1;
2; jcgryext.asm - grayscale colorspace conversion (SSE2)
3;
4; Copyright (C) 2011, D. R. Commander.
5;
6; Based on the x86 SIMD extension for IJG JPEG library
7; Copyright (C) 1999-2006, MIYASAKA Masaru.
8; For conditions of distribution and use, see copyright notice in jsimdext.inc
9;
10; This file should be assembled with NASM (Netwide Assembler),
11; can *not* be assembled with Microsoft's MASM or any compatible
12; assembler (including Borland's Turbo Assembler).
13; NASM is available from http://nasm.sourceforge.net/ or
14; http://sourceforge.net/project/showfiles.php?group_id=6208
15;
16; [TAB8]
17
18%include "jcolsamp.inc"
19
20; --------------------------------------------------------------------------
21;
22; Convert some rows of samples to the output colorspace.
23;
24; GLOBAL(void)
25; jsimd_rgb_gray_convert_sse2 (JDIMENSION img_width,
26;                              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
27;                              JDIMENSION output_row, int num_rows);
28;
29
30%define img_width(b)    (b)+8           ; JDIMENSION img_width
31%define input_buf(b)    (b)+12          ; JSAMPARRAY input_buf
32%define output_buf(b)   (b)+16          ; JSAMPIMAGE output_buf
33%define output_row(b)   (b)+20          ; JDIMENSION output_row
34%define num_rows(b)     (b)+24          ; int num_rows
35
36%define original_ebp    ebp+0
37%define wk(i)           ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
38%define WK_NUM          2
39%define gotptr          wk(0)-SIZEOF_POINTER    ; void * gotptr
40
41        align   16
42
43        global  EXTN(jsimd_rgb_gray_convert_sse2)
44
45EXTN(jsimd_rgb_gray_convert_sse2):
46        push    ebp
47        mov     eax,esp                         ; eax = original ebp
48        sub     esp, byte 4
49        and     esp, byte (-SIZEOF_XMMWORD)     ; align to 128 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 [img_width(eax)]
64        test    ecx,ecx
65        jz      near .return
66
67        push    ecx
68
69        mov     esi, JSAMPIMAGE [output_buf(eax)]
70        mov     ecx, JDIMENSION [output_row(eax)]
71        mov     edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
72        lea     edi, [edi+ecx*SIZEOF_JSAMPROW]
73
74        pop     ecx
75
76        mov     esi, JSAMPARRAY [input_buf(eax)]
77        mov     eax, INT [num_rows(eax)]
78        test    eax,eax
79        jle     near .return
80        alignx  16,7
81.rowloop:
82        pushpic eax
83        push    edi
84        push    esi
85        push    ecx                     ; col
86
87        mov     esi, JSAMPROW [esi]     ; inptr
88        mov     edi, JSAMPROW [edi]     ; outptr0
89        movpic  eax, POINTER [gotptr]   ; load GOT address (eax)
90
91        cmp     ecx, byte SIZEOF_XMMWORD
92        jae     near .columnloop
93        alignx  16,7
94
95%if RGB_PIXELSIZE == 3 ; ---------------
96
97.column_ld1:
98        push    eax
99        push    edx
100        lea     ecx,[ecx+ecx*2]         ; imul ecx,RGB_PIXELSIZE
101        test    cl, SIZEOF_BYTE
102        jz      short .column_ld2
103        sub     ecx, byte SIZEOF_BYTE
104        movzx   eax, BYTE [esi+ecx]
105.column_ld2:
106        test    cl, SIZEOF_WORD
107        jz      short .column_ld4
108        sub     ecx, byte SIZEOF_WORD
109        movzx   edx, WORD [esi+ecx]
110        shl     eax, WORD_BIT
111        or      eax,edx
112.column_ld4:
113        movd    xmmA,eax
114        pop     edx
115        pop     eax
116        test    cl, SIZEOF_DWORD
117        jz      short .column_ld8
118        sub     ecx, byte SIZEOF_DWORD
119        movd    xmmF, XMM_DWORD [esi+ecx]
120        pslldq  xmmA, SIZEOF_DWORD
121        por     xmmA,xmmF
122.column_ld8:
123        test    cl, SIZEOF_MMWORD
124        jz      short .column_ld16
