|
1 ; |
|
2 ; jcsamss2.asm - downsampling (SSE2) |
|
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 "jsimdext.inc" |
|
20 |
|
21 ; -------------------------------------------------------------------------- |
|
22 SECTION SEG_TEXT |
|
23 BITS 32 |
|
24 ; |
|
25 ; Downsample pixel values of a single component. |
|
26 ; This version handles the common case of 2:1 horizontal and 1:1 vertical, |
|
27 ; without smoothing. |
|
28 ; |
|
29 ; GLOBAL(void) |
|
30 ; jsimd_h2v1_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor, |
|
31 ; JDIMENSION v_samp_factor, JDIMENSION width_blocks, |
|
32 ; JSAMPARRAY input_data, JSAMPARRAY output_data); |
|
33 ; |
|
34 |
|
35 %define img_width(b) (b)+8 ; JDIMENSION image_width |
|
36 %define max_v_samp(b) (b)+12 ; int max_v_samp_factor |
|
37 %define v_samp(b) (b)+16 ; JDIMENSION v_samp_factor |
|
38 %define width_blks(b) (b)+20 ; JDIMENSION width_blocks |
|
39 %define input_data(b) (b)+24 ; JSAMPARRAY input_data |
|
40 %define output_data(b) (b)+28 ; JSAMPARRAY output_data |
|
41 |
|
42 align 16 |
|
43 global EXTN(jsimd_h2v1_downsample_sse2) |
|
44 |
|
45 EXTN(jsimd_h2v1_downsample_sse2): |
|
46 push ebp |
|
47 mov ebp,esp |
|
48 ; push ebx ; unused |
|
49 ; push ecx ; need not be preserved |
|
50 ; push edx ; need not be preserved |
|
51 push esi |
|
52 push edi |
|
53 |
|
54 mov ecx, JDIMENSION [width_blks(ebp)] |
|
55 shl ecx,3 ; imul ecx,DCTSIZE (ecx = output_cols) |
|
56 jz near .return |
|
57 |
|
58 mov edx, JDIMENSION [img_width(ebp)] |
|
59 |
|
60 ; -- expand_right_edge |
|
61 |
|
62 push ecx |
|
63 shl ecx,1 ; output_cols * 2 |
|
64 sub ecx,edx |
|
65 jle short .expand_end |
|
66 |
|
67 mov eax, INT [max_v_samp(ebp)] |
|
68 test eax,eax |
|
69 jle short .expand_end |
|
70 |
|
71 cld |
|
72 mov esi, JSAMPARRAY [input_data(ebp)] ; input_data |
|
73 alignx 16,7 |
|
74 .expandloop: |
|
75 push eax |
|
76 push ecx |
|
77 |
|
78 mov edi, JSAMPROW [esi] |
|
79 add edi,edx |
|
80 mov al, JSAMPLE [edi-1] |
|
81 |
|
82 rep stosb |
|
83 |
|
84 pop ecx |
|
85 pop eax |
|
86 |
|
87 add esi, byte SIZEOF_JSAMPROW |
|
88 dec eax |
|
89 jg short .expandloop |
|
90 |
|
91 .expand_end: |
|
92 pop ecx ; output_cols |
|
93 |
|
94 ; -- h2v1_downsample |
|
95 |
|
96 mov eax, JDIMENSION [v_samp(ebp)] ; rowctr |
|
97 test eax,eax |
|
98 jle near .return |
|
99 |
|
100 mov edx, 0x00010000 ; bias pattern |
|
101 movd xmm7,edx |
|
102 pcmpeqw xmm6,xmm6 |
|
103 pshufd xmm7,xmm7,0x00 ; xmm7={0, 1, 0, 1, 0, 1, 0, 1} |
|
104 psrlw xmm6,BYTE_BIT ; xmm6={0xFF 0x00 0xFF 0x00 ..} |
|
105 |
|
106 mov esi, JSAMPARRAY [input_data(ebp)] ; input_data |
|
107 mov edi, JSAMPARRAY [output_data(ebp)] ; output_data |
|
108 alignx 16,7 |
|
109 .rowloop: |
|
110 push ecx |
|
111 push edi |
|
112 push esi |
|
113 |
|
114 mov esi, JSAMPROW [esi] ; inptr |
|
115 mov edi, JSAMPROW [edi] ; outptr |
|
116 |
|
117 cmp ecx, byte SIZEOF_XMMWORD |
|
118 jae short .columnloop |
|
119 alignx 16,7 |
|
120 |
|
121 .columnloop_r8: |
|
122 movdqa xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD] |
|
123 pxor xmm1,xmm1 |
|
124 mov ecx, SIZEOF_XMMWORD |
|
125 jmp short .downsample |
|
126 alignx 16,7 |
|
127 |
|
128 .columnloop: |
