[FFmpeg-devel] [PATCH] lavc/aarch64: new optimization for 8-bit hevc_pel_uni_w_pixels, qpel_uni_w_h, qpel_uni_w_v, qpel_uni_w_hv and qpel_h

[Martin Storsjö]

Hi,

Overall these patches seem mostly ok, but I've got a few minor points to 
make:

- The usdot instruction requires the i8mm extension (part of armv8.6-a), 
while udot or sdot would require the dotprod extension (available in 
armv8.4-a). If you could manage with udot or sdot, these functions would 
be usable on a wider set of CPUs.

Therefore, the current guards are wrong. Also, I finally got support 
implemented for optionally using these cpu extensions, even if the 
baseline of the compile don't include it, by runtime enabling it. See the 
patchset at https://patchwork.ffmpeg.org/project/ffmpeg/list/?series=9009.

To adapt your patches on top of this, see the two topmost commits at 
https://github.com/mstorsjo/ffmpeg/commits/archext.

- The indentation is inconsistent; in the first patch, you have some 
instructions written like this:

+        sqadd   v1.4s, v1.4s, v29.4s

While you later use this style:

+        dup             v1.16b, v28.b[1]

The latter seems to match the style we commonly use; please reformat your 
code to match that consistently.

With some macro invocations in the first patch, you also seem to have too 
much indentation in some places. See e.g. this:

+1:      ldr             q23, [x2, x3]
+        add             x2, x2, x3, lsl #1
+        QPEL_FILTER_B     v26, v16, v17, v18, v19, v20, v21, v22, v23
+        QPEL_FILTER_B2    v27, v16, v17, v18, v19, v20, v21, v22, v23
+        QPEL_UNI_W_V_16
+        subs            w4, w4, #1
+        b.eq            2f

(If the macro name is too long, that's ok, but here there's no need to 
have those lines unaligned.)

- In the third patch, you've got multiple parameters from the stack like 
this:

+        ldp             x14, x15, [sp]          // mx, my
+        ldr             w13, [sp, #16]          // width

I see that the mx an my parameters are intptr_t; that's good, since if 
they would be 32 bit integers, the ABI for such parameters on the stack 
differ between macOS/Darwin and Linux. But as long as they're intptr_t 
they behave the same.

- At the same place, you're backing up a bunch of registers:

+        stp             x20, x21, [sp, #-16]!
+        stp             x22, x23, [sp, #-16]!
+        stp             x24, x25, [sp, #-16]!
+        stp             x26, x27, [sp, #-16]!
+        stp             x28, x30, [sp, #-16]!

This is inefficient; instead, do this:

+        stp             x28, x30, [sp, #-80]!
+        stp             x20, x21, [sp, #16]
+        stp             x22, x23, [sp, #32]
+        stp             x24, x25, [sp, #48]
+        stp             x26, x27, [sp, #64]

Also, following that, I see that you back up the stack pointer in x28. Why 
do you use x28 for that? Using x29 would be customary as frame pointer.

Aside for that, I think the rest of the patches is acceptable.

// Martin