• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jidctfst.asm - fast integer IDCT (SSE2)
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; [TAB8]
23
24%include "jsimdext.inc"
25%include "jdct.inc"
26
27; --------------------------------------------------------------------------
28
29%define CONST_BITS  8  ; 14 is also OK.
30%define PASS1_BITS  2
31
32%if IFAST_SCALE_BITS != PASS1_BITS
33%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
34%endif
35
36%if CONST_BITS == 8
37F_1_082 equ 277              ; FIX(1.082392200)
38F_1_414 equ 362              ; FIX(1.414213562)
39F_1_847 equ 473              ; FIX(1.847759065)
40F_2_613 equ 669              ; FIX(2.613125930)
41F_1_613 equ (F_2_613 - 256)  ; FIX(2.613125930) - FIX(1)
42%else
43; NASM cannot do compile-time arithmetic on floating-point constants.
44%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
45F_1_082 equ DESCALE(1162209775, 30 - CONST_BITS)  ; FIX(1.082392200)
46F_1_414 equ DESCALE(1518500249, 30 - CONST_BITS)  ; FIX(1.414213562)
47F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
48F_2_613 equ DESCALE(2805822602, 30 - CONST_BITS)  ; FIX(2.613125930)
49F_1_613 equ (F_2_613 - (1 << CONST_BITS))       ; FIX(2.613125930) - FIX(1)
50%endif
51
52; --------------------------------------------------------------------------
53    SECTION     SEG_CONST
54
55; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
56; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
57
58%define PRE_MULTIPLY_SCALE_BITS  2
59%define CONST_SHIFT              (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
60
61    alignz      32
62    GLOBAL_DATA(jconst_idct_ifast_sse2)
63
64EXTN(jconst_idct_ifast_sse2):
65
66PW_F1414       times 8  dw  F_1_414 << CONST_SHIFT
67PW_F1847       times 8  dw  F_1_847 << CONST_SHIFT
68PW_MF1613      times 8  dw -F_1_613 << CONST_SHIFT
69PW_F1082       times 8  dw  F_1_082 << CONST_SHIFT
70PB_CENTERJSAMP times 16 db  CENTERJSAMPLE
71
72    alignz      32
73
74; --------------------------------------------------------------------------
75    SECTION     SEG_TEXT
76    BITS        32
77;
78; Perform dequantization and inverse DCT on one block of coefficients.
79;
80; GLOBAL(void)
81; jsimd_idct_ifast_sse2(void *dct_table, JCOEFPTR coef_block,
82;                       JSAMPARRAY output_buf, JDIMENSION output_col)
83;
84
85%define dct_table(b)   (b) + 8          ; jpeg_component_info *compptr
86%define coef_block(b)  (b) + 12         ; JCOEFPTR coef_block
87%define output_buf(b)  (b) + 16         ; JSAMPARRAY output_buf
88%define output_col(b)  (b) + 20         ; JDIMENSION output_col
89
90%define original_ebp   ebp + 0
91%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD
92                                        ; xmmword wk[WK_NUM]
93%define WK_NUM         2
94
95    align       32
96    GLOBAL_FUNCTION(jsimd_idct_ifast_sse2)
97
98EXTN(jsimd_idct_ifast_sse2):
99    push        ebp
100    mov         eax, esp                     ; eax = original ebp
101    sub         esp, byte 4
102    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
103    mov         [esp], eax
104    mov         ebp, esp                     ; ebp = aligned ebp
105    lea         esp, [wk(0)]
106    pushpic     ebx
107;   push        ecx                     ; unused
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.
