1;
2; jidctfst.asm - fast integer IDCT (64-bit SSE2)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright 2009 D. R. Commander
6;
7; Based on
8; x86 SIMD extension for IJG JPEG library
9; Copyright (C) 1999-2006, MIYASAKA Masaru.
10; For conditions of distribution and use, see copyright notice in jsimdext.inc
11;
12; This file should be assembled with NASM (Netwide Assembler),
13; can *not* be assembled with Microsoft's MASM or any compatible
14; assembler (including Borland's Turbo Assembler).
15; NASM is available from http://nasm.sourceforge.net/ or
16; http://sourceforge.net/project/showfiles.php?group_id=6208
17;
18; This file contains a fast, not so accurate integer implementation of
19; the inverse DCT (Discrete Cosine Transform). The following code is
20; based directly on the IJG's original jidctfst.c; see the jidctfst.c
21; for more details.
22;
23; [TAB8]
24
25%include "jsimdext.inc"
26%include "jdct.inc"
27
28; --------------------------------------------------------------------------
29
30%define CONST_BITS      8       ; 14 is also OK.
31%define PASS1_BITS      2
32
33%if IFAST_SCALE_BITS != PASS1_BITS
34%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
35%endif
36
37%if CONST_BITS == 8
38F_1_082 equ     277             ; FIX(1.082392200)
39F_1_414 equ     362             ; FIX(1.414213562)
40F_1_847 equ     473             ; FIX(1.847759065)
41F_2_613 equ     669             ; FIX(2.613125930)
42F_1_613 equ     (F_2_613 - 256) ; FIX(2.613125930) - FIX(1)
43%else
44; NASM cannot do compile-time arithmetic on floating-point constants.
45%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
46F_1_082 equ     DESCALE(1162209775,30-CONST_BITS)       ; FIX(1.082392200)
47F_1_414 equ     DESCALE(1518500249,30-CONST_BITS)       ; FIX(1.414213562)
48F_1_847 equ     DESCALE(1984016188,30-CONST_BITS)       ; FIX(1.847759065)
49F_2_613 equ     DESCALE(2805822602,30-CONST_BITS)       ; FIX(2.613125930)
50F_1_613 equ     (F_2_613 - (1 << CONST_BITS))   ; FIX(2.613125930) - FIX(1)
51%endif
52
53; --------------------------------------------------------------------------
54        SECTION SEG_CONST
55
56; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
57; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
58
59%define PRE_MULTIPLY_SCALE_BITS   2
60%define CONST_SHIFT     (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
61
62        alignz  16
63        global  EXTN(jconst_idct_ifast_sse2)
64
65EXTN(jconst_idct_ifast_sse2):
66
67PW_F1414        times 8 dw  F_1_414 << CONST_SHIFT
68PW_F1847        times 8 dw  F_1_847 << CONST_SHIFT
69PW_MF1613       times 8 dw -F_1_613 << CONST_SHIFT
70PW_F1082        times 8 dw  F_1_082 << CONST_SHIFT
71PB_CENTERJSAMP  times 16 db CENTERJSAMPLE
72
73        alignz  16
74
75; --------------------------------------------------------------------------
76        SECTION SEG_TEXT
77        BITS    64
78;
79; Perform dequantization and inverse DCT on one block of coefficients.
80;
81; GLOBAL(void)
82; jsimd_idct_ifast_sse2 (void * dct_table, JCOEFPTR coef_block,
83;                       JSAMPARRAY output_buf, JDIMENSION output_col)
84;
85
86; r10 = jpeg_component_info * compptr
87; r11 = JCOEFPTR coef_block
88; r12 = JSAMPARRAY output_buf
89; r13 = JDIMENSION output_col
90
91%define original_rbp    rbp+0
92%define wk(i)           rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
93%define WK_NUM          2
94
95        align   16
96        global  EXTN(jsimd_idct_ifast_sse2)
97
98EXTN(jsimd_idct_ifast_sse2):
99        push    rbp
100        mov     rax,rsp                         ; rax = original rbp
101        sub     rsp, byte 4
102        and     rsp, byte (-SIZEOF_XMMWORD)     ; align to 128 bits
103        mov     [rsp],rax
104        mov     rbp,rsp                         ; rbp = aligned rbp
105        lea     rsp, [wk(0)]
106        collect_args
107
108        ; ---- Pass 1: process columns from input.
