1;
2; jdmrgext.asm - merged upsampling/color conversion (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; [TAB8]
18
19%include "jcolsamp.inc"
20
21; --------------------------------------------------------------------------
22;
23; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
24;
25; GLOBAL(void)
26; jsimd_h2v1_merged_upsample_mmx (JDIMENSION output_width,
27;                                 JSAMPIMAGE input_buf,
28;                                 JDIMENSION in_row_group_ctr,
29;                                 JSAMPARRAY output_buf);
30;
31
32%define output_width(b) (b)+8                   ; JDIMENSION output_width
33%define input_buf(b)            (b)+12          ; JSAMPIMAGE input_buf
34%define in_row_group_ctr(b)     (b)+16          ; JDIMENSION in_row_group_ctr
35%define output_buf(b)           (b)+20          ; JSAMPARRAY output_buf
36
37%define original_ebp    ebp+0
38%define wk(i)           ebp-(WK_NUM-(i))*SIZEOF_MMWORD  ; mmword wk[WK_NUM]
39%define WK_NUM          3
40%define gotptr          wk(0)-SIZEOF_POINTER    ; void * gotptr
41
42        align   16
43        global  EXTN(jsimd_h2v1_merged_upsample_mmx)
44
45EXTN(jsimd_h2v1_merged_upsample_mmx):
46        push    ebp
47        mov     eax,esp                         ; eax = original ebp
48        sub     esp, byte 4
49        and     esp, byte (-SIZEOF_MMWORD)      ; align to 64 bits
50        mov     [esp],eax
51        mov     ebp,esp                         ; ebp = aligned ebp
52        lea     esp, [wk(0)]
53        pushpic eax             ; make a room for GOT address
54        push    ebx
55;       push    ecx             ; need not be preserved
56;       push    edx             ; need not be preserved
57        push    esi
58        push    edi
59
60        get_GOT ebx                     ; get GOT address
61        movpic  POINTER [gotptr], ebx   ; save GOT address
62
63        mov     ecx, JDIMENSION [output_width(eax)]     ; col
64        test    ecx,ecx
65        jz      near .return
66
67        push    ecx
68
69        mov     edi, JSAMPIMAGE [input_buf(eax)]
70        mov     ecx, JDIMENSION [in_row_group_ctr(eax)]
71        mov     esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
72        mov     ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
73        mov     edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
74        mov     edi, JSAMPARRAY [output_buf(eax)]
75        mov     esi, JSAMPROW [esi+ecx*SIZEOF_JSAMPROW]         ; inptr0
76        mov     ebx, JSAMPROW [ebx+ecx*SIZEOF_JSAMPROW]         ; inptr1
77        mov     edx, JSAMPROW [edx+ecx*SIZEOF_JSAMPROW]         ; inptr2
78        mov     edi, JSAMPROW [edi]                             ; outptr
79
80        pop     ecx                     ; col
81
82        alignx  16,7
83.columnloop:
84        movpic  eax, POINTER [gotptr]   ; load GOT address (eax)
85
86        movq      mm6, MMWORD [ebx]     ; mm6=Cb(01234567)
87        movq      mm7, MMWORD [edx]     ; mm7=Cr(01234567)
88
89        pxor      mm1,mm1               ; mm1=(all 0's)
90        pcmpeqw   mm3,mm3
91        psllw     mm3,7                 ; mm3={0xFF80 0xFF80 0xFF80 0xFF80}
92
93        movq      mm4,mm6
94        punpckhbw mm6,mm1               ; mm6=Cb(4567)=CbH
95        punpcklbw mm4,mm1               ; mm4=Cb(0123)=CbL
96        movq      mm0,mm7
97        punpckhbw mm7,mm1               ; mm7=Cr(4567)=CrH
98        punpcklbw mm0,mm1               ; mm0=Cr(0123)=CrL
99
100        paddw     mm6,mm3
101        paddw     mm4,mm3
102        paddw     mm7,mm3
103        paddw     mm0,mm3
104
105        ; (Original)
106        ; R = Y                + 1.40200 * Cr
107        ; G = Y - 0.34414 * Cb - 0.71414 * Cr
108        ; B = Y + 1.77200 * Cb
109        ;
110        ; (This implementation)