115
116;   mov         eax, [original_ebp]
117    mov         edx, POINTER [dct_table(eax)]    ; quantptr
118    mov         esi, JCOEFPTR [coef_block(eax)]  ; inptr
119
120%ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2
121    mov         eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
122    or          eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
123    jnz         near .columnDCT
124
125    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
126    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
127    por         xmm0, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
128    por         xmm1, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)]
129    por         xmm0, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
130    por         xmm1, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
131    por         xmm0, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
132    por         xmm1, xmm0
133    packsswb    xmm1, xmm1
134    packsswb    xmm1, xmm1
135    movd        eax, xmm1
136    test        eax, eax
137    jnz         short .columnDCT
138
139    ; -- AC terms all zero
140
141    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
142    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
143
144    movdqa      xmm7, xmm0              ; xmm0=in0=(00 01 02 03 04 05 06 07)
145    punpcklwd   xmm0, xmm0              ; xmm0=(00 00 01 01 02 02 03 03)
146    punpckhwd   xmm7, xmm7              ; xmm7=(04 04 05 05 06 06 07 07)
147
148    pshufd      xmm6, xmm0, 0x00        ; xmm6=col0=(00 00 00 00 00 00 00 00)
149    pshufd      xmm2, xmm0, 0x55        ; xmm2=col1=(01 01 01 01 01 01 01 01)
150    pshufd      xmm5, xmm0, 0xAA        ; xmm5=col2=(02 02 02 02 02 02 02 02)
151    pshufd      xmm0, xmm0, 0xFF        ; xmm0=col3=(03 03 03 03 03 03 03 03)
152    pshufd      xmm1, xmm7, 0x00        ; xmm1=col4=(04 04 04 04 04 04 04 04)
153    pshufd      xmm4, xmm7, 0x55        ; xmm4=col5=(05 05 05 05 05 05 05 05)
154    pshufd      xmm3, xmm7, 0xAA        ; xmm3=col6=(06 06 06 06 06 06 06 06)
155    pshufd      xmm7, xmm7, 0xFF        ; xmm7=col7=(07 07 07 07 07 07 07 07)
156
157    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=col1
158    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=col3
159    jmp         near .column_end
160    alignx      16, 7
161%endif
162.columnDCT:
163
164    ; -- Even part
165
166    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
167    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
168    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)]
169    pmullw      xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_IFAST_MULT_TYPE)]
170    movdqa      xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)]
171    movdqa      xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
172    pmullw      xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_IFAST_MULT_TYPE)]
173    pmullw      xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_IFAST_MULT_TYPE)]
174
175    movdqa      xmm4, xmm0
176    movdqa      xmm5, xmm1
177    psubw       xmm0, xmm2              ; xmm0=tmp11
178    psubw       xmm1, xmm3
179    paddw       xmm4, xmm2              ; xmm4=tmp10
180    paddw       xmm5, xmm3              ; xmm5=tmp13
181
182    psllw       xmm1, PRE_MULTIPLY_SCALE_BITS
183    pmulhw      xmm1, [GOTOFF(ebx,PW_F1414)]
184    psubw       xmm1, xmm5              ; xmm1=tmp12
185
186    movdqa      xmm6, xmm4
187    movdqa      xmm7, xmm0
188    psubw       xmm4, xmm5              ; xmm4=tmp3
189    psubw       xmm0, xmm1              ; xmm0=tmp2
190    paddw       xmm6, xmm5              ; xmm6=tmp0
191    paddw       xmm7, xmm1              ; xmm7=tmp1
192
193    movdqa      XMMWORD [wk(1)], xmm4   ; wk(1)=tmp3
194    movdqa      XMMWORD [wk(0)], xmm0   ; wk(0)=tmp2
195
196    ; -- Odd part
197
198    movdqa      xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
199    movdqa      xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
200    pmullw      xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_IFAST_MULT_TYPE)]
201    pmullw      xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_IFAST_MULT_TYPE)]
202    movdqa      xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
203    movdqa      xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
204    pmullw      xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_IFAST_MULT_TYPE)]
205    pmullw      xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_IFAST_MULT_TYPE)]
206
207    movdqa      xmm4, xmm2
208    movdqa      xmm0, xmm5
209    psubw       xmm2, xmm1              ; xmm2=z12
210    psubw       xmm5, xmm3              ; xmm5=z10
211    paddw       xmm4, xmm1              ; xmm4=z11
212    paddw       xmm0, xmm3              ; xmm0=z13
213
214    movdqa      xmm1, xmm5              ; xmm1=z10(unscaled)
215    psllw       xmm2, PRE_MULTIPLY_SCALE_BITS
216    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
217
218    movdqa      xmm3, xmm4
219    psubw       xmm4, xmm0
220    paddw       xmm3, xmm0              ; xmm3=tmp7
221
222    psllw       xmm4, PRE_MULTIPLY_SCALE_BITS
223    pmulhw      xmm4, [GOTOFF(ebx,PW_F1414)]  ; xmm4=tmp11
224
225    ; To avoid overflow...
