media/libjpeg/simd/jfss2int.asm@b8a032363ba2
media/libjpeg/simd/jfss2int.asm
Thu, 22 Jan 2015 13:21:57 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Thu, 22 Jan 2015 13:21:57 +0100
- branch
- TOR_BUG_9701
- changeset 15
- b8a032363ba2
- permissions
- -rw-r--r--
Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6
1 ;
2 ; jfss2int.asm - accurate integer FDCT (SSE2)
3 ;
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ;
6 ; Based on
7 ; x86 SIMD extension for IJG JPEG library
8 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
9 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
10 ;
11 ; This file should be assembled with NASM (Netwide Assembler),
12 ; can *not* be assembled with Microsoft's MASM or any compatible
13 ; assembler (including Borland's Turbo Assembler).
14 ; NASM is available from http://nasm.sourceforge.net/ or
15 ; http://sourceforge.net/project/showfiles.php?group_id=6208
16 ;
17 ; This file contains a slow-but-accurate integer implementation of the
18 ; forward DCT (Discrete Cosine Transform). The following code is based
19 ; directly on the IJG's original jfdctint.c; see the jfdctint.c for
20 ; more details.
21 ;
22 ; [TAB8]
24 %include "jsimdext.inc"
25 %include "jdct.inc"
27 ; --------------------------------------------------------------------------
29 %define CONST_BITS 13
30 %define PASS1_BITS 2
32 %define DESCALE_P1 (CONST_BITS-PASS1_BITS)
33 %define DESCALE_P2 (CONST_BITS+PASS1_BITS)
35 %if CONST_BITS == 13
36 F_0_298 equ 2446 ; FIX(0.298631336)
37 F_0_390 equ 3196 ; FIX(0.390180644)
38 F_0_541 equ 4433 ; FIX(0.541196100)
39 F_0_765 equ 6270 ; FIX(0.765366865)
40 F_0_899 equ 7373 ; FIX(0.899976223)
41 F_1_175 equ 9633 ; FIX(1.175875602)
42 F_1_501 equ 12299 ; FIX(1.501321110)
43 F_1_847 equ 15137 ; FIX(1.847759065)
44 F_1_961 equ 16069 ; FIX(1.961570560)
45 F_2_053 equ 16819 ; FIX(2.053119869)
46 F_2_562 equ 20995 ; FIX(2.562915447)
47 F_3_072 equ 25172 ; FIX(3.072711026)
48 %else
49 ; NASM cannot do compile-time arithmetic on floating-point constants.
50 %define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
51 F_0_298 equ DESCALE( 320652955,30-CONST_BITS) ; FIX(0.298631336)
52 F_0_390 equ DESCALE( 418953276,30-CONST_BITS) ; FIX(0.390180644)
53 F_0_541 equ DESCALE( 581104887,30-CONST_BITS) ; FIX(0.541196100)
54 F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
55 F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
56 F_1_175 equ DESCALE(1262586813,30-CONST_BITS) ; FIX(1.175875602)
57 F_1_501 equ DESCALE(1612031267,30-CONST_BITS) ; FIX(1.501321110)
58 F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
59 F_1_961 equ DESCALE(2106220350,30-CONST_BITS) ; FIX(1.961570560)
60 F_2_053 equ DESCALE(2204520673,30-CONST_BITS) ; FIX(2.053119869)
61 F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
62 F_3_072 equ DESCALE(3299298341,30-CONST_BITS) ; FIX(3.072711026)
63 %endif
65 ; --------------------------------------------------------------------------
66 SECTION SEG_CONST
68 alignz 16
69 global EXTN(jconst_fdct_islow_sse2)
71 EXTN(jconst_fdct_islow_sse2):
73 PW_F130_F054 times 4 dw (F_0_541+F_0_765), F_0_541
74 PW_F054_MF130 times 4 dw F_0_541, (F_0_541-F_1_847)
75 PW_MF078_F117 times 4 dw (F_1_175-F_1_961), F_1_175
76 PW_F117_F078 times 4 dw F_1_175, (F_1_175-F_0_390)
77 PW_MF060_MF089 times 4 dw (F_0_298-F_0_899),-F_0_899
78 PW_MF089_F060 times 4 dw -F_0_899, (F_1_501-F_0_899)
79 PW_MF050_MF256 times 4 dw (F_2_053-F_2_562),-F_2_562
80 PW_MF256_F050 times 4 dw -F_2_562, (F_3_072-F_2_562)
81 PD_DESCALE_P1 times 4 dd 1 << (DESCALE_P1-1)
82 PD_DESCALE_P2 times 4 dd 1 << (DESCALE_P2-1)
83 PW_DESCALE_P2X times 8 dw 1 << (PASS1_BITS-1)
85 alignz 16
87 ; --------------------------------------------------------------------------
88 SECTION SEG_TEXT
89 BITS 32
90 ;
91 ; Perform the forward DCT on one block of samples.
