compilation of extension field in MSM is too slow

[cliff0412]

No description provided.

[doutv]

Build time when enable G2 MSM: 300s

cmake -B build -DCMAKE_BUILD_TYPE=Release -DCURVE=BN254 -DG2_ENABLED=ON -DBUILD_MSM=ON
time cmake --build build
cmake --build build  302.83s user 7.47s system 99% cpu 5:10.47 total

Look into details

/usr/local/cuda/bin/nvcc --time - --verbose -forward-unknown-to-host-compiler -DFEATURE_BN254 -DG2_ENABLED -DBUILD_MSM -DEXPOSE_C_INTERFACE -DTAKE_RESPONSIBILITY_FOR_ERROR_MESSAGE -DUSE_CUDA -D__ADX__ --options-file CMakeFiles/cryptocudacuda.dir/includes_CUDA.rsp -O3 -DNDEBUG -std=c++17 "--generate-code=arch=compute_89,code=[compute_89,sm_89]" -Xcompiler=-fPIC -dc -MD -MT CMakeFiles/cryptocudacuda.dir/msm/msm.cu.o -MF CMakeFiles/cryptocudacuda.dir/msm/msm.cu.o.d -x cu -rdc=true -c /home/okxdex/data/zkdex-pap/workspace/jason-huang/cryptography_cuda/native/msm/msm.cu -o CMakeFiles/cryptocudacuda.dir/msm/msm.cu.o

#$ cicc --c++17 --gnu_version=110400 --display_error_number --orig_src_file_name "/home/okxdex/data/zkdex-pap/workspace/jason-huang/cryptography_cuda/native/msm/msm.cu" --orig_src_path_name "/home/okxdex/data/zkdex-pap/workspace/jason-huang/cryptography_cuda/native/msm/msm.cu" --allow_managed  --device-c   -arch compute_89 -m64 --no-version-ident -ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 --include_file_name "tmpxft_000f2741_00000000-3_msm.fatbin.c" -tused --module_id_file_name "/tmp/tmpxft_000f2741_00000000-4_msm.module_id" --gen_c_file_name "/tmp/tmpxft_000f2741_00000000-6_msm.cudafe1.c" --stub_file_name "/tmp/tmpxft_000f2741_00000000-6_msm.cudafe1.stub.c" --gen_device_file_name "/tmp/tmpxft_000f2741_00000000-6_msm.cudafe1.gpu"  "/tmp/tmpxft_000f2741_00000000-7_msm.cpp1.ii" -o "/tmp/tmpxft_000f2741_00000000-6_msm.ptx"
/home/okxdex/data/zkdex-pap/workspace/jason-huang/cryptography_cuda/native/msm/msm.cu , cicc ,  /tmp/tmpxft_000f2741_00000000-7_msm.cpp1.ii  , /tmp/tmpxft_000f2741_00000000-6_msm.ptx , compute_89 , nvcc , 260733.2656 , ms
#$ ptxas -arch=sm_89 -m64 --compile-only  "/tmp/tmpxft_000f2741_00000000-6_msm.ptx"  -o "/tmp/tmpxft_000f2741_00000000-8_msm.sm_89.cubin" 
/home/okxdex/data/zkdex-pap/workspace/jason-huang/cryptography_cuda/native/msm/msm.cu , ptxas , /tmp/tmpxft_000f2741_00000000-6_msm.ptx  , /tmp/tmpxft_000f2741_00000000-8_msm.sm_89.cubin , sm_89 , nvcc , 14485.4258 , ms

[doutv]

After delete __forceinline__ in extension_field.cuh, the compile time became much faster
i.e. do not inline functions related to G2 extension field
cicc: 260s -> 54s
However, I'm not sure whether it would affect the runtime speed.

#$ cicc --c++17 --gnu_version=110400 --display_error_number --orig_src_file_name "/home/okxdex/data/zkdex-pap/workspace/jason-huang/cryptography_cuda/native/msm/msm.cu" --orig_src_path_name "/home/okxdex/data/zkdex-pap/workspace/jason-huang/cryptography_cuda/native/msm/msm.cu" --allow_managed  --device-c   -arch compute_60 -m64 --no-version-ident -ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 --include_file_name "tmpxft_000f5135_00000000-3_msm.fatbin.c" -tused --module_id_file_name "/tmp/tmpxft_000f5135_00000000-4_msm.module_id" --gen_c_file_name "/tmp/tmpxft_000f5135_00000000-6_msm.cudafe1.c" --stub_file_name "/tmp/tmpxft_000f5135_00000000-6_msm.cudafe1.stub.c" --gen_device_file_name "/tmp/tmpxft_000f5135_00000000-6_msm.cudafe1.gpu"  "/tmp/tmpxft_000f5135_00000000-7_msm.cpp1.ii" -o "/tmp/tmpxft_000f5135_00000000-6_msm.ptx"
/home/okxdex/data/zkdex-pap/workspace/jason-huang/cryptography_cuda/native/msm/msm.cu , cicc ,  /tmp/tmpxft_000f5135_00000000-7_msm.cpp1.ii  , /tmp/tmpxft_000f5135_00000000-6_msm.ptx , compute_60 , nvcc , 54353.8086 , ms
#$ ptxas -arch=sm_60 -m64 --compile-only  "/tmp/tmpxft_000f5135_00000000-6_msm.ptx"  -o "/tmp/tmpxft_000f5135_00000000-8_msm.sm_60.cubin" 
/home/okxdex/data/zkdex-pap/workspace/jason-huang/cryptography_cuda/native/msm/msm.cu , ptxas , /tmp/tmpxft_000f5135_00000000-6_msm.ptx  , /tmp/tmpxft_000f5135_00000000-8_msm.sm_60.cubin , sm_60 , nvcc , 18578.5273 , ms