109
110        mov     rdx, r10                ; quantptr
111        mov     rsi, r11                ; inptr
112
113%ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2
114        mov     eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
115        or      eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
116        jnz     near .columnDCT
117
118        movdqa  xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
119        movdqa  xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
120        por     xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
121        por     xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
122        por     xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
123        por     xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
124        por     xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
125        por     xmm1,xmm0
126        packsswb xmm1,xmm1
127        packsswb xmm1,xmm1
128        movd    eax,xmm1
129        test    rax,rax
130        jnz     short .columnDCT
131
132        ; -- AC terms all zero
133
134        movdqa  xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
135        pmullw  xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
136
137        movdqa    xmm7,xmm0             ; xmm0=in0=(00 01 02 03 04 05 06 07)
138        punpcklwd xmm0,xmm0             ; xmm0=(00 00 01 01 02 02 03 03)
139        punpckhwd xmm7,xmm7             ; xmm7=(04 04 05 05 06 06 07 07)
140
141        pshufd  xmm6,xmm0,0x00          ; xmm6=col0=(00 00 00 00 00 00 00 00)
142        pshufd  xmm2,xmm0,0x55          ; xmm2=col1=(01 01 01 01 01 01 01 01)
143        pshufd  xmm5,xmm0,0xAA          ; xmm5=col2=(02 02 02 02 02 02 02 02)
144        pshufd  xmm0,xmm0,0xFF          ; xmm0=col3=(03 03 03 03 03 03 03 03)
145        pshufd  xmm1,xmm7,0x00          ; xmm1=col4=(04 04 04 04 04 04 04 04)
146        pshufd  xmm4,xmm7,0x55          ; xmm4=col5=(05 05 05 05 05 05 05 05)
147        pshufd  xmm3,xmm7,0xAA          ; xmm3=col6=(06 06 06 06 06 06 06 06)
148        pshufd  xmm7,xmm7,0xFF          ; xmm7=col7=(07 07 07 07 07 07 07 07)
149
150        movdqa  XMMWORD [wk(0)], xmm2   ; wk(0)=col1
151        movdqa  XMMWORD [wk(1)], xmm0   ; wk(1)=col3
152        jmp     near .column_end
153%endif
154.columnDCT:
155
156        ; -- Even part
157
158        movdqa  xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
159        movdqa  xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
160        pmullw  xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
161        pmullw  xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
162        movdqa  xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
163        movdqa  xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
164        pmullw  xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
165        pmullw  xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
166
167        movdqa  xmm4,xmm0
168        movdqa  xmm5,xmm1
169        psubw   xmm0,xmm2               ; xmm0=tmp11
170        psubw   xmm1,xmm3
171        paddw   xmm4,xmm2               ; xmm4=tmp10
172        paddw   xmm5,xmm3               ; xmm5=tmp13
173
174        psllw   xmm1,PRE_MULTIPLY_SCALE_BITS
175        pmulhw  xmm1,[rel PW_F1414]
176        psubw   xmm1,xmm5               ; xmm1=tmp12
177
178        movdqa  xmm6,xmm4
179        movdqa  xmm7,xmm0
180        psubw   xmm4,xmm5               ; xmm4=tmp3
181        psubw   xmm0,xmm1               ; xmm0=tmp2
182        paddw   xmm6,xmm5               ; xmm6=tmp0
183        paddw   xmm7,xmm1               ; xmm7=tmp1
184
185        movdqa  XMMWORD [wk(1)], xmm4   ; wk(1)=tmp3
186        movdqa  XMMWORD [wk(0)], xmm0   ; wk(0)=tmp2
187
188        ; -- Odd part
189
190        movdqa  xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
191        movdqa  xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
192        pmullw  xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
193        pmullw  xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
194        movdqa  xmm5, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
195        movdqa  xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
196        pmullw  xmm5, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
197        pmullw  xmm1, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
198
199        movdqa  xmm4,xmm2
200        movdqa  xmm0,xmm5
201        psubw   xmm2,xmm1               ; xmm2=z12
202        psubw   xmm5,xmm3               ; xmm5=z10
203        paddw   xmm4,xmm1               ; xmm4=z11
204        paddw   xmm0,xmm3               ; xmm0=z13
205
206        movdqa  xmm1,xmm5               ; xmm1=z10(unscaled)
207        psllw   xmm2,PRE_MULTIPLY_SCALE_BITS
208        psllw   xmm5,PRE_MULTIPLY_SCALE_BITS
209
210        movdqa  xmm3,xmm4
211        psubw   xmm4,xmm0
212        paddw   xmm3,xmm0               ; xmm3=tmp7
213
214        psllw   xmm4,PRE_MULTIPLY_SCALE_BITS
215        pmulhw  xmm4,[rel PW_F1414]     ; xmm4=tmp11
216
217        ; To avoid overflow...