226    ;
227    ; (Original)
228    ; tmp12 = -2.613125930 * z10 + z5;
229    ;
230    ; (This implementation)
231    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
232    ;       = -1.613125930 * z10 - z10 + z5;
233
234    movdqa      xmm0, xmm5
235    paddw       xmm5, xmm2
236    pmulhw      xmm5, [GOTOFF(ebx,PW_F1847)]   ; xmm5=z5
237    pmulhw      xmm0, [GOTOFF(ebx,PW_MF1613)]
238    pmulhw      xmm2, [GOTOFF(ebx,PW_F1082)]
239    psubw       xmm0, xmm1
240    psubw       xmm2, xmm5              ; xmm2=tmp10
241    paddw       xmm0, xmm5              ; xmm0=tmp12
242
243    ; -- Final output stage
244
245    psubw       xmm0, xmm3              ; xmm0=tmp6
246    movdqa      xmm1, xmm6
247    movdqa      xmm5, xmm7
248    paddw       xmm6, xmm3              ; xmm6=data0=(00 01 02 03 04 05 06 07)
249    paddw       xmm7, xmm0              ; xmm7=data1=(10 11 12 13 14 15 16 17)
250    psubw       xmm1, xmm3              ; xmm1=data7=(70 71 72 73 74 75 76 77)
251    psubw       xmm5, xmm0              ; xmm5=data6=(60 61 62 63 64 65 66 67)
252    psubw       xmm4, xmm0              ; xmm4=tmp5
253
254    movdqa      xmm3, xmm6              ; transpose coefficients(phase 1)
255    punpcklwd   xmm6, xmm7              ; xmm6=(00 10 01 11 02 12 03 13)
256    punpckhwd   xmm3, xmm7              ; xmm3=(04 14 05 15 06 16 07 17)
257    movdqa      xmm0, xmm5              ; transpose coefficients(phase 1)
258    punpcklwd   xmm5, xmm1              ; xmm5=(60 70 61 71 62 72 63 73)
259    punpckhwd   xmm0, xmm1              ; xmm0=(64 74 65 75 66 76 67 77)
260
261    movdqa      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
262    movdqa      xmm1, XMMWORD [wk(1)]   ; xmm1=tmp3
263
264    movdqa      XMMWORD [wk(0)], xmm5   ; wk(0)=(60 70 61 71 62 72 63 73)
265    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=(64 74 65 75 66 76 67 77)
266
267    paddw       xmm2, xmm4              ; xmm2=tmp4
268    movdqa      xmm5, xmm7
269    movdqa      xmm0, xmm1
270    paddw       xmm7, xmm4              ; xmm7=data2=(20 21 22 23 24 25 26 27)
271    paddw       xmm1, xmm2              ; xmm1=data4=(40 41 42 43 44 45 46 47)
272    psubw       xmm5, xmm4              ; xmm5=data5=(50 51 52 53 54 55 56 57)
273    psubw       xmm0, xmm2              ; xmm0=data3=(30 31 32 33 34 35 36 37)
274
275    movdqa      xmm4, xmm7              ; transpose coefficients(phase 1)
276    punpcklwd   xmm7, xmm0              ; xmm7=(20 30 21 31 22 32 23 33)
277    punpckhwd   xmm4, xmm0              ; xmm4=(24 34 25 35 26 36 27 37)
278    movdqa      xmm2, xmm1              ; transpose coefficients(phase 1)
279    punpcklwd   xmm1, xmm5              ; xmm1=(40 50 41 51 42 52 43 53)
280    punpckhwd   xmm2, xmm5              ; xmm2=(44 54 45 55 46 56 47 57)
281
282    movdqa      xmm0, xmm3              ; transpose coefficients(phase 2)
