Mercurial > libpostproc.hg
comparison postprocess_template.c @ 95:c24dab9bca80 libpostproc
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
author | diego |
---|---|
date | Sat, 22 Mar 2008 15:46:34 +0000 |
parents | a8ed701a0873 |
children | 404a656698ff |
comparison
equal
deleted
inserted
replaced
94:094413c45b0f | 95:c24dab9bca80 |
---|---|
40 | 40 |
41 #ifdef HAVE_MMX2 | 41 #ifdef HAVE_MMX2 |
42 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t" | 42 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t" |
43 #elif defined (HAVE_MMX) | 43 #elif defined (HAVE_MMX) |
44 #define PMINUB(b,a,t) \ | 44 #define PMINUB(b,a,t) \ |
45 "movq " #a ", " #t " \n\t"\ | 45 "movq " #a ", " #t " \n\t"\ |
46 "psubusb " #b ", " #t " \n\t"\ | 46 "psubusb " #b ", " #t " \n\t"\ |
47 "psubb " #t ", " #a " \n\t" | 47 "psubb " #t ", " #a " \n\t" |
48 #endif | 48 #endif |
49 | 49 |
50 #ifdef HAVE_MMX2 | 50 #ifdef HAVE_MMX2 |
51 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t" | 51 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t" |
52 #elif defined (HAVE_MMX) | 52 #elif defined (HAVE_MMX) |
53 #define PMAXUB(a,b) \ | 53 #define PMAXUB(a,b) \ |
54 "psubusb " #a ", " #b " \n\t"\ | 54 "psubusb " #a ", " #b " \n\t"\ |
55 "paddb " #a ", " #b " \n\t" | 55 "paddb " #a ", " #b " \n\t" |
56 #endif | 56 #endif |
57 | 57 |
58 //FIXME? |255-0| = 1 (should not be a problem ...) | 58 //FIXME? |255-0| = 1 (should not be a problem ...) |
59 #ifdef HAVE_MMX | 59 #ifdef HAVE_MMX |
60 /** | 60 /** |
61 * Check if the middle 8x8 Block in the given 8x16 block is flat | 61 * Check if the middle 8x8 Block in the given 8x16 block is flat |
62 */ | 62 */ |
63 static inline int RENAME(vertClassify)(uint8_t src[], int stride, PPContext *c){ | 63 static inline int RENAME(vertClassify)(uint8_t src[], int stride, PPContext *c){ |
64 int numEq= 0, dcOk; | 64 int numEq= 0, dcOk; |
65 src+= stride*4; // src points to begin of the 8x8 Block | 65 src+= stride*4; // src points to begin of the 8x8 Block |
66 asm volatile( | 66 asm volatile( |
67 "movq %0, %%mm7 \n\t" | 67 "movq %0, %%mm7 \n\t" |
68 "movq %1, %%mm6 \n\t" | 68 "movq %1, %%mm6 \n\t" |
69 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP]) | 69 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP]) |
70 ); | 70 ); |
71 | 71 |
72 asm volatile( | 72 asm volatile( |
73 "lea (%2, %3), %%"REG_a" \n\t" | 73 "lea (%2, %3), %%"REG_a" \n\t" |
74 // 0 1 2 3 4 5 6 7 8 9 | 74 // 0 1 2 3 4 5 6 7 8 9 |
75 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2 | 75 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2 |
76 | 76 |
77 "movq (%2), %%mm0 \n\t" | 77 "movq (%2), %%mm0 \n\t" |
78 "movq (%%"REG_a"), %%mm1 \n\t" | 78 "movq (%%"REG_a"), %%mm1 \n\t" |
79 "movq %%mm0, %%mm3 \n\t" | 79 "movq %%mm0, %%mm3 \n\t" |
80 "movq %%mm0, %%mm4 \n\t" | 80 "movq %%mm0, %%mm4 \n\t" |
81 PMAXUB(%%mm1, %%mm4) | 81 PMAXUB(%%mm1, %%mm4) |
82 PMINUB(%%mm1, %%mm3, %%mm5) | 82 PMINUB(%%mm1, %%mm3, %%mm5) |
83 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece | 83 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
84 "paddb %%mm7, %%mm0 \n\t" | 84 "paddb %%mm7, %%mm0 \n\t" |
85 "pcmpgtb %%mm6, %%mm0 \n\t" | 85 "pcmpgtb %%mm6, %%mm0 \n\t" |
86 | 86 |
87 "movq (%%"REG_a",%3), %%mm2 \n\t" | 87 "movq (%%"REG_a",%3), %%mm2 \n\t" |
88 PMAXUB(%%mm2, %%mm4) | 88 PMAXUB(%%mm2, %%mm4) |
89 PMINUB(%%mm2, %%mm3, %%mm5) | 89 PMINUB(%%mm2, %%mm3, %%mm5) |
90 "psubb %%mm2, %%mm1 \n\t" | 90 "psubb %%mm2, %%mm1 \n\t" |
91 "paddb %%mm7, %%mm1 \n\t" | 91 "paddb %%mm7, %%mm1 \n\t" |
92 "pcmpgtb %%mm6, %%mm1 \n\t" | 92 "pcmpgtb %%mm6, %%mm1 \n\t" |
93 "paddb %%mm1, %%mm0 \n\t" | 93 "paddb %%mm1, %%mm0 \n\t" |
94 | 94 |
95 "movq (%%"REG_a", %3, 2), %%mm1 \n\t" | 95 "movq (%%"REG_a", %3, 2), %%mm1 \n\t" |
96 PMAXUB(%%mm1, %%mm4) | 96 PMAXUB(%%mm1, %%mm4) |
97 PMINUB(%%mm1, %%mm3, %%mm5) | 97 PMINUB(%%mm1, %%mm3, %%mm5) |
98 "psubb %%mm1, %%mm2 \n\t" | 98 "psubb %%mm1, %%mm2 \n\t" |
99 "paddb %%mm7, %%mm2 \n\t" | 99 "paddb %%mm7, %%mm2 \n\t" |
100 "pcmpgtb %%mm6, %%mm2 \n\t" | 100 "pcmpgtb %%mm6, %%mm2 \n\t" |
101 "paddb %%mm2, %%mm0 \n\t" | 101 "paddb %%mm2, %%mm0 \n\t" |
102 | 102 |
103 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t" | 103 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t" |
104 | 104 |
105 "movq (%2, %3, 4), %%mm2 \n\t" | 105 "movq (%2, %3, 4), %%mm2 \n\t" |
106 PMAXUB(%%mm2, %%mm4) | 106 PMAXUB(%%mm2, %%mm4) |
107 PMINUB(%%mm2, %%mm3, %%mm5) | 107 PMINUB(%%mm2, %%mm3, %%mm5) |
108 "psubb %%mm2, %%mm1 \n\t" | 108 "psubb %%mm2, %%mm1 \n\t" |
109 "paddb %%mm7, %%mm1 \n\t" | 109 "paddb %%mm7, %%mm1 \n\t" |
110 "pcmpgtb %%mm6, %%mm1 \n\t" | 110 "pcmpgtb %%mm6, %%mm1 \n\t" |
111 "paddb %%mm1, %%mm0 \n\t" | 111 "paddb %%mm1, %%mm0 \n\t" |
112 | 112 |
113 "movq (%%"REG_a"), %%mm1 \n\t" | 113 "movq (%%"REG_a"), %%mm1 \n\t" |
114 PMAXUB(%%mm1, %%mm4) | 114 PMAXUB(%%mm1, %%mm4) |
115 PMINUB(%%mm1, %%mm3, %%mm5) | 115 PMINUB(%%mm1, %%mm3, %%mm5) |
116 "psubb %%mm1, %%mm2 \n\t" | 116 "psubb %%mm1, %%mm2 \n\t" |
117 "paddb %%mm7, %%mm2 \n\t" | 117 "paddb %%mm7, %%mm2 \n\t" |
118 "pcmpgtb %%mm6, %%mm2 \n\t" | 118 "pcmpgtb %%mm6, %%mm2 \n\t" |
119 "paddb %%mm2, %%mm0 \n\t" | 119 "paddb %%mm2, %%mm0 \n\t" |
120 | 120 |
121 "movq (%%"REG_a", %3), %%mm2 \n\t" | 121 "movq (%%"REG_a", %3), %%mm2 \n\t" |
122 PMAXUB(%%mm2, %%mm4) | 122 PMAXUB(%%mm2, %%mm4) |
123 PMINUB(%%mm2, %%mm3, %%mm5) | 123 PMINUB(%%mm2, %%mm3, %%mm5) |
124 "psubb %%mm2, %%mm1 \n\t" | 124 "psubb %%mm2, %%mm1 \n\t" |
125 "paddb %%mm7, %%mm1 \n\t" | 125 "paddb %%mm7, %%mm1 \n\t" |
126 "pcmpgtb %%mm6, %%mm1 \n\t" | 126 "pcmpgtb %%mm6, %%mm1 \n\t" |
127 "paddb %%mm1, %%mm0 \n\t" | 127 "paddb %%mm1, %%mm0 \n\t" |
128 | 128 |
129 "movq (%%"REG_a", %3, 2), %%mm1 \n\t" | 129 "movq (%%"REG_a", %3, 2), %%mm1 \n\t" |
130 PMAXUB(%%mm1, %%mm4) | 130 PMAXUB(%%mm1, %%mm4) |
131 PMINUB(%%mm1, %%mm3, %%mm5) | 131 PMINUB(%%mm1, %%mm3, %%mm5) |
132 "psubb %%mm1, %%mm2 \n\t" | 132 "psubb %%mm1, %%mm2 \n\t" |
133 "paddb %%mm7, %%mm2 \n\t" | 133 "paddb %%mm7, %%mm2 \n\t" |
134 "pcmpgtb %%mm6, %%mm2 \n\t" | 134 "pcmpgtb %%mm6, %%mm2 \n\t" |
135 "paddb %%mm2, %%mm0 \n\t" | 135 "paddb %%mm2, %%mm0 \n\t" |
136 "psubusb %%mm3, %%mm4 \n\t" | 136 "psubusb %%mm3, %%mm4 \n\t" |
137 | 137 |
138 " \n\t" | 138 " \n\t" |
139 #ifdef HAVE_MMX2 | 139 #ifdef HAVE_MMX2 |
140 "pxor %%mm7, %%mm7 \n\t" | 140 "pxor %%mm7, %%mm7 \n\t" |
141 "psadbw %%mm7, %%mm0 \n\t" | 141 "psadbw %%mm7, %%mm0 \n\t" |
142 #else | 142 #else |
143 "movq %%mm0, %%mm1 \n\t" | 143 "movq %%mm0, %%mm1 \n\t" |
144 "psrlw $8, %%mm0 \n\t" | 144 "psrlw $8, %%mm0 \n\t" |
145 "paddb %%mm1, %%mm0 \n\t" | 145 "paddb %%mm1, %%mm0 \n\t" |
146 "movq %%mm0, %%mm1 \n\t" | 146 "movq %%mm0, %%mm1 \n\t" |
147 "psrlq $16, %%mm0 \n\t" | 147 "psrlq $16, %%mm0 \n\t" |
148 "paddb %%mm1, %%mm0 \n\t" | 148 "paddb %%mm1, %%mm0 \n\t" |
149 "movq %%mm0, %%mm1 \n\t" | 149 "movq %%mm0, %%mm1 \n\t" |
150 "psrlq $32, %%mm0 \n\t" | 150 "psrlq $32, %%mm0 \n\t" |
151 "paddb %%mm1, %%mm0 \n\t" | 151 "paddb %%mm1, %%mm0 \n\t" |
152 #endif | 152 #endif |
153 "movq %4, %%mm7 \n\t" // QP,..., QP | 153 "movq %4, %%mm7 \n\t" // QP,..., QP |
154 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP | 154 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
155 "psubusb %%mm7, %%mm4 \n\t" // Diff <= 2QP -> 0 | 155 "psubusb %%mm7, %%mm4 \n\t" // Diff <= 2QP -> 0 |
156 "packssdw %%mm4, %%mm4 \n\t" | 156 "packssdw %%mm4, %%mm4 \n\t" |
157 "movd %%mm0, %0 \n\t" | 157 "movd %%mm0, %0 \n\t" |
158 "movd %%mm4, %1 \n\t" | 158 "movd %%mm4, %1 \n\t" |
159 | 159 |
160 : "=r" (numEq), "=r" (dcOk) | 160 : "=r" (numEq), "=r" (dcOk) |
161 : "r" (src), "r" ((long)stride), "m" (c->pQPb) | 161 : "r" (src), "r" ((long)stride), "m" (c->pQPb) |
162 : "%"REG_a | 162 : "%"REG_a |
163 ); | 163 ); |
164 | 164 |
165 numEq= (-numEq) &0xFF; | 165 numEq= (-numEq) &0xFF; |
166 if(numEq > c->ppMode.flatnessThreshold){ | 166 if(numEq > c->ppMode.flatnessThreshold){ |
167 if(dcOk) return 0; | 167 if(dcOk) return 0; |
168 else return 1; | 168 else return 1; |
169 }else{ | 169 }else{ |
170 return 2; | 170 return 2; |
171 } | 171 } |
172 } | 172 } |
173 #endif //HAVE_MMX | 173 #endif //HAVE_MMX |
174 | 174 |
175 /** | 175 /** |
176 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle) | 176 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle) |
178 */ | 178 */ |
179 #ifndef HAVE_ALTIVEC | 179 #ifndef HAVE_ALTIVEC |
180 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c) | 180 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c) |
181 { | 181 { |
182 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 182 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
183 src+= stride*3; | 183 src+= stride*3; |
184 asm volatile( //"movv %0 %1 %2\n\t" | 184 asm volatile( //"movv %0 %1 %2\n\t" |
185 "movq %2, %%mm0 \n\t" // QP,..., QP | 185 "movq %2, %%mm0 \n\t" // QP,..., QP |
186 "pxor %%mm4, %%mm4 \n\t" | 186 "pxor %%mm4, %%mm4 \n\t" |
187 | 187 |
188 "movq (%0), %%mm6 \n\t" | 188 "movq (%0), %%mm6 \n\t" |
189 "movq (%0, %1), %%mm5 \n\t" | 189 "movq (%0, %1), %%mm5 \n\t" |
190 "movq %%mm5, %%mm1 \n\t" | 190 "movq %%mm5, %%mm1 \n\t" |
191 "movq %%mm6, %%mm2 \n\t" | 191 "movq %%mm6, %%mm2 \n\t" |
192 "psubusb %%mm6, %%mm5 \n\t" | 192 "psubusb %%mm6, %%mm5 \n\t" |
193 "psubusb %%mm1, %%mm2 \n\t" | 193 "psubusb %%mm1, %%mm2 \n\t" |
194 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines | 194 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
195 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 | 195 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
196 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF | 196 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF |
197 | 197 |
198 "pand %%mm2, %%mm6 \n\t" | 198 "pand %%mm2, %%mm6 \n\t" |
199 "pandn %%mm1, %%mm2 \n\t" | 199 "pandn %%mm1, %%mm2 \n\t" |
200 "por %%mm2, %%mm6 \n\t"// First Line to Filter | 200 "por %%mm2, %%mm6 \n\t"// First Line to Filter |
201 | 201 |
202 "movq (%0, %1, 8), %%mm5 \n\t" | 202 "movq (%0, %1, 8), %%mm5 \n\t" |
203 "lea (%0, %1, 4), %%"REG_a" \n\t" | 203 "lea (%0, %1, 4), %%"REG_a" \n\t" |
204 "lea (%0, %1, 8), %%"REG_c" \n\t" | 204 "lea (%0, %1, 8), %%"REG_c" \n\t" |
205 "sub %1, %%"REG_c" \n\t" | 205 "sub %1, %%"REG_c" \n\t" |
206 "add %1, %0 \n\t" // %0 points to line 1 not 0 | 206 "add %1, %0 \n\t" // %0 points to line 1 not 0 |
207 "movq (%0, %1, 8), %%mm7 \n\t" | 207 "movq (%0, %1, 8), %%mm7 \n\t" |
208 "movq %%mm5, %%mm1 \n\t" | 208 "movq %%mm5, %%mm1 \n\t" |
209 "movq %%mm7, %%mm2 \n\t" | 209 "movq %%mm7, %%mm2 \n\t" |
210 "psubusb %%mm7, %%mm5 \n\t" | 210 "psubusb %%mm7, %%mm5 \n\t" |
211 "psubusb %%mm1, %%mm2 \n\t" | 211 "psubusb %%mm1, %%mm2 \n\t" |
212 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines | 212 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
213 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 | 213 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
214 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF | 214 "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF |
215 | 215 |
216 "pand %%mm2, %%mm7 \n\t" | 216 "pand %%mm2, %%mm7 \n\t" |
217 "pandn %%mm1, %%mm2 \n\t" | 217 "pandn %%mm1, %%mm2 \n\t" |
218 "por %%mm2, %%mm7 \n\t" // First Line to Filter | 218 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
219 | 219 |
220 | 220 |
221 // 1 2 3 4 5 6 7 8 | 221 // 1 2 3 4 5 6 7 8 |
222 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1 | 222 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1 |
223 // 6 4 2 2 1 1 | 223 // 6 4 2 2 1 1 |
224 // 6 4 4 2 | 224 // 6 4 4 2 |
225 // 6 8 2 | 225 // 6 8 2 |
226 | 226 |
227 "movq (%0, %1), %%mm0 \n\t" // 1 | 227 "movq (%0, %1), %%mm0 \n\t" // 1 |
228 "movq %%mm0, %%mm1 \n\t" // 1 | 228 "movq %%mm0, %%mm1 \n\t" // 1 |
229 PAVGB(%%mm6, %%mm0) //1 1 /2 | 229 PAVGB(%%mm6, %%mm0) //1 1 /2 |
230 PAVGB(%%mm6, %%mm0) //3 1 /4 | 230 PAVGB(%%mm6, %%mm0) //3 1 /4 |
231 | 231 |
232 "movq (%0, %1, 4), %%mm2 \n\t" // 1 | 232 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
233 "movq %%mm2, %%mm5 \n\t" // 1 | 233 "movq %%mm2, %%mm5 \n\t" // 1 |
234 PAVGB((%%REGa), %%mm2) // 11 /2 | 234 PAVGB((%%REGa), %%mm2) // 11 /2 |
235 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | 235 PAVGB((%0, %1, 2), %%mm2) // 211 /4 |
236 "movq %%mm2, %%mm3 \n\t" // 211 /4 | 236 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
237 "movq (%0), %%mm4 \n\t" // 1 | 237 "movq (%0), %%mm4 \n\t" // 1 |
238 PAVGB(%%mm4, %%mm3) // 4 211 /8 | 238 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
239 PAVGB(%%mm0, %%mm3) //642211 /16 | 239 PAVGB(%%mm0, %%mm3) //642211 /16 |
240 "movq %%mm3, (%0) \n\t" // X | 240 "movq %%mm3, (%0) \n\t" // X |
241 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9 | 241 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9 |
242 "movq %%mm1, %%mm0 \n\t" // 1 | 242 "movq %%mm1, %%mm0 \n\t" // 1 |
243 PAVGB(%%mm6, %%mm0) //1 1 /2 | 243 PAVGB(%%mm6, %%mm0) //1 1 /2 |
244 "movq %%mm4, %%mm3 \n\t" // 1 | 244 "movq %%mm4, %%mm3 \n\t" // 1 |
245 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 | 245 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 |
246 PAVGB((%%REGa,%1,2), %%mm5) // 11 /2 | 246 PAVGB((%%REGa,%1,2), %%mm5) // 11 /2 |
247 PAVGB((%%REGa), %%mm5) // 211 /4 | 247 PAVGB((%%REGa), %%mm5) // 211 /4 |
248 PAVGB(%%mm5, %%mm3) // 2 2211 /8 | 248 PAVGB(%%mm5, %%mm3) // 2 2211 /8 |
249 PAVGB(%%mm0, %%mm3) //4242211 /16 | 249 PAVGB(%%mm0, %%mm3) //4242211 /16 |
250 "movq %%mm3, (%0,%1) \n\t" // X | 250 "movq %%mm3, (%0,%1) \n\t" // X |
251 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 | 251 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
252 PAVGB(%%mm4, %%mm6) //11 /2 | 252 PAVGB(%%mm4, %%mm6) //11 /2 |
253 "movq (%%"REG_c"), %%mm0 \n\t" // 1 | 253 "movq (%%"REG_c"), %%mm0 \n\t" // 1 |
254 PAVGB((%%REGa, %1, 2), %%mm0) // 11/2 | 254 PAVGB((%%REGa, %1, 2), %%mm0) // 11/2 |
255 "movq %%mm0, %%mm3 \n\t" // 11/2 | 255 "movq %%mm0, %%mm3 \n\t" // 11/2 |
256 PAVGB(%%mm1, %%mm0) // 2 11/4 | 256 PAVGB(%%mm1, %%mm0) // 2 11/4 |
257 PAVGB(%%mm6, %%mm0) //222 11/8 | 257 PAVGB(%%mm6, %%mm0) //222 11/8 |
258 PAVGB(%%mm2, %%mm0) //22242211/16 | 258 PAVGB(%%mm2, %%mm0) //22242211/16 |
259 "movq (%0, %1, 2), %%mm2 \n\t" // 1 | 259 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
260 "movq %%mm0, (%0, %1, 2) \n\t" // X | 260 "movq %%mm0, (%0, %1, 2) \n\t" // X |
261 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9 | 261 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9 |
262 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1 | 262 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1 |
263 PAVGB((%%REGc), %%mm0) // 11 /2 | 263 PAVGB((%%REGc), %%mm0) // 11 /2 |
264 PAVGB(%%mm0, %%mm6) //11 11 /4 | 264 PAVGB(%%mm0, %%mm6) //11 11 /4 |
265 PAVGB(%%mm1, %%mm4) // 11 /2 | 265 PAVGB(%%mm1, %%mm4) // 11 /2 |
