• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jidctfst.asm - fast integer IDCT (64-bit SSE2)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright (C) 2009, 2016, D. R. Commander.
6; Copyright (C) 2018, Matthias Räncker.
7;
8; Based on the 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%include "jsimdext.inc"
24%include "jdct.inc"
25
26; --------------------------------------------------------------------------
27
28%define CONST_BITS  8  ; 14 is also OK.
29%define PASS1_BITS  2
30
31%if IFAST_SCALE_BITS != PASS1_BITS
32%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
33%endif
34
35%if CONST_BITS == 8
36F_1_082 equ 277              ; FIX(1.082392200)
37F_1_414 equ 362              ; FIX(1.414213562)
38F_1_847 equ 473              ; FIX(1.847759065)
39F_2_613 equ 669              ; FIX(2.613125930)
40F_1_613 equ (F_2_613 - 256)  ; FIX(2.613125930) - FIX(1)
41%else
42; NASM cannot do compile-time arithmetic on floating-point constants.
43%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
44F_1_082 equ DESCALE(1162209775, 30 - CONST_BITS)  ; FIX(1.082392200)
45F_1_414 equ DESCALE(1518500249, 30 - CONST_BITS)  ; FIX(1.414213562)
46F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
47F_2_613 equ DESCALE(2805822602, 30 - CONST_BITS)  ; FIX(2.613125930)
48F_1_613 equ (F_2_613 - (1 << CONST_BITS))         ; FIX(2.613125930) - FIX(1)
49%endif
50
51; --------------------------------------------------------------------------
52    SECTION     SEG_CONST
53
54; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
55; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
56
57%define PRE_MULTIPLY_SCALE_BITS  2
58%define CONST_SHIFT              (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
59
60    alignz      32
61    GLOBAL_DATA(jconst_idct_ifast_sse2)
62
63EXTN(jconst_idct_ifast_sse2):
64
65PW_F1414       times 8  dw  F_1_414 << CONST_SHIFT
66PW_F1847       times 8  dw  F_1_847 << CONST_SHIFT
67PW_MF1613      times 8  dw -F_1_613 << CONST_SHIFT
68PW_F1082       times 8  dw  F_1_082 << CONST_SHIFT
69PB_CENTERJSAMP times 16 db  CENTERJSAMPLE
70
71    alignz      32
72
73; --------------------------------------------------------------------------
74    SECTION     SEG_TEXT
75    BITS        64
76;
77; Perform dequantization and inverse DCT on one block of coefficients.
78;
79; GLOBAL(void)
80; jsimd_idct_ifast_sse2(void *dct_table, JCOEFPTR coef_block,
81;                      JSAMPARRAY output_buf, JDIMENSION output_col)
82;
83
84; r10 = jpeg_component_info *compptr
85; r11 = JCOEFPTR coef_block
86; r12 = JSAMPARRAY output_buf
87; r13d = JDIMENSION output_col
88
89%define original_rbp  rbp + 0
90%define wk(i)         rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD
91                                        ; xmmword wk[WK_NUM]
92%define WK_NUM        2
93
94    align       32
95    GLOBAL_FUNCTION(jsimd_idct_ifast_sse2)
96
97EXTN(jsimd_idct_ifast_sse2):
98    push        rbp
99    mov         rax, rsp                     ; rax = original rbp
100    sub         rsp, byte 4
101    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
102    mov         [rsp], rax
103    mov         rbp, rsp                     ; rbp = aligned rbp
104    lea         rsp, [wk(0)]
105    collect_args 4
106
107    ; ---- Pass 1: process columns from input.
