summaryrefslogtreecommitdiff
path: root/media/libjpeg/simd/i386/jidctflt-sse2.asm
blob: c646eaef76ef31bff57a31c9e7ad7ca746e9eff1 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
;
; jidctflt.asm - floating-point IDCT (SSE & SSE2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2016, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
; For conditions of distribution and use, see copyright notice in jsimdext.inc
;
; This file should be assembled with NASM (Netwide Assembler),
; can *not* be assembled with Microsoft's MASM or any compatible
; assembler (including Borland's Turbo Assembler).
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
; This file contains a floating-point implementation of the inverse DCT
; (Discrete Cosine Transform). The following code is based directly on
; the IJG's original jidctflt.c; see the jidctflt.c for more details.

%include "jsimdext.inc"
%include "jdct.inc"

; --------------------------------------------------------------------------

%macro unpcklps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
    shufps      %1, %2, 0x44
%endmacro

%macro unpckhps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
    shufps      %1, %2, 0xEE
%endmacro

; --------------------------------------------------------------------------
    SECTION     SEG_CONST

    alignz      32
    GLOBAL_DATA(jconst_idct_float_sse2)

EXTN(jconst_idct_float_sse2):

PD_1_414        times 4  dd  1.414213562373095048801689
PD_1_847        times 4  dd  1.847759065022573512256366
PD_1_082        times 4  dd  1.082392200292393968799446
PD_M2_613       times 4  dd -2.613125929752753055713286
PD_RNDINT_MAGIC times 4  dd  100663296.0  ; (float)(0x00C00000 << 3)
PB_CENTERJSAMP  times 16 db  CENTERJSAMPLE

    alignz      32

; --------------------------------------------------------------------------
    SECTION     SEG_TEXT
    BITS        32
;
; Perform dequantization and inverse DCT on one block of coefficients.
;
; GLOBAL(void)
; jsimd_idct_float_sse2(void *dct_table, JCOEFPTR coef_block,
;                       JSAMPARRAY output_buf, JDIMENSION output_col)
;

%define dct_table(b)   (b) + 8          ; void *dct_table
%define coef_block(b)  (b) + 12         ; JCOEFPTR coef_block
%define output_buf(b)  (b) + 16         ; JSAMPARRAY output_buf
%define output_col(b)  (b) + 20         ; JDIMENSION output_col

%define original_ebp   ebp + 0
%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD
                                        ; xmmword wk[WK_NUM]
%define WK_NUM         2
%define workspace      wk(0) - DCTSIZE2 * SIZEOF_FAST_FLOAT
                                        ; FAST_FLOAT workspace[DCTSIZE2]

    align       32
    GLOBAL_FUNCTION(jsimd_idct_float_sse2)

EXTN(jsimd_idct_float_sse2):
    push        ebp
    mov         eax, esp                     ; eax = original ebp
    sub         esp, byte 4
    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
    mov         [esp], eax
    mov         ebp, esp                     ; ebp = aligned ebp
    lea         esp, [workspace]
    push        ebx
;   push        ecx                     ; need not be preserved
;   push        edx                     ; need not be preserved
    push        esi
    push        edi

    get_GOT     ebx                     ; get GOT address

    ; ---- Pass 1: process columns from input, store into work array.

;   mov         eax, [original_ebp]
    mov         edx, POINTER [dct_table(eax)]    ; quantptr
    mov         esi, JCOEFPTR [coef_block(eax)]  ; inptr
    lea         edi, [workspace]                 ; FAST_FLOAT *wsptr
    mov         ecx, DCTSIZE/4                   ; ctr
    alignx      16, 7
.columnloop:
%ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
    mov         eax, dword [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
    or          eax, dword [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
    jnz         near .columnDCT

    movq        xmm1, XMM_MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
    movq        xmm2, XMM_MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
    movq        xmm3, XMM_MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
    movq        xmm4, XMM_MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
    movq        xmm5, XMM_MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
    movq        xmm6, XMM_MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
    movq        xmm7, XMM_MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
    por         xmm1, xmm2
    por         xmm3, xmm4
    por         xmm5, xmm6
    por         xmm1, xmm3
    por         xmm5, xmm7
    por         xmm1, xmm5
    packsswb    xmm1, xmm1
    movd        eax, xmm1
    test        eax, eax
    jnz         short .columnDCT