92 ;
93 ; GLOBAL(void)
94 ; jsimd_fdct_islow_sse2 (DCTELEM * data)
95 ;
97 %define data(b) (b)+8 ; DCTELEM * data
99 %define original_ebp ebp+0
100 %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
101 %define WK_NUM 6
103 align 16
104 global EXTN(jsimd_fdct_islow_sse2)
106 EXTN(jsimd_fdct_islow_sse2):
107 push ebp
108 mov eax,esp ; eax = original ebp
109 sub esp, byte 4
110 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
111 mov [esp],eax
112 mov ebp,esp ; ebp = aligned ebp
113 lea esp, [wk(0)]
114 pushpic ebx
115 ; push ecx ; unused
116 ; push edx ; need not be preserved
117 ; push esi ; unused
118 ; push edi ; unused
120 get_GOT ebx ; get GOT address
122 ; ---- Pass 1: process rows.
124 mov edx, POINTER [data(eax)] ; (DCTELEM *)
126 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
127 movdqa xmm1, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
128 movdqa xmm2, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
129 movdqa xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
131 ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
132 ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
134 movdqa xmm4,xmm0 ; transpose coefficients(phase 1)
135 punpcklwd xmm0,xmm1 ; xmm0=(00 10 01 11 02 12 03 13)
136 punpckhwd xmm4,xmm1 ; xmm4=(04 14 05 15 06 16 07 17)
137 movdqa xmm5,xmm2 ; transpose coefficients(phase 1)
138 punpcklwd xmm2,xmm3 ; xmm2=(20 30 21 31 22 32 23 33)
139 punpckhwd xmm5,xmm3 ; xmm5=(24 34 25 35 26 36 27 37)
141 movdqa xmm6, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
142 movdqa xmm7, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
143 movdqa xmm1, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
144 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
146 ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
147 ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
149 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33)
150 movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37)
152 movdqa xmm2,xmm6 ; transpose coefficients(phase 1)
153 punpcklwd xmm6,xmm7 ; xmm6=(40 50 41 51 42 52 43 53)
154 punpckhwd xmm2,xmm7 ; xmm2=(44 54 45 55 46 56 47 57)
155 movdqa xmm5,xmm1 ; transpose coefficients(phase 1)
156 punpcklwd xmm1,xmm3 ; xmm1=(60 70 61 71 62 72 63 73)
157 punpckhwd xmm5,xmm3 ; xmm5=(64 74 65 75 66 76 67 77)
159 movdqa xmm7,xmm6 ; transpose coefficients(phase 2)
160 punpckldq xmm6,xmm1 ; xmm6=(40 50 60 70 41 51 61 71)
161 punpckhdq xmm7,xmm1 ; xmm7=(42 52 62 72 43 53 63 73)
162 movdqa xmm3,xmm2 ; transpose coefficients(phase 2)
163 punpckldq xmm2,xmm5 ; xmm2=(44 54 64 74 45 55 65 75)
164 punpckhdq xmm3,xmm5 ; xmm3=(46 56 66 76 47 57 67 77)
166 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33)
167 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37)
