• 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;
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_sse2)
61
62EXTN(jconst_idct_ifast_sse2):
63
64PW_F1414       times 8  dw  F_1_414 << CONST_SHIFT
65PW_F1847       times 8  dw  F_1_847 << CONST_SHIFT
66PW_MF1613      times 8  dw -F_1_613 << CONST_SHIFT
67PW_F1082       times 8  dw  F_1_082 << CONST_SHIFT
68PB_CENTERJSAMP times 16 db  CENTERJSAMPLE
69
70    alignz      32
71
72; --------------------------------------------------------------------------
73    SECTION     SEG_TEXT
74    BITS        64
75;
76; Perform dequantization and inverse DCT on one block of coefficients.
77;
78; GLOBAL(void)
79; jsimd_idct_ifast_sse2(void *dct_table, JCOEFPTR coef_block,
80;                      JSAMPARRAY output_buf, JDIMENSION output_col)
81;
82
83; r10 = jpeg_component_info *compptr
84; r11 = JCOEFPTR coef_block
85; r12 = JSAMPARRAY output_buf
86; r13d = JDIMENSION output_col
87
88%define original_rbp  rbp + 0
89%define wk(i)         rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD
90                                        ; xmmword wk[WK_NUM]
91%define WK_NUM        2
92
93    align       32
94    GLOBAL_FUNCTION(jsimd_idct_ifast_sse2)
95
96EXTN(jsimd_idct_ifast_sse2):
97    push        rbp
98    mov         rax, rsp                     ; rax = original rbp
99    sub         rsp, byte 4
100    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
101    mov         [rsp], rax
102    mov         rbp, rsp                     ; rbp = aligned rbp
103    lea         rsp, [wk(0)]
104    collect_args 4
105
106    ; ---- Pass 1: process columns from input.
107
108    mov         rdx, r10                ; quantptr
109    mov         rsi, r11                ; inptr
110
111%ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2
112    mov         eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
113    or          eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
114    jnz         near .columnDCT
115
116    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
117    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
118    por         xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
119    por         xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
120    por         xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
121    por         xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
122    por         xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
123    por         xmm1, xmm0
124    packsswb    xmm1, xmm1
125    packsswb    xmm1, xmm1
126    movd        eax, xmm1
127    test        rax, rax
128    jnz         short .columnDCT
129
130    ; -- AC terms all zero
131
132    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
133    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
134
135    movdqa      xmm7, xmm0              ; xmm0=in0=(00 01 02 03 04 05 06 07)
136    punpcklwd   xmm0, xmm0              ; xmm0=(00 00 01 01 02 02 03 03)
137    punpckhwd   xmm7, xmm7              ; xmm7=(04 04 05 05 06 06 07 07)
138
139    pshufd      xmm6, xmm0, 0x00        ; xmm6=col0=(00 00 00 00 00 00 00 00)
140    pshufd      xmm2, xmm0, 0x55        ; xmm2=col1=(01 01 01 01 01 01 01 01)
141    pshufd      xmm5, xmm0, 0xAA        ; xmm5=col2=(02 02 02 02 02 02 02 02)
142    pshufd      xmm0, xmm0, 0xFF        ; xmm0=col3=(03 03 03 03 03 03 03 03)
143    pshufd      xmm1, xmm7, 0x00        ; xmm1=col4=(04 04 04 04 04 04 04 04)
144    pshufd      xmm4, xmm7, 0x55        ; xmm4=col5=(05 05 05 05 05 05 05 05)
145    pshufd      xmm3, xmm7, 0xAA        ; xmm3=col6=(06 06 06 06 06 06 06 06)
146    pshufd      xmm7, xmm7, 0xFF        ; xmm7=col7=(07 07 07 07 07 07 07 07)
147
148    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=col1
149    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=col3
150    jmp         near .column_end
151%endif
152.columnDCT:
153
154    ; -- Even part
155
156    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
157    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
158    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