111        ; R = Y                + 0.40200 * Cr + Cr
112        ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
113        ; B = Y - 0.22800 * Cb + Cb + Cb
114
115        movq    mm5,mm6                 ; mm5=CbH
116        movq    mm2,mm4                 ; mm2=CbL
117        paddw   mm6,mm6                 ; mm6=2*CbH
118        paddw   mm4,mm4                 ; mm4=2*CbL
119        movq    mm1,mm7                 ; mm1=CrH
120        movq    mm3,mm0                 ; mm3=CrL
121        paddw   mm7,mm7                 ; mm7=2*CrH
122        paddw   mm0,mm0                 ; mm0=2*CrL
123
124        pmulhw  mm6,[GOTOFF(eax,PW_MF0228)]     ; mm6=(2*CbH * -FIX(0.22800))
125        pmulhw  mm4,[GOTOFF(eax,PW_MF0228)]     ; mm4=(2*CbL * -FIX(0.22800))
126        pmulhw  mm7,[GOTOFF(eax,PW_F0402)]      ; mm7=(2*CrH * FIX(0.40200))
127        pmulhw  mm0,[GOTOFF(eax,PW_F0402)]      ; mm0=(2*CrL * FIX(0.40200))
128
129        paddw   mm6,[GOTOFF(eax,PW_ONE)]
130        paddw   mm4,[GOTOFF(eax,PW_ONE)]
131        psraw   mm6,1                   ; mm6=(CbH * -FIX(0.22800))
132        psraw   mm4,1                   ; mm4=(CbL * -FIX(0.22800))
133        paddw   mm7,[GOTOFF(eax,PW_ONE)]
134        paddw   mm0,[GOTOFF(eax,PW_ONE)]
135        psraw   mm7,1                   ; mm7=(CrH * FIX(0.40200))
136        psraw   mm0,1                   ; mm0=(CrL * FIX(0.40200))
137
138        paddw   mm6,mm5
139        paddw   mm4,mm2
140        paddw   mm6,mm5                 ; mm6=(CbH * FIX(1.77200))=(B-Y)H
141        paddw   mm4,mm2                 ; mm4=(CbL * FIX(1.77200))=(B-Y)L
142        paddw   mm7,mm1                 ; mm7=(CrH * FIX(1.40200))=(R-Y)H
143        paddw   mm0,mm3                 ; mm0=(CrL * FIX(1.40200))=(R-Y)L
144
145        movq    MMWORD [wk(0)], mm6     ; wk(0)=(B-Y)H
146        movq    MMWORD [wk(1)], mm7     ; wk(1)=(R-Y)H
147
148        movq      mm6,mm5
149        movq      mm7,mm2
150        punpcklwd mm5,mm1
151        punpckhwd mm6,mm1
152        pmaddwd   mm5,[GOTOFF(eax,PW_MF0344_F0285)]
153        pmaddwd   mm6,[GOTOFF(eax,PW_MF0344_F0285)]
154        punpcklwd mm2,mm3
155        punpckhwd mm7,mm3
156        pmaddwd   mm2,[GOTOFF(eax,PW_MF0344_F0285)]
157        pmaddwd   mm7,[GOTOFF(eax,PW_MF0344_F0285)]
158
159        paddd     mm5,[GOTOFF(eax,PD_ONEHALF)]
160        paddd     mm6,[GOTOFF(eax,PD_ONEHALF)]
161        psrad     mm5,SCALEBITS
162        psrad     mm6,SCALEBITS
163        paddd     mm2,[GOTOFF(eax,PD_ONEHALF)]
164        paddd     mm7,[GOTOFF(eax,PD_ONEHALF)]
165        psrad     mm2,SCALEBITS
166        psrad     mm7,SCALEBITS
167
168        packssdw  mm5,mm6       ; mm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
169        packssdw  mm2,mm7       ; mm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
170        psubw     mm5,mm1       ; mm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
171        psubw     mm2,mm3       ; mm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
172
173        movq    MMWORD [wk(2)], mm5     ; wk(2)=(G-Y)H
174
175        mov     al,2                    ; Yctr
176        jmp     short .Yloop_1st
177        alignx  16,7
178
179.Yloop_2nd:
180        movq    mm0, MMWORD [wk(1)]     ; mm0=(R-Y)H
181        movq    mm2, MMWORD [wk(2)]     ; mm2=(G-Y)H
182        movq    mm4, MMWORD [wk(0)]     ; mm4=(B-Y)H
183        alignx  16,7
184
185.Yloop_1st:
186        movq    mm7, MMWORD [esi]       ; mm7=Y(01234567)
187
188        pcmpeqw mm6,mm6
189        psrlw   mm6,BYTE_BIT            ; mm6={0xFF 0x00 0xFF 0x00 ..}
190        pand    mm6,mm7                 ; mm6=Y(0246)=YE
191        psrlw   mm7,BYTE_BIT            ; mm7=Y(1357)=YO
192
193        movq    mm1,mm0                 ; mm1=mm0=(R-Y)(L/H)
194        movq    mm3,mm2                 ; mm3=mm2=(G-Y)(L/H)