168 movdqa XMMWORD [wk(2)], xmm7 ; wk(2)=(42 52 62 72 43 53 63 73)
169 movdqa XMMWORD [wk(3)], xmm2 ; wk(3)=(44 54 64 74 45 55 65 75)
171 movdqa xmm7,xmm0 ; transpose coefficients(phase 2)
172 punpckldq xmm0,xmm1 ; xmm0=(00 10 20 30 01 11 21 31)
173 punpckhdq xmm7,xmm1 ; xmm7=(02 12 22 32 03 13 23 33)
174 movdqa xmm2,xmm4 ; transpose coefficients(phase 2)
175 punpckldq xmm4,xmm5 ; xmm4=(04 14 24 34 05 15 25 35)
176 punpckhdq xmm2,xmm5 ; xmm2=(06 16 26 36 07 17 27 37)
178 movdqa xmm1,xmm0 ; transpose coefficients(phase 3)
179 punpcklqdq xmm0,xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0
180 punpckhqdq xmm1,xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1
181 movdqa xmm5,xmm2 ; transpose coefficients(phase 3)
182 punpcklqdq xmm2,xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6
183 punpckhqdq xmm5,xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7
185 movdqa xmm6,xmm1
186 movdqa xmm3,xmm0
187 psubw xmm1,xmm2 ; xmm1=data1-data6=tmp6
188 psubw xmm0,xmm5 ; xmm0=data0-data7=tmp7
189 paddw xmm6,xmm2 ; xmm6=data1+data6=tmp1
190 paddw xmm3,xmm5 ; xmm3=data0+data7=tmp0
192 movdqa xmm2, XMMWORD [wk(2)] ; xmm2=(42 52 62 72 43 53 63 73)
193 movdqa xmm5, XMMWORD [wk(3)] ; xmm5=(44 54 64 74 45 55 65 75)
194 movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6
195 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7
197 movdqa xmm1,xmm7 ; transpose coefficients(phase 3)
198 punpcklqdq xmm7,xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2
199 punpckhqdq xmm1,xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3
200 movdqa xmm0,xmm4 ; transpose coefficients(phase 3)
201 punpcklqdq xmm4,xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4
202 punpckhqdq xmm0,xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5
204 movdqa xmm2,xmm1
205 movdqa xmm5,xmm7
206 paddw xmm1,xmm4 ; xmm1=data3+data4=tmp3
207 paddw xmm7,xmm0 ; xmm7=data2+data5=tmp2
208 psubw xmm2,xmm4 ; xmm2=data3-data4=tmp4
209 psubw xmm5,xmm0 ; xmm5=data2-data5=tmp5
211 ; -- Even part
213 movdqa xmm4,xmm3
214 movdqa xmm0,xmm6
215 paddw xmm3,xmm1 ; xmm3=tmp10
216 paddw xmm6,xmm7 ; xmm6=tmp11
217 psubw xmm4,xmm1 ; xmm4=tmp13
218 psubw xmm0,xmm7 ; xmm0=tmp12
220 movdqa xmm1,xmm3
221 paddw xmm3,xmm6 ; xmm3=tmp10+tmp11
222 psubw xmm1,xmm6 ; xmm1=tmp10-tmp11
224 psllw xmm3,PASS1_BITS ; xmm3=data0
225 psllw xmm1,PASS1_BITS ; xmm1=data4
227 movdqa XMMWORD [wk(2)], xmm3 ; wk(2)=data0
228 movdqa XMMWORD [wk(3)], xmm1 ; wk(3)=data4
230 ; (Original)
231 ; z1 = (tmp12 + tmp13) * 0.541196100;
232 ; data2 = z1 + tmp13 * 0.765366865;
233 ; data6 = z1 + tmp12 * -1.847759065;
234 ;
235 ; (This implementation)
236 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
237 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
239 movdqa xmm7,xmm4 ; xmm4=tmp13
240 movdqa xmm6,xmm4