108
109    mov         rdx, r10                ; quantptr
110    mov         rsi, r11                ; inptr
111
112%ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2
113    mov         eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
114    or          eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
115    jnz         near .columnDCT
116
117    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
118    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
119    por         xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
120    por         xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
121    por         xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
122    por         xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
123    por         xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
124    por         xmm1, xmm0
125    packsswb    xmm1, xmm1
126    packsswb    xmm1, xmm1
127    movd        eax, xmm1
128    test        rax, rax
129    jnz         short .columnDCT
130
131    ; -- AC terms all zero
132
133    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
134    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
135
136    movdqa      xmm7, xmm0              ; xmm0=in0=(00 01 02 03 04 05 06 07)
137    punpcklwd   xmm0, xmm0              ; xmm0=(00 00 01 01 02 02 03 03)
138    punpckhwd   xmm7, xmm7              ; xmm7=(04 04 05 05 06 06 07 07)
139
140    pshufd      xmm6, xmm0, 0x00        ; xmm6=col0=(00 00 00 00 00 00 00 00)
141    pshufd      xmm2, xmm0, 0x55        ; xmm2=col1=(01 01 01 01 01 01 01 01)
142    pshufd      xmm5, xmm0, 0xAA        ; xmm5=col2=(02 02 02 02 02 02 02 02)
143    pshufd      xmm0, xmm0, 0xFF        ; xmm0=col3=(03 03 03 03 03 03 03 03)
144    pshufd      xmm1, xmm7, 0x00        ; xmm1=col4=(04 04 04 04 04 04 04 04)
145    pshufd      xmm4, xmm7, 0x55        ; xmm4=col5=(05 05 05 05 05 05 05 05)
146    pshufd      xmm3, xmm7, 0xAA        ; xmm3=col6=(06 06 06 06 06 06 06 06)
147    pshufd      xmm7, xmm7, 0xFF        ; xmm7=col7=(07 07 07 07 07 07 07 07)
148
149    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=col1
150    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=col3
151    jmp         near .column_end
152%endif
153.columnDCT:
154
155    ; -- Even part
156
157    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
158    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
159    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
160    pmullw      xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
161    movdqa      xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
162    movdqa      xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
163    pmullw      xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
164    pmullw      xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
165
166    movdqa      xmm4, xmm0
167    movdqa      xmm5, xmm1
168    psubw       xmm0, xmm2              ; xmm0=tmp11
169    psubw       xmm1, xmm3
170    paddw       xmm4, xmm2              ; xmm4=tmp10
171    paddw       xmm5, xmm3              ; xmm5=tmp13
172
173    psllw       xmm1, PRE_MULTIPLY_SCALE_BITS
174    pmulhw      xmm1, [rel PW_F1414]
175    psubw       xmm1, xmm5              ; xmm1=tmp12
176
177    movdqa      xmm6, xmm4
178    movdqa      xmm7, xmm0
179    psubw       xmm4, xmm5              ; xmm4=tmp3
180    psubw       xmm0, xmm1              ; xmm0=tmp2
181    paddw       xmm6, xmm5              ; xmm6=tmp0
182    paddw       xmm7, xmm1              ; xmm7=tmp1
183
184    movdqa      XMMWORD [wk(1)], xmm4   ; wk(1)=tmp3
185    movdqa      XMMWORD [wk(0)], xmm0   ; wk(0)=tmp2
186
187    ; -- Odd part
188
189    movdqa      xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
190    movdqa      xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
191    pmullw      xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
192    pmullw      xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
193    movdqa      xmm5, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
194    movdqa      xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
195    pmullw      xmm5, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
196    pmullw      xmm1, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
197
198    movdqa      xmm4, xmm2
199    movdqa      xmm0, xmm5
200    psubw       xmm2, xmm1              ; xmm2=z12
201    psubw       xmm5, xmm3              ; xmm5=z10
202    paddw       xmm4, xmm1              ; xmm4=z11
203    paddw       xmm0, xmm3              ; xmm0=z13
204
205    movdqa      xmm1, xmm5              ; xmm1=z10(unscaled)
206    psllw       xmm2, PRE_MULTIPLY_SCALE_BITS
207    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
208
209    movdqa      xmm3, xmm4
210    psubw       xmm4, xmm0
211    paddw       xmm3, xmm0              ; xmm3=tmp7
212
213    psllw       xmm4, PRE_MULTIPLY_SCALE_BITS
214    pmulhw      xmm4, [rel PW_F1414]    ; xmm4=tmp11
215
216    ; To avoid overflow...