266 PAVGB(%%mm2, %%mm1) // 11 /2 | 266 PAVGB(%%mm2, %%mm1) // 11 /2 |
267 PAVGB(%%mm1, %%mm6) //1122 11 /8 | 267 PAVGB(%%mm1, %%mm6) //1122 11 /8 |
268 PAVGB(%%mm5, %%mm6) //112242211 /16 | 268 PAVGB(%%mm5, %%mm6) //112242211 /16 |
269 "movq (%%"REG_a"), %%mm5 \n\t" // 1 | 269 "movq (%%"REG_a"), %%mm5 \n\t" // 1 |
270 "movq %%mm6, (%%"REG_a") \n\t" // X | 270 "movq %%mm6, (%%"REG_a") \n\t" // X |
271 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9 | 271 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9 |
272 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" // 1 | 272 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" // 1 |
273 PAVGB(%%mm7, %%mm6) // 11 /2 | 273 PAVGB(%%mm7, %%mm6) // 11 /2 |
274 PAVGB(%%mm4, %%mm6) // 11 11 /4 | 274 PAVGB(%%mm4, %%mm6) // 11 11 /4 |
275 PAVGB(%%mm3, %%mm6) // 11 2211 /8 | 275 PAVGB(%%mm3, %%mm6) // 11 2211 /8 |
276 PAVGB(%%mm5, %%mm2) // 11 /2 | 276 PAVGB(%%mm5, %%mm2) // 11 /2 |
277 "movq (%0, %1, 4), %%mm4 \n\t" // 1 | 277 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
278 PAVGB(%%mm4, %%mm2) // 112 /4 | 278 PAVGB(%%mm4, %%mm2) // 112 /4 |
279 PAVGB(%%mm2, %%mm6) // 112242211 /16 | 279 PAVGB(%%mm2, %%mm6) // 112242211 /16 |
280 "movq %%mm6, (%0, %1, 4) \n\t" // X | 280 "movq %%mm6, (%0, %1, 4) \n\t" // X |
281 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 | 281 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
282 PAVGB(%%mm7, %%mm1) // 11 2 /4 | 282 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
283 PAVGB(%%mm4, %%mm5) // 11 /2 | 283 PAVGB(%%mm4, %%mm5) // 11 /2 |
284 PAVGB(%%mm5, %%mm0) // 11 11 /4 | 284 PAVGB(%%mm5, %%mm0) // 11 11 /4 |
285 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" // 1 | 285 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" // 1 |
286 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 | 286 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
287 PAVGB(%%mm0, %%mm1) // 11224222 /16 | 287 PAVGB(%%mm0, %%mm1) // 11224222 /16 |
288 "movq %%mm1, (%%"REG_a", %1, 2) \n\t" // X | 288 "movq %%mm1, (%%"REG_a", %1, 2) \n\t" // X |
289 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 | 289 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
290 PAVGB((%%REGc), %%mm2) // 112 4 /8 | 290 PAVGB((%%REGc), %%mm2) // 112 4 /8 |
291 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1 | 291 "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1 |
292 PAVGB(%%mm0, %%mm6) // 1 1 /2 | 292 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
293 PAVGB(%%mm7, %%mm6) // 1 12 /4 | 293 PAVGB(%%mm7, %%mm6) // 1 12 /4 |
294 PAVGB(%%mm2, %%mm6) // 1122424 /4 | 294 PAVGB(%%mm2, %%mm6) // 1122424 /4 |
295 "movq %%mm6, (%%"REG_c") \n\t" // X | 295 "movq %%mm6, (%%"REG_c") \n\t" // X |
296 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 | 296 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
297 PAVGB(%%mm7, %%mm5) // 11 2 /4 | 297 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
298 PAVGB(%%mm7, %%mm5) // 11 6 /8 | 298 PAVGB(%%mm7, %%mm5) // 11 6 /8 |
299 | 299 |
300 PAVGB(%%mm3, %%mm0) // 112 /4 | 300 PAVGB(%%mm3, %%mm0) // 112 /4 |
301 PAVGB(%%mm0, %%mm5) // 112246 /16 | 301 PAVGB(%%mm0, %%mm5) // 112246 /16 |
302 "movq %%mm5, (%%"REG_a", %1, 4) \n\t" // X | 302 "movq %%mm5, (%%"REG_a", %1, 4) \n\t" // X |
303 "sub %1, %0 \n\t" | 303 "sub %1, %0 \n\t" |
304 | 304 |
305 : | 305 : |
306 : "r" (src), "r" ((long)stride), "m" (c->pQPb) | 306 : "r" (src), "r" ((long)stride), "m" (c->pQPb) |
307 : "%"REG_a, "%"REG_c | 307 : "%"REG_a, "%"REG_c |
308 ); | 308 ); |
309 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 309 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
310 const int l1= stride; | 310 const int l1= stride; |
311 const int l2= stride + l1; | 311 const int l2= stride + l1; |
312 const int l3= stride + l2; | 312 const int l3= stride + l2; |
313 const int l4= stride + l3; | 313 const int l4= stride + l3; |
314 const int l5= stride + l4; | 314 const int l5= stride + l4; |
315 const int l6= stride + l5; | 315 const int l6= stride + l5; |
316 const int l7= stride + l6; | 316 const int l7= stride + l6; |
317 const int l8= stride + l7; | 317 const int l8= stride + l7; |
318 const int l9= stride + l8; | 318 const int l9= stride + l8; |
319 int x; | 319 int x; |
320 src+= stride*3; | 320 src+= stride*3; |
321 for(x=0; x<BLOCK_SIZE; x++) | 321 for(x=0; x<BLOCK_SIZE; x++){ |
322 { | 322 const int first= FFABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1]; |
323 const int first= FFABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1]; | 323 const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8]; |
324 const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8]; | 324 |
325 | 325 int sums[10]; |
326 int sums[10]; | 326 sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4; |
327 sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4; | 327 sums[1] = sums[0] - first + src[l4]; |
328 sums[1] = sums[0] - first + src[l4]; | 328 sums[2] = sums[1] - first + src[l5]; |
329 sums[2] = sums[1] - first + src[l5]; | 329 sums[3] = sums[2] - first + src[l6]; |
330 sums[3] = sums[2] - first + src[l6]; | 330 sums[4] = sums[3] - first + src[l7]; |
331 sums[4] = sums[3] - first + src[l7]; | 331 sums[5] = sums[4] - src[l1] + src[l8]; |
332 sums[5] = sums[4] - src[l1] + src[l8]; | 332 sums[6] = sums[5] - src[l2] + last; |
333 sums[6] = sums[5] - src[l2] + last; | 333 sums[7] = sums[6] - src[l3] + last; |
334 sums[7] = sums[6] - src[l3] + last; | 334 sums[8] = sums[7] - src[l4] + last; |
335 sums[8] = sums[7] - src[l4] + last; | 335 sums[9] = sums[8] - src[l5] + last; |
336 sums[9] = sums[8] - src[l5] + last; | 336 |
337 | 337 src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4; |
338 src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4; | 338 src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4; |
339 src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4; | 339 src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4; |
340 src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4; | 340 src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4; |
341 src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4; | 341 src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4; |
342 src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4; | 342 src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4; |
343 src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4; | 343 src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4; |
344 src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4; | 344 src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4; |
345 src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4; | 345 |
346 | 346 src++; |
347 src++; | 347 } |
348 } | |
349 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 348 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
350 } | 349 } |
351 #endif //HAVE_ALTIVEC | 350 #endif //HAVE_ALTIVEC |
352 | 351 |
353 #if 0 | 352 #if 0 |
364 1 12 12 23 | 363 1 12 12 23 |
365 */ | 364 */ |
366 static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP) | 365 static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP) |
367 { | 366 { |
368 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 367 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
369 src+= stride*3; | 368 src+= stride*3; |
370 // FIXME rounding | 369 // FIXME rounding |
371 asm volatile( | 370 asm volatile( |
372 "pxor %%mm7, %%mm7 \n\t" // 0 | 371 "pxor %%mm7, %%mm7 \n\t" // 0 |
373 "movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE | 372 "movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE |
374 "leal (%0, %1), %%"REG_a" \n\t" | 373 "leal (%0, %1), %%"REG_a" \n\t" |
375 "leal (%%"REG_a", %1, 4), %%"REG_c" \n\t" | 374 "leal (%%"REG_a", %1, 4), %%"REG_c" \n\t" |
376 // 0 1 2 3 4 5 6 7 8 9 | 375 // 0 1 2 3 4 5 6 7 8 9 |
377 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1 | 376 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1 |
378 "movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP | 377 "movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP |
379 "movq %%mm0, %%mm1 \n\t" // QP,..., QP | 378 "movq %%mm0, %%mm1 \n\t" // QP,..., QP |
380 "paddusb "MANGLE(b02)", %%mm0 \n\t" | 379 "paddusb "MANGLE(b02)", %%mm0 \n\t" |
381 "psrlw $2, %%mm0 \n\t" | 380 "psrlw $2, %%mm0 \n\t" |
382 "pand "MANGLE(b3F)", %%mm0 \n\t" // QP/4,..., QP/4 | 381 "pand "MANGLE(b3F)", %%mm0 \n\t" // QP/4,..., QP/4 |
383 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... | 382 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... |
384 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 | 383 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 |
385 "movq (%%"REG_c"), %%mm3 \n\t" // line 5 | 384 "movq (%%"REG_c"), %%mm3 \n\t" // line 5 |
386 "movq %%mm2, %%mm4 \n\t" // line 4 | 385 "movq %%mm2, %%mm4 \n\t" // line 4 |
387 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 | 386 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 |
388 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 | 387 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 |
389 PAVGB(%%mm3, %%mm5) | 388 PAVGB(%%mm3, %%mm5) |
390 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 | 389 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 |
391 "psubusb %%mm3, %%mm4 \n\t" | 390 "psubusb %%mm3, %%mm4 \n\t" |
392 "psubusb %%mm2, %%mm3 \n\t" | 391 "psubusb %%mm2, %%mm3 \n\t" |
393 "por %%mm3, %%mm4 \n\t" // |l4 - l5| | 392 "por %%mm3, %%mm4 \n\t" // |l4 - l5| |
394 "psubusb %%mm0, %%mm4 \n\t" | 393 "psubusb %%mm0, %%mm4 \n\t" |
395 "pcmpeqb %%mm7, %%mm4 \n\t" | 394 "pcmpeqb %%mm7, %%mm4 \n\t" |
396 "pand %%mm4, %%mm5 \n\t" // d/2 | 395 "pand %%mm4, %%mm5 \n\t" // d/2 |
397 | 396 |
398 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 | 397 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 |
399 "paddb %%mm5, %%mm2 \n\t" | 398 "paddb %%mm5, %%mm2 \n\t" |
400 // "psubb %%mm6, %%mm2 \n\t" | 399 // "psubb %%mm6, %%mm2 \n\t" |
401 "movq %%mm2, (%0,%1, 4) \n\t" | 400 "movq %%mm2, (%0,%1, 4) \n\t" |
402 | 401 |
403 "movq (%%"REG_c"), %%mm2 \n\t" | 402 "movq (%%"REG_c"), %%mm2 \n\t" |
404 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 | 403 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 |
405 "psubb %%mm5, %%mm2 \n\t" | 404 "psubb %%mm5, %%mm2 \n\t" |
406 // "psubb %%mm6, %%mm2 \n\t" | 405 // "psubb %%mm6, %%mm2 \n\t" |
407 "movq %%mm2, (%%"REG_c") \n\t" | 406 "movq %%mm2, (%%"REG_c") \n\t" |
408 | 407 |
409 "paddb %%mm6, %%mm5 \n\t" | 408 "paddb %%mm6, %%mm5 \n\t" |
410 "psrlw $2, %%mm5 \n\t" | 409 "psrlw $2, %%mm5 \n\t" |
411 "pand "MANGLE(b3F)", %%mm5 \n\t" | 410 "pand "MANGLE(b3F)", %%mm5 \n\t" |
412 "psubb "MANGLE(b20)", %%mm5 \n\t" // (l5-l4)/8 | 411 "psubb "MANGLE(b20)", %%mm5 \n\t" // (l5-l4)/8 |
413 | 412 |
414 "movq (%%"REG_a", %1, 2), %%mm2 \n\t" | 413 "movq (%%"REG_a", %1, 2), %%mm2 \n\t" |
415 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 | 414 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 |
416 "paddsb %%mm5, %%mm2 \n\t" | 415 "paddsb %%mm5, %%mm2 \n\t" |
417 "psubb %%mm6, %%mm2 \n\t" | 416 "psubb %%mm6, %%mm2 \n\t" |
418 "movq %%mm2, (%%"REG_a", %1, 2) \n\t" | 417 "movq %%mm2, (%%"REG_a", %1, 2) \n\t" |
419 | 418 |
420 "movq (%%"REG_c", %1), %%mm2 \n\t" | 419 "movq (%%"REG_c", %1), %%mm2 \n\t" |
421 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 | 420 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 |
422 "psubsb %%mm5, %%mm2 \n\t" | 421 "psubsb %%mm5, %%mm2 \n\t" |
423 "psubb %%mm6, %%mm2 \n\t" | 422 "psubb %%mm6, %%mm2 \n\t" |
424 "movq %%mm2, (%%"REG_c", %1) \n\t" | 423 "movq %%mm2, (%%"REG_c", %1) \n\t" |
425 | 424 |
426 : | 425 : |
427 : "r" (src), "r" ((long)stride) | 426 : "r" (src), "r" ((long)stride) |
428 : "%"REG_a, "%"REG_c | 427 : "%"REG_a, "%"REG_c |
429 ); | 428 ); |
430 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 429 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
431 const int l1= stride; | 430 const int l1= stride; |
432 const int l2= stride + l1; | 431 const int l2= stride + l1; |
433 const int l3= stride + l2; | 432 const int l3= stride + l2; |
434 const int l4= stride + l3; | 433 const int l4= stride + l3; |
435 const int l5= stride + l4; | 434 const int l5= stride + l4; |
436 const int l6= stride + l5; | 435 const int l6= stride + l5; |
437 // const int l7= stride + l6; | 436 // const int l7= stride + l6; |
438 // const int l8= stride + l7; | 437 // const int l8= stride + l7; |
439 // const int l9= stride + l8; | 438 // const int l9= stride + l8; |
440 int x; | 439 int x; |
441 const int QP15= QP + (QP>>2); | 440 const int QP15= QP + (QP>>2); |
442 src+= stride*3; | 441 src+= stride*3; |
443 for(x=0; x<BLOCK_SIZE; x++) | 442 for(x=0; x<BLOCK_SIZE; x++){ |
444 { | 443 const int v = (src[x+l5] - src[x+l4]); |
445 const int v = (src[x+l5] - src[x+l4]); | 444 if(FFABS(v) < QP15){ |
446 if(FFABS(v) < QP15) | 445 src[x+l3] +=v>>3; |
447 { | 446 src[x+l4] +=v>>1; |
448 src[x+l3] +=v>>3; | 447 src[x+l5] -=v>>1; |
449 src[x+l4] +=v>>1; | 448 src[x+l6] -=v>>3; |
450 src[x+l5] -=v>>1; | |
451 src[x+l6] -=v>>3; | |
452 | |
453 } | |
454 } | 449 } |
450 } | |
455 | 451 |
456 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 452 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
457 } | 453 } |
458 #endif //0 | 454 #endif //0 |
459 | 455 |
465 * MMX2 version does correct clipping C version does not | 461 * MMX2 version does correct clipping C version does not |
466 */ | 462 */ |
467 static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co) | 463 static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co) |
468 { | 464 { |
469 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 465 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
470 src+= stride*3; | 466 src+= stride*3; |
471 | 467 |
472 asm volatile( | 468 asm volatile( |
473 "pxor %%mm7, %%mm7 \n\t" // 0 | 469 "pxor %%mm7, %%mm7 \n\t" // 0 |
474 "lea (%0, %1), %%"REG_a" \n\t" | 470 "lea (%0, %1), %%"REG_a" \n\t" |
475 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" | 471 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" |
476 // 0 1 2 3 4 5 6 7 8 9 | 472 // 0 1 2 3 4 5 6 7 8 9 |
477 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1 | 473 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1 |
478 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3 | 474 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3 |
479 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 | 475 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 |
480 "movq %%mm1, %%mm2 \n\t" // line 4 | 476 "movq %%mm1, %%mm2 \n\t" // line 4 |
481 "psubusb %%mm0, %%mm1 \n\t" | 477 "psubusb %%mm0, %%mm1 \n\t" |
482 "psubusb %%mm2, %%mm0 \n\t" | 478 "psubusb %%mm2, %%mm0 \n\t" |
483 "por %%mm1, %%mm0 \n\t" // |l2 - l3| | 479 "por %%mm1, %%mm0 \n\t" // |l2 - l3| |
