• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jidctfst.asm - fast integer IDCT (MMX)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright (C) 2016, D. R. Commander.
6;
7; Based on the 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 fast, not so accurate integer implementation of
18; the inverse DCT (Discrete Cosine Transform). The following code is
19; based directly on the IJG's original jidctfst.c; see the jidctfst.c
20; for more details.
21
22%include "jsimdext.inc"
23%include "jdct.inc"
24
25; --------------------------------------------------------------------------
26
27%define CONST_BITS  8  ; 14 is also OK.
28%define PASS1_BITS  2
29
30%if IFAST_SCALE_BITS != PASS1_BITS
31%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
32%endif
33
34%if CONST_BITS == 8
35F_1_082 equ 277              ; FIX(1.082392200)
36F_1_414 equ 362              ; FIX(1.414213562)
37F_1_847 equ 473              ; FIX(1.847759065)
38F_2_613 equ 669              ; FIX(2.613125930)
39F_1_613 equ (F_2_613 - 256)  ; FIX(2.613125930) - FIX(1)
40%else
41; NASM cannot do compile-time arithmetic on floating-point constants.
42%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
43F_1_082 equ DESCALE(1162209775, 30 - CONST_BITS)  ; FIX(1.082392200)
44F_1_414 equ DESCALE(1518500249, 30 - CONST_BITS)  ; FIX(1.414213562)
45F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
46F_2_613 equ DESCALE(2805822602, 30 - CONST_BITS)  ; FIX(2.613125930)
47F_1_613 equ (F_2_613 - (1 << CONST_BITS))       ; FIX(2.613125930) - FIX(1)
48%endif
49
50; --------------------------------------------------------------------------
51    SECTION     SEG_CONST
52
53; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
54; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
55
56%define PRE_MULTIPLY_SCALE_BITS  2
57%define CONST_SHIFT              (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
58
59    alignz      32
60    GLOBAL_DATA(jconst_idct_ifast_mmx)
61
62EXTN(jconst_idct_ifast_mmx):
63
64PW_F1414       times 4 dw  F_1_414 << CONST_SHIFT
65PW_F1847       times 4 dw  F_1_847 << CONST_SHIFT
66PW_MF1613      times 4 dw -F_1_613 << CONST_SHIFT
67PW_F1082       times 4 dw  F_1_082 << CONST_SHIFT
68PB_CENTERJSAMP times 8 db  CENTERJSAMPLE
69
70    alignz      32
71
72; --------------------------------------------------------------------------
73    SECTION     SEG_TEXT
74    BITS        32
75;
76; Perform dequantization and inverse DCT on one block of coefficients.
77;
78; GLOBAL(void)
79; jsimd_idct_ifast_mmx(void *dct_table, JCOEFPTR coef_block,
80;                      JSAMPARRAY output_buf, JDIMENSION output_col)
81;
82
83%define dct_table(b)   (b) + 8          ; jpeg_component_info *compptr
84%define coef_block(b)  (b) + 12         ; JCOEFPTR coef_block
85%define output_buf(b)  (b) + 16         ; JSAMPARRAY output_buf
86%define output_col(b)  (b) + 20         ; JDIMENSION output_col
87
88%define original_ebp   ebp + 0
89%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_MMWORD
90                                        ; mmword wk[WK_NUM]
91%define WK_NUM         2
92%define workspace      wk(0) - DCTSIZE2 * SIZEOF_JCOEF
93                                        ; JCOEF workspace[DCTSIZE2]
94
95    align       32
96    GLOBAL_FUNCTION(jsimd_idct_ifast_mmx)
97
98EXTN(jsimd_idct_ifast_mmx):
99    push        ebp
100    mov         eax, esp                    ; eax = original ebp
101    sub         esp, byte 4
102    and         esp, byte (-SIZEOF_MMWORD)  ; align to 64 bits
103    mov         [esp], eax
104    mov         ebp, esp                    ; ebp = aligned ebp
105    lea         esp, [workspace]
106    push        ebx
107;   push        ecx                     ; need not be preserved
108;   push        edx                     ; need not be preserved
109    push        esi
110    push        edi
111
112    get_GOT     ebx                     ; get GOT address
113
114    ; ---- Pass 1: process columns from input, store into work array.
