ryan/mozjpeg
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Attempt to improve performance by refactoring the compression-side color conversion and DCT algorithms so that they take full advantage of the additional registers available with 64-bit SSE2. This produces a somewhat yawn-worthy speedup of 2-3%, but at least the code is a lot more readable now.

Browse Source 
git-svn-id: svn+ssh://svn.code.sf.net/p/libjpeg-turbo/code/trunk@1335 632fc199-4ca6-4c93-a231-07263d6284db
This commit is contained in:
DRC
2014-08-09 14:30:28 +00:00
parent 3728aa01d8
commit a8ab3424c7
4 changed files with 576 additions and 725 deletions
Download Patch File Download Diff File Expand all files Collapse all files

123
simd/jccolext-sse2-64.asm
View File

@@ -3,7 +3,7 @@
;
; x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
; Copyright (C) 2009, D. R. Commander.
; Copyright (C) 2009, 2014, D. R. Commander.
; For conditions of distribution and use, see copyright notice in jsimdext.inc
;
; This file should be assembled with NASM (Netwide Assembler),
@@ -32,21 +32,11 @@
; r13 = JDIMENSION output_row
; r14 = int num_rows
%define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
%define WK_NUM 8
align 16
global EXTN(jsimd_rgb_ycc_convert_sse2)
EXTN(jsimd_rgb_ycc_convert_sse2):
push rbp
mov rax,rsp ; rax = original rbp
sub rsp, byte 4
and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
mov [rsp],rax
mov rbp,rsp ; rbp = aligned rbp
lea rsp, [wk(0)]
collect_args
push rbx
@@ -290,34 +280,31 @@ EXTN(jsimd_rgb_ycc_convert_sse2):
; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=RE
movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=RO
movdqa XMMWORD [wk(2)], xmm4 ; wk(2)=BE
movdqa XMMWORD [wk(3)], xmm5 ; wk(3)=BO
movdqa xmm8,xmm0 ; xmm8=RE
movdqa xmm9,xmm1 ; xmm9=RO
movdqa xmm10,xmm4 ; xmm10=BE
movdqa xmm11,xmm5 ; xmm11=BO
movdqa xmm6,xmm1
movdqa xmm13,xmm1
punpcklwd xmm1,xmm3
punpckhwd xmm6,xmm3
punpckhwd xmm13,xmm3
movdqa xmm7,xmm1
movdqa xmm4,xmm6
movdqa xmm4,xmm13
pmaddwd xmm1,[rel PW_F0299_F0337] ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
pmaddwd xmm6,[rel PW_F0299_F0337] ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
pmaddwd xmm13,[rel PW_F0299_F0337] ; xmm13=ROH*FIX(0.299)+GOH*FIX(0.337)
pmaddwd xmm7,[rel PW_MF016_MF033] ; xmm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
pmaddwd xmm4,[rel PW_MF016_MF033] ; xmm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
movdqa XMMWORD [wk(4)], xmm1 ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
movdqa XMMWORD [wk(5)], xmm6 ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
pxor xmm1,xmm1
pxor xmm12,xmm12
pxor xmm6,xmm6
punpcklwd xmm1,xmm5 ; xmm1=BOL
punpcklwd xmm12,xmm5 ; xmm12=BOL
punpckhwd xmm6,xmm5 ; xmm6=BOH
psrld xmm1,1 ; xmm1=BOL*FIX(0.500)
psrld xmm12,1 ; xmm12=BOL*FIX(0.500)
psrld xmm6,1 ; xmm6=BOH*FIX(0.500)
movdqa xmm5,[rel PD_ONEHALFM1_CJ] ; xmm5=[PD_ONEHALFM1_CJ]
paddd xmm7,xmm1
paddd xmm7,xmm12
paddd xmm4,xmm6
paddd xmm7,xmm5
paddd xmm4,xmm5
@@ -325,34 +312,29 @@ EXTN(jsimd_rgb_ycc_convert_sse2):
psrld xmm4,SCALEBITS ; xmm4=CbOH
packssdw xmm7,xmm4 ; xmm7=CbO
movdqa xmm1, XMMWORD [wk(2)] ; xmm1=BE
movdqa xmm6,xmm0
movdqa xmm15,xmm0
punpcklwd xmm0,xmm2
punpckhwd xmm6,xmm2
punpckhwd xmm15,xmm2
movdqa xmm5,xmm0
movdqa xmm4,xmm6
movdqa xmm4,xmm15
pmaddwd xmm0,[rel PW_F0299_F0337] ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
pmaddwd xmm6,[rel PW_F0299_F0337] ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
pmaddwd xmm15,[rel PW_F0299_F0337] ; xmm15=REH*FIX(0.299)+GEH*FIX(0.337)
pmaddwd xmm5,[rel PW_MF016_MF033] ; xmm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
pmaddwd xmm4,[rel PW_MF016_MF033] ; xmm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
movdqa XMMWORD [wk(6)], xmm0 ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
movdqa XMMWORD [wk(7)], xmm6 ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
pxor xmm0,xmm0
pxor xmm14,xmm14
pxor xmm6,xmm6
punpcklwd xmm0,xmm1 ; xmm0=BEL
punpckhwd xmm6,xmm1 ; xmm6=BEH
psrld xmm0,1 ; xmm0=BEL*FIX(0.500)
punpcklwd xmm14,xmm10 ; xmm14=BEL
punpckhwd xmm6,xmm10 ; xmm6=BEH
psrld xmm14,1 ; xmm14=BEL*FIX(0.500)
psrld xmm6,1 ; xmm6=BEH*FIX(0.500)
movdqa xmm1,[rel PD_ONEHALFM1_CJ] ; xmm1=[PD_ONEHALFM1_CJ]
movdqa xmm12,[rel PD_ONEHALFM1_CJ] ; xmm12=[PD_ONEHALFM1_CJ]
paddd xmm5,xmm0
paddd xmm5,xmm14
paddd xmm4,xmm6
paddd xmm5,xmm1
paddd xmm4,xmm1
paddd xmm5,xmm12
paddd xmm4,xmm12
psrld xmm5,SCALEBITS ; xmm5=CbEL
psrld xmm4,SCALEBITS ; xmm4=CbEH
packssdw xmm5,xmm4 ; xmm5=CbE
@@ -361,34 +343,30 @@ EXTN(jsimd_rgb_ycc_convert_sse2):
por xmm5,xmm7 ; xmm5=Cb
movdqa XMMWORD [rbx], xmm5 ; Save Cb
movdqa xmm0, XMMWORD [wk(3)] ; xmm0=BO
movdqa xmm6, XMMWORD [wk(2)] ; xmm6=BE
movdqa xmm1, XMMWORD [wk(1)] ; xmm1=RO
movdqa xmm4,xmm0
punpcklwd xmm0,xmm3
movdqa xmm4,xmm11
punpcklwd xmm11,xmm3
punpckhwd xmm4,xmm3
movdqa xmm7,xmm0
movdqa xmm7,xmm11
movdqa xmm5,xmm4
pmaddwd xmm0,[rel PW_F0114_F0250] ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
pmaddwd xmm11,[rel PW_F0114_F0250] ; xmm11=BOL*FIX(0.114)+GOL*FIX(0.250)
pmaddwd xmm4,[rel PW_F0114_F0250] ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
pmaddwd xmm7,[rel PW_MF008_MF041] ; xmm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
pmaddwd xmm5,[rel PW_MF008_MF041] ; xmm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
movdqa xmm3,[rel PD_ONEHALF] ; xmm3=[PD_ONEHALF]
paddd xmm0, XMMWORD [wk(4)]
paddd xmm4, XMMWORD [wk(5)]
paddd xmm0,xmm3
paddd xmm11, xmm1
paddd xmm4, xmm13
paddd xmm11,xmm3
paddd xmm4,xmm3
psrld xmm0,SCALEBITS ; xmm0=YOL
psrld xmm11,SCALEBITS ; xmm11=YOL
psrld xmm4,SCALEBITS ; xmm4=YOH
packssdw xmm0,xmm4 ; xmm0=YO
packssdw xmm11,xmm4 ; xmm11=YO
pxor xmm3,xmm3
pxor xmm4,xmm4
punpcklwd xmm3,xmm1 ; xmm3=ROL
punpckhwd xmm4,xmm1 ; xmm4=ROH
punpcklwd xmm3,xmm9 ; xmm3=ROL
punpckhwd xmm4,xmm9 ; xmm4=ROH
psrld xmm3,1 ; xmm3=ROL*FIX(0.500)
psrld xmm4,1 ; xmm4=ROH*FIX(0.500)
@@ -402,31 +380,31 @@ EXTN(jsimd_rgb_ycc_convert_sse2):
psrld xmm5,SCALEBITS ; xmm5=CrOH
packssdw xmm7,xmm5 ; xmm7=CrO
movdqa xmm3, XMMWORD [wk(0)] ; xmm3=RE
movdqa xmm3, xmm8 ; xmm3=RE
movdqa xmm4,xmm6
punpcklwd xmm6,xmm2
movdqa xmm4,xmm10
punpcklwd xmm10,xmm2
punpckhwd xmm4,xmm2
movdqa xmm1,xmm6
movdqa xmm1,xmm10
movdqa xmm5,xmm4
pmaddwd xmm6,[rel PW_F0114_F0250] ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
pmaddwd xmm10,[rel PW_F0114_F0250] ; xmm10=BEL*FIX(0.114)+GEL*FIX(0.250)
pmaddwd xmm4,[rel PW_F0114_F0250] ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
pmaddwd xmm1,[rel PW_MF008_MF041] ; xmm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
pmaddwd xmm5,[rel PW_MF008_MF041] ; xmm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
movdqa xmm2,[rel PD_ONEHALF] ; xmm2=[PD_ONEHALF]
paddd xmm6, XMMWORD [wk(6)]
paddd xmm4, XMMWORD [wk(7)]
paddd xmm6,xmm2
paddd xmm10, xmm0
paddd xmm4, xmm15
paddd xmm10,xmm2
paddd xmm4,xmm2
psrld xmm6,SCALEBITS ; xmm6=YEL
psrld xmm10,SCALEBITS ; xmm10=YEL
psrld xmm4,SCALEBITS ; xmm4=YEH
packssdw xmm6,xmm4 ; xmm6=YE
packssdw xmm10,xmm4 ; xmm10=YE
psllw xmm0,BYTE_BIT
por xmm6,xmm0 ; xmm6=Y
movdqa XMMWORD [rdi], xmm6 ; Save Y
psllw xmm11,BYTE_BIT
por xmm10,xmm11 ; xmm10=Y
movdqa XMMWORD [rdi], xmm10 ; Save Y
pxor xmm2,xmm2
pxor xmm4,xmm4
@@ -475,9 +453,6 @@ EXTN(jsimd_rgb_ycc_convert_sse2):
.return:
pop rbx
uncollect_args
mov rsp,rbp ; rsp <- aligned rbp
pop rsp ; rsp <- original rbp
pop rbp
ret
; For some reason, the OS X linker does not honor the request to align the

75
simd/jcgryext-sse2-64.asm
View File

@@ -3,7 +3,7 @@
;
; x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
; Copyright (C) 2011, D. R. Commander.
; Copyright (C) 2011, 2014, D. R. Commander.