241 punpcklwd xmm7,xmm0 ; xmm0=tmp12
242 punpckhwd xmm6,xmm0
243 movdqa xmm4,xmm7
244 movdqa xmm0,xmm6
245 pmaddwd xmm7,[GOTOFF(ebx,PW_F130_F054)] ; xmm7=data2L
246 pmaddwd xmm6,[GOTOFF(ebx,PW_F130_F054)] ; xmm6=data2H
247 pmaddwd xmm4,[GOTOFF(ebx,PW_F054_MF130)] ; xmm4=data6L
248 pmaddwd xmm0,[GOTOFF(ebx,PW_F054_MF130)] ; xmm0=data6H
250 paddd xmm7,[GOTOFF(ebx,PD_DESCALE_P1)]
251 paddd xmm6,[GOTOFF(ebx,PD_DESCALE_P1)]
252 psrad xmm7,DESCALE_P1
253 psrad xmm6,DESCALE_P1
254 paddd xmm4,[GOTOFF(ebx,PD_DESCALE_P1)]
255 paddd xmm0,[GOTOFF(ebx,PD_DESCALE_P1)]
256 psrad xmm4,DESCALE_P1
257 psrad xmm0,DESCALE_P1
259 packssdw xmm7,xmm6 ; xmm7=data2
260 packssdw xmm4,xmm0 ; xmm4=data6
262 movdqa XMMWORD [wk(4)], xmm7 ; wk(4)=data2
263 movdqa XMMWORD [wk(5)], xmm4 ; wk(5)=data6
265 ; -- Odd part
267 movdqa xmm3, XMMWORD [wk(0)] ; xmm3=tmp6
268 movdqa xmm1, XMMWORD [wk(1)] ; xmm1=tmp7
270 movdqa xmm6,xmm2 ; xmm2=tmp4
271 movdqa xmm0,xmm5 ; xmm5=tmp5
272 paddw xmm6,xmm3 ; xmm6=z3
273 paddw xmm0,xmm1 ; xmm0=z4
275 ; (Original)
276 ; z5 = (z3 + z4) * 1.175875602;
277 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
278 ; z3 += z5; z4 += z5;
279 ;
280 ; (This implementation)
281 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
282 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
284 movdqa xmm7,xmm6
285 movdqa xmm4,xmm6
286 punpcklwd xmm7,xmm0
287 punpckhwd xmm4,xmm0
288 movdqa xmm6,xmm7
289 movdqa xmm0,xmm4
290 pmaddwd xmm7,[GOTOFF(ebx,PW_MF078_F117)] ; xmm7=z3L
291 pmaddwd xmm4,[GOTOFF(ebx,PW_MF078_F117)] ; xmm4=z3H
292 pmaddwd xmm6,[GOTOFF(ebx,PW_F117_F078)] ; xmm6=z4L
293 pmaddwd xmm0,[GOTOFF(ebx,PW_F117_F078)] ; xmm0=z4H
295 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=z3L
296 movdqa XMMWORD [wk(1)], xmm4 ; wk(1)=z3H
298 ; (Original)
299 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
300 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
301 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
302 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
303 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
304 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
305 ;
306 ; (This implementation)
307 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
308 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
309 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
310 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
311 ; data7 = tmp4 + z3; data5 = tmp5 + z4;
312 ; data3 = tmp6 + z3; data1 = tmp7 + z4;
314 movdqa xmm7,xmm2
315 movdqa xmm4,xmm2
316 punpcklwd xmm7,xmm1
317 punpckhwd xmm4,xmm1
318 movdqa xmm2,xmm7
319 movdqa xmm1,xmm4
320 pmaddwd xmm7,[GOTOFF(ebx,PW_MF060_MF089)] ; xmm7=tmp4L
321 pmaddwd xmm4,[GOTOFF(ebx,PW_MF060_MF089)] ; xmm4=tmp4H
322 pmaddwd xmm2,[GOTOFF(ebx,PW_MF089_F060)] ; xmm2=tmp7L
