• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jidctflt.asm - floating-point IDCT (SSE & MMX)
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 floating-point implementation of the inverse DCT
18; (Discrete Cosine Transform). The following code is based directly on
19; the IJG's original jidctflt.c; see the jidctflt.c for more details.
20
21%include "jsimdext.inc"
22%include "jdct.inc"
23
24; --------------------------------------------------------------------------
25
26%macro unpcklps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
27    shufps      %1, %2, 0x44
28%endmacro
29
30%macro unpckhps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
31    shufps      %1, %2, 0xEE
32%endmacro
33
34; --------------------------------------------------------------------------
35    SECTION     SEG_CONST
36
37    alignz      32
38    GLOBAL_DATA(jconst_idct_float_sse)
39
40EXTN(jconst_idct_float_sse):
41
42PD_1_414       times 4 dd  1.414213562373095048801689
43PD_1_847       times 4 dd  1.847759065022573512256366
44PD_1_082       times 4 dd  1.082392200292393968799446
45PD_M2_613      times 4 dd -2.613125929752753055713286
46PD_0_125       times 4 dd  0.125        ; 1/8
47PB_CENTERJSAMP times 8 db  CENTERJSAMPLE
48
49    alignz      32
50
51; --------------------------------------------------------------------------
52    SECTION     SEG_TEXT
53    BITS        32
54;
55; Perform dequantization and inverse DCT on one block of coefficients.
56;
57; GLOBAL(void)
58; jsimd_idct_float_sse(void *dct_table, JCOEFPTR coef_block,
59;                      JSAMPARRAY output_buf, JDIMENSION output_col)
60;
61
62%define dct_table(b)   (b) + 8          ; void *dct_table
63%define coef_block(b)  (b) + 12         ; JCOEFPTR coef_block
64%define output_buf(b)  (b) + 16         ; JSAMPARRAY output_buf
65%define output_col(b)  (b) + 20         ; JDIMENSION output_col
66
67%define original_ebp   ebp + 0
68%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD
69                                        ; xmmword wk[WK_NUM]
70%define WK_NUM         2
71%define workspace      wk(0) - DCTSIZE2 * SIZEOF_FAST_FLOAT
72                                        ; FAST_FLOAT workspace[DCTSIZE2]
73
74    align       32
75    GLOBAL_FUNCTION(jsimd_idct_float_sse)
76
77EXTN(jsimd_idct_float_sse):
78    push        ebp
79    mov         eax, esp                     ; eax = original ebp
80    sub         esp, byte 4
81    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
82    mov         [esp], eax
83    mov         ebp, esp                     ; ebp = aligned ebp
84    lea         esp, [workspace]
85    push        ebx
86;   push        ecx                     ; need not be preserved
87;   push        edx                     ; need not be preserved
88    push        esi
89    push        edi
90
91    get_GOT     ebx                     ; get GOT address
92
93    ; ---- Pass 1: process columns from input, store into work array.