283    punpckldq   xmm3, xmm4              ; xmm3=(04 14 24 34 05 15 25 35)
284    punpckhdq   xmm0, xmm4              ; xmm0=(06 16 26 36 07 17 27 37)
285    movdqa      xmm5, xmm6              ; transpose coefficients(phase 2)
286    punpckldq   xmm6, xmm7              ; xmm6=(00 10 20 30 01 11 21 31)
287    punpckhdq   xmm5, xmm7              ; xmm5=(02 12 22 32 03 13 23 33)
288
289    movdqa      xmm4, XMMWORD [wk(0)]   ; xmm4=(60 70 61 71 62 72 63 73)
290    movdqa      xmm7, XMMWORD [wk(1)]   ; xmm7=(64 74 65 75 66 76 67 77)
291
292    movdqa      XMMWORD [wk(0)], xmm3   ; wk(0)=(04 14 24 34 05 15 25 35)
293    movdqa      XMMWORD [wk(1)], xmm0   ; wk(1)=(06 16 26 36 07 17 27 37)
294
295    movdqa      xmm3, xmm1              ; transpose coefficients(phase 2)
296    punpckldq   xmm1, xmm4              ; xmm1=(40 50 60 70 41 51 61 71)
297    punpckhdq   xmm3, xmm4              ; xmm3=(42 52 62 72 43 53 63 73)
298    movdqa      xmm0, xmm2              ; transpose coefficients(phase 2)
299    punpckldq   xmm2, xmm7              ; xmm2=(44 54 64 74 45 55 65 75)
300    punpckhdq   xmm0, xmm7              ; xmm0=(46 56 66 76 47 57 67 77)
301
302    movdqa      xmm4, xmm6              ; transpose coefficients(phase 3)
303    punpcklqdq  xmm6, xmm1              ; xmm6=col0=(00 10 20 30 40 50 60 70)
304    punpckhqdq  xmm4, xmm1              ; xmm4=col1=(01 11 21 31 41 51 61 71)
305    movdqa      xmm7, xmm5              ; transpose coefficients(phase 3)
306    punpcklqdq  xmm5, xmm3              ; xmm5=col2=(02 12 22 32 42 52 62 72)
307    punpckhqdq  xmm7, xmm3              ; xmm7=col3=(03 13 23 33 43 53 63 73)
308
309    movdqa      xmm1, XMMWORD [wk(0)]   ; xmm1=(04 14 24 34 05 15 25 35)
310    movdqa      xmm3, XMMWORD [wk(1)]   ; xmm3=(06 16 26 36 07 17 27 37)
311
312    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=col1
313    movdqa      XMMWORD [wk(1)], xmm7   ; wk(1)=col3
314
315    movdqa      xmm4, xmm1              ; transpose coefficients(phase 3)
316    punpcklqdq  xmm1, xmm2              ; xmm1=col4=(04 14 24 34 44 54 64 74)
317    punpckhqdq  xmm4, xmm2              ; xmm4=col5=(05 15 25 35 45 55 65 75)
318    movdqa      xmm7, xmm3              ; transpose coefficients(phase 3)
319    punpcklqdq  xmm3, xmm0              ; xmm3=col6=(06 16 26 36 46 56 66 76)
320    punpckhqdq  xmm7, xmm0              ; xmm7=col7=(07 17 27 37 47 57 67 77)
321.column_end:
322
323    ; -- Prefetch the next coefficient block
324
325    prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
326    prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
327    prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
328    prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
329
330    ; ---- Pass 2: process rows from work array, store into output array.