217    ;
218    ; (Original)
219    ; tmp12 = -2.613125930 * z10 + z5;
220    ;
221    ; (This implementation)
222    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
223    ;       = -1.613125930 * z10 - z10 + z5;
224
225    movdqa      xmm0, xmm5
226    paddw       xmm5, xmm2
227    pmulhw      xmm5, [rel PW_F1847]    ; xmm5=z5
228    pmulhw      xmm0, [rel PW_MF1613]
229    pmulhw      xmm2, [rel PW_F1082]
230    psubw       xmm0, xmm1
231    psubw       xmm2, xmm5              ; xmm2=tmp10
232    paddw       xmm0, xmm5              ; xmm0=tmp12
233
234    ; -- Final output stage
235
236    psubw       xmm0, xmm3              ; xmm0=tmp6
237    movdqa      xmm1, xmm6
238    movdqa      xmm5, xmm7
239    paddw       xmm6, xmm3              ; xmm6=data0=(00 01 02 03 04 05 06 07)
240    paddw       xmm7, xmm0              ; xmm7=data1=(10 11 12 13 14 15 16 17)
241    psubw       xmm1, xmm3              ; xmm1=data7=(70 71 72 73 74 75 76 77)
242    psubw       xmm5, xmm0              ; xmm5=data6=(60 61 62 63 64 65 66 67)
243    psubw       xmm4, xmm0              ; xmm4=tmp5
244
245    movdqa      xmm3, xmm6              ; transpose coefficients(phase 1)
246    punpcklwd   xmm6, xmm7              ; xmm6=(00 10 01 11 02 12 03 13)
247    punpckhwd   xmm3, xmm7              ; xmm3=(04 14 05 15 06 16 07 17)
248    movdqa      xmm0, xmm5              ; transpose coefficients(phase 1)
249    punpcklwd   xmm5, xmm1              ; xmm5=(60 70 61 71 62 72 63 73)
250    punpckhwd   xmm0, xmm1              ; xmm0=(64 74 65 75 66 76 67 77)
251
252    movdqa      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
253    movdqa      xmm1, XMMWORD [wk(1)]   ; xmm1=tmp3
254
255    movdqa      XMMWORD [wk(0)], xmm5   ; wk(0)=(60 70 61 71 62 72 63 73)
256    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=(64 74 65 75 66 76 67 77)
257
258    paddw       xmm2, xmm4              ; xmm2=tmp4
259    movdqa      xmm5, xmm7
260    movdqa      xmm0, xmm1
261    paddw       xmm7, xmm4              ; xmm7=data2=(20 21 22 23 24 25 26 27)
262    paddw       xmm1, xmm2              ; xmm1=data4=(40 41 42 43 44 45 46 47)
263    psubw       xmm5, xmm4              ; xmm5=data5=(50 51 52 53 54 55 56 57)
264    psubw       xmm0, xmm2              ; xmm0=data3=(30 31 32 33 34 35 36 37)
265
266    movdqa      xmm4, xmm7              ; transpose coefficients(phase 1)
267    punpcklwd   xmm7, xmm0              ; xmm7=(20 30 21 31 22 32 23 33)
268    punpckhwd   xmm4, xmm0              ; xmm4=(24 34 25 35 26 36 27 37)
269    movdqa      xmm2, xmm1              ; transpose coefficients(phase 1)
270    punpcklwd   xmm1, xmm5              ; xmm1=(40 50 41 51 42 52 43 53)
271    punpckhwd   xmm2, xmm5              ; xmm2=(44 54 45 55 46 56 47 57)
272
273    movdqa      xmm0, xmm3              ; transpose coefficients(phase 2)
274    punpckldq   xmm3, xmm4              ; xmm3=(04 14 24 34 05 15 25 35)
275    punpckhdq   xmm0, xmm4              ; xmm0=(06 16 26 36 07 17 27 37)
276    movdqa      xmm5, xmm6              ; transpose coefficients(phase 2)
277    punpckldq   xmm6, xmm7              ; xmm6=(00 10 20 30 01 11 21 31)
278    punpckhdq   xmm5, xmm7              ; xmm5=(02 12 22 32 03 13 23 33)
279
280    movdqa      xmm4, XMMWORD [wk(0)]   ; xmm4=(60 70 61 71 62 72 63 73)
281    movdqa      xmm7, XMMWORD [wk(1)]   ; xmm7=(64 74 65 75 66 76 67 77)
282
283    movdqa      XMMWORD [wk(0)], xmm3   ; wk(0)=(04 14 24 34 05 15 25 35)
284    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=(06 16 26 36 07 17 27 37)
285
286    movdqa      xmm3, xmm1              ; transpose coefficients(phase 2)
287    punpckldq   xmm1, xmm4              ; xmm1=(40 50 60 70 41 51 61 71)
288    punpckhdq   xmm3, xmm4              ; xmm3=(42 52 62 72 43 53 63 73)