115
116;   mov         eax, [original_ebp]
117    mov         edx, POINTER [dct_table(eax)]    ; quantptr
118    mov         esi, JCOEFPTR [coef_block(eax)]  ; inptr
119    lea         edi, [workspace]                 ; JCOEF *wsptr
120    mov         ecx, DCTSIZE/4                   ; ctr
121    alignx      16, 7
122.columnloop:
123%ifndef NO_ZERO_COLUMN_TEST_IFAST_MMX
124    mov         eax, dword [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
125    or          eax, dword [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
126    jnz         short .columnDCT
127
128    movq        mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
129    movq        mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
130    por         mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
131    por         mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
132    por         mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
133    por         mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
134    por         mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
135    por         mm1, mm0
136    packsswb    mm1, mm1
137    movd        eax, mm1
138    test        eax, eax
139    jnz         short .columnDCT
140
141    ; -- AC terms all zero
142
143    movq        mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
144    pmullw      mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)]
145
146    movq        mm2, mm0                ; mm0=in0=(00 01 02 03)
147    punpcklwd   mm0, mm0                ; mm0=(00 00 01 01)
148    punpckhwd   mm2, mm2                ; mm2=(02 02 03 03)
149
150    movq        mm1, mm0
151    punpckldq   mm0, mm0                ; mm0=(00 00 00 00)
152    punpckhdq   mm1, mm1                ; mm1=(01 01 01 01)
153    movq        mm3, mm2
154    punpckldq   mm2, mm2                ; mm2=(02 02 02 02)
155    punpckhdq   mm3, mm3                ; mm3=(03 03 03 03)
156
157    movq        MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
158    movq        MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0
159    movq        MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
160    movq        MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm1
161    movq        MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
162    movq        MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2
163    movq        MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
164    movq        MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm3
165    jmp         near .nextcolumn
166    alignx      16, 7
167%endif
168.columnDCT:
169
170    ; -- Even part
171
172    movq        mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
173    movq        mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
174    pmullw      mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)]
175    pmullw      mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_IFAST_MULT_TYPE)]
176    movq        mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
177    movq        mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
178    pmullw      mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_IFAST_MULT_TYPE)]
179    pmullw      mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_IFAST_MULT_TYPE)]
180
181    movq        mm4, mm0
182    movq        mm5, mm1
183    psubw       mm0, mm2                ; mm0=tmp11
184    psubw       mm1, mm3
185    paddw       mm4, mm2                ; mm4=tmp10
186    paddw       mm5, mm3                ; mm5=tmp13
187
188    psllw       mm1, PRE_MULTIPLY_SCALE_BITS
189    pmulhw      mm1, [GOTOFF(ebx,PW_F1414)]
190    psubw       mm1, mm5                ; mm1=tmp12
191
192    movq        mm6, mm4
193    movq        mm7, mm0
194    psubw       mm4, mm5                ; mm4=tmp3
195    psubw       mm0, mm1                ; mm0=tmp2
196    paddw       mm6, mm5                ; mm6=tmp0
197    paddw       mm7, mm1                ; mm7=tmp1
198
199    movq        MMWORD [wk(1)], mm4     ; wk(1)=tmp3
200    movq        MMWORD [wk(0)], mm0     ; wk(0)=tmp2
201
202    ; -- Odd part
203
204    movq        mm2, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
205    movq        mm3, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
206    pmullw      mm2, MMWORD [MMBLOCK(1,0,edx,SIZEOF_IFAST_MULT_TYPE)]
207    pmullw      mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_IFAST_MULT_TYPE)]
208    movq        mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
209    movq        mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
210    pmullw      mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_IFAST_MULT_TYPE)]
211    pmullw      mm1, MMWORD [MMBLOCK(7,0,edx,SIZEOF_IFAST_MULT_TYPE)]
212
213    movq        mm4, mm2
214    movq        mm0, mm5
215    psubw       mm2, mm1                ; mm2=z12
216    psubw       mm5, mm3                ; mm5=z10
217    paddw       mm4, mm1                ; mm4=z11
218    paddw       mm0, mm3                ; mm0=z13
219
220    movq        mm1, mm5                ; mm1=z10(unscaled)
221    psllw       mm2, PRE_MULTIPLY_SCALE_BITS
222    psllw       mm5, PRE_MULTIPLY_SCALE_BITS
223
224    movq        mm3, mm4
225    psubw       mm4, mm0
226    paddw       mm3, mm0                ; mm3=tmp7
227
228    psllw       mm4, PRE_MULTIPLY_SCALE_BITS
229    pmulhw      mm4, [GOTOFF(ebx,PW_F1414)]  ; mm4=tmp11
230
231    ; To avoid overflow...
