[FFmpeg-devel] [PATCH 4/5] lavc/aarch64: new optimization for 8-bit hevc_epel_h

[Logan.Lyu]

Hi, Martin,

I modified it according to your comments. Please review again.

And here are the checkasm benchmark results of the related functions:

put_hevc_epel_h4_8_c: 67.1
put_hevc_epel_h4_8_i8mm: 21.1
put_hevc_epel_h6_8_c: 147.1
put_hevc_epel_h6_8_i8mm: 45.1
put_hevc_epel_h8_8_c: 237.4
put_hevc_epel_h8_8_i8mm: 72.1
put_hevc_epel_h12_8_c: 527.4
put_hevc_epel_h12_8_i8mm: 115.4
put_hevc_epel_h16_8_c: 943.6
put_hevc_epel_h16_8_i8mm: 153.9
put_hevc_epel_h24_8_c: 2105.4
put_hevc_epel_h24_8_i8mm: 384.4
put_hevc_epel_h32_8_c: 3631.4
put_hevc_epel_h32_8_i8mm: 519.9
put_hevc_epel_h48_8_c: 8082.1
put_hevc_epel_h48_8_i8mm: 1110.4
put_hevc_epel_h64_8_c: 14400.6
put_hevc_epel_h64_8_i8mm: 2057.1

put_hevc_qpel_h4_8_c: 124.9
put_hevc_qpel_h4_8_neon: 43.1
put_hevc_qpel_h4_8_i8mm: 33.1
put_hevc_qpel_h6_8_c: 269.4
put_hevc_qpel_h6_8_neon: 90.6
put_hevc_qpel_h6_8_i8mm: 61.4
put_hevc_qpel_h8_8_c: 477.6
put_hevc_qpel_h8_8_neon: 82.1
put_hevc_qpel_h8_8_i8mm: 99.9
put_hevc_qpel_h12_8_c: 1062.4
put_hevc_qpel_h12_8_neon: 226.9
put_hevc_qpel_h12_8_i8mm: 170.9
put_hevc_qpel_h16_8_c: 1880.6
put_hevc_qpel_h16_8_neon: 302.9
put_hevc_qpel_h16_8_i8mm: 251.4
put_hevc_qpel_h24_8_c: 4221.9
put_hevc_qpel_h24_8_neon: 893.9
put_hevc_qpel_h24_8_i8mm: 626.1
put_hevc_qpel_h32_8_c: 7437.6
put_hevc_qpel_h32_8_neon: 1189.9
put_hevc_qpel_h32_8_i8mm: 959.1
put_hevc_qpel_h48_8_c: 16838.4
put_hevc_qpel_h48_8_neon: 2727.9
put_hevc_qpel_h48_8_i8mm: 2163.9
put_hevc_qpel_h64_8_c: 29982.1
put_hevc_qpel_h64_8_neon: 4777.6


在 2023/6/12 16:12, Martin Storsjö 写道:
> On Sun, 4 Jun 2023, Logan.Lyu at myais.com.cn wrote:
>
>> From: Logan Lyu <Logan.Lyu at myais.com.cn>
>>
>> Signed-off-by: Logan Lyu <Logan.Lyu at myais.com.cn>
>> ---
>> libavcodec/aarch64/hevcdsp_epel_neon.S    | 343 ++++++++++++++++++++++
>> libavcodec/aarch64/hevcdsp_init_aarch64.c |   7 +-
>> 2 files changed, 349 insertions(+), 1 deletion(-)
>
>
>> +        st2             {v20.8h, v21.8h}, [x7]
>> +        subs            w3, w3, #1   // height
>> +        b.ne            1b
>> +        ret
>
> In general, place the loop counter decrement somewhere else than 
> exactly before the branch that depends on the result. E.g. after the 
> initial loads is usually a good place, or between the st1/2 
> instructions and the instructions that calculate the final output values.
>
> The same goes probably for all places in all these patches.
>
>> @@ -283,13 +287,14 @@ av_cold void 
>> ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth)
>>         NEON8_FNASSIGN_PARTIAL_4(c->put_hevc_qpel_uni_w, 1, 0, 
>> qpel_uni_w_v,);
>>
>>         if (have_i8mm(cpu_flags)) {
>> +            NEON8_FNASSIGN(c->put_hevc_epel, 0, 1, epel_h, _i8mm);
>>             NEON8_FNASSIGN(c->put_hevc_epel_uni_w, 0, 1, epel_uni_w_h 
>> ,_i8mm);
>>             NEON8_FNASSIGN(c->put_hevc_qpel, 0, 1, qpel_h, _i8mm);
>>             NEON8_FNASSIGN(c->put_hevc_qpel_uni_w, 0, 1, 
>> qpel_uni_w_h, _i8mm);
>>             NEON8_FNASSIGN_PARTIAL_5(c->put_hevc_qpel_uni_w, 1, 1, 
>> qpel_uni_w_hv, _i8mm);
>>         }
>> -
>>     }
>> +
>>     if (bit_depth == 10) {
>
> Here are some stray unrelated whitespace changes.
>
> Other than that, this patch looks mostly reasonable.
>
> // Martin
>