1;
2; jidctflt.asm - floating-point IDCT (SSE & MMX)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5;
6; Based on
7; 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; [TAB8]
22
23%include "jsimdext.inc"
24%include "jdct.inc"
25
26; --------------------------------------------------------------------------
27
28%macro  unpcklps2 2     ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
29        shufps  %1,%2,0x44
30%endmacro
31
32%macro  unpckhps2 2     ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
33        shufps  %1,%2,0xEE
34%endmacro
35
36; --------------------------------------------------------------------------
37        SECTION SEG_CONST
38
39        alignz  16
40        global  EXTN(jconst_idct_float_sse)
41
42EXTN(jconst_idct_float_sse):
43
44PD_1_414        times 4 dd  1.414213562373095048801689
45PD_1_847        times 4 dd  1.847759065022573512256366
46PD_1_082        times 4 dd  1.082392200292393968799446
47PD_M2_613       times 4 dd -2.613125929752753055713286
48PD_0_125        times 4 dd  0.125       ; 1/8
49PB_CENTERJSAMP  times 8 db  CENTERJSAMPLE
50
51        alignz  16
52
53; --------------------------------------------------------------------------
54        SECTION SEG_TEXT
55        BITS    32
56;
57; Perform dequantization and inverse DCT on one block of coefficients.
58;
59; GLOBAL(void)
60; jsimd_idct_float_sse (void * dct_table, JCOEFPTR coef_block,
61;                       JSAMPARRAY output_buf, JDIMENSION output_col)
62;
63
64%define dct_table(b)    (b)+8           ; void * dct_table
65%define coef_block(b)   (b)+12          ; JCOEFPTR coef_block
66%define output_buf(b)   (b)+16          ; JSAMPARRAY output_buf
67%define output_col(b)   (b)+20          ; JDIMENSION output_col
68
69%define original_ebp    ebp+0
70%define wk(i)           ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
71%define WK_NUM          2
72%define workspace       wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT
73                                        ; FAST_FLOAT workspace[DCTSIZE2]
74
75        align   16
76        global  EXTN(jsimd_idct_float_sse)
77
78EXTN(jsimd_idct_float_sse):
79        push    ebp
80        mov     eax,esp                         ; eax = original ebp
81        sub     esp, byte 4
82        and     esp, byte (-SIZEOF_XMMWORD)     ; align to 128 bits
83        mov     [esp],eax
84        mov     ebp,esp                         ; ebp = aligned ebp
85        lea     esp, [workspace]
86        push    ebx
87;       push    ecx             ; need not be preserved
88;       push    edx             ; need not be preserved
89        push    esi
90        push    edi
91
92        get_GOT ebx             ; get GOT address
93
94        ; ---- Pass 1: process columns from input, store into work array.
95
96;       mov     eax, [original_ebp]
97        mov     edx, POINTER [dct_table(eax)]           ; quantptr
98        mov     esi, JCOEFPTR [coef_block(eax)]         ; inptr
99        lea     edi, [workspace]                        ; FAST_FLOAT * wsptr
100        mov     ecx, DCTSIZE/4                          ; ctr
101        alignx  16,7
102.columnloop:
103%ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
104        mov     eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
105        or      eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
106        jnz     near .columnDCT
107
108        movq    mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
109        movq    mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
110        por     mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
111        por     mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
112        por     mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
113        por     mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