; For conditions of distribution and use, see copyright notice in jsimdext.inc
;
; This file should be assembled with NASM (Netwide Assembler),
@@ -32,21 +32,11 @@
; r13 = JDIMENSION output_row
; r14 = int num_rows
%define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
%define WK_NUM 2
align 16
global EXTN(jsimd_rgb_gray_convert_sse2)
EXTN(jsimd_rgb_gray_convert_sse2):
push rbp
mov rax,rsp ; rax = original rbp
sub rsp, byte 4
and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
mov [rsp],rax
mov rbp,rsp ; rbp = aligned rbp
lea rsp, [wk(0)]
collect_args
push rbx
@@ -284,7 +274,7 @@ EXTN(jsimd_rgb_gray_convert_sse2):
pmaddwd xmm1,[rel PW_F0299_F0337] ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
pmaddwd xmm6,[rel PW_F0299_F0337] ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
movdqa xmm7, xmm6 ; xmm7=ROH*FIX(0.299)+GOH*FIX(0.337)
movdqa xmm7, xmm6 ; xmm7=ROH*FIX(0.299)+GOH*FIX(0.337)
movdqa xmm6,xmm0
punpcklwd xmm0,xmm2
@@ -292,47 +282,41 @@ EXTN(jsimd_rgb_gray_convert_sse2):
pmaddwd xmm0,[rel PW_F0299_F0337] ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
pmaddwd xmm6,[rel PW_F0299_F0337] ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337)
movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337)
movdqa xmm0, xmm5 ; xmm0=BO
movdqa xmm6, xmm4 ; xmm6=BE
movdqa xmm4,xmm0
punpcklwd xmm0,xmm3
punpckhwd xmm4,xmm3
pmaddwd xmm0,[rel PW_F0114_F0250] ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
pmaddwd xmm4,[rel PW_F0114_F0250] ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
movdqa xmm12,xmm5
punpcklwd xmm5,xmm3
punpckhwd xmm12,xmm3
pmaddwd xmm5,[rel PW_F0114_F0250] ; xmm5=BOL*FIX(0.114)+GOL*FIX(0.250)
pmaddwd xmm12,[rel PW_F0114_F0250] ; xmm12=BOH*FIX(0.114)+GOH*FIX(0.250)
movdqa xmm3,[rel PD_ONEHALF] ; xmm3=[PD_ONEHALF]
paddd xmm0, xmm1
paddd xmm4, xmm7
paddd xmm0,xmm3
paddd xmm4,xmm3
psrld xmm0,SCALEBITS ; xmm0=YOL
psrld xmm4,SCALEBITS ; xmm4=YOH
packssdw xmm0,xmm4 ; xmm0=YO
paddd xmm5, xmm1
paddd xmm12, xmm7
paddd xmm5,xmm3
paddd xmm12,xmm3
psrld xmm5,SCALEBITS ; xmm5=YOL
psrld xmm12,SCALEBITS ; xmm12=YOH
packssdw xmm5,xmm12 ; xmm5=YO
movdqa xmm4,xmm6
punpcklwd xmm6,xmm2
punpckhwd xmm4,xmm2
pmaddwd xmm6,[rel PW_F0114_F0250] ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
pmaddwd xmm4,[rel PW_F0114_F0250] ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
movdqa xmm12,xmm4
punpcklwd xmm4,xmm2
punpckhwd xmm12,xmm2
pmaddwd xmm4,[rel PW_F0114_F0250] ; xmm4=BEL*FIX(0.114)+GEL*FIX(0.250)
pmaddwd xmm12,[rel PW_F0114_F0250] ; xmm12=BEH*FIX(0.114)+GEH*FIX(0.250)
movdqa xmm2,[rel PD_ONEHALF] ; xmm2=[PD_ONEHALF]
paddd xmm6, XMMWORD [wk(0)]
paddd xmm4, XMMWORD [wk(1)]
paddd xmm6,xmm2
paddd xmm4,xmm0 ; xmm4=REL*FIX(0.299)+GEL*FIX(0.337)
paddd xmm12,xmm6 ; xmm12=REH*FIX(0.299)+GEH*FIX(0.337)
paddd xmm4,xmm2
psrld xmm6,SCALEBITS ; xmm6=YEL
psrld xmm4,SCALEBITS ; xmm4=YEH
packssdw xmm6,xmm4 ; xmm6=YE
paddd xmm12,xmm2
psrld xmm4,SCALEBITS ; xmm4=YEL
psrld xmm12,SCALEBITS ; xmm12=YEH
packssdw xmm4,xmm12 ; xmm4=YE
psllw xmm0,BYTE_BIT
por xmm6,xmm0 ; xmm6=Y
movdqa XMMWORD [rdi], xmm6 ; Save Y
psllw xmm5,BYTE_BIT
por xmm4,xmm5 ; xmm4=Y
movdqa XMMWORD [rdi], xmm4 ; Save Y
sub rcx, byte SIZEOF_XMMWORD
add rsi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; inptr
@@ -354,9 +338,6 @@ EXTN(jsimd_rgb_gray_convert_sse2):
.return:
pop rbx
uncollect_args
mov rsp,rbp ; rsp <- aligned rbp
pop rsp ; rsp <- original rbp
pop rbp
ret
; For some reason, the OS X linker does not honor the request to align the

452
simd/jfdctfst-sse2-64.asm
View File

@@ -2,7 +2,7 @@
; jfdctfst.asm - fast integer FDCT (64-bit SSE2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright 2009 D. R. Commander
; Copyright 2009, 2014, D. R. Commander
;
; Based on
; x86 SIMD extension for IJG JPEG library
@@ -76,315 +76,273 @@ PW_F1306 times 8 dw F_1_306 << CONST_SHIFT
; r10 = DCTELEM * data
%define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
%define WK_NUM 2
align 16
global EXTN(jsimd_fdct_ifast_sse2)
EXTN(jsimd_fdct_ifast_sse2):
push rbp
mov rax,rsp ; rax = original rbp
sub rsp, byte 4
and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
mov [rsp],rax
mov rbp,rsp ; rbp = aligned rbp
lea rsp, [wk(0)]
collect_args
; ---- Pass 1: process rows.
mov rdx, r10 ; (DCTELEM *)
movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm8, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm9, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm11, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm4, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm13, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm5, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm15, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
; xmm8=(00 01 02 03 04 05 06 07), xmm9=(10 11 12 13 14 15 16 17)
; xmm2=(20 21 22 23 24 25 26 27), xmm11=(30 31 32 33 34 35 36 37)
; xmm4=(40 41 42 43 44 45 46 47), xmm13=(50 51 52 53 54 55 56 57)
; xmm5=(60 61 62 63 64 65 66 67), xmm15=(70 71 72 73 74 75 76 77)
movdqa xmm4,xmm0 ; transpose coefficients(phase 1)
punpcklwd xmm0,xmm1 ; xmm0=(00 10 01 11 02 12 03 13)
punpckhwd xmm4,xmm1 ; xmm4=(04 14 05 15 06 16 07 17)
movdqa xmm5,xmm2 ; transpose coefficients(phase 1)
punpcklwd xmm2,xmm3 ; xmm2=(20 30 21 31 22 32 23 33)
punpckhwd xmm5,xmm3 ; xmm5=(24 34 25 35 26 36 27 37)
movdqa xmm12,xmm8 ; transpose coefficients(phase 1)
punpcklwd xmm8,xmm9 ; xmm8=(00 10 01 11 02 12 03 13)
punpckhwd xmm12,xmm9 ; xmm12=(04 14 05 15 06 16 07 17)
movdqa xmm1,xmm2 ; transpose coefficients(phase 1)
punpcklwd xmm2,xmm11 ; xmm2=(20 30 21 31 22 32 23 33)
punpckhwd xmm1,xmm11 ; xmm1=(24 34 25 35 26 36 27 37)
movdqa xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm0,xmm4 ; transpose coefficients(phase 1)
punpcklwd xmm4,xmm13 ; xmm4=(40 50 41 51 42 52 43 53)
punpckhwd xmm0,xmm13 ; xmm0=(44 54 45 55 46 56 47 57)
movdqa xmm3,xmm5 ; transpose coefficients(phase 1)
punpcklwd xmm5,xmm15 ; xmm5=(60 70 61 71 62 72 63 73)
punpckhwd xmm3,xmm15 ; xmm3=(64 74 65 75 66 76 67 77)
; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33)
movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37)
movdqa xmm2,xmm6 ; transpose coefficients(phase 1)
punpcklwd xmm6,xmm7 ; xmm6=(40 50 41 51 42 52 43 53)
punpckhwd xmm2,xmm7 ; xmm2=(44 54 45 55 46 56 47 57)
movdqa xmm5,xmm1 ; transpose coefficients(phase 1)
punpcklwd xmm1,xmm3 ; xmm1=(60 70 61 71 62 72 63 73)
punpckhwd xmm5,xmm3 ; xmm5=(64 74 65 75 66 76 67 77)
movdqa xmm7,xmm6 ; transpose coefficients(phase 2)
punpckldq xmm6,xmm1 ; xmm6=(40 50 60 70 41 51 61 71)
punpckhdq xmm7,xmm1 ; xmm7=(42 52 62 72 43 53 63 73)
