[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 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.

// Martin