94
95;   mov         eax, [original_ebp]
96    mov         edx, POINTER [dct_table(eax)]    ; quantptr
97    mov         esi, JCOEFPTR [coef_block(eax)]  ; inptr
98    lea         edi, [workspace]                 ; FAST_FLOAT *wsptr
99    mov         ecx, DCTSIZE/4                   ; ctr
100    alignx      16, 7
101.columnloop:
102%ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
103    mov         eax, dword [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
104    or          eax, dword [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
105    jnz         near .columnDCT
106
107    movq        mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
108    movq        mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
109    por         mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
110    por         mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
111    por         mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
112    por         mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
113    por         mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
114    por         mm1, mm0
115    packsswb    mm1, mm1
116    movd        eax, mm1
117    test        eax, eax
118    jnz         short .columnDCT
119
120    ; -- AC terms all zero
121
122    movq        mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
123
124    punpckhwd   mm1, mm0                   ; mm1=(** 02 ** 03)
125    punpcklwd   mm0, mm0                   ; mm0=(00 00 01 01)
126    psrad       mm1, (DWORD_BIT-WORD_BIT)  ; mm1=in0H=(02 03)
127    psrad       mm0, (DWORD_BIT-WORD_BIT)  ; mm0=in0L=(00 01)
128    cvtpi2ps    xmm3, mm1                  ; xmm3=(02 03 ** **)
129    cvtpi2ps    xmm0, mm0                  ; xmm0=(00 01 ** **)
130    movlhps     xmm0, xmm3                 ; xmm0=in0=(00 01 02 03)
131
132    mulps       xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
133
134    movaps      xmm1, xmm0
135    movaps      xmm2, xmm0
136    movaps      xmm3, xmm0
137
138    shufps      xmm0, xmm0, 0x00        ; xmm0=(00 00 00 00)
139    shufps      xmm1, xmm1, 0x55        ; xmm1=(01 01 01 01)
140    shufps      xmm2, xmm2, 0xAA        ; xmm2=(02 02 02 02)
141    shufps      xmm3, xmm3, 0xFF        ; xmm3=(03 03 03 03)
142
143    movaps      XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
144    movaps      XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
145    movaps      XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm1
146    movaps      XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
147    movaps      XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm2
148    movaps      XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm2
149    movaps      XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm3
150    movaps      XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
151    jmp         near .nextcolumn
152    alignx      16, 7
153%endif
154.columnDCT:
155
156    ; -- Even part
157
158    movq        mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
159    movq        mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
160    movq        mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
161    movq        mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
162
163    punpckhwd   mm4, mm0                ; mm4=(** 02 ** 03)
164    punpcklwd   mm0, mm0                ; mm0=(00 00 01 01)
165    punpckhwd   mm5, mm1                ; mm5=(** 22 ** 23)
166    punpcklwd   mm1, mm1                ; mm1=(20 20 21 21)
167
168    psrad       mm4, (DWORD_BIT-WORD_BIT)  ; mm4=in0H=(02 03)
169    psrad       mm0, (DWORD_BIT-WORD_BIT)  ; mm0=in0L=(00 01)
170    cvtpi2ps    xmm4, mm4                  ; xmm4=(02 03 ** **)
171    cvtpi2ps    xmm0, mm0                  ; xmm0=(00 01 ** **)
172    psrad       mm5, (DWORD_BIT-WORD_BIT)  ; mm5=in2H=(22 23)
173    psrad       mm1, (DWORD_BIT-WORD_BIT)  ; mm1=in2L=(20 21)
