Why is NSight Compute "missing" my program's kernel launches?
I'm using NSight Compute to profile a program which launches some CUDA kernels. I know for certain that they are launched; but when I press the "play" button in NSight Compute, despite having enabled the profiling of all kernels - the program profiling concludes (no crash/failure) - with no kernels profiled. Why is that?
To determine why this is, don't just "play" to the end. Instead, press the "->:" button, which gets you to the next kernel launch, then perhas advance again to the next API call (with "->*").
After one of these you are likely to see some error reported under the "API Stream" listed.
Example:
In this case, we've gotten (the very common first encountered) ERR_NVGPUCTRPERM error - your user doesn't have permissions to access the GPU performance counters. You can add such permissions by following the instructions here.
- Strange errors of compiling pcl1.8 with cuda 11.3
- How do I use compiler intrinsic __fmul_?
- NVIDIA NVCC changes compile time constant when using template trait types
- Polymorphism alternatives when working with CUDA in C++
- How to avoid find_package_handle_standard_args package name warning
- What is the difference between cuda vs tensor cores?
- How to measure the inner kernel time in NVIDIA CUDA?
- Alternative reasons for "out of memory" error than lack of free global memory?
- Why is this CUDA program silently failing? (cudaMemcpyDeviceToHost always results in zeros)
- Eigen decomposition of Hermitian Matrix using CuSolver does not match the result with matlab
- How to find and link CUDA libraries using CMake 3.15 and later?
- ImportError: libcuda.so.1: cannot open shared object file
- Trying to 'Make' CUDA SDK, ld cannot find library, ldconfig says it can
- In CUDA, what is memory coalescing, and how is it achieved?
- Is it possible to run CUDA on AMD GPUs?
- CUDA process sharing memory(in GPU) among two separate CUDA process
- Can not find GPU devices in a data center node
- Why does transform_reduce result differ from the result of transform & reduce?
- How to get the format of nvdec video frame from ffmpeg decoding ? How to use it in cuda structures?
- Does __shfl_sync in CUDA always operate on registers, or does it involve shared memory or global memory in some certain situations?
- C++ and CUDA name mangling inconsistent
- Speed up nested loops and bitwise operations with Alea GPU
- Allocating CUDA Unified Memory by overloading operator new causes illegal access
- NVML Header File Missing
- Is it possible to execute multiple instances of a CUDA program on a multi-GPU machine?
- Cuda Large Block Number Causes Garbage Printf() Values
- Is there any way that I can convert my cpu parallel code to the cuda?
- nvidia-smi Failed to initialize NVML: GPU access blocked by the operating system
- undefined reference to `cuCtxGetCurrent` while getting CUDA context for OptiX
- What's the capacity of a CUDA stream (=queue)?