323 pmaddwd xmm1,[GOTOFF(ebx,PW_MF089_F060)] ; xmm1=tmp7H
325 paddd xmm7, XMMWORD [wk(0)] ; xmm7=data7L
326 paddd xmm4, XMMWORD [wk(1)] ; xmm4=data7H
327 paddd xmm2,xmm6 ; xmm2=data1L
328 paddd xmm1,xmm0 ; xmm1=data1H
330 paddd xmm7,[GOTOFF(ebx,PD_DESCALE_P1)]
331 paddd xmm4,[GOTOFF(ebx,PD_DESCALE_P1)]
332 psrad xmm7,DESCALE_P1
333 psrad xmm4,DESCALE_P1
334 paddd xmm2,[GOTOFF(ebx,PD_DESCALE_P1)]
335 paddd xmm1,[GOTOFF(ebx,PD_DESCALE_P1)]
336 psrad xmm2,DESCALE_P1
337 psrad xmm1,DESCALE_P1
339 packssdw xmm7,xmm4 ; xmm7=data7
340 packssdw xmm2,xmm1 ; xmm2=data1
342 movdqa xmm4,xmm5
343 movdqa xmm1,xmm5
344 punpcklwd xmm4,xmm3
345 punpckhwd xmm1,xmm3
346 movdqa xmm5,xmm4
347 movdqa xmm3,xmm1
348 pmaddwd xmm4,[GOTOFF(ebx,PW_MF050_MF256)] ; xmm4=tmp5L
349 pmaddwd xmm1,[GOTOFF(ebx,PW_MF050_MF256)] ; xmm1=tmp5H
350 pmaddwd xmm5,[GOTOFF(ebx,PW_MF256_F050)] ; xmm5=tmp6L
351 pmaddwd xmm3,[GOTOFF(ebx,PW_MF256_F050)] ; xmm3=tmp6H
353 paddd xmm4,xmm6 ; xmm4=data5L
354 paddd xmm1,xmm0 ; xmm1=data5H
355 paddd xmm5, XMMWORD [wk(0)] ; xmm5=data3L
356 paddd xmm3, XMMWORD [wk(1)] ; xmm3=data3H
358 paddd xmm4,[GOTOFF(ebx,PD_DESCALE_P1)]
359 paddd xmm1,[GOTOFF(ebx,PD_DESCALE_P1)]
360 psrad xmm4,DESCALE_P1
361 psrad xmm1,DESCALE_P1
362 paddd xmm5,[GOTOFF(ebx,PD_DESCALE_P1)]
363 paddd xmm3,[GOTOFF(ebx,PD_DESCALE_P1)]
364 psrad xmm5,DESCALE_P1
365 psrad xmm3,DESCALE_P1
367 packssdw xmm4,xmm1 ; xmm4=data5
368 packssdw xmm5,xmm3 ; xmm5=data3
370 ; ---- Pass 2: process columns.
372 ; mov edx, POINTER [data(eax)] ; (DCTELEM *)
374 movdqa xmm6, XMMWORD [wk(2)] ; xmm6=col0
375 movdqa xmm0, XMMWORD [wk(4)] ; xmm0=col2
377 ; xmm6=(00 10 20 30 40 50 60 70), xmm0=(02 12 22 32 42 52 62 72)
378 ; xmm2=(01 11 21 31 41 51 61 71), xmm5=(03 13 23 33 43 53 63 73)
380 movdqa xmm1,xmm6 ; transpose coefficients(phase 1)
381 punpcklwd xmm6,xmm2 ; xmm6=(00 01 10 11 20 21 30 31)
382 punpckhwd xmm1,xmm2 ; xmm1=(40 41 50 51 60 61 70 71)
383 movdqa xmm3,xmm0 ; transpose coefficients(phase 1)
384 punpcklwd xmm0,xmm5 ; xmm0=(02 03 12 13 22 23 32 33)
385 punpckhwd xmm3,xmm5 ; xmm3=(42 43 52 53 62 63 72 73)
387 movdqa xmm2, XMMWORD [wk(3)] ; xmm2=col4
388 movdqa xmm5, XMMWORD [wk(5)] ; xmm5=col6
390 ; xmm2=(04 14 24 34 44 54 64 74), xmm5=(06 16 26 36 46 56 66 76)
391 ; xmm4=(05 15 25 35 45 55 65 75), xmm7=(07 17 27 37 47 57 67 77)
393 movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=(02 03 12 13 22 23 32 33)
394 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=(42 43 52 53 62 63 72 73)
396 movdqa xmm0,xmm2 ; transpose coefficients(phase 1)
397 punpcklwd xmm2,xmm4 ; xmm2=(04 05 14 15 24 25 34 35)
398 punpckhwd xmm0,xmm4 ; xmm0=(44 45 54 55 64 65 74 75)
399 movdqa xmm3,xmm5 ; transpose coefficients(phase 1)
400 punpcklwd xmm5,xmm7 ; xmm5=(06 07 16 17 26 27 36 37)