232    ;
233    ; (Original)
234    ; tmp12 = -2.613125930 * z10 + z5;
235    ;
236    ; (This implementation)
237    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
238    ;       = -1.613125930 * z10 - z10 + z5;
239
240    movq        mm0, mm5
241    paddw       mm5, mm2
242    pmulhw      mm5, [GOTOFF(ebx,PW_F1847)]   ; mm5=z5
243    pmulhw      mm0, [GOTOFF(ebx,PW_MF1613)]
244    pmulhw      mm2, [GOTOFF(ebx,PW_F1082)]
245    psubw       mm0, mm1
246    psubw       mm2, mm5                ; mm2=tmp10
247    paddw       mm0, mm5                ; mm0=tmp12
248
249    ; -- Final output stage
250
251    psubw       mm0, mm3                ; mm0=tmp6
252    movq        mm1, mm6
253    movq        mm5, mm7
254    paddw       mm6, mm3                ; mm6=data0=(00 01 02 03)
255    paddw       mm7, mm0                ; mm7=data1=(10 11 12 13)
256    psubw       mm1, mm3                ; mm1=data7=(70 71 72 73)
257    psubw       mm5, mm0                ; mm5=data6=(60 61 62 63)
258    psubw       mm4, mm0                ; mm4=tmp5
259
260    movq        mm3, mm6                ; transpose coefficients(phase 1)
261    punpcklwd   mm6, mm7                ; mm6=(00 10 01 11)
262    punpckhwd   mm3, mm7                ; mm3=(02 12 03 13)
263    movq        mm0, mm5                ; transpose coefficients(phase 1)
264    punpcklwd   mm5, mm1                ; mm5=(60 70 61 71)
265    punpckhwd   mm0, mm1                ; mm0=(62 72 63 73)
266
267    movq        mm7, MMWORD [wk(0)]     ; mm7=tmp2
268    movq        mm1, MMWORD [wk(1)]     ; mm1=tmp3
269
270    movq        MMWORD [wk(0)], mm5     ; wk(0)=(60 70 61 71)
271    movq        MMWORD [wk(1)], mm0     ; wk(1)=(62 72 63 73)
272
273    paddw       mm2, mm4                ; mm2=tmp4
274    movq        mm5, mm7
275    movq        mm0, mm1
276    paddw       mm7, mm4                ; mm7=data2=(20 21 22 23)
277    paddw       mm1, mm2                ; mm1=data4=(40 41 42 43)
278    psubw       mm5, mm4                ; mm5=data5=(50 51 52 53)
279    psubw       mm0, mm2                ; mm0=data3=(30 31 32 33)
280
281    movq        mm4, mm7                ; transpose coefficients(phase 1)
282    punpcklwd   mm7, mm0                ; mm7=(20 30 21 31)
283    punpckhwd   mm4, mm0                ; mm4=(22 32 23 33)
284    movq        mm2, mm1                ; transpose coefficients(phase 1)
285    punpcklwd   mm1, mm5                ; mm1=(40 50 41 51)
286    punpckhwd   mm2, mm5                ; mm2=(42 52 43 53)
287
288    movq        mm0, mm6                ; transpose coefficients(phase 2)
289    punpckldq   mm6, mm7                ; mm6=(00 10 20 30)
290    punpckhdq   mm0, mm7                ; mm0=(01 11 21 31)
291    movq        mm5, mm3                ; transpose coefficients(phase 2)
292    punpckldq   mm3, mm4                ; mm3=(02 12 22 32)
293    punpckhdq   mm5, mm4                ; mm5=(03 13 23 33)
294
295    movq        mm7, MMWORD [wk(0)]     ; mm7=(60 70 61 71)
296    movq        mm4, MMWORD [wk(1)]     ; mm4=(62 72 63 73)
297
298    movq        MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm6
299    movq        MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm0
300    movq        MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm3
301    movq        MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm5
302
303    movq        mm6, mm1                ; transpose coefficients(phase 2)
304    punpckldq   mm1, mm7                ; mm1=(40 50 60 70)
305    punpckhdq   mm6, mm7                ; mm6=(41 51 61 71)
306    movq        mm0, mm2                ; transpose coefficients(phase 2)
307    punpckldq   mm2, mm4                ; mm2=(42 52 62 72)
308    punpckhdq   mm0, mm4                ; mm0=(43 53 63 73)
309
310    movq        MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm1
311    movq        MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm6
312    movq        MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2
313    movq        MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm0
314
315.nextcolumn:
316    add         esi, byte 4*SIZEOF_JCOEF            ; coef_block
317    add         edx, byte 4*SIZEOF_IFAST_MULT_TYPE  ; quantptr
318    add         edi, byte 4*DCTSIZE*SIZEOF_JCOEF    ; wsptr
319    dec         ecx                                 ; ctr
320    jnz         near .columnloop
321
322    ; ---- Pass 2: process rows from work array, store into output array.