174    cvtpi2ps    xmm5, mm5                  ; xmm5=(22 23 ** **)
175    cvtpi2ps    xmm1, mm1                  ; xmm1=(20 21 ** **)
176
177    punpckhwd   mm6, mm2                ; mm6=(** 42 ** 43)
178    punpcklwd   mm2, mm2                ; mm2=(40 40 41 41)
179    punpckhwd   mm7, mm3                ; mm7=(** 62 ** 63)
180    punpcklwd   mm3, mm3                ; mm3=(60 60 61 61)
181
182    psrad       mm6, (DWORD_BIT-WORD_BIT)  ; mm6=in4H=(42 43)
183    psrad       mm2, (DWORD_BIT-WORD_BIT)  ; mm2=in4L=(40 41)
184    cvtpi2ps    xmm6, mm6                  ; xmm6=(42 43 ** **)
185    cvtpi2ps    xmm2, mm2                  ; xmm2=(40 41 ** **)
186    psrad       mm7, (DWORD_BIT-WORD_BIT)  ; mm7=in6H=(62 63)
187    psrad       mm3, (DWORD_BIT-WORD_BIT)  ; mm3=in6L=(60 61)
188    cvtpi2ps    xmm7, mm7                  ; xmm7=(62 63 ** **)
189    cvtpi2ps    xmm3, mm3                  ; xmm3=(60 61 ** **)
190
191    movlhps     xmm0, xmm4              ; xmm0=in0=(00 01 02 03)
192    movlhps     xmm1, xmm5              ; xmm1=in2=(20 21 22 23)
193    mulps       xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
194    mulps       xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
195
196    movlhps     xmm2, xmm6              ; xmm2=in4=(40 41 42 43)
197    movlhps     xmm3, xmm7              ; xmm3=in6=(60 61 62 63)
198    mulps       xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
199    mulps       xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
200
201    movaps      xmm4, xmm0
202    movaps      xmm5, xmm1
203    subps       xmm0, xmm2              ; xmm0=tmp11
204    subps       xmm1, xmm3
205    addps       xmm4, xmm2              ; xmm4=tmp10
206    addps       xmm5, xmm3              ; xmm5=tmp13
207
208    mulps       xmm1, [GOTOFF(ebx,PD_1_414)]
209    subps       xmm1, xmm5              ; xmm1=tmp12
210
211    movaps      xmm6, xmm4
212    movaps      xmm7, xmm0
213    subps       xmm4, xmm5              ; xmm4=tmp3
214    subps       xmm0, xmm1              ; xmm0=tmp2
215    addps       xmm6, xmm5              ; xmm6=tmp0
216    addps       xmm7, xmm1              ; xmm7=tmp1
217
218    movaps      XMMWORD [wk(1)], xmm4   ; tmp3
219    movaps      XMMWORD [wk(0)], xmm0   ; tmp2
220
221    ; -- Odd part
222
223    movq        mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
224    movq        mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
225    movq        mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
226    movq        mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
227
228    punpckhwd   mm6, mm4                ; mm6=(** 12 ** 13)
229    punpcklwd   mm4, mm4                ; mm4=(10 10 11 11)
230    punpckhwd   mm2, mm0                ; mm2=(** 32 ** 33)
231    punpcklwd   mm0, mm0                ; mm0=(30 30 31 31)
232
233    psrad       mm6, (DWORD_BIT-WORD_BIT)  ; mm6=in1H=(12 13)
234    psrad       mm4, (DWORD_BIT-WORD_BIT)  ; mm4=in1L=(10 11)
235    cvtpi2ps    xmm4, mm6                  ; xmm4=(12 13 ** **)
236    cvtpi2ps    xmm2, mm4                  ; xmm2=(10 11 ** **)
237    psrad       mm2, (DWORD_BIT-WORD_BIT)  ; mm2=in3H=(32 33)
238    psrad       mm0, (DWORD_BIT-WORD_BIT)  ; mm0=in3L=(30 31)
239    cvtpi2ps    xmm0, mm2                  ; xmm0=(32 33 ** **)
240    cvtpi2ps    xmm3, mm0                  ; xmm3=(30 31 ** **)
241
242    punpckhwd   mm7, mm5                ; mm7=(** 52 ** 53)
243    punpcklwd   mm5, mm5                ; mm5=(50 50 51 51)
244    punpckhwd   mm3, mm1                ; mm3=(** 72 ** 73)
245    punpcklwd   mm1, mm1                ; mm1=(70 70 71 71)
246
247    movlhps     xmm2, xmm4              ; xmm2=in1=(10 11 12 13)
248    movlhps     xmm3, xmm0              ; xmm3=in3=(30 31 32 33)
249
250    psrad       mm7, (DWORD_BIT-WORD_BIT)  ; mm7=in5H=(52 53)
251    psrad       mm5, (DWORD_BIT-WORD_BIT)  ; mm5=in5L=(50 51)
