blob: a7e73f73a2045676cf1ecfa1c6d6c7bc23592028 [file] [log] [blame]
MIYASAKA Masarua2e6a9d2006-02-04 00:00:00 +00001;
2; jfmmxint.asm - accurate integer FDCT (MMX)
3;
Pierre Ossman5eb84ff2009-03-09 13:25:30 +00004; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5;
6; Based on
MIYASAKA Masarua2e6a9d2006-02-04 00:00:00 +00007; 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;
MIYASAKA Masarua2e6a9d2006-02-04 00:00:00 +000022; [TAB8]
23
Pierre Ossman3a65ef42009-03-16 13:34:18 +000024%include "jsimdext.inc"
25%include "jdct.inc"
MIYASAKA Masarua2e6a9d2006-02-04 00:00:00 +000026
27; --------------------------------------------------------------------------
28
29%define CONST_BITS 13
30%define PASS1_BITS 2
31
32%define DESCALE_P1 (CONST_BITS-PASS1_BITS)
33%define DESCALE_P2 (CONST_BITS+PASS1_BITS)
34
35%if CONST_BITS == 13
36F_0_298 equ 2446 ; FIX(0.298631336)
37F_0_390 equ 3196 ; FIX(0.390180644)
38F_0_541 equ 4433 ; FIX(0.541196100)
39F_0_765 equ 6270 ; FIX(0.765366865)
40F_0_899 equ 7373 ; FIX(0.899976223)
41F_1_175 equ 9633 ; FIX(1.175875602)
42F_1_501 equ 12299 ; FIX(1.501321110)
43F_1_847 equ 15137 ; FIX(1.847759065)
44F_1_961 equ 16069 ; FIX(1.961570560)
45F_2_053 equ 16819 ; FIX(2.053119869)
46F_2_562 equ 20995 ; FIX(2.562915447)
47F_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))
51F_0_298 equ DESCALE( 320652955,30-CONST_BITS) ; FIX(0.298631336)
52F_0_390 equ DESCALE( 418953276,30-CONST_BITS) ; FIX(0.390180644)
53F_0_541 equ DESCALE( 581104887,30-CONST_BITS) ; FIX(0.541196100)
54F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
55F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
56F_1_175 equ DESCALE(1262586813,30-CONST_BITS) ; FIX(1.175875602)
57F_1_501 equ DESCALE(1612031267,30-CONST_BITS) ; FIX(1.501321110)
58F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
59F_1_961 equ DESCALE(2106220350,30-CONST_BITS) ; FIX(1.961570560)
60F_2_053 equ DESCALE(2204520673,30-CONST_BITS) ; FIX(2.053119869)
61F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
62F_3_072 equ DESCALE(3299298341,30-CONST_BITS) ; FIX(3.072711026)
63%endif
64
65; --------------------------------------------------------------------------
66 SECTION SEG_CONST
67
68 alignz 16
69 global EXTN(jconst_fdct_islow_mmx)
70
71EXTN(jconst_fdct_islow_mmx):
72
73PW_F130_F054 times 2 dw (F_0_541+F_0_765), F_0_541
74PW_F054_MF130 times 2 dw F_0_541, (F_0_541-F_1_847)
75PW_MF078_F117 times 2 dw (F_1_175-F_1_961), F_1_175
76PW_F117_F078 times 2 dw F_1_175, (F_1_175-F_0_390)
77PW_MF060_MF089 times 2 dw (F_0_298-F_0_899),-F_0_899
78PW_MF089_F060 times 2 dw -F_0_899, (F_1_501-F_0_899)
79PW_MF050_MF256 times 2 dw (F_2_053-F_2_562),-F_2_562
80PW_MF256_F050 times 2 dw -F_2_562, (F_3_072-F_2_562)
81PD_DESCALE_P1 times 2 dd 1 << (DESCALE_P1-1)
82PD_DESCALE_P2 times 2 dd 1 << (DESCALE_P2-1)
83PW_DESCALE_P2X times 4 dw 1 << (PASS1_BITS-1)
84
85 alignz 16
86
87; --------------------------------------------------------------------------
88 SECTION SEG_TEXT
89 BITS 32
90;
91; Perform the forward DCT on one block of samples.