401 punpckhwd xmm3,xmm7 ; xmm3=(46 47 56 57 66 67 76 77)
403 movdqa xmm4,xmm2 ; transpose coefficients(phase 2)
404 punpckldq xmm2,xmm5 ; xmm2=(04 05 06 07 14 15 16 17)
405 punpckhdq xmm4,xmm5 ; xmm4=(24 25 26 27 34 35 36 37)
406 movdqa xmm7,xmm0 ; transpose coefficients(phase 2)
407 punpckldq xmm0,xmm3 ; xmm0=(44 45 46 47 54 55 56 57)
408 punpckhdq xmm7,xmm3 ; xmm7=(64 65 66 67 74 75 76 77)
410 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=(02 03 12 13 22 23 32 33)
411 movdqa xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53 62 63 72 73)
412 movdqa XMMWORD [wk(2)], xmm4 ; wk(2)=(24 25 26 27 34 35 36 37)
413 movdqa XMMWORD [wk(3)], xmm0 ; wk(3)=(44 45 46 47 54 55 56 57)
415 movdqa xmm4,xmm6 ; transpose coefficients(phase 2)
416 punpckldq xmm6,xmm5 ; xmm6=(00 01 02 03 10 11 12 13)
417 punpckhdq xmm4,xmm5 ; xmm4=(20 21 22 23 30 31 32 33)
418 movdqa xmm0,xmm1 ; transpose coefficients(phase 2)
419 punpckldq xmm1,xmm3 ; xmm1=(40 41 42 43 50 51 52 53)
420 punpckhdq xmm0,xmm3 ; xmm0=(60 61 62 63 70 71 72 73)
422 movdqa xmm5,xmm6 ; transpose coefficients(phase 3)
423 punpcklqdq xmm6,xmm2 ; xmm6=(00 01 02 03 04 05 06 07)=data0
424 punpckhqdq xmm5,xmm2 ; xmm5=(10 11 12 13 14 15 16 17)=data1
425 movdqa xmm3,xmm0 ; transpose coefficients(phase 3)
426 punpcklqdq xmm0,xmm7 ; xmm0=(60 61 62 63 64 65 66 67)=data6
427 punpckhqdq xmm3,xmm7 ; xmm3=(70 71 72 73 74 75 76 77)=data7
429 movdqa xmm2,xmm5
430 movdqa xmm7,xmm6
431 psubw xmm5,xmm0 ; xmm5=data1-data6=tmp6
432 psubw xmm6,xmm3 ; xmm6=data0-data7=tmp7
433 paddw xmm2,xmm0 ; xmm2=data1+data6=tmp1
434 paddw xmm7,xmm3 ; xmm7=data0+data7=tmp0
436 movdqa xmm0, XMMWORD [wk(2)] ; xmm0=(24 25 26 27 34 35 36 37)
437 movdqa xmm3, XMMWORD [wk(3)] ; xmm3=(44 45 46 47 54 55 56 57)
438 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=tmp6
439 movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7
441 movdqa xmm5,xmm4 ; transpose coefficients(phase 3)
442 punpcklqdq xmm4,xmm0 ; xmm4=(20 21 22 23 24 25 26 27)=data2
443 punpckhqdq xmm5,xmm0 ; xmm5=(30 31 32 33 34 35 36 37)=data3
444 movdqa xmm6,xmm1 ; transpose coefficients(phase 3)
445 punpcklqdq xmm1,xmm3 ; xmm1=(40 41 42 43 44 45 46 47)=data4
446 punpckhqdq xmm6,xmm3 ; xmm6=(50 51 52 53 54 55 56 57)=data5
448 movdqa xmm0,xmm5
449 movdqa xmm3,xmm4
450 paddw xmm5,xmm1 ; xmm5=data3+data4=tmp3
451 paddw xmm4,xmm6 ; xmm4=data2+data5=tmp2
452 psubw xmm0,xmm1 ; xmm0=data3-data4=tmp4
453 psubw xmm3,xmm6 ; xmm3=data2-data5=tmp5
455 ; -- Even part
457 movdqa xmm1,xmm7
458 movdqa xmm6,xmm2
459 paddw xmm7,xmm5 ; xmm7=tmp10
460 paddw xmm2,xmm4 ; xmm2=tmp11
461 psubw xmm1,xmm5 ; xmm1=tmp13
462 psubw xmm6,xmm4 ; xmm6=tmp12
464 movdqa xmm5,xmm7
465 paddw xmm7,xmm2 ; xmm7=tmp10+tmp11
466 psubw xmm5,xmm2 ; xmm5=tmp10-tmp11
468 paddw xmm7,[GOTOFF(ebx,PW_DESCALE_P2X)]