    ; -- AC terms all zero

    movq        xmm0, XMM_MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]

    punpcklwd   xmm0, xmm0                  ; xmm0=(00 00 01 01 02 02 03 03)
    psrad       xmm0, (DWORD_BIT-WORD_BIT)  ; xmm0=in0=(00 01 02 03)
    cvtdq2ps    xmm0, xmm0                  ; xmm0=in0=(00 01 02 03)

    mulps       xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]

    movaps      xmm1, xmm0
    movaps      xmm2, xmm0
    movaps      xmm3, xmm0

    shufps      xmm0, xmm0, 0x00        ; xmm0=(00 00 00 00)
    shufps      xmm1, xmm1, 0x55        ; xmm1=(01 01 01 01)
    shufps      xmm2, xmm2, 0xAA        ; xmm2=(02 02 02 02)
    shufps      xmm3, xmm3, 0xFF        ; xmm3=(03 03 03 03)

    movaps      XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
    movaps      XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
    movaps      XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm1
    movaps      XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
    movaps      XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm2
    movaps      XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm2
    movaps      XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm3
    movaps      XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
    jmp         near .nextcolumn
    alignx      16, 7
%endif
.columnDCT:

    ; -- Even part

    movq        xmm0, XMM_MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
    movq        xmm1, XMM_MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
    movq        xmm2, XMM_MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
    movq        xmm3, XMM_MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]

    punpcklwd   xmm0, xmm0                  ; xmm0=(00 00 01 01 02 02 03 03)
    punpcklwd   xmm1, xmm1                  ; xmm1=(20 20 21 21 22 22 23 23)
    psrad       xmm0, (DWORD_BIT-WORD_BIT)  ; xmm0=in0=(00 01 02 03)
    psrad       xmm1, (DWORD_BIT-WORD_BIT)  ; xmm1=in2=(20 21 22 23)
    cvtdq2ps    xmm0, xmm0                  ; xmm0=in0=(00 01 02 03)
    cvtdq2ps    xmm1, xmm1                  ; xmm1=in2=(20 21 22 23)

    punpcklwd   xmm2, xmm2                  ; xmm2=(40 40 41 41 42 42 43 43)
    punpcklwd   xmm3, xmm3                  ; xmm3=(60 60 61 61 62 62 63 63)
    psrad       xmm2, (DWORD_BIT-WORD_BIT)  ; xmm2=in4=(40 41 42 43)
    psrad       xmm3, (DWORD_BIT-WORD_BIT)  ; xmm3=in6=(60 61 62 63)
    cvtdq2ps    xmm2, xmm2                  ; xmm2=in4=(40 41 42 43)
    cvtdq2ps    xmm3, xmm3                  ; xmm3=in6=(60 61 62 63)

    mulps       xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
    mulps       xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
    mulps       xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
    mulps       xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]

    movaps      xmm4, xmm0
    movaps      xmm5, xmm1
    subps       xmm0, xmm2              ; xmm0=tmp11
    subps       xmm1, xmm3
    addps       xmm4, xmm2              ; xmm4=tmp10
    addps       xmm5, xmm3              ; xmm5=tmp13

    mulps       xmm1, [GOTOFF(ebx,PD_1_414)]
    subps       xmm1, xmm5              ; xmm1=tmp12

    movaps      xmm6, xmm4
    movaps      xmm7, xmm0
    subps       xmm4, xmm5              ; xmm4=tmp3
    subps       xmm0, xmm1              ; xmm0=tmp2
    addps       xmm6, xmm5              ; xmm6=tmp0
    addps       xmm7, xmm1              ; xmm7=tmp1

    movaps      XMMWORD [wk(1)], xmm4   ; tmp3
    movaps      XMMWORD [wk(0)], xmm0   ; tmp2