|
129 movdqa xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD] |
|
130 movdqa xmm1, XMMWORD [esi+1*SIZEOF_XMMWORD] |
|
131 |
|
132 .downsample: |
|
133 movdqa xmm2,xmm0 |
|
134 movdqa xmm3,xmm1 |
|
135 |
|
136 pand xmm0,xmm6 |
|
137 psrlw xmm2,BYTE_BIT |
|
138 pand xmm1,xmm6 |
|
139 psrlw xmm3,BYTE_BIT |
|
140 |
|
141 paddw xmm0,xmm2 |
|
142 paddw xmm1,xmm3 |
|
143 paddw xmm0,xmm7 |
|
144 paddw xmm1,xmm7 |
|
145 psrlw xmm0,1 |
|
146 psrlw xmm1,1 |
|
147 |
|
148 packuswb xmm0,xmm1 |
|
149 |
|
150 movdqa XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0 |
|
151 |
|
152 sub ecx, byte SIZEOF_XMMWORD ; outcol |
|
153 add esi, byte 2*SIZEOF_XMMWORD ; inptr |
|
154 add edi, byte 1*SIZEOF_XMMWORD ; outptr |
|
155 cmp ecx, byte SIZEOF_XMMWORD |
|
156 jae short .columnloop |
|
157 test ecx,ecx |
|
158 jnz short .columnloop_r8 |
|
159 |
|
160 pop esi |
|
161 pop edi |
|
162 pop ecx |
|
163 |
|
164 add esi, byte SIZEOF_JSAMPROW ; input_data |
|
165 add edi, byte SIZEOF_JSAMPROW ; output_data |
|
166 dec eax ; rowctr |
|
167 jg near .rowloop |
|
168 |
|
169 .return: |
|
170 pop edi |
|
171 pop esi |
|
172 ; pop edx ; need not be preserved |
|
173 ; pop ecx ; need not be preserved |
|
174 ; pop ebx ; unused |
|
175 pop ebp |
|
176 ret |
|
177 |
|
178 ; -------------------------------------------------------------------------- |
|
179 ; |
|
180 ; Downsample pixel values of a single component. |
|
181 ; This version handles the standard case of 2:1 horizontal and 2:1 vertical, |
|
182 ; without smoothing. |
|
183 ; |
|
184 ; GLOBAL(void) |
|
185 ; jsimd_h2v2_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor, |
|
186 ; JDIMENSION v_samp_factor, JDIMENSION width_blocks, |
|
187 ; JSAMPARRAY input_data, JSAMPARRAY output_data); |
|
188 ; |
|
189 |
|
190 %define img_width(b) (b)+8 ; JDIMENSION image_width |
|
191 %define max_v_samp(b) (b)+12 ; int max_v_samp_factor |
|
192 %define v_samp(b) (b)+16 ; JDIMENSION v_samp_factor |
|
193 %define width_blks(b) (b)+20 ; JDIMENSION width_blocks |
|
194 %define input_data(b) (b)+24 ; JSAMPARRAY input_data |
|
195 %define output_data(b) (b)+28 ; JSAMPARRAY output_data |
|
196 |
|
197 align 16 |
|
198 global EXTN(jsimd_h2v2_downsample_sse2) |
|
199 |
|
200 EXTN(jsimd_h2v2_downsample_sse2): |
|
201 push ebp |
|
202 mov ebp,esp |
|
203 ; push ebx ; unused |
|
204 ; push ecx ; need not be preserved |
|
205 ; push edx ; need not be preserved |
|
206 push esi |
|
207 push edi |
|
208 |
|
209 mov ecx, JDIMENSION [width_blks(ebp)] |
|
210 shl ecx,3 ; imul ecx,DCTSIZE (ecx = output_cols) |
|
211 jz near .return |
|
212 |
|
213 mov edx, JDIMENSION [img_width(ebp)] |
|
214 |
|
215 ; -- expand_right_edge |
|
216 |
|
217 push ecx |
|
218 shl ecx,1 ; output_cols * 2 |
|
219 sub ecx,edx |
|
220 jle short .expand_end |
|
221 |
|
222 mov eax, INT [max_v_samp(ebp)] |
|
223 test eax,eax |
|
224 jle short .expand_end |
|
225 |
|
226 cld |
|
227 mov esi, JSAMPARRAY [input_data(ebp)] ; input_data |
|
228 alignx 16,7 |
|
229 .expandloop: |
|
230 push eax |
|
231 push ecx |
|
232 |
|
233 mov edi, JSAMPROW [esi] |
|
234 add edi,edx |
|
235 mov al, JSAMPLE [edi-1] |
|
236 |
|
237 rep stosb |
|
238 |