484 "movq (%%"REG_c"), %%mm3 \n\t" // line 5 | 480 "movq (%%"REG_c"), %%mm3 \n\t" // line 5 |
485 "movq (%%"REG_c", %1), %%mm4 \n\t" // line 6 | 481 "movq (%%"REG_c", %1), %%mm4 \n\t" // line 6 |
486 "movq %%mm3, %%mm5 \n\t" // line 5 | 482 "movq %%mm3, %%mm5 \n\t" // line 5 |
487 "psubusb %%mm4, %%mm3 \n\t" | 483 "psubusb %%mm4, %%mm3 \n\t" |
488 "psubusb %%mm5, %%mm4 \n\t" | 484 "psubusb %%mm5, %%mm4 \n\t" |
489 "por %%mm4, %%mm3 \n\t" // |l5 - l6| | 485 "por %%mm4, %%mm3 \n\t" // |l5 - l6| |
490 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 | 486 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 |
491 "movq %%mm2, %%mm1 \n\t" // line 4 | 487 "movq %%mm2, %%mm1 \n\t" // line 4 |
492 "psubusb %%mm5, %%mm2 \n\t" | 488 "psubusb %%mm5, %%mm2 \n\t" |
493 "movq %%mm2, %%mm4 \n\t" | 489 "movq %%mm2, %%mm4 \n\t" |
494 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 | 490 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 |
495 "psubusb %%mm1, %%mm5 \n\t" | 491 "psubusb %%mm1, %%mm5 \n\t" |
496 "por %%mm5, %%mm4 \n\t" // |l4 - l5| | 492 "por %%mm5, %%mm4 \n\t" // |l4 - l5| |
497 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2) | 493 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2) |
498 "movq %%mm4, %%mm3 \n\t" // d | 494 "movq %%mm4, %%mm3 \n\t" // d |
499 "movq %2, %%mm0 \n\t" | 495 "movq %2, %%mm0 \n\t" |
500 "paddusb %%mm0, %%mm0 \n\t" | 496 "paddusb %%mm0, %%mm0 \n\t" |
501 "psubusb %%mm0, %%mm4 \n\t" | 497 "psubusb %%mm0, %%mm4 \n\t" |
502 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 | 498 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 |
503 "psubusb "MANGLE(b01)", %%mm3 \n\t" | 499 "psubusb "MANGLE(b01)", %%mm3 \n\t" |
504 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 | 500 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 |
505 | 501 |
506 PAVGB(%%mm7, %%mm3) // d/2 | 502 PAVGB(%%mm7, %%mm3) // d/2 |
507 "movq %%mm3, %%mm1 \n\t" // d/2 | 503 "movq %%mm3, %%mm1 \n\t" // d/2 |
508 PAVGB(%%mm7, %%mm3) // d/4 | 504 PAVGB(%%mm7, %%mm3) // d/4 |
509 PAVGB(%%mm1, %%mm3) // 3*d/8 | 505 PAVGB(%%mm1, %%mm3) // 3*d/8 |
510 | 506 |
511 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 | 507 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 |
512 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 | 508 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
513 "psubusb %%mm3, %%mm0 \n\t" | 509 "psubusb %%mm3, %%mm0 \n\t" |
514 "pxor %%mm2, %%mm0 \n\t" | 510 "pxor %%mm2, %%mm0 \n\t" |
515 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 | 511 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 |
516 | 512 |
517 "movq (%%"REG_c"), %%mm0 \n\t" // line 5 | 513 "movq (%%"REG_c"), %%mm0 \n\t" // line 5 |
518 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 | 514 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
519 "paddusb %%mm3, %%mm0 \n\t" | 515 "paddusb %%mm3, %%mm0 \n\t" |
520 "pxor %%mm2, %%mm0 \n\t" | 516 "pxor %%mm2, %%mm0 \n\t" |
521 "movq %%mm0, (%%"REG_c") \n\t" // line 5 | 517 "movq %%mm0, (%%"REG_c") \n\t" // line 5 |
522 | 518 |
523 PAVGB(%%mm7, %%mm1) // d/4 | 519 PAVGB(%%mm7, %%mm1) // d/4 |
524 | 520 |
525 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3 | 521 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3 |
526 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 | 522 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
527 "psubusb %%mm1, %%mm0 \n\t" | 523 "psubusb %%mm1, %%mm0 \n\t" |
528 "pxor %%mm2, %%mm0 \n\t" | 524 "pxor %%mm2, %%mm0 \n\t" |
529 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" // line 3 | 525 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" // line 3 |
530 | 526 |
531 "movq (%%"REG_c", %1), %%mm0 \n\t" // line 6 | 527 "movq (%%"REG_c", %1), %%mm0 \n\t" // line 6 |
532 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 | 528 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
533 "paddusb %%mm1, %%mm0 \n\t" | 529 "paddusb %%mm1, %%mm0 \n\t" |
534 "pxor %%mm2, %%mm0 \n\t" | 530 "pxor %%mm2, %%mm0 \n\t" |
535 "movq %%mm0, (%%"REG_c", %1) \n\t" // line 6 | 531 "movq %%mm0, (%%"REG_c", %1) \n\t" // line 6 |
536 | 532 |
537 PAVGB(%%mm7, %%mm1) // d/8 | 533 PAVGB(%%mm7, %%mm1) // d/8 |
538 | 534 |
539 "movq (%%"REG_a", %1), %%mm0 \n\t" // line 2 | 535 "movq (%%"REG_a", %1), %%mm0 \n\t" // line 2 |
540 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 | 536 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 |
541 "psubusb %%mm1, %%mm0 \n\t" | 537 "psubusb %%mm1, %%mm0 \n\t" |
542 "pxor %%mm2, %%mm0 \n\t" | 538 "pxor %%mm2, %%mm0 \n\t" |
543 "movq %%mm0, (%%"REG_a", %1) \n\t" // line 2 | 539 "movq %%mm0, (%%"REG_a", %1) \n\t" // line 2 |
544 | 540 |
545 "movq (%%"REG_c", %1, 2), %%mm0 \n\t" // line 7 | 541 "movq (%%"REG_c", %1, 2), %%mm0 \n\t" // line 7 |
546 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 | 542 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 |
547 "paddusb %%mm1, %%mm0 \n\t" | 543 "paddusb %%mm1, %%mm0 \n\t" |
548 "pxor %%mm2, %%mm0 \n\t" | 544 "pxor %%mm2, %%mm0 \n\t" |
549 "movq %%mm0, (%%"REG_c", %1, 2) \n\t" // line 7 | 545 "movq %%mm0, (%%"REG_c", %1, 2) \n\t" // line 7 |
550 | 546 |
551 : | 547 : |
552 : "r" (src), "r" ((long)stride), "m" (co->pQPb) | 548 : "r" (src), "r" ((long)stride), "m" (co->pQPb) |
553 : "%"REG_a, "%"REG_c | 549 : "%"REG_a, "%"REG_c |
554 ); | 550 ); |
555 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 551 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
556 | 552 |
557 const int l1= stride; | 553 const int l1= stride; |
558 const int l2= stride + l1; | 554 const int l2= stride + l1; |
559 const int l3= stride + l2; | 555 const int l3= stride + l2; |
560 const int l4= stride + l3; | 556 const int l4= stride + l3; |
561 const int l5= stride + l4; | 557 const int l5= stride + l4; |
562 const int l6= stride + l5; | 558 const int l6= stride + l5; |
563 const int l7= stride + l6; | 559 const int l7= stride + l6; |
564 // const int l8= stride + l7; | 560 // const int l8= stride + l7; |
565 // const int l9= stride + l8; | 561 // const int l9= stride + l8; |
566 int x; | 562 int x; |
567 | 563 |
568 src+= stride*3; | 564 src+= stride*3; |
569 for(x=0; x<BLOCK_SIZE; x++) | 565 for(x=0; x<BLOCK_SIZE; x++){ |
570 { | 566 int a= src[l3] - src[l4]; |
571 int a= src[l3] - src[l4]; | 567 int b= src[l4] - src[l5]; |
572 int b= src[l4] - src[l5]; | 568 int c= src[l5] - src[l6]; |
573 int c= src[l5] - src[l6]; | 569 |
574 | 570 int d= FFABS(b) - ((FFABS(a) + FFABS(c))>>1); |
575 int d= FFABS(b) - ((FFABS(a) + FFABS(c))>>1); | 571 d= FFMAX(d, 0); |
576 d= FFMAX(d, 0); | 572 |
577 | 573 if(d < co->QP*2){ |
578 if(d < co->QP*2) | 574 int v = d * FFSIGN(-b); |
579 { | 575 |
580 int v = d * FFSIGN(-b); | 576 src[l2] +=v>>3; |
581 | 577 src[l3] +=v>>2; |
582 src[l2] +=v>>3; | 578 src[l4] +=(3*v)>>3; |
583 src[l3] +=v>>2; | 579 src[l5] -=(3*v)>>3; |
584 src[l4] +=(3*v)>>3; | 580 src[l6] -=v>>2; |
585 src[l5] -=(3*v)>>3; | 581 src[l7] -=v>>3; |
586 src[l6] -=v>>2; | |
587 src[l7] -=v>>3; | |
588 | |
589 } | |
590 src++; | |
591 } | 582 } |
583 src++; | |
584 } | |
592 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 585 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
593 } | 586 } |
594 | 587 |
595 #ifndef HAVE_ALTIVEC | 588 #ifndef HAVE_ALTIVEC |
596 static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c) | 589 static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c) |
597 { | 590 { |
598 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 591 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
599 /* | 592 /* |
600 uint8_t tmp[16]; | 593 uint8_t tmp[16]; |
601 const int l1= stride; | 594 const int l1= stride; |
602 const int l2= stride + l1; | 595 const int l2= stride + l1; |
603 const int l3= stride + l2; | 596 const int l3= stride + l2; |
604 const int l4= (int)tmp - (int)src - stride*3; | 597 const int l4= (int)tmp - (int)src - stride*3; |
605 const int l5= (int)tmp - (int)src - stride*3 + 8; | 598 const int l5= (int)tmp - (int)src - stride*3 + 8; |
606 const int l6= stride*3 + l3; | 599 const int l6= stride*3 + l3; |
607 const int l7= stride + l6; | 600 const int l7= stride + l6; |
608 const int l8= stride + l7; | 601 const int l8= stride + l7; |
609 | 602 |
610 memcpy(tmp, src+stride*7, 8); | 603 memcpy(tmp, src+stride*7, 8); |
611 memcpy(tmp+8, src+stride*8, 8); | 604 memcpy(tmp+8, src+stride*8, 8); |
612 */ | 605 */ |
613 src+= stride*4; | 606 src+= stride*4; |
614 asm volatile( | 607 asm volatile( |
615 | 608 |
616 #if 0 //sligtly more accurate and slightly slower | 609 #if 0 //sligtly more accurate and slightly slower |
617 "pxor %%mm7, %%mm7 \n\t" // 0 | 610 "pxor %%mm7, %%mm7 \n\t" // 0 |
618 "lea (%0, %1), %%"REG_a" \n\t" | 611 "lea (%0, %1), %%"REG_a" \n\t" |
619 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" | 612 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" |
620 // 0 1 2 3 4 5 6 7 | 613 // 0 1 2 3 4 5 6 7 |
621 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1 | 614 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1 |
622 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 | 615 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 |
623 | 616 |
624 | 617 |
625 "movq (%0, %1, 2), %%mm0 \n\t" // l2 | 618 "movq (%0, %1, 2), %%mm0 \n\t" // l2 |
626 "movq (%0), %%mm1 \n\t" // l0 | 619 "movq (%0), %%mm1 \n\t" // l0 |
627 "movq %%mm0, %%mm2 \n\t" // l2 | 620 "movq %%mm0, %%mm2 \n\t" // l2 |
628 PAVGB(%%mm7, %%mm0) // ~l2/2 | 621 PAVGB(%%mm7, %%mm0) // ~l2/2 |
629 PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4 | 622 PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4 |
630 PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8 | 623 PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8 |
631 | 624 |
632 "movq (%%"REG_a"), %%mm1 \n\t" // l1 | 625 "movq (%%"REG_a"), %%mm1 \n\t" // l1 |
633 "movq (%%"REG_a", %1, 2), %%mm3 \n\t" // l3 | 626 "movq (%%"REG_a", %1, 2), %%mm3 \n\t" // l3 |
634 "movq %%mm1, %%mm4 \n\t" // l1 | 627 "movq %%mm1, %%mm4 \n\t" // l1 |
635 PAVGB(%%mm7, %%mm1) // ~l1/2 | 628 PAVGB(%%mm7, %%mm1) // ~l1/2 |
636 PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4 | 629 PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4 |
637 PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8 | 630 PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8 |
638 | 631 |
639 "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8 | 632 "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8 |
640 "psubusb %%mm1, %%mm0 \n\t" | 633 "psubusb %%mm1, %%mm0 \n\t" |
641 "psubusb %%mm4, %%mm1 \n\t" | 634 "psubusb %%mm4, %%mm1 \n\t" |
642 "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8 | 635 "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8 |
643 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0 | 636 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0 |
644 | 637 |
645 "movq (%0, %1, 4), %%mm0 \n\t" // l4 | 638 "movq (%0, %1, 4), %%mm0 \n\t" // l4 |
646 "movq %%mm0, %%mm4 \n\t" // l4 | 639 "movq %%mm0, %%mm4 \n\t" // l4 |
647 PAVGB(%%mm7, %%mm0) // ~l4/2 | 640 PAVGB(%%mm7, %%mm0) // ~l4/2 |
648 PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4 | 641 PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4 |
649 PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8 | 642 PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8 |
650 | 643 |
651 "movq (%%"REG_c"), %%mm2 \n\t" // l5 | 644 "movq (%%"REG_c"), %%mm2 \n\t" // l5 |
652 "movq %%mm3, %%mm5 \n\t" // l3 | 645 "movq %%mm3, %%mm5 \n\t" // l3 |
653 PAVGB(%%mm7, %%mm3) // ~l3/2 | 646 PAVGB(%%mm7, %%mm3) // ~l3/2 |
654 PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4 | 647 PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4 |
655 PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8 | 648 PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8 |
656 | 649 |
657 "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8 | 650 "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8 |
658 "psubusb %%mm3, %%mm0 \n\t" | 651 "psubusb %%mm3, %%mm0 \n\t" |
659 "psubusb %%mm6, %%mm3 \n\t" | 652 "psubusb %%mm6, %%mm3 \n\t" |
660 "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8 | 653 "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8 |
661 "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5) | 654 "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5) |
662 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0 | 655 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0 |
663 | 656 |
664 "movq (%%"REG_c", %1), %%mm6 \n\t" // l6 | 657 "movq (%%"REG_c", %1), %%mm6 \n\t" // l6 |
665 "movq %%mm6, %%mm5 \n\t" // l6 | 658 "movq %%mm6, %%mm5 \n\t" // l6 |
666 PAVGB(%%mm7, %%mm6) // ~l6/2 | 659 PAVGB(%%mm7, %%mm6) // ~l6/2 |
667 PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4 | 660 PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4 |
668 PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8 | 661 PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8 |
669 | 662 |
670 "movq (%%"REG_c", %1, 2), %%mm5 \n\t" // l7 | 663 "movq (%%"REG_c", %1, 2), %%mm5 \n\t" // l7 |
671 "movq %%mm2, %%mm4 \n\t" // l5 | 664 "movq %%mm2, %%mm4 \n\t" // l5 |
672 PAVGB(%%mm7, %%mm2) // ~l5/2 | 665 PAVGB(%%mm7, %%mm2) // ~l5/2 |
673 PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4 | 666 PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4 |
674 PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8 | 667 PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8 |
675 | 668 |
676 "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8 | 669 "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8 |
677 "psubusb %%mm2, %%mm6 \n\t" | 670 "psubusb %%mm2, %%mm6 \n\t" |
678 "psubusb %%mm4, %%mm2 \n\t" | 671 "psubusb %%mm4, %%mm2 \n\t" |
679 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8 | 672 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8 |
680 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0 | 673 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0 |
681 | 674 |
682 | 675 |
683 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8 | 676 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8 |
684 "movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ? | 677 "movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ? |
685 "paddusb "MANGLE(b01)", %%mm4 \n\t" | 678 "paddusb "MANGLE(b01)", %%mm4 \n\t" |
686 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP | 679 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP |
687 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8 | 680 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8 |
688 "pand %%mm4, %%mm3 \n\t" | 681 "pand %%mm4, %%mm3 \n\t" |
689 | 682 |
690 "movq %%mm3, %%mm1 \n\t" | 683 "movq %%mm3, %%mm1 \n\t" |
691 // "psubusb "MANGLE(b01)", %%mm3 \n\t" | 684 // "psubusb "MANGLE(b01)", %%mm3 \n\t" |