92;
93; GLOBAL(void)
Pierre Ossman5eb84ff2009-03-09 13:25:30 +000094; jsimd_fdct_islow_mmx (DCTELEM * data)
MIYASAKA Masarua2e6a9d2006-02-04 00:00:00 +000095;
96
97%define data(b) (b)+8 ; DCTELEM * data
98
99%define original_ebp ebp+0
100%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM]
101%define WK_NUM 2
102
103 align 16
Pierre Ossman5eb84ff2009-03-09 13:25:30 +0000104 global EXTN(jsimd_fdct_islow_mmx)
MIYASAKA Masarua2e6a9d2006-02-04 00:00:00 +0000105
Pierre Ossman5eb84ff2009-03-09 13:25:30 +0000106EXTN(jsimd_fdct_islow_mmx):
MIYASAKA Masarua2e6a9d2006-02-04 00:00:00 +0000107 push ebp
108 mov eax,esp ; eax = original ebp
109 sub esp, byte 4
110 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
111 mov [esp],eax
112 mov ebp,esp ; ebp = aligned ebp
113 lea esp, [wk(0)]
114 pushpic ebx
115; push ecx ; need not be preserved
116; push edx ; need not be preserved
117; push esi ; unused
118; push edi ; unused
119
120 get_GOT ebx ; get GOT address
121
122 ; ---- Pass 1: process rows.
123
124 mov edx, POINTER [data(eax)] ; (DCTELEM *)
125 mov ecx, DCTSIZE/4
126 alignx 16,7
127.rowloop:
128
129 movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
130 movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
131 movq mm2, MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)]
132 movq mm3, MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)]
133
134 ; mm0=(20 21 22 23), mm2=(24 25 26 27)
135 ; mm1=(30 31 32 33), mm3=(34 35 36 37)
136
137 movq mm4,mm0 ; transpose coefficients(phase 1)
138 punpcklwd mm0,mm1 ; mm0=(20 30 21 31)
139 punpckhwd mm4,mm1 ; mm4=(22 32 23 33)
140 movq mm5,mm2 ; transpose coefficients(phase 1)
141 punpcklwd mm2,mm3 ; mm2=(24 34 25 35)
142 punpckhwd mm5,mm3 ; mm5=(26 36 27 37)
143
144 movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
145 movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
146 movq mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)]
147 movq mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)]
148
149 ; mm6=(00 01 02 03), mm1=(04 05 06 07)
150 ; mm7=(10 11 12 13), mm3=(14 15 16 17)
151
152 movq MMWORD [wk(0)], mm4 ; wk(0)=(22 32 23 33)
153 movq MMWORD [wk(1)], mm2 ; wk(1)=(24 34 25 35)
154
155 movq mm4,mm6 ; transpose coefficients(phase 1)
156 punpcklwd mm6,mm7 ; mm6=(00 10 01 11)
157 punpckhwd mm4,mm7 ; mm4=(02 12 03 13)
158 movq mm2,mm1 ; transpose coefficients(phase 1)
159 punpcklwd mm1,mm3 ; mm1=(04 14 05 15)
160 punpckhwd mm2,mm3 ; mm2=(06 16 07 17)
161
162 movq mm7,mm6 ; transpose coefficients(phase 2)
163 punpckldq mm6,mm0 ; mm6=(00 10 20 30)=data0
164 punpckhdq mm7,mm0 ; mm7=(01 11 21 31)=data1
165 movq mm3,mm2 ; transpose coefficients(phase 2)
166 punpckldq mm2,mm5 ; mm2=(06 16 26 36)=data6
167 punpckhdq mm3,mm5 ; mm3=(07 17 27 37)=data7
168
169 movq mm0,mm7
170 movq mm5,mm6
171 psubw mm7,mm2 ; mm7=data1-data6=tmp6
172 psubw mm6,mm3 ; mm6=data0-data7=tmp7
173 paddw mm0,mm2 ; mm0=data1+data6=tmp1
174 paddw mm5,mm3 ; mm5=data0+data7=tmp0
175
176 movq mm2, MMWORD [wk(0)] ; mm2=(22 32 23 33)