469 paddw xmm5,[GOTOFF(ebx,PW_DESCALE_P2X)]
470 psraw xmm7,PASS1_BITS ; xmm7=data0
471 psraw xmm5,PASS1_BITS ; xmm5=data4
473 movdqa XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)], xmm7
474 movdqa XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)], xmm5
476 ; (Original)
477 ; z1 = (tmp12 + tmp13) * 0.541196100;
478 ; data2 = z1 + tmp13 * 0.765366865;
479 ; data6 = z1 + tmp12 * -1.847759065;
480 ;
481 ; (This implementation)
482 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
483 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
485 movdqa xmm4,xmm1 ; xmm1=tmp13
486 movdqa xmm2,xmm1
487 punpcklwd xmm4,xmm6 ; xmm6=tmp12
488 punpckhwd xmm2,xmm6
489 movdqa xmm1,xmm4
490 movdqa xmm6,xmm2
491 pmaddwd xmm4,[GOTOFF(ebx,PW_F130_F054)] ; xmm4=data2L
492 pmaddwd xmm2,[GOTOFF(ebx,PW_F130_F054)] ; xmm2=data2H
493 pmaddwd xmm1,[GOTOFF(ebx,PW_F054_MF130)] ; xmm1=data6L
494 pmaddwd xmm6,[GOTOFF(ebx,PW_F054_MF130)] ; xmm6=data6H
496 paddd xmm4,[GOTOFF(ebx,PD_DESCALE_P2)]
497 paddd xmm2,[GOTOFF(ebx,PD_DESCALE_P2)]
498 psrad xmm4,DESCALE_P2
499 psrad xmm2,DESCALE_P2
500 paddd xmm1,[GOTOFF(ebx,PD_DESCALE_P2)]
501 paddd xmm6,[GOTOFF(ebx,PD_DESCALE_P2)]
502 psrad xmm1,DESCALE_P2
503 psrad xmm6,DESCALE_P2
505 packssdw xmm4,xmm2 ; xmm4=data2
506 packssdw xmm1,xmm6 ; xmm1=data6
508 movdqa XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)], xmm4
509 movdqa XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)], xmm1
511 ; -- Odd part
513 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp6
514 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7
516 movdqa xmm2,xmm0 ; xmm0=tmp4
517 movdqa xmm6,xmm3 ; xmm3=tmp5
518 paddw xmm2,xmm7 ; xmm2=z3
519 paddw xmm6,xmm5 ; xmm6=z4
521 ; (Original)
522 ; z5 = (z3 + z4) * 1.175875602;
523 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
524 ; z3 += z5; z4 += z5;
525 ;
526 ; (This implementation)
527 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
528 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
530 movdqa xmm4,xmm2
531 movdqa xmm1,xmm2
532 punpcklwd xmm4,xmm6
533 punpckhwd xmm1,xmm6
534 movdqa xmm2,xmm4
535 movdqa xmm6,xmm1
536 pmaddwd xmm4,[GOTOFF(ebx,PW_MF078_F117)] ; xmm4=z3L
537 pmaddwd xmm1,[GOTOFF(ebx,PW_MF078_F117)] ; xmm1=z3H
538 pmaddwd xmm2,[GOTOFF(ebx,PW_F117_F078)] ; xmm2=z4L
539 pmaddwd xmm6,[GOTOFF(ebx,PW_F117_F078)] ; xmm6=z4H
541 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=z3L
542 movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=z3H
544 ; (Original)
545 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
546 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
547 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
548 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
549 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