692 PAVGB(%%mm7, %%mm3) | 685 PAVGB(%%mm7, %%mm3) |
693 PAVGB(%%mm7, %%mm3) | 686 PAVGB(%%mm7, %%mm3) |
694 "paddusb %%mm1, %%mm3 \n\t" | 687 "paddusb %%mm1, %%mm3 \n\t" |
695 // "paddusb "MANGLE(b01)", %%mm3 \n\t" | 688 // "paddusb "MANGLE(b01)", %%mm3 \n\t" |
696 | 689 |
697 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" //l3 | 690 "movq (%%"REG_a", %1, 2), %%mm6 \n\t" //l3 |
698 "movq (%0, %1, 4), %%mm5 \n\t" //l4 | 691 "movq (%0, %1, 4), %%mm5 \n\t" //l4 |
699 "movq (%0, %1, 4), %%mm4 \n\t" //l4 | 692 "movq (%0, %1, 4), %%mm4 \n\t" //l4 |
700 "psubusb %%mm6, %%mm5 \n\t" | 693 "psubusb %%mm6, %%mm5 \n\t" |
701 "psubusb %%mm4, %%mm6 \n\t" | 694 "psubusb %%mm4, %%mm6 \n\t" |
702 "por %%mm6, %%mm5 \n\t" // |l3-l4| | 695 "por %%mm6, %%mm5 \n\t" // |l3-l4| |
703 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4) | 696 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4) |
704 "pxor %%mm6, %%mm0 \n\t" | 697 "pxor %%mm6, %%mm0 \n\t" |
705 "pand %%mm0, %%mm3 \n\t" | 698 "pand %%mm0, %%mm3 \n\t" |
706 PMINUB(%%mm5, %%mm3, %%mm0) | 699 PMINUB(%%mm5, %%mm3, %%mm0) |
707 | 700 |
708 "psubusb "MANGLE(b01)", %%mm3 \n\t" | 701 "psubusb "MANGLE(b01)", %%mm3 \n\t" |
709 PAVGB(%%mm7, %%mm3) | 702 PAVGB(%%mm7, %%mm3) |
710 | 703 |
711 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" | 704 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" |
712 "movq (%0, %1, 4), %%mm2 \n\t" | 705 "movq (%0, %1, 4), %%mm2 \n\t" |
713 "pxor %%mm6, %%mm0 \n\t" | 706 "pxor %%mm6, %%mm0 \n\t" |
714 "pxor %%mm6, %%mm2 \n\t" | 707 "pxor %%mm6, %%mm2 \n\t" |
715 "psubb %%mm3, %%mm0 \n\t" | 708 "psubb %%mm3, %%mm0 \n\t" |
716 "paddb %%mm3, %%mm2 \n\t" | 709 "paddb %%mm3, %%mm2 \n\t" |
717 "pxor %%mm6, %%mm0 \n\t" | 710 "pxor %%mm6, %%mm0 \n\t" |
718 "pxor %%mm6, %%mm2 \n\t" | 711 "pxor %%mm6, %%mm2 \n\t" |
719 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" | 712 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" |
720 "movq %%mm2, (%0, %1, 4) \n\t" | 713 "movq %%mm2, (%0, %1, 4) \n\t" |
721 #endif //0 | 714 #endif //0 |
722 | 715 |
723 "lea (%0, %1), %%"REG_a" \n\t" | 716 "lea (%0, %1), %%"REG_a" \n\t" |
724 "pcmpeqb %%mm6, %%mm6 \n\t" // -1 | 717 "pcmpeqb %%mm6, %%mm6 \n\t" // -1 |
725 // 0 1 2 3 4 5 6 7 | 718 // 0 1 2 3 4 5 6 7 |
726 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1 | 719 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1 |
727 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 | 720 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 |
728 | 721 |
729 | 722 |
730 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // l3 | 723 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // l3 |
731 "movq (%0, %1, 4), %%mm0 \n\t" // l4 | 724 "movq (%0, %1, 4), %%mm0 \n\t" // l4 |
732 "pxor %%mm6, %%mm1 \n\t" // -l3-1 | 725 "pxor %%mm6, %%mm1 \n\t" // -l3-1 |
733 PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2 | 726 PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2 |
734 // mm1=-l3-1, mm0=128-q | 727 // mm1=-l3-1, mm0=128-q |
735 | 728 |
736 "movq (%%"REG_a", %1, 4), %%mm2 \n\t" // l5 | 729 "movq (%%"REG_a", %1, 4), %%mm2 \n\t" // l5 |
737 "movq (%%"REG_a", %1), %%mm3 \n\t" // l2 | 730 "movq (%%"REG_a", %1), %%mm3 \n\t" // l2 |
738 "pxor %%mm6, %%mm2 \n\t" // -l5-1 | 731 "pxor %%mm6, %%mm2 \n\t" // -l5-1 |
739 "movq %%mm2, %%mm5 \n\t" // -l5-1 | 732 "movq %%mm2, %%mm5 \n\t" // -l5-1 |
740 "movq "MANGLE(b80)", %%mm4 \n\t" // 128 | 733 "movq "MANGLE(b80)", %%mm4 \n\t" // 128 |
741 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" | 734 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" |
742 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2 | 735 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2 |
743 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128 | 736 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128 |
744 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128 | 737 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128 |
745 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128 | 738 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128 |
746 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1 | 739 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1 |
747 | 740 |
748 "movq (%%"REG_a"), %%mm2 \n\t" // l1 | 741 "movq (%%"REG_a"), %%mm2 \n\t" // l1 |
749 "pxor %%mm6, %%mm2 \n\t" // -l1-1 | 742 "pxor %%mm6, %%mm2 \n\t" // -l1-1 |
750 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2 | 743 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2 |
751 PAVGB((%0), %%mm1) // (l0-l3+256)/2 | 744 PAVGB((%0), %%mm1) // (l0-l3+256)/2 |
752 "movq "MANGLE(b80)", %%mm3 \n\t" // 128 | 745 "movq "MANGLE(b80)", %%mm3 \n\t" // 128 |
753 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128 | 746 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128 |
754 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128 | 747 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128 |
755 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128 | 748 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128 |
756 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1 | 749 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1 |
757 | 750 |
758 PAVGB((%%REGc, %1), %%mm5) // (l6-l5+256)/2 | 751 PAVGB((%%REGc, %1), %%mm5) // (l6-l5+256)/2 |
759 "movq (%%"REG_c", %1, 2), %%mm1 \n\t" // l7 | 752 "movq (%%"REG_c", %1, 2), %%mm1 \n\t" // l7 |
760 "pxor %%mm6, %%mm1 \n\t" // -l7-1 | 753 "pxor %%mm6, %%mm1 \n\t" // -l7-1 |
761 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2 | 754 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2 |
762 "movq "MANGLE(b80)", %%mm2 \n\t" // 128 | 755 "movq "MANGLE(b80)", %%mm2 \n\t" // 128 |
763 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128 | 756 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128 |
764 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128 | 757 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128 |
765 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128 | 758 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128 |
766 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128 | 759 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128 |
767 | 760 |
768 "movq "MANGLE(b00)", %%mm1 \n\t" // 0 | 761 "movq "MANGLE(b00)", %%mm1 \n\t" // 0 |
769 "movq "MANGLE(b00)", %%mm5 \n\t" // 0 | 762 "movq "MANGLE(b00)", %%mm5 \n\t" // 0 |
770 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16 | 763 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16 |
771 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16 | 764 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16 |
772 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16| | 765 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16| |
773 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16| | 766 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16| |
774 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16 | 767 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16 |
775 | 768 |
776 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128 | 769 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128 |
777 | 770 |
778 "movq "MANGLE(b00)", %%mm7 \n\t" // 0 | 771 "movq "MANGLE(b00)", %%mm7 \n\t" // 0 |
779 "movq %2, %%mm2 \n\t" // QP | 772 "movq %2, %%mm2 \n\t" // QP |
780 PAVGB(%%mm6, %%mm2) // 128 + QP/2 | 773 PAVGB(%%mm6, %%mm2) // 128 + QP/2 |
781 "psubb %%mm6, %%mm2 \n\t" | 774 "psubb %%mm6, %%mm2 \n\t" |
782 | 775 |
783 "movq %%mm4, %%mm1 \n\t" | 776 "movq %%mm4, %%mm1 \n\t" |
784 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy) | 777 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy) |
785 "pxor %%mm1, %%mm4 \n\t" | 778 "pxor %%mm1, %%mm4 \n\t" |
786 "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16 | 779 "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16 |
787 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2 | 780 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2 |
788 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16 | 781 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16 |
789 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16 | 782 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16 |
790 | 783 |
791 "movq %%mm4, %%mm3 \n\t" // d | 784 "movq %%mm4, %%mm3 \n\t" // d |
792 "psubusb "MANGLE(b01)", %%mm4 \n\t" | 785 "psubusb "MANGLE(b01)", %%mm4 \n\t" |
793 PAVGB(%%mm7, %%mm4) // d/32 | 786 PAVGB(%%mm7, %%mm4) // d/32 |
794 PAVGB(%%mm7, %%mm4) // (d + 32)/64 | 787 PAVGB(%%mm7, %%mm4) // (d + 32)/64 |
795 "paddb %%mm3, %%mm4 \n\t" // 5d/64 | 788 "paddb %%mm3, %%mm4 \n\t" // 5d/64 |
796 "pand %%mm2, %%mm4 \n\t" | 789 "pand %%mm2, %%mm4 \n\t" |
797 | 790 |
798 "movq "MANGLE(b80)", %%mm5 \n\t" // 128 | 791 "movq "MANGLE(b80)", %%mm5 \n\t" // 128 |
799 "psubb %%mm0, %%mm5 \n\t" // q | 792 "psubb %%mm0, %%mm5 \n\t" // q |
800 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding | 793 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding |
801 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q) | 794 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q) |
802 "pxor %%mm7, %%mm5 \n\t" | 795 "pxor %%mm7, %%mm5 \n\t" |
803 | 796 |
804 PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64) | 797 PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64) |
805 "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q) | 798 "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q) |
806 | 799 |
807 "pand %%mm7, %%mm4 \n\t" | 800 "pand %%mm7, %%mm4 \n\t" |
808 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" | 801 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" |
809 "movq (%0, %1, 4), %%mm2 \n\t" | 802 "movq (%0, %1, 4), %%mm2 \n\t" |
810 "pxor %%mm1, %%mm0 \n\t" | 803 "pxor %%mm1, %%mm0 \n\t" |
811 "pxor %%mm1, %%mm2 \n\t" | 804 "pxor %%mm1, %%mm2 \n\t" |
812 "paddb %%mm4, %%mm0 \n\t" | 805 "paddb %%mm4, %%mm0 \n\t" |
813 "psubb %%mm4, %%mm2 \n\t" | 806 "psubb %%mm4, %%mm2 \n\t" |
814 "pxor %%mm1, %%mm0 \n\t" | 807 "pxor %%mm1, %%mm0 \n\t" |
815 "pxor %%mm1, %%mm2 \n\t" | 808 "pxor %%mm1, %%mm2 \n\t" |
816 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" | 809 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" |
817 "movq %%mm2, (%0, %1, 4) \n\t" | 810 "movq %%mm2, (%0, %1, 4) \n\t" |
818 | 811 |
819 : | 812 : |
820 : "r" (src), "r" ((long)stride), "m" (c->pQPb) | 813 : "r" (src), "r" ((long)stride), "m" (c->pQPb) |
821 : "%"REG_a, "%"REG_c | 814 : "%"REG_a, "%"REG_c |
822 ); | 815 ); |
823 | 816 |
824 /* | 817 /* |
825 { | 818 { |
826 int x; | 819 int x; |
827 src-= stride; | 820 src-= stride; |
828 for(x=0; x<BLOCK_SIZE; x++) | 821 for(x=0; x<BLOCK_SIZE; x++){ |
829 { | 822 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]); |
830 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]); | 823 if(FFABS(middleEnergy)< 8*QP){ |
831 if(FFABS(middleEnergy)< 8*QP) | 824 const int q=(src[l4] - src[l5])/2; |
832 { | 825 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); |
833 const int q=(src[l4] - src[l5])/2; | 826 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); |
834 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); | 827 |
835 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); | 828 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) ); |
836 | 829 d= FFMAX(d, 0); |
837 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) ); | 830 |
838 d= FFMAX(d, 0); | 831 d= (5*d + 32) >> 6; |
839 | 832 d*= FFSIGN(-middleEnergy); |
840 d= (5*d + 32) >> 6; | 833 |
841 d*= FFSIGN(-middleEnergy); | 834 if(q>0){ |
842 | 835 d= d<0 ? 0 : d; |
843 if(q>0) | 836 d= d>q ? q : d; |
844 { | 837 }else{ |
845 d= d<0 ? 0 : d; | 838 d= d>0 ? 0 : d; |
846 d= d>q ? q : d; | 839 d= d<q ? q : d; |
847 } | 840 } |
848 else | 841 |
849 { | 842 src[l4]-= d; |
850 d= d>0 ? 0 : d; | 843 src[l5]+= d; |
851 d= d<q ? q : d; | |
852 } | |
853 | |
854 src[l4]-= d; | |
855 src[l5]+= d; | |
856 } | |
857 src++; | |
858 } | 844 } |
859 src-=8; | 845 src++; |
860 for(x=0; x<8; x++) | 846 } |
861 { | 847 src-=8; |
862 int y; | 848 for(x=0; x<8; x++){ |
863 for(y=4; y<6; y++) | 849 int y; |
864 { | 850 for(y=4; y<6; y++){ |
865 int d= src[x+y*stride] - tmp[x+(y-4)*8]; | 851 int d= src[x+y*stride] - tmp[x+(y-4)*8]; |
866 int ad= FFABS(d); | 852 int ad= FFABS(d); |
867 static int max=0; | 853 static int max=0; |
868 static int sum=0; | 854 static int sum=0; |
869 static int num=0; | 855 static int num=0; |
870 static int bias=0; | 856 static int bias=0; |
871 | 857 |
872 if(max<ad) max=ad; | 858 if(max<ad) max=ad; |
873 sum+= ad>3 ? 1 : 0; | 859 sum+= ad>3 ? 1 : 0; |
874 if(ad>3) | 860 if(ad>3){ |
875 { | 861 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255; |
876 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255; | 862 } |
877 } | 863 if(y==4) bias+=d; |
878 if(y==4) bias+=d; | 864 num++; |
879 num++; | 865 if(num%1000000 == 0){ |
880 if(num%1000000 == 0) | 866 av_log(c, AV_LOG_INFO, " %d %d %d %d\n", num, sum, max, bias); |
881 { | 867 } |
882 av_log(c, AV_LOG_INFO, " %d %d %d %d\n", num, sum, max, bias); | |
883 } | |
884 } | |
885 } | 868 } |
869 } | |
886 } | 870 } |
887 */ | 871 */ |
888 #elif defined (HAVE_MMX) | 872 #elif defined (HAVE_MMX) |
889 src+= stride*4; | 873 src+= stride*4; |
890 asm volatile( | 874 asm volatile( |
891 "pxor %%mm7, %%mm7 \n\t" | 875 "pxor %%mm7, %%mm7 \n\t" |
892 "lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars | 876 "lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars |
893 "and "ALIGN_MASK", %%"REG_c" \n\t" // align | 877 "and "ALIGN_MASK", %%"REG_c" \n\t" // align |
894 // 0 1 2 3 4 5 6 7 | 878 // 0 1 2 3 4 5 6 7 |
895 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1 | 879 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1 |
896 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 | 880 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 |
897 | 881 |
898 "movq (%0), %%mm0 \n\t" | 882 "movq (%0), %%mm0 \n\t" |
899 "movq %%mm0, %%mm1 \n\t" | 883 "movq %%mm0, %%mm1 \n\t" |
900 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 | 884 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
901 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 | 885 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
902 | 886 |
903 "movq (%0, %1), %%mm2 \n\t" | 887 "movq (%0, %1), %%mm2 \n\t" |
904 "lea (%0, %1, 2), %%"REG_a" \n\t" | 888 "lea (%0, %1, 2), %%"REG_a" \n\t" |
905 "movq %%mm2, %%mm3 \n\t" | 889 "movq %%mm2, %%mm3 \n\t" |
906 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 | 890 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
907 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 | 891 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
908 | 892 |
909 "movq (%%"REG_a"), %%mm4 \n\t" | 893 "movq (%%"REG_a"), %%mm4 \n\t" |
910 "movq %%mm4, %%mm5 \n\t" | 894 "movq %%mm4, %%mm5 \n\t" |
