[FFmpeg-devel] [PATCH 06/15] vp9/x86: store unpacked intermediates for filter6/14 on stack.

Ronald S. Bultje rsbultje at gmail.com
Sat Dec 27 20:54:05 CET 2014


Hi,

On Sat, Dec 27, 2014 at 12:54 PM, Clément Bœsch <u at pkh.me> wrote:

> On Sat, Dec 27, 2014 at 11:02:41AM -0500, Ronald S. Bultje wrote:
> > filter16 goes from 508 to 482 (h) or 346 to 314 (v) cycles; filter88
> > goes from 240 to 238 (h) or 174 to 165 (v) cycles, measured on TOS.
> > ---
> >  libavcodec/x86/vp9lpf.asm | 151
> ++++++++++++++++++++++++----------------------
> >  1 file changed, 79 insertions(+), 72 deletions(-)
> >
> > diff --git a/libavcodec/x86/vp9lpf.asm b/libavcodec/x86/vp9lpf.asm
> > index 613a104..e117798 100644
> > --- a/libavcodec/x86/vp9lpf.asm
> > +++ b/libavcodec/x86/vp9lpf.asm
> > @@ -79,39 +79,42 @@ SECTION .text
> >      por                 %1, %4              ; new&mask | old&~mask
> >  %endmacro
> >
> > -%macro FILTER_SUBx2_ADDx2 8 ; %1=dst %2=h/l %3=cache %4=sub1 %5=sub2
> %6=add1 %7=add2 %8=rshift
> > -    punpck%2bw          %3, %4, m0
> > -    psubw               %1, %3
> > -    punpck%2bw          %3, %5, m0
> > -    psubw               %1, %3
> > -    punpck%2bw          %3, %6, m0
> > -    paddw               %1, %3
> > -    punpck%2bw          %3, %7, m0
> > +%macro FILTER_SUBx2_ADDx2 9-10 "" ; %1=dst %2=h/l %3=cache %4=stack_off
> %5=sub1 %6=sub2 %7=add1 %8=add2 %9=rshift, [unpack]
> > +    psubw               %3, [rsp+%4+%5*32]
> > +    psubw               %3, [rsp+%4+%6*32]
> > +    paddw               %3, [rsp+%4+%7*32]
> > +%ifnidn %10, ""
> > +    punpck%2bw          %1, %10, m0
> > +    mova    [rsp+%4+%8*32], %1
> >      paddw               %3, %1
> > -    psraw               %1, %3, %8
> > +%else
> > +    paddw               %3, [rsp+%4+%8*32]
> > +%endif
> > +    psraw               %1, %3, %9
> >  %endmacro
> >
> > -%macro FILTER_INIT 8 ; tmp1, tmp2, cacheL, cacheH, dstp, filterid,
> mask, source
> > -    FILTER%6_INIT       %1, l, %3
> > -    FILTER%6_INIT       %2, h, %4
> > +; FIXME interleave l/h better (for instruction pairing)
> > +%macro FILTER_INIT 9 ; tmp1, tmp2, cacheL, cacheH, dstp, stack_off,
> filterid, mask, source
> > +    FILTER%7_INIT       %1, l, %3, %6 +  0
> > +    FILTER%7_INIT       %2, h, %4, %6 + 16
> >      packuswb            %1, %2
> > -    MASK_APPLY          %1, %8, %7, %2
> > +    MASK_APPLY          %1, %9, %8, %2
> >      mova                %5, %1
> >  %endmacro
> >
> > -%macro FILTER_UPDATE 11-14 ; tmp1, tmp2, cacheL, cacheH, dstp, -, -, +,
> +, rshift, mask, [source], [preload reg + value]
> > -%if %0 == 13 ; no source + preload
> > -    mova                %12, %13
> > -%elif %0 == 14 ; source + preload
> > -    mova                %13, %14
> > +
> > +%macro FILTER_UPDATE 12-15 "", "" ; tmp1, tmp2, cacheL, cacheH, dstp,
> stack_off, -, -, +, +, rshift, mask, [source], [unpack]