114        por     mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
115        por     mm1,mm0
116        packsswb mm1,mm1
117        movd    eax,mm1
118        test    eax,eax
119        jnz     short .columnDCT
120
121        ; -- AC terms all zero
122
123        movq      mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
124
125        punpckhwd mm1,mm0                       ; mm1=(** 02 ** 03)
126        punpcklwd mm0,mm0                       ; mm0=(00 00 01 01)
127        psrad     mm1,(DWORD_BIT-WORD_BIT)      ; mm1=in0H=(02 03)
128        psrad     mm0,(DWORD_BIT-WORD_BIT)      ; mm0=in0L=(00 01)
129        cvtpi2ps  xmm3,mm1                      ; xmm3=(02 03 ** **)
130        cvtpi2ps  xmm0,mm0                      ; xmm0=(00 01 ** **)
131        movlhps   xmm0,xmm3                     ; xmm0=in0=(00 01 02 03)
132
133        mulps   xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
134
135        movaps  xmm1,xmm0
136        movaps  xmm2,xmm0
137        movaps  xmm3,xmm0
138
139        shufps  xmm0,xmm0,0x00                  ; xmm0=(00 00 00 00)
140        shufps  xmm1,xmm1,0x55                  ; xmm1=(01 01 01 01)
141        shufps  xmm2,xmm2,0xAA                  ; xmm2=(02 02 02 02)
142        shufps  xmm3,xmm3,0xFF                  ; xmm3=(03 03 03 03)
143
144        movaps  XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
145        movaps  XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
146        movaps  XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm1
147        movaps  XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
148        movaps  XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm2
149        movaps  XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm2
150        movaps  XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm3
151        movaps  XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
152        jmp     near .nextcolumn
153        alignx  16,7
154%endif
155.columnDCT:
156
157        ; -- Even part
158
159        movq      mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
160        movq      mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
161        movq      mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
162        movq      mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
163
164        punpckhwd mm4,mm0                       ; mm4=(** 02 ** 03)
165        punpcklwd mm0,mm0                       ; mm0=(00 00 01 01)
166        punpckhwd mm5,mm1                       ; mm5=(** 22 ** 23)
167        punpcklwd mm1,mm1                       ; mm1=(20 20 21 21)
168
169        psrad     mm4,(DWORD_BIT-WORD_BIT)      ; mm4=in0H=(02 03)
170        psrad     mm0,(DWORD_BIT-WORD_BIT)      ; mm0=in0L=(00 01)
171        cvtpi2ps  xmm4,mm4                      ; xmm4=(02 03 ** **)
172        cvtpi2ps  xmm0,mm0                      ; xmm0=(00 01 ** **)
173        psrad     mm5,(DWORD_BIT-WORD_BIT)      ; mm5=in2H=(22 23)
174        psrad     mm1,(DWORD_BIT-WORD_BIT)      ; mm1=in2L=(20 21)
175        cvtpi2ps  xmm5,mm5                      ; xmm5=(22 23 ** **)
176        cvtpi2ps  xmm1,mm1                      ; xmm1=(20 21 ** **)
177
178        punpckhwd mm6,mm2                       ; mm6=(** 42 ** 43)
179        punpcklwd mm2,mm2                       ; mm2=(40 40 41 41)
180        punpckhwd mm7,mm3                       ; mm7=(** 62 ** 63)
181        punpcklwd mm3,mm3                       ; mm3=(60 60 61 61)
182
183        psrad     mm6,(DWORD_BIT-WORD_BIT)      ; mm6=in4H=(42 43)
184        psrad     mm2,(DWORD_BIT-WORD_BIT)      ; mm2=in4L=(40 41)
185        cvtpi2ps  xmm6,mm6                      ; xmm6=(42 43 ** **)
186        cvtpi2ps  xmm2,mm2                      ; xmm2=(40 41 ** **)