252    cvtpi2ps    xmm4, mm7                  ; xmm4=(52 53 ** **)
253    cvtpi2ps    xmm5, mm5                  ; xmm5=(50 51 ** **)
254    psrad       mm3, (DWORD_BIT-WORD_BIT)  ; mm3=in7H=(72 73)
255    psrad       mm1, (DWORD_BIT-WORD_BIT)  ; mm1=in7L=(70 71)
256    cvtpi2ps    xmm0, mm3                  ; xmm0=(72 73 ** **)
257    cvtpi2ps    xmm1, mm1                  ; xmm1=(70 71 ** **)
258
259    mulps       xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
260    mulps       xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
261
262    movlhps     xmm5, xmm4              ; xmm5=in5=(50 51 52 53)
263    movlhps     xmm1, xmm0              ; xmm1=in7=(70 71 72 73)
264    mulps       xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
265    mulps       xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
266
267    movaps      xmm4, xmm2
268    movaps      xmm0, xmm5
269    addps       xmm2, xmm1              ; xmm2=z11
270    addps       xmm5, xmm3              ; xmm5=z13
271    subps       xmm4, xmm1              ; xmm4=z12
272    subps       xmm0, xmm3              ; xmm0=z10
273
274    movaps      xmm1, xmm2
275    subps       xmm2, xmm5
276    addps       xmm1, xmm5              ; xmm1=tmp7
277
278    mulps       xmm2, [GOTOFF(ebx,PD_1_414)]  ; xmm2=tmp11
279
280    movaps      xmm3, xmm0
281    addps       xmm0, xmm4
282    mulps       xmm0, [GOTOFF(ebx,PD_1_847)]   ; xmm0=z5
283    mulps       xmm3, [GOTOFF(ebx,PD_M2_613)]  ; xmm3=(z10 * -2.613125930)
284    mulps       xmm4, [GOTOFF(ebx,PD_1_082)]   ; xmm4=(z12 * 1.082392200)
285    addps       xmm3, xmm0                     ; xmm3=tmp12
286    subps       xmm4, xmm0                     ; xmm4=tmp10
287
288    ; -- Final output stage
289
290    subps       xmm3, xmm1              ; xmm3=tmp6
291    movaps      xmm5, xmm6
292    movaps      xmm0, xmm7
293    addps       xmm6, xmm1              ; xmm6=data0=(00 01 02 03)
294    addps       xmm7, xmm3              ; xmm7=data1=(10 11 12 13)
295    subps       xmm5, xmm1              ; xmm5=data7=(70 71 72 73)
296    subps       xmm0, xmm3              ; xmm0=data6=(60 61 62 63)
297    subps       xmm2, xmm3              ; xmm2=tmp5
298
299    movaps      xmm1, xmm6              ; transpose coefficients(phase 1)
300    unpcklps    xmm6, xmm7              ; xmm6=(00 10 01 11)
301    unpckhps    xmm1, xmm7              ; xmm1=(02 12 03 13)
302    movaps      xmm3, xmm0              ; transpose coefficients(phase 1)
303    unpcklps    xmm0, xmm5              ; xmm0=(60 70 61 71)
304    unpckhps    xmm3, xmm5              ; xmm3=(62 72 63 73)
305
306    movaps      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
307    movaps      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp3
308
309    movaps      XMMWORD [wk(0)], xmm0   ; wk(0)=(60 70 61 71)
310    movaps      XMMWORD [wk(1)], xmm3   ; wk(1)=(62 72 63 73)
311
312    addps       xmm4, xmm2              ; xmm4=tmp4
313    movaps      xmm0, xmm7
314    movaps      xmm3, xmm5
315    addps       xmm7, xmm2              ; xmm7=data2=(20 21 22 23)
316    addps       xmm5, xmm4              ; xmm5=data4=(40 41 42 43)
317    subps       xmm0, xmm2              ; xmm0=data5=(50 51 52 53)
318    subps       xmm3, xmm4              ; xmm3=data3=(30 31 32 33)
319
320    movaps      xmm2, xmm7              ; transpose coefficients(phase 1)
321    unpcklps    xmm7, xmm3              ; xmm7=(20 30 21 31)
322    unpckhps    xmm2, xmm3              ; xmm2=(22 32 23 33)
323    movaps      xmm4, xmm5              ; transpose coefficients(phase 1)
324    unpcklps    xmm5, xmm0              ; xmm5=(40 50 41 51)
325    unpckhps    xmm4, xmm0              ; xmm4=(42 52 43 53)
326
327    movaps      xmm3, xmm6              ; transpose coefficients(phase 2)
328    unpcklps2   xmm6, xmm7              ; xmm6=(00 10 20 30)
329    unpckhps2   xmm3, xmm7              ; xmm3=(01 11 21 31)