movdqa xmm3,xmm2 ; transpose coefficients(phase 2)
punpckldq xmm2,xmm5 ; xmm2=(44 54 64 74 45 55 65 75)
punpckhdq xmm3,xmm5 ; xmm3=(46 56 66 76 47 57 67 77)
movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33)
movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37)
movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(42 52 62 72 43 53 63 73)
movdqa XMMWORD [wk(1)], xmm2 ; wk(1)=(44 54 64 74 45 55 65 75)
movdqa xmm10,xmm8 ; transpose coefficients(phase 2)
punpckldq xmm8,xmm2 ; xmm8=(00 10 20 30 01 11 21 31)
punpckhdq xmm10,xmm2 ; xmm10=(02 12 22 32 03 13 23 33)
movdqa xmm14,xmm12 ; transpose coefficients(phase 2)
punpckldq xmm12,xmm1 ; xmm12=(04 14 24 34 05 15 25 35)
punpckhdq xmm14,xmm1 ; xmm14=(06 16 26 36 07 17 27 37)
movdqa xmm6,xmm4 ; transpose coefficients(phase 2)
punpckldq xmm4,xmm5 ; xmm4=(40 50 60 70 41 51 61 71)
punpckhdq xmm6,xmm5 ; xmm6=(42 52 62 72 43 53 63 73)
movdqa xmm7,xmm0 ; transpose coefficients(phase 2)
punpckldq xmm0,xmm1 ; xmm0=(00 10 20 30 01 11 21 31)
punpckhdq xmm7,xmm1 ; xmm7=(02 12 22 32 03 13 23 33)
movdqa xmm2,xmm4 ; transpose coefficients(phase 2)
punpckldq xmm4,xmm5 ; xmm4=(04 14 24 34 05 15 25 35)
punpckhdq xmm2,xmm5 ; xmm2=(06 16 26 36 07 17 27 37)
punpckldq xmm0,xmm3 ; xmm0=(44 54 64 74 45 55 65 75)
punpckhdq xmm7,xmm3 ; xmm7=(46 56 66 76 47 57 67 77)
movdqa xmm1,xmm0 ; transpose coefficients(phase 3)
punpcklqdq xmm0,xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0
punpckhqdq xmm1,xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1
movdqa xmm5,xmm2 ; transpose coefficients(phase 3)
punpcklqdq xmm2,xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6
punpckhqdq xmm5,xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7
movdqa xmm9,xmm8 ; transpose coefficients(phase 3)
punpcklqdq xmm8,xmm4 ; xmm8=(00 10 20 30 40 50 60 70)=data0
punpckhqdq xmm9,xmm4 ; xmm9=(01 11 21 31 41 51 61 71)=data1
movdqa xmm11,xmm10 ; transpose coefficients(phase 3)
punpcklqdq xmm10,xmm6 ; xmm10=(02 12 22 32 42 52 62 72)=data2
punpckhqdq xmm11,xmm6 ; xmm11=(03 13 23 33 43 53 63 73)=data3
movdqa xmm6,xmm1
movdqa xmm3,xmm0
psubw xmm1,xmm2 ; xmm1=data1-data6=tmp6
psubw xmm0,xmm5 ; xmm0=data0-data7=tmp7
paddw xmm6,xmm2 ; xmm6=data1+data6=tmp1
paddw xmm3,xmm5 ; xmm3=data0+data7=tmp0
movdqa xmm13,xmm12 ; transpose coefficients(phase 3)
punpcklqdq xmm12,xmm0 ; xmm12=(04 14 24 34 44 54 64 74)=data4
punpckhqdq xmm13,xmm0 ; xmm13=(05 15 25 35 45 55 65 75)=data5
movdqa xmm15,xmm14 ; transpose coefficients(phase 3)
punpcklqdq xmm14,xmm7 ; xmm14=(06 16 26 36 46 56 66 76)=data6
punpckhqdq xmm15,xmm7 ; xmm15=(07 17 27 37 47 57 67 77)=data7
movdqa xmm2, XMMWORD [wk(0)] ; xmm2=(42 52 62 72 43 53 63 73)
movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(44 54 64 74 45 55 65 75)
movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6
movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7
movdqa xmm0,xmm8
paddw xmm0,xmm15 ; xmm0=data0+data7=tmp0
movdqa xmm1,xmm9
paddw xmm1,xmm14 ; xmm1=data1+data6=tmp1
movdqa xmm2,xmm10
paddw xmm2,xmm13 ; xmm2=data2+data5=tmp2
movdqa xmm3,xmm11
paddw xmm3,xmm12 ; xmm3=data3+data4=tmp3
movdqa xmm1,xmm7 ; transpose coefficients(phase 3)
punpcklqdq xmm7,xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2
punpckhqdq xmm1,xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3
movdqa xmm0,xmm4 ; transpose coefficients(phase 3)
punpcklqdq xmm4,xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4
punpckhqdq xmm0,xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5
movdqa xmm2,xmm1
movdqa xmm5,xmm7
paddw xmm1,xmm4 ; xmm1=data3+data4=tmp3
paddw xmm7,xmm0 ; xmm7=data2+data5=tmp2
psubw xmm2,xmm4 ; xmm2=data3-data4=tmp4
psubw xmm5,xmm0 ; xmm5=data2-data5=tmp5
psubw xmm11,xmm12 ; xmm11=data3-data4=tmp4
psubw xmm10,xmm13 ; xmm10=data2-data5=tmp5
psubw xmm9,xmm14 ; xmm9=data1-data6=tmp6
psubw xmm8,xmm15 ; xmm8=data0-data7=tmp7
; -- Even part
movdqa xmm4,xmm3
movdqa xmm0,xmm6
psubw xmm3,xmm1 ; xmm3=tmp13
psubw xmm6,xmm7 ; xmm6=tmp12
paddw xmm4,xmm1 ; xmm4=tmp10
paddw xmm0,xmm7 ; xmm0=tmp11
movdqa xmm4,xmm0
paddw xmm4,xmm3 ; xmm4=tmp0+tmp3=tmp10
movdqa xmm5,xmm1
paddw xmm5,xmm2 ; xmm5=tmp1+tmp2=tmp11
psubw xmm1,xmm2 ; xmm1=tmp1-tmp2=tmp12
psubw xmm0,xmm3 ; xmm0=tmp0-tmp3=tmp13
paddw xmm6,xmm3
psllw xmm6,PRE_MULTIPLY_SCALE_BITS
pmulhw xmm6,[rel PW_F0707] ; xmm6=z1
paddw xmm1,xmm0 ; xmm1=tmp12+tmp13
psllw xmm1,PRE_MULTIPLY_SCALE_BITS
pmulhw xmm1,[rel PW_F0707] ; xmm1=z1
movdqa xmm1,xmm4
movdqa xmm7,xmm3
psubw xmm4,xmm0 ; xmm4=data4
psubw xmm3,xmm6 ; xmm3=data6
paddw xmm1,xmm0 ; xmm1=data0
paddw xmm7,xmm6 ; xmm7=data2
movdqa xmm0, XMMWORD [wk(0)] ; xmm0=tmp6
movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp7
movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=data4
movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=data6
movdqa xmm12,xmm4
paddw xmm12,xmm5 ; xmm12=tmp10+tmp11=out0
movdqa xmm14,xmm0
paddw xmm14,xmm1 ; xmm14=tmp13+z1=out2
psubw xmm4,xmm5 ; xmm4=tmp10-tmp11=out4
psubw xmm0,xmm1 ; xmm0=tmp13-z1=out6
; -- Odd part
paddw xmm2,xmm5 ; xmm2=tmp10
paddw xmm5,xmm0 ; xmm5=tmp11
paddw xmm0,xmm6 ; xmm0=tmp12, xmm6=tmp7
paddw xmm11,xmm10 ; xmm11=tmp4+tmp5=tmp10
paddw xmm10,xmm9 ; xmm10=tmp5+tmp6=tmp11
paddw xmm9,xmm8 ; xmm9=tmp6+tmp7=tmp12
psllw xmm2,PRE_MULTIPLY_SCALE_BITS
psllw xmm0,PRE_MULTIPLY_SCALE_BITS
psllw xmm11,PRE_MULTIPLY_SCALE_BITS
psllw xmm10,PRE_MULTIPLY_SCALE_BITS
psllw xmm9,PRE_MULTIPLY_SCALE_BITS
psllw xmm5,PRE_MULTIPLY_SCALE_BITS
pmulhw xmm5,[rel PW_F0707] ; xmm5=z3
movdqa xmm5,xmm11
psubw xmm5,xmm9 ; xmm5=tmp10-tmp12
pmulhw xmm5,[rel PW_F0382] ; xmm5=z5
movdqa xmm4,xmm2 ; xmm4=tmp10
psubw xmm2,xmm0
pmulhw xmm2,[rel PW_F0382] ; xmm2=z5
pmulhw xmm4,[rel PW_F0541] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
pmulhw xmm0,[rel PW_F1306] ; xmm0=MULTIPLY(tmp12,FIX_1_306562)
paddw xmm4,xmm2 ; xmm4=z2
paddw xmm0,xmm2 ; xmm0=z4
pmulhw xmm11,[rel PW_F0541] ; xmm11=MULTIPLY(tmp10,FIX_0_541196)
paddw xmm11,xmm5 ; +z5=z2
movdqa xmm3,xmm6
psubw xmm6,xmm5 ; xmm6=z13
paddw xmm3,xmm5 ; xmm3=z11
pmulhw xmm9,[rel PW_F1306] ; xmm9=MULTIPLY(tmp12,FIX_1_306562)
paddw xmm9,xmm5 ; +z5=z4
movdqa xmm2,xmm6
movdqa xmm5,xmm3
psubw xmm6,xmm4 ; xmm6=data3
psubw xmm3,xmm0 ; xmm3=data7
paddw xmm2,xmm4 ; xmm2=data5
paddw xmm5,xmm0 ; xmm5=data1
pmulhw xmm10,[rel PW_F0707] ; xmm10=MULTIPLY(tmp11,FIX_1_306562)=z3
movdqa xmm1,xmm8
paddw xmm1,xmm10 ; xmm1=tmp7+z3=z11
psubw xmm8,xmm10 ; xmm8=tmp7-z3=z13
movdqa xmm13,xmm8
paddw xmm13,xmm11 ; xmm13=z13+z2=out5
movdqa xmm15,xmm1
psubw xmm15,xmm9 ; xmm15=z11-z4=out7
paddw xmm9,xmm1 ; xmm9=z11+z4=out1
psubw xmm8,xmm11 ; xmm8=z13-z2=out3
; ---- Pass 2: process columns.
; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72)
; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73)
; Re-order registers so we can reuse the same transpose code
movdqa xmm11,xmm8
movdqa xmm8,xmm12
movdqa xmm2,xmm14
movdqa xmm5,xmm0
movdqa xmm4,xmm1 ; transpose coefficients(phase 1)
punpcklwd xmm1,xmm5 ; xmm1=(00 01 10 11 20 21 30 31)
punpckhwd xmm4,xmm5 ; xmm4=(40 41 50 51 60 61 70 71)
movdqa xmm0,xmm7 ; transpose coefficients(phase 1)
punpcklwd xmm7,xmm6 ; xmm7=(02 03 12 13 22 23 32 33)
punpckhwd xmm0,xmm6 ; xmm0=(42 43 52 53 62 63 72 73)
; xmm8=(00 10 20 30 40 50 60 70), xmm9=(01 11 21 31 41 51 61 71)
; xmm2=(02 12 22 32 42 52 62 72), xmm11=(03 13 23 33 43 53 63 73)
; xmm4=(04 14 24 34 44 54 64 74), xmm13=(05 15 25 35 45 55 65 75)
; xmm5=(06 16 26 36 46 56 66 76), xmm15=(07 17 27 37 47 57 67 77)
movdqa xmm5, XMMWORD [wk(0)] ; xmm5=col4
movdqa xmm6, XMMWORD [wk(1)] ; xmm6=col6
movdqa xmm12,xmm8 ; transpose coefficients(phase 1)
punpcklwd xmm8,xmm9 ; xmm8=(00 01 10 11 20 21 30 31)
punpckhwd xmm12,xmm9 ; xmm12=(40 41 50 51 60 61 70 71)
movdqa xmm1,xmm2 ; transpose coefficients(phase 1)
punpcklwd xmm2,xmm11 ; xmm2=(02 03 12 13 22 23 32 33)
punpckhwd xmm1,xmm11 ; xmm1=(42 43 52 53 62 63 72 73)
; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76)
; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77)
movdqa xmm0,xmm4 ; transpose coefficients(phase 1)
punpcklwd xmm4,xmm13 ; xmm4=(04 05 14 15 24 25 34 35)
punpckhwd xmm0,xmm13 ; xmm0=(44 45 54 55 64 65 74 75)
movdqa xmm3,xmm5 ; transpose coefficients(phase 1)
punpcklwd xmm5,xmm15 ; xmm5=(06 07 16 17 26 27 36 37)
punpckhwd xmm3,xmm15 ; xmm3=(46 47 56 57 66 67 76 77)
movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(02 03 12 13 22 23 32 33)
movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=(42 43 52 53 62 63 72 73)
movdqa xmm10,xmm8 ; transpose coefficients(phase 2)
punpckldq xmm8,xmm2 ; xmm8=(00 01 02 03 10 11 12 13)
punpckhdq xmm10,xmm2 ; xmm10=(20 21 22 23 30 31 32 33)
movdqa xmm14,xmm12 ; transpose coefficients(phase 2)
punpckldq xmm12,xmm1 ; xmm12=(40 41 42 43 50 51 52 53)
punpckhdq xmm14,xmm1 ; xmm14=(60 61 62 63 70 71 72 73)
movdqa xmm7,xmm5 ; transpose coefficients(phase 1)
punpcklwd xmm5,xmm2 ; xmm5=(04 05 14 15 24 25 34 35)
punpckhwd xmm7,xmm2 ; xmm7=(44 45 54 55 64 65 74 75)
movdqa xmm0,xmm6 ; transpose coefficients(phase 1)
punpcklwd xmm6,xmm3 ; xmm6=(06 07 16 17 26 27 36 37)
punpckhwd xmm0,xmm3 ; xmm0=(46 47 56 57 66 67 76 77)
movdqa xmm6,xmm4 ; transpose coefficients(phase 2)
punpckldq xmm4,xmm5 ; xmm4=(04 05 06 07 14 15 16 17)
punpckhdq xmm6,xmm5 ; xmm6=(24 25 26 27 34 35 36 37)
movdqa xmm7,xmm0 ; transpose coefficients(phase 2)
punpckldq xmm0,xmm3 ; xmm0=(44 45 46 47 54 55 56 57)
punpckhdq xmm7,xmm3 ; xmm7=(64 65 66 67 74 75 76 77)
movdqa xmm2,xmm5 ; transpose coefficients(phase 2)
punpckldq xmm5,xmm6 ; xmm5=(04 05 06 07 14 15 16 17)
punpckhdq xmm2,xmm6 ; xmm2=(24 25 26 27 34 35 36 37)
movdqa xmm3,xmm7 ; transpose coefficients(phase 2)
punpckldq xmm7,xmm0 ; xmm7=(44 45 46 47 54 55 56 57)
punpckhdq xmm3,xmm0 ; xmm3=(64 65 66 67 74 75 76 77)
movdqa xmm9,xmm8 ; transpose coefficients(phase 3)
punpcklqdq xmm8,xmm4 ; xmm8=(00 01 02 03 04 05 06 07)=data0
punpckhqdq xmm9,xmm4 ; xmm9=(10 11 12 13 14 15 16 17)=data1
movdqa xmm11,xmm10 ; transpose coefficients(phase 3)
punpcklqdq xmm10,xmm6 ; xmm10=(20 21 22 23 24 25 26 27)=data2
punpckhqdq xmm11,xmm6 ; xmm11=(30 31 32 33 34 35 36 37)=data3
movdqa xmm6, XMMWORD [wk(0)] ; xmm6=(02 03 12 13 22 23 32 33)
movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(42 43 52 53 62 63 72 73)
movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(24 25 26 27 34 35 36 37)
movdqa XMMWORD [wk(1)], xmm7 ; wk(1)=(44 45 46 47 54 55 56 57)
movdqa xmm13,xmm12 ; transpose coefficients(phase 3)
punpcklqdq xmm12,xmm0 ; xmm12=(40 41 42 43 44 45 46 47)=data4
punpckhqdq xmm13,xmm0 ; xmm13=(50 51 52 53 54 55 56 57)=data5
movdqa xmm15,xmm14 ; transpose coefficients(phase 3)
punpcklqdq xmm14,xmm7 ; xmm14=(60 61 62 63 64 65 66 67)=data6
punpckhqdq xmm15,xmm7 ; xmm15=(70 71 72 73 74 75 76 77)=data7
movdqa xmm2,xmm1 ; transpose coefficients(phase 2)
punpckldq xmm1,xmm6 ; xmm1=(00 01 02 03 10 11 12 13)
punpckhdq xmm2,xmm6 ; xmm2=(20 21 22 23 30 31 32 33)
movdqa xmm7,xmm4 ; transpose coefficients(phase 2)
punpckldq xmm4,xmm0 ; xmm4=(40 41 42 43 50 51 52 53)
punpckhdq xmm7,xmm0 ; xmm7=(60 61 62 63 70 71 72 73)
movdqa xmm0,xmm8
paddw xmm0,xmm15 ; xmm0=data0+data7=tmp0
movdqa xmm1,xmm9
paddw xmm1,xmm14 ; xmm1=data1+data6=tmp1
movdqa xmm2,xmm10
paddw xmm2,xmm13 ; xmm2=data2+data5=tmp2
movdqa xmm3,xmm11
paddw xmm3,xmm12 ; xmm3=data3+data4=tmp3
movdqa xmm6,xmm1 ; transpose coefficients(phase 3)
punpcklqdq xmm1,xmm5 ; xmm1=(00 01 02 03 04 05 06 07)=data0
punpckhqdq xmm6,xmm5 ; xmm6=(10 11 12 13 14 15 16 17)=data1
movdqa xmm0,xmm7 ; transpose coefficients(phase 3)
punpcklqdq xmm7,xmm3 ; xmm7=(60 61 62 63 64 65 66 67)=data6
punpckhqdq xmm0,xmm3 ; xmm0=(70 71 72 73 74 75 76 77)=data7
movdqa xmm5,xmm6
movdqa xmm3,xmm1
psubw xmm6,xmm7 ; xmm6=data1-data6=tmp6
psubw xmm1,xmm0 ; xmm1=data0-data7=tmp7
paddw xmm5,xmm7 ; xmm5=data1+data6=tmp1
paddw xmm3,xmm0 ; xmm3=data0+data7=tmp0
movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(24 25 26 27 34 35 36 37)
movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(44 45 46 47 54 55 56 57)
movdqa XMMWORD [wk(0)], xmm6 ; wk(0)=tmp6
movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=tmp7
movdqa xmm6,xmm2 ; transpose coefficients(phase 3)
punpcklqdq xmm2,xmm7 ; xmm2=(20 21 22 23 24 25 26 27)=data2
punpckhqdq xmm6,xmm7 ; xmm6=(30 31 32 33 34 35 36 37)=data3
movdqa xmm1,xmm4 ; transpose coefficients(phase 3)
punpcklqdq xmm4,xmm0 ; xmm4=(40 41 42 43 44 45 46 47)=data4
punpckhqdq xmm1,xmm0 ; xmm1=(50 51 52 53 54 55 56 57)=data5
movdqa xmm7,xmm6
movdqa xmm0,xmm2
paddw xmm6,xmm4 ; xmm6=data3+data4=tmp3
paddw xmm2,xmm1 ; xmm2=data2+data5=tmp2
psubw xmm7,xmm4 ; xmm7=data3-data4=tmp4
psubw xmm0,xmm1 ; xmm0=data2-data5=tmp5
psubw xmm11,xmm12 ; xmm11=data3-data4=tmp4
psubw xmm10,xmm13 ; xmm10=data2-data5=tmp5
psubw xmm9,xmm14 ; xmm9=data1-data6=tmp6
psubw xmm8,xmm15 ; xmm8=data0-data7=tmp7
; -- Even part
movdqa xmm4,xmm3
movdqa xmm1,xmm5
psubw xmm3,xmm6 ; xmm3=tmp13
psubw xmm5,xmm2 ; xmm5=tmp12
paddw xmm4,xmm6 ; xmm4=tmp10
paddw xmm1,xmm2 ; xmm1=tmp11
movdqa xmm4,xmm0
paddw xmm4,xmm3 ; xmm4=tmp0+tmp3=tmp10
movdqa xmm5,xmm1
paddw xmm5,xmm2 ; xmm5=tmp1+tmp2=tmp11
psubw xmm1,xmm2 ; xmm1=tmp1-tmp2=tmp12
psubw xmm0,xmm3 ; xmm0=tmp0-tmp3=tmp13
paddw xmm5,xmm3
psllw xmm5,PRE_MULTIPLY_SCALE_BITS
pmulhw xmm5,[rel PW_F0707] ; xmm5=z1
paddw xmm1,xmm0 ; xmm1=tmp12+tmp13
psllw xmm1,PRE_MULTIPLY_SCALE_BITS
pmulhw xmm1,[rel PW_F0707] ; xmm1=z1
movdqa xmm6,xmm4
movdqa xmm2,xmm3
psubw xmm4,xmm1 ; xmm4=data4
psubw xmm3,xmm5 ; xmm3=data6
paddw xmm6,xmm1 ; xmm6=data0
paddw xmm2,xmm5 ; xmm2=data2
movdqa XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm4
movdqa XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm3
movdqa XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm6
movdqa XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm2
movdqa xmm12,xmm4
paddw xmm12,xmm5 ; xmm12=tmp10+tmp11=out0
movdqa xmm14,xmm0
paddw xmm14,xmm1 ; xmm14=tmp13+z1=out2
psubw xmm4,xmm5 ; xmm4=tmp10-tmp11=out4
psubw xmm0,xmm1 ; xmm0=tmp13-z1=out6
; -- Odd part
movdqa xmm1, XMMWORD [wk(0)] ; xmm1=tmp6
movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7
paddw xmm11,xmm10 ; xmm11=tmp4+tmp5=tmp10
paddw xmm10,xmm9 ; xmm10=tmp5+tmp6=tmp11
paddw xmm9,xmm8 ; xmm9=tmp6+tmp7=tmp12
paddw xmm7,xmm0 ; xmm7=tmp10
paddw xmm0,xmm1 ; xmm0=tmp11
paddw xmm1,xmm5 ; xmm1=tmp12, xmm5=tmp7
psllw xmm11,PRE_MULTIPLY_SCALE_BITS
psllw xmm10,PRE_MULTIPLY_SCALE_BITS
psllw xmm9,PRE_MULTIPLY_SCALE_BITS
psllw xmm7,PRE_MULTIPLY_SCALE_BITS
psllw xmm1,PRE_MULTIPLY_SCALE_BITS
movdqa xmm5,xmm11
psubw xmm5,xmm9 ; xmm5=tmp10-tmp12
pmulhw xmm5,[rel PW_F0382] ; xmm5=z5
psllw xmm0,PRE_MULTIPLY_SCALE_BITS
pmulhw xmm0,[rel PW_F0707] ; xmm0=z3
pmulhw xmm11,[rel PW_F0541] ; xmm11=MULTIPLY(tmp10,FIX_0_541196)
paddw xmm11,xmm5 ; +z5=z2
movdqa xmm4,xmm7 ; xmm4=tmp10
psubw xmm7,xmm1
pmulhw xmm7,[rel PW_F0382] ; xmm7=z5
pmulhw xmm4,[rel PW_F0541] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
pmulhw xmm1,[rel PW_F1306] ; xmm1=MULTIPLY(tmp12,FIX_1_306562)
paddw xmm4,xmm7 ; xmm4=z2
paddw xmm1,xmm7 ; xmm1=z4
pmulhw xmm9,[rel PW_F1306] ; xmm9=MULTIPLY(tmp12,FIX_1_306562)
paddw xmm9,xmm5 ; +z5=z4
movdqa xmm3,xmm5
psubw xmm5,xmm0 ; xmm5=z13
paddw xmm3,xmm0 ; xmm3=z11
pmulhw xmm10,[rel PW_F0707] ; xmm10=MULTIPLY(tmp11,FIX_1_306562)=z3
movdqa xmm6,xmm5
movdqa xmm2,xmm3
psubw xmm5,xmm4 ; xmm5=data3
psubw xmm3,xmm1 ; xmm3=data7
paddw xmm6,xmm4 ; xmm6=data5
paddw xmm2,xmm1 ; xmm2=data1
movdqa xmm1,xmm8
paddw xmm1,xmm10 ; xmm1=tmp7+z3=z11
psubw xmm8,xmm10 ; xmm8=tmp7-z3=z13
movdqa XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm5
movdqa XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm3
movdqa XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm6
movdqa XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm2
movdqa xmm13,xmm8
paddw xmm13,xmm11 ; xmm13=z13+z2=out5
movdqa xmm15,xmm1
psubw xmm15,xmm9 ; xmm15=z11-z4=out7
paddw xmm9,xmm1 ; xmm9=z11+z4=out1
psubw xmm8,xmm11 ; xmm8=z13-z2=out3
; -- Write result
movdqa XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm12
movdqa XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm9
movdqa XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm14
movdqa XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm8
movdqa XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm4
movdqa XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm13
movdqa XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm0
movdqa XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm15
uncollect_args
mov rsp,rbp ; rsp <- aligned rbp
pop rsp ; rsp <- original rbp
pop rbp
ret
; For some reason, the OS X linker does not honor the request to align the

651
simd/jfdctint-sse2-64.asm
View File

@@ -2,7 +2,7 @@
; jfdctint.asm - accurate integer FDCT (64-bit SSE2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright 2009 D. R. Commander
; Copyright 2009, 2014, D. R. Commander
;
; Based on
; x86 SIMD extension for IJG JPEG library
@@ -97,174 +97,140 @@ PW_DESCALE_P2X times 8 dw 1 << (PASS1_BITS-1)
; r10 = DCTELEM * data
%define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
%define WK_NUM 6
align 16
global EXTN(jsimd_fdct_islow_sse2)
EXTN(jsimd_fdct_islow_sse2):
push rbp
mov rax,rsp ; rax = original rbp
sub rsp, byte 4
and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
mov [rsp],rax
mov rbp,rsp ; rbp = aligned rbp
lea rsp, [wk(0)]
collect_args
; ---- Pass 1: process rows.
