ZLUDA v3.8.8

[lshqqytiger]

Change logs

• Update README.
• Add missing cudaDeviceSynchronize and cudaGetErrorString for cuda runtime api.
• Fix bug of cuda runtime api.
• [Nightly] Add nvcuda.zluda_get_nightly_flag.
  It returns 1 if ZLUDA was built with --nightly flag. Otherwise, it returns 0.
• [Nightly] Unlock cuBLASLt matmul compute type.
• [Nightly] Experimental cuDNN support on Windows.

About cuDNN

cuDNN is required by various high-performance softwares.

Luckily, there are a few of them that allows us to disable it.

However, we cannot run others that does not. (e.g. ONNX Runtime)

You'll need nightly ZLUDA and a very specific build of MIOpen.

Actually, I would like not to share it because it is virtually unusable and useless at this moment.

As I tested, because MIOpen itself is very unstable and incomplete, there are some limitations as described below.

• Only the custom build of MIOpen without rocMLIR and composable kernel is tested.
  Other unofficial builds will not work.
• Only FP32 is supported for Conv2d in gfx1100.
• There is small memory leak issue due to technical difficulties.
• The performance is poor due to the reasons I described above.

Despite of the issues, I'm merging this into master because the branches are diverging so much.