search

google / XNNPACK

High-efficiency floating-point neural network inference operators for mobile, server, and Web
Other
1.89k stars 376 forks source link

ARM build on gcc-13 failed with src/reference/unary-elementwise.cc:125:14: error: invalid ‘static_cast’ from type ‘xnn_bfloat16’ to type ‘_Float16’ #7489

Open xwang233 opened 1 day ago

xwang233 commented 1 day ago

Original issue from pytorch: https://github.com/pytorch/pytorch/issues/141083

The XNNPACK commit is at https://github.com/google/XNNPACK/commit/4ea82e595b36106653175dcb04b2aa532660d0d8

Build error:

[3174/5315] Building CXX object confu-deps/XNNPACK/CMakeFiles/reference-ukernels.dir/src/reference/unary-elementwise.cc.o
FAILED: confu-deps/XNNPACK/CMakeFiles/reference-ukernels.dir/src/reference/unary-elementwise.cc.o
/usr/bin/c++ -DCAFFE2_PERF_WITH_SVE=1 -DXNN_ENABLE_ARM_BF16=0 -DXNN_ENABLE_ARM_DOTPROD=1 -DXNN_ENABLE_ARM_FP16_SCALAR=1 -DXNN_ENABLE_ARM_FP16_VECTOR=1 -DXNN_ENABLE_ARM_I8MM=0 -DXNN_ENABLE_ARM_SME2=1 -DXNN_ENABLE_ARM_SME=1 -DXNN_ENABLE_ASSEMBLY=1 -DXNN_ENABLE_AVX256SKX=1 -DXNN_ENABLE_AVX256VNNI=1 -DXNN_ENABLE_AVX256VNNIGFNI=1 -DXNN_ENABLE_AVX512AMX=1 -DXNN_ENABLE_AVX512F=1 -DXNN_ENABLE_AVX512FP16=1 -DXNN_ENABLE_AVX512SKX=1 -DXNN_ENABLE_AVX512VBMI=1 -DXNN_ENABLE_AVX512VNNI=1 -DXNN_ENABLE_AVX512VNNIGFNI=1 -DXNN_ENABLE_AVXVNNI=0 -DXNN_ENABLE_AVXVNNIINT8=0 -DXNN_ENABLE_CPUINFO=1 -DXNN_ENABLE_DWCONV_MULTIPASS=0 -DXNN_ENABLE_GEMM_M_SPECIALIZATION=1 -DXNN_ENABLE_HVX=1 -DXNN_ENABLE_KLEIDIAI=0 -DXNN_ENABLE_MEMOPT=1 -DXNN_ENABLE_RISCV_VECTOR=1 -DXNN_ENABLE_SPARSE=1 -DXNN_ENABLE_VSX=1 -I/opt/pytorch/pytorch/third_party/XNNPACK/include -I/opt/pytorch/pytorch/third_party/XNNPACK/src -I/opt/pytorch/pytorch/third_party/pthreadpool/include -isystem /opt/pytorch/pytorch/third_party/protobuf/src -DNO_CUDNN_DESTROY_HANDLE -fno-gnu-unique -D_GLIBCXX_USE_CXX11_ABI=1 -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -O3 -DNDEBUG -std=c++14 -fPIC -Wno-psabi -O2 -pthread -MD -MT confu-deps/XNNPACK/CMakeFiles/reference-ukernels.dir/src/reference/unary-elementwise.cc.o -MF confu-deps/XNNPACK/CMakeFiles/reference-ukernels.dir/src/reference/unary-elementwise.cc.o.d -o confu-deps/XNNPACK/CMakeFiles/reference-ukernels.dir/src/reference/unary-elementwise.cc.o -c /opt/pytorch/pytorch/third_party/XNNPACK/src/reference/unary-elementwise.cc
/opt/pytorch/pytorch/third_party/XNNPACK/src/reference/unary-elementwise.cc: In instantiation of ‘TOut {anonymous}::ConvertOp<TIn, TOut>::operator()(TIn) const [with TIn = xnn_bfloat16; TOut = _Float16]’:
/opt/pytorch/pytorch/third_party/XNNPACK/src/reference/unary-elementwise.cc:39:32:   required from ‘void {anonymous}::unary_ukernel_unquantized(size_t, const TIn*, TOut*, const xnn_unary_uparams*) [with TIn = xnn_bfloat16; TOut = _Float16; Operator = ConvertOp<xnn_bfloat16, _Float16>; size_t = long unsigned int]’
/opt/pytorch/pytorch/third_party/XNNPACK/src/reference/unary-elementwise.cc:167:7:   required from ‘const xnn_unary_elementwise_config* {anonymous}::get_convert_config(std::false_type, std::false_type) [with TIn = xnn_bfloat16; TOut = _Float16; std::false_type = std::integral_constant<bool, false>]’
/opt/pytorch/pytorch/third_party/XNNPACK/src/reference/unary-elementwise.cc:183:50:   required from ‘const xnn_unary_elementwise_config* {anonymous}::get_convert_config(xnn_datatype, InputQuantized) [with TIn = xnn_bfloat16; InputQuantized = std::integral_constant<bool, false>]’
/opt/pytorch/pytorch/third_party/XNNPACK/src/reference/unary-elementwise.cc:217:46:   required from here
/opt/pytorch/pytorch/third_party/XNNPACK/src/reference/unary-elementwise.cc:125:14: error: invalid ‘static_cast’ from type ‘xnn_bfloat16’ to type ‘_Float16’
  125 |       return static_cast<TOut>(x);
      |              ^~~~~~~~~~~~~~~~~~~~

Environment:

CPU is neoverse-v2, or Nvidia Grace

It was reported that gcc-11 builds ok in the original pytorch issue.

steven-johnson commented 1 day ago

Investigating.