> > +; FIXME interleave this properly with the subx2/addx2
> > +%if %0 == 15
> > +    mova               %14, %15
> >  %endif
> > -    FILTER_SUBx2_ADDx2  %1, l, %3, %6, %7, %8, %9, %10
> > -    FILTER_SUBx2_ADDx2  %2, h, %4, %6, %7, %8, %9, %10
> > +    FILTER_SUBx2_ADDx2  %1, l, %3, %6 +  0, %7, %8, %9, %10, %11, %14
> > +    FILTER_SUBx2_ADDx2  %2, h, %4, %6 + 16, %7, %8, %9, %10, %11, %14
> >      packuswb            %1, %2
> > -%if %0 == 12 || %0 == 14
> > -    MASK_APPLY          %1, %12, %11, %2
> > +%ifnidn %13, ""
> > +    MASK_APPLY          %1, %13, %12, %2
> >  %else
> > -    MASK_APPLY          %1, %5, %11, %2
> > +    MASK_APPLY          %1, %5, %12, %2
> >  %endif
> >      mova                %5, %1
> >  %endmacro
> > @@ -151,44 +154,48 @@ SECTION .text
> >      paddusb             %1, %4                          ; add the
> negatives
> >  %endmacro
> >
> > -%macro FILTER6_INIT 3 ; %1=dst %2=h/l %3=cache
> > +%macro FILTER6_INIT 4 ; %1=dst %2=h/l %3=cache, %4=stack_off
> >      punpck%2bw          %1, m14, m0                     ; p3: B->W
> > +    mova     [rsp+%4+0*32], %1
> >      paddw               %3, %1, %1                      ; p3*2
> >      paddw               %3, %1                          ; p3*3
> >      punpck%2bw          %1, m15, m0                     ; p2: B->W
> > +    mova     [rsp+%4+1*32], %1
> >      paddw               %3, %1                          ; p3*3 + p2
> >      paddw               %3, %1                          ; p3*3 + p2*2
> >      punpck%2bw          %1, m10, m0                     ; p1: B->W
> > +    mova     [rsp+%4+2*32], %1
> >      paddw               %3, %1                          ; p3*3 + p2*2 +
> p1
> >      punpck%2bw          %1, m11, m0                     ; p0: B->W
> > +    mova     [rsp+%4+3*32], %1
> >      paddw               %3, %1                          ; p3*3 + p2*2 +
> p1 + p0
> >      punpck%2bw          %1, m12, m0                     ; q0: B->W
> > +    mova     [rsp+%4+4*32], %1
> >      paddw               %3, %1                          ; p3*3 + p2*2 +
> p1 + p0 + q0
> >      paddw               %3, [pw_4]                      ; p3*3 + p2*2 +
> p1 + p0 + q0 + 4
> >      psraw               %1, %3, 3                       ; (p3*3 + p2*2
> + p1 + p0 + q0 + 4) >> 3
> >  %endmacro
> >
> > -%macro FILTER14_INIT 3 ; %1=dst %2=h/l %3=cache
> > +%macro FILTER14_INIT 4 ; %1=dst %2=h/l %3=cache, %4=stack_off
> >      punpck%2bw          %1, m2, m0                      ; p7: B->W
> > +    mova    [rsp+%4+ 8*32], %1
> >      psllw               %3, %1, 3                       ; p7*8
> >      psubw               %3, %1                          ; p7*7
> >      punpck%2bw          %1, m3, m0                      ; p6: B->W
> > +    mova    [rsp+%4+ 9*32], %1
> >      paddw               %3, %1                          ; p7*7 + p6
> >      paddw               %3, %1                          ; p7*7 + p6*2