177 movq mm3, MMWORD [wk(1)] ; mm3=(24 34 25 35)
178 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6
179 movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7
180
181 movq mm7,mm4 ; transpose coefficients(phase 2)
182 punpckldq mm4,mm2 ; mm4=(02 12 22 32)=data2
183 punpckhdq mm7,mm2 ; mm7=(03 13 23 33)=data3
184 movq mm6,mm1 ; transpose coefficients(phase 2)
185 punpckldq mm1,mm3 ; mm1=(04 14 24 34)=data4
186 punpckhdq mm6,mm3 ; mm6=(05 15 25 35)=data5
187
188 movq mm2,mm7
189 movq mm3,mm4
190 paddw mm7,mm1 ; mm7=data3+data4=tmp3
191 paddw mm4,mm6 ; mm4=data2+data5=tmp2
192 psubw mm2,mm1 ; mm2=data3-data4=tmp4
193 psubw mm3,mm6 ; mm3=data2-data5=tmp5
194
195 ; -- Even part
196
197 movq mm1,mm5
198 movq mm6,mm0
199 paddw mm5,mm7 ; mm5=tmp10
200 paddw mm0,mm4 ; mm0=tmp11
201 psubw mm1,mm7 ; mm1=tmp13
202 psubw mm6,mm4 ; mm6=tmp12
203
204 movq mm7,mm5
205 paddw mm5,mm0 ; mm5=tmp10+tmp11
206 psubw mm7,mm0 ; mm7=tmp10-tmp11
207
208 psllw mm5,PASS1_BITS ; mm5=data0
209 psllw mm7,PASS1_BITS ; mm7=data4
210
211 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm5
212 movq MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)], mm7
213
214 ; (Original)
215 ; z1 = (tmp12 + tmp13) * 0.541196100;
216 ; data2 = z1 + tmp13 * 0.765366865;
217 ; data6 = z1 + tmp12 * -1.847759065;
218 ;
219 ; (This implementation)
220 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
221 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
222
223 movq mm4,mm1 ; mm1=tmp13
224 movq mm0,mm1
225 punpcklwd mm4,mm6 ; mm6=tmp12
226 punpckhwd mm0,mm6
227 movq mm1,mm4
228 movq mm6,mm0
229 pmaddwd mm4,[GOTOFF(ebx,PW_F130_F054)] ; mm4=data2L
230 pmaddwd mm0,[GOTOFF(ebx,PW_F130_F054)] ; mm0=data2H
231 pmaddwd mm1,[GOTOFF(ebx,PW_F054_MF130)] ; mm1=data6L
232 pmaddwd mm6,[GOTOFF(ebx,PW_F054_MF130)] ; mm6=data6H
233
234 paddd mm4,[GOTOFF(ebx,PD_DESCALE_P1)]
235 paddd mm0,[GOTOFF(ebx,PD_DESCALE_P1)]
236 psrad mm4,DESCALE_P1
237 psrad mm0,DESCALE_P1
238 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P1)]
239 paddd mm6,[GOTOFF(ebx,PD_DESCALE_P1)]
240 psrad mm1,DESCALE_P1
241 psrad mm6,DESCALE_P1
242
243 packssdw mm4,mm0 ; mm4=data2
244 packssdw mm1,mm6 ; mm1=data6
245
246 movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
247 movq MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)], mm1
248
249 ; -- Odd part
250
251 movq mm5, MMWORD [wk(0)] ; mm5=tmp6
252 movq mm7, MMWORD [wk(1)] ; mm7=tmp7
253
254 movq mm0,mm2 ; mm2=tmp4
255 movq mm6,mm3 ; mm3=tmp5
256 paddw mm0,mm5 ; mm0=z3
257 paddw mm6,mm7 ; mm6=z4
258
259 ; (Original)
260 ; z5 = (z3 + z4) * 1.175875602;
261 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
262 ; z3 += z5; z4 += z5;
263 ;
264 ; (This implementation)
265 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
266 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
267
268 movq mm4,mm0
269 movq mm1,mm0
270 punpcklwd mm4,mm6
271 punpckhwd mm1,mm6
272 movq mm0,mm4
273 movq mm6,mm1
274 pmaddwd mm4,[GOTOFF(ebx,PW_MF078_F117)] ; mm4=z3L
275 pmaddwd mm1,[GOTOFF(ebx,PW_MF078_F117)] ; mm1=z3H
