[FFmpeg-devel] [PATCH] riscv: vc1dsp: Don't check vlenb before checking the CPU flags

[Martin Storsjö]

On Fri, 15 Dec 2023, Rémi Denis-Courmont wrote:

> Le 15 décembre 2023 17:39:48 GMT+02:00, "Martin Storsjö" <martin at martin.st> a écrit :
>> On Fri, 15 Dec 2023, Rémi Denis-Courmont wrote:
>>
>>> Le 15 décembre 2023 15:02:04 GMT+02:00, "Martin Storsjö" <martin at martin.st> a écrit :
>>>> We can't call ff_get_rv_vlenb() if we don't have RVV available
>>>> at all.
>>>> 
>>>> Due to the SIGILL signal handler in checkasm catching it, in an
>>>> unexpected place, this caused checkasm to hang instead of reporting
>>>> the issue.
>>>> ---
>>>> libavcodec/riscv/vc1dsp_init.c | 16 +++++++---------
>>>> 1 file changed, 7 insertions(+), 9 deletions(-)
>>>> 
>>>> diff --git a/libavcodec/riscv/vc1dsp_init.c b/libavcodec/riscv/vc1dsp_init.c
>>>> index 0d22d28f4d..2bb7e7fe8f 100644
>>>> --- a/libavcodec/riscv/vc1dsp_init.c
>>>> +++ b/libavcodec/riscv/vc1dsp_init.c
>>>> @@ -35,15 +35,13 @@ av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp)
>>>> #if HAVE_RVV
>>>>     int flags = av_get_cpu_flags();
>>>> 
>>>> -    if (ff_get_rv_vlenb() >= 16) {
>>>> -        if (flags & AV_CPU_FLAG_RVV_I64) {
>>>> -            dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv;
>>>> -            dsp->vc1_inv_trans_8x4_dc = ff_vc1_inv_trans_8x4_dc_rvv;
>>>> -        }
>>>> -        if (flags & AV_CPU_FLAG_RVV_I32) {
>>>> -            dsp->vc1_inv_trans_4x8_dc = ff_vc1_inv_trans_4x8_dc_rvv;
>>>> -            dsp->vc1_inv_trans_4x4_dc = ff_vc1_inv_trans_4x4_dc_rvv;
>>>> -        }
>>>> +    if (flags & AV_CPU_FLAG_RVV_I64 && ff_get_rv_vlenb() >= 16) {
>>>> +        dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv;
>>>> +        dsp->vc1_inv_trans_8x4_dc = ff_vc1_inv_trans_8x4_dc_rvv;
>>>> +    }
>>>> +    if (flags & AV_CPU_FLAG_RVV_I32 && ff_get_rv_vlenb() >= 16) {
>>>> +        dsp->vc1_inv_trans_4x8_dc = ff_vc1_inv_trans_4x8_dc_rvv;
>>>> +        dsp->vc1_inv_trans_4x4_dc = ff_vc1_inv_trans_4x4_dc_rvv;
>>> 
>>> I64 implies I32 so it is not necessary to check vlenb twice. That's what I was going for originally in my then review comments but then woopsie.
>>
>> Sure, fixed.
>>
>> FWIW I see that vc1_inv_trans_8x4_dc_rvv_i64 seems to fail the checkasm test most of the time as well.
>
> Hmm, I didn't write those optimisations but I thought I tested them 
> before pushing. Is this subtly dependent on the vector length, maybe? 
> Currently only 128-bit hardware is commercially available but QEMU can 
> also emulate 256, 512 and 1014.

Ah, yes, it succeeds with 128 bit vectors, but fails with 256 bit.

// Martin