[bernhardmgruber]

We previously only aligned to alignof(T) and not alignof(T) * vec_size. This wasn't noticed since the compiler just keeps the data in registers anyway. However, in a debug build, the arrays are spilled to local memory so the alignment matters.

Small SASS changes for cub.test.device.transform.lid_0 on sm86 and sm120 but only a few immediates:

<         /*00b0*/                   IMAD.MOV.U32 R8, RZ, RZ, 0x3e2 ;                           /* 0x000003e2ff087424 */
---
>         /*00b0*/                   IMAD.MOV.U32 R8, RZ, RZ, 0x3e6 ;                           /* 0x000003e6ff087424 */
29401c29401
<         /*0220*/                   IMAD.MOV.U32 R8, RZ, RZ, 0x230 ;                           /* 0x00000230ff087424 */
---
>         /*0220*/                   IMAD.MOV.U32 R8, RZ, RZ, 0x234 ;                           /* 0x00000234ff087424 */
29477c29477
<         /*0480*/                   IMAD.MOV.U32 R8, RZ, RZ, 0x213 ;                           /* 0x00000213ff087424 */
---
>         /*0480*/                   IMAD.MOV.U32 R8, RZ, RZ, 0x217 ;                           /* 0x00000217ff087424 */

I am not worried, given we also need the fix for correctness.

Fixes: #7160

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 21h 24m: Pass: 100%/128 | Total: 6d 20h | Max: 5h 00m | Hits: 57%/176785

See results here.

[github-actions]

Successfully created backport PR for branch/3.2.x:

• [Backport branch/3.2.x] Align local vector storage arrays in vec transform #7163