[x264-devel] [PATCH 11/24] arm: Implement neon 8x16c intra predict functions

[Martin Storsjö]

On Wed, 19 Aug 2015, Janne Grunau wrote:

> On 2015-08-13 23:59:32 +0300, Martin Storsjö wrote:
>> This implements the same functions as are implemented for 8x8c
>> and as for 8x16c on aarch64.
>>
>> Some of the simpler ones actually turn out to be slower than the
>> plain C version, at least on some CPUs.
>
> See 'arm64: optimize various intra_predict asm functions'
> (<1439822360-17282-1-git-send-email-janne-x264 at jannau.net>)
>
> That makes all intra_predict functions at least as fast as the C version
> on a cortex-a53 in arm64 mode.
>
>> checkasm timing       Cortex-A7      A8     A9
>> intra_predict_8x16c_dc_c     1347    910    1017
>> intra_predict_8x16c_dc_neon  1271    1366   1247
>> intra_predict_8x16c_dcl_c    859     677    692
>> intra_predict_8x16c_dcl_neon 1006    1209   1065
>> intra_predict_8x16c_dct_c    871     540    590
>> intra_predict_8x16c_dct_neon 672     511    657
>> intra_predict_8x16c_h_c      937     712    719
>> intra_predict_8x16c_h_neon   722     682    672
>> intra_predict_8x16c_p_c      10184   9967   8652
>> intra_predict_8x16c_p_neon   2617    1973   1983
>> intra_predict_8x16c_v_c      610     380    429
>> intra_predict_8x16c_v_neon   570     513    507
>> ---
>>  common/arm/predict-a.S |  158 ++++++++++++++++++++++++++++++++++++++++++++++++
>>  common/arm/predict-c.c |   15 +++++
>>  common/arm/predict.h   |    8 +++
>>  common/predict.c       |    4 ++
>>  4 files changed, 185 insertions(+)
>>
>> diff --git a/common/arm/predict-a.S b/common/arm/predict-a.S
>> index 7e5d9d3..228fd2e 100644
>> --- a/common/arm/predict-a.S
>> +++ b/common/arm/predict-a.S
>> @@ -5,6 +5,7 @@
>>   *
>>   * Authors: David Conrad <lessen42 at gmail.com>
>>   *          Mans Rullgard <mans at mansr.com>
>> + *          Martin Storsjo <martin at martin.st>
>>   *
>>   * This program is free software; you can redistribute it and/or modify
>>   * it under the terms of the GNU General Public License as published by
>> @@ -552,6 +553,163 @@ function x264_predict_8x8c_p_neon
>>  endfunc
>>
>>
>> +function x264_predict_8x16c_dc_top_neon
>> +    sub         r2,  r0,  #FDEC_STRIDE
>> +    mov         r1,  #FDEC_STRIDE
>> +    vld1.8      {d0}, [r2,:64]
>> +    vpaddl.u8   d0,  d0
>> +    vpadd.u16   d0,  d0,  d0
>> +    vrshrn.u16  d0,  q0,  #2
>> +    vdup.8      d1,  d0[1]
>> +    vdup.8      d0,  d0[0]
>> +    vtrn.32     d0,  d1
>
> vmov d1, d0

Hmm, I'm not quite sure what you suggest here, should I replace the vtrn 
with vmov d1, d0, or use it instead of the vmov q1, q0 below? (After this 
vtrn, d1 and d0 should be identical, right?)

>
>> +    vmov        q1,  q0
>> +    b           pred8x16_dc_end
>
> since we need every cycle to it probably makes sense to avoid the branch
> and vmov

So copypaste (or macroize) pred8x16_dc_end here, and make it use the same 
input register for all writes to avoid the vmov q1, q0?

I'll try your other suggestions below and see what I can squeeze out of 
it.

// Martin