911 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 | 895 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
912 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 | 896 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
913 | 897 |
914 "paddw %%mm0, %%mm0 \n\t" // 2L0 | 898 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
915 "paddw %%mm1, %%mm1 \n\t" // 2H0 | 899 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
916 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 | 900 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
917 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 | 901 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
918 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 | 902 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
919 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 | 903 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
920 | 904 |
921 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 | 905 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
922 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 | 906 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
923 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 | 907 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
924 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 | 908 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
925 | 909 |
926 "movq (%%"REG_a", %1), %%mm2 \n\t" | 910 "movq (%%"REG_a", %1), %%mm2 \n\t" |
927 "movq %%mm2, %%mm3 \n\t" | 911 "movq %%mm2, %%mm3 \n\t" |
928 "punpcklbw %%mm7, %%mm2 \n\t" // L3 | 912 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
929 "punpckhbw %%mm7, %%mm3 \n\t" // H3 | 913 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
930 | 914 |
931 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 | 915 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
932 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 | 916 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
933 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 | 917 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
934 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 | 918 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
935 "movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3 | 919 "movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
936 "movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3 | 920 "movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
937 | 921 |
938 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" | 922 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" |
939 "movq %%mm0, %%mm1 \n\t" | 923 "movq %%mm0, %%mm1 \n\t" |
940 "punpcklbw %%mm7, %%mm0 \n\t" // L4 | 924 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
941 "punpckhbw %%mm7, %%mm1 \n\t" // H4 | 925 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
942 | 926 |
943 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 | 927 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
944 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 | 928 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
945 "movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4 | 929 "movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4 |
946 "movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4 | 930 "movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4 |
947 "paddw %%mm4, %%mm4 \n\t" // 2L2 | 931 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
948 "paddw %%mm5, %%mm5 \n\t" // 2H2 | 932 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
949 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 | 933 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
950 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 | 934 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
951 | 935 |
952 "lea (%%"REG_a", %1), %0 \n\t" | 936 "lea (%%"REG_a", %1), %0 \n\t" |
953 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 | 937 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
954 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 | 938 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
955 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 | 939 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
956 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 | 940 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
957 //50 opcodes so far | 941 //50 opcodes so far |
958 "movq (%0, %1, 2), %%mm2 \n\t" | 942 "movq (%0, %1, 2), %%mm2 \n\t" |
959 "movq %%mm2, %%mm3 \n\t" | 943 "movq %%mm2, %%mm3 \n\t" |
960 "punpcklbw %%mm7, %%mm2 \n\t" // L5 | 944 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
961 "punpckhbw %%mm7, %%mm3 \n\t" // H5 | 945 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
962 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 | 946 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
963 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 | 947 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
964 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 | 948 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
965 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 | 949 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
966 | 950 |
967 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" | 951 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" |
968 "punpcklbw %%mm7, %%mm6 \n\t" // L6 | 952 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
969 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 | 953 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
970 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" | 954 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" |
971 "punpckhbw %%mm7, %%mm6 \n\t" // H6 | 955 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
972 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 | 956 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
973 | 957 |
974 "paddw %%mm0, %%mm0 \n\t" // 2L4 | 958 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
975 "paddw %%mm1, %%mm1 \n\t" // 2H4 | 959 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
976 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 | 960 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
977 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 | 961 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
978 | 962 |
979 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 | 963 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
980 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 | 964 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
981 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 | 965 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
982 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 | 966 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
983 | 967 |
984 "movq (%0, %1, 4), %%mm2 \n\t" | 968 "movq (%0, %1, 4), %%mm2 \n\t" |
985 "movq %%mm2, %%mm3 \n\t" | 969 "movq %%mm2, %%mm3 \n\t" |
986 "punpcklbw %%mm7, %%mm2 \n\t" // L7 | 970 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
987 "punpckhbw %%mm7, %%mm3 \n\t" // H7 | 971 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
988 | 972 |
989 "paddw %%mm2, %%mm2 \n\t" // 2L7 | 973 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
990 "paddw %%mm3, %%mm3 \n\t" // 2H7 | 974 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
991 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 | 975 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
992 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 | 976 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
993 | 977 |
994 "movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 | 978 "movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
995 "movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 | 979 "movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
996 | 980 |
997 #ifdef HAVE_MMX2 | 981 #ifdef HAVE_MMX2 |
998 "movq %%mm7, %%mm6 \n\t" // 0 | 982 "movq %%mm7, %%mm6 \n\t" // 0 |
999 "psubw %%mm0, %%mm6 \n\t" | 983 "psubw %%mm0, %%mm6 \n\t" |
1000 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| | 984 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
1001 "movq %%mm7, %%mm6 \n\t" // 0 | 985 "movq %%mm7, %%mm6 \n\t" // 0 |
1002 "psubw %%mm1, %%mm6 \n\t" | 986 "psubw %%mm1, %%mm6 \n\t" |
1003 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| | 987 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
1004 "movq %%mm7, %%mm6 \n\t" // 0 | 988 "movq %%mm7, %%mm6 \n\t" // 0 |
1005 "psubw %%mm2, %%mm6 \n\t" | 989 "psubw %%mm2, %%mm6 \n\t" |
1006 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| | 990 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
1007 "movq %%mm7, %%mm6 \n\t" // 0 | 991 "movq %%mm7, %%mm6 \n\t" // 0 |
1008 "psubw %%mm3, %%mm6 \n\t" | 992 "psubw %%mm3, %%mm6 \n\t" |
1009 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| | 993 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
1010 #else | 994 #else |
1011 "movq %%mm7, %%mm6 \n\t" // 0 | 995 "movq %%mm7, %%mm6 \n\t" // 0 |
1012 "pcmpgtw %%mm0, %%mm6 \n\t" | 996 "pcmpgtw %%mm0, %%mm6 \n\t" |
1013 "pxor %%mm6, %%mm0 \n\t" | 997 "pxor %%mm6, %%mm0 \n\t" |
1014 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| | 998 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
1015 "movq %%mm7, %%mm6 \n\t" // 0 | 999 "movq %%mm7, %%mm6 \n\t" // 0 |
1016 "pcmpgtw %%mm1, %%mm6 \n\t" | 1000 "pcmpgtw %%mm1, %%mm6 \n\t" |
1017 "pxor %%mm6, %%mm1 \n\t" | 1001 "pxor %%mm6, %%mm1 \n\t" |
1018 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| | 1002 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
1019 "movq %%mm7, %%mm6 \n\t" // 0 | 1003 "movq %%mm7, %%mm6 \n\t" // 0 |
1020 "pcmpgtw %%mm2, %%mm6 \n\t" | 1004 "pcmpgtw %%mm2, %%mm6 \n\t" |
1021 "pxor %%mm6, %%mm2 \n\t" | 1005 "pxor %%mm6, %%mm2 \n\t" |
1022 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| | 1006 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
1023 "movq %%mm7, %%mm6 \n\t" // 0 | 1007 "movq %%mm7, %%mm6 \n\t" // 0 |
1024 "pcmpgtw %%mm3, %%mm6 \n\t" | 1008 "pcmpgtw %%mm3, %%mm6 \n\t" |
1025 "pxor %%mm6, %%mm3 \n\t" | 1009 "pxor %%mm6, %%mm3 \n\t" |
1026 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| | 1010 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
1027 #endif | 1011 #endif |
1028 | 1012 |
1029 #ifdef HAVE_MMX2 | 1013 #ifdef HAVE_MMX2 |
1030 "pminsw %%mm2, %%mm0 \n\t" | 1014 "pminsw %%mm2, %%mm0 \n\t" |
1031 "pminsw %%mm3, %%mm1 \n\t" | 1015 "pminsw %%mm3, %%mm1 \n\t" |
1032 #else | 1016 #else |
1033 "movq %%mm0, %%mm6 \n\t" | 1017 "movq %%mm0, %%mm6 \n\t" |
1034 "psubusw %%mm2, %%mm6 \n\t" | 1018 "psubusw %%mm2, %%mm6 \n\t" |
1035 "psubw %%mm6, %%mm0 \n\t" | 1019 "psubw %%mm6, %%mm0 \n\t" |
1036 "movq %%mm1, %%mm6 \n\t" | 1020 "movq %%mm1, %%mm6 \n\t" |
1037 "psubusw %%mm3, %%mm6 \n\t" | 1021 "psubusw %%mm3, %%mm6 \n\t" |
1038 "psubw %%mm6, %%mm1 \n\t" | 1022 "psubw %%mm6, %%mm1 \n\t" |
1039 #endif | 1023 #endif |
1040 | 1024 |
1041 "movd %2, %%mm2 \n\t" // QP | 1025 "movd %2, %%mm2 \n\t" // QP |
1042 "punpcklbw %%mm7, %%mm2 \n\t" | 1026 "punpcklbw %%mm7, %%mm2 \n\t" |
1043 | 1027 |
1044 "movq %%mm7, %%mm6 \n\t" // 0 | 1028 "movq %%mm7, %%mm6 \n\t" // 0 |
1045 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) | 1029 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
1046 "pxor %%mm6, %%mm4 \n\t" | 1030 "pxor %%mm6, %%mm4 \n\t" |
1047 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| | 1031 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
1048 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) | 1032 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
1049 "pxor %%mm7, %%mm5 \n\t" | 1033 "pxor %%mm7, %%mm5 \n\t" |
1050 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| | 1034 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
1051 // 100 opcodes | 1035 // 100 opcodes |
1052 "psllw $3, %%mm2 \n\t" // 8QP | 1036 "psllw $3, %%mm2 \n\t" // 8QP |
1053 "movq %%mm2, %%mm3 \n\t" // 8QP | 1037 "movq %%mm2, %%mm3 \n\t" // 8QP |
1054 "pcmpgtw %%mm4, %%mm2 \n\t" | 1038 "pcmpgtw %%mm4, %%mm2 \n\t" |
1055 "pcmpgtw %%mm5, %%mm3 \n\t" | 1039 "pcmpgtw %%mm5, %%mm3 \n\t" |
1056 "pand %%mm2, %%mm4 \n\t" | 1040 "pand %%mm2, %%mm4 \n\t" |
1057 "pand %%mm3, %%mm5 \n\t" | 1041 "pand %%mm3, %%mm5 \n\t" |
1058 | 1042 |
1059 | 1043 |
1060 "psubusw %%mm0, %%mm4 \n\t" // hd | 1044 "psubusw %%mm0, %%mm4 \n\t" // hd |
1061 "psubusw %%mm1, %%mm5 \n\t" // ld | 1045 "psubusw %%mm1, %%mm5 \n\t" // ld |
1062 | 1046 |
1063 | 1047 |
1064 "movq "MANGLE(w05)", %%mm2 \n\t" // 5 | 1048 "movq "MANGLE(w05)", %%mm2 \n\t" // 5 |
1065 "pmullw %%mm2, %%mm4 \n\t" | 1049 "pmullw %%mm2, %%mm4 \n\t" |
1066 "pmullw %%mm2, %%mm5 \n\t" | 1050 "pmullw %%mm2, %%mm5 \n\t" |
1067 "movq "MANGLE(w20)", %%mm2 \n\t" // 32 | 1051 "movq "MANGLE(w20)", %%mm2 \n\t" // 32 |
1068 "paddw %%mm2, %%mm4 \n\t" | 1052 "paddw %%mm2, %%mm4 \n\t" |
1069 "paddw %%mm2, %%mm5 \n\t" | 1053 "paddw %%mm2, %%mm5 \n\t" |
1070 "psrlw $6, %%mm4 \n\t" | 1054 "psrlw $6, %%mm4 \n\t" |
1071 "psrlw $6, %%mm5 \n\t" | 1055 "psrlw $6, %%mm5 \n\t" |
1072 | 1056 |
1073 "movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4 | 1057 "movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4 |
1074 "movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4 | 1058 "movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4 |
1075 | 1059 |
1076 "pxor %%mm2, %%mm2 \n\t" | 1060 "pxor %%mm2, %%mm2 \n\t" |
1077 "pxor %%mm3, %%mm3 \n\t" | 1061 "pxor %%mm3, %%mm3 \n\t" |
1078 | 1062 |
1079 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) | 1063 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
1080 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) | 1064 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
1081 "pxor %%mm2, %%mm0 \n\t" | 1065 "pxor %%mm2, %%mm0 \n\t" |
1082 "pxor %%mm3, %%mm1 \n\t" | 1066 "pxor %%mm3, %%mm1 \n\t" |
1083 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| | 1067 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
1084 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| | 1068 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| |
1085 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 | 1069 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 |
1086 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 | 1070 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
1087 | 1071 |
1088 "pxor %%mm6, %%mm2 \n\t" | 1072 "pxor %%mm6, %%mm2 \n\t" |
1089 "pxor %%mm7, %%mm3 \n\t" | 1073 "pxor %%mm7, %%mm3 \n\t" |
1090 "pand %%mm2, %%mm4 \n\t" | 1074 "pand %%mm2, %%mm4 \n\t" |
1091 "pand %%mm3, %%mm5 \n\t" | 1075 "pand %%mm3, %%mm5 \n\t" |
1092 | 1076 |
1093 #ifdef HAVE_MMX2 | 1077 #ifdef HAVE_MMX2 |
1094 "pminsw %%mm0, %%mm4 \n\t" | 1078 "pminsw %%mm0, %%mm4 \n\t" |