187        psrad     mm7,(DWORD_BIT-WORD_BIT)      ; mm7=in6H=(62 63)
188        psrad     mm3,(DWORD_BIT-WORD_BIT)      ; mm3=in6L=(60 61)
189        cvtpi2ps  xmm7,mm7                      ; xmm7=(62 63 ** **)
190        cvtpi2ps  xmm3,mm3                      ; xmm3=(60 61 ** **)
191
192        movlhps   xmm0,xmm4                     ; xmm0=in0=(00 01 02 03)
193        movlhps   xmm1,xmm5                     ; xmm1=in2=(20 21 22 23)
194        mulps     xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
195        mulps     xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
196
197        movlhps   xmm2,xmm6                     ; xmm2=in4=(40 41 42 43)
198        movlhps   xmm3,xmm7                     ; xmm3=in6=(60 61 62 63)
199        mulps     xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
200        mulps     xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
201
202        movaps  xmm4,xmm0
203        movaps  xmm5,xmm1
204        subps   xmm0,xmm2               ; xmm0=tmp11
205        subps   xmm1,xmm3
206        addps   xmm4,xmm2               ; xmm4=tmp10
207        addps   xmm5,xmm3               ; xmm5=tmp13
208
209        mulps   xmm1,[GOTOFF(ebx,PD_1_414)]
210        subps   xmm1,xmm5               ; xmm1=tmp12
211
212        movaps  xmm6,xmm4
213        movaps  xmm7,xmm0
214        subps   xmm4,xmm5               ; xmm4=tmp3
215        subps   xmm0,xmm1               ; xmm0=tmp2
216        addps   xmm6,xmm5               ; xmm6=tmp0
217        addps   xmm7,xmm1               ; xmm7=tmp1
218
219        movaps  XMMWORD [wk(1)], xmm4   ; tmp3
220        movaps  XMMWORD [wk(0)], xmm0   ; tmp2
221
222        ; -- Odd part
223
224        movq      mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
225        movq      mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
226        movq      mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
227        movq      mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
228
229        punpckhwd mm6,mm4                       ; mm6=(** 12 ** 13)
230        punpcklwd mm4,mm4                       ; mm4=(10 10 11 11)
231        punpckhwd mm2,mm0                       ; mm2=(** 32 ** 33)
232        punpcklwd mm0,mm0                       ; mm0=(30 30 31 31)
233
234        psrad     mm6,(DWORD_BIT-WORD_BIT)      ; mm6=in1H=(12 13)
235        psrad     mm4,(DWORD_BIT-WORD_BIT)      ; mm4=in1L=(10 11)
236        cvtpi2ps  xmm4,mm6                      ; xmm4=(12 13 ** **)
237        cvtpi2ps  xmm2,mm4                      ; xmm2=(10 11 ** **)
238        psrad     mm2,(DWORD_BIT-WORD_BIT)      ; mm2=in3H=(32 33)
239        psrad     mm0,(DWORD_BIT-WORD_BIT)      ; mm0=in3L=(30 31)
240        cvtpi2ps  xmm0,mm2                      ; xmm0=(32 33 ** **)
241        cvtpi2ps  xmm3,mm0                      ; xmm3=(30 31 ** **)
242
243        punpckhwd mm7,mm5                       ; mm7=(** 52 ** 53)
244        punpcklwd mm5,mm5                       ; mm5=(50 50 51 51)
245        punpckhwd mm3,mm1                       ; mm3=(** 72 ** 73)
246        punpcklwd mm1,mm1                       ; mm1=(70 70 71 71)
247
248        movlhps   xmm2,xmm4                     ; xmm2=in1=(10 11 12 13)
249        movlhps   xmm3,xmm0                     ; xmm3=in3=(30 31 32 33)
250
251        psrad     mm7,(DWORD_BIT-WORD_BIT)      ; mm7=in5H=(52 53)
252        psrad     mm5,(DWORD_BIT-WORD_BIT)      ; mm5=in5L=(50 51)
253        cvtpi2ps  xmm4,mm7                      ; xmm4=(52 53 ** **)
254        cvtpi2ps  xmm5,mm5                      ; xmm5=(50 51 ** **)
255        psrad     mm3,(DWORD_BIT-WORD_BIT)      ; mm3=in7H=(72 73)
256        psrad     mm1,(DWORD_BIT-WORD_BIT)      ; mm1=in7L=(70 71)
257        cvtpi2ps  xmm0,mm3                      ; xmm0=(72 73 ** **)
258        cvtpi2ps  xmm1,mm1                      ; xmm1=(70 71 ** **)
