[ET-VK] Removing tile input storage variable in conv_pw op and fetching the data in main loop. Also unrolling the main loop for performance improvement.

[trivedivivek]

Stack from ghstack (oldest at bottom):

• [ET-VK] Reduced int precision for texture coordinates in conv2d_pw op, to reduce shader register pressure. #6468
• -> [ET-VK] Removing tile input storage variable in conv_pw op and fetching the data in main loop. Also unrolling the main loop for performance improvement. #6467
• [ET-VK] Reduced int precision for texture coordinates in conv2d_pw op, to reduce shader register pressure and slightly improve performance. #6466
• [ET-VK] Reduced int precision for all int storage in q linear op, and reducing some texture coordinate storage variables to improve performance. #6465

This diff removes the tile input storage array in_tex in the conv_pw op and fetches the data in the main loop for performance improvement.
The main loop has also been unrolled for performance improvement.

Differential Revision: D64767314

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6467

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• [PRE-EMPTIVE] Experimenting with new runners linux.aws.a100 on inductor-perf-compare.yml

✅ No Failures

As of commit c695c67 with merge base fa30e80 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[facebook-github-bot]

This pull request was exported from Phabricator. Differential Revision: D64767314