195        movq    mm5,mm4                 ; mm5=mm4=(B-Y)(L/H)
196
197        paddw     mm0,mm6               ; mm0=((R-Y)+YE)=RE=(R0 R2 R4 R6)
198        paddw     mm1,mm7               ; mm1=((R-Y)+YO)=RO=(R1 R3 R5 R7)
199        packuswb  mm0,mm0               ; mm0=(R0 R2 R4 R6 ** ** ** **)
200        packuswb  mm1,mm1               ; mm1=(R1 R3 R5 R7 ** ** ** **)
201
202        paddw     mm2,mm6               ; mm2=((G-Y)+YE)=GE=(G0 G2 G4 G6)
203        paddw     mm3,mm7               ; mm3=((G-Y)+YO)=GO=(G1 G3 G5 G7)
204        packuswb  mm2,mm2               ; mm2=(G0 G2 G4 G6 ** ** ** **)
205        packuswb  mm3,mm3               ; mm3=(G1 G3 G5 G7 ** ** ** **)
206
207        paddw     mm4,mm6               ; mm4=((B-Y)+YE)=BE=(B0 B2 B4 B6)
208        paddw     mm5,mm7               ; mm5=((B-Y)+YO)=BO=(B1 B3 B5 B7)
209        packuswb  mm4,mm4               ; mm4=(B0 B2 B4 B6 ** ** ** **)
210        packuswb  mm5,mm5               ; mm5=(B1 B3 B5 B7 ** ** ** **)
211
212%if RGB_PIXELSIZE == 3 ; ---------------
213
214        ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
215        ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
216        ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
217        ; mmG=(** ** ** ** ** ** ** **), mmH=(** ** ** ** ** ** ** **)
218
219        punpcklbw mmA,mmC               ; mmA=(00 10 02 12 04 14 06 16)
220        punpcklbw mmE,mmB               ; mmE=(20 01 22 03 24 05 26 07)
221        punpcklbw mmD,mmF               ; mmD=(11 21 13 23 15 25 17 27)
222
223        movq      mmG,mmA
224        movq      mmH,mmA
225        punpcklwd mmA,mmE               ; mmA=(00 10 20 01 02 12 22 03)
226        punpckhwd mmG,mmE               ; mmG=(04 14 24 05 06 16 26 07)
227
228        psrlq     mmH,2*BYTE_BIT        ; mmH=(02 12 04 14 06 16 -- --)
229        psrlq     mmE,2*BYTE_BIT        ; mmE=(22 03 24 05 26 07 -- --)
230
231        movq      mmC,mmD
232        movq      mmB,mmD
233        punpcklwd mmD,mmH               ; mmD=(11 21 02 12 13 23 04 14)
234        punpckhwd mmC,mmH               ; mmC=(15 25 06 16 17 27 -- --)
235
236        psrlq     mmB,2*BYTE_BIT        ; mmB=(13 23 15 25 17 27 -- --)
237
238        movq      mmF,mmE
239        punpcklwd mmE,mmB               ; mmE=(22 03 13 23 24 05 15 25)
240        punpckhwd mmF,mmB               ; mmF=(26 07 17 27 -- -- -- --)
241
242        punpckldq mmA,mmD               ; mmA=(00 10 20 01 11 21 02 12)
243        punpckldq mmE,mmG               ; mmE=(22 03 13 23 04 14 24 05)
244        punpckldq mmC,mmF               ; mmC=(15 25 06 16 26 07 17 27)
245
246        cmp     ecx, byte SIZEOF_MMWORD
247        jb      short .column_st16
248
249        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
250        movq    MMWORD [edi+1*SIZEOF_MMWORD], mmE
251        movq    MMWORD [edi+2*SIZEOF_MMWORD], mmC
252
253        sub     ecx, byte SIZEOF_MMWORD
254        jz      near .endcolumn
255
256        add     edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD   ; outptr
257        add     esi, byte SIZEOF_MMWORD                 ; inptr0
258        dec     al                      ; Yctr
259        jnz     near .Yloop_2nd
260
261        add     ebx, byte SIZEOF_MMWORD                 ; inptr1
262        add     edx, byte SIZEOF_MMWORD                 ; inptr2
263        jmp     near .columnloop
264        alignx  16,7
265
266.column_st16:
267        lea     ecx, [ecx+ecx*2]        ; imul ecx, RGB_PIXELSIZE
268        cmp     ecx, byte 2*SIZEOF_MMWORD
269        jb      short .column_st8
270        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
271        movq    MMWORD [edi+1*SIZEOF_MMWORD], mmE
272        movq    mmA,mmC
273        sub     ecx, byte 2*SIZEOF_MMWORD
274        add     edi, byte 2*SIZEOF_MMWORD