125        sub     ecx, byte SIZEOF_MMWORD
126        movq    xmmB, XMM_MMWORD [esi+ecx]
127        pslldq  xmmA, SIZEOF_MMWORD
128        por     xmmA,xmmB
129.column_ld16:
130        test    cl, SIZEOF_XMMWORD
131        jz      short .column_ld32
132        movdqa  xmmF,xmmA
133        movdqu  xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
134        mov     ecx, SIZEOF_XMMWORD
135        jmp     short .rgb_gray_cnv
136.column_ld32:
137        test    cl, 2*SIZEOF_XMMWORD
138        mov     ecx, SIZEOF_XMMWORD
139        jz      short .rgb_gray_cnv
140        movdqa  xmmB,xmmA
141        movdqu  xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
142        movdqu  xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
143        jmp     short .rgb_gray_cnv
144        alignx  16,7
145
146.columnloop:
147        movdqu  xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
148        movdqu  xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
149        movdqu  xmmB, XMMWORD [esi+2*SIZEOF_XMMWORD]
150
151.rgb_gray_cnv:
152        ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
153        ; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
154        ; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
155
156        movdqa    xmmG,xmmA
157        pslldq    xmmA,8        ; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12)
158        psrldq    xmmG,8        ; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --)
159
160        punpckhbw xmmA,xmmF     ; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A)
161        pslldq    xmmF,8        ; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27)
162
163        punpcklbw xmmG,xmmB     ; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D)
164        punpckhbw xmmF,xmmB     ; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F)
165
166        movdqa    xmmD,xmmA
167        pslldq    xmmA,8        ; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09)
168        psrldq    xmmD,8        ; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --)
169
170        punpckhbw xmmA,xmmG     ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D)
171        pslldq    xmmG,8        ; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B)
172
173        punpcklbw xmmD,xmmF     ; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E)
174        punpckhbw xmmG,xmmF     ; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F)
175
176        movdqa    xmmE,xmmA
177        pslldq    xmmA,8        ; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C)
178        psrldq    xmmE,8        ; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --)
179
180        punpckhbw xmmA,xmmD     ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
181        pslldq    xmmD,8        ; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D)
182
183        punpcklbw xmmE,xmmG     ; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F)
184        punpckhbw xmmD,xmmG     ; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F)
185
186        pxor      xmmH,xmmH
187
188        movdqa    xmmC,xmmA
189        punpcklbw xmmA,xmmH     ; xmmA=(00 02 04 06 08 0A 0C 0E)
190        punpckhbw xmmC,xmmH     ; xmmC=(10 12 14 16 18 1A 1C 1E)
191
192        movdqa    xmmB,xmmE
193        punpcklbw xmmE,xmmH     ; xmmE=(20 22 24 26 28 2A 2C 2E)
194        punpckhbw xmmB,xmmH     ; xmmB=(01 03 05 07 09 0B 0D 0F)
195
196        movdqa    xmmF,xmmD
197        punpcklbw xmmD,xmmH     ; xmmD=(11 13 15 17 19 1B 1D 1F)
198        punpckhbw xmmF,xmmH     ; xmmF=(21 23 25 27 29 2B 2D 2F)
199
200%else ; RGB_PIXELSIZE == 4 ; -----------
201
202.column_ld1:
203        test    cl, SIZEOF_XMMWORD/16
204        jz      short .column_ld2
205        sub     ecx, byte SIZEOF_XMMWORD/16
206        movd    xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE]