    ; -- Odd part

    movq        xmm2, XMM_MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
    movq        xmm3, XMM_MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
    movq        xmm5, XMM_MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
    movq        xmm1, XMM_MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]

    punpcklwd   xmm2, xmm2                  ; xmm2=(10 10 11 11 12 12 13 13)
    punpcklwd   xmm3, xmm3                  ; xmm3=(30 30 31 31 32 32 33 33)
    psrad       xmm2, (DWORD_BIT-WORD_BIT)  ; xmm2=in1=(10 11 12 13)
    psrad       xmm3, (DWORD_BIT-WORD_BIT)  ; xmm3=in3=(30 31 32 33)
    cvtdq2ps    xmm2, xmm2                  ; xmm2=in1=(10 11 12 13)
    cvtdq2ps    xmm3, xmm3                  ; xmm3=in3=(30 31 32 33)

    punpcklwd   xmm5, xmm5                  ; xmm5=(50 50 51 51 52 52 53 53)
    punpcklwd   xmm1, xmm1                  ; xmm1=(70 70 71 71 72 72 73 73)
    psrad       xmm5, (DWORD_BIT-WORD_BIT)  ; xmm5=in5=(50 51 52 53)
    psrad       xmm1, (DWORD_BIT-WORD_BIT)  ; xmm1=in7=(70 71 72 73)
    cvtdq2ps    xmm5, xmm5                  ; xmm5=in5=(50 51 52 53)
    cvtdq2ps    xmm1, xmm1                  ; xmm1=in7=(70 71 72 73)

    mulps       xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
    mulps       xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
    mulps       xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
    mulps       xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]

    movaps      xmm4, xmm2
    movaps      xmm0, xmm5
    addps       xmm2, xmm1              ; xmm2=z11
    addps       xmm5, xmm3              ; xmm5=z13
    subps       xmm4, xmm1              ; xmm4=z12
    subps       xmm0, xmm3              ; xmm0=z10

    movaps      xmm1, xmm2
    subps       xmm2, xmm5
    addps       xmm1, xmm5              ; xmm1=tmp7

    mulps       xmm2, [GOTOFF(ebx,PD_1_414)]  ; xmm2=tmp11

    movaps      xmm3, xmm0
    addps       xmm0, xmm4
    mulps       xmm0, [GOTOFF(ebx,PD_1_847)]   ; xmm0=z5
    mulps       xmm3, [GOTOFF(ebx,PD_M2_613)]  ; xmm3=(z10 * -2.613125930)
    mulps       xmm4, [GOTOFF(ebx,PD_1_082)]   ; xmm4=(z12 * 1.082392200)
    addps       xmm3, xmm0                     ; xmm3=tmp12
    subps       xmm4, xmm0                     ; xmm4=tmp10

    ; -- Final output stage

    subps       xmm3, xmm1              ; xmm3=tmp6
    movaps      xmm5, xmm6
    movaps      xmm0, xmm7
    addps       xmm6, xmm1              ; xmm6=data0=(00 01 02 03)
    addps       xmm7, xmm3              ; xmm7=data1=(10 11 12 13)
    subps       xmm5, xmm1              ; xmm5=data7=(70 71 72 73)
    subps       xmm0, xmm3              ; xmm0=data6=(60 61 62 63)
    subps       xmm2, xmm3              ; xmm2=tmp5

    movaps      xmm1, xmm6              ; transpose coefficients(phase 1)
    unpcklps    xmm6, xmm7              ; xmm6=(00 10 01 11)
    unpckhps    xmm1, xmm7              ; xmm1=(02 12 03 13)
    movaps      xmm3, xmm0              ; transpose coefficients(phase 1)
    unpcklps    xmm0, xmm5              ; xmm0=(60 70 61 71)
    unpckhps    xmm3, xmm5              ; xmm3=(62 72 63 73)

    movaps      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
    movaps      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp3