259
260        mulps     xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
261        mulps     xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
262
263        movlhps   xmm5,xmm4                     ; xmm5=in5=(50 51 52 53)
264        movlhps   xmm1,xmm0                     ; xmm1=in7=(70 71 72 73)
265        mulps     xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
266        mulps     xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
267
268        movaps  xmm4,xmm2
269        movaps  xmm0,xmm5
270        addps   xmm2,xmm1               ; xmm2=z11
271        addps   xmm5,xmm3               ; xmm5=z13
272        subps   xmm4,xmm1               ; xmm4=z12
273        subps   xmm0,xmm3               ; xmm0=z10
274
275        movaps  xmm1,xmm2
276        subps   xmm2,xmm5
277        addps   xmm1,xmm5               ; xmm1=tmp7
278
279        mulps   xmm2,[GOTOFF(ebx,PD_1_414)]     ; xmm2=tmp11
280
281        movaps  xmm3,xmm0
282        addps   xmm0,xmm4
283        mulps   xmm0,[GOTOFF(ebx,PD_1_847)]     ; xmm0=z5
284        mulps   xmm3,[GOTOFF(ebx,PD_M2_613)]    ; xmm3=(z10 * -2.613125930)
285        mulps   xmm4,[GOTOFF(ebx,PD_1_082)]     ; xmm4=(z12 * 1.082392200)
286        addps   xmm3,xmm0               ; xmm3=tmp12
287        subps   xmm4,xmm0               ; xmm4=tmp10
288
289        ; -- Final output stage
290
291        subps   xmm3,xmm1               ; xmm3=tmp6
292        movaps  xmm5,xmm6
293        movaps  xmm0,xmm7
294        addps   xmm6,xmm1               ; xmm6=data0=(00 01 02 03)
295        addps   xmm7,xmm3               ; xmm7=data1=(10 11 12 13)
296        subps   xmm5,xmm1               ; xmm5=data7=(70 71 72 73)
297        subps   xmm0,xmm3               ; xmm0=data6=(60 61 62 63)
298        subps   xmm2,xmm3               ; xmm2=tmp5
299
300        movaps    xmm1,xmm6             ; transpose coefficients(phase 1)
301        unpcklps  xmm6,xmm7             ; xmm6=(00 10 01 11)
302        unpckhps  xmm1,xmm7             ; xmm1=(02 12 03 13)
303        movaps    xmm3,xmm0             ; transpose coefficients(phase 1)
304        unpcklps  xmm0,xmm5             ; xmm0=(60 70 61 71)
305        unpckhps  xmm3,xmm5             ; xmm3=(62 72 63 73)
306
307        movaps  xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
308        movaps  xmm5, XMMWORD [wk(1)]   ; xmm5=tmp3
309
310        movaps  XMMWORD [wk(0)], xmm0   ; wk(0)=(60 70 61 71)
311        movaps  XMMWORD [wk(1)], xmm3   ; wk(1)=(62 72 63 73)
312
313        addps   xmm4,xmm2               ; xmm4=tmp4
314        movaps  xmm0,xmm7
315        movaps  xmm3,xmm5
316        addps   xmm7,xmm2               ; xmm7=data2=(20 21 22 23)
317        addps   xmm5,xmm4               ; xmm5=data4=(40 41 42 43)
318        subps   xmm0,xmm2               ; xmm0=data5=(50 51 52 53)
319        subps   xmm3,xmm4               ; xmm3=data3=(30 31 32 33)
320
321        movaps    xmm2,xmm7             ; transpose coefficients(phase 1)
322        unpcklps  xmm7,xmm3             ; xmm7=(20 30 21 31)
323        unpckhps  xmm2,xmm3             ; xmm2=(22 32 23 33)
324        movaps    xmm4,xmm5             ; transpose coefficients(phase 1)
325        unpcklps  xmm5,xmm0             ; xmm5=(40 50 41 51)
326        unpckhps  xmm4,xmm0             ; xmm4=(42 52 43 53)
327
328        movaps    xmm3,xmm6             ; transpose coefficients(phase 2)
329        unpcklps2 xmm6,xmm7             ; xmm6=(00 10 20 30)
330        unpckhps2 xmm3,xmm7             ; xmm3=(01 11 21 31)
331        movaps    xmm0,xmm1             ; transpose coefficients(phase 2)
332        unpcklps2 xmm1,xmm2             ; xmm1=(02 12 22 32)
333        unpckhps2 xmm0,xmm2             ; xmm0=(03 13 23 33)
334
335        movaps  xmm7, XMMWORD [wk(0)]   ; xmm7=(60 70 61 71)