mov rdx, r10 ; (DCTELEM *)
movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm8, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm9, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm11, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm4, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm13, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm5, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm15, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
; xmm8=(00 01 02 03 04 05 06 07), xmm9=(10 11 12 13 14 15 16 17)
; xmm2=(20 21 22 23 24 25 26 27), xmm11=(30 31 32 33 34 35 36 37)
; xmm4=(40 41 42 43 44 45 46 47), xmm13=(50 51 52 53 54 55 56 57)
; xmm5=(60 61 62 63 64 65 66 67), xmm15=(70 71 72 73 74 75 76 77)
movdqa xmm4,xmm0 ; transpose coefficients(phase 1)
punpcklwd xmm0,xmm1 ; xmm0=(00 10 01 11 02 12 03 13)
punpckhwd xmm4,xmm1 ; xmm4=(04 14 05 15 06 16 07 17)
movdqa xmm5,xmm2 ; transpose coefficients(phase 1)
punpcklwd xmm2,xmm3 ; xmm2=(20 30 21 31 22 32 23 33)
punpckhwd xmm5,xmm3 ; xmm5=(24 34 25 35 26 36 27 37)
movdqa xmm12,xmm8 ; transpose coefficients(phase 1)
punpcklwd xmm8,xmm9 ; xmm8=(00 10 01 11 02 12 03 13)
punpckhwd xmm12,xmm9 ; xmm12=(04 14 05 15 06 16 07 17)
movdqa xmm1,xmm2 ; transpose coefficients(phase 1)
punpcklwd xmm2,xmm11 ; xmm2=(20 30 21 31 22 32 23 33)
punpckhwd xmm1,xmm11 ; xmm1=(24 34 25 35 26 36 27 37)
movdqa xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
movdqa xmm0,xmm4 ; transpose coefficients(phase 1)
punpcklwd xmm4,xmm13 ; xmm4=(40 50 41 51 42 52 43 53)
punpckhwd xmm0,xmm13 ; xmm0=(44 54 45 55 46 56 47 57)
movdqa xmm3,xmm5 ; transpose coefficients(phase 1)
punpcklwd xmm5,xmm15 ; xmm5=(60 70 61 71 62 72 63 73)
punpckhwd xmm3,xmm15 ; xmm3=(64 74 65 75 66 76 67 77)
; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33)
movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37)
movdqa xmm2,xmm6 ; transpose coefficients(phase 1)
punpcklwd xmm6,xmm7 ; xmm6=(40 50 41 51 42 52 43 53)
punpckhwd xmm2,xmm7 ; xmm2=(44 54 45 55 46 56 47 57)
movdqa xmm5,xmm1 ; transpose coefficients(phase 1)
punpcklwd xmm1,xmm3 ; xmm1=(60 70 61 71 62 72 63 73)
punpckhwd xmm5,xmm3 ; xmm5=(64 74 65 75 66 76 67 77)
movdqa xmm7,xmm6 ; transpose coefficients(phase 2)
punpckldq xmm6,xmm1 ; xmm6=(40 50 60 70 41 51 61 71)
punpckhdq xmm7,xmm1 ; xmm7=(42 52 62 72 43 53 63 73)
movdqa xmm3,xmm2 ; transpose coefficients(phase 2)
punpckldq xmm2,xmm5 ; xmm2=(44 54 64 74 45 55 65 75)
punpckhdq xmm3,xmm5 ; xmm3=(46 56 66 76 47 57 67 77)
movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33)
movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37)
movdqa XMMWORD [wk(2)], xmm7 ; wk(2)=(42 52 62 72 43 53 63 73)
movdqa XMMWORD [wk(3)], xmm2 ; wk(3)=(44 54 64 74 45 55 65 75)
movdqa xmm10,xmm8 ; transpose coefficients(phase 2)
punpckldq xmm8,xmm2 ; xmm8=(00 10 20 30 01 11 21 31)
punpckhdq xmm10,xmm2 ; xmm10=(02 12 22 32 03 13 23 33)
movdqa xmm14,xmm12 ; transpose coefficients(phase 2)
punpckldq xmm12,xmm1 ; xmm12=(04 14 24 34 05 15 25 35)
punpckhdq xmm14,xmm1 ; xmm14=(06 16 26 36 07 17 27 37)
movdqa xmm6,xmm4 ; transpose coefficients(phase 2)
punpckldq xmm4,xmm5 ; xmm4=(40 50 60 70 41 51 61 71)
punpckhdq xmm6,xmm5 ; xmm6=(42 52 62 72 43 53 63 73)
movdqa xmm7,xmm0 ; transpose coefficients(phase 2)
punpckldq xmm0,xmm1 ; xmm0=(00 10 20 30 01 11 21 31)
punpckhdq xmm7,xmm1 ; xmm7=(02 12 22 32 03 13 23 33)
movdqa xmm2,xmm4 ; transpose coefficients(phase 2)
punpckldq xmm4,xmm5 ; xmm4=(04 14 24 34 05 15 25 35)
punpckhdq xmm2,xmm5 ; xmm2=(06 16 26 36 07 17 27 37)
punpckldq xmm0,xmm3 ; xmm0=(44 54 64 74 45 55 65 75)
punpckhdq xmm7,xmm3 ; xmm7=(46 56 66 76 47 57 67 77)
movdqa xmm1,xmm0 ; transpose coefficients(phase 3)
punpcklqdq xmm0,xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0
punpckhqdq xmm1,xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1
movdqa xmm5,xmm2 ; transpose coefficients(phase 3)
punpcklqdq xmm2,xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6
punpckhqdq xmm5,xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7
movdqa xmm9,xmm8 ; transpose coefficients(phase 3)
punpcklqdq xmm8,xmm4 ; xmm8=(00 10 20 30 40 50 60 70)=data0
punpckhqdq xmm9,xmm4 ; xmm9=(01 11 21 31 41 51 61 71)=data1
movdqa xmm11,xmm10 ; transpose coefficients(phase 3)
punpcklqdq xmm10,xmm6 ; xmm10=(02 12 22 32 42 52 62 72)=data2
punpckhqdq xmm11,xmm6 ; xmm11=(03 13 23 33 43 53 63 73)=data3
movdqa xmm6,xmm1
movdqa xmm3,xmm0
psubw xmm1,xmm2 ; xmm1=data1-data6=tmp6
psubw xmm0,xmm5 ; xmm0=data0-data7=tmp7
paddw xmm6,xmm2 ; xmm6=data1+data6=tmp1
paddw xmm3,xmm5 ; xmm3=data0+data7=tmp0
movdqa xmm13,xmm12 ; transpose coefficients(phase 3)
punpcklqdq xmm12,xmm0 ; xmm12=(04 14 24 34 44 54 64 74)=data4
punpckhqdq xmm13,xmm0 ; xmm13=(05 15 25 35 45 55 65 75)=data5
movdqa xmm15,xmm14 ; transpose coefficients(phase 3)
punpcklqdq xmm14,xmm7 ; xmm14=(06 16 26 36 46 56 66 76)=data6
punpckhqdq xmm15,xmm7 ; xmm15=(07 17 27 37 47 57 67 77)=data7
movdqa xmm2, XMMWORD [wk(2)] ; xmm2=(42 52 62 72 43 53 63 73)
movdqa xmm5, XMMWORD [wk(3)] ; xmm5=(44 54 64 74 45 55 65 75)
movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6
movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7
movdqa xmm0,xmm8
paddw xmm0,xmm15 ; xmm0=data0+data7=tmp0
movdqa xmm1,xmm9
paddw xmm1,xmm14 ; xmm1=data1+data6=tmp1
movdqa xmm2,xmm10
paddw xmm2,xmm13 ; xmm2=data2+data5=tmp2
movdqa xmm3,xmm11
paddw xmm3,xmm12 ; xmm3=data3+data4=tmp3
movdqa xmm1,xmm7 ; transpose coefficients(phase 3)
punpcklqdq xmm7,xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2
punpckhqdq xmm1,xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3
movdqa xmm0,xmm4 ; transpose coefficients(phase 3)
punpcklqdq xmm4,xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4
punpckhqdq xmm0,xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5
movdqa xmm2,xmm1
movdqa xmm5,xmm7
paddw xmm1,xmm4 ; xmm1=data3+data4=tmp3
paddw xmm7,xmm0 ; xmm7=data2+data5=tmp2
psubw xmm2,xmm4 ; xmm2=data3-data4=tmp4
psubw xmm5,xmm0 ; xmm5=data2-data5=tmp5
psubw xmm11,xmm12 ; xmm11=data3-data4=tmp4
psubw xmm10,xmm13 ; xmm10=data2-data5=tmp5
psubw xmm9,xmm14 ; xmm9=data1-data6=tmp6
psubw xmm8,xmm15 ; xmm8=data0-data7=tmp7
; -- Even part
movdqa xmm4,xmm3
movdqa xmm0,xmm6
paddw xmm3,xmm1 ; xmm3=tmp10
