Re: [FFmpeg-devel] [PATCH] swscale/x86/rgb2rgb: add SSE2 shuffle_bytes functions

[Andreas Rheinhardt <andreas.rheinhardt@outlook.com>]

James Almer:
> On 6/6/2024 11:48 AM, Andreas Rheinhardt wrote:
>> James Almer:
>>> And remove shuffle_bytes_2103_mmxext.
>>>
>>> shuffle_bytes_0321_c: 28.1
>>> shuffle_bytes_0321_sse2: 13.6
>>> shuffle_bytes_0321_ssse3: 9.6
>>> shuffle_bytes_0321_avx2: 7.1
>>> shuffle_bytes_1230_c: 52.6
>>> shuffle_bytes_1230_sse2: 12.1
>>> shuffle_bytes_1230_ssse3: 8.6
>>> shuffle_bytes_1230_avx2: 6.6
>>> shuffle_bytes_2103_c: 29.1
>>> shuffle_bytes_2103_mmxext: 29.3 // removed
>>> shuffle_bytes_2103_sse2: 12.5
>>> shuffle_bytes_2103_ssse3: 8.6
>>> shuffle_bytes_2103_avx2: 7.1
>>> shuffle_bytes_3012_c: 52.1
>>> shuffle_bytes_3012_sse2: 12.1
>>> shuffle_bytes_3012_ssse3: 8.6
>>> shuffle_bytes_3012_avx2: 7.1
>>> shuffle_bytes_3210_c: 50.6
>>> shuffle_bytes_3210_sse2: 14.6
>>> shuffle_bytes_3210_ssse3: 8.6
>>> shuffle_bytes_3210_avx2: 7.1
>>>
>>> Signed-off-by: James Almer <jamrial@gmail.com>
>>> ---
>>>   libswscale/x86/rgb2rgb.c     | 14 ++++--
>>>   libswscale/x86/rgb_2_rgb.asm | 83 +++++++++++++++++++++++++-----------
>>>   2 files changed, 69 insertions(+), 28 deletions(-)
>>>
>>> diff --git a/libswscale/x86/rgb2rgb.c b/libswscale/x86/rgb2rgb.c
>>> index 21ccfafe51..9f6c8efc72 100644
>>> --- a/libswscale/x86/rgb2rgb.c
>>> +++ b/libswscale/x86/rgb2rgb.c
>>> @@ -116,7 +116,11 @@ DECLARE_ALIGNED(8, extern const uint64_t,
>>> ff_bgr2UVOffset);
>>>     #endif /* HAVE_INLINE_ASM */
>>>   -void ff_shuffle_bytes_2103_mmxext(const uint8_t *src, uint8_t
>>> *dst, int src_size);
>>> +void ff_shuffle_bytes_2103_sse2(const uint8_t *src, uint8_t *dst,
>>> int src_size);
>>> +void ff_shuffle_bytes_0321_sse2(const uint8_t *src, uint8_t *dst,
>>> int src_size);
>>> +void ff_shuffle_bytes_1230_sse2(const uint8_t *src, uint8_t *dst,
>>> int src_size);
>>> +void ff_shuffle_bytes_3012_sse2(const uint8_t *src, uint8_t *dst,
>>> int src_size);
>>> +void ff_shuffle_bytes_3210_sse2(const uint8_t *src, uint8_t *dst,
>>> int src_size);
>>>   void ff_shuffle_bytes_2103_ssse3(const uint8_t *src, uint8_t *dst,
>>> int src_size);
>>>   void ff_shuffle_bytes_0321_ssse3(const uint8_t *src, uint8_t *dst,
>>> int src_size);
>>>   void ff_shuffle_bytes_1230_ssse3(const uint8_t *src, uint8_t *dst,
>>> int src_size);
>>> @@ -154,10 +158,12 @@ av_cold void rgb2rgb_init_x86(void)
>>>           rgb2rgb_init_avx();
>>>   #endif /* HAVE_INLINE_ASM */
>>>   -    if (EXTERNAL_MMXEXT(cpu_flags)) {
>>> -        shuffle_bytes_2103 = ff_shuffle_bytes_2103_mmxext;