336        movaps  xmm2, XMMWORD [wk(1)]   ; xmm2=(62 72 63 73)
337
338        movaps  XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm6
339        movaps  XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
340        movaps  XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm1
341        movaps  XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm0
342
343        movaps    xmm6,xmm5             ; transpose coefficients(phase 2)
344        unpcklps2 xmm5,xmm7             ; xmm5=(40 50 60 70)
345        unpckhps2 xmm6,xmm7             ; xmm6=(41 51 61 71)
346        movaps    xmm3,xmm4             ; transpose coefficients(phase 2)
347        unpcklps2 xmm4,xmm2             ; xmm4=(42 52 62 72)
348        unpckhps2 xmm3,xmm2             ; xmm3=(43 53 63 73)
349
350        movaps  XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm5
351        movaps  XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
352        movaps  XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm4
353        movaps  XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
354
355.nextcolumn:
356        add     esi, byte 4*SIZEOF_JCOEF                ; coef_block
357        add     edx, byte 4*SIZEOF_FLOAT_MULT_TYPE      ; quantptr
358        add     edi,      4*DCTSIZE*SIZEOF_FAST_FLOAT   ; wsptr
359        dec     ecx                                     ; ctr
360        jnz     near .columnloop
361
362        ; -- Prefetch the next coefficient block
363
364        prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
365        prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
366        prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
367        prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
368
369        ; ---- Pass 2: process rows from work array, store into output array.
370
371        mov     eax, [original_ebp]
372        lea     esi, [workspace]                        ; FAST_FLOAT * wsptr
373        mov     edi, JSAMPARRAY [output_buf(eax)]       ; (JSAMPROW *)
374        mov     eax, JDIMENSION [output_col(eax)]
375        mov     ecx, DCTSIZE/4                          ; ctr
376        alignx  16,7
377.rowloop:
378
379        ; -- Even part
380
381        movaps  xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
382        movaps  xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
383        movaps  xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
384        movaps  xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
385
386        movaps  xmm4,xmm0
387        movaps  xmm5,xmm1
388        subps   xmm0,xmm2               ; xmm0=tmp11
389        subps   xmm1,xmm3
390        addps   xmm4,xmm2               ; xmm4=tmp10
391        addps   xmm5,xmm3               ; xmm5=tmp13
392
393        mulps   xmm1,[GOTOFF(ebx,PD_1_414)]
394        subps   xmm1,xmm5               ; xmm1=tmp12
395
396        movaps  xmm6,xmm4
397        movaps  xmm7,xmm0
398        subps   xmm4,xmm5               ; xmm4=tmp3
399        subps   xmm0,xmm1               ; xmm0=tmp2
400        addps   xmm6,xmm5               ; xmm6=tmp0
401        addps   xmm7,xmm1               ; xmm7=tmp1
402
403        movaps  XMMWORD [wk(1)], xmm4   ; tmp3
404        movaps  XMMWORD [wk(0)], xmm0   ; tmp2
405
406        ; -- Odd part
407
408        movaps  xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
409        movaps  xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
410        movaps  xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
411        movaps  xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
412
413        movaps  xmm4,xmm2
414        movaps  xmm0,xmm5
415        addps   xmm2,xmm1               ; xmm2=z11
416        addps   xmm5,xmm3               ; xmm5=z13
417        subps   xmm4,xmm1               ; xmm4=z12
418        subps   xmm0,xmm3               ; xmm0=z10