207.column_ld2:
208        test    cl, SIZEOF_XMMWORD/8
209        jz      short .column_ld4
210        sub     ecx, byte SIZEOF_XMMWORD/8
211        movq    xmmE, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE]
212        pslldq  xmmA, SIZEOF_MMWORD
213        por     xmmA,xmmE
214.column_ld4:
215        test    cl, SIZEOF_XMMWORD/4
216        jz      short .column_ld8
217        sub     ecx, byte SIZEOF_XMMWORD/4
218        movdqa  xmmE,xmmA
219        movdqu  xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE]
220.column_ld8:
221        test    cl, SIZEOF_XMMWORD/2
222        mov     ecx, SIZEOF_XMMWORD
223        jz      short .rgb_gray_cnv
224        movdqa  xmmF,xmmA
225        movdqa  xmmH,xmmE
226        movdqu  xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
227        movdqu  xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
228        jmp     short .rgb_gray_cnv
229        alignx  16,7
230
231.columnloop:
232        movdqu  xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
233        movdqu  xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
234        movdqu  xmmF, XMMWORD [esi+2*SIZEOF_XMMWORD]
235        movdqu  xmmH, XMMWORD [esi+3*SIZEOF_XMMWORD]
236
237.rgb_gray_cnv:
238        ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
239        ; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
240        ; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
241        ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
242
243        movdqa    xmmD,xmmA
244        punpcklbw xmmA,xmmE     ; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35)
245        punpckhbw xmmD,xmmE     ; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37)
246
247        movdqa    xmmC,xmmF
248        punpcklbw xmmF,xmmH     ; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D)
249        punpckhbw xmmC,xmmH     ; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F)
250
251        movdqa    xmmB,xmmA
252        punpcklwd xmmA,xmmF     ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C)
253        punpckhwd xmmB,xmmF     ; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D)
254
255        movdqa    xmmG,xmmD
256        punpcklwd xmmD,xmmC     ; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E)
257        punpckhwd xmmG,xmmC     ; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F)
258
259        movdqa    xmmE,xmmA
260        punpcklbw xmmA,xmmD     ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
261        punpckhbw xmmE,xmmD     ; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E)
262
263        movdqa    xmmH,xmmB
264        punpcklbw xmmB,xmmG     ; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F)
265        punpckhbw xmmH,xmmG     ; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F)
266
267        pxor      xmmF,xmmF
268
269        movdqa    xmmC,xmmA
270        punpcklbw xmmA,xmmF     ; xmmA=(00 02 04 06 08 0A 0C 0E)
271        punpckhbw xmmC,xmmF     ; xmmC=(10 12 14 16 18 1A 1C 1E)
272
273        movdqa    xmmD,xmmB
274        punpcklbw xmmB,xmmF     ; xmmB=(01 03 05 07 09 0B 0D 0F)
275        punpckhbw xmmD,xmmF     ; xmmD=(11 13 15 17 19 1B 1D 1F)
276
277        movdqa    xmmG,xmmE
278        punpcklbw xmmE,xmmF     ; xmmE=(20 22 24 26 28 2A 2C 2E)
279        punpckhbw xmmG,xmmF     ; xmmG=(30 32 34 36 38 3A 3C 3E)
280
281        punpcklbw xmmF,xmmH
282        punpckhbw xmmH,xmmH
283        psrlw     xmmF,BYTE_BIT ; xmmF=(21 23 25 27 29 2B 2D 2F)
284        psrlw     xmmH,BYTE_BIT ; xmmH=(31 33 35 37 39 3B 3D 3F)
285
286%endif ; RGB_PIXELSIZE ; ---------------
287
288        ; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
289        ; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
290
291        ; (Original)
292        ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