218        ;
219        ; (Original)
220        ; tmp12 = -2.613125930 * z10 + z5;
221        ;
222        ; (This implementation)
223        ; tmp12 = (-1.613125930 - 1) * z10 + z5;
224        ;       = -1.613125930 * z10 - z10 + z5;
225
226        movdqa  xmm0,xmm5
227        paddw   xmm5,xmm2
228        pmulhw  xmm5,[rel PW_F1847]     ; xmm5=z5
229        pmulhw  xmm0,[rel PW_MF1613]
230        pmulhw  xmm2,[rel PW_F1082]
231        psubw   xmm0,xmm1
232        psubw   xmm2,xmm5               ; xmm2=tmp10
233        paddw   xmm0,xmm5               ; xmm0=tmp12
234
235        ; -- Final output stage
236
237        psubw   xmm0,xmm3               ; xmm0=tmp6
238        movdqa  xmm1,xmm6
239        movdqa  xmm5,xmm7
240        paddw   xmm6,xmm3               ; xmm6=data0=(00 01 02 03 04 05 06 07)
241        paddw   xmm7,xmm0               ; xmm7=data1=(10 11 12 13 14 15 16 17)
242        psubw   xmm1,xmm3               ; xmm1=data7=(70 71 72 73 74 75 76 77)
243        psubw   xmm5,xmm0               ; xmm5=data6=(60 61 62 63 64 65 66 67)
244        psubw   xmm4,xmm0               ; xmm4=tmp5
245
246        movdqa    xmm3,xmm6             ; transpose coefficients(phase 1)
247        punpcklwd xmm6,xmm7             ; xmm6=(00 10 01 11 02 12 03 13)
248        punpckhwd xmm3,xmm7             ; xmm3=(04 14 05 15 06 16 07 17)
249        movdqa    xmm0,xmm5             ; transpose coefficients(phase 1)
250        punpcklwd xmm5,xmm1             ; xmm5=(60 70 61 71 62 72 63 73)
251        punpckhwd xmm0,xmm1             ; xmm0=(64 74 65 75 66 76 67 77)
252
253        movdqa  xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
254        movdqa  xmm1, XMMWORD [wk(1)]   ; xmm1=tmp3
255
256        movdqa  XMMWORD [wk(0)], xmm5   ; wk(0)=(60 70 61 71 62 72 63 73)
257        movdqa  XMMWORD [wk(1)], xmm0   ; wk(1)=(64 74 65 75 66 76 67 77)
258
259        paddw   xmm2,xmm4               ; xmm2=tmp4
260        movdqa  xmm5,xmm7
261        movdqa  xmm0,xmm1
262        paddw   xmm7,xmm4               ; xmm7=data2=(20 21 22 23 24 25 26 27)
263        paddw   xmm1,xmm2               ; xmm1=data4=(40 41 42 43 44 45 46 47)
264        psubw   xmm5,xmm4               ; xmm5=data5=(50 51 52 53 54 55 56 57)
265        psubw   xmm0,xmm2               ; xmm0=data3=(30 31 32 33 34 35 36 37)
266
267        movdqa    xmm4,xmm7             ; transpose coefficients(phase 1)
268        punpcklwd xmm7,xmm0             ; xmm7=(20 30 21 31 22 32 23 33)
269        punpckhwd xmm4,xmm0             ; xmm4=(24 34 25 35 26 36 27 37)
270        movdqa    xmm2,xmm1             ; transpose coefficients(phase 1)
271        punpcklwd xmm1,xmm5             ; xmm1=(40 50 41 51 42 52 43 53)
272        punpckhwd xmm2,xmm5             ; xmm2=(44 54 45 55 46 56 47 57)
273
274        movdqa    xmm0,xmm3             ; transpose coefficients(phase 2)