|
239 pop ecx |
|
240 pop eax |
|
241 |
|
242 add esi, byte SIZEOF_JSAMPROW |
|
243 dec eax |
|
244 jg short .expandloop |
|
245 |
|
246 .expand_end: |
|
247 pop ecx ; output_cols |
|
248 |
|
249 ; -- h2v2_downsample |
|
250 |
|
251 mov eax, JDIMENSION [v_samp(ebp)] ; rowctr |
|
252 test eax,eax |
|
253 jle near .return |
|
254 |
|
255 mov edx, 0x00020001 ; bias pattern |
|
256 movd xmm7,edx |
|
257 pcmpeqw xmm6,xmm6 |
|
258 pshufd xmm7,xmm7,0x00 ; xmm7={1, 2, 1, 2, 1, 2, 1, 2} |
|
259 psrlw xmm6,BYTE_BIT ; xmm6={0xFF 0x00 0xFF 0x00 ..} |
|
260 |
|
261 mov esi, JSAMPARRAY [input_data(ebp)] ; input_data |
|
262 mov edi, JSAMPARRAY [output_data(ebp)] ; output_data |
|
263 alignx 16,7 |
|
264 .rowloop: |
|
265 push ecx |
|
266 push edi |
|
267 push esi |
|
268 |
|
269 mov edx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; inptr0 |
|
270 mov esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; inptr1 |
|
271 mov edi, JSAMPROW [edi] ; outptr |
|
272 |
|
273 cmp ecx, byte SIZEOF_XMMWORD |
|
274 jae short .columnloop |
|
275 alignx 16,7 |
|
276 |
|
277 .columnloop_r8: |
|
278 movdqa xmm0, XMMWORD [edx+0*SIZEOF_XMMWORD] |
|
279 movdqa xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD] |
|
280 pxor xmm2,xmm2 |
|
281 pxor xmm3,xmm3 |
|
282 mov ecx, SIZEOF_XMMWORD |
|
283 jmp short .downsample |
|
284 alignx 16,7 |
|
285 |
|
286 .columnloop: |
|
287 movdqa xmm0, XMMWORD [edx+0*SIZEOF_XMMWORD] |
|
288 movdqa xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD] |
|
289 movdqa xmm2, XMMWORD [edx+1*SIZEOF_XMMWORD] |
|
290 movdqa xmm3, XMMWORD [esi+1*SIZEOF_XMMWORD] |
|
291 |
|
292 .downsample: |
|
293 movdqa xmm4,xmm0 |
|
294 movdqa xmm5,xmm1 |
|
295 pand xmm0,xmm6 |
|
296 psrlw xmm4,BYTE_BIT |
|
297 pand xmm1,xmm6 |
|
298 psrlw xmm5,BYTE_BIT |
|
299 paddw xmm0,xmm4 |
|
300 paddw xmm1,xmm5 |
|
301 |
|
302 movdqa xmm4,xmm2 |
|
303 movdqa xmm5,xmm3 |
|
304 pand xmm2,xmm6 |
|
305 psrlw xmm4,BYTE_BIT |
|
306 pand xmm3,xmm6 |
|
307 psrlw xmm5,BYTE_BIT |
|
308 paddw xmm2,xmm4 |
|
309 paddw xmm3,xmm5 |
|
310 |
|
311 paddw xmm0,xmm1 |
|
312 paddw xmm2,xmm3 |
|
313 paddw xmm0,xmm7 |
|
314 paddw xmm2,xmm7 |
|
315 psrlw xmm0,2 |
|
316 psrlw xmm2,2 |
|
317 |
|
318 packuswb xmm0,xmm2 |
|
319 |
|
320 movdqa XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0 |
|
321 |
|
322 sub ecx, byte SIZEOF_XMMWORD ; outcol |
|
323 add edx, byte 2*SIZEOF_XMMWORD ; inptr0 |
|
324 add esi, byte 2*SIZEOF_XMMWORD ; inptr1 |
|
325 add edi, byte 1*SIZEOF_XMMWORD ; outptr |
|
326 cmp ecx, byte SIZEOF_XMMWORD |
|
327 jae near .columnloop |
|
328 test ecx,ecx |
|
329 jnz near .columnloop_r8 |
|
330 |
|
331 pop esi |
|
332 pop edi |
|
333 pop ecx |
|
334 |
|
335 add esi, byte 2*SIZEOF_JSAMPROW ; input_data |
|
336 add edi, byte 1*SIZEOF_JSAMPROW ; output_data |
|
337 dec eax ; rowctr |
|
338 jg near .rowloop |
|
339 |
|
340 .return: |
|
341 pop edi |
|
342 pop esi |
|
343 ; pop edx ; need not be preserved |
|
344 ; pop ecx ; need not be preserved |
|
345 ; pop ebx ; unused |
|
346 pop ebp |
|
347 ret |
|
348 |
|
349 ; For some reason, the OS X linker does not honor the request to align the |
|
350 ; segment unless we do this. |
|
351 align 16 |