> >      punpck%2bw          %1, m8, m0                      ; p5: B->W
> > +    mova    [rsp+%4+10*32], %1
> >      paddw               %3, %1                          ; p7*7 + p6*2 +
> p5
> >      punpck%2bw          %1, m9, m0                      ; p4: B->W
> > +    mova    [rsp+%4+11*32], %1
> >      paddw               %3, %1                          ; p7*7 + p6*2 +
> p5 + p4
> > -    punpck%2bw          %1, m14, m0                     ; p3: B->W
> > -    paddw               %3, %1                          ; p7*7 + p6*2 +
> p5 + p4 + p3
> > -    punpck%2bw          %1, m15, m0                     ; p2: B->W
> > -    paddw               %3, %1                          ; p7*7 + p6*2 +
> p5 + .. + p2
> > -    punpck%2bw          %1, m10, m0                     ; p1: B->W
> > -    paddw               %3, %1                          ; p7*7 + p6*2 +
> p5 + .. + p1
> > -    punpck%2bw          %1, m11, m0                     ; p0: B->W
> > -    paddw               %3, %1                          ; p7*7 + p6*2 +
> p5 + .. + p0
> > -    punpck%2bw          %1, m12, m0                     ; q0: B->W
> > -    paddw               %3, %1                          ; p7*7 + p6*2 +
> p5 + .. + p0 + q0
> > +    paddw               %3, [rsp+%4+ 0*32]              ; p7*7 + p6*2 +
> p5 + p4 + p3
> > +    paddw               %3, [rsp+%4+ 1*32]              ; p7*7 + p6*2 +
> p5 + .. + p2
> > +    paddw               %3, [rsp+%4+ 2*32]              ; p7*7 + p6*2 +
> p5 + .. + p1
> > +    paddw               %3, [rsp+%4+ 3*32]              ; p7*7 + p6*2 +
> p5 + .. + p0
> > +    paddw               %3, [rsp+%4+ 4*32]              ; p7*7 + p6*2 +
> p5 + .. + p0 + q0
> >      paddw               %3, [pw_8]                      ; p7*7 + p6*2 +
> p5 + .. + p0 + q0 + 8
> >      psraw               %1, %3, 4                       ; (p7*7 + p6*2
> + p5 + .. + p0 + q0 + 8) >> 4
> >  %endmacro
> > @@ -306,14 +313,14 @@ SECTION .text
> >  %endif
> >  %endmacro
> >
> > -%macro LOOPFILTER 3 ; %1=v/h %2=size1 %3=stack
> > +%macro LOOPFILTER 4 ; %1=v/h %2=size1 %3+%4=stack
> >  %if UNIX64
> > -cglobal vp9_loop_filter_%1_%2_16, 5, 9, 16, %3, dst, stride, E, I, H,
> mstride, dst2, stride3, mstride3
> > +cglobal vp9_loop_filter_%1_%2_16, 5, 9, 16, %3 + %4, dst, stride, E, I,
> H, mstride, dst2, stride3, mstride3
> >  %else
> >  %if WIN64
> > -cglobal vp9_loop_filter_%1_%2_16, 4, 8, 16, %3, dst, stride, E, I,
> mstride, dst2, stride3, mstride3
> > +cglobal vp9_loop_filter_%1_%2_16, 4, 8, 16, %3 + %4, dst, stride, E, I,
> mstride, dst2, stride3, mstride3
> >  %else
> > -cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3, dst, stride, mstride,
> dst2, stride3, mstride3
> > +cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4, dst, stride,
> mstride, dst2, stride3, mstride3
> >  %define Ed dword r2m
> >  %define Id dword r3m
> >  %endif
> > @@ -630,12 +637,12 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3,