289    movdqa      xmm0, xmm2              ; transpose coefficients(phase 2)
290    punpckldq   xmm2, xmm7              ; xmm2=(44 54 64 74 45 55 65 75)
291    punpckhdq   xmm0, xmm7              ; xmm0=(46 56 66 76 47 57 67 77)
292
293    movdqa      xmm4, xmm6              ; transpose coefficients(phase 3)
294    punpcklqdq  xmm6, xmm1              ; xmm6=col0=(00 10 20 30 40 50 60 70)
295    punpckhqdq  xmm4, xmm1              ; xmm4=col1=(01 11 21 31 41 51 61 71)
296    movdqa      xmm7, xmm5              ; transpose coefficients(phase 3)
297    punpcklqdq  xmm5, xmm3              ; xmm5=col2=(02 12 22 32 42 52 62 72)
298    punpckhqdq  xmm7, xmm3              ; xmm7=col3=(03 13 23 33 43 53 63 73)
299
300    movdqa      xmm1, XMMWORD [wk(0)]   ; xmm1=(04 14 24 34 05 15 25 35)
301    movdqa      xmm3, XMMWORD [wk(1)]   ; xmm3=(06 16 26 36 07 17 27 37)
302
303    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=col1
304    movdqa      XMMWORD [wk(1)], xmm7   ; wk(1)=col3
305
306    movdqa      xmm4, xmm1              ; transpose coefficients(phase 3)
307    punpcklqdq  xmm1, xmm2              ; xmm1=col4=(04 14 24 34 44 54 64 74)
308    punpckhqdq  xmm4, xmm2              ; xmm4=col5=(05 15 25 35 45 55 65 75)
309    movdqa      xmm7, xmm3              ; transpose coefficients(phase 3)
310    punpcklqdq  xmm3, xmm0              ; xmm3=col6=(06 16 26 36 46 56 66 76)
311    punpckhqdq  xmm7, xmm0              ; xmm7=col7=(07 17 27 37 47 57 67 77)
312.column_end:
313
314    ; -- Prefetch the next coefficient block
315
316    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
317    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
318    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
319    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
320
321    ; ---- Pass 2: process rows from work array, store into output array.
322
323    mov         rax, [original_rbp]
324    mov         rdi, r12                ; (JSAMPROW *)
325    mov         eax, r13d
326
327    ; -- Even part
328
329    ; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6
330
331    movdqa      xmm2, xmm6
332    movdqa      xmm0, xmm5
333    psubw       xmm6, xmm1              ; xmm6=tmp11
334    psubw       xmm5, xmm3
335    paddw       xmm2, xmm1              ; xmm2=tmp10
336    paddw       xmm0, xmm3              ; xmm0=tmp13
337
338    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
339    pmulhw      xmm5, [rel PW_F1414]
340    psubw       xmm5, xmm0              ; xmm5=tmp12
341
342    movdqa      xmm1, xmm2
343    movdqa      xmm3, xmm6
344    psubw       xmm2, xmm0              ; xmm2=tmp3
345    psubw       xmm6, xmm5              ; xmm6=tmp2
346    paddw       xmm1, xmm0              ; xmm1=tmp0
347    paddw       xmm3, xmm5              ; xmm3=tmp1
348
349    movdqa      xmm0, XMMWORD [wk(0)]   ; xmm0=col1
350    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=col3
351
352    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=tmp3
353    movdqa      XMMWORD [wk(1)], xmm6   ; wk(1)=tmp2
354
355    ; -- Odd part
356
357    ; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7
358
359    movdqa      xmm2, xmm0
360    movdqa      xmm6, xmm4
361    psubw       xmm0, xmm7              ; xmm0=z12
362    psubw       xmm4, xmm5              ; xmm4=z10
363    paddw       xmm2, xmm7              ; xmm2=z11
364    paddw       xmm6, xmm5              ; xmm6=z13
365
366    movdqa      xmm7, xmm4              ; xmm7=z10(unscaled)
367    psllw       xmm0, PRE_MULTIPLY_SCALE_BITS
368    psllw       xmm4, PRE_MULTIPLY_SCALE_BITS
369
370    movdqa      xmm5, xmm2
371    psubw       xmm2, xmm6
372    paddw       xmm5, xmm6              ; xmm5=tmp7
373
374    psllw       xmm2, PRE_MULTIPLY_SCALE_BITS
375    pmulhw      xmm2, [rel PW_F1414]    ; xmm2=tmp11
376
377    ; To avoid overflow...