paddw xmm6,xmm7 ; xmm6=tmp11
psubw xmm4,xmm1 ; xmm4=tmp13
psubw xmm0,xmm7 ; xmm0=tmp12
movdqa xmm4,xmm0
paddw xmm4,xmm3 ; xmm4=tmp0+tmp3=tmp10
movdqa xmm5,xmm1
paddw xmm5,xmm2 ; xmm5=tmp1+tmp2=tmp11
psubw xmm1,xmm2 ; xmm1=tmp1-tmp2=tmp12
psubw xmm0,xmm3 ; xmm0=tmp0-tmp3=tmp13
movdqa xmm1,xmm3
paddw xmm3,xmm6 ; xmm3=tmp10+tmp11
psubw xmm1,xmm6 ; xmm1=tmp10-tmp11
movdqa xmm12,xmm4
paddw xmm12,xmm5 ; xmm12=tmp10+tmp11
psubw xmm4,xmm5 ; xmm4=tmp10-tmp11
psllw xmm3,PASS1_BITS ; xmm3=data0
psllw xmm1,PASS1_BITS ; xmm1=data4
movdqa XMMWORD [wk(2)], xmm3 ; wk(2)=data0
movdqa XMMWORD [wk(3)], xmm1 ; wk(3)=data4
psllw xmm12,PASS1_BITS ; xmm12=out0
psllw xmm4,PASS1_BITS ; xmm4=out4
; (Original)
; z1 = (tmp12 + tmp13) * 0.541196100;
; data2 = z1 + tmp13 * 0.765366865;
; data6 = z1 + tmp12 * -1.847759065;
; out2 = z1 + tmp13 * 0.765366865;
; out6 = z1 + tmp12 * -1.847759065;
;
; (This implementation)
; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
; out2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
; out6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
movdqa xmm7,xmm4 ; xmm4=tmp13
movdqa xmm6,xmm4
punpcklwd xmm7,xmm0 ; xmm0=tmp12
punpckhwd xmm6,xmm0
movdqa xmm4,xmm7
movdqa xmm2,xmm0 ; xmm0=tmp13
movdqa xmm6,xmm0
punpcklwd xmm2,xmm1 ; xmm1=tmp12
punpckhwd xmm6,xmm1
movdqa xmm5,xmm2
movdqa xmm0,xmm6
pmaddwd xmm7,[rel PW_F130_F054] ; xmm7=data2L
pmaddwd xmm6,[rel PW_F130_F054] ; xmm6=data2H
pmaddwd xmm4,[rel PW_F054_MF130] ; xmm4=data6L
pmaddwd xmm0,[rel PW_F054_MF130] ; xmm0=data6H
pmaddwd xmm2,[rel PW_F130_F054] ; xmm2=out2L
pmaddwd xmm6,[rel PW_F130_F054] ; xmm6=out2H
pmaddwd xmm5,[rel PW_F054_MF130] ; xmm5=out6L
pmaddwd xmm0,[rel PW_F054_MF130] ; xmm0=out6H
paddd xmm7,[rel PD_DESCALE_P1]
paddd xmm2,[rel PD_DESCALE_P1]
paddd xmm6,[rel PD_DESCALE_P1]
psrad xmm7,DESCALE_P1
psrad xmm2,DESCALE_P1
psrad xmm6,DESCALE_P1
paddd xmm4,[rel PD_DESCALE_P1]
paddd xmm5,[rel PD_DESCALE_P1]
paddd xmm0,[rel PD_DESCALE_P1]
psrad xmm4,DESCALE_P1
psrad xmm5,DESCALE_P1
psrad xmm0,DESCALE_P1
packssdw xmm7,xmm6 ; xmm7=data2
packssdw xmm4,xmm0 ; xmm4=data6
movdqa XMMWORD [wk(4)], xmm7 ; wk(4)=data2
movdqa XMMWORD [wk(5)], xmm4 ; wk(5)=data6
packssdw xmm2,xmm6 ; xmm2=out2
packssdw xmm5,xmm0 ; xmm5=out6
; -- Odd part
movdqa xmm3, XMMWORD [wk(0)] ; xmm3=tmp6
movdqa xmm1, XMMWORD [wk(1)] ; xmm1=tmp7
movdqa xmm6,xmm2 ; xmm2=tmp4
movdqa xmm0,xmm5 ; xmm5=tmp5
paddw xmm6,xmm3 ; xmm6=z3
paddw xmm0,xmm1 ; xmm0=z4
movdqa xmm6,xmm11 ; xmm11=tmp4
movdqa xmm0,xmm10 ; xmm10=tmp5
paddw xmm6,xmm9 ; xmm6=tmp4+tmp6=z3
paddw xmm0,xmm8 ; xmm0=tmp5+tmp7=z4
; (Original)
; z5 = (z3 + z4) * 1.175875602;
@@ -276,239 +242,211 @@ EXTN(jsimd_fdct_islow_sse2):
; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
movdqa xmm7,xmm6
movdqa xmm4,xmm6
movdqa xmm14,xmm6
punpcklwd xmm7,xmm0
punpckhwd xmm4,xmm0
punpckhwd xmm14,xmm0
movdqa xmm6,xmm7
movdqa xmm0,xmm4
movdqa xmm0,xmm14
pmaddwd xmm7,[rel PW_MF078_F117] ; xmm7=z3L
pmaddwd xmm4,[rel PW_MF078_F117] ; xmm4=z3H
pmaddwd xmm14,[rel PW_MF078_F117] ; xmm14=z3H
pmaddwd xmm6,[rel PW_F117_F078] ; xmm6=z4L
pmaddwd xmm0,[rel PW_F117_F078] ; xmm0=z4H
movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=z3L
movdqa XMMWORD [wk(1)], xmm4 ; wk(1)=z3H
; (Original)
; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
; out7 = tmp4 + z1 + z3; out5 = tmp5 + z2 + z4;
; out3 = tmp6 + z2 + z3; out1 = tmp7 + z1 + z4;
;
; (This implementation)
; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
; data7 = tmp4 + z3; data5 = tmp5 + z4;
; data3 = tmp6 + z3; data1 = tmp7 + z4;
; out7 = tmp4 + z3; out5 = tmp5 + z4;
; out3 = tmp6 + z3; out1 = tmp7 + z4;
movdqa xmm7,xmm2
movdqa xmm4,xmm2
punpcklwd xmm7,xmm1
punpckhwd xmm4,xmm1
movdqa xmm2,xmm7
movdqa xmm1,xmm4
pmaddwd xmm7,[rel PW_MF060_MF089] ; xmm7=tmp4L
pmaddwd xmm4,[rel PW_MF060_MF089] ; xmm4=tmp4H
pmaddwd xmm2,[rel PW_MF089_F060] ; xmm2=tmp7L
movdqa xmm15,xmm11 ; xmm11=tmp4
movdqa xmm13,xmm11
punpcklwd xmm15,xmm8 ; xmm8=tmp7
punpckhwd xmm13,xmm8
movdqa xmm8,xmm15
movdqa xmm1,xmm13
pmaddwd xmm15,[rel PW_MF060_MF089] ; xmm15=tmp4L
pmaddwd xmm13,[rel PW_MF060_MF089] ; xmm13=tmp4H
pmaddwd xmm8,[rel PW_MF089_F060] ; xmm8=tmp7L
pmaddwd xmm1,[rel PW_MF089_F060] ; xmm1=tmp7H
paddd xmm7, XMMWORD [wk(0)] ; xmm7=data7L
paddd xmm4, XMMWORD [wk(1)] ; xmm4=data7H
paddd xmm2,xmm6 ; xmm2=data1L
paddd xmm1,xmm0 ; xmm1=data1H
paddd xmm15,xmm7 ; xmm15=out7L
paddd xmm13,xmm14 ; xmm13=out7H
paddd xmm8,xmm6 ; xmm8=out1L
paddd xmm1,xmm0 ; xmm1=out1H
paddd xmm7,[rel PD_DESCALE_P1]
paddd xmm4,[rel PD_DESCALE_P1]
psrad xmm7,DESCALE_P1
psrad xmm4,DESCALE_P1
paddd xmm2,[rel PD_DESCALE_P1]
paddd xmm15,[rel PD_DESCALE_P1]
paddd xmm13,[rel PD_DESCALE_P1]
psrad xmm15,DESCALE_P1
psrad xmm13,DESCALE_P1
paddd xmm8,[rel PD_DESCALE_P1]
paddd xmm1,[rel PD_DESCALE_P1]
psrad xmm2,DESCALE_P1
psrad xmm8,DESCALE_P1
psrad xmm1,DESCALE_P1
packssdw xmm7,xmm4 ; xmm7=data7
packssdw xmm2,xmm1 ; xmm2=data1
packssdw xmm15,xmm13 ; xmm15=out7
packssdw xmm8,xmm1 ; xmm8=out1
movdqa xmm4,xmm5
movdqa xmm1,xmm5
punpcklwd xmm4,xmm3
punpckhwd xmm1,xmm3
movdqa xmm5,xmm4
movdqa xmm13,xmm10 ; xmm10=tmp5
movdqa xmm1,xmm10
punpcklwd xmm13,xmm9 ; xmm9=tmp6
punpckhwd xmm1,xmm9
movdqa xmm11,xmm13
movdqa xmm3,xmm1
pmaddwd xmm4,[rel PW_MF050_MF256] ; xmm4=tmp5L
pmaddwd xmm13,[rel PW_MF050_MF256] ; xmm13=tmp5L
pmaddwd xmm1,[rel PW_MF050_MF256] ; xmm1=tmp5H
pmaddwd xmm5,[rel PW_MF256_F050] ; xmm5=tmp6L
pmaddwd xmm11,[rel PW_MF256_F050] ; xmm11=tmp6L
pmaddwd xmm3,[rel PW_MF256_F050] ; xmm3=tmp6H
paddd xmm4,xmm6 ; xmm4=data5L
paddd xmm1,xmm0 ; xmm1=data5H
paddd xmm5, XMMWORD [wk(0)] ; xmm5=data3L
paddd xmm3, XMMWORD [wk(1)] ; xmm3=data3H
paddd xmm13,xmm6 ; xmm13=out5L
paddd xmm1,xmm0 ; xmm1=out5H
paddd xmm11,xmm7 ; xmm11=out3L
paddd xmm3,xmm14 ; xmm3=out3H
paddd xmm4,[rel PD_DESCALE_P1]
paddd xmm13,[rel PD_DESCALE_P1]
paddd xmm1,[rel PD_DESCALE_P1]
psrad xmm4,DESCALE_P1
psrad xmm13,DESCALE_P1
psrad xmm1,DESCALE_P1
paddd xmm5,[rel PD_DESCALE_P1]
paddd xmm11,[rel PD_DESCALE_P1]
paddd xmm3,[rel PD_DESCALE_P1]
psrad xmm5,DESCALE_P1
psrad xmm11,DESCALE_P1
psrad xmm3,DESCALE_P1
packssdw xmm4,xmm1 ; xmm4=data5
packssdw xmm5,xmm3 ; xmm5=data3
packssdw xmm13,xmm1 ; xmm13=out5
packssdw xmm11,xmm3 ; xmm11=out3
; ---- Pass 2: process columns.