331
332    mov         eax, [original_ebp]
333    mov         edi, JSAMPARRAY [output_buf(eax)]  ; (JSAMPROW *)
334    mov         eax, JDIMENSION [output_col(eax)]
335
336    ; -- Even part
337
338    ; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6
339
340    movdqa      xmm2, xmm6
341    movdqa      xmm0, xmm5
342    psubw       xmm6, xmm1              ; xmm6=tmp11
343    psubw       xmm5, xmm3
344    paddw       xmm2, xmm1              ; xmm2=tmp10
345    paddw       xmm0, xmm3              ; xmm0=tmp13
346
347    psllw       xmm5, PRE_MULTIPLY_SCALE_BITS
348    pmulhw      xmm5, [GOTOFF(ebx,PW_F1414)]
349    psubw       xmm5, xmm0              ; xmm5=tmp12
350
351    movdqa      xmm1, xmm2
352    movdqa      xmm3, xmm6
353    psubw       xmm2, xmm0              ; xmm2=tmp3
354    psubw       xmm6, xmm5              ; xmm6=tmp2
355    paddw       xmm1, xmm0              ; xmm1=tmp0
356    paddw       xmm3, xmm5              ; xmm3=tmp1
357
358    movdqa      xmm0, XMMWORD [wk(0)]   ; xmm0=col1
359    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=col3
360
361    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=tmp3
362    movdqa      XMMWORD [wk(1)], xmm6   ; wk(1)=tmp2
363
364    ; -- Odd part
365
366    ; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7
367
368    movdqa      xmm2, xmm0
369    movdqa      xmm6, xmm4
370    psubw       xmm0, xmm7              ; xmm0=z12
371    psubw       xmm4, xmm5              ; xmm4=z10
372    paddw       xmm2, xmm7              ; xmm2=z11
373    paddw       xmm6, xmm5              ; xmm6=z13
374
375    movdqa      xmm7, xmm4              ; xmm7=z10(unscaled)
376    psllw       xmm0, PRE_MULTIPLY_SCALE_BITS
377    psllw       xmm4, PRE_MULTIPLY_SCALE_BITS
378
379    movdqa      xmm5, xmm2
380    psubw       xmm2, xmm6
381    paddw       xmm5, xmm6              ; xmm5=tmp7
382
383    psllw       xmm2, PRE_MULTIPLY_SCALE_BITS
384    pmulhw      xmm2, [GOTOFF(ebx,PW_F1414)]  ; xmm2=tmp11
385
386    ; To avoid overflow...
387    ;
388    ; (Original)
389    ; tmp12 = -2.613125930 * z10 + z5;
390    ;
391    ; (This implementation)
392    ; tmp12 = (-1.613125930 - 1) * z10 + z5;
393    ;       = -1.613125930 * z10 - z10 + z5;
394
395    movdqa      xmm6, xmm4
396    paddw       xmm4, xmm0
397    pmulhw      xmm4, [GOTOFF(ebx,PW_F1847)]   ; xmm4=z5
398    pmulhw      xmm6, [GOTOFF(ebx,PW_MF1613)]
399    pmulhw      xmm0, [GOTOFF(ebx,PW_F1082)]
400    psubw       xmm6, xmm7
401    psubw       xmm0, xmm4              ; xmm0=tmp10
402    paddw       xmm6, xmm4              ; xmm6=tmp12
403
404    ; -- Final output stage
405
406    psubw       xmm6, xmm5              ; xmm6=tmp6
407    movdqa      xmm7, xmm1
408    movdqa      xmm4, xmm3
409    paddw       xmm1, xmm5              ; xmm1=data0=(00 10 20 30 40 50 60 70)
410    paddw       xmm3, xmm6              ; xmm3=data1=(01 11 21 31 41 51 61 71)
411    psraw       xmm1, (PASS1_BITS+3)    ; descale
412    psraw       xmm3, (PASS1_BITS+3)    ; descale
413    psubw       xmm7, xmm5              ; xmm7=data7=(07 17 27 37 47 57 67 77)
414    psubw       xmm4, xmm6              ; xmm4=data6=(06 16 26 36 46 56 66 76)
415    psraw       xmm7, (PASS1_BITS+3)    ; descale
416    psraw       xmm4, (PASS1_BITS+3)    ; descale
417    psubw       xmm2, xmm6              ; xmm2=tmp5
418
419    packsswb    xmm1, xmm4        ; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
420    packsswb    xmm3, xmm7        ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
421
422    movdqa      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp2
423    movdqa      xmm6, XMMWORD [wk(0)]   ; xmm6=tmp3
424
425    paddw       xmm0, xmm2              ; xmm0=tmp4
426    movdqa      xmm4, xmm5
427    movdqa      xmm7, xmm6
428    paddw       xmm5, xmm2              ; xmm5=data2=(02 12 22 32 42 52 62 72)
429    paddw       xmm6, xmm0              ; xmm6=data4=(04 14 24 34 44 54 64 74)
430    psraw       xmm5, (PASS1_BITS+3)    ; descale
431    psraw       xmm6, (PASS1_BITS+3)    ; descale
432    psubw       xmm4, xmm2              ; xmm4=data5=(05 15 25 35 45 55 65 75)