>>> -    }
>>>       if (EXTERNAL_SSE2(cpu_flags)) {
>>> +        shuffle_bytes_2103 = ff_shuffle_bytes_2103_sse2;
>>> +        shuffle_bytes_0321 = ff_shuffle_bytes_0321_sse2;
>>> +        shuffle_bytes_1230 = ff_shuffle_bytes_1230_sse2;
>>> +        shuffle_bytes_3012 = ff_shuffle_bytes_3012_sse2;
>>> +        shuffle_bytes_3210 = ff_shuffle_bytes_3210_sse2;
>>>   #if ARCH_X86_64
>>>           uyvytoyuv422 = ff_uyvytoyuv422_sse2;
>>>   #endif
>>> diff --git a/libswscale/x86/rgb_2_rgb.asm b/libswscale/x86/rgb_2_rgb.asm
>>> index 0bf1278718..9fc1974389 100644
>>> --- a/libswscale/x86/rgb_2_rgb.asm
>>> +++ b/libswscale/x86/rgb_2_rgb.asm
>>> @@ -25,7 +25,6 @@
>>>     SECTION_RODATA
>>>   -pb_mask_shuffle2103_mmx times 8 dw 255
>>>   pb_shuffle2103: db 2, 1, 0, 3, 6, 5, 4, 7, 10, 9, 8, 11, 14, 13,
>>> 12, 15
>>>   pb_shuffle0321: db 0, 3, 2, 1, 4, 7, 6, 5, 8, 11, 10, 9, 12, 15,
>>> 14, 13
>>>   pb_shuffle1230: db 1, 2, 3, 0, 5, 6, 7, 4, 9, 10, 11, 8, 13, 14,
>>> 15, 12
>>> @@ -50,11 +49,50 @@ SECTION .text
>>>  
>>> ;------------------------------------------------------------------------------
>>>   ; shuffle_bytes_2103_mmext (const uint8_t *src, uint8_t *dst, int
>>> src_size)
>>>  
>>> ;------------------------------------------------------------------------------
>>> -INIT_MMX mmxext
>>> -cglobal shuffle_bytes_2103, 3, 5, 8, src, dst, w, tmp, x
>>> -    mova   m6, [pb_mask_shuffle2103_mmx]
>>> -    mova   m7, m6
>>> -    psllq  m7, 8
>>> +
>>> +%macro SHUFFLE2103_SSE2 0
>>> +    pshuflw   m1, m0, 0xb1
>>> +    pshufhw   m1, m1, 0xb1
>>> +
>>> +    pand     m0, m3
>>> +    pand     m1, m2
>>> +%endmacro
>>> +
>>> +%macro SHUFFLE0321_SSE2 0
>>> +    pshuflw   m1, m0, 0xb1
>>> +    pshufhw   m1, m1, 0xb1
>>> +
>>> +    pand     m0, m2
>>> +    pand     m1, m3
>>> +%endmacro
>>> +
>>> +%macro SHUFFLE1230_SSE2 0
>>> +    pslld    m1, m0, 24
>>> +    psrld    m0, 8
>>> +%endmacro
>>> +
>>> +%macro SHUFFLE3012_SSE2 0
>>> +    pslld    m1, m0, 8
>>> +    psrld    m0, 24
>>> +%endmacro
>>> +
>>> +%macro SHUFFLE3210_SSE2 0
>>> +    pshuflw   m1, m0, 0xb1
>>> +    pshufhw   m1, m1, 0xb1
>>> +
>>> +    psrlw     m0, m1, 8
>>> +    psllw     m1, 8
>>> +%endmacro
>>> +
>>> +; %1-4 index shuffle
>>> +; %5 load mask
>>> +%macro SHUFFLE_BYTES_SSE2 5
>>> +cglobal shuffle_bytes_%1%2%3%4, 3, 5, 4, src, dst, w, tmp, x
>>> +%if %5
>>> +    pcmpeqw        m2, m2
>>> +    psllw          m3, m2, 8 ; (word) { 0xff00 } x4
>>> +    psrlw          m2, 8     ; (word) { 0x00ff } x4
>>> +%endif
>>>         movsxdifnidn wq, wd
