Starting with CUDA 12.4, nvcc can't deduce a template type in template function under weird conditions

Accelerated Computing CUDA CUDA NVCC Compiler
1

The following code compiles with nvcc 12.3 and older

Does not compile under nvcc 12.4 and 12.5

This issue affects the OpenSource project: morousg/cvGPUSpeedup: A faster implementation of OpenCV-CUDA that uses OpenCV objects, and more! (github.com)

Platforms: Windows 11 23H2, Ubuntu 22.04, Compiler Explorer (Compiler Explorer)

With only one of the 4 changes proposed in the code example, the code compiles. None of them should be necessary.

CODE:

// Why does this code not compile?
#include <type_traits>

// There are 4 ways of making it compile
// 1) Compile with nvcc 12.3 or older

template <bool results>
constexpr bool variable = results;

template <typename T>
constexpr bool getValue(const T& value) {
    return true;
}

struct SomeType {
    int member;
};

constexpr bool function1() {
    constexpr SomeType test{ 1 };

    // 2) Replace this return with: 
    // return variable<getValue<SomeType>(test)>;
    // 3) Replace this return with: 
    // return variable<getValue(5)>;
    return variable<getValue(test)>;
}

constexpr bool function2() {
    // 4) Replace this retur with:
    // return true;
    return variable<true>;
}

int main() {
    return 0;
}
1 Like
2

More info. This is the error message from nvcc 12.4:

#$ _NVVM_BRANCH_=nvvm
#$ _NVVM_BRANCH_SUFFIX_=
#$ _SPACE_= 
#$ _CUDART_=cudart
#$ _HERE_=/opt/compiler-explorer/cuda/12.4.1/bin
#$ _THERE_=/opt/compiler-explorer/cuda/12.4.1/bin
#$ _TARGET_SIZE_=
#$ _TARGET_DIR_=
#$ _TARGET_DIR_=targets/x86_64-linux
#$ TOP=/opt/compiler-explorer/cuda/12.4.1/bin/..
#$ NVVMIR_LIBRARY_DIR=/opt/compiler-explorer/cuda/12.4.1/bin/../nvvm/libdevice
#$ LD_LIBRARY_PATH=/opt/compiler-explorer/cuda/12.4.1/bin/../lib:/opt/compiler-explorer/cuda/12.4.1/lib:/opt/compiler-explorer/cuda/12.4.1/lib32:/opt/compiler-explorer/cuda/12.4.1/lib64
#$ PATH=/opt/compiler-explorer/cuda/12.4.1/bin/../nvvm/bin:/opt/compiler-explorer/cuda/12.4.1/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/snap/bin
#$ INCLUDES="-I/opt/compiler-explorer/cuda/12.4.1/bin/../targets/x86_64-linux/include"  
#$ LIBRARIES=  "-L/opt/compiler-explorer/cuda/12.4.1/bin/../targets/x86_64-linux/lib/stubs" "-L/opt/compiler-explorer/cuda/12.4.1/bin/../targets/x86_64-linux/lib"
#$ CUDAFE_FLAGS=
#$ PTXAS_FLAGS=
#$ "/opt/compiler-explorer/gcc-10.2.0/bin"/gcc -D__CUDA_ARCH_LIST__=890 -D__NV_LEGACY_LAUNCH -E -x c++ -D__CUDACC__ -D__NVCC__  -S "-I/opt/compiler-explorer/cuda/12.4.1/bin/../targets/x86_64-linux/include"    -D__CUDACC_VER_MAJOR__=12 -D__CUDACC_VER_MINOR__=4 -D__CUDACC_VER_BUILD__=131 -D__CUDA_API_VER_MAJOR__=12 -D__CUDA_API_VER_MINOR__=4 -D__NVCC_DIAG_PRAGMA_SUPPORT__=1 -include "cuda_runtime.h" -m64 -g "<source>" -o "/app/example.cpp4.ii" 
#$ cudafe++ --c++14 --gnu_version=100200 --display_error_number --orig_src_file_name "<source>" --orig_src_path_name "<source>" --allow_managed --keep-device-functions  --m64 --parse_templates --gen_c_file_name "<source>dafe1.cpp" --stub_file_name "example.cudafe1.stub.c" --gen_module_id_file --module_id_file_name "/app/example.module_id" "/app/example.cpp4.ii" 
#$ "/opt/compiler-explorer/gcc-10.2.0/bin"/gcc -D__CUDA_ARCH__=890 -D__CUDA_ARCH_LIST__=890 -D__NV_LEGACY_LAUNCH -E -x c++  -DCUDA_DOUBLE_MATH_FUNCTIONS -D__CUDACC__ -D__NVCC__  -S "-I/opt/compiler-explorer/cuda/12.4.1/bin/../targets/x86_64-linux/include"    -D__CUDACC_VER_MAJOR__=12 -D__CUDACC_VER_MINOR__=4 -D__CUDACC_VER_BUILD__=131 -D__CUDA_API_VER_MAJOR__=12 -D__CUDA_API_VER_MINOR__=4 -D__NVCC_DIAG_PRAGMA_SUPPORT__=1 -include "cuda_runtime.h" -m64 -g "<source>" -o "/app/example.cpp1.ii" 
#$ cicc --c++14 --gnu_version=100200 --display_error_number --orig_src_file_name "<source>" --orig_src_path_name "<source>" --allow_managed --keep-device-functions   -arch compute_89 -m64 --no-version-ident -ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 --include_file_name "example.fatbin.c" -generate-line-info -tused --module_id_file_name "/app/example.module_id" --gen_c_file_name "<source>dafe1.c" --stub_file_name "<source>dafe1.stub.c" --gen_device_file_name "<source>dafe1.gpu"  "/app/example.cpp1.ii" -o "/app/example.ptx"
#$ ptxas -arch=sm_89 -m64 -v  --generate-line-info "/app/example.ptx"  -o "<source>bin" 
ptxas info    : 0 bytes gmem
#$ fatbinary --create="/app/example.fatbin" -64 --ident="<source>" --cmdline="-v  --generate-line-info " --cicc-cmdline="-generate-line-info -ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 " "--image3=kind=elf,sm=89,file=<source>bin" --embedded-fatbin="/app/example.fatbin.c" 
#$ "/opt/compiler-explorer/gcc-10.2.0/bin"/gcc -D__CUDA_ARCH__=890 -D__CUDA_ARCH_LIST__=890 -D__NV_LEGACY_LAUNCH -c -x c++  -DCUDA_DOUBLE_MATH_FUNCTIONS -S -Wno-psabi "-I/opt/compiler-explorer/cuda/12.4.1/bin/../targets/x86_64-linux/include"   -m64 -g "<source>dafe1.cpp" -o "/app/output.s" 
<source>: In function 'constexpr bool function2()':
<source>:34:30: error: no matching function for call to 'getValue(<brace-enclosed initializer list>)'
   34 |     return variable<true>;
      |                              ^
<source>:11:1: note: candidate: 'template<class T> constexpr bool getValue(const T&)'
   11 | constexpr bool getValue(const T& value) {
      | ^~~~~~~~
<source>:11:1: note:   template argument deduction/substitution failed:
<source>:34:30: note:   couldn't deduce template parameter 'T'
   34 |     return variable<true>;
      |                              ^
<source>:34:8: error: template argument 1 is invalid
   34 |     return variable<true>;
      |        ^~~~~~~~~~~~~~~~~~~     
# --error 0x1 --
Compiler returned: 1
3

I suggest filing a bug.

1 Like
4

I already did via partners portal. Bug ID: 4717351

Should I do it also via the link you provided?

Thanks!

5

No, your bug ID 4717351 should be sufficient.

7

BUG 4717351 is resolved and verified on this case . This will be fixed in a future CUDA release .

8

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.