378    ;
379    ; (Original)
380    ; tmp12 = -2.613125930 * z10 + z5;
381    ;
382    ; (This implementation)
383    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
384    ;       = -1.613125930 * z10 - z10 + z5;
385
386    movdqa      xmm6, xmm4
387    paddw       xmm4, xmm0
388    pmulhw      xmm4, [rel PW_F1847]    ; xmm4=z5
389    pmulhw      xmm6, [rel PW_MF1613]
390    pmulhw      xmm0, [rel PW_F1082]
391    psubw       xmm6, xmm7
392    psubw       xmm0, xmm4              ; xmm0=tmp10
393    paddw       xmm6, xmm4              ; xmm6=tmp12
394
395    ; -- Final output stage
396
397    psubw       xmm6, xmm5              ; xmm6=tmp6
398    movdqa      xmm7, xmm1
399    movdqa      xmm4, xmm3
400    paddw       xmm1, xmm5              ; xmm1=data0=(00 10 20 30 40 50 60 70)
401    paddw       xmm3, xmm6              ; xmm3=data1=(01 11 21 31 41 51 61 71)
402    psraw       xmm1, (PASS1_BITS+3)    ; descale
403    psraw       xmm3, (PASS1_BITS+3)    ; descale
404    psubw       xmm7, xmm5              ; xmm7=data7=(07 17 27 37 47 57 67 77)
405    psubw       xmm4, xmm6              ; xmm4=data6=(06 16 26 36 46 56 66 76)
406    psraw       xmm7, (PASS1_BITS+3)    ; descale
407    psraw       xmm4, (PASS1_BITS+3)    ; descale
408    psubw       xmm2, xmm6              ; xmm2=tmp5
409
410    packsswb    xmm1, xmm4        ; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
411    packsswb    xmm3, xmm7        ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
412
413    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp2
414    movdqa      xmm6, XMMWORD [wk(0)]   ; xmm6=tmp3
415
416    paddw       xmm0, xmm2              ; xmm0=tmp4
417    movdqa      xmm4, xmm5
418    movdqa      xmm7, xmm6
419    paddw       xmm5, xmm2              ; xmm5=data2=(02 12 22 32 42 52 62 72)
420    paddw       xmm6, xmm0              ; xmm6=data4=(04 14 24 34 44 54 64 74)
421    psraw       xmm5, (PASS1_BITS+3)    ; descale
422    psraw       xmm6, (PASS1_BITS+3)    ; descale
423    psubw       xmm4, xmm2              ; xmm4=data5=(05 15 25 35 45 55 65 75)
424    psubw       xmm7, xmm0              ; xmm7=data3=(03 13 23 33 43 53 63 73)
425    psraw       xmm4, (PASS1_BITS+3)    ; descale
426    psraw       xmm7, (PASS1_BITS+3)    ; descale
427
428    movdqa      xmm2, [rel PB_CENTERJSAMP]  ; xmm2=[rel PB_CENTERJSAMP]
429
430    packsswb    xmm5, xmm6        ; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
431    packsswb    xmm7, xmm4        ; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
432
433    paddb       xmm1, xmm2
434    paddb       xmm3, xmm2
435    paddb       xmm5, xmm2
436    paddb       xmm7, xmm2
437
438    movdqa      xmm0, xmm1        ; transpose coefficients(phase 1)
439    punpcklbw   xmm1, xmm3        ; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
440    punpckhbw   xmm0, xmm3        ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
441    movdqa      xmm6, xmm5        ; transpose coefficients(phase 1)
442    punpcklbw   xmm5, xmm7        ; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
443    punpckhbw   xmm6, xmm7        ; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
444
445    movdqa      xmm4, xmm1        ; transpose coefficients(phase 2)
446    punpcklwd   xmm1, xmm5        ; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
447    punpckhwd   xmm4, xmm5        ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
448    movdqa      xmm2, xmm6        ; transpose coefficients(phase 2)
449    punpcklwd   xmm6, xmm0        ; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
450    punpckhwd   xmm2, xmm0        ; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
451
452    movdqa      xmm3, xmm1        ; transpose coefficients(phase 3)
453    punpckldq   xmm1, xmm6        ; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
454    punpckhdq   xmm3, xmm6        ; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
455    movdqa      xmm7, xmm4        ; transpose coefficients(phase 3)
456    punpckldq   xmm4, xmm2        ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
457    punpckhdq   xmm7, xmm2        ; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
458
459    pshufd      xmm5, xmm1, 0x4E  ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
460    pshufd      xmm0, xmm3, 0x4E  ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
461    pshufd      xmm6, xmm4, 0x4E  ; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
462    pshufd      xmm2, xmm7, 0x4E  ; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
463
464    mov         rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
465    mov         rsip, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
466    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
467    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
468    mov         rdxp, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
469    mov         rsip, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
470    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
471    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
472
473    mov         rdxp, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
474    mov         rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
475    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
476    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
477    mov         rdxp, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
478    mov         rsip, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
479    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
480    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
481
482    uncollect_args 4
483    mov         rsp, rbp                ; rsp <- aligned rbp
484    pop         rsp                     ; rsp <- original rbp
485    pop         rbp
486    ret
487    ret
488
489; For some reason, the OS X linker does not honor the request to align the
490; segment unless we do this.
491    align       32
492