SASS inspection kit for the NVIDIA RTX 5060 Laptop GPU

The .sass files are read-only cuobjdump listings. The corresponding CUDA C++
and PTX files are the editable inputs. The captures in this kit were made with
CUDA Toolkit 13.3 for sm_120. Other compilers or target GPUs can allocate
different registers and emit different instruction sequences.

On Windows, open x64 Developer PowerShell with the CUDA Toolkit tools on PATH,
change into this directory, and build one kernel:

    ptxas -v -arch=sm_120 increment.ptx -o increment.cubin
    cuobjdump --dump-sass increment.cubin
    nvdisasm increment.cubin
    nvcc -std=c++17 run.cpp -lcuda -o run.exe
    .\run.exe increment.cubin increment

The runner verifies nine lengths against CPU values, including guard words.
Use the same recipe with pair.ptx / pair, then mirror.ptx / mirror. For the
one-warp Tensor Core example, use:

    nvcc -arch=sm_120 mma.cu -o mma.exe
    .\mma.exe
    cuobjdump --dump-sass mma.exe

For the extra samples used in the questions:

    ptxas -arch=sm_120 reading_samples.ptx -o reading_samples.cubin
    cuobjdump --dump-sass reading_samples.cubin
    py check_reading_samples.py reading_samples.cubin

Use the compiled cubin when comparing the executed program to an offline
listing. Loading a .ptx module can exercise the driver's separate JIT path.

The browser's algorithm tests run a deliberately bounded instruction subset.
They model register and memory values under the contract of each task.
Use the real CUDA tools above to inspect hardware code and use CUDA events
or Nsight Compute for timing and profiling.