    movaps      XMMWORD [wk(0)], xmm0   ; wk(0)=(60 70 61 71)
    movaps      XMMWORD [wk(1)], xmm3   ; wk(1)=(62 72 63 73)

    addps       xmm4, xmm2              ; xmm4=tmp4
    movaps      xmm0, xmm7
    movaps      xmm3, xmm5
    addps       xmm7, xmm2              ; xmm7=data2=(20 21 22 23)
    addps       xmm5, xmm4              ; xmm5=data4=(40 41 42 43)
    subps       xmm0, xmm2              ; xmm0=data5=(50 51 52 53)
    subps       xmm3, xmm4              ; xmm3=data3=(30 31 32 33)

    movaps      xmm2, xmm7              ; transpose coefficients(phase 1)
    unpcklps    xmm7, xmm3              ; xmm7=(20 30 21 31)
    unpckhps    xmm2, xmm3              ; xmm2=(22 32 23 33)
    movaps      xmm4, xmm5              ; transpose coefficients(phase 1)
    unpcklps    xmm5, xmm0              ; xmm5=(40 50 41 51)
    unpckhps    xmm4, xmm0              ; xmm4=(42 52 43 53)

    movaps      xmm3, xmm6              ; transpose coefficients(phase 2)
    unpcklps2   xmm6, xmm7              ; xmm6=(00 10 20 30)
    unpckhps2   xmm3, xmm7              ; xmm3=(01 11 21 31)
    movaps      xmm0, xmm1              ; transpose coefficients(phase 2)
    unpcklps2   xmm1, xmm2              ; xmm1=(02 12 22 32)
    unpckhps2   xmm0, xmm2              ; xmm0=(03 13 23 33)

    movaps      xmm7, XMMWORD [wk(0)]   ; xmm7=(60 70 61 71)
    movaps      xmm2, XMMWORD [wk(1)]   ; xmm2=(62 72 63 73)

    movaps      XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm6
    movaps      XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
    movaps      XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm1
    movaps      XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm0

    movaps      xmm6, xmm5              ; transpose coefficients(phase 2)
    unpcklps2   xmm5, xmm7              ; xmm5=(40 50 60 70)
    unpckhps2   xmm6, xmm7              ; xmm6=(41 51 61 71)
    movaps      xmm3, xmm4              ; transpose coefficients(phase 2)
    unpcklps2   xmm4, xmm2              ; xmm4=(42 52 62 72)
    unpckhps2   xmm3, xmm2              ; xmm3=(43 53 63 73)

    movaps      XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm5
    movaps      XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
    movaps      XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm4
    movaps      XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3

.nextcolumn:
    add         esi, byte 4*SIZEOF_JCOEF               ; coef_block
    add         edx, byte 4*SIZEOF_FLOAT_MULT_TYPE     ; quantptr
    add         edi,      4*DCTSIZE*SIZEOF_FAST_FLOAT  ; wsptr
    dec         ecx                                    ; ctr
    jnz         near .columnloop

    ; -- Prefetch the next coefficient block

    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]

    ; ---- Pass 2: process rows from work array, store into output array.

    mov         eax, [original_ebp]
    lea         esi, [workspace]                   ; FAST_FLOAT *wsptr
    mov         edi, JSAMPARRAY [output_buf(eax)]  ; (JSAMPROW *)
    mov         eax, JDIMENSION [output_col(eax)]
    mov         ecx, DCTSIZE/4                     ; ctr
    alignx      16, 7
.rowloop:

    ; -- Even part

    movaps      xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
    movaps      xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
    movaps      xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
    movaps      xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]

    movaps      xmm4, xmm0
    movaps      xmm5, xmm1
    subps       xmm0, xmm2              ; xmm0=tmp11
    subps       xmm1, xmm3
    addps       xmm4, xmm2              ; xmm4=tmp10
    addps       xmm5, xmm3              ; xmm5=tmp13

    mulps       xmm1, [GOTOFF(ebx,PD_1_414)]
    subps       xmm1, xmm5              ; xmm1=tmp12

    movaps      xmm6, xmm4
    movaps      xmm7, xmm0
    subps       xmm4, xmm5              ; xmm4=tmp3
    subps       xmm0, xmm1              ; xmm0=tmp2
    addps       xmm6, xmm5              ; xmm6=tmp0
    addps       xmm7, xmm1              ; xmm7=tmp1

    movaps      XMMWORD [wk(1)], xmm4   ; tmp3
    movaps      XMMWORD [wk(0)], xmm0   ; tmp2