275        punpckldq xmm3,xmm4             ; xmm3=(04 14 24 34 05 15 25 35)
276        punpckhdq xmm0,xmm4             ; xmm0=(06 16 26 36 07 17 27 37)
277        movdqa    xmm5,xmm6             ; transpose coefficients(phase 2)
278        punpckldq xmm6,xmm7             ; xmm6=(00 10 20 30 01 11 21 31)
279        punpckhdq xmm5,xmm7             ; xmm5=(02 12 22 32 03 13 23 33)
280
281        movdqa  xmm4, XMMWORD [wk(0)]   ; xmm4=(60 70 61 71 62 72 63 73)
282        movdqa  xmm7, XMMWORD [wk(1)]   ; xmm7=(64 74 65 75 66 76 67 77)
283
284        movdqa  XMMWORD [wk(0)], xmm3   ; wk(0)=(04 14 24 34 05 15 25 35)
285        movdqa  XMMWORD [wk(1)], xmm0   ; wk(1)=(06 16 26 36 07 17 27 37)
286
287        movdqa    xmm3,xmm1             ; transpose coefficients(phase 2)
288        punpckldq xmm1,xmm4             ; xmm1=(40 50 60 70 41 51 61 71)
289        punpckhdq xmm3,xmm4             ; xmm3=(42 52 62 72 43 53 63 73)
290        movdqa    xmm0,xmm2             ; transpose coefficients(phase 2)
291        punpckldq xmm2,xmm7             ; xmm2=(44 54 64 74 45 55 65 75)
292        punpckhdq xmm0,xmm7             ; xmm0=(46 56 66 76 47 57 67 77)
293
294        movdqa     xmm4,xmm6            ; transpose coefficients(phase 3)
295        punpcklqdq xmm6,xmm1            ; xmm6=col0=(00 10 20 30 40 50 60 70)
296        punpckhqdq xmm4,xmm1            ; xmm4=col1=(01 11 21 31 41 51 61 71)
297        movdqa     xmm7,xmm5            ; transpose coefficients(phase 3)
298        punpcklqdq xmm5,xmm3            ; xmm5=col2=(02 12 22 32 42 52 62 72)
299        punpckhqdq xmm7,xmm3            ; xmm7=col3=(03 13 23 33 43 53 63 73)
300
301        movdqa  xmm1, XMMWORD [wk(0)]   ; xmm1=(04 14 24 34 05 15 25 35)
302        movdqa  xmm3, XMMWORD [wk(1)]   ; xmm3=(06 16 26 36 07 17 27 37)
303
304        movdqa  XMMWORD [wk(0)], xmm4   ; wk(0)=col1
305        movdqa  XMMWORD [wk(1)], xmm7   ; wk(1)=col3
306
307        movdqa     xmm4,xmm1            ; transpose coefficients(phase 3)
308        punpcklqdq xmm1,xmm2            ; xmm1=col4=(04 14 24 34 44 54 64 74)
309        punpckhqdq xmm4,xmm2            ; xmm4=col5=(05 15 25 35 45 55 65 75)
310        movdqa     xmm7,xmm3            ; transpose coefficients(phase 3)
311        punpcklqdq xmm3,xmm0            ; xmm3=col6=(06 16 26 36 46 56 66 76)
312        punpckhqdq xmm7,xmm0            ; xmm7=col7=(07 17 27 37 47 57 67 77)
313.column_end:
314
315        ; -- Prefetch the next coefficient block
316
317        prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
318        prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
319        prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
320        prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
321
322        ; ---- Pass 2: process rows from work array, store into output array.
