One more benchmark added to try to achieve hardware(T4 GPU used currently) fp32 peak performance in examples/bench directory.

---

TuringAs

An open source SASS assembler for NVIDIA Volta, Turing, and Ampere GPUs.

Requirements:

• Python >= 3.6

Supported hardware:

All NVIDIA Volta (SM70), Turing (SM75), and Ampere (SM80) GPUs.

Other features:

• Include files.
• Inline python code.

Install the library

python setup.py install

Use the library

python -m turingas.main -i <input.sass> -o <output.cubin> -arch <arch>

# E.g.
python -m turingas.main -i input.sass -o output.cubin -arch 75 # 75 for Turing
                                                               # or 70 (Volta), 80 (Ampere)

Related projects

AsFermi, an SASS assembler for NVIDIA Fermi GPUs. By Hou Yunqing.

MaxAs, an SASS assembler for NVIDIA Maxwell and Pascal. By Scott Gray.

KeplerAs, an SASS assembler for NVIDIA Kepler. By Xiuxia Zhang.

TODO list

To support following instructions:

• Type conversion instructions (I2I, I2F, F2I, F2F)
• MUFU # Multifunction. E.g., sin, cos.
• LDSM # Load matrix from shared memory.
• Texture instructions
• Surface instructions
• Unified data path instructions
• Other (ISCADD, CALL, JMP ...)
• New Ampere instructions.

Citation

If you find this tool helpful, please cite:

@inproceedings{yan2020winograd-conv,
  author    = {Da Yan and
               Wei Wang and
               Xiaowen Chu},
  title     = {Optimizing Batched Winograd Convolution on GPUs},
  booktitle = {25th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP '20)},
  year      = {2020},
  address   = {San Diego, CA, USA},
  publisher = {ACM},
}

This project is released under the MIT License.

-- Da Yan