    ; -- Odd part

    movaps      xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
    movaps      xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
    movaps      xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
    movaps      xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]

    movaps      xmm4, xmm2
    movaps      xmm0, xmm5
    addps       xmm2, xmm1              ; xmm2=z11
    addps       xmm5, xmm3              ; xmm5=z13
    subps       xmm4, xmm1              ; xmm4=z12
    subps       xmm0, xmm3              ; xmm0=z10

    movaps      xmm1, xmm2
    subps       xmm2, xmm5
    addps       xmm1, xmm5              ; xmm1=tmp7

    mulps       xmm2, [GOTOFF(ebx,PD_1_414)]  ; xmm2=tmp11

    movaps      xmm3, xmm0
    addps       xmm0, xmm4
    mulps       xmm0, [GOTOFF(ebx,PD_1_847)]   ; xmm0=z5
    mulps       xmm3, [GOTOFF(ebx,PD_M2_613)]  ; xmm3=(z10 * -2.613125930)
    mulps       xmm4, [GOTOFF(ebx,PD_1_082)]   ; xmm4=(z12 * 1.082392200)
    addps       xmm3, xmm0                     ; xmm3=tmp12
    subps       xmm4, xmm0                     ; xmm4=tmp10

    ; -- Final output stage

    subps       xmm3, xmm1              ; xmm3=tmp6
    movaps      xmm5, xmm6
    movaps      xmm0, xmm7
    addps       xmm6, xmm1              ; xmm6=data0=(00 10 20 30)
    addps       xmm7, xmm3              ; xmm7=data1=(01 11 21 31)
    subps       xmm5, xmm1              ; xmm5=data7=(07 17 27 37)
    subps       xmm0, xmm3              ; xmm0=data6=(06 16 26 36)
    subps       xmm2, xmm3              ; xmm2=tmp5

    movaps      xmm1, [GOTOFF(ebx,PD_RNDINT_MAGIC)]  ; xmm1=[PD_RNDINT_MAGIC]
    pcmpeqd     xmm3, xmm3
    psrld       xmm3, WORD_BIT          ; xmm3={0xFFFF 0x0000 0xFFFF 0x0000 ..}

    addps       xmm6, xmm1              ; xmm6=roundint(data0/8)=(00 ** 10 ** 20 ** 30 **)
    addps       xmm7, xmm1              ; xmm7=roundint(data1/8)=(01 ** 11 ** 21 ** 31 **)
    addps       xmm0, xmm1              ; xmm0=roundint(data6/8)=(06 ** 16 ** 26 ** 36 **)
    addps       xmm5, xmm1              ; xmm5=roundint(data7/8)=(07 ** 17 ** 27 ** 37 **)

    pand        xmm6, xmm3              ; xmm6=(00 -- 10 -- 20 -- 30 --)
    pslld       xmm7, WORD_BIT          ; xmm7=(-- 01 -- 11 -- 21 -- 31)
    pand        xmm0, xmm3              ; xmm0=(06 -- 16 -- 26 -- 36 --)
    pslld       xmm5, WORD_BIT          ; xmm5=(-- 07 -- 17 -- 27 -- 37)
    por         xmm6, xmm7              ; xmm6=(00 01 10 11 20 21 30 31)
    por         xmm0, xmm5              ; xmm0=(06 07 16 17 26 27 36 37)

    movaps      xmm1, XMMWORD [wk(0)]   ; xmm1=tmp2
    movaps      xmm3, XMMWORD [wk(1)]   ; xmm3=tmp3