323
324        mov     rax, [original_rbp]
325        mov     rdi, r12        ; (JSAMPROW *)
326        mov     eax, r13d
327
328        ; -- Even part
329
330        ; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6
331
332        movdqa  xmm2,xmm6
333        movdqa  xmm0,xmm5
334        psubw   xmm6,xmm1               ; xmm6=tmp11
335        psubw   xmm5,xmm3
336        paddw   xmm2,xmm1               ; xmm2=tmp10
337        paddw   xmm0,xmm3               ; xmm0=tmp13
338
339        psllw   xmm5,PRE_MULTIPLY_SCALE_BITS
340        pmulhw  xmm5,[rel PW_F1414]
341        psubw   xmm5,xmm0               ; xmm5=tmp12
342
343        movdqa  xmm1,xmm2
344        movdqa  xmm3,xmm6
345        psubw   xmm2,xmm0               ; xmm2=tmp3
346        psubw   xmm6,xmm5               ; xmm6=tmp2
347        paddw   xmm1,xmm0               ; xmm1=tmp0
348        paddw   xmm3,xmm5               ; xmm3=tmp1
349
350        movdqa  xmm0, XMMWORD [wk(0)]   ; xmm0=col1
351        movdqa  xmm5, XMMWORD [wk(1)]   ; xmm5=col3
352
353        movdqa  XMMWORD [wk(0)], xmm2   ; wk(0)=tmp3
354        movdqa  XMMWORD [wk(1)], xmm6   ; wk(1)=tmp2
355
356        ; -- Odd part
357
358        ; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7
359
360        movdqa  xmm2,xmm0
361        movdqa  xmm6,xmm4
362        psubw   xmm0,xmm7               ; xmm0=z12
363        psubw   xmm4,xmm5               ; xmm4=z10
364        paddw   xmm2,xmm7               ; xmm2=z11
365        paddw   xmm6,xmm5               ; xmm6=z13
366
367        movdqa  xmm7,xmm4               ; xmm7=z10(unscaled)
368        psllw   xmm0,PRE_MULTIPLY_SCALE_BITS
369        psllw   xmm4,PRE_MULTIPLY_SCALE_BITS
370
371        movdqa  xmm5,xmm2
372        psubw   xmm2,xmm6
373        paddw   xmm5,xmm6               ; xmm5=tmp7
374
375        psllw   xmm2,PRE_MULTIPLY_SCALE_BITS
376        pmulhw  xmm2,[rel PW_F1414]     ; xmm2=tmp11
377
378        ; To avoid overflow...
379        ;
380        ; (Original)
381        ; tmp12 = -2.613125930 * z10 + z5;
382        ;
383        ; (This implementation)
384        ; tmp12 = (-1.613125930 - 1) * z10 + z5;
385        ;       = -1.613125930 * z10 - z10 + z5;
386
387        movdqa  xmm6,xmm4
388        paddw   xmm4,xmm0
389        pmulhw  xmm4,[rel PW_F1847]     ; xmm4=z5
390        pmulhw  xmm6,[rel PW_MF1613]
391        pmulhw  xmm0,[rel PW_F1082]
392        psubw   xmm6,xmm7
393        psubw   xmm0,xmm4               ; xmm0=tmp10
394        paddw   xmm6,xmm4               ; xmm6=tmp12
395
396        ; -- Final output stage
397
398        psubw   xmm6,xmm5               ; xmm6=tmp6
399        movdqa  xmm7,xmm1
400        movdqa  xmm4,xmm3
401        paddw   xmm1,xmm5               ; xmm1=data0=(00 10 20 30 40 50 60 70)
402        paddw   xmm3,xmm6               ; xmm3=data1=(01 11 21 31 41 51 61 71)
403        psraw   xmm1,(PASS1_BITS+3)     ; descale
404        psraw   xmm3,(PASS1_BITS+3)     ; descale
405        psubw   xmm7,xmm5               ; xmm7=data7=(07 17 27 37 47 57 67 77)
406        psubw   xmm4,xmm6               ; xmm4=data6=(06 16 26 36 46 56 66 76)
407        psraw   xmm7,(PASS1_BITS+3)     ; descale
408        psraw   xmm4,(PASS1_BITS+3)     ; descale
409        psubw   xmm2,xmm6               ; xmm2=tmp5
410
411        packsswb  xmm1,xmm4     ; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
412        packsswb  xmm3,xmm7     ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
413
414        movdqa  xmm5, XMMWORD [wk(1)]   ; xmm5=tmp2
415        movdqa  xmm6, XMMWORD [wk(0)]   ; xmm6=tmp3
416
417        paddw   xmm0,xmm2               ; xmm0=tmp4
418        movdqa  xmm4,xmm5
419        movdqa  xmm7,xmm6
420        paddw   xmm5,xmm2               ; xmm5=data2=(02 12 22 32 42 52 62 72)
421        paddw   xmm6,xmm0               ; xmm6=data4=(04 14 24 34 44 54 64 74)
422        psraw   xmm5,(PASS1_BITS+3)     ; descale
423        psraw   xmm6,(PASS1_BITS+3)     ; descale
424        psubw   xmm4,xmm2               ; xmm4=data5=(05 15 25 35 45 55 65 75)