323
324    mov         eax, [original_ebp]
325    lea         esi, [workspace]                   ; JCOEF *wsptr
326    mov         edi, JSAMPARRAY [output_buf(eax)]  ; (JSAMPROW *)
327    mov         eax, JDIMENSION [output_col(eax)]
328    mov         ecx, DCTSIZE/4                     ; ctr
329    alignx      16, 7
330.rowloop:
331
332    ; -- Even part
333
334    movq        mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
335    movq        mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
336    movq        mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
337    movq        mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
338
339    movq        mm4, mm0
340    movq        mm5, mm1
341    psubw       mm0, mm2                ; mm0=tmp11
342    psubw       mm1, mm3
343    paddw       mm4, mm2                ; mm4=tmp10
344    paddw       mm5, mm3                ; mm5=tmp13
345
346    psllw       mm1, PRE_MULTIPLY_SCALE_BITS
347    pmulhw      mm1, [GOTOFF(ebx,PW_F1414)]
348    psubw       mm1, mm5                ; mm1=tmp12
349
350    movq        mm6, mm4
351    movq        mm7, mm0
352    psubw       mm4, mm5                ; mm4=tmp3
353    psubw       mm0, mm1                ; mm0=tmp2
354    paddw       mm6, mm5                ; mm6=tmp0
355    paddw       mm7, mm1                ; mm7=tmp1
356
357    movq        MMWORD [wk(1)], mm4     ; wk(1)=tmp3
358    movq        MMWORD [wk(0)], mm0     ; wk(0)=tmp2
359
360    ; -- Odd part
361
362    movq        mm2, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
363    movq        mm3, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
364    movq        mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
365    movq        mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
366
367    movq        mm4, mm2
368    movq        mm0, mm5
369    psubw       mm2, mm1                ; mm2=z12
370    psubw       mm5, mm3                ; mm5=z10
371    paddw       mm4, mm1                ; mm4=z11
372    paddw       mm0, mm3                ; mm0=z13
373
374    movq        mm1, mm5                ; mm1=z10(unscaled)
375    psllw       mm2, PRE_MULTIPLY_SCALE_BITS
376    psllw       mm5, PRE_MULTIPLY_SCALE_BITS
377
378    movq        mm3, mm4
379    psubw       mm4, mm0
380    paddw       mm3, mm0                ; mm3=tmp7
381
382    psllw       mm4, PRE_MULTIPLY_SCALE_BITS
383    pmulhw      mm4, [GOTOFF(ebx,PW_F1414)]  ; mm4=tmp11
384
385    ; To avoid overflow...
386    ;
387    ; (Original)
388    ; tmp12 = -2.613125930 * z10 + z5;
389    ;
390    ; (This implementation)
391    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
392    ;       = -1.613125930 * z10 - z10 + z5;
393
394    movq        mm0, mm5
395    paddw       mm5, mm2
396    pmulhw      mm5, [GOTOFF(ebx,PW_F1847)]   ; mm5=z5
397    pmulhw      mm0, [GOTOFF(ebx,PW_MF1613)]
398    pmulhw      mm2, [GOTOFF(ebx,PW_F1082)]
399    psubw       mm0, mm1
400    psubw       mm2, mm5                ; mm2=tmp10
401    paddw       mm0, mm5                ; mm0=tmp12
402
403    ; -- Final output stage
404
405    psubw       mm0, mm3                ; mm0=tmp6
406    movq        mm1, mm6
407    movq        mm5, mm7
408    paddw       mm6, mm3                ; mm6=data0=(00 10 20 30)
409    paddw       mm7, mm0                ; mm7=data1=(01 11 21 31)
410    psraw       mm6, (PASS1_BITS+3)     ; descale
411    psraw       mm7, (PASS1_BITS+3)     ; descale
412    psubw       mm1, mm3                ; mm1=data7=(07 17 27 37)
413    psubw       mm5, mm0                ; mm5=data6=(06 16 26 36)
414    psraw       mm1, (PASS1_BITS+3)     ; descale
415    psraw       mm5, (PASS1_BITS+3)     ; descale
416    psubw       mm4, mm0                ; mm4=tmp5
417
418    packsswb    mm6, mm5                ; mm6=(00 10 20 30 06 16 26 36)
419    packsswb    mm7, mm1                ; mm7=(01 11 21 31 07 17 27 37)
420
