The first 32 elements of each row were correct, while the last 16 were scrambled. This hasn't been noticed, because the checkasm test erroneously only checked half of the output (for 8 bit functions), and apparently none of the samples as part of "fate-hevc" seem to trigger this specific function.
Showing the changes to libavcodec (1 of 1 file). Show all
Unified · Split
| @@ -1572,6 +1572,7 @@ function ff_hevc_put_hevc_epel_h48_8_neon_i8mm, export=1 · 1571 unmodified lines | |||
| 1572 | xtn2 v22.8h, v26.4s | 1572 | xtn2 v22.8h, v26.4s |
| 1573 | xtn v23.4h, v23.4s | 1573 | xtn v23.4h, v23.4s |
| 1574 | xtn2 v23.8h, v27.4s | 1574 | xtn2 v23.8h, v27.4s |
| 1575 | add x7, x0, #64 | ||
| 1575 | st4 {v20.8h, v21.8h, v22.8h, v23.8h}, [x0], x10 | 1576 | st4 {v20.8h, v21.8h, v22.8h, v23.8h}, [x0], x10 |
| 1576 | ext v4.16b, v2.16b, v3.16b, #1 | 1577 | ext v4.16b, v2.16b, v3.16b, #1 |
| 1577 | ext v5.16b, v2.16b, v3.16b, #2 | 1578 | ext v5.16b, v2.16b, v3.16b, #2 |
| @@ -1584,11 +1585,14 @@ function ff_hevc_put_hevc_epel_h48_8_neon_i8mm, export=1 · 6 unmodified lines | |||
| 1584 | usdot v21.4s, v4.16b, v30.16b | 1585 | usdot v21.4s, v4.16b, v30.16b |
| 1585 | usdot v22.4s, v5.16b, v30.16b | 1586 | usdot v22.4s, v5.16b, v30.16b |
| 1586 | usdot v23.4s, v6.16b, v30.16b | 1587 | usdot v23.4s, v6.16b, v30.16b |
| 1587 | xtn v20.4h, v20.4s | 1588 | zip1 v24.4s, v20.4s, v22.4s |
| 1588 | xtn2 v20.8h, v22.4s | 1589 | zip2 v25.4s, v20.4s, v22.4s |
| 1589 | xtn v21.4h, v21.4s | 1590 | zip1 v26.4s, v21.4s, v23.4s |
| 1590 | xtn2 v21.8h, v23.4s | 1591 | zip2 v27.4s, v21.4s, v23.4s |
| 1591 | add x7, x0, #64 | 1592 | xtn v20.4h, v24.4s |
| 1593 | xtn2 v20.8h, v25.4s | ||
| 1594 | xtn v21.4h, v26.4s | ||
| 1595 | xtn2 v21.8h, v27.4s | ||
| 1592 | st2 {v20.8h, v21.8h}, [x7] | 1596 | st2 {v20.8h, v21.8h}, [x7] |
| 1593 | b.ne 1b | 1597 | b.ne 1b |
| 1594 | ret | 1598 | ret |