425        psubw   xmm7,xmm0               ; xmm7=data3=(03 13 23 33 43 53 63 73)
426        psraw   xmm4,(PASS1_BITS+3)     ; descale
427        psraw   xmm7,(PASS1_BITS+3)     ; descale
428
429        movdqa    xmm2,[rel PB_CENTERJSAMP]     ; xmm2=[rel PB_CENTERJSAMP]
430
431        packsswb  xmm5,xmm6     ; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
432        packsswb  xmm7,xmm4     ; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
433
434        paddb     xmm1,xmm2
435        paddb     xmm3,xmm2
436        paddb     xmm5,xmm2
437        paddb     xmm7,xmm2
438
439        movdqa    xmm0,xmm1     ; transpose coefficients(phase 1)
440        punpcklbw xmm1,xmm3     ; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
441        punpckhbw xmm0,xmm3     ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
442        movdqa    xmm6,xmm5     ; transpose coefficients(phase 1)
443        punpcklbw xmm5,xmm7     ; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
444        punpckhbw xmm6,xmm7     ; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
445
446        movdqa    xmm4,xmm1     ; transpose coefficients(phase 2)
447        punpcklwd xmm1,xmm5     ; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
448        punpckhwd xmm4,xmm5     ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
449        movdqa    xmm2,xmm6     ; transpose coefficients(phase 2)
450        punpcklwd xmm6,xmm0     ; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
451        punpckhwd xmm2,xmm0     ; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
452
453        movdqa    xmm3,xmm1     ; transpose coefficients(phase 3)
454        punpckldq xmm1,xmm6     ; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
455        punpckhdq xmm3,xmm6     ; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
456        movdqa    xmm7,xmm4     ; transpose coefficients(phase 3)
457        punpckldq xmm4,xmm2     ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
458        punpckhdq xmm7,xmm2     ; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
459
460        pshufd  xmm5,xmm1,0x4E  ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
461        pshufd  xmm0,xmm3,0x4E  ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
462        pshufd  xmm6,xmm4,0x4E  ; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
463        pshufd  xmm2,xmm7,0x4E  ; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
464
465        mov     rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
466        mov     rsi, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
467        movq    XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
468        movq    XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
469        mov     rdx, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
470        mov     rsi, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
471        movq    XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
472        movq    XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
473
474        mov     rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
475        mov     rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
476        movq    XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
477        movq    XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
478        mov     rdx, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
479        mov     rsi, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
480        movq    XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
481        movq    XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
482
483        uncollect_args
484        mov     rsp,rbp         ; rsp <- aligned rbp
485        pop     rsp             ; rsp <- original rbp
486        pop     rbp
487        ret
488        ret
489
490; For some reason, the OS X linker does not honor the request to align the
491; segment unless we do this.
492        align   16
493