421    movq        mm3, MMWORD [wk(0)]     ; mm3=tmp2
422    movq        mm0, MMWORD [wk(1)]     ; mm0=tmp3
423
424    paddw       mm2, mm4                ; mm2=tmp4
425    movq        mm5, mm3
426    movq        mm1, mm0
427    paddw       mm3, mm4                ; mm3=data2=(02 12 22 32)
428    paddw       mm0, mm2                ; mm0=data4=(04 14 24 34)
429    psraw       mm3, (PASS1_BITS+3)     ; descale
430    psraw       mm0, (PASS1_BITS+3)     ; descale
431    psubw       mm5, mm4                ; mm5=data5=(05 15 25 35)
432    psubw       mm1, mm2                ; mm1=data3=(03 13 23 33)
433    psraw       mm5, (PASS1_BITS+3)     ; descale
434    psraw       mm1, (PASS1_BITS+3)     ; descale
435
436    movq        mm4, [GOTOFF(ebx,PB_CENTERJSAMP)]  ; mm4=[PB_CENTERJSAMP]
437
438    packsswb    mm3, mm0                ; mm3=(02 12 22 32 04 14 24 34)
439    packsswb    mm1, mm5                ; mm1=(03 13 23 33 05 15 25 35)
440
441    paddb       mm6, mm4
442    paddb       mm7, mm4
443    paddb       mm3, mm4
444    paddb       mm1, mm4
445
446    movq        mm2, mm6                ; transpose coefficients(phase 1)
447    punpcklbw   mm6, mm7                ; mm6=(00 01 10 11 20 21 30 31)
448    punpckhbw   mm2, mm7                ; mm2=(06 07 16 17 26 27 36 37)
449    movq        mm0, mm3                ; transpose coefficients(phase 1)
450    punpcklbw   mm3, mm1                ; mm3=(02 03 12 13 22 23 32 33)
451    punpckhbw   mm0, mm1                ; mm0=(04 05 14 15 24 25 34 35)
452
453    movq        mm5, mm6                ; transpose coefficients(phase 2)
454    punpcklwd   mm6, mm3                ; mm6=(00 01 02 03 10 11 12 13)
455    punpckhwd   mm5, mm3                ; mm5=(20 21 22 23 30 31 32 33)
456    movq        mm4, mm0                ; transpose coefficients(phase 2)
457    punpcklwd   mm0, mm2                ; mm0=(04 05 06 07 14 15 16 17)
458    punpckhwd   mm4, mm2                ; mm4=(24 25 26 27 34 35 36 37)
459
460    movq        mm7, mm6                ; transpose coefficients(phase 3)
461    punpckldq   mm6, mm0                ; mm6=(00 01 02 03 04 05 06 07)
462    punpckhdq   mm7, mm0                ; mm7=(10 11 12 13 14 15 16 17)
463    movq        mm1, mm5                ; transpose coefficients(phase 3)
464    punpckldq   mm5, mm4                ; mm5=(20 21 22 23 24 25 26 27)
465    punpckhdq   mm1, mm4                ; mm1=(30 31 32 33 34 35 36 37)
466
467    pushpic     ebx                     ; save GOT address
468
469    mov         edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
470    mov         ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
471    movq        MMWORD [edx+eax*SIZEOF_JSAMPLE], mm6
472    movq        MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm7
473    mov         edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
474    mov         ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
475    movq        MMWORD [edx+eax*SIZEOF_JSAMPLE], mm5
476    movq        MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm1
477
478    poppic      ebx                     ; restore GOT address
479
480    add         esi, byte 4*SIZEOF_JCOEF     ; wsptr
481    add         edi, byte 4*SIZEOF_JSAMPROW
482    dec         ecx                          ; ctr
483    jnz         near .rowloop
484
485    emms                                ; empty MMX state
486
487    pop         edi
488    pop         esi
489;   pop         edx                     ; need not be preserved
490;   pop         ecx                     ; need not be preserved
491    pop         ebx
492    mov         esp, ebp                ; esp <- aligned ebp
493    pop         esp                     ; esp <- original ebp
494    pop         ebp
495    ret
496
497; For some reason, the OS X linker does not honor the request to align the
498; segment unless we do this.
499    align       32
500