Built on CUDA 11.5 and run several NVIDIA architectures on 8.6 (Ampere) GPU (runtime compile of virtuals takes minutes, empty images on virtual Volta and below)

[ProbablyMartian]

I have tried to build and run all architectures as separate sd-cli files - from ones that got included (on my system) when following default CUDA build instructions - from 50 (Maxwell) to 86 (Ampere).

Disclaimer: tests done on one GPU (Ampere) + CPU (no AVX2), on one model (Z-Image-Turbo Q4/Q8) and one CUDA version (11.5), OS (Linux Mint 21 aka LM21). If your observations show different picture, please let me know.

Observations:

1. Long pause for virtuals during runtime after "build cached graph cut plan done" before 1st "mapping tensors" line - minutes (app is using one CPU core, took ~15 minutes for Volta), initially thought virtuals do not work, then got patient and they do work. Might be misleading 1st time one seems it,
   I suggest sd.cpp add some line in debug mode (or even higher) when starting to compile virtual CUDA.
2. Virtuals after compiled (on 2nd run) run with about same speed from Maxwell to Ampere, BUT Volta (70) and below produce empty white images (empty green in tae previews) instead of prompted.
   So hopes of forward-compatibility go out the window, correct? Well, at least for CUDA, how to check Vulkan? I guess we have no sd.cpp builds from Maxwell days, do we?
3. I have also Maxwell and Pascal GPUs, I have run same CUDA builds on a Pascal GPU and got correct images for Maxwell and Pascal virtual builds (unlike blank squares when run on the Ampere GPU).
4. Runtime PTX complication on more modern PC took 2 min vs. 2.5 on older computer for Maxwell / Pascal. For architectures starting from Volta (Tensors?) it took ~3-4 times longer (Volta in particular longer even more).

I have made other minor observations, if you are interested in more details, let me know.

Details on testing:

1. sd.cpp git downloaded some time September 2026.
2. for f in CUDA-folder/sd-cli-*; do for (( i=1;i<=3;i++ )) do
3. I removed ~/.nv/ComputeCache between loops runs (I have not used CUDA_FORCE_PTX_JIT=1 sdi-cli cause I wanted to compare with compile time first to next compiled time runs).

P.S. related: I did build and run on CUDA 12 (LM 22) too a bit on my Ampere RTX. On LM21 - Vulkan's sd.cpp release does not work, but works on LM22. Speed of CUDA's is ~1.2x of Vulkan's, but I liked results (images) of Vulkan's produced better as same settings.