330    movaps      xmm0, xmm1              ; transpose coefficients(phase 2)
331    unpcklps2   xmm1, xmm2              ; xmm1=(02 12 22 32)
332    unpckhps2   xmm0, xmm2              ; xmm0=(03 13 23 33)
333
334    movaps      xmm7, XMMWORD [wk(0)]   ; xmm7=(60 70 61 71)
335    movaps      xmm2, XMMWORD [wk(1)]   ; xmm2=(62 72 63 73)
336
337    movaps      XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm6
338    movaps      XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
339    movaps      XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm1
340    movaps      XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm0
341
342    movaps      xmm6, xmm5              ; transpose coefficients(phase 2)
343    unpcklps2   xmm5, xmm7              ; xmm5=(40 50 60 70)
344    unpckhps2   xmm6, xmm7              ; xmm6=(41 51 61 71)
345    movaps      xmm3, xmm4              ; transpose coefficients(phase 2)
346    unpcklps2   xmm4, xmm2              ; xmm4=(42 52 62 72)
347    unpckhps2   xmm3, xmm2              ; xmm3=(43 53 63 73)
348
349    movaps      XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm5
350    movaps      XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
351    movaps      XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm4
352    movaps      XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
353
354.nextcolumn:
355    add         esi, byte 4*SIZEOF_JCOEF               ; coef_block
356    add         edx, byte 4*SIZEOF_FLOAT_MULT_TYPE     ; quantptr
357    add         edi,      4*DCTSIZE*SIZEOF_FAST_FLOAT  ; wsptr
358    dec         ecx                                    ; ctr
359    jnz         near .columnloop
360
361    ; -- Prefetch the next coefficient block
362
363    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
364    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
365    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
366    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
367
368    ; ---- Pass 2: process rows from work array, store into output array.
369
370    mov         eax, [original_ebp]
371    lea         esi, [workspace]                   ; FAST_FLOAT *wsptr
372    mov         edi, JSAMPARRAY [output_buf(eax)]  ; (JSAMPROW *)
373    mov         eax, JDIMENSION [output_col(eax)]
374    mov         ecx, DCTSIZE/4                     ; ctr
375    alignx      16, 7
376.rowloop:
377
378    ; -- Even part
379
380    movaps      xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
381    movaps      xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
382    movaps      xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
383    movaps      xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
384
385    movaps      xmm4, xmm0
386    movaps      xmm5, xmm1
387    subps       xmm0, xmm2              ; xmm0=tmp11
388    subps       xmm1, xmm3
389    addps       xmm4, xmm2              ; xmm4=tmp10
390    addps       xmm5, xmm3              ; xmm5=tmp13
391
392    mulps       xmm1, [GOTOFF(ebx,PD_1_414)]
393    subps       xmm1, xmm5              ; xmm1=tmp12
394
395    movaps      xmm6, xmm4
396    movaps      xmm7, xmm0
397    subps       xmm4, xmm5              ; xmm4=tmp3
398    subps       xmm0, xmm1              ; xmm0=tmp2
399    addps       xmm6, xmm5              ; xmm6=tmp0
400    addps       xmm7, xmm1              ; xmm7=tmp1
401
402    movaps      XMMWORD [wk(1)], xmm4   ; tmp3
403    movaps      XMMWORD [wk(0)], xmm0   ; tmp2
404
405    ; -- Odd part
406
407    movaps      xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
408    movaps      xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
409    movaps      xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
410    movaps      xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
411
412    movaps      xmm4, xmm2
413    movaps      xmm0, xmm5
414    addps       xmm2, xmm1              ; xmm2=z11