550 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
551 ;
552 ; (This implementation)
553 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
554 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
555 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
556 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
557 ; data7 = tmp4 + z3; data5 = tmp5 + z4;
558 ; data3 = tmp6 + z3; data1 = tmp7 + z4;
560 movdqa xmm4,xmm0
561 movdqa xmm1,xmm0
562 punpcklwd xmm4,xmm5
563 punpckhwd xmm1,xmm5
564 movdqa xmm0,xmm4
565 movdqa xmm5,xmm1
566 pmaddwd xmm4,[GOTOFF(ebx,PW_MF060_MF089)] ; xmm4=tmp4L
567 pmaddwd xmm1,[GOTOFF(ebx,PW_MF060_MF089)] ; xmm1=tmp4H
568 pmaddwd xmm0,[GOTOFF(ebx,PW_MF089_F060)] ; xmm0=tmp7L
569 pmaddwd xmm5,[GOTOFF(ebx,PW_MF089_F060)] ; xmm5=tmp7H
571 paddd xmm4, XMMWORD [wk(0)] ; xmm4=data7L
572 paddd xmm1, XMMWORD [wk(1)] ; xmm1=data7H
573 paddd xmm0,xmm2 ; xmm0=data1L
574 paddd xmm5,xmm6 ; xmm5=data1H
576 paddd xmm4,[GOTOFF(ebx,PD_DESCALE_P2)]
577 paddd xmm1,[GOTOFF(ebx,PD_DESCALE_P2)]
578 psrad xmm4,DESCALE_P2
579 psrad xmm1,DESCALE_P2
580 paddd xmm0,[GOTOFF(ebx,PD_DESCALE_P2)]
581 paddd xmm5,[GOTOFF(ebx,PD_DESCALE_P2)]
582 psrad xmm0,DESCALE_P2
583 psrad xmm5,DESCALE_P2
585 packssdw xmm4,xmm1 ; xmm4=data7
586 packssdw xmm0,xmm5 ; xmm0=data1
588 movdqa XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)], xmm4
589 movdqa XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)], xmm0
591 movdqa xmm1,xmm3
592 movdqa xmm5,xmm3
593 punpcklwd xmm1,xmm7
594 punpckhwd xmm5,xmm7
595 movdqa xmm3,xmm1
596 movdqa xmm7,xmm5
597 pmaddwd xmm1,[GOTOFF(ebx,PW_MF050_MF256)] ; xmm1=tmp5L
598 pmaddwd xmm5,[GOTOFF(ebx,PW_MF050_MF256)] ; xmm5=tmp5H
599 pmaddwd xmm3,[GOTOFF(ebx,PW_MF256_F050)] ; xmm3=tmp6L
600 pmaddwd xmm7,[GOTOFF(ebx,PW_MF256_F050)] ; xmm7=tmp6H
602 paddd xmm1,xmm2 ; xmm1=data5L
603 paddd xmm5,xmm6 ; xmm5=data5H
604 paddd xmm3, XMMWORD [wk(0)] ; xmm3=data3L
605 paddd xmm7, XMMWORD [wk(1)] ; xmm7=data3H
607 paddd xmm1,[GOTOFF(ebx,PD_DESCALE_P2)]
608 paddd xmm5,[GOTOFF(ebx,PD_DESCALE_P2)]
609 psrad xmm1,DESCALE_P2
610 psrad xmm5,DESCALE_P2
611 paddd xmm3,[GOTOFF(ebx,PD_DESCALE_P2)]
612 paddd xmm7,[GOTOFF(ebx,PD_DESCALE_P2)]
613 psrad xmm3,DESCALE_P2
614 psrad xmm7,DESCALE_P2
616 packssdw xmm1,xmm5 ; xmm1=data5
617 packssdw xmm3,xmm7 ; xmm3=data3
619 movdqa XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)], xmm1
620 movdqa XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)], xmm3
622 ; pop edi ; unused
623 ; pop esi ; unused
624 ; pop edx ; need not be preserved
625 ; pop ecx ; unused
626 poppic ebx
627 mov esp,ebp ; esp <- aligned ebp
628 pop esp ; esp <- original ebp
629 pop ebp
630 ret
632 ; For some reason, the OS X linker does not honor the request to align the
633 ; segment unless we do this.
634 align 16