>>>       mov xq, wq
>>> @@ -68,13 +106,13 @@ cglobal shuffle_bytes_2103, 3, 5, 8, src, dst,
>>> w, tmp, x
>>>       je .loop_simd
>>>     .loop_scalar:
>>> -   mov          tmpb, [srcq + wq + 2]
>>> +   mov          tmpb, [srcq + wq + %1]
>>>      mov [dstq+wq + 0], tmpb
>>> -   mov          tmpb, [srcq + wq + 1]
>>> +   mov          tmpb, [srcq + wq + %2]
>>>      mov [dstq+wq + 1], tmpb
>>> -   mov          tmpb, [srcq + wq + 0]
>>> +   mov          tmpb, [srcq + wq + %3]
>>>      mov [dstq+wq + 2], tmpb
>>> -   mov          tmpb, [srcq + wq + 3]
>>> +   mov          tmpb, [srcq + wq + %4]
>>>      mov [dstq+wq + 3], tmpb
>>>      add            wq, 4
>>>      sub            xq, 4
>>> @@ -86,29 +124,26 @@ jge .end
>>>     .loop_simd:
>>>       movu     m0, [srcq+wq]
>>> -    movu     m1, [srcq+wq+8]
>>> -
>>> -    pshufw   m3, m0, 177
>>> -    pshufw   m5, m1, 177
>>> -
>>> -    pand     m0, m7
>>> -    pand     m3, m6
>>>   -    pand     m1, m7
>>> -    pand     m5, m6
>>> +    SHUFFLE%1%2%3%4_SSE2
>>>   -    por      m0, m3
>>> -    por      m1, m5
>>> +    por      m0, m1
>>>         movu      [dstq+wq], m0
>>> -    movu  [dstq+wq + 8], m1
>>>   -    add              wq, mmsize*2
>>> +    add              wq, mmsize
>>>       jl .loop_simd
>>>     .end:
>>> -    emms
>>>       RET
>>> +%endmacro
>>> +
>>> +INIT_XMM sse2
>>> +SHUFFLE_BYTES_SSE2 2, 1, 0, 3, 1
>>> +SHUFFLE_BYTES_SSE2 0, 3, 2, 1, 1
>>> +SHUFFLE_BYTES_SSE2 1, 2, 3, 0, 0
>>> +SHUFFLE_BYTES_SSE2 3, 0, 1, 2, 0
>>> +SHUFFLE_BYTES_SSE2 3, 2, 1, 0, 0
>>>    
>>> ;------------------------------------------------------------------------------
>>>   ; shuffle_bytes_## (const uint8_t *src, uint8_t *dst, int src_size)
>>
>> How old are the youngest processors with SSE2, but without SSSE3?
> 
> AMD Phenom/K10.
> 
>> According to Wikipedia, nearly 15 years. Which makes me believe that the
>> SSE2 versions are not worth it (how many of these CPUs will use a new
>> FFmpeg anyway?).
> 
> Simply by using the latest version of a video player that uses ffmpeg is
> enough to be able to run the newest code.

I asked "how many", not "how".

> It was easy to write and i don't feel particularly interested enough to
> argue, so if you think it's not worth adding, i can just remove the
> mmxext version and skip adding anything.

I think we should not optimize for CPUs that do not even have x86-64 v2.
So I would not add these SSE2 versions. But the one missing SSSE3
version (shuffle_bytes_2103_ssse3) is of course worth it.

- Andreas

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".