419
420        movaps  xmm1,xmm2
421        subps   xmm2,xmm5
422        addps   xmm1,xmm5               ; xmm1=tmp7
423
424        mulps   xmm2,[GOTOFF(ebx,PD_1_414)]     ; xmm2=tmp11
425
426        movaps  xmm3,xmm0
427        addps   xmm0,xmm4
428        mulps   xmm0,[GOTOFF(ebx,PD_1_847)]     ; xmm0=z5
429        mulps   xmm3,[GOTOFF(ebx,PD_M2_613)]    ; xmm3=(z10 * -2.613125930)
430        mulps   xmm4,[GOTOFF(ebx,PD_1_082)]     ; xmm4=(z12 * 1.082392200)
431        addps   xmm3,xmm0               ; xmm3=tmp12
432        subps   xmm4,xmm0               ; xmm4=tmp10
433
434        ; -- Final output stage
435
436        subps   xmm3,xmm1               ; xmm3=tmp6
437        movaps  xmm5,xmm6
438        movaps  xmm0,xmm7
439        addps   xmm6,xmm1               ; xmm6=data0=(00 10 20 30)
440        addps   xmm7,xmm3               ; xmm7=data1=(01 11 21 31)
441        subps   xmm5,xmm1               ; xmm5=data7=(07 17 27 37)
442        subps   xmm0,xmm3               ; xmm0=data6=(06 16 26 36)
443        subps   xmm2,xmm3               ; xmm2=tmp5
444
445        movaps  xmm1,[GOTOFF(ebx,PD_0_125)]     ; xmm1=[PD_0_125]
446
447        mulps   xmm6,xmm1               ; descale(1/8)
448        mulps   xmm7,xmm1               ; descale(1/8)
449        mulps   xmm5,xmm1               ; descale(1/8)
450        mulps   xmm0,xmm1               ; descale(1/8)
451
452        movhlps   xmm3,xmm6
453        movhlps   xmm1,xmm7
454        cvtps2pi  mm0,xmm6              ; round to int32, mm0=data0L=(00 10)
455        cvtps2pi  mm1,xmm7              ; round to int32, mm1=data1L=(01 11)
456        cvtps2pi  mm2,xmm3              ; round to int32, mm2=data0H=(20 30)
457        cvtps2pi  mm3,xmm1              ; round to int32, mm3=data1H=(21 31)
458        packssdw  mm0,mm2               ; mm0=data0=(00 10 20 30)
459        packssdw  mm1,mm3               ; mm1=data1=(01 11 21 31)
460
461        movhlps   xmm6,xmm5
462        movhlps   xmm7,xmm0
463        cvtps2pi  mm4,xmm5              ; round to int32, mm4=data7L=(07 17)
464        cvtps2pi  mm5,xmm0              ; round to int32, mm5=data6L=(06 16)
465        cvtps2pi  mm6,xmm6              ; round to int32, mm6=data7H=(27 37)
466        cvtps2pi  mm7,xmm7              ; round to int32, mm7=data6H=(26 36)
467        packssdw  mm4,mm6               ; mm4=data7=(07 17 27 37)
468        packssdw  mm5,mm7               ; mm5=data6=(06 16 26 36)
469
470        packsswb  mm0,mm5               ; mm0=(00 10 20 30 06 16 26 36)
471        packsswb  mm1,mm4               ; mm1=(01 11 21 31 07 17 27 37)
472
473        movaps  xmm3, XMMWORD [wk(0)]   ; xmm3=tmp2
474        movaps  xmm1, XMMWORD [wk(1)]   ; xmm1=tmp3
475
476        movaps  xmm6,[GOTOFF(ebx,PD_0_125)]     ; xmm6=[PD_0_125]
477
478        addps   xmm4,xmm2               ; xmm4=tmp4
479        movaps  xmm5,xmm3
480        movaps  xmm0,xmm1
481        addps   xmm3,xmm2               ; xmm3=data2=(02 12 22 32)
482        addps   xmm1,xmm4               ; xmm1=data4=(04 14 24 34)
483        subps   xmm5,xmm2               ; xmm5=data5=(05 15 25 35)
484        subps   xmm0,xmm4               ; xmm0=data3=(03 13 23 33)
485
486        mulps   xmm3,xmm6               ; descale(1/8)
487        mulps   xmm1,xmm6               ; descale(1/8)
488        mulps   xmm5,xmm6               ; descale(1/8)
489        mulps   xmm0,xmm6               ; descale(1/8)
490
491        movhlps   xmm7,xmm3
492        movhlps   xmm2,xmm1
493        cvtps2pi  mm2,xmm3              ; round to int32, mm2=data2L=(02 12)
494        cvtps2pi  mm3,xmm1              ; round to int32, mm3=data4L=(04 14)
495        cvtps2pi  mm6,xmm7              ; round to int32, mm6=data2H=(22 32)
496        cvtps2pi  mm7,xmm2              ; round to int32, mm7=data4H=(24 34)