159    pmullw      xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
160    movdqa      xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
161    movdqa      xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
162    pmullw      xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
163    pmullw      xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
164
165    movdqa      xmm4, xmm0
166    movdqa      xmm5, xmm1
167    psubw       xmm0, xmm2              ; xmm0=tmp11
168    psubw       xmm1, xmm3
169    paddw       xmm4, xmm2              ; xmm4=tmp10
170    paddw       xmm5, xmm3              ; xmm5=tmp13
171
172    psllw       xmm1, PRE_MULTIPLY_SCALE_BITS
173    pmulhw      xmm1, [rel PW_F1414]
174    psubw       xmm1, xmm5              ; xmm1=tmp12
175
176    movdqa      xmm6, xmm4
177    movdqa      xmm7, xmm0
178    psubw       xmm4, xmm5              ; xmm4=tmp3
179    psubw       xmm0, xmm1              ; xmm0=tmp2
180    paddw       xmm6, xmm5              ; xmm6=tmp0
181    paddw       xmm7, xmm1              ; xmm7=tmp1
182
183    movdqa      XMMWORD [wk(1)], xmm4   ; wk(1)=tmp3
184    movdqa      XMMWORD [wk(0)], xmm0   ; wk(0)=tmp2
185
186    ; -- Odd part
187
188    movdqa      xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
189    movdqa      xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
190    pmullw      xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
191    pmullw      xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
192    movdqa      xmm5, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
193    movdqa      xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
194    pmullw      xmm5, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
195    pmullw      xmm1, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
196
197    movdqa      xmm4, xmm2
198    movdqa      xmm0, xmm5
199    psubw       xmm2, xmm1              ; xmm2=z12
200    psubw       xmm5, xmm3              ; xmm5=z10
201    paddw       xmm4, xmm1              ; xmm4=z11
202    paddw       xmm0, xmm3              ; xmm0=z13
203
204    movdqa      xmm1, xmm5              ; xmm1=z10(unscaled)
205    psllw       xmm2, PRE_MULTIPLY_SCALE_BITS
206    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
207
208    movdqa      xmm3, xmm4
209    psubw       xmm4, xmm0
210    paddw       xmm3, xmm0              ; xmm3=tmp7
211
212    psllw       xmm4, PRE_MULTIPLY_SCALE_BITS
213    pmulhw      xmm4, [rel PW_F1414]    ; xmm4=tmp11
214
215    ; To avoid overflow...
216    ;
217    ; (Original)
218    ; tmp12 = -2.613125930 * z10 + z5;
219    ;
220    ; (This implementation)
221    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
222    ;       = -1.613125930 * z10 - z10 + z5;
223
224    movdqa      xmm0, xmm5
225    paddw       xmm5, xmm2
226    pmulhw      xmm5, [rel PW_F1847]    ; xmm5=z5
227    pmulhw      xmm0, [rel PW_MF1613]
228    pmulhw      xmm2, [rel PW_F1082]
229    psubw       xmm0, xmm1
230    psubw       xmm2, xmm5              ; xmm2=tmp10
231    paddw       xmm0, xmm5              ; xmm0=tmp12
232
233    ; -- Final output stage
234
235    psubw       xmm0, xmm3              ; xmm0=tmp6
236    movdqa      xmm1, xmm6
237    movdqa      xmm5, xmm7
238    paddw       xmm6, xmm3              ; xmm6=data0=(00 01 02 03 04 05 06 07)
239    paddw       xmm7, xmm0              ; xmm7=data1=(10 11 12 13 14 15 16 17)
240    psubw       xmm1, xmm3              ; xmm1=data7=(70 71 72 73 74 75 76 77)
241    psubw       xmm5, xmm0              ; xmm5=data6=(60 61 62 63 64 65 66 67)
242    psubw       xmm4, xmm0              ; xmm4=tmp5
243
244    movdqa      xmm3, xmm6              ; transpose coefficients(phase 1)
245    punpcklwd   xmm6, xmm7              ; xmm6=(00 10 01 11 02 12 03 13)
246    punpckhwd   xmm3, xmm7              ; xmm3=(04 14 05 15 06 16 07 17)
247    movdqa      xmm0, xmm5              ; transpose coefficients(phase 1)
248    punpcklwd   xmm5, xmm1              ; xmm5=(60 70 61 71 62 72 63 73)
249    punpckhwd   xmm0, xmm1              ; xmm0=(64 74 65 75 66 76 67 77)
