double>::complex(const std::enable_if<std::is_same<T1,float>::value,phi::dtype::complex<double>>::type &)'
D:\Paddle\paddle/phi/common/complex.h(113): note: or       'phi::dtype::complex<double>::complex(const T1 &)'
D:\Paddle\paddle/phi/common/complex.h(79): note: or       'phi::dtype::complex<double>::complex(const cuda::std::__4::complex<_Xp> &)'
D:\Paddle\paddle/phi/common/complex.h(72): note: or       'phi::dtype::complex<double>::complex(const thrust::THRUST_200200___CUDA_ARCH_LIST___NS::complex<U> &)'
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(33): note: while trying to match the argument list '(InType)'
        with
        [
            InType=phi::dtype::float8_e5m2
        ]
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(33): note: the template instantiation context (the oldest one first) is
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(207): note: see reference to function template instantiation 'void phi::VisitDataType<phi::CastDataType<phi::dtype::float8_e5m2>>(paddle::framework::proto::VarType::Type,Visitor)' being compiled
        with
        [
            Visitor=phi::CastDataType<phi::dtype::float8_e5m2>
        ]
D:\Paddle\paddle/phi/core/framework/var_type_helper.h(156): note: see reference to function template instantiation 'void phi::CastDataType<phi::dtype::float8_e5m2>::apply<phi::dtype::complex<double>>(void)' being compiled
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(106): note: see reference to class template instantiation 'phi::CastDataTypeFunctor<InType,OutType>' being compiled
        with
        [
            InType=phi::dtype::float8_e5m2,
            OutType=phi::dtype::complex<double>
        ]
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(32): note: while compiling class template member function 'OutType phi::CastDataTypeFunctor<InType,OutType>::operator ()(InType) const'
        with
        [
            OutType=phi::dtype::complex<double>,
            InType=phi::dtype::float8_e5m2
        ]
D:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\include\algorithm(3799): note: see the first reference to 'phi::CastDataTypeFunctor<InType,OutType>::operator ()' in 'std::transform'
        with
        [
            InType=phi::dtype::float8_e5m2,
            OutType=phi::dtype::complex<double>
        ]
D:\Paddle\paddle/phi/common/transform.h(78): note: see the first reference to 'std::transform' in 'phi::Transform<phi::CPUContext>::operator ()'
D:\Paddle\paddle\phi\core\framework\data_type_transform.cc(102): note: see the first reference to 'phi::Transform<phi::CPUContext>::operator ()' in 'phi::CastDataType<phi::dtype::float8_e5m2>::apply'
D:\Paddle\paddle/phi/core/framework/var_type_helper.h(156): note: see the first reference to 'phi::CastDataType<phi::dtype::float8_e5m2>::apply' in 'phi::VisitDataType'
[290/3505] Building CXX object paddle\phi\CMakeFiles\phi.dir\core\vocab\string_array.cc.obj
ninja: build stopped: subcommand failed.

• issue: paddle.utils.run_check()报错 #75848
• 拆分自 fix bugs when compiling with VS2022 #75611
• 前置pr [Bug Fix] Fix CastKernel for low-precision to complex type conversions #75930

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[youge325]

/re-run all-failed

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 16 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@3a94489). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
paddle/phi/core/framework/data_type_transform.cc	0.00%	16 Missing ⚠️

❌ Your patch status has failed because the patch coverage (0.00%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #75929   +/-   ##
==========================================
  Coverage           ?    0.00%           
==========================================
  Files              ?        1           
  Lines              ?       16           
  Branches           ?        0           
==========================================
  Hits               ?        0           
  Misses             ?       16           
  Partials           ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[youge325]

/re-run all-failed

[youge325]

/re-run all-failed

[youge325]

/re-run all-failed

[youge325]

/re-run all-failed

[youge325]

/re-run all-failed

[luotao1]

单侧覆盖率不够，需要补充单侧。#75929 (comment)

[youge325]

单侧覆盖率不够，需要补充单侧。#75929 (comment)

明白了，之前强推的含单测的提交 6aae326 流水线没过，可能是因为我之前有个对cast_impl.h修改的pr #75930 没有合入，导致单测没有生效，我这次再试试

[youge325]

/re-run all-failed

[youge325]

#76185 已经解决该问题