1095 "pminsw %%mm1, %%mm5 \n\t" | 1079 "pminsw %%mm1, %%mm5 \n\t" |
1096 #else | 1080 #else |
1097 "movq %%mm4, %%mm2 \n\t" | 1081 "movq %%mm4, %%mm2 \n\t" |
1098 "psubusw %%mm0, %%mm2 \n\t" | 1082 "psubusw %%mm0, %%mm2 \n\t" |
1099 "psubw %%mm2, %%mm4 \n\t" | 1083 "psubw %%mm2, %%mm4 \n\t" |
1100 "movq %%mm5, %%mm2 \n\t" | 1084 "movq %%mm5, %%mm2 \n\t" |
1101 "psubusw %%mm1, %%mm2 \n\t" | 1085 "psubusw %%mm1, %%mm2 \n\t" |
1102 "psubw %%mm2, %%mm5 \n\t" | 1086 "psubw %%mm2, %%mm5 \n\t" |
1103 #endif | 1087 #endif |
1104 "pxor %%mm6, %%mm4 \n\t" | 1088 "pxor %%mm6, %%mm4 \n\t" |
1105 "pxor %%mm7, %%mm5 \n\t" | 1089 "pxor %%mm7, %%mm5 \n\t" |
1106 "psubw %%mm6, %%mm4 \n\t" | 1090 "psubw %%mm6, %%mm4 \n\t" |
1107 "psubw %%mm7, %%mm5 \n\t" | 1091 "psubw %%mm7, %%mm5 \n\t" |
1108 "packsswb %%mm5, %%mm4 \n\t" | 1092 "packsswb %%mm5, %%mm4 \n\t" |
1109 "movq (%0), %%mm0 \n\t" | 1093 "movq (%0), %%mm0 \n\t" |
1110 "paddb %%mm4, %%mm0 \n\t" | 1094 "paddb %%mm4, %%mm0 \n\t" |
1111 "movq %%mm0, (%0) \n\t" | 1095 "movq %%mm0, (%0) \n\t" |
1112 "movq (%0, %1), %%mm0 \n\t" | 1096 "movq (%0, %1), %%mm0 \n\t" |
1113 "psubb %%mm4, %%mm0 \n\t" | 1097 "psubb %%mm4, %%mm0 \n\t" |
1114 "movq %%mm0, (%0, %1) \n\t" | 1098 "movq %%mm0, (%0, %1) \n\t" |
1115 | 1099 |
1116 : "+r" (src) | 1100 : "+r" (src) |
1117 : "r" ((long)stride), "m" (c->pQPb) | 1101 : "r" ((long)stride), "m" (c->pQPb) |
1118 : "%"REG_a, "%"REG_c | 1102 : "%"REG_a, "%"REG_c |
1119 ); | 1103 ); |
1120 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1104 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1121 const int l1= stride; | 1105 const int l1= stride; |
1122 const int l2= stride + l1; | 1106 const int l2= stride + l1; |
1123 const int l3= stride + l2; | 1107 const int l3= stride + l2; |
1124 const int l4= stride + l3; | 1108 const int l4= stride + l3; |
1125 const int l5= stride + l4; | 1109 const int l5= stride + l4; |
1126 const int l6= stride + l5; | 1110 const int l6= stride + l5; |
1127 const int l7= stride + l6; | 1111 const int l7= stride + l6; |
1128 const int l8= stride + l7; | 1112 const int l8= stride + l7; |
1129 // const int l9= stride + l8; | 1113 // const int l9= stride + l8; |
1130 int x; | 1114 int x; |
1131 src+= stride*3; | 1115 src+= stride*3; |
1132 for(x=0; x<BLOCK_SIZE; x++) | 1116 for(x=0; x<BLOCK_SIZE; x++){ |
1133 { | 1117 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]); |
1134 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]); | 1118 if(FFABS(middleEnergy) < 8*c->QP){ |
1135 if(FFABS(middleEnergy) < 8*c->QP) | 1119 const int q=(src[l4] - src[l5])/2; |
1136 { | 1120 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); |
1137 const int q=(src[l4] - src[l5])/2; | 1121 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); |
1138 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); | 1122 |
1139 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); | 1123 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) ); |
1140 | 1124 d= FFMAX(d, 0); |
1141 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) ); | 1125 |
1142 d= FFMAX(d, 0); | 1126 d= (5*d + 32) >> 6; |
1143 | 1127 d*= FFSIGN(-middleEnergy); |
1144 d= (5*d + 32) >> 6; | 1128 |
1145 d*= FFSIGN(-middleEnergy); | 1129 if(q>0){ |
1146 | 1130 d= d<0 ? 0 : d; |
1147 if(q>0) | 1131 d= d>q ? q : d; |
1148 { | 1132 }else{ |
1149 d= d<0 ? 0 : d; | 1133 d= d>0 ? 0 : d; |
1150 d= d>q ? q : d; | 1134 d= d<q ? q : d; |
1151 } | 1135 } |
1152 else | 1136 |
1153 { | 1137 src[l4]-= d; |
1154 d= d>0 ? 0 : d; | 1138 src[l5]+= d; |
1155 d= d<q ? q : d; | |
1156 } | |
1157 | |
1158 src[l4]-= d; | |
1159 src[l5]+= d; | |
1160 } | |
1161 src++; | |
1162 } | 1139 } |
1140 src++; | |
1141 } | |
1163 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1142 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1164 } | 1143 } |
1165 #endif //HAVE_ALTIVEC | 1144 #endif //HAVE_ALTIVEC |
1166 | 1145 |
1167 #ifndef HAVE_ALTIVEC | 1146 #ifndef HAVE_ALTIVEC |
1168 static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c) | 1147 static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c) |
1169 { | 1148 { |
1170 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1149 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1171 asm volatile( | 1150 asm volatile( |
1172 "pxor %%mm6, %%mm6 \n\t" | 1151 "pxor %%mm6, %%mm6 \n\t" |
1173 "pcmpeqb %%mm7, %%mm7 \n\t" | 1152 "pcmpeqb %%mm7, %%mm7 \n\t" |
1174 "movq %2, %%mm0 \n\t" | 1153 "movq %2, %%mm0 \n\t" |
1175 "punpcklbw %%mm6, %%mm0 \n\t" | 1154 "punpcklbw %%mm6, %%mm0 \n\t" |
1176 "psrlw $1, %%mm0 \n\t" | 1155 "psrlw $1, %%mm0 \n\t" |
1177 "psubw %%mm7, %%mm0 \n\t" | 1156 "psubw %%mm7, %%mm0 \n\t" |
1178 "packuswb %%mm0, %%mm0 \n\t" | 1157 "packuswb %%mm0, %%mm0 \n\t" |
1179 "movq %%mm0, %3 \n\t" | 1158 "movq %%mm0, %3 \n\t" |
1180 | 1159 |
1181 "lea (%0, %1), %%"REG_a" \n\t" | 1160 "lea (%0, %1), %%"REG_a" \n\t" |
1182 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" | 1161 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" |
1183 | 1162 |
1184 // 0 1 2 3 4 5 6 7 8 9 | 1163 // 0 1 2 3 4 5 6 7 8 9 |
1185 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 | 1164 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 |
1186 | 1165 |
1187 #undef FIND_MIN_MAX | 1166 #undef FIND_MIN_MAX |
1188 #ifdef HAVE_MMX2 | 1167 #ifdef HAVE_MMX2 |
1189 #define REAL_FIND_MIN_MAX(addr)\ | 1168 #define REAL_FIND_MIN_MAX(addr)\ |
1190 "movq " #addr ", %%mm0 \n\t"\ | 1169 "movq " #addr ", %%mm0 \n\t"\ |
1191 "pminub %%mm0, %%mm7 \n\t"\ | 1170 "pminub %%mm0, %%mm7 \n\t"\ |
1192 "pmaxub %%mm0, %%mm6 \n\t" | 1171 "pmaxub %%mm0, %%mm6 \n\t" |
1193 #else | 1172 #else |
1194 #define REAL_FIND_MIN_MAX(addr)\ | 1173 #define REAL_FIND_MIN_MAX(addr)\ |
1195 "movq " #addr ", %%mm0 \n\t"\ | 1174 "movq " #addr ", %%mm0 \n\t"\ |
1196 "movq %%mm7, %%mm1 \n\t"\ | 1175 "movq %%mm7, %%mm1 \n\t"\ |
1197 "psubusb %%mm0, %%mm6 \n\t"\ | 1176 "psubusb %%mm0, %%mm6 \n\t"\ |
1198 "paddb %%mm0, %%mm6 \n\t"\ | 1177 "paddb %%mm0, %%mm6 \n\t"\ |
1199 "psubusb %%mm0, %%mm1 \n\t"\ | 1178 "psubusb %%mm0, %%mm1 \n\t"\ |
1200 "psubb %%mm1, %%mm7 \n\t" | 1179 "psubb %%mm1, %%mm7 \n\t" |
1201 #endif | 1180 #endif |
1202 #define FIND_MIN_MAX(addr) REAL_FIND_MIN_MAX(addr) | 1181 #define FIND_MIN_MAX(addr) REAL_FIND_MIN_MAX(addr) |
1203 | 1182 |
1204 FIND_MIN_MAX((%%REGa)) | 1183 FIND_MIN_MAX((%%REGa)) |
1205 FIND_MIN_MAX((%%REGa, %1)) | 1184 FIND_MIN_MAX((%%REGa, %1)) |
1208 FIND_MIN_MAX((%%REGd)) | 1187 FIND_MIN_MAX((%%REGd)) |
1209 FIND_MIN_MAX((%%REGd, %1)) | 1188 FIND_MIN_MAX((%%REGd, %1)) |
1210 FIND_MIN_MAX((%%REGd, %1, 2)) | 1189 FIND_MIN_MAX((%%REGd, %1, 2)) |
1211 FIND_MIN_MAX((%0, %1, 8)) | 1190 FIND_MIN_MAX((%0, %1, 8)) |
1212 | 1191 |
1213 "movq %%mm7, %%mm4 \n\t" | 1192 "movq %%mm7, %%mm4 \n\t" |
1214 "psrlq $8, %%mm7 \n\t" | 1193 "psrlq $8, %%mm7 \n\t" |
1215 #ifdef HAVE_MMX2 | 1194 #ifdef HAVE_MMX2 |
1216 "pminub %%mm4, %%mm7 \n\t" // min of pixels | 1195 "pminub %%mm4, %%mm7 \n\t" // min of pixels |
1217 "pshufw $0xF9, %%mm7, %%mm4 \n\t" | 1196 "pshufw $0xF9, %%mm7, %%mm4 \n\t" |
1218 "pminub %%mm4, %%mm7 \n\t" // min of pixels | 1197 "pminub %%mm4, %%mm7 \n\t" // min of pixels |
1219 "pshufw $0xFE, %%mm7, %%mm4 \n\t" | 1198 "pshufw $0xFE, %%mm7, %%mm4 \n\t" |
1220 "pminub %%mm4, %%mm7 \n\t" | 1199 "pminub %%mm4, %%mm7 \n\t" |
1221 #else | 1200 #else |
1222 "movq %%mm7, %%mm1 \n\t" | 1201 "movq %%mm7, %%mm1 \n\t" |
1223 "psubusb %%mm4, %%mm1 \n\t" | 1202 "psubusb %%mm4, %%mm1 \n\t" |
1224 "psubb %%mm1, %%mm7 \n\t" | 1203 "psubb %%mm1, %%mm7 \n\t" |
1225 "movq %%mm7, %%mm4 \n\t" | 1204 "movq %%mm7, %%mm4 \n\t" |
1226 "psrlq $16, %%mm7 \n\t" | 1205 "psrlq $16, %%mm7 \n\t" |
1227 "movq %%mm7, %%mm1 \n\t" | 1206 "movq %%mm7, %%mm1 \n\t" |
1228 "psubusb %%mm4, %%mm1 \n\t" | 1207 "psubusb %%mm4, %%mm1 \n\t" |
1229 "psubb %%mm1, %%mm7 \n\t" | 1208 "psubb %%mm1, %%mm7 \n\t" |
1230 "movq %%mm7, %%mm4 \n\t" | 1209 "movq %%mm7, %%mm4 \n\t" |
1231 "psrlq $32, %%mm7 \n\t" | 1210 "psrlq $32, %%mm7 \n\t" |
1232 "movq %%mm7, %%mm1 \n\t" | 1211 "movq %%mm7, %%mm1 \n\t" |
1233 "psubusb %%mm4, %%mm1 \n\t" | 1212 "psubusb %%mm4, %%mm1 \n\t" |
1234 "psubb %%mm1, %%mm7 \n\t" | 1213 "psubb %%mm1, %%mm7 \n\t" |
1235 #endif | 1214 #endif |
1236 | 1215 |
1237 | 1216 |
1238 "movq %%mm6, %%mm4 \n\t" | 1217 "movq %%mm6, %%mm4 \n\t" |
1239 "psrlq $8, %%mm6 \n\t" | 1218 "psrlq $8, %%mm6 \n\t" |
1240 #ifdef HAVE_MMX2 | 1219 #ifdef HAVE_MMX2 |
1241 "pmaxub %%mm4, %%mm6 \n\t" // max of pixels | 1220 "pmaxub %%mm4, %%mm6 \n\t" // max of pixels |
1242 "pshufw $0xF9, %%mm6, %%mm4 \n\t" | 1221 "pshufw $0xF9, %%mm6, %%mm4 \n\t" |
1243 "pmaxub %%mm4, %%mm6 \n\t" | 1222 "pmaxub %%mm4, %%mm6 \n\t" |
1244 "pshufw $0xFE, %%mm6, %%mm4 \n\t" | 1223 "pshufw $0xFE, %%mm6, %%mm4 \n\t" |
1245 "pmaxub %%mm4, %%mm6 \n\t" | 1224 "pmaxub %%mm4, %%mm6 \n\t" |
1246 #else | 1225 #else |
1247 "psubusb %%mm4, %%mm6 \n\t" | 1226 "psubusb %%mm4, %%mm6 \n\t" |
1248 "paddb %%mm4, %%mm6 \n\t" | 1227 "paddb %%mm4, %%mm6 \n\t" |
1249 "movq %%mm6, %%mm4 \n\t" | 1228 "movq %%mm6, %%mm4 \n\t" |
1250 "psrlq $16, %%mm6 \n\t" | 1229 "psrlq $16, %%mm6 \n\t" |
1251 "psubusb %%mm4, %%mm6 \n\t" | 1230 "psubusb %%mm4, %%mm6 \n\t" |
1252 "paddb %%mm4, %%mm6 \n\t" | 1231 "paddb %%mm4, %%mm6 \n\t" |
1253 "movq %%mm6, %%mm4 \n\t" | 1232 "movq %%mm6, %%mm4 \n\t" |
1254 "psrlq $32, %%mm6 \n\t" | 1233 "psrlq $32, %%mm6 \n\t" |
1255 "psubusb %%mm4, %%mm6 \n\t" | 1234 "psubusb %%mm4, %%mm6 \n\t" |
1256 "paddb %%mm4, %%mm6 \n\t" | 1235 "paddb %%mm4, %%mm6 \n\t" |
1257 #endif | 1236 #endif |
1258 "movq %%mm6, %%mm0 \n\t" // max | 1237 "movq %%mm6, %%mm0 \n\t" // max |
1259 "psubb %%mm7, %%mm6 \n\t" // max - min | 1238 "psubb %%mm7, %%mm6 \n\t" // max - min |
1260 "movd %%mm6, %%ecx \n\t" | 1239 "movd %%mm6, %%ecx \n\t" |
1261 "cmpb "MANGLE(deringThreshold)", %%cl \n\t" | 1240 "cmpb "MANGLE(deringThreshold)", %%cl \n\t" |
1262 " jb 1f \n\t" | 1241 " jb 1f \n\t" |
1263 "lea -24(%%"REG_SP"), %%"REG_c" \n\t" | 1242 "lea -24(%%"REG_SP"), %%"REG_c" \n\t" |
1264 "and "ALIGN_MASK", %%"REG_c" \n\t" | 1243 "and "ALIGN_MASK", %%"REG_c" \n\t" |
1265 PAVGB(%%mm0, %%mm7) // a=(max + min)/2 | 1244 PAVGB(%%mm0, %%mm7) // a=(max + min)/2 |
1266 "punpcklbw %%mm7, %%mm7 \n\t" | 1245 "punpcklbw %%mm7, %%mm7 \n\t" |
1267 "punpcklbw %%mm7, %%mm7 \n\t" | 1246 "punpcklbw %%mm7, %%mm7 \n\t" |
1268 "punpcklbw %%mm7, %%mm7 \n\t" | 1247 "punpcklbw %%mm7, %%mm7 \n\t" |
1269 "movq %%mm7, (%%"REG_c") \n\t" | 1248 "movq %%mm7, (%%"REG_c") \n\t" |
1270 | 1249 |
1271 "movq (%0), %%mm0 \n\t" // L10 | 1250 "movq (%0), %%mm0 \n\t" // L10 |
1272 "movq %%mm0, %%mm1 \n\t" // L10 | 1251 "movq %%mm0, %%mm1 \n\t" // L10 |
1273 "movq %%mm0, %%mm2 \n\t" // L10 | 1252 "movq %%mm0, %%mm2 \n\t" // L10 |
1274 "psllq $8, %%mm1 \n\t" | 1253 "psllq $8, %%mm1 \n\t" |
1275 "psrlq $8, %%mm2 \n\t" | 1254 "psrlq $8, %%mm2 \n\t" |
1276 "movd -4(%0), %%mm3 \n\t" | 1255 "movd -4(%0), %%mm3 \n\t" |
1277 "movd 8(%0), %%mm4 \n\t" | 1256 "movd 8(%0), %%mm4 \n\t" |
1278 "psrlq $24, %%mm3 \n\t" | 1257 "psrlq $24, %%mm3 \n\t" |
1279 "psllq $56, %%mm4 \n\t" | 1258 "psllq $56, %%mm4 \n\t" |
1280 "por %%mm3, %%mm1 \n\t" // L00 | 1259 "por %%mm3, %%mm1 \n\t" // L00 |
1281 "por %%mm4, %%mm2 \n\t" // L20 | 1260 "por %%mm4, %%mm2 \n\t" // L20 |
1282 "movq %%mm1, %%mm3 \n\t" // L00 | 1261 "movq %%mm1, %%mm3 \n\t" // L00 |
1283 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2 | 1262 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2 |
1284 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4 | 1263 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4 |
1285 "psubusb %%mm7, %%mm0 \n\t" | 1264 "psubusb %%mm7, %%mm0 \n\t" |
1286 "psubusb %%mm7, %%mm2 \n\t" | 1265 "psubusb %%mm7, %%mm2 \n\t" |
1287 "psubusb %%mm7, %%mm3 \n\t" | 1266 "psubusb %%mm7, %%mm3 \n\t" |
1288 "pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1 | 1267 "pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1 |
1289 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1 | 1268 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1 |
1290 "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1 | 1269 "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1 |
1291 "paddb %%mm2, %%mm0 \n\t" | 1270 "paddb %%mm2, %%mm0 \n\t" |
1292 "paddb %%mm3, %%mm0 \n\t" | 1271 "paddb %%mm3, %%mm0 \n\t" |
1293 | 1272 |
1294 "movq (%%"REG_a"), %%mm2 \n\t" // L11 | 1273 "movq (%%"REG_a"), %%mm2 \n\t" // L11 |
1295 "movq %%mm2, %%mm3 \n\t" // L11 | 1274 "movq %%mm2, %%mm3 \n\t" // L11 |
1296 "movq %%mm2, %%mm4 \n\t" // L11 | 1275 "movq %%mm2, %%mm4 \n\t" // L11 |
1297 "psllq $8, %%mm3 \n\t" | 1276 "psllq $8, %%mm3 \n\t" |
1298 "psrlq $8, %%mm4 \n\t" | 1277 "psrlq $8, %%mm4 \n\t" |
1299 "movd -4(%%"REG_a"), %%mm5 \n\t" | 1278 "movd -4(%%"REG_a"), %%mm5 \n\t" |
1300 "movd 8(%%"REG_a"), %%mm6 \n\t" | 1279 "movd 8(%%"REG_a"), %%mm6 \n\t" |
1301 "psrlq $24, %%mm5 \n\t" | 1280 "psrlq $24, %%mm5 \n\t" |
1302 "psllq $56, %%mm6 \n\t" | 1281 "psllq $56, %%mm6 \n\t" |
1303 "por %%mm5, %%mm3 \n\t" // L01 | 1282 "por %%mm5, %%mm3 \n\t" // L01 |
1304 "por %%mm6, %%mm4 \n\t" // L21 | 1283 "por %%mm6, %%mm4 \n\t" // L21 |
1305 "movq %%mm3, %%mm5 \n\t" // L01 | 1284 "movq %%mm3, %%mm5 \n\t" // L01 |
1306 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2 | 1285 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2 |
1307 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4 | 1286 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4 |
1308 "psubusb %%mm7, %%mm2 \n\t" | 1287 "psubusb %%mm7, %%mm2 \n\t" |
1309 "psubusb %%mm7, %%mm4 \n\t" | 1288 "psubusb %%mm7, %%mm4 \n\t" |
1310 "psubusb %%mm7, %%mm5 \n\t" | 1289 "psubusb %%mm7, %%mm5 \n\t" |
1311 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1 | 1290 "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1 |
1312 "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1 | 1291 "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1 |
1313 "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1 | 1292 "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1 |
1314 "paddb %%mm4, %%mm2 \n\t" | 1293 "paddb %%mm4, %%mm2 \n\t" |
1315 "paddb %%mm5, %%mm2 \n\t" | 1294 "paddb %%mm5, %%mm2 \n\t" |
1316 // 0, 2, 3, 1 | 1295 // 0, 2, 3, 1 |
1317 #define REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ | 1296 #define REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ |
1318 "movq " #src ", " #sx " \n\t" /* src[0] */\ | 1297 "movq " #src ", " #sx " \n\t" /* src[0] */\ |
1319 "movq " #sx ", " #lx " \n\t" /* src[0] */\ | 1298 "movq " #sx ", " #lx " \n\t" /* src[0] */\ |
1320 "movq " #sx ", " #t0 " \n\t" /* src[0] */\ | 1299 "movq " #sx ", " #t0 " \n\t" /* src[0] */\ |
1321 "psllq $8, " #lx " \n\t"\ | 1300 "psllq $8, " #lx " \n\t"\ |
1322 "psrlq $8, " #t0 " \n\t"\ | 1301 "psrlq $8, " #t0 " \n\t"\ |
1323 "movd -4" #src ", " #t1 " \n\t"\ | 1302 "movd -4" #src ", " #t1 " \n\t"\ |