415    addps       xmm5, xmm3              ; xmm5=z13
416    subps       xmm4, xmm1              ; xmm4=z12
417    subps       xmm0, xmm3              ; xmm0=z10
418
419    movaps      xmm1, xmm2
420    subps       xmm2, xmm5
421    addps       xmm1, xmm5              ; xmm1=tmp7
422
423    mulps       xmm2, [GOTOFF(ebx,PD_1_414)]  ; xmm2=tmp11
424
425    movaps      xmm3, xmm0
426    addps       xmm0, xmm4
427    mulps       xmm0, [GOTOFF(ebx,PD_1_847)]   ; xmm0=z5
428    mulps       xmm3, [GOTOFF(ebx,PD_M2_613)]  ; xmm3=(z10 * -2.613125930)
429    mulps       xmm4, [GOTOFF(ebx,PD_1_082)]   ; xmm4=(z12 * 1.082392200)
430    addps       xmm3, xmm0                     ; xmm3=tmp12
431    subps       xmm4, xmm0                     ; xmm4=tmp10
432
433    ; -- Final output stage
434
435    subps       xmm3, xmm1              ; xmm3=tmp6
436    movaps      xmm5, xmm6
437    movaps      xmm0, xmm7
438    addps       xmm6, xmm1              ; xmm6=data0=(00 10 20 30)
439    addps       xmm7, xmm3              ; xmm7=data1=(01 11 21 31)
440    subps       xmm5, xmm1              ; xmm5=data7=(07 17 27 37)
441    subps       xmm0, xmm3              ; xmm0=data6=(06 16 26 36)
442    subps       xmm2, xmm3              ; xmm2=tmp5
443
444    movaps      xmm1, [GOTOFF(ebx,PD_0_125)]  ; xmm1=[PD_0_125]
445
446    mulps       xmm6, xmm1              ; descale(1/8)
447    mulps       xmm7, xmm1              ; descale(1/8)
448    mulps       xmm5, xmm1              ; descale(1/8)
449    mulps       xmm0, xmm1              ; descale(1/8)
450
451    movhlps     xmm3, xmm6
452    movhlps     xmm1, xmm7
453    cvtps2pi    mm0, xmm6               ; round to int32, mm0=data0L=(00 10)
454    cvtps2pi    mm1, xmm7               ; round to int32, mm1=data1L=(01 11)
455    cvtps2pi    mm2, xmm3               ; round to int32, mm2=data0H=(20 30)
456    cvtps2pi    mm3, xmm1               ; round to int32, mm3=data1H=(21 31)
457    packssdw    mm0, mm2                ; mm0=data0=(00 10 20 30)
458    packssdw    mm1, mm3                ; mm1=data1=(01 11 21 31)
459
460    movhlps     xmm6, xmm5
461    movhlps     xmm7, xmm0
462    cvtps2pi    mm4, xmm5               ; round to int32, mm4=data7L=(07 17)
463    cvtps2pi    mm5, xmm0               ; round to int32, mm5=data6L=(06 16)
464    cvtps2pi    mm6, xmm6               ; round to int32, mm6=data7H=(27 37)
465    cvtps2pi    mm7, xmm7               ; round to int32, mm7=data6H=(26 36)
466    packssdw    mm4, mm6                ; mm4=data7=(07 17 27 37)
467    packssdw    mm5, mm7                ; mm5=data6=(06 16 26 36)
468
469    packsswb    mm0, mm5                ; mm0=(00 10 20 30 06 16 26 36)
470    packsswb    mm1, mm4                ; mm1=(01 11 21 31 07 17 27 37)
471
472    movaps      xmm3, XMMWORD [wk(0)]   ; xmm3=tmp2
473    movaps      xmm1, XMMWORD [wk(1)]   ; xmm1=tmp3
474
475    movaps      xmm6, [GOTOFF(ebx,PD_0_125)]  ; xmm6=[PD_0_125]
476
477    addps       xmm4, xmm2              ; xmm4=tmp4
478    movaps      xmm5, xmm3
479    movaps      xmm0, xmm1
480    addps       xmm3, xmm2              ; xmm3=data2=(02 12 22 32)
481    addps       xmm1, xmm4              ; xmm1=data4=(04 14 24 34)
482    subps       xmm5, xmm2              ; xmm5=data5=(05 15 25 35)
483    subps       xmm0, xmm4              ; xmm0=data3=(03 13 23 33)
484
485    mulps       xmm3, xmm6              ; descale(1/8)
486    mulps       xmm1, xmm6              ; descale(1/8)
487    mulps       xmm5, xmm6              ; descale(1/8)
488    mulps       xmm0, xmm6              ; descale(1/8)
489
490    movhlps     xmm7, xmm3
491    movhlps     xmm2, xmm1
492    cvtps2pi    mm2, xmm3               ; round to int32, mm2=data2L=(02 12)
493    cvtps2pi    mm3, xmm1               ; round to int32, mm3=data4L=(04 14)
494    cvtps2pi    mm6, xmm7               ; round to int32, mm6=data2H=(22 32)