    addps       xmm4, xmm2              ; xmm4=tmp4
    movaps      xmm7, xmm1
    movaps      xmm5, xmm3
    addps       xmm1, xmm2              ; xmm1=data2=(02 12 22 32)
    addps       xmm3, xmm4              ; xmm3=data4=(04 14 24 34)
    subps       xmm7, xmm2              ; xmm7=data5=(05 15 25 35)
    subps       xmm5, xmm4              ; xmm5=data3=(03 13 23 33)

    movaps      xmm2, [GOTOFF(ebx,PD_RNDINT_MAGIC)]  ; xmm2=[PD_RNDINT_MAGIC]
    pcmpeqd     xmm4, xmm4
    psrld       xmm4, WORD_BIT          ; xmm4={0xFFFF 0x0000 0xFFFF 0x0000 ..}

    addps       xmm3, xmm2              ; xmm3=roundint(data4/8)=(04 ** 14 ** 24 ** 34 **)
    addps       xmm7, xmm2              ; xmm7=roundint(data5/8)=(05 ** 15 ** 25 ** 35 **)
    addps       xmm1, xmm2              ; xmm1=roundint(data2/8)=(02 ** 12 ** 22 ** 32 **)
    addps       xmm5, xmm2              ; xmm5=roundint(data3/8)=(03 ** 13 ** 23 ** 33 **)

    pand        xmm3, xmm4              ; xmm3=(04 -- 14 -- 24 -- 34 --)
    pslld       xmm7, WORD_BIT          ; xmm7=(-- 05 -- 15 -- 25 -- 35)
    pand        xmm1, xmm4              ; xmm1=(02 -- 12 -- 22 -- 32 --)
    pslld       xmm5, WORD_BIT          ; xmm5=(-- 03 -- 13 -- 23 -- 33)
    por         xmm3, xmm7              ; xmm3=(04 05 14 15 24 25 34 35)
    por         xmm1, xmm5              ; xmm1=(02 03 12 13 22 23 32 33)

    movdqa      xmm2, [GOTOFF(ebx,PB_CENTERJSAMP)]  ; xmm2=[PB_CENTERJSAMP]

    packsswb    xmm6, xmm3        ; xmm6=(00 01 10 11 20 21 30 31 04 05 14 15 24 25 34 35)
    packsswb    xmm1, xmm0        ; xmm1=(02 03 12 13 22 23 32 33 06 07 16 17 26 27 36 37)
    paddb       xmm6, xmm2
    paddb       xmm1, xmm2

    movdqa      xmm4, xmm6        ; transpose coefficients(phase 2)
    punpcklwd   xmm6, xmm1        ; xmm6=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
    punpckhwd   xmm4, xmm1        ; xmm4=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)

    movdqa      xmm7, xmm6        ; transpose coefficients(phase 3)
    punpckldq   xmm6, xmm4        ; xmm6=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
    punpckhdq   xmm7, xmm4        ; xmm7=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)

    pshufd      xmm5, xmm6, 0x4E  ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
    pshufd      xmm3, xmm7, 0x4E  ; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)

    pushpic     ebx                     ; save GOT address

    mov         edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
    mov         ebx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
    movq        XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm6
    movq        XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE], xmm7
    mov         edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
    mov         ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
    movq        XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm5
    movq        XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE], xmm3

    poppic      ebx                     ; restore GOT address

    add         esi, byte 4*SIZEOF_FAST_FLOAT  ; wsptr
    add         edi, byte 4*SIZEOF_JSAMPROW
    dec         ecx                            ; ctr
    jnz         near .rowloop

    pop         edi
    pop         esi
;   pop         edx                     ; need not be preserved
;   pop         ecx                     ; need not be preserved
    pop         ebx
    mov         esp, ebp                ; esp <- aligned ebp
    pop         esp                     ; esp <- original ebp
    pop         ebp
    ret

; For some reason, the OS X linker does not honor the request to align the
; segment unless we do this.
    align       32