1324 "psrlq $24, " #t1 " \n\t"\ | 1303 "psrlq $24, " #t1 " \n\t"\ |
1325 "por " #t1 ", " #lx " \n\t" /* src[-1] */\ | 1304 "por " #t1 ", " #lx " \n\t" /* src[-1] */\ |
1326 "movd 8" #src ", " #t1 " \n\t"\ | 1305 "movd 8" #src ", " #t1 " \n\t"\ |
1327 "psllq $56, " #t1 " \n\t"\ | 1306 "psllq $56, " #t1 " \n\t"\ |
1328 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\ | 1307 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\ |
1329 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\ | 1308 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\ |
1330 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\ | 1309 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\ |
1331 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\ | 1310 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\ |
1332 PAVGB(lx, pplx) \ | 1311 PAVGB(lx, pplx) \ |
1333 "movq " #lx ", 8(%%"REG_c") \n\t"\ | 1312 "movq " #lx ", 8(%%"REG_c") \n\t"\ |
1334 "movq (%%"REG_c"), " #lx " \n\t"\ | 1313 "movq (%%"REG_c"), " #lx " \n\t"\ |
1335 "psubusb " #lx ", " #t1 " \n\t"\ | 1314 "psubusb " #lx ", " #t1 " \n\t"\ |
1336 "psubusb " #lx ", " #t0 " \n\t"\ | 1315 "psubusb " #lx ", " #t0 " \n\t"\ |
1337 "psubusb " #lx ", " #sx " \n\t"\ | 1316 "psubusb " #lx ", " #sx " \n\t"\ |
1338 "movq "MANGLE(b00)", " #lx " \n\t"\ | 1317 "movq "MANGLE(b00)", " #lx " \n\t"\ |
1339 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\ | 1318 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\ |
1340 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\ | 1319 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\ |
1341 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\ | 1320 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\ |
1342 "paddb " #t1 ", " #t0 " \n\t"\ | 1321 "paddb " #t1 ", " #t0 " \n\t"\ |
1343 "paddb " #t0 ", " #sx " \n\t"\ | 1322 "paddb " #t0 ", " #sx " \n\t"\ |
1344 \ | 1323 \ |
1345 PAVGB(plx, pplx) /* filtered */\ | 1324 PAVGB(plx, pplx) /* filtered */\ |
1346 "movq " #dst ", " #t0 " \n\t" /* dst */\ | 1325 "movq " #dst ", " #t0 " \n\t" /* dst */\ |
1347 "movq " #t0 ", " #t1 " \n\t" /* dst */\ | 1326 "movq " #t0 ", " #t1 " \n\t" /* dst */\ |
1348 "psubusb %3, " #t0 " \n\t"\ | 1327 "psubusb %3, " #t0 " \n\t"\ |
1349 "paddusb %3, " #t1 " \n\t"\ | 1328 "paddusb %3, " #t1 " \n\t"\ |
1350 PMAXUB(t0, pplx)\ | 1329 PMAXUB(t0, pplx)\ |
1351 PMINUB(t1, pplx, t0)\ | 1330 PMINUB(t1, pplx, t0)\ |
1352 "paddb " #sx ", " #ppsx " \n\t"\ | 1331 "paddb " #sx ", " #ppsx " \n\t"\ |
1353 "paddb " #psx ", " #ppsx " \n\t"\ | 1332 "paddb " #psx ", " #ppsx " \n\t"\ |
1354 "#paddb "MANGLE(b02)", " #ppsx " \n\t"\ | 1333 "#paddb "MANGLE(b02)", " #ppsx " \n\t"\ |
1355 "pand "MANGLE(b08)", " #ppsx " \n\t"\ | 1334 "pand "MANGLE(b08)", " #ppsx " \n\t"\ |
1356 "pcmpeqb " #lx ", " #ppsx " \n\t"\ | 1335 "pcmpeqb " #lx ", " #ppsx " \n\t"\ |
1357 "pand " #ppsx ", " #pplx " \n\t"\ | 1336 "pand " #ppsx ", " #pplx " \n\t"\ |
1358 "pandn " #dst ", " #ppsx " \n\t"\ | 1337 "pandn " #dst ", " #ppsx " \n\t"\ |
1359 "por " #pplx ", " #ppsx " \n\t"\ | 1338 "por " #pplx ", " #ppsx " \n\t"\ |
1360 "movq " #ppsx ", " #dst " \n\t"\ | 1339 "movq " #ppsx ", " #dst " \n\t"\ |
1361 "movq 8(%%"REG_c"), " #lx " \n\t" | 1340 "movq 8(%%"REG_c"), " #lx " \n\t" |
1362 | 1341 |
1363 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ | 1342 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ |
1364 REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) | 1343 REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) |
1365 /* | 1344 /* |
1366 0000000 | 1345 0000000 |
1385 DERING_CORE((%%REGd) ,(%%REGd, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | 1364 DERING_CORE((%%REGd) ,(%%REGd, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) |
1386 DERING_CORE((%%REGd, %1) ,(%%REGd, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | 1365 DERING_CORE((%%REGd, %1) ,(%%REGd, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) |
1387 DERING_CORE((%%REGd, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | 1366 DERING_CORE((%%REGd, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) |
1388 DERING_CORE((%0, %1, 8) ,(%%REGd, %1, 4),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | 1367 DERING_CORE((%0, %1, 8) ,(%%REGd, %1, 4),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) |
1389 | 1368 |
1390 "1: \n\t" | 1369 "1: \n\t" |
1391 : : "r" (src), "r" ((long)stride), "m" (c->pQPb), "m"(c->pQPb2) | 1370 : : "r" (src), "r" ((long)stride), "m" (c->pQPb), "m"(c->pQPb2) |
1392 : "%"REG_a, "%"REG_d, "%"REG_c | 1371 : "%"REG_a, "%"REG_d, "%"REG_c |
1393 ); | 1372 ); |
1394 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1373 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1395 int y; | 1374 int y; |
1396 int min=255; | 1375 int min=255; |
1397 int max=0; | 1376 int max=0; |
1398 int avg; | 1377 int avg; |
1399 uint8_t *p; | 1378 uint8_t *p; |
1400 int s[10]; | 1379 int s[10]; |
1401 const int QP2= c->QP/2 + 1; | 1380 const int QP2= c->QP/2 + 1; |
1402 | 1381 |
1403 for(y=1; y<9; y++) | 1382 for(y=1; y<9; y++){ |
1404 { | 1383 int x; |
1405 int x; | 1384 p= src + stride*y; |
1406 p= src + stride*y; | 1385 for(x=1; x<9; x++){ |
1407 for(x=1; x<9; x++) | 1386 p++; |
1408 { | 1387 if(*p > max) max= *p; |
1409 p++; | 1388 if(*p < min) min= *p; |
1410 if(*p > max) max= *p; | |
1411 if(*p < min) min= *p; | |
1412 } | |
1413 } | 1389 } |
1414 avg= (min + max + 1)>>1; | 1390 } |
1415 | 1391 avg= (min + max + 1)>>1; |
1416 if(max - min <deringThreshold) return; | 1392 |
1417 | 1393 if(max - min <deringThreshold) return; |
1418 for(y=0; y<10; y++) | 1394 |
1419 { | 1395 for(y=0; y<10; y++){ |
1420 int t = 0; | 1396 int t = 0; |
1421 | 1397 |
1422 if(src[stride*y + 0] > avg) t+= 1; | 1398 if(src[stride*y + 0] > avg) t+= 1; |
1423 if(src[stride*y + 1] > avg) t+= 2; | 1399 if(src[stride*y + 1] > avg) t+= 2; |
1424 if(src[stride*y + 2] > avg) t+= 4; | 1400 if(src[stride*y + 2] > avg) t+= 4; |
1425 if(src[stride*y + 3] > avg) t+= 8; | 1401 if(src[stride*y + 3] > avg) t+= 8; |
1426 if(src[stride*y + 4] > avg) t+= 16; | 1402 if(src[stride*y + 4] > avg) t+= 16; |
1427 if(src[stride*y + 5] > avg) t+= 32; | 1403 if(src[stride*y + 5] > avg) t+= 32; |
1428 if(src[stride*y + 6] > avg) t+= 64; | 1404 if(src[stride*y + 6] > avg) t+= 64; |
1429 if(src[stride*y + 7] > avg) t+= 128; | 1405 if(src[stride*y + 7] > avg) t+= 128; |
1430 if(src[stride*y + 8] > avg) t+= 256; | 1406 if(src[stride*y + 8] > avg) t+= 256; |
1431 if(src[stride*y + 9] > avg) t+= 512; | 1407 if(src[stride*y + 9] > avg) t+= 512; |
1432 | 1408 |
1433 t |= (~t)<<16; | 1409 t |= (~t)<<16; |
1434 t &= (t<<1) & (t>>1); | 1410 t &= (t<<1) & (t>>1); |
1435 s[y] = t; | 1411 s[y] = t; |
1412 } | |
1413 | |
1414 for(y=1; y<9; y++){ | |
1415 int t = s[y-1] & s[y] & s[y+1]; | |
1416 t|= t>>16; | |
1417 s[y-1]= t; | |
1418 } | |
1419 | |
1420 for(y=1; y<9; y++){ | |
1421 int x; | |
1422 int t = s[y-1]; | |
1423 | |
1424 p= src + stride*y; | |
1425 for(x=1; x<9; x++){ | |
1426 p++; | |
1427 if(t & (1<<x)){ | |
1428 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1)) | |
1429 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1)) | |
1430 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1)); | |
1431 f= (f + 8)>>4; | |
1432 | |
1433 #ifdef DEBUG_DERING_THRESHOLD | |
1434 asm volatile("emms\n\t":); | |
1435 { | |
1436 static long long numPixels=0; | |
1437 if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++; | |
1438 // if((max-min)<20 || (max-min)*QP<200) | |
1439 // if((max-min)*QP < 500) | |
1440 // if(max-min<QP/2) | |
1441 if(max-min < 20){ | |
1442 static int numSkiped=0; | |
1443 static int errorSum=0; | |
1444 static int worstQP=0; | |
1445 static int worstRange=0; | |
1446 static int worstDiff=0; | |
1447 int diff= (f - *p); | |
1448 int absDiff= FFABS(diff); | |
1449 int error= diff*diff; | |
1450 | |
1451 if(x==1 || x==8 || y==1 || y==8) continue; | |
1452 | |
1453 numSkiped++; | |
1454 if(absDiff > worstDiff){ | |
1455 worstDiff= absDiff; | |
1456 worstQP= QP; | |
1457 worstRange= max-min; | |
1458 } | |
1459 errorSum+= error; | |
1460 | |
1461 if(1024LL*1024LL*1024LL % numSkiped == 0){ | |
1462 av_log(c, AV_LOG_INFO, "sum:%1.3f, skip:%d, wQP:%d, " | |
1463 "wRange:%d, wDiff:%d, relSkip:%1.3f\n", | |
1464 (float)errorSum/numSkiped, numSkiped, worstQP, worstRange, | |
1465 worstDiff, (float)numSkiped/numPixels); | |
1466 } | |
1467 } | |
1468 } | |
1469 #endif | |
1470 if (*p + QP2 < f) *p= *p + QP2; | |
1471 else if(*p - QP2 > f) *p= *p - QP2; | |
1472 else *p=f; | |
1473 } | |
1436 } | 1474 } |
1437 | 1475 } |
1438 for(y=1; y<9; y++) | 1476 #ifdef DEBUG_DERING_THRESHOLD |
1439 { | 1477 if(max-min < 20){ |
1440 int t = s[y-1] & s[y] & s[y+1]; | 1478 for(y=1; y<9; y++){ |
1441 t|= t>>16; | 1479 int x; |
1442 s[y-1]= t; | 1480 int t = 0; |
1481 p= src + stride*y; | |
1482 for(x=1; x<9; x++){ | |
1483 p++; | |
1484 *p = FFMIN(*p + 20, 255); | |
1485 } | |
1443 } | 1486 } |
1444 | 1487 // src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255; |
1445 for(y=1; y<9; y++) | 1488 } |
1446 { | |
1447 int x; | |
1448 int t = s[y-1]; | |
1449 | |
1450 p= src + stride*y; | |
1451 for(x=1; x<9; x++) | |
1452 { | |
1453 p++; | |
1454 if(t & (1<<x)) | |
1455 { | |
1456 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1)) | |
1457 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1)) | |
1458 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1)); | |
1459 f= (f + 8)>>4; | |
1460 | |
1461 #ifdef DEBUG_DERING_THRESHOLD | |
1462 asm volatile("emms\n\t":); | |
1463 { | |
1464 static long long numPixels=0; | |
1465 if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++; | |
1466 // if((max-min)<20 || (max-min)*QP<200) | |
1467 // if((max-min)*QP < 500) | |
1468 // if(max-min<QP/2) | |
1469 if(max-min < 20) | |
1470 { | |
1471 static int numSkiped=0; | |
1472 static int errorSum=0; | |
1473 static int worstQP=0; | |
1474 static int worstRange=0; | |
1475 static int worstDiff=0; | |
1476 int diff= (f - *p); | |
1477 int absDiff= FFABS(diff); | |
1478 int error= diff*diff; | |
1479 | |
1480 if(x==1 || x==8 || y==1 || y==8) continue; | |
1481 | |
1482 numSkiped++; | |
1483 if(absDiff > worstDiff) | |
1484 { | |
1485 worstDiff= absDiff; | |
1486 worstQP= QP; | |
1487 worstRange= max-min; | |
1488 } | |
1489 errorSum+= error; | |
1490 | |
1491 if(1024LL*1024LL*1024LL % numSkiped == 0) | |
1492 { | |
1493 av_log(c, AV_LOG_INFO, "sum:%1.3f, skip:%d, wQP:%d, " | |
1494 "wRange:%d, wDiff:%d, relSkip:%1.3f\n", | |
1495 (float)errorSum/numSkiped, numSkiped, worstQP, worstRange, | |
1496 worstDiff, (float)numSkiped/numPixels); | |
1497 } | |
1498 } | |
1499 } | |
1500 #endif | |
1501 if (*p + QP2 < f) *p= *p + QP2; | |
1502 else if(*p - QP2 > f) *p= *p - QP2; | |
1503 else *p=f; | |
1504 } | |
1505 } | |
1506 } | |
1507 #ifdef DEBUG_DERING_THRESHOLD | |
1508 if(max-min < 20) | |
1509 { | |
1510 for(y=1; y<9; y++) | |
1511 { | |
1512 int x; | |
1513 int t = 0; | |
1514 p= src + stride*y; | |
1515 for(x=1; x<9; x++) | |
1516 { | |
1517 p++; | |
1518 *p = FFMIN(*p + 20, 255); | |
1519 } | |
1520 } | |
1521 // src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255; | |
1522 } | |
1523 #endif | 1489 #endif |
1524 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1490 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1525 } | 1491 } |
1526 #endif //HAVE_ALTIVEC | 1492 #endif //HAVE_ALTIVEC |
1527 | 1493 |
1532 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | 1498 * lines 4-12 will be read into the deblocking filter and should be deinterlaced |
1533 */ | 1499 */ |
1534 static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride) | 1500 static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride) |
1535 { | 1501 { |
1536 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1502 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1537 src+= 4*stride; | 1503 src+= 4*stride; |
1538 asm volatile( | 1504 asm volatile( |
1539 "lea (%0, %1), %%"REG_a" \n\t" | 1505 "lea (%0, %1), %%"REG_a" \n\t" |
1540 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" | 1506 "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" |
1541 // 0 1 2 3 4 5 6 7 8 9 | 1507 // 0 1 2 3 4 5 6 7 8 9 |
1542 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1 | 1508 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1 |
1543 | 1509 |
1544 "movq (%0), %%mm0 \n\t" | 1510 "movq (%0), %%mm0 \n\t" |
1545 "movq (%%"REG_a", %1), %%mm1 \n\t" | 1511 "movq (%%"REG_a", %1), %%mm1 \n\t" |
1546 PAVGB(%%mm1, %%mm0) | 1512 PAVGB(%%mm1, %%mm0) |
1547 "movq %%mm0, (%%"REG_a") \n\t" | 1513 "movq %%mm0, (%%"REG_a") \n\t" |
1548 "movq (%0, %1, 4), %%mm0 \n\t" | 1514 "movq (%0, %1, 4), %%mm0 \n\t" |
1549 PAVGB(%%mm0, %%mm1) | 1515 PAVGB(%%mm0, %%mm1) |
1550 "movq %%mm1, (%%"REG_a", %1, 2) \n\t" | 1516 "movq %%mm1, (%%"REG_a", %1, 2) \n\t" |
1551 "movq (%%"REG_c", %1), %%mm1 \n\t" | 1517 "movq (%%"REG_c", %1), %%mm1 \n\t" |
1552 PAVGB(%%mm1, %%mm0) | 1518 PAVGB(%%mm1, %%mm0) |
1553 "movq %%mm0, (%%"REG_c") \n\t" | 1519 "movq %%mm0, (%%"REG_c") \n\t" |
1554 "movq (%0, %1, 8), %%mm0 \n\t" | 1520 "movq (%0, %1, 8), %%mm0 \n\t" |
1555 PAVGB(%%mm0, %%mm1) | 1521 PAVGB(%%mm0, %%mm1) |
1556 "movq %%mm1, (%%"REG_c", %1, 2) \n\t" | 1522 "movq %%mm1, (%%"REG_c", %1, 2) \n\t" |
1557 | 1523 |
1558 : : "r" (src), "r" ((long)stride) | 1524 : : "r" (src), "r" ((long)stride) |
1559 : "%"REG_a, "%"REG_c | 1525 : "%"REG_a, "%"REG_c |
1560 ); | 1526 ); |
1561 #else | 1527 #else |
1562 int a, b, x; | 1528 int a, b, x; |
1563 src+= 4*stride; | 1529 src+= 4*stride; |
1564 | 1530 |
1565 for(x=0; x<2; x++){ | 1531 for(x=0; x<2; x++){ |
1566 a= *(uint32_t*)&src[stride*0]; | 1532 a= *(uint32_t*)&src[stride*0]; |
1567 b= *(uint32_t*)&src[stride*2]; | 1533 b= *(uint32_t*)&src[stride*2]; |
1568 *(uint32_t*)&src[stride*1]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); | 1534 *(uint32_t*)&src[stride*1]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); |
1569 a= *(uint32_t*)&src[stride*4]; | 1535 a= *(uint32_t*)&src[stride*4]; |
1570 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); | 1536 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); |
1571 b= *(uint32_t*)&src[stride*6]; | 1537 b= *(uint32_t*)&src[stride*6]; |
1572 *(uint32_t*)&src[stride*5]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); | 1538 *(uint32_t*)&src[stride*5]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); |
1573 a= *(uint32_t*)&src[stride*8]; | 1539 a= *(uint32_t*)&src[stride*8]; |
1574 *(uint32_t*)&src[stride*7]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); | 1540 *(uint32_t*)&src[stride*7]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); |
1575 src += 4; | 1541 src += 4; |
1576 } | 1542 } |
1577 #endif | 1543 #endif |
1578 } | 1544 } |
1579 | 1545 |
1580 /** | 1546 /** |