495    cvtps2pi    mm7, xmm2               ; round to int32, mm7=data4H=(24 34)
496    packssdw    mm2, mm6                ; mm2=data2=(02 12 22 32)
497    packssdw    mm3, mm7                ; mm3=data4=(04 14 24 34)
498
499    movhlps     xmm4, xmm5
500    movhlps     xmm6, xmm0
501    cvtps2pi    mm5, xmm5               ; round to int32, mm5=data5L=(05 15)
502    cvtps2pi    mm4, xmm0               ; round to int32, mm4=data3L=(03 13)
503    cvtps2pi    mm6, xmm4               ; round to int32, mm6=data5H=(25 35)
504    cvtps2pi    mm7, xmm6               ; round to int32, mm7=data3H=(23 33)
505    packssdw    mm5, mm6                ; mm5=data5=(05 15 25 35)
506    packssdw    mm4, mm7                ; mm4=data3=(03 13 23 33)
507
508    movq        mm6, [GOTOFF(ebx,PB_CENTERJSAMP)]  ; mm6=[PB_CENTERJSAMP]
509
510    packsswb    mm2, mm3                ; mm2=(02 12 22 32 04 14 24 34)
511    packsswb    mm4, mm5                ; mm4=(03 13 23 33 05 15 25 35)
512
513    paddb       mm0, mm6
514    paddb       mm1, mm6
515    paddb       mm2, mm6
516    paddb       mm4, mm6
517
518    movq        mm7, mm0                ; transpose coefficients(phase 1)
519    punpcklbw   mm0, mm1                ; mm0=(00 01 10 11 20 21 30 31)
520    punpckhbw   mm7, mm1                ; mm7=(06 07 16 17 26 27 36 37)
521    movq        mm3, mm2                ; transpose coefficients(phase 1)
522    punpcklbw   mm2, mm4                ; mm2=(02 03 12 13 22 23 32 33)
523    punpckhbw   mm3, mm4                ; mm3=(04 05 14 15 24 25 34 35)
524
525    movq        mm5, mm0                ; transpose coefficients(phase 2)
526    punpcklwd   mm0, mm2                ; mm0=(00 01 02 03 10 11 12 13)
527    punpckhwd   mm5, mm2                ; mm5=(20 21 22 23 30 31 32 33)
528    movq        mm6, mm3                ; transpose coefficients(phase 2)
529    punpcklwd   mm3, mm7                ; mm3=(04 05 06 07 14 15 16 17)
530    punpckhwd   mm6, mm7                ; mm6=(24 25 26 27 34 35 36 37)
531
532    movq        mm1, mm0                ; transpose coefficients(phase 3)
533    punpckldq   mm0, mm3                ; mm0=(00 01 02 03 04 05 06 07)
534    punpckhdq   mm1, mm3                ; mm1=(10 11 12 13 14 15 16 17)
535    movq        mm4, mm5                ; transpose coefficients(phase 3)
536    punpckldq   mm5, mm6                ; mm5=(20 21 22 23 24 25 26 27)
537    punpckhdq   mm4, mm6                ; mm4=(30 31 32 33 34 35 36 37)
538
539    pushpic     ebx                     ; save GOT address
540
541    mov         edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
542    mov         ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
543    movq        MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0
544    movq        MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm1
545    mov         edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
546    mov         ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
547    movq        MMWORD [edx+eax*SIZEOF_JSAMPLE], mm5
548    movq        MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4
549
550    poppic      ebx                     ; restore GOT address
551
552    add         esi, byte 4*SIZEOF_FAST_FLOAT  ; wsptr
553    add         edi, byte 4*SIZEOF_JSAMPROW
554    dec         ecx                            ; ctr
555    jnz         near .rowloop
556
557    emms                                ; empty MMX state
558
559    pop         edi
560    pop         esi
561;   pop         edx                     ; need not be preserved
562;   pop         ecx                     ; need not be preserved
563    pop         ebx
564    mov         esp, ebp                ; esp <- aligned ebp
565    pop         esp                     ; esp <- original ebp
566    pop         ebp
567    ret
568
569; For some reason, the OS X linker does not honor the request to align the
570; segment unless we do this.
571    align       32
572