293        ;
294        ; (This implementation)
295        ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
296
297        movdqa    xmm6,xmm1
298        punpcklwd xmm1,xmm3
299        punpckhwd xmm6,xmm3
300        pmaddwd   xmm1,[GOTOFF(eax,PW_F0299_F0337)] ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
301        pmaddwd   xmm6,[GOTOFF(eax,PW_F0299_F0337)] ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
302
303        movdqa    xmm7, xmm6    ; xmm7=ROH*FIX(0.299)+GOH*FIX(0.337)
304
305        movdqa    xmm6,xmm0
306        punpcklwd xmm0,xmm2
307        punpckhwd xmm6,xmm2
308        pmaddwd   xmm0,[GOTOFF(eax,PW_F0299_F0337)] ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
309        pmaddwd   xmm6,[GOTOFF(eax,PW_F0299_F0337)] ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
310
311        movdqa    XMMWORD [wk(0)], xmm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337)
312        movdqa    XMMWORD [wk(1)], xmm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337)
313
314        movdqa    xmm0, xmm5    ; xmm0=BO
315        movdqa    xmm6, xmm4    ; xmm6=BE
316
317        movdqa    xmm4,xmm0
318        punpcklwd xmm0,xmm3
319        punpckhwd xmm4,xmm3
320        pmaddwd   xmm0,[GOTOFF(eax,PW_F0114_F0250)] ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
321        pmaddwd   xmm4,[GOTOFF(eax,PW_F0114_F0250)] ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
322
323        movdqa    xmm3,[GOTOFF(eax,PD_ONEHALF)] ; xmm3=[PD_ONEHALF]
324
325        paddd     xmm0, xmm1
326        paddd     xmm4, xmm7
327        paddd     xmm0,xmm3
328        paddd     xmm4,xmm3
329        psrld     xmm0,SCALEBITS        ; xmm0=YOL
330        psrld     xmm4,SCALEBITS        ; xmm4=YOH
331        packssdw  xmm0,xmm4             ; xmm0=YO
332
333        movdqa    xmm4,xmm6
334        punpcklwd xmm6,xmm2
335        punpckhwd xmm4,xmm2
336        pmaddwd   xmm6,[GOTOFF(eax,PW_F0114_F0250)] ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
337        pmaddwd   xmm4,[GOTOFF(eax,PW_F0114_F0250)] ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
338
339        movdqa    xmm2,[GOTOFF(eax,PD_ONEHALF)] ; xmm2=[PD_ONEHALF]
340
341        paddd     xmm6, XMMWORD [wk(0)]
342        paddd     xmm4, XMMWORD [wk(1)]
343        paddd     xmm6,xmm2
344        paddd     xmm4,xmm2
345        psrld     xmm6,SCALEBITS        ; xmm6=YEL
346        psrld     xmm4,SCALEBITS        ; xmm4=YEH
347        packssdw  xmm6,xmm4             ; xmm6=YE
348
349        psllw     xmm0,BYTE_BIT
350        por       xmm6,xmm0             ; xmm6=Y
351        movdqa    XMMWORD [edi], xmm6   ; Save Y
352
353        sub     ecx, byte SIZEOF_XMMWORD
354        add     esi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; inptr
355        add     edi, byte SIZEOF_XMMWORD                ; outptr0
356        cmp     ecx, byte SIZEOF_XMMWORD
357        jae     near .columnloop
358        test    ecx,ecx
359        jnz     near .column_ld1
360
361        pop     ecx                     ; col
362        pop     esi
363        pop     edi
364        poppic  eax
365
366        add     esi, byte SIZEOF_JSAMPROW       ; input_buf
367        add     edi, byte SIZEOF_JSAMPROW
368        dec     eax                             ; num_rows
369        jg      near .rowloop
370
371.return:
372        pop     edi
373        pop     esi
374;       pop     edx             ; need not be preserved
375;       pop     ecx             ; need not be preserved
376        pop     ebx
377        mov     esp,ebp         ; esp <- aligned ebp
378        pop     esp             ; esp <- original ebp
379        pop     ebp
380        ret
381
382; For some reason, the OS X linker does not honor the request to align the
383; segment unless we do this.
384        align   16
385