275        jmp     short .column_st4
276.column_st8:
277        cmp     ecx, byte SIZEOF_MMWORD
278        jb      short .column_st4
279        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
280        movq    mmA,mmE
281        sub     ecx, byte SIZEOF_MMWORD
282        add     edi, byte SIZEOF_MMWORD
283.column_st4:
284        movd    eax,mmA
285        cmp     ecx, byte SIZEOF_DWORD
286        jb      short .column_st2
287        mov     DWORD [edi+0*SIZEOF_DWORD], eax
288        psrlq   mmA,DWORD_BIT
289        movd    eax,mmA
290        sub     ecx, byte SIZEOF_DWORD
291        add     edi, byte SIZEOF_DWORD
292.column_st2:
293        cmp     ecx, byte SIZEOF_WORD
294        jb      short .column_st1
295        mov     WORD [edi+0*SIZEOF_WORD], ax
296        shr     eax,WORD_BIT
297        sub     ecx, byte SIZEOF_WORD
298        add     edi, byte SIZEOF_WORD
299.column_st1:
300        cmp     ecx, byte SIZEOF_BYTE
301        jb      short .endcolumn
302        mov     BYTE [edi+0*SIZEOF_BYTE], al
303
304%else ; RGB_PIXELSIZE == 4 ; -----------
305
306%ifdef RGBX_FILLER_0XFF
307        pcmpeqb   mm6,mm6               ; mm6=(X0 X2 X4 X6 ** ** ** **)
308        pcmpeqb   mm7,mm7               ; mm7=(X1 X3 X5 X7 ** ** ** **)
309%else
310        pxor      mm6,mm6               ; mm6=(X0 X2 X4 X6 ** ** ** **)
311        pxor      mm7,mm7               ; mm7=(X1 X3 X5 X7 ** ** ** **)
312%endif
313        ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
314        ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
315        ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
316        ; mmG=(30 32 34 36 ** ** ** **), mmH=(31 33 35 37 ** ** ** **)
317
318        punpcklbw mmA,mmC               ; mmA=(00 10 02 12 04 14 06 16)
319        punpcklbw mmE,mmG               ; mmE=(20 30 22 32 24 34 26 36)
320        punpcklbw mmB,mmD               ; mmB=(01 11 03 13 05 15 07 17)
321        punpcklbw mmF,mmH               ; mmF=(21 31 23 33 25 35 27 37)
322
323        movq      mmC,mmA
324        punpcklwd mmA,mmE               ; mmA=(00 10 20 30 02 12 22 32)
325        punpckhwd mmC,mmE               ; mmC=(04 14 24 34 06 16 26 36)
326        movq      mmG,mmB
327        punpcklwd mmB,mmF               ; mmB=(01 11 21 31 03 13 23 33)
328        punpckhwd mmG,mmF               ; mmG=(05 15 25 35 07 17 27 37)
329
330        movq      mmD,mmA
331        punpckldq mmA,mmB               ; mmA=(00 10 20 30 01 11 21 31)
332        punpckhdq mmD,mmB               ; mmD=(02 12 22 32 03 13 23 33)
333        movq      mmH,mmC
334        punpckldq mmC,mmG               ; mmC=(04 14 24 34 05 15 25 35)
335        punpckhdq mmH,mmG               ; mmH=(06 16 26 36 07 17 27 37)
336
337        cmp     ecx, byte SIZEOF_MMWORD
338        jb      short .column_st16
339
340        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
341        movq    MMWORD [edi+1*SIZEOF_MMWORD], mmD
342        movq    MMWORD [edi+2*SIZEOF_MMWORD], mmC
343        movq    MMWORD [edi+3*SIZEOF_MMWORD], mmH
344
345        sub     ecx, byte SIZEOF_MMWORD
346        jz      short .endcolumn
347
348        add     edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD   ; outptr
349        add     esi, byte SIZEOF_MMWORD                 ; inptr0
350        dec     al                      ; Yctr
351        jnz     near .Yloop_2nd
352
353        add     ebx, byte SIZEOF_MMWORD                 ; inptr1
354        add     edx, byte SIZEOF_MMWORD                 ; inptr2
355        jmp     near .columnloop
356        alignx  16,7
357
358.column_st16:
359        cmp     ecx, byte SIZEOF_MMWORD/2
360        jb      short .column_st8
361        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
362        movq    MMWORD [edi+1*SIZEOF_MMWORD], mmD
363        movq    mmA,mmC
364        movq    mmD,mmH