276 pmaddwd mm0,[GOTOFF(ebx,PW_F117_F078)] ; mm0=z4L
277 pmaddwd mm6,[GOTOFF(ebx,PW_F117_F078)] ; mm6=z4H
278
279 movq MMWORD [wk(0)], mm4 ; wk(0)=z3L
280 movq MMWORD [wk(1)], mm1 ; wk(1)=z3H
281
282 ; (Original)
283 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
284 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
285 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
286 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
287 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
288 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
289 ;
290 ; (This implementation)
291 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
292 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
293 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
294 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
295 ; data7 = tmp4 + z3; data5 = tmp5 + z4;
296 ; data3 = tmp6 + z3; data1 = tmp7 + z4;
297
298 movq mm4,mm2
299 movq mm1,mm2
300 punpcklwd mm4,mm7
301 punpckhwd mm1,mm7
302 movq mm2,mm4
303 movq mm7,mm1
304 pmaddwd mm4,[GOTOFF(ebx,PW_MF060_MF089)] ; mm4=tmp4L
305 pmaddwd mm1,[GOTOFF(ebx,PW_MF060_MF089)] ; mm1=tmp4H
306 pmaddwd mm2,[GOTOFF(ebx,PW_MF089_F060)] ; mm2=tmp7L
307 pmaddwd mm7,[GOTOFF(ebx,PW_MF089_F060)] ; mm7=tmp7H
308
309 paddd mm4, MMWORD [wk(0)] ; mm4=data7L
310 paddd mm1, MMWORD [wk(1)] ; mm1=data7H
311 paddd mm2,mm0 ; mm2=data1L
312 paddd mm7,mm6 ; mm7=data1H
313
314 paddd mm4,[GOTOFF(ebx,PD_DESCALE_P1)]
315 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P1)]
316 psrad mm4,DESCALE_P1
317 psrad mm1,DESCALE_P1
318 paddd mm2,[GOTOFF(ebx,PD_DESCALE_P1)]
319 paddd mm7,[GOTOFF(ebx,PD_DESCALE_P1)]
320 psrad mm2,DESCALE_P1
321 psrad mm7,DESCALE_P1
322
323 packssdw mm4,mm1 ; mm4=data7
324 packssdw mm2,mm7 ; mm2=data1
325
326 movq MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)], mm4
327 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm2
328
329 movq mm1,mm3
330 movq mm7,mm3
331 punpcklwd mm1,mm5
332 punpckhwd mm7,mm5
333 movq mm3,mm1
334 movq mm5,mm7
335 pmaddwd mm1,[GOTOFF(ebx,PW_MF050_MF256)] ; mm1=tmp5L
336 pmaddwd mm7,[GOTOFF(ebx,PW_MF050_MF256)] ; mm7=tmp5H
337 pmaddwd mm3,[GOTOFF(ebx,PW_MF256_F050)] ; mm3=tmp6L
338 pmaddwd mm5,[GOTOFF(ebx,PW_MF256_F050)] ; mm5=tmp6H
339
340 paddd mm1,mm0 ; mm1=data5L
341 paddd mm7,mm6 ; mm7=data5H
342 paddd mm3, MMWORD [wk(0)] ; mm3=data3L
343 paddd mm5, MMWORD [wk(1)] ; mm5=data3H
344
345 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P1)]
346 paddd mm7,[GOTOFF(ebx,PD_DESCALE_P1)]
347 psrad mm1,DESCALE_P1
348 psrad mm7,DESCALE_P1
349 paddd mm3,[GOTOFF(ebx,PD_DESCALE_P1)]
350 paddd mm5,[GOTOFF(ebx,PD_DESCALE_P1)]
351 psrad mm3,DESCALE_P1
352 psrad mm5,DESCALE_P1
353
354 packssdw mm1,mm7 ; mm1=data5
355 packssdw mm3,mm5 ; mm3=data3
356
357 movq MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)], mm1
358 movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm3
359
360 add edx, byte 4*DCTSIZE*SIZEOF_DCTELEM
361 dec ecx
362 jnz near .rowloop
363
364 ; ---- Pass 2: process columns.