1581 * Deinterlaces the given block by cubic interpolating every second line. | 1547 * Deinterlaces the given block by cubic interpolating every second line. |
1585 * this filter will read lines 3-15 and write 7-13 | 1551 * this filter will read lines 3-15 and write 7-13 |
1586 */ | 1552 */ |
1587 static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride) | 1553 static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride) |
1588 { | 1554 { |
1589 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1555 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1590 src+= stride*3; | 1556 src+= stride*3; |
1591 asm volatile( | 1557 asm volatile( |
1592 "lea (%0, %1), %%"REG_a" \n\t" | 1558 "lea (%0, %1), %%"REG_a" \n\t" |
1593 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" | 1559 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" |
1594 "lea (%%"REG_d", %1, 4), %%"REG_c" \n\t" | 1560 "lea (%%"REG_d", %1, 4), %%"REG_c" \n\t" |
1595 "add %1, %%"REG_c" \n\t" | 1561 "add %1, %%"REG_c" \n\t" |
1596 "pxor %%mm7, %%mm7 \n\t" | 1562 "pxor %%mm7, %%mm7 \n\t" |
1597 // 0 1 2 3 4 5 6 7 8 9 10 | 1563 // 0 1 2 3 4 5 6 7 8 9 10 |
1598 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx | 1564 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx |
1599 | 1565 |
1600 #define REAL_DEINT_CUBIC(a,b,c,d,e)\ | 1566 #define REAL_DEINT_CUBIC(a,b,c,d,e)\ |
1601 "movq " #a ", %%mm0 \n\t"\ | 1567 "movq " #a ", %%mm0 \n\t"\ |
1602 "movq " #b ", %%mm1 \n\t"\ | 1568 "movq " #b ", %%mm1 \n\t"\ |
1603 "movq " #d ", %%mm2 \n\t"\ | 1569 "movq " #d ", %%mm2 \n\t"\ |
1604 "movq " #e ", %%mm3 \n\t"\ | 1570 "movq " #e ", %%mm3 \n\t"\ |
1605 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\ | 1571 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\ |
1606 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\ | 1572 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\ |
1607 "movq %%mm0, %%mm2 \n\t"\ | 1573 "movq %%mm0, %%mm2 \n\t"\ |
1608 "punpcklbw %%mm7, %%mm0 \n\t"\ | 1574 "punpcklbw %%mm7, %%mm0 \n\t"\ |
1609 "punpckhbw %%mm7, %%mm2 \n\t"\ | 1575 "punpckhbw %%mm7, %%mm2 \n\t"\ |
1610 "movq %%mm1, %%mm3 \n\t"\ | 1576 "movq %%mm1, %%mm3 \n\t"\ |
1611 "punpcklbw %%mm7, %%mm1 \n\t"\ | 1577 "punpcklbw %%mm7, %%mm1 \n\t"\ |
1612 "punpckhbw %%mm7, %%mm3 \n\t"\ | 1578 "punpckhbw %%mm7, %%mm3 \n\t"\ |
1613 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\ | 1579 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\ |
1614 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\ | 1580 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\ |
1615 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\ | 1581 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\ |
1616 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\ | 1582 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\ |
1617 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\ | 1583 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\ |
1618 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\ | 1584 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\ |
1619 "packuswb %%mm3, %%mm1 \n\t"\ | 1585 "packuswb %%mm3, %%mm1 \n\t"\ |
1620 "movq %%mm1, " #c " \n\t" | 1586 "movq %%mm1, " #c " \n\t" |
1621 #define DEINT_CUBIC(a,b,c,d,e) REAL_DEINT_CUBIC(a,b,c,d,e) | 1587 #define DEINT_CUBIC(a,b,c,d,e) REAL_DEINT_CUBIC(a,b,c,d,e) |
1622 | 1588 |
1623 DEINT_CUBIC((%0) , (%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd, %1)) | 1589 DEINT_CUBIC((%0) , (%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd, %1)) |
1624 DEINT_CUBIC((%%REGa, %1), (%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%0, %1, 8)) | 1590 DEINT_CUBIC((%%REGa, %1), (%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%0, %1, 8)) |
1625 DEINT_CUBIC((%0, %1, 4) , (%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGc)) | 1591 DEINT_CUBIC((%0, %1, 4) , (%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGc)) |
1626 DEINT_CUBIC((%%REGd, %1), (%0, %1, 8) , (%%REGd, %1, 4), (%%REGc) , (%%REGc, %1, 2)) | 1592 DEINT_CUBIC((%%REGd, %1), (%0, %1, 8) , (%%REGd, %1, 4), (%%REGc) , (%%REGc, %1, 2)) |
1627 | 1593 |
1628 : : "r" (src), "r" ((long)stride) | 1594 : : "r" (src), "r" ((long)stride) |
1629 : "%"REG_a, "%"REG_d, "%"REG_c | 1595 : "%"REG_a, "%"REG_d, "%"REG_c |
1630 ); | 1596 ); |
1631 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1597 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1632 int x; | 1598 int x; |
1633 src+= stride*3; | 1599 src+= stride*3; |
1634 for(x=0; x<8; x++) | 1600 for(x=0; x<8; x++){ |
1635 { | 1601 src[stride*3] = CLIP((-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4); |
1636 src[stride*3] = CLIP((-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4); | 1602 src[stride*5] = CLIP((-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4); |
1637 src[stride*5] = CLIP((-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4); | 1603 src[stride*7] = CLIP((-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4); |
1638 src[stride*7] = CLIP((-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4); | 1604 src[stride*9] = CLIP((-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4); |
1639 src[stride*9] = CLIP((-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4); | 1605 src++; |
1640 src++; | 1606 } |
1641 } | |
1642 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1607 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1643 } | 1608 } |
1644 | 1609 |
1645 /** | 1610 /** |
1646 * Deinterlaces the given block by filtering every second line with a (-1 4 2 4 -1) filter. | 1611 * Deinterlaces the given block by filtering every second line with a (-1 4 2 4 -1) filter. |
1650 * this filter will read lines 4-13 and write 5-11 | 1615 * this filter will read lines 4-13 and write 5-11 |
1651 */ | 1616 */ |
1652 static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp) | 1617 static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp) |
1653 { | 1618 { |
1654 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1619 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1655 src+= stride*4; | 1620 src+= stride*4; |
1656 asm volatile( | 1621 asm volatile( |
1657 "lea (%0, %1), %%"REG_a" \n\t" | 1622 "lea (%0, %1), %%"REG_a" \n\t" |
1658 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" | 1623 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" |
1659 "pxor %%mm7, %%mm7 \n\t" | 1624 "pxor %%mm7, %%mm7 \n\t" |
1660 "movq (%2), %%mm0 \n\t" | 1625 "movq (%2), %%mm0 \n\t" |
1661 // 0 1 2 3 4 5 6 7 8 9 10 | 1626 // 0 1 2 3 4 5 6 7 8 9 10 |
1662 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx | 1627 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx |
1663 | 1628 |
1664 #define REAL_DEINT_FF(a,b,c,d)\ | 1629 #define REAL_DEINT_FF(a,b,c,d)\ |
1665 "movq " #a ", %%mm1 \n\t"\ | 1630 "movq " #a ", %%mm1 \n\t"\ |
1666 "movq " #b ", %%mm2 \n\t"\ | 1631 "movq " #b ", %%mm2 \n\t"\ |
1667 "movq " #c ", %%mm3 \n\t"\ | 1632 "movq " #c ", %%mm3 \n\t"\ |
1668 "movq " #d ", %%mm4 \n\t"\ | 1633 "movq " #d ", %%mm4 \n\t"\ |
1669 PAVGB(%%mm3, %%mm1) \ | 1634 PAVGB(%%mm3, %%mm1) \ |
1670 PAVGB(%%mm4, %%mm0) \ | 1635 PAVGB(%%mm4, %%mm0) \ |
1671 "movq %%mm0, %%mm3 \n\t"\ | 1636 "movq %%mm0, %%mm3 \n\t"\ |
1672 "punpcklbw %%mm7, %%mm0 \n\t"\ | 1637 "punpcklbw %%mm7, %%mm0 \n\t"\ |
1673 "punpckhbw %%mm7, %%mm3 \n\t"\ | 1638 "punpckhbw %%mm7, %%mm3 \n\t"\ |
1674 "movq %%mm1, %%mm4 \n\t"\ | 1639 "movq %%mm1, %%mm4 \n\t"\ |
1675 "punpcklbw %%mm7, %%mm1 \n\t"\ | 1640 "punpcklbw %%mm7, %%mm1 \n\t"\ |
1676 "punpckhbw %%mm7, %%mm4 \n\t"\ | 1641 "punpckhbw %%mm7, %%mm4 \n\t"\ |
1677 "psllw $2, %%mm1 \n\t"\ | 1642 "psllw $2, %%mm1 \n\t"\ |
1678 "psllw $2, %%mm4 \n\t"\ | 1643 "psllw $2, %%mm4 \n\t"\ |
1679 "psubw %%mm0, %%mm1 \n\t"\ | 1644 "psubw %%mm0, %%mm1 \n\t"\ |
1680 "psubw %%mm3, %%mm4 \n\t"\ | 1645 "psubw %%mm3, %%mm4 \n\t"\ |
1681 "movq %%mm2, %%mm5 \n\t"\ | 1646 "movq %%mm2, %%mm5 \n\t"\ |
1682 "movq %%mm2, %%mm0 \n\t"\ | 1647 "movq %%mm2, %%mm0 \n\t"\ |
1683 "punpcklbw %%mm7, %%mm2 \n\t"\ | 1648 "punpcklbw %%mm7, %%mm2 \n\t"\ |
1684 "punpckhbw %%mm7, %%mm5 \n\t"\ | 1649 "punpckhbw %%mm7, %%mm5 \n\t"\ |
1685 "paddw %%mm2, %%mm1 \n\t"\ | 1650 "paddw %%mm2, %%mm1 \n\t"\ |
1686 "paddw %%mm5, %%mm4 \n\t"\ | 1651 "paddw %%mm5, %%mm4 \n\t"\ |
1687 "psraw $2, %%mm1 \n\t"\ | 1652 "psraw $2, %%mm1 \n\t"\ |
1688 "psraw $2, %%mm4 \n\t"\ | 1653 "psraw $2, %%mm4 \n\t"\ |
1689 "packuswb %%mm4, %%mm1 \n\t"\ | 1654 "packuswb %%mm4, %%mm1 \n\t"\ |
1690 "movq %%mm1, " #b " \n\t"\ | 1655 "movq %%mm1, " #b " \n\t"\ |
1691 | 1656 |
1692 #define DEINT_FF(a,b,c,d) REAL_DEINT_FF(a,b,c,d) | 1657 #define DEINT_FF(a,b,c,d) REAL_DEINT_FF(a,b,c,d) |
1693 | 1658 |
1694 DEINT_FF((%0) , (%%REGa) , (%%REGa, %1), (%%REGa, %1, 2)) | 1659 DEINT_FF((%0) , (%%REGa) , (%%REGa, %1), (%%REGa, %1, 2)) |
1695 DEINT_FF((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) ) | 1660 DEINT_FF((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) ) |
1696 DEINT_FF((%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%%REGd, %1, 2)) | 1661 DEINT_FF((%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%%REGd, %1, 2)) |
1697 DEINT_FF((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4)) | 1662 DEINT_FF((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4)) |
1698 | 1663 |
1699 "movq %%mm0, (%2) \n\t" | 1664 "movq %%mm0, (%2) \n\t" |
1700 : : "r" (src), "r" ((long)stride), "r"(tmp) | 1665 : : "r" (src), "r" ((long)stride), "r"(tmp) |
1701 : "%"REG_a, "%"REG_d | 1666 : "%"REG_a, "%"REG_d |
1702 ); | 1667 ); |
1703 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1668 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1704 int x; | 1669 int x; |
1705 src+= stride*4; | 1670 src+= stride*4; |
1706 for(x=0; x<8; x++) | 1671 for(x=0; x<8; x++){ |
1707 { | 1672 int t1= tmp[x]; |
1708 int t1= tmp[x]; | 1673 int t2= src[stride*1]; |
1709 int t2= src[stride*1]; | 1674 |
1710 | 1675 src[stride*1]= CLIP((-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3); |
1711 src[stride*1]= CLIP((-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3); | 1676 t1= src[stride*4]; |
1712 t1= src[stride*4]; | 1677 src[stride*3]= CLIP((-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3); |
1713 src[stride*3]= CLIP((-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3); | 1678 t2= src[stride*6]; |
1714 t2= src[stride*6]; | 1679 src[stride*5]= CLIP((-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3); |
1715 src[stride*5]= CLIP((-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3); | 1680 t1= src[stride*8]; |
1716 t1= src[stride*8]; | 1681 src[stride*7]= CLIP((-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3); |
1717 src[stride*7]= CLIP((-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3); | 1682 tmp[x]= t1; |
1718 tmp[x]= t1; | 1683 |
1719 | 1684 src++; |
1720 src++; | 1685 } |
1721 } | |
1722 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1686 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1723 } | 1687 } |
1724 | 1688 |
1725 /** | 1689 /** |
1726 * Deinterlaces the given block by filtering every line with a (-1 2 6 2 -1) filter. | 1690 * Deinterlaces the given block by filtering every line with a (-1 2 6 2 -1) filter. |
1730 * this filter will read lines 4-13 and write 4-11 | 1694 * this filter will read lines 4-13 and write 4-11 |
1731 */ | 1695 */ |
1732 static inline void RENAME(deInterlaceL5)(uint8_t src[], int stride, uint8_t *tmp, uint8_t *tmp2) | 1696 static inline void RENAME(deInterlaceL5)(uint8_t src[], int stride, uint8_t *tmp, uint8_t *tmp2) |
1733 { | 1697 { |
1734 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1698 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1735 src+= stride*4; | 1699 src+= stride*4; |
1736 asm volatile( | 1700 asm volatile( |
1737 "lea (%0, %1), %%"REG_a" \n\t" | 1701 "lea (%0, %1), %%"REG_a" \n\t" |
1738 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" | 1702 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" |
1739 "pxor %%mm7, %%mm7 \n\t" | 1703 "pxor %%mm7, %%mm7 \n\t" |
1740 "movq (%2), %%mm0 \n\t" | 1704 "movq (%2), %%mm0 \n\t" |
1741 "movq (%3), %%mm1 \n\t" | 1705 "movq (%3), %%mm1 \n\t" |
1742 // 0 1 2 3 4 5 6 7 8 9 10 | 1706 // 0 1 2 3 4 5 6 7 8 9 10 |
1743 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx | 1707 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx |
1744 | 1708 |
1745 #define REAL_DEINT_L5(t1,t2,a,b,c)\ | 1709 #define REAL_DEINT_L5(t1,t2,a,b,c)\ |
1746 "movq " #a ", %%mm2 \n\t"\ | 1710 "movq " #a ", %%mm2 \n\t"\ |
1747 "movq " #b ", %%mm3 \n\t"\ | 1711 "movq " #b ", %%mm3 \n\t"\ |
1748 "movq " #c ", %%mm4 \n\t"\ | 1712 "movq " #c ", %%mm4 \n\t"\ |
1749 PAVGB(t2, %%mm3) \ | 1713 PAVGB(t2, %%mm3) \ |
1750 PAVGB(t1, %%mm4) \ | 1714 PAVGB(t1, %%mm4) \ |
1751 "movq %%mm2, %%mm5 \n\t"\ | 1715 "movq %%mm2, %%mm5 \n\t"\ |
1752 "movq %%mm2, " #t1 " \n\t"\ | 1716 "movq %%mm2, " #t1 " \n\t"\ |
1753 "punpcklbw %%mm7, %%mm2 \n\t"\ | 1717 "punpcklbw %%mm7, %%mm2 \n\t"\ |
1754 "punpckhbw %%mm7, %%mm5 \n\t"\ | 1718 "punpckhbw %%mm7, %%mm5 \n\t"\ |
1755 "movq %%mm2, %%mm6 \n\t"\ | 1719 "movq %%mm2, %%mm6 \n\t"\ |
1756 "paddw %%mm2, %%mm2 \n\t"\ | 1720 "paddw %%mm2, %%mm2 \n\t"\ |
1757 "paddw %%mm6, %%mm2 \n\t"\ | 1721 "paddw %%mm6, %%mm2 \n\t"\ |
1758 "movq %%mm5, %%mm6 \n\t"\ | 1722 "movq %%mm5, %%mm6 \n\t"\ |
1759 "paddw %%mm5, %%mm5 \n\t"\ | 1723 "paddw %%mm5, %%mm5 \n\t"\ |
1760 "paddw %%mm6, %%mm5 \n\t"\ | 1724 "paddw %%mm6, %%mm5 \n\t"\ |
1761 "movq %%mm3, %%mm6 \n\t"\ | 1725 "movq %%mm3, %%mm6 \n\t"\ |
1762 "punpcklbw %%mm7, %%mm3 \n\t"\ | 1726 "punpcklbw %%mm7, %%mm3 \n\t"\ |
1763 "punpckhbw %%mm7, %%mm6 \n\t"\ | 1727 "punpckhbw %%mm7, %%mm6 \n\t"\ |
1764 "paddw %%mm3, %%mm3 \n\t"\ | 1728 "paddw %%mm3, %%mm3 \n\t"\ |
1765 "paddw %%mm6, %%mm6 \n\t"\ | 1729 "paddw %%mm6, %%mm6 \n\t"\ |
1766 "paddw %%mm3, %%mm2 \n\t"\ | 1730 "paddw %%mm3, %%mm2 \n\t"\ |
1767 "paddw %%mm6, %%mm5 \n\t"\ | 1731 "paddw %%mm6, %%mm5 \n\t"\ |
1768 "movq %%mm4, %%mm6 \n\t"\ | 1732 "movq %%mm4, %%mm6 \n\t"\ |
1769 "punpcklbw %%mm7, %%mm4 \n\t"\ | 1733 "punpcklbw %%mm7, %%mm4 \n\t"\ |
1770 "punpckhbw %%mm7, %%mm6 \n\t"\ | 1734 "punpckhbw %%mm7, %%mm6 \n\t"\ |
1771 "psubw %%mm4, %%mm2 \n\t"\ | 1735 "psubw %%mm4, %%mm2 \n\t"\ |
1772 "psubw %%mm6, %%mm5 \n\t"\ | 1736 "psubw %%mm6, %%mm5 \n\t"\ |
1773 "psraw $2, %%mm2 \n\t"\ | 1737 "psraw $2, %%mm2 \n\t"\ |
1774 "psraw $2, %%mm5 \n\t"\ | 1738 "psraw $2, %%mm5 \n\t"\ |
1775 "packuswb %%mm5, %%mm2 \n\t"\ | 1739 "packuswb %%mm5, %%mm2 \n\t"\ |
1776 "movq %%mm2, " #a " \n\t"\ | 1740 "movq %%mm2, " #a " \n\t"\ |
1777 | 1741 |
1778 #define DEINT_L5(t1,t2,a,b,c) REAL_DEINT_L5(t1,t2,a,b,c) | 1742 #define DEINT_L5(t1,t2,a,b,c) REAL_DEINT_L5(t1,t2,a,b,c) |
1779 | 1743 |
1780 DEINT_L5(%%mm0, %%mm1, (%0) , (%%REGa) , (%%REGa, %1) ) | 1744 DEINT_L5(%%mm0, %%mm1, (%0) , (%%REGa) , (%%REGa, %1) ) |
1781 DEINT_L5(%%mm1, %%mm0, (%%REGa) , (%%REGa, %1) , (%%REGa, %1, 2