> dst, stride, mstride, dst2, stri
> >      mova               m15, [P2]
> >      mova                m8, [Q2]
> >      mova                m9, [Q3]
> > -    FILTER_INIT         m4, m5, m6, m7, [P2], 6,
>  m3, m15    ; [p2]
> > -    FILTER_UPDATE       m6, m7, m4, m5, [P1], m14, m15, m10, m13, 3,
> m3         ; [p1] -p3 -p2 +p1 +q1
> > -    FILTER_UPDATE       m4, m5, m6, m7, [P0], m14, m10, m11,  m8, 3,
> m3         ; [p0] -p3 -p1 +p0 +q2
> > -    FILTER_UPDATE       m6, m7, m4, m5, [Q0], m14, m11, m12,  m9, 3,
> m3         ; [q0] -p3 -p0 +q0 +q3
> > -    FILTER_UPDATE       m4, m5, m6, m7, [Q1], m15, m12, m13,  m9, 3,
> m3         ; [q1] -p2 -q0 +q1 +q3
> > -    FILTER_UPDATE       m6, m7, m4, m5, [Q2], m10, m13,  m8,  m9, 3,
> m3,  m8    ; [q2] -p1 -q1 +q2 +q3
> > +    FILTER_INIT         m4, m5, m6, m7, [P2], %4, 6,             m3,
> m15      ; [p2]
> > +    FILTER_UPDATE       m4, m5, m6, m7, [P1], %4, 0, 1, 2, 5, 3, m3,
> "", m13 ; [p1] -p3 -p2 +p1 +q1
> > +    FILTER_UPDATE       m4, m5, m6, m7, [P0], %4, 0, 2, 3, 6, 3, m3,
> "", m8  ; [p0] -p3 -p1 +p0 +q2
> > +    FILTER_UPDATE       m4, m5, m6, m7, [Q0], %4, 0, 3, 4, 7, 3, m3,
> "", m9  ; [q0] -p3 -p0 +q0 +q3
> > +    FILTER_UPDATE       m4, m5, m6, m7, [Q1], %4, 1, 4, 5, 7, 3, m3,
> ""      ; [q1] -p2 -q0 +q1 +q3
> > +    FILTER_UPDATE       m4, m5, m6, m7, [Q2], %4, 2, 5, 6, 7, 3, m3,
> m8      ; [q2] -p1 -q1 +q2 +q3
> >  %endif
> >
> >      ; (m0: 0, [m1: flat8out], m2: fm & flat8in, m8..15: q2 q3 p1 p0 q0
> q1 p3 p2)
> > @@ -665,20 +672,20 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3,
> dst, stride, mstride, dst2, stri
> >      mova            m3, [P6]
> >      mova            m8, [P5]
> >      mova            m9, [P4]
> > -    FILTER_INIT     m4, m5, m6, m7, [P6],  14,                   m1,  m3
> > -    FILTER_UPDATE   m6, m7, m4, m5, [P5],  m2,  m3,  m8, m13, 4, m1,
> m8                ; [p5] -p7 -p6 +p5 +q1
> > -    FILTER_UPDATE   m4, m5, m6, m7, [P4],  m2,  m8,  m9, m13, 4, m1,
> m9, m13, [Q2]     ; [p4] -p7 -p5 +p4 +q2
> > -    FILTER_UPDATE   m6, m7, m4, m5, [P3],  m2,  m9, m14, m13, 4, m1,
> m14, m13, [Q3]     ; [p3] -p7 -p4 +p3 +q3
> > -    FILTER_UPDATE   m4, m5, m6, m7, [P2],  m2, m14, m15, m13, 4, m1,
>   m13, [Q4]     ; [p2] -p7 -p3 +p2 +q4
> > -    FILTER_UPDATE   m6, m7, m4, m5, [P1],  m2, m15, m10, m13, 4, m1,
>   m13, [Q5]     ; [p1] -p7 -p2 +p1 +q5
> > -    FILTER_UPDATE   m4, m5, m6, m7, [P0],  m2, m10, m11, m13, 4, m1,
>   m13, [Q6]     ; [p0] -p7 -p1 +p0 +q6
> > -    FILTER_UPDATE   m6, m7, m4, m5, [Q0],  m2, m11, m12, m13, 4, m1,
>   m13, [Q7]     ; [q0] -p7 -p0 +q0 +q7
> > -    FILTER_UPDATE   m4, m5, m6, m7, [Q1],  m3, m12,  m2, m13, 4, m1,
>    m2, [Q1]     ; [q1] -p6 -q0 +q1 +q7
> > -    FILTER_UPDATE   m6, m7, m4, m5, [Q2],  m8,  m2,  m3, m13, 4, m1,
>    m3, [Q2]     ; [q2] -p5 -q1 +q2 +q7