365
366 mov edx, POINTER [data(eax)] ; (DCTELEM *)
367 mov ecx, DCTSIZE/4
368 alignx 16,7
369.columnloop:
370
371 movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
372 movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
373 movq mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
374 movq mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
375
376 ; mm0=(02 12 22 32), mm2=(42 52 62 72)
377 ; mm1=(03 13 23 33), mm3=(43 53 63 73)
378
379 movq mm4,mm0 ; transpose coefficients(phase 1)
380 punpcklwd mm0,mm1 ; mm0=(02 03 12 13)
381 punpckhwd mm4,mm1 ; mm4=(22 23 32 33)
382 movq mm5,mm2 ; transpose coefficients(phase 1)
383 punpcklwd mm2,mm3 ; mm2=(42 43 52 53)
384 punpckhwd mm5,mm3 ; mm5=(62 63 72 73)
385
386 movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
387 movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
388 movq mm1, MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
389 movq mm3, MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
390
391 ; mm6=(00 10 20 30), mm1=(40 50 60 70)
392 ; mm7=(01 11 21 31), mm3=(41 51 61 71)
393
394 movq MMWORD [wk(0)], mm4 ; wk(0)=(22 23 32 33)
395 movq MMWORD [wk(1)], mm2 ; wk(1)=(42 43 52 53)
396
397 movq mm4,mm6 ; transpose coefficients(phase 1)
398 punpcklwd mm6,mm7 ; mm6=(00 01 10 11)
399 punpckhwd mm4,mm7 ; mm4=(20 21 30 31)
400 movq mm2,mm1 ; transpose coefficients(phase 1)
401 punpcklwd mm1,mm3 ; mm1=(40 41 50 51)
402 punpckhwd mm2,mm3 ; mm2=(60 61 70 71)
403
404 movq mm7,mm6 ; transpose coefficients(phase 2)
405 punpckldq mm6,mm0 ; mm6=(00 01 02 03)=data0
406 punpckhdq mm7,mm0 ; mm7=(10 11 12 13)=data1
407 movq mm3,mm2 ; transpose coefficients(phase 2)
408 punpckldq mm2,mm5 ; mm2=(60 61 62 63)=data6
409 punpckhdq mm3,mm5 ; mm3=(70 71 72 73)=data7
410
411 movq mm0,mm7
412 movq mm5,mm6
413 psubw mm7,mm2 ; mm7=data1-data6=tmp6
414 psubw mm6,mm3 ; mm6=data0-data7=tmp7
415 paddw mm0,mm2 ; mm0=data1+data6=tmp1
416 paddw mm5,mm3 ; mm5=data0+data7=tmp0
417
418 movq mm2, MMWORD [wk(0)] ; mm2=(22 23 32 33)
419 movq mm3, MMWORD [wk(1)] ; mm3=(42 43 52 53)
420 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6
421 movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7
422
423 movq mm7,mm4 ; transpose coefficients(phase 2)
424 punpckldq mm4,mm2 ; mm4=(20 21 22 23)=data2
425 punpckhdq mm7,mm2 ; mm7=(30 31 32 33)=data3
426 movq mm6,mm1 ; transpose coefficients(phase 2)
427 punpckldq mm1,mm3 ; mm1=(40 41 42 43)=data4
428 punpckhdq mm6,mm3 ; mm6=(50 51 52 53)=data5
429
430 movq mm2,mm7
431 movq mm3,mm4
432 paddw mm7,mm1 ; mm7=data3+data4=tmp3
433 paddw mm4,mm6 ; mm4=data2+data5=tmp2
434 psubw mm2,mm1 ; mm2=data3-data4=tmp4
435 psubw mm3,mm6 ; mm3=data2-data5=tmp5
436
437 ; -- Even part
438
439 movq mm1,mm5
440 movq mm6,mm0
441 paddw mm5,mm7 ; mm5=tmp10
442 paddw mm0,mm4 ; mm0=tmp11
443 psubw mm1,mm7 ; mm1=tmp13
444 psubw mm6,mm4 ; mm6=tmp12