497        packssdw  mm2,mm6               ; mm2=data2=(02 12 22 32)
498        packssdw  mm3,mm7               ; mm3=data4=(04 14 24 34)
499
500        movhlps   xmm4,xmm5
501        movhlps   xmm6,xmm0
502        cvtps2pi  mm5,xmm5              ; round to int32, mm5=data5L=(05 15)
503        cvtps2pi  mm4,xmm0              ; round to int32, mm4=data3L=(03 13)
504        cvtps2pi  mm6,xmm4              ; round to int32, mm6=data5H=(25 35)
505        cvtps2pi  mm7,xmm6              ; round to int32, mm7=data3H=(23 33)
506        packssdw  mm5,mm6               ; mm5=data5=(05 15 25 35)
507        packssdw  mm4,mm7               ; mm4=data3=(03 13 23 33)
508
509        movq      mm6,[GOTOFF(ebx,PB_CENTERJSAMP)]      ; mm6=[PB_CENTERJSAMP]
510
511        packsswb  mm2,mm3               ; mm2=(02 12 22 32 04 14 24 34)
512        packsswb  mm4,mm5               ; mm4=(03 13 23 33 05 15 25 35)
513
514        paddb     mm0,mm6
515        paddb     mm1,mm6
516        paddb     mm2,mm6
517        paddb     mm4,mm6
518
519        movq      mm7,mm0               ; transpose coefficients(phase 1)
520        punpcklbw mm0,mm1               ; mm0=(00 01 10 11 20 21 30 31)
521        punpckhbw mm7,mm1               ; mm7=(06 07 16 17 26 27 36 37)
522        movq      mm3,mm2               ; transpose coefficients(phase 1)
523        punpcklbw mm2,mm4               ; mm2=(02 03 12 13 22 23 32 33)
524        punpckhbw mm3,mm4               ; mm3=(04 05 14 15 24 25 34 35)
525
526        movq      mm5,mm0               ; transpose coefficients(phase 2)
527        punpcklwd mm0,mm2               ; mm0=(00 01 02 03 10 11 12 13)
528        punpckhwd mm5,mm2               ; mm5=(20 21 22 23 30 31 32 33)
529        movq      mm6,mm3               ; transpose coefficients(phase 2)
530        punpcklwd mm3,mm7               ; mm3=(04 05 06 07 14 15 16 17)
531        punpckhwd mm6,mm7               ; mm6=(24 25 26 27 34 35 36 37)
532
533        movq      mm1,mm0               ; transpose coefficients(phase 3)
534        punpckldq mm0,mm3               ; mm0=(00 01 02 03 04 05 06 07)
535        punpckhdq mm1,mm3               ; mm1=(10 11 12 13 14 15 16 17)
536        movq      mm4,mm5               ; transpose coefficients(phase 3)
537        punpckldq mm5,mm6               ; mm5=(20 21 22 23 24 25 26 27)
538        punpckhdq mm4,mm6               ; mm4=(30 31 32 33 34 35 36 37)
539
540        pushpic ebx                     ; save GOT address
541
542        mov     edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
543        mov     ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
544        movq    MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0
545        movq    MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm1
546        mov     edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
547        mov     ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
548        movq    MMWORD [edx+eax*SIZEOF_JSAMPLE], mm5
549        movq    MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4
550
551        poppic  ebx                     ; restore GOT address
552
553        add     esi, byte 4*SIZEOF_FAST_FLOAT   ; wsptr
554        add     edi, byte 4*SIZEOF_JSAMPROW
555        dec     ecx                             ; ctr
556        jnz     near .rowloop
557
558        emms            ; empty MMX state
559
560        pop     edi
561        pop     esi
562;       pop     edx             ; need not be preserved
563;       pop     ecx             ; need not be preserved
564        pop     ebx
565        mov     esp,ebp         ; esp <- aligned ebp
566        pop     esp             ; esp <- original ebp
567        pop     ebp
568        ret
569
570; For some reason, the OS X linker does not honor the request to align the
571; segment unless we do this.
572        align   16
573