Identifier "FLT_MAX" is undefined

[danijel3]

Not sure if something hasn't changed recently, but seems something isn't letting cudadecoder compile properly:

/usr/local/cuda/bin/nvcc -c cuda-decoder-kernels.cu -o cuda-decoder-kernels.o -I/usr/local/cuda/include -I/opt/kaldi/tools/cub-1.8.0 -I.. -isystem /opt/kaldi/tools/openfst-1.7.2/include --compiler-options -fPIC --machine 64 -DHAVE_CUDA -ccbin c++ -DKALDI_DOUBLEPRECISION=0 -std=c++14 -DCUDA_API_PER_THREAD_DEFAULT_STREAM -lineinfo --verbose -Wno-deprecated-gpu-targets -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90 -I../ -I/opt/kaldi/tools/openfst-1.7.2/include
#$ _NVVM_BRANCH_=nvvm
#$ _SPACE_=
#$ _CUDART_=cudart
#$ _HERE_=/usr/local/cuda/bin
#$ _THERE_=/usr/local/cuda/bin
#$ _TARGET_SIZE_=
#$ _TARGET_DIR_=
#$ _TARGET_DIR_=targets/x86_64-linux
#$ TOP=/usr/local/cuda/bin/..
#$ CICC_PATH=/usr/local/cuda/bin/../nvvm/bin
#$ CICC_NEXT_PATH=/usr/local/cuda/bin/../nvvm-next/bin
#$ NVVMIR_LIBRARY_DIR=/usr/local/cuda/bin/../nvvm/libdevice
#$ LD_LIBRARY_PATH=/usr/local/cuda/bin/../lib:
#$ PATH=/usr/local/cuda/bin/../nvvm/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin
#$ INCLUDES="-I/usr/local/cuda/bin/../targets/x86_64-linux/include"
#$ LIBRARIES=  "-L/usr/local/cuda/bin/../targets/x86_64-linux/lib/stubs" "-L/usr/local/cuda/bin/../targets/x86_64-linux/lib"
#$ CUDAFE_FLAGS=
#$ PTXAS_FLAGS=
#$ c++ -std=c++14 -D__CUDA_ARCH_LIST__=500,520,600,610,700,750,800,860,900 -D__NV_LEGACY_LAUNCH -E -x c++ -D__CUDACC__ -D__NVCC__  -fPIC -I"/usr/local/cuda/include" -I"/opt/kaldi/tools/cub-1.8.0" -I".." -I"../" -I"/opt/kaldi/tools/openfst-1.7.2/include" "-I/usr/local/cuda/bin/../targets/x86_64-linux/include"   -isystem "/opt/kaldi/tools/openfst-1.7.2/include"  -D "HAVE_CUDA" -D "KALDI_DOUBLEPRECISION=0" -D "CUDA_API_PER_THREAD_DEFAULT_STREAM" -D__CUDACC_VER_MAJOR__=12 -D__CUDACC_VER_MINOR__=5 -D__CUDACC_VER_BUILD__=40 -D__CUDA_API_VER_MAJOR__=12 -D__CUDA_API_VER_MINOR__=5 -D__NVCC_DIAG_PRAGMA_SUPPORT__=1 -include "cuda_runtime.h" -m64 "cuda-decoder-kernels.cu" -o "/tmp/tmpxft_00006ccd_00000000-5_cuda-decoder-kernels.cpp4.ii"
#$ cudafe++ --c++14 --gnu_version=110400 --display_error_number --orig_src_file_name "cuda-decoder-kernels.cu" --orig_src_path_name "/opt/kaldi/src/cudadecoder/cuda-decoder-kernels.cu" --allow_managed  --m64 --parse_templates --gen_c_file_name "/tmp/tmpxft_00006ccd_00000000-6_cuda-decoder-kernels.compute_90.cudafe1.cpp" --stub_file_name "tmpxft_00006ccd_00000000-6_cuda-decoder-kernels.compute_90.cudafe1.stub.c" --gen_module_id_file --module_id_file_name "/tmp/tmpxft_00006ccd_00000000-4_cuda-decoder-kernels.module_id" "/tmp/tmpxft_00006ccd_00000000-5_cuda-decoder-kernels.cpp4.ii"
cuda-decoder-kernels.cu(539): error: identifier "FLT_MAX" is undefined
      CostType total_cost = FLT_MAX;

My version of NVCC:

$ /usr/local/cuda/bin/nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2024 NVIDIA Corporation
Built on Wed_Apr_17_19:19:55_PDT_2024
Cuda compilation tools, release 12.5, V12.5.40
Build cuda_12.5.r12.5/compiler.34177558_0

And g++: g++ (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0

[danijel3]

My quick and dirty solution was to temporarily add:

const float FLT_MAX=3.402823466e+38F;

at the start of cudadecoder/cuda-decoder-kernels.cu, but it would nice to know what caused it. Is this something only on my Ubuntu machine or are there others facing the same problem?

[jtrmal]

inluding might solve it another way -- not sure what has changed, maybe you have a new release of cuda when they changed strategy of includes? y.

On Thu, May 23, 2024 at 1:57 PM Danijel Koržinek @.***> wrote:

My quick and dirty solution was to temporarily add:

const float FLT_MAX=3.402823466e+38F;

at the start of cudadecoder/cuda-decoder-kernels.cu, but it would nice to know what caused it. Is this something only on my Ubuntu machine or are there others facing the same problem?

— Reply to this email directly, view it on GitHub https://github.com/kaldi-asr/kaldi/issues/4914#issuecomment-2126927252, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACUKYX7ZEWMBT6H4BRSUNBLZDXKRVAVCNFSM6AAAAABIFO7M2CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMRWHEZDOMRVGI . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[danpovey]

Yeah we probably committed the sin of relying on things that are indirectly included. @danijel3 is there any chance you could commit a fix? It looks like we are supposed to

include

... I think your approach would cause a syntax error if float.h were included, as it's done via a define statement.

[danijel3]

limits.h unfortunately doesn't help and float.h is already included, but doesn't work for some reason. I'll keep digging at it. It may be an installation issue or something, but I can at least figure it out in case it happens to someone else...

[danpovey]

It would be fine to just replace FLT_MAX in the code with that constant, and put // FLT_MAX as a comment.

[danijel3]

I confirmed the issue using a blank Ubuntu Docker image. From what I can see, it is related only to CUDA SDK (so the cu files) and not the standard C++ libs. The install uses the latest (as of now) version of the SDK from the NVidia repo on the Ubuntu 22.04:

deb [signed-by=/usr/share/keyrings/cuda-archive-keyring.gpg] https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/ /

The PR above is harmless and hopefully sufficient. I've checked and it does resolve the issue.