[FFmpeg-devel] [PATCH] lavc/ac3: add R-V Zbb extract_exponents

[Rémi Denis-Courmont]

Le 3 décembre 2023 19:50:18 GMT+02:00, Zhao Zhili <quinkblack at foxmail.com> a écrit :
>
>
>> On Oct 3, 2023, at 00:47, Rémi Denis-Courmont <remi at remlab.net> wrote:
>> 
>> 
>> diff --git a/libavcodec/riscv/ac3dsp_rvb.S b/libavcodec/riscv/ac3dsp_rvb.S
>> new file mode 100644
>> index 0000000000..48f8bb101e
>> --- /dev/null
>> +++ b/libavcodec/riscv/ac3dsp_rvb.S
>> 
>> +func ff_extract_exponents_rvb, zbb
>> +1:
>> +        lw       t0, (a1)
>> +        addi     a0, a0, 1
>> +        neg      t1, t0
>> +        addi     a1, a1, 4
>> +        max      t0, t0, t1
>> +        addi     a2, a2, -1
>> +        clz      t0, t0
>> +        addi     t0, t0, 24 - __riscv_xlen
>> +        sb       t0, -1(a0)
>> +        bgtz    a2, 1b
>> +
>> +        ret
>> +endfunc
>> —
>
>Got build failure with clang 14:
>
><instantiation>:6:21: warning: unknown option, expected 'push', 'pop', 'rvc', 'norvc', 'relax' or 'norelax'
>            .option arch, +zbb
>                    ^
>src/libavcodec/riscv/ac3dsp_rvb.S:24:1: note: while in macro instantiation
>func ff_extract_exponents_rvb, zbb
>^
>src/libavcodec/riscv/ac3dsp_rvb.S:30:9: error: instruction requires the following: 'Zbb' (Basic Bit-Manipulation)
>        max t0, t0, t1
>        ^
>src/libavcodec/riscv/ac3dsp_rvb.S:32:9: error: instruction requires the following: 'Zbb' (Basic Bit-Manipulation)
>        clz t0, t0
>        ^
>make: *** [/home/quink/work/ffmpeg/ffbuild/common.mak:93: libavcodec/riscv/ac3dsp_rvb.o] Error 1
>make: *** Waiting for unfinished jobs....
><instantiation>:6:21: warning: unknown option, expected 'push', 'pop', 'rvc', 'norvc', 'relax' or 'norelax'
>            .option arch, +f
>                    ^
>src/libavcodec/riscv/audiodsp_rvf.S:23:1: note: while in macro instantiation
>func ff_vector_clipf_rvf, f
>^
>
>Someone says clang 14 has Zbb extensions support. I don’t know what’s going on.

It's not practical to support such a broken assembler as LLVM built-in's until they get their act together. You can add tests in FFmpeg configure but that's just going to turn all optimisations off. You could also disable the integrated assembler and use binutils, but then you'll hit the limitation of FFmpeg's configure whereby it tests the inline assembler rather than the outline one.

So really you're better off with GCC. RISC-V support on LLVM is pretty sad, TBH.