433    psubw       xmm7, xmm0              ; xmm7=data3=(03 13 23 33 43 53 63 73)
434    psraw       xmm4, (PASS1_BITS+3)    ; descale
435    psraw       xmm7, (PASS1_BITS+3)    ; descale
436
437    movdqa      xmm2, [GOTOFF(ebx,PB_CENTERJSAMP)]  ; xmm2=[PB_CENTERJSAMP]
438
439    packsswb    xmm5, xmm6        ; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
440    packsswb    xmm7, xmm4        ; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
441
442    paddb       xmm1, xmm2
443    paddb       xmm3, xmm2
444    paddb       xmm5, xmm2
445    paddb       xmm7, xmm2
446
447    movdqa      xmm0, xmm1        ; transpose coefficients(phase 1)
448    punpcklbw   xmm1, xmm3        ; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
449    punpckhbw   xmm0, xmm3        ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
450    movdqa      xmm6, xmm5        ; transpose coefficients(phase 1)
451    punpcklbw   xmm5, xmm7        ; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
452    punpckhbw   xmm6, xmm7        ; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
453
454    movdqa      xmm4, xmm1        ; transpose coefficients(phase 2)
455    punpcklwd   xmm1, xmm5        ; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
456    punpckhwd   xmm4, xmm5        ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
457    movdqa      xmm2, xmm6        ; transpose coefficients(phase 2)
458    punpcklwd   xmm6, xmm0        ; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
459    punpckhwd   xmm2, xmm0        ; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
460
461    movdqa      xmm3, xmm1        ; transpose coefficients(phase 3)
462    punpckldq   xmm1, xmm6        ; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
463    punpckhdq   xmm3, xmm6        ; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
464    movdqa      xmm7, xmm4        ; transpose coefficients(phase 3)
465    punpckldq   xmm4, xmm2        ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
466    punpckhdq   xmm7, xmm2        ; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
467
468    pshufd      xmm5, xmm1, 0x4E  ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
469    pshufd      xmm0, xmm3, 0x4E  ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
470    pshufd      xmm6, xmm4, 0x4E  ; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
471    pshufd      xmm2, xmm7, 0x4E  ; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
472
473    mov         edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
474    mov         esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
475    movq        XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm1
476    movq        XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm3
477    mov         edx, JSAMPROW [edi+4*SIZEOF_JSAMPROW]
478    mov         esi, JSAMPROW [edi+6*SIZEOF_JSAMPROW]
479    movq        XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm4
480    movq        XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm7
481
482    mov         edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
483    mov         esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
484    movq        XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm5
485    movq        XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm0
486    mov         edx, JSAMPROW [edi+5*SIZEOF_JSAMPROW]
487    mov         esi, JSAMPROW [edi+7*SIZEOF_JSAMPROW]
488    movq        XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm6
489    movq        XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm2
490
491    pop         edi
492    pop         esi
493;   pop         edx                     ; need not be preserved
494;   pop         ecx                     ; unused
495    poppic      ebx
496    mov         esp, ebp                ; esp <- aligned ebp
497    pop         esp                     ; esp <- original ebp
498    pop         ebp
499    ret
500
501; For some reason, the OS X linker does not honor the request to align the
502; segment unless we do this.
503    align       32
504