250
251    movdqa      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
252    movdqa      xmm1, XMMWORD [wk(1)]   ; xmm1=tmp3
253
254    movdqa      XMMWORD [wk(0)], xmm5   ; wk(0)=(60 70 61 71 62 72 63 73)
255    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=(64 74 65 75 66 76 67 77)
256
257    paddw       xmm2, xmm4              ; xmm2=tmp4
258    movdqa      xmm5, xmm7
259    movdqa      xmm0, xmm1
260    paddw       xmm7, xmm4              ; xmm7=data2=(20 21 22 23 24 25 26 27)
261    paddw       xmm1, xmm2              ; xmm1=data4=(40 41 42 43 44 45 46 47)
262    psubw       xmm5, xmm4              ; xmm5=data5=(50 51 52 53 54 55 56 57)
263    psubw       xmm0, xmm2              ; xmm0=data3=(30 31 32 33 34 35 36 37)
264
265    movdqa      xmm4, xmm7              ; transpose coefficients(phase 1)
266    punpcklwd   xmm7, xmm0              ; xmm7=(20 30 21 31 22 32 23 33)
267    punpckhwd   xmm4, xmm0              ; xmm4=(24 34 25 35 26 36 27 37)
268    movdqa      xmm2, xmm1              ; transpose coefficients(phase 1)
269    punpcklwd   xmm1, xmm5              ; xmm1=(40 50 41 51 42 52 43 53)
270    punpckhwd   xmm2, xmm5              ; xmm2=(44 54 45 55 46 56 47 57)
271
272    movdqa      xmm0, xmm3              ; transpose coefficients(phase 2)
273    punpckldq   xmm3, xmm4              ; xmm3=(04 14 24 34 05 15 25 35)
274    punpckhdq   xmm0, xmm4              ; xmm0=(06 16 26 36 07 17 27 37)
275    movdqa      xmm5, xmm6              ; transpose coefficients(phase 2)
276    punpckldq   xmm6, xmm7              ; xmm6=(00 10 20 30 01 11 21 31)
277    punpckhdq   xmm5, xmm7              ; xmm5=(02 12 22 32 03 13 23 33)
278
279    movdqa      xmm4, XMMWORD [wk(0)]   ; xmm4=(60 70 61 71 62 72 63 73)
280    movdqa      xmm7, XMMWORD [wk(1)]   ; xmm7=(64 74 65 75 66 76 67 77)
281
282    movdqa      XMMWORD [wk(0)], xmm3   ; wk(0)=(04 14 24 34 05 15 25 35)
283    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=(06 16 26 36 07 17 27 37)
284
285    movdqa      xmm3, xmm1              ; transpose coefficients(phase 2)
286    punpckldq   xmm1, xmm4              ; xmm1=(40 50 60 70 41 51 61 71)
287    punpckhdq   xmm3, xmm4              ; xmm3=(42 52 62 72 43 53 63 73)
288    movdqa      xmm0, xmm2              ; transpose coefficients(phase 2)
289    punpckldq   xmm2, xmm7              ; xmm2=(44 54 64 74 45 55 65 75)
290    punpckhdq   xmm0, xmm7              ; xmm0=(46 56 66 76 47 57 67 77)
291
292    movdqa      xmm4, xmm6              ; transpose coefficients(phase 3)
293    punpcklqdq  xmm6, xmm1              ; xmm6=col0=(00 10 20 30 40 50 60 70)
294    punpckhqdq  xmm4, xmm1              ; xmm4=col1=(01 11 21 31 41 51 61 71)
295    movdqa      xmm7, xmm5              ; transpose coefficients(phase 3)
296    punpcklqdq  xmm5, xmm3              ; xmm5=col2=(02 12 22 32 42 52 62 72)
297    punpckhqdq  xmm7, xmm3              ; xmm7=col3=(03 13 23 33 43 53 63 73)
298
299    movdqa      xmm1, XMMWORD [wk(0)]   ; xmm1=(04 14 24 34 05 15 25 35)
300    movdqa      xmm3, XMMWORD [wk(1)]   ; xmm3=(06 16 26 36 07 17 27 37)
301
302    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=col1
303    movdqa      XMMWORD [wk(1)], xmm7   ; wk(1)=col3
304
305    movdqa      xmm4, xmm1              ; transpose coefficients(phase 3)
306    punpcklqdq  xmm1, xmm2              ; xmm1=col4=(04 14 24 34 44 54 64 74)
307    punpckhqdq  xmm4, xmm2              ; xmm4=col5=(05 15 25 35 45 55 65 75)
308    movdqa      xmm7, xmm3              ; transpose coefficients(phase 3)
309    punpcklqdq  xmm3, xmm0              ; xmm3=col6=(06 16 26 36 46 56 66 76)
310    punpckhqdq  xmm7, xmm0              ; xmm7=col7=(07 17 27 37 47 57 67 77)
311.column_end:
312
313    ; -- Prefetch the next coefficient block
314
315    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
316    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
317    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
318    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
319
320    ; ---- Pass 2: process rows from work array, store into output array.