> > -    FILTER_UPDATE   m4, m5, m6, m7, [Q3],  m9,  m3,  m8, m13, 4, m1,
> m8,  m8, [Q3]     ; [q3] -p4 -q2 +q3 +q7
> > -    FILTER_UPDATE   m6, m7, m4, m5, [Q4], m14,  m8,  m9, m13, 4, m1,
> m9,  m9, [Q4]     ; [q4] -p3 -q3 +q4 +q7
> > -    FILTER_UPDATE   m4, m5, m6, m7, [Q5], m15,  m9, m14, m13, 4, m1,
> m14, m14, [Q5]     ; [q5] -p2 -q4 +q5 +q7
> > -    FILTER_UPDATE   m6, m7, m4, m5, [Q6], m10, m14, m15, m13, 4, m1,
> m15, m15, [Q6]     ; [q6] -p1 -q5 +q6 +q7
> > +    FILTER_INIT     m4, m5, m6, m7, [P6], %4, 14,                m1,
> m3            ; [p6]
> > +    FILTER_UPDATE   m4, m5, m6, m7, [P5], %4,  8,  9, 10,  5, 4, m1,
> m8            ; [p5] -p7 -p6 +p5 +q1
> > +    FILTER_UPDATE   m4, m5, m6, m7, [P4], %4,  8, 10, 11,  6, 4, m1,
> m9            ; [p4] -p7 -p5 +p4 +q2
> > +    FILTER_UPDATE   m4, m5, m6, m7, [P3], %4,  8, 11,  0,  7, 4, m1,
> m14            ; [p3] -p7 -p4 +p3 +q3
> > +    FILTER_UPDATE   m4, m5, m6, m7, [P2], %4,  8,  0,  1, 12, 4, m1,
> "",  m8, [Q4] ; [p2] -p7 -p3 +p2 +q4
> > +    FILTER_UPDATE   m4, m5, m6, m7, [P1], %4,  8,  1,  2, 13, 4, m1,
> "",  m9, [Q5] ; [p1] -p7 -p2 +p1 +q5
> > +    FILTER_UPDATE   m4, m5, m6, m7, [P0], %4,  8,  2,  3, 14, 4, m1,
> "", m14, [Q6] ; [p0] -p7 -p1 +p0 +q6
> > +    FILTER_UPDATE   m4, m5, m6, m7, [Q0], %4,  8,  3,  4, 15, 4, m1,
> "", m15, [Q7] ; [q0] -p7 -p0 +q0 +q7
> > +    FILTER_UPDATE   m4, m5, m6, m7, [Q1], %4,  9,  4,  5, 15, 4, m1,
> ""            ; [q1] -p6 -q0 +q1 +q7
> > +    FILTER_UPDATE   m4, m5, m6, m7, [Q2], %4, 10,  5,  6, 15, 4, m1,
> ""            ; [q2] -p5 -q1 +q2 +q7
> > +    FILTER_UPDATE   m4, m5, m6, m7, [Q3], %4, 11,  6,  7, 15, 4, m1,
> ""            ; [q3] -p4 -q2 +q3 +q7
> > +    FILTER_UPDATE   m4, m5, m6, m7, [Q4], %4,  0,  7, 12, 15, 4, m1,
> m8            ; [q4] -p3 -q3 +q4 +q7
> > +    FILTER_UPDATE   m4, m5, m6, m7, [Q5], %4,  1, 12, 13, 15, 4, m1,
> m9            ; [q5] -p2 -q4 +q5 +q7
> > +    FILTER_UPDATE   m4, m5, m6, m7, [Q6], %4,  2, 13, 14, 15, 4, m1,
> m14            ; [q6] -p1 -q5 +q6 +q7
>
> Does this change require update of the ASCII schematic above?
>
> In the schematic, horizontal axis are the cached line content for the
> given register, and vertical axis is each destination, one after another.
> A dot means the register content is unchanged. If the new code (caching
> results onto the stack) doesn't fit into this schematic, feel free to drop
> it, or make a new one
>

No, the ascii part didn't change. What I changed was that you unpacked each
value per addition or subtraction. I unpack it only once, store the result
on stack and reuse it for future additions and subtractions. So where you
unpack 4x and then sub/add from registers, I unpack 1x, sub/add from
register (since it's already there), store on stack, and add/sub ~3x from
stack. However, the sequence didn't change from your schematic. :)

Ronald


More information about the ffmpeg-devel mailing list