445
446 movq mm7,mm5
447 paddw mm5,mm0 ; mm5=tmp10+tmp11
448 psubw mm7,mm0 ; mm7=tmp10-tmp11
449
450 paddw mm5,[GOTOFF(ebx,PW_DESCALE_P2X)]
451 paddw mm7,[GOTOFF(ebx,PW_DESCALE_P2X)]
452 psraw mm5,PASS1_BITS ; mm5=data0
453 psraw mm7,PASS1_BITS ; mm7=data4
454
455 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm5
456 movq MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)], mm7
457
458 ; (Original)
459 ; z1 = (tmp12 + tmp13) * 0.541196100;
460 ; data2 = z1 + tmp13 * 0.765366865;
461 ; data6 = z1 + tmp12 * -1.847759065;
462 ;
463 ; (This implementation)
464 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
465 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
466
467 movq mm4,mm1 ; mm1=tmp13
468 movq mm0,mm1
469 punpcklwd mm4,mm6 ; mm6=tmp12
470 punpckhwd mm0,mm6
471 movq mm1,mm4
472 movq mm6,mm0
473 pmaddwd mm4,[GOTOFF(ebx,PW_F130_F054)] ; mm4=data2L
474 pmaddwd mm0,[GOTOFF(ebx,PW_F130_F054)] ; mm0=data2H
475 pmaddwd mm1,[GOTOFF(ebx,PW_F054_MF130)] ; mm1=data6L
476 pmaddwd mm6,[GOTOFF(ebx,PW_F054_MF130)] ; mm6=data6H
477
478 paddd mm4,[GOTOFF(ebx,PD_DESCALE_P2)]
479 paddd mm0,[GOTOFF(ebx,PD_DESCALE_P2)]
480 psrad mm4,DESCALE_P2
481 psrad mm0,DESCALE_P2
482 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P2)]
483 paddd mm6,[GOTOFF(ebx,PD_DESCALE_P2)]
484 psrad mm1,DESCALE_P2
485 psrad mm6,DESCALE_P2
486
487 packssdw mm4,mm0 ; mm4=data2
488 packssdw mm1,mm6 ; mm1=data6
489
490 movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
491 movq MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)], mm1
492
493 ; -- Odd part
494
495 movq mm5, MMWORD [wk(0)] ; mm5=tmp6
496 movq mm7, MMWORD [wk(1)] ; mm7=tmp7
497
498 movq mm0,mm2 ; mm2=tmp4
499 movq mm6,mm3 ; mm3=tmp5
500 paddw mm0,mm5 ; mm0=z3
501 paddw mm6,mm7 ; mm6=z4
502
503 ; (Original)
504 ; z5 = (z3 + z4) * 1.175875602;
505 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
506 ; z3 += z5; z4 += z5;
507 ;
508 ; (This implementation)
509 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
510 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
511
512 movq mm4,mm0
513 movq mm1,mm0
514 punpcklwd mm4,mm6
515 punpckhwd mm1,mm6
516 movq mm0,mm4
517 movq mm6,mm1
518 pmaddwd mm4,[GOTOFF(ebx,PW_MF078_F117)] ; mm4=z3L
519 pmaddwd mm1,[GOTOFF(ebx,PW_MF078_F117)] ; mm1=z3H
520 pmaddwd mm0,[GOTOFF(ebx,PW_F117_F078)] ; mm0=z4L
521 pmaddwd mm6,[GOTOFF(ebx,PW_F117_F078)] ; mm6=z4H
522
523 movq MMWORD [wk(0)], mm4 ; wk(0)=z3L
524 movq MMWORD [wk(1)], mm1 ; wk(1)=z3H
525
526 ; (Original)
527 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
528 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
529 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
530 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
531 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
532 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