321
322    mov         rax, [original_rbp]
323    mov         rdi, r12                ; (JSAMPROW *)
324    mov         eax, r13d
325
326    ; -- Even part
327
328    ; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6
329
330    movdqa      xmm2, xmm6
331    movdqa      xmm0, xmm5
332    psubw       xmm6, xmm1              ; xmm6=tmp11
333    psubw       xmm5, xmm3
334    paddw       xmm2, xmm1              ; xmm2=tmp10
335    paddw       xmm0, xmm3              ; xmm0=tmp13
336
337    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
338    pmulhw      xmm5, [rel PW_F1414]
339    psubw       xmm5, xmm0              ; xmm5=tmp12
340
341    movdqa      xmm1, xmm2
342    movdqa      xmm3, xmm6
343    psubw       xmm2, xmm0              ; xmm2=tmp3
344    psubw       xmm6, xmm5              ; xmm6=tmp2
345    paddw       xmm1, xmm0              ; xmm1=tmp0
346    paddw       xmm3, xmm5              ; xmm3=tmp1
347
348    movdqa      xmm0, XMMWORD [wk(0)]   ; xmm0=col1
349    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=col3
350
351    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=tmp3
352    movdqa      XMMWORD [wk(1)], xmm6   ; wk(1)=tmp2
353
354    ; -- Odd part
355
356    ; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7
357
358    movdqa      xmm2, xmm0
359    movdqa      xmm6, xmm4
360    psubw       xmm0, xmm7              ; xmm0=z12
361    psubw       xmm4, xmm5              ; xmm4=z10
362    paddw       xmm2, xmm7              ; xmm2=z11
363    paddw       xmm6, xmm5              ; xmm6=z13
364
365    movdqa      xmm7, xmm4              ; xmm7=z10(unscaled)
366    psllw       xmm0, PRE_MULTIPLY_SCALE_BITS
367    psllw       xmm4, PRE_MULTIPLY_SCALE_BITS
368
369    movdqa      xmm5, xmm2
370    psubw       xmm2, xmm6
371    paddw       xmm5, xmm6              ; xmm5=tmp7
372
373    psllw       xmm2, PRE_MULTIPLY_SCALE_BITS
374    pmulhw      xmm2, [rel PW_F1414]    ; xmm2=tmp11
375
376    ; To avoid overflow...