365        sub     ecx, byte SIZEOF_MMWORD/2
366        add     edi, byte 2*SIZEOF_MMWORD
367.column_st8:
368        cmp     ecx, byte SIZEOF_MMWORD/4
369        jb      short .column_st4
370        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
371        movq    mmA,mmD
372        sub     ecx, byte SIZEOF_MMWORD/4
373        add     edi, byte 1*SIZEOF_MMWORD
374.column_st4:
375        cmp     ecx, byte SIZEOF_MMWORD/8
376        jb      short .endcolumn
377        movd    DWORD [edi+0*SIZEOF_DWORD], mmA
378
379%endif ; RGB_PIXELSIZE ; ---------------
380
381.endcolumn:
382        emms            ; empty MMX state
383
384.return:
385        pop     edi
386        pop     esi
387;       pop     edx             ; need not be preserved
388;       pop     ecx             ; need not be preserved
389        pop     ebx
390        mov     esp,ebp         ; esp <- aligned ebp
391        pop     esp             ; esp <- original ebp
392        pop     ebp
393        ret
394
395; --------------------------------------------------------------------------
396;
397; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
398;
399; GLOBAL(void)
400; jsimd_h2v2_merged_upsample_mmx (JDIMENSION output_width,
401;                                 JSAMPIMAGE input_buf,
402;                                 JDIMENSION in_row_group_ctr,
403;                                 JSAMPARRAY output_buf);
404;
405
406%define output_width(b) (b)+8                   ; JDIMENSION output_width
407%define input_buf(b)            (b)+12          ; JSAMPIMAGE input_buf
408%define in_row_group_ctr(b)     (b)+16          ; JDIMENSION in_row_group_ctr
409%define output_buf(b)           (b)+20          ; JSAMPARRAY output_buf
410
411        align   16
412        global  EXTN(jsimd_h2v2_merged_upsample_mmx)
413
414EXTN(jsimd_h2v2_merged_upsample_mmx):
415        push    ebp
416        mov     ebp,esp
417        push    ebx
418;       push    ecx             ; need not be preserved
419;       push    edx             ; need not be preserved
420        push    esi
421        push    edi
422
423        mov     eax, JDIMENSION [output_width(ebp)]
424
425        mov     edi, JSAMPIMAGE [input_buf(ebp)]
426        mov     ecx, JDIMENSION [in_row_group_ctr(ebp)]
427        mov     esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
428        mov     ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
429        mov     edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
430        mov     edi, JSAMPARRAY [output_buf(ebp)]
431        lea     esi, [esi+ecx*SIZEOF_JSAMPROW]
432
433        push    edx                     ; inptr2
434        push    ebx                     ; inptr1
435        push    esi                     ; inptr00
436        mov     ebx,esp
437
438        push    edi                     ; output_buf (outptr0)
439        push    ecx                     ; in_row_group_ctr
440        push    ebx                     ; input_buf
441        push    eax                     ; output_width
442
443        call    near EXTN(jsimd_h2v1_merged_upsample_mmx)
444
445        add     esi, byte SIZEOF_JSAMPROW       ; inptr01
446        add     edi, byte SIZEOF_JSAMPROW       ; outptr1
447        mov     POINTER [ebx+0*SIZEOF_POINTER], esi
448        mov     POINTER [ebx-1*SIZEOF_POINTER], edi
449
450        call    near EXTN(jsimd_h2v1_merged_upsample_mmx)
451
452        add     esp, byte 7*SIZEOF_DWORD
453
454        pop     edi
455        pop     esi
456;       pop     edx             ; need not be preserved
457;       pop     ecx             ; need not be preserved
458        pop     ebx
459        pop     ebp
460        ret
461
462; For some reason, the OS X linker does not honor the request to align the
463; segment unless we do this.
464        align   16
465