533 ;
534 ; (This implementation)
535 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
536 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
537 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
538 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
539 ; data7 = tmp4 + z3; data5 = tmp5 + z4;
540 ; data3 = tmp6 + z3; data1 = tmp7 + z4;
541
542 movq mm4,mm2
543 movq mm1,mm2
544 punpcklwd mm4,mm7
545 punpckhwd mm1,mm7
546 movq mm2,mm4
547 movq mm7,mm1
548 pmaddwd mm4,[GOTOFF(ebx,PW_MF060_MF089)] ; mm4=tmp4L
549 pmaddwd mm1,[GOTOFF(ebx,PW_MF060_MF089)] ; mm1=tmp4H
550 pmaddwd mm2,[GOTOFF(ebx,PW_MF089_F060)] ; mm2=tmp7L
551 pmaddwd mm7,[GOTOFF(ebx,PW_MF089_F060)] ; mm7=tmp7H
552
553 paddd mm4, MMWORD [wk(0)] ; mm4=data7L
554 paddd mm1, MMWORD [wk(1)] ; mm1=data7H
555 paddd mm2,mm0 ; mm2=data1L
556 paddd mm7,mm6 ; mm7=data1H
557
558 paddd mm4,[GOTOFF(ebx,PD_DESCALE_P2)]
559 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P2)]
560 psrad mm4,DESCALE_P2
561 psrad mm1,DESCALE_P2
562 paddd mm2,[GOTOFF(ebx,PD_DESCALE_P2)]
563 paddd mm7,[GOTOFF(ebx,PD_DESCALE_P2)]
564 psrad mm2,DESCALE_P2
565 psrad mm7,DESCALE_P2
566
567 packssdw mm4,mm1 ; mm4=data7
568 packssdw mm2,mm7 ; mm2=data1
569
570 movq MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)], mm4
571 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm2
572
573 movq mm1,mm3
574 movq mm7,mm3
575 punpcklwd mm1,mm5
576 punpckhwd mm7,mm5
577 movq mm3,mm1
578 movq mm5,mm7
579 pmaddwd mm1,[GOTOFF(ebx,PW_MF050_MF256)] ; mm1=tmp5L
580 pmaddwd mm7,[GOTOFF(ebx,PW_MF050_MF256)] ; mm7=tmp5H
581 pmaddwd mm3,[GOTOFF(ebx,PW_MF256_F050)] ; mm3=tmp6L
582 pmaddwd mm5,[GOTOFF(ebx,PW_MF256_F050)] ; mm5=tmp6H
583
584 paddd mm1,mm0 ; mm1=data5L
585 paddd mm7,mm6 ; mm7=data5H
586 paddd mm3, MMWORD [wk(0)] ; mm3=data3L
587 paddd mm5, MMWORD [wk(1)] ; mm5=data3H
588
589 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P2)]
590 paddd mm7,[GOTOFF(ebx,PD_DESCALE_P2)]
591 psrad mm1,DESCALE_P2
592 psrad mm7,DESCALE_P2
593 paddd mm3,[GOTOFF(ebx,PD_DESCALE_P2)]
594 paddd mm5,[GOTOFF(ebx,PD_DESCALE_P2)]
595 psrad mm3,DESCALE_P2
596 psrad mm5,DESCALE_P2
597
598 packssdw mm1,mm7 ; mm1=data5
599 packssdw mm3,mm5 ; mm3=data3
600
601 movq MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)], mm1
602 movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm3
603
604 add edx, byte 4*SIZEOF_DCTELEM
605 dec ecx
606 jnz near .columnloop
607
608 emms ; empty MMX state
609
610; pop edi ; unused
611; pop esi ; unused
612; pop edx ; need not be preserved
613; pop ecx ; need not be preserved
614 poppic ebx
615 mov esp,ebp ; esp <- aligned ebp
616 pop esp ; esp <- original ebp
617 pop ebp
618 ret
619
DRC132b5fd2009-10-08 09:04:56 +0000620; For some reason, the OS X linker does not honor the request to align the
621; segment unless we do this.
622 align 16