377    ;
378    ; (Original)
379    ; tmp12 = -2.613125930 * z10 + z5;
380    ;
381    ; (This implementation)
382    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
383    ;       = -1.613125930 * z10 - z10 + z5;
384
385    movdqa      xmm6, xmm4
386    paddw       xmm4, xmm0
387    pmulhw      xmm4, [rel PW_F1847]    ; xmm4=z5
388    pmulhw      xmm6, [rel PW_MF1613]
389    pmulhw      xmm0, [rel PW_F1082]
390    psubw       xmm6, xmm7
391    psubw       xmm0, xmm4              ; xmm0=tmp10
392    paddw       xmm6, xmm4              ; xmm6=tmp12
393
394    ; -- Final output stage
395
396    psubw       xmm6, xmm5              ; xmm6=tmp6
397    movdqa      xmm7, xmm1
398    movdqa      xmm4, xmm3
399    paddw       xmm1, xmm5              ; xmm1=data0=(00 10 20 30 40 50 60 70)
400    paddw       xmm3, xmm6              ; xmm3=data1=(01 11 21 31 41 51 61 71)
401    psraw       xmm1, (PASS1_BITS+3)    ; descale
402    psraw       xmm3, (PASS1_BITS+3)    ; descale
403    psubw       xmm7, xmm5              ; xmm7=data7=(07 17 27 37 47 57 67 77)
404    psubw       xmm4, xmm6              ; xmm4=data6=(06 16 26 36 46 56 66 76)
405    psraw       xmm7, (PASS1_BITS+3)    ; descale
406    psraw       xmm4, (PASS1_BITS+3)    ; descale
407    psubw       xmm2, xmm6              ; xmm2=tmp5
408
409    packsswb    xmm1, xmm4        ; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
410    packsswb    xmm3, xmm7        ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
411
412    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp2
413    movdqa      xmm6, XMMWORD [wk(0)]   ; xmm6=tmp3
414
415    paddw       xmm0, xmm2              ; xmm0=tmp4
416    movdqa      xmm4, xmm5
417    movdqa      xmm7, xmm6
418    paddw       xmm5, xmm2              ; xmm5=data2=(02 12 22 32 42 52 62 72)
419    paddw       xmm6, xmm0              ; xmm6=data4=(04 14 24 34 44 54 64 74)
420    psraw       xmm5, (PASS1_BITS+3)    ; descale
421    psraw       xmm6, (PASS1_BITS+3)    ; descale
422    psubw       xmm4, xmm2              ; xmm4=data5=(05 15 25 35 45 55 65 75)
423    psubw       xmm7, xmm0              ; xmm7=data3=(03 13 23 33 43 53 63 73)
424    psraw       xmm4, (PASS1_BITS+3)    ; descale
425    psraw       xmm7, (PASS1_BITS+3)    ; descale
426
427    movdqa      xmm2, [rel PB_CENTERJSAMP]  ; xmm2=[rel PB_CENTERJSAMP]
428
429    packsswb    xmm5, xmm6        ; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
430    packsswb    xmm7, xmm4        ; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
431
432    paddb       xmm1, xmm2
433    paddb       xmm3, xmm2
434    paddb       xmm5, xmm2
435    paddb       xmm7, xmm2
436
437    movdqa      xmm0, xmm1        ; transpose coefficients(phase 1)
438    punpcklbw   xmm1, xmm3        ; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
439    punpckhbw   xmm0, xmm3        ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
440    movdqa      xmm6, xmm5        ; transpose coefficients(phase 1)
441    punpcklbw   xmm5, xmm7        ; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
442    punpckhbw   xmm6, xmm7        ; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
443
444    movdqa      xmm4, xmm1        ; transpose coefficients(phase 2)
445    punpcklwd   xmm1, xmm5        ; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
446    punpckhwd   xmm4, xmm5        ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
447    movdqa      xmm2, xmm6        ; transpose coefficients(phase 2)
448    punpcklwd   xmm6, xmm0        ; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
449    punpckhwd   xmm2, xmm0        ; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
450
451    movdqa      xmm3, xmm1        ; transpose coefficients(phase 3)
452    punpckldq   xmm1, xmm6        ; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
453    punpckhdq   xmm3, xmm6        ; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
454    movdqa      xmm7, xmm4        ; transpose coefficients(phase 3)
455    punpckldq   xmm4, xmm2        ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
456    punpckhdq   xmm7, xmm2        ; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
457
458    pshufd      xmm5, xmm1, 0x4E  ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
459    pshufd      xmm0, xmm3, 0x4E  ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
460    pshufd      xmm6, xmm4, 0x4E  ; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
461    pshufd      xmm2, xmm7, 0x4E  ; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
462
463    mov         rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
464    mov         rsi, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
465    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
466    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
467    mov         rdx, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
468    mov         rsi, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
469    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
470    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
471
472    mov         rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
473    mov         rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
474    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
475    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
476    mov         rdx, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
477    mov         rsi, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
478    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
479    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
480
481    uncollect_args 4
482    mov         rsp, rbp                ; rsp <- aligned rbp
483    pop         rsp                     ; rsp <- original rbp
484    pop         rbp
485    ret
486    ret
487
488; For some reason, the OS X linker does not honor the request to align the
489; segment unless we do this.
490    align       32
491