[FFmpeg-devel] [PATCH] faster vp6 decoding

[Sebastien Lucas]

On Wed, Feb 11, 2009 at 4:28 PM, Zuxy Meng <zuxy.meng at gmail.com> wrote:
> Hi,
>
> 2009/2/9 Jason Garrett-Glaser <darkshikari at gmail.com>:
>> +    "punpcklbw %%mm7, %%mm0\n\t"                                \
>> +    "punpcklbw %%mm7, %%mm1\n\t"                                \
>> +    "punpckhbw %%mm7, %%mm3\n\t"                                \
>> +    "punpckhbw %%mm7, %%mm4\n\t"                                \
>> +    "pmullw  0(%2), %%mm0\n\t" /* src[x-8 ] * biweight [0] */   \
>> +    "pmullw  8(%2), %%mm1\n\t" /* src[x   ] * biweight [1] */   \
>> +    "pmullw  0(%2), %%mm3\n\t" /* src[x-8 ] * biweight [0] */   \
>> +    "pmullw  8(%2), %%mm4\n\t" /* src[x   ] * biweight [1] */   \
>> +    "paddw %%mm1, %%mm0\n\t"                                    \
>> +    "paddw %%mm4, %%mm3\n\t"                                    \
>>
>> This can be done faster with pmaddubsw (SSSE3-only, but worth making
>> another version surely).
>
> Sure but that would require weights to be stored as arrays of int8_t
> instead of int16_t?

Yes and that's almost possible, but not.

>> Worthwhile if you make an SSE version.
>
> SSE2?
>
>> Works by interleaving the weights, allowing you to avoid the unpacks,
>> use only two multiplies, and avoid the adds, too, I think.  If I'm
>> right, that makes the entire thing quite a bit less than half the
>> instructions.
>
> I tried something like below and it's about 15% faster on my Pentium
> M. The speed up should be more prominent on modern CPUs with 128 bit
> FADD unit:
>
> #define DIAG4_SSE2(in1,in2,in3,in4)                             \
>    "movq "#in1"(%0), %%xmm0\n\t"                                \
>    "movq "#in2"(%0), %%xmm1\n\t"                                \
>    "punpcklbw %%xmm7, %%xmm0\n\t"                                \
>    "punpcklbw %%xmm7, %%xmm1\n\t"                                \
>    "pmullw  %%xmm4, %%xmm0\n\t" /* src[x-8 ] * biweight [0] */   \
>    "pmullw  %%xmm5, %%xmm1\n\t" /* src[x   ] * biweight [1] */   \
>    "paddw %%xmm1, %%xmm0\n\t"                                    \
>    "movq  "#in3"(%0), %%xmm1\n\t"                               \
>    "movq  "#in4"(%0), %%xmm2\n\t"                               \
>    "punpcklbw %%xmm7, %%xmm1\n\t"                                \
>    "punpcklbw %%xmm7, %%xmm2\n\t"                                \
>    "pmullw %%xmm6, %%xmm1\n\t" /* src[x+8 ] * biweight [2] */   \
>    "pmullw %%xmm3, %%xmm2\n\t" /* src[x+16] * biweight [3] */   \
>    "paddw %%xmm2, %%xmm1\n\t"                                    \
>    "paddw %%xmm1, %%xmm0\n\t"                                    \
>    "paddw _ff_diag4_round, %%xmm0\n\t" /* Add 64 */             \
>    "psrlw $7, %%xmm0\n\t"                                       \
>    "packuswb %%xmm0, %%xmm0\n\t"                                 \
>    "movq %%xmm0, (%1)\n\t"
>
> static void ff_vp6_filter_diag4_sse2(uint8_t *dst, uint8_t *src, int stride,
>                             const int16_t *h_weights,const int16_t *v_weights)
> {
>    uint8_t tmp[8*11];
>    uint8_t *t = tmp;
>    src -= stride;
>
>    asm (
>    "pxor %%xmm7, %%xmm7\n\t"
>    "movq %4, %%xmm3\n\t"
>    "pshuflw $0, %%xmm3, %%xmm4\n\t"
>    "punpcklqdq %%xmm4, %%xmm4\n\t"
>    "pshuflw $85, %%xmm3, %%xmm5\n\t"
>    "punpcklqdq %%xmm5, %%xmm5\n\t"
>    "pshuflw $170, %%xmm3, %%xmm6\n\t"
>    "punpcklqdq %%xmm6, %%xmm6\n\t"
>    "pshuflw $255, %%xmm3, %%xmm3\n\t"
>    "punpcklqdq %%xmm3, %%xmm3\n\t"
>    "1:\n\t"
>    DIAG4_SSE2(-1,0,1,2)
>    "addl $8, %1\n\t"
>    "addl %2, %0\n\t"
>    "decl %3\n\t"
>    "jnz 1b\n\t"
>            :
>            : "r" (src), "r" (t), "g" (stride), "r" (11),
> "m"(*(int64_t*)h_weights)
>            : "memory"
>            );
>
>    t = tmp + 8;
>
>    asm (
>    "movq %4, %%xmm3\n\t"
>    "pshuflw $0, %%xmm3, %%xmm4\n\t"
>    "punpcklqdq %%xmm4, %%xmm4\n\t"
>    "pshuflw $85, %%xmm3, %%xmm5\n\t"
>    "punpcklqdq %%xmm5, %%xmm5\n\t"
>    "pshuflw $170, %%xmm3, %%xmm6\n\t"
>    "punpcklqdq %%xmm6, %%xmm6\n\t"
>    "pshuflw $255, %%xmm3, %%xmm3\n\t"
>    "punpcklqdq %%xmm3, %%xmm3\n\t"
>    "1:\n\t"
>    DIAG4_SSE2(-8,0,8,16)
>    "addl $8, %0\n\t"
>    "addl %2, %1\n\t"
>    "decl %3\n\t"
>    "jnz 1b\n\t"
>            :
>            : "r" (t), "r" (dst), "g" (stride), "r" (8),
> "m"(*(int64_t*)v_weights)
>            : "memory"
>    );
> }


Thanks for your time, I guess that mean the patch is actually working
(not crashing and even bitexact output ?).
I also used pshufw in my (still unsent) MMXEXT version of the patch.
I'm still tweaking it.

I have a small fix to the MMX code to test, I'll update your code accordingly.

It leaves me with the X86_64 problem which I don't know how to fix.

S?bastien