Closed rosethompson closed 1 month ago
I am making notes here so I don't lose this data. I am currently unable to replicate this branch predictor issue. I ran the branch embench benchmarks on commits from dec 18, 2023, jan 18, 2024, and feb 19, 2024 and there is only a tiny change in the accuracy between dec and jan < 0.01%. My notes show I original found the issue in early feb 2024. I am wondering if the issue is compiler related. The programs I ran are compiled using the latest gcc 13.2.0. I used the embench source code checked into the repo from dec 18, 2023 for each of the three simulations.
@ross144 if you wanted to test older compiler versions, there are downloadable, compiled binaries (not with quite the same multilib setup but should work) listed by date archived here: https://github.com/riscv-collab/riscv-gnu-toolchain/releases.
How about we write this off as a fluke because it is working now, close the issue, and move on to the hotter issues. We can return to these notes if we see it resurface.
On Aug 19, 2024, at 7:47 PM, Rose Thompson @.***> wrote:
I am making notes here so I don't lose this data. I am currently unable to replicate this branch predictor issue. I ran the branch embench benchmarks on commits from dec 18, 2023, jan 18, 2024, and feb 19, 2024 and there is only a tiny change in the accuracy between dec and jan < 0.01%. My notes show I original found the issue in early feb 2024. I am wondering if the issue is compiler related. The programs I ran are compiled using the latest gcc 13.2.0. I used the embench source code checked into the repo from dec 18, 2023 for each of the three simulations.
— Reply to this email directly, view it on GitHub https://github.com/openhwgroup/cvw/issues/768#issuecomment-2297860375, or unsubscribe https://github.com/notifications/unsubscribe-auth/AR4AA34KFDFK7Q4OE7QSER3ZSKU3FAVCNFSM6AAAAABHEBKIH2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDEOJXHA3DAMZXGU. You are receiving this because you are subscribed to this thread.
Changes since Dec 18, 2023 result in lower branch predictor accuracy by 1 to 2%. We must resolve before TLR5 is completed.