movdqa xmm6, XMMWORD [wk(2)] ; xmm6=col0
movdqa xmm0, XMMWORD [wk(4)] ; xmm0=col2
; Re-order registers so we can reuse the same transpose code
movdqa xmm9,xmm8
movdqa xmm8,xmm12
; xmm6=(00 10 20 30 40 50 60 70), xmm0=(02 12 22 32 42 52 62 72)
; xmm2=(01 11 21 31 41 51 61 71), xmm5=(03 13 23 33 43 53 63 73)
; xmm8=(00 10 20 30 40 50 60 70), xmm9=(01 11 21 31 41 51 61 71)
; xmm2=(02 12 22 32 42 52 62 72), xmm11=(03 13 23 33 43 53 63 73)
; xmm4=(04 14 24 34 44 54 64 74), xmm13=(05 15 25 35 45 55 65 75)
; xmm5=(06 16 26 36 46 56 66 76), xmm15=(07 17 27 37 47 57 67 77)
movdqa xmm1,xmm6 ; transpose coefficients(phase 1)
punpcklwd xmm6,xmm2 ; xmm6=(00 01 10 11 20 21 30 31)
punpckhwd xmm1,xmm2 ; xmm1=(40 41 50 51 60 61 70 71)
movdqa xmm3,xmm0 ; transpose coefficients(phase 1)
punpcklwd xmm0,xmm5 ; xmm0=(02 03 12 13 22 23 32 33)
punpckhwd xmm3,xmm5 ; xmm3=(42 43 52 53 62 63 72 73)
movdqa xmm12,xmm8 ; transpose coefficients(phase 1)
punpcklwd xmm8,xmm9 ; xmm8=(00 01 10 11 20 21 30 31)
punpckhwd xmm12,xmm9 ; xmm12=(40 41 50 51 60 61 70 71)
movdqa xmm1,xmm2 ; transpose coefficients(phase 1)
punpcklwd xmm2,xmm11 ; xmm2=(02 03 12 13 22 23 32 33)
punpckhwd xmm1,xmm11 ; xmm1=(42 43 52 53 62 63 72 73)
movdqa xmm2, XMMWORD [wk(3)] ; xmm2=col4
movdqa xmm5, XMMWORD [wk(5)] ; xmm5=col6
; xmm2=(04 14 24 34 44 54 64 74), xmm5=(06 16 26 36 46 56 66 76)
; xmm4=(05 15 25 35 45 55 65 75), xmm7=(07 17 27 37 47 57 67 77)
movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=(02 03 12 13 22 23 32 33)
movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=(42 43 52 53 62 63 72 73)
movdqa xmm0,xmm2 ; transpose coefficients(phase 1)
punpcklwd xmm2,xmm4 ; xmm2=(04 05 14 15 24 25 34 35)
punpckhwd xmm0,xmm4 ; xmm0=(44 45 54 55 64 65 74 75)
movdqa xmm0,xmm4 ; transpose coefficients(phase 1)
punpcklwd xmm4,xmm13 ; xmm4=(04 05 14 15 24 25 34 35)
punpckhwd xmm0,xmm13 ; xmm0=(44 45 54 55 64 65 74 75)
movdqa xmm3,xmm5 ; transpose coefficients(phase 1)
punpcklwd xmm5,xmm7 ; xmm5=(06 07 16 17 26 27 36 37)
punpckhwd xmm3,xmm7 ; xmm3=(46 47 56 57 66 67 76 77)
punpcklwd xmm5,xmm15 ; xmm5=(06 07 16 17 26 27 36 37)
punpckhwd xmm3,xmm15 ; xmm3=(46 47 56 57 66 67 76 77)
movdqa xmm4,xmm2 ; transpose coefficients(phase 2)
punpckldq xmm2,xmm5 ; xmm2=(04 05 06 07 14 15 16 17)
punpckhdq xmm4,xmm5 ; xmm4=(24 25 26 27 34 35 36 37)
movdqa xmm10,xmm8 ; transpose coefficients(phase 2)
punpckldq xmm8,xmm2 ; xmm8=(00 01 02 03 10 11 12 13)
punpckhdq xmm10,xmm2 ; xmm10=(20 21 22 23 30 31 32 33)
movdqa xmm14,xmm12 ; transpose coefficients(phase 2)
punpckldq xmm12,xmm1 ; xmm12=(40 41 42 43 50 51 52 53)
punpckhdq xmm14,xmm1 ; xmm14=(60 61 62 63 70 71 72 73)
movdqa xmm6,xmm4 ; transpose coefficients(phase 2)
punpckldq xmm4,xmm5 ; xmm4=(04 05 06 07 14 15 16 17)
punpckhdq xmm6,xmm5 ; xmm6=(24 25 26 27 34 35 36 37)
movdqa xmm7,xmm0 ; transpose coefficients(phase 2)
punpckldq xmm0,xmm3 ; xmm0=(44 45 46 47 54 55 56 57)
punpckhdq xmm7,xmm3 ; xmm7=(64 65 66 67 74 75 76 77)
movdqa xmm5, XMMWORD [wk(0)] ; xmm5=(02 03 12 13 22 23 32 33)
movdqa xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53 62 63 72 73)
movdqa XMMWORD [wk(2)], xmm4 ; wk(2)=(24 25 26 27 34 35 36 37)
movdqa XMMWORD [wk(3)], xmm0 ; wk(3)=(44 45 46 47 54 55 56 57)
movdqa xmm9,xmm8 ; transpose coefficients(phase 3)
punpcklqdq xmm8,xmm4 ; xmm8=(00 01 02 03 04 05 06 07)=data0
punpckhqdq xmm9,xmm4 ; xmm9=(10 11 12 13 14 15 16 17)=data1
movdqa xmm11,xmm10 ; transpose coefficients(phase 3)
punpcklqdq xmm10,xmm6 ; xmm10=(20 21 22 23 24 25 26 27)=data2
punpckhqdq xmm11,xmm6 ; xmm11=(30 31 32 33 34 35 36 37)=data3
movdqa xmm4,xmm6 ; transpose coefficients(phase 2)
punpckldq xmm6,xmm5 ; xmm6=(00 01 02 03 10 11 12 13)
punpckhdq xmm4,xmm5 ; xmm4=(20 21 22 23 30 31 32 33)
movdqa xmm0,xmm1 ; transpose coefficients(phase 2)
punpckldq xmm1,xmm3 ; xmm1=(40 41 42 43 50 51 52 53)
punpckhdq xmm0,xmm3 ; xmm0=(60 61 62 63 70 71 72 73)
movdqa xmm13,xmm12 ; transpose coefficients(phase 3)
punpcklqdq xmm12,xmm0 ; xmm12=(40 41 42 43 44 45 46 47)=data4
punpckhqdq xmm13,xmm0 ; xmm13=(50 51 52 53 54 55 56 57)=data5
movdqa xmm15,xmm14 ; transpose coefficients(phase 3)
punpcklqdq xmm14,xmm7 ; xmm14=(60 61 62 63 64 65 66 67)=data6
punpckhqdq xmm15,xmm7 ; xmm15=(70 71 72 73 74 75 76 77)=data7
movdqa xmm5,xmm6 ; transpose coefficients(phase 3)
punpcklqdq xmm6,xmm2 ; xmm6=(00 01 02 03 04 05 06 07)=data0
punpckhqdq xmm5,xmm2 ; xmm5=(10 11 12 13 14 15 16 17)=data1
movdqa xmm3,xmm0 ; transpose coefficients(phase 3)
punpcklqdq xmm0,xmm7 ; xmm0=(60 61 62 63 64 65 66 67)=data6
punpckhqdq xmm3,xmm7 ; xmm3=(70 71 72 73 74 75 76 77)=data7
movdqa xmm0,xmm8
paddw xmm0,xmm15 ; xmm0=data0+data7=tmp0
movdqa xmm1,xmm9
paddw xmm1,xmm14 ; xmm1=data1+data6=tmp1
movdqa xmm2,xmm10
paddw xmm2,xmm13 ; xmm2=data2+data5=tmp2
movdqa xmm3,xmm11
paddw xmm3,xmm12 ; xmm3=data3+data4=tmp3
movdqa xmm2,xmm5
movdqa xmm7,xmm6
psubw xmm5,xmm0 ; xmm5=data1-data6=tmp6
psubw xmm6,xmm3 ; xmm6=data0-data7=tmp7
paddw xmm2,xmm0 ; xmm2=data1+data6=tmp1
paddw xmm7,xmm3 ; xmm7=data0+data7=tmp0
movdqa xmm0, XMMWORD [wk(2)] ; xmm0=(24 25 26 27 34 35 36 37)
movdqa xmm3, XMMWORD [wk(3)] ; xmm3=(44 45 46 47 54 55 56 57)
movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=tmp6
movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7
movdqa xmm5,xmm4 ; transpose coefficients(phase 3)
punpcklqdq xmm4,xmm0 ; xmm4=(20 21 22 23 24 25 26 27)=data2
punpckhqdq xmm5,xmm0 ; xmm5=(30 31 32 33 34 35 36 37)=data3
movdqa xmm6,xmm1 ; transpose coefficients(phase 3)
punpcklqdq xmm1,xmm3 ; xmm1=(40 41 42 43 44 45 46 47)=data4
punpckhqdq xmm6,xmm3 ; xmm6=(50 51 52 53 54 55 56 57)=data5
movdqa xmm0,xmm5
movdqa xmm3,xmm4
paddw xmm5,xmm1 ; xmm5=data3+data4=tmp3
paddw xmm4,xmm6 ; xmm4=data2+data5=tmp2
psubw xmm0,xmm1 ; xmm0=data3-data4=tmp4
psubw xmm3,xmm6 ; xmm3=data2-data5=tmp5
psubw xmm11,xmm12 ; xmm11=data3-data4=tmp4
psubw xmm10,xmm13 ; xmm10=data2-data5=tmp5
psubw xmm9,xmm14 ; xmm9=data1-data6=tmp6
psubw xmm8,xmm15 ; xmm8=data0-data7=tmp7
; -- Even part
movdqa xmm1,xmm7
movdqa xmm6,xmm2
paddw xmm7,xmm5 ; xmm7=tmp10
paddw xmm2,xmm4 ; xmm2=tmp11
psubw xmm1,xmm5 ; xmm1=tmp13
psubw xmm6,xmm4 ; xmm6=tmp12
movdqa xmm4,xmm0
paddw xmm4,xmm3 ; xmm4=tmp0+tmp3=tmp10
movdqa xmm5,xmm1
paddw xmm5,xmm2 ; xmm5=tmp1+tmp2=tmp11
psubw xmm1,xmm2 ; xmm1=tmp1-tmp2=tmp12
psubw xmm0,xmm3 ; xmm0=tmp0-tmp3=tmp13
movdqa xmm5,xmm7
paddw xmm7,xmm2 ; xmm7=tmp10+tmp11
psubw xmm5,xmm2 ; xmm5=tmp10-tmp11
movdqa xmm12,xmm4
paddw xmm12,xmm5 ; xmm12=tmp10+tmp11
psubw xmm4,xmm5 ; xmm4=tmp10-tmp11
paddw xmm7,[rel PW_DESCALE_P2X]
paddw xmm5,[rel PW_DESCALE_P2X]
psraw xmm7,PASS1_BITS ; xmm7=data0
psraw xmm5,PASS1_BITS ; xmm5=data4
movdqa XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm7
movdqa XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm5
paddw xmm12,[rel PW_DESCALE_P2X]
paddw xmm4,[rel PW_DESCALE_P2X]
psraw xmm12,PASS1_BITS ; xmm12=out0
psraw xmm4,PASS1_BITS ; xmm4=out4
; (Original)
; z1 = (tmp12 + tmp13) * 0.541196100;
; data2 = z1 + tmp13 * 0.765366865;
; data6 = z1 + tmp12 * -1.847759065;
; out2 = z1 + tmp13 * 0.765366865;
; out6 = z1 + tmp12 * -1.847759065;
;
; (This implementation)
; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
; out2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
; out6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
movdqa xmm4,xmm1 ; xmm1=tmp13
movdqa xmm2,xmm1
punpcklwd xmm4,xmm6 ; xmm6=tmp12
punpckhwd xmm2,xmm6
movdqa xmm1,xmm4
movdqa xmm6,xmm2
pmaddwd xmm4,[rel PW_F130_F054] ; xmm4=data2L
pmaddwd xmm2,[rel PW_F130_F054] ; xmm2=data2H
pmaddwd xmm1,[rel PW_F054_MF130] ; xmm1=data6L
pmaddwd xmm6,[rel PW_F054_MF130] ; xmm6=data6H
movdqa xmm2,xmm0 ; xmm0=tmp13
movdqa xmm6,xmm0
punpcklwd xmm2,xmm1 ; xmm1=tmp12
punpckhwd xmm6,xmm1
movdqa xmm5,xmm2
movdqa xmm0,xmm6
pmaddwd xmm2,[rel PW_F130_F054] ; xmm2=out2L
pmaddwd xmm6,[rel PW_F130_F054] ; xmm6=out2H
pmaddwd xmm5,[rel PW_F054_MF130] ; xmm5=out6L
pmaddwd xmm0,[rel PW_F054_MF130] ; xmm0=out6H
paddd xmm4,[rel PD_DESCALE_P2]
paddd xmm2,[rel PD_DESCALE_P2]
psrad xmm4,DESCALE_P2
psrad xmm2,DESCALE_P2
paddd xmm1,[rel PD_DESCALE_P2]
paddd xmm6,[rel PD_DESCALE_P2]
psrad xmm1,DESCALE_P2
psrad xmm2,DESCALE_P2
psrad xmm6,DESCALE_P2
paddd xmm5,[rel PD_DESCALE_P2]
paddd xmm0,[rel PD_DESCALE_P2]
psrad xmm5,DESCALE_P2
psrad xmm0,DESCALE_P2
packssdw xmm4,xmm2 ; xmm4=data2
packssdw xmm1,xmm6 ; xmm1=data6
movdqa XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm4
movdqa XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm1
packssdw xmm2,xmm6 ; xmm2=out2
packssdw xmm5,xmm0 ; xmm5=out6
; -- Odd part
movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp6
movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7
movdqa xmm2,xmm0 ; xmm0=tmp4
movdqa xmm6,xmm3 ; xmm3=tmp5
paddw xmm2,xmm7 ; xmm2=z3
paddw xmm6,xmm5 ; xmm6=z4
movdqa xmm6,xmm11 ; xmm11=tmp4
movdqa xmm0,xmm10 ; xmm10=tmp5
paddw xmm6,xmm9 ; xmm6=tmp4+tmp6=z3
paddw xmm0,xmm8 ; xmm0=tmp5+tmp7=z4
; (Original)
; z5 = (z3 + z4) * 1.175875602;
@@ -519,102 +457,101 @@ EXTN(jsimd_fdct_islow_sse2):
; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
movdqa xmm4,xmm2
movdqa xmm1,xmm2
punpcklwd xmm4,xmm6
punpckhwd xmm1,xmm6
movdqa xmm2,xmm4
movdqa xmm6,xmm1
pmaddwd xmm4,[rel PW_MF078_F117] ; xmm4=z3L
pmaddwd xmm1,[rel PW_MF078_F117] ; xmm1=z3H
pmaddwd xmm2,[rel PW_F117_F078] ; xmm2=z4L
pmaddwd xmm6,[rel PW_F117_F078] ; xmm6=z4H
movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=z3L
movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=z3H
movdqa xmm7,xmm6
movdqa xmm14,xmm6
punpcklwd xmm7,xmm0
punpckhwd xmm14,xmm0
movdqa xmm6,xmm7
movdqa xmm0,xmm14
pmaddwd xmm7,[rel PW_MF078_F117] ; xmm7=z3L
pmaddwd xmm14,[rel PW_MF078_F117] ; xmm14=z3H
pmaddwd xmm6,[rel PW_F117_F078] ; xmm6=z4L
pmaddwd xmm0,[rel PW_F117_F078] ; xmm0=z4H
; (Original)
; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
; out7 = tmp4 + z1 + z3; out5 = tmp5 + z2 + z4;
; out3 = tmp6 + z2 + z3; out1 = tmp7 + z1 + z4;
;
; (This implementation)
; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
; data7 = tmp4 + z3; data5 = tmp5 + z4;
; data3 = tmp6 + z3; data1 = tmp7 + z4;
; out7 = tmp4 + z3; out5 = tmp5 + z4;
; out3 = tmp6 + z3; out1 = tmp7 + z4;
movdqa xmm4,xmm0
movdqa xmm1,xmm0
punpcklwd xmm4,xmm5
punpckhwd xmm1,xmm5
movdqa xmm0,xmm4
movdqa xmm5,xmm1
pmaddwd xmm4,[rel PW_MF060_MF089] ; xmm4=tmp4L
pmaddwd xmm1,[rel PW_MF060_MF089] ; xmm1=tmp4H
pmaddwd xmm0,[rel PW_MF089_F060] ; xmm0=tmp7L
pmaddwd xmm5,[rel PW_MF089_F060] ; xmm5=tmp7H
movdqa xmm15,xmm11 ; xmm11=tmp4
movdqa xmm13,xmm11
punpcklwd xmm15,xmm8 ; xmm8=tmp7
punpckhwd xmm13,xmm8
movdqa xmm8,xmm15
movdqa xmm1,xmm13
pmaddwd xmm15,[rel PW_MF060_MF089] ; xmm15=tmp4L
pmaddwd xmm13,[rel PW_MF060_MF089] ; xmm13=tmp4H
pmaddwd xmm8,[rel PW_MF089_F060] ; xmm8=tmp7L
pmaddwd xmm1,[rel PW_MF089_F060] ; xmm1=tmp7H
paddd xmm4, XMMWORD [wk(0)] ; xmm4=data7L
paddd xmm1, XMMWORD [wk(1)] ; xmm1=data7H
paddd xmm0,xmm2 ; xmm0=data1L
paddd xmm5,xmm6 ; xmm5=data1H
paddd xmm15,xmm7 ; xmm15=out7L
paddd xmm13,xmm14 ; xmm13=out7H
paddd xmm8,xmm6 ; xmm8=out1L
paddd xmm1,xmm0 ; xmm1=out1H
paddd xmm4,[rel PD_DESCALE_P2]
paddd xmm15,[rel PD_DESCALE_P2]
paddd xmm13,[rel PD_DESCALE_P2]
psrad xmm15,DESCALE_P2
psrad xmm13,DESCALE_P2
paddd xmm8,[rel PD_DESCALE_P2]
paddd xmm1,[rel PD_DESCALE_P2]
psrad xmm4,DESCALE_P2
psrad xmm8,DESCALE_P2
psrad xmm1,DESCALE_P2
paddd xmm0,[rel PD_DESCALE_P2]
paddd xmm5,[rel PD_DESCALE_P2]
psrad xmm0,DESCALE_P2
psrad xmm5,DESCALE_P2
packssdw xmm4,xmm1 ; xmm4=data7
packssdw xmm0,xmm5 ; xmm0=data1
packssdw xmm15,xmm13 ; xmm15=out7
packssdw xmm8,xmm1 ; xmm8=out1
movdqa XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm4
movdqa XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm0
movdqa xmm1,xmm3
movdqa xmm5,xmm3
punpcklwd xmm1,xmm7
punpckhwd xmm5,xmm7
movdqa xmm13,xmm10 ; xmm10=tmp5
movdqa xmm1,xmm10
punpcklwd xmm13,xmm9 ; xmm9=tmp6
punpckhwd xmm1,xmm9
movdqa xmm11,xmm13
movdqa xmm3,xmm1
movdqa xmm7,xmm5
pmaddwd xmm1,[rel PW_MF050_MF256] ; xmm1=tmp5L
pmaddwd xmm5,[rel PW_MF050_MF256] ; xmm5=tmp5H
pmaddwd xmm3,[rel PW_MF256_F050] ; xmm3=tmp6L
pmaddwd xmm7,[rel PW_MF256_F050] ; xmm7=tmp6H
pmaddwd xmm13,[rel PW_MF050_MF256] ; xmm13=tmp5L
pmaddwd xmm1,[rel PW_MF050_MF256] ; xmm1=tmp5H
pmaddwd xmm11,[rel PW_MF256_F050] ; xmm11=tmp6L
pmaddwd xmm3,[rel PW_MF256_F050] ; xmm3=tmp6H
paddd xmm1,xmm2 ; xmm1=data5L
paddd xmm5,xmm6 ; xmm5=data5H
paddd xmm3, XMMWORD [wk(0)] ; xmm3=data3L
paddd xmm7, XMMWORD [wk(1)] ; xmm7=data3H
paddd xmm13,xmm6 ; xmm13=out5L
paddd xmm1,xmm0 ; xmm1=out5H
paddd xmm11,xmm7 ; xmm11=out3L
paddd xmm3,xmm14 ; xmm3=out3H
paddd xmm13,[rel PD_DESCALE_P2]
paddd xmm1,[rel PD_DESCALE_P2]
paddd xmm5,[rel PD_DESCALE_P2]
psrad xmm13,DESCALE_P2
psrad xmm1,DESCALE_P2
psrad xmm5,DESCALE_P2
paddd xmm11,[rel PD_DESCALE_P2]
paddd xmm3,[rel PD_DESCALE_P2]
paddd xmm7,[rel PD_DESCALE_P2]
psrad xmm11,DESCALE_P2
psrad xmm3,DESCALE_P2
psrad xmm7,DESCALE_P2
packssdw xmm1,xmm5 ; xmm1=data5
packssdw xmm3,xmm7 ; xmm3=data3
packssdw xmm13,xmm1 ; xmm13=out5
packssdw xmm11,xmm3 ; xmm11=out3
movdqa XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm1
movdqa XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm3
; -- Write result
movdqa XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm12
movdqa XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm8
movdqa XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm2
movdqa XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm11
movdqa XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm4
movdqa XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm13
movdqa XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm5
movdqa XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm15
uncollect_args
mov rsp,rbp ; rsp <- aligned rbp
pop rsp ; rsp <- original rbp
pop rbp
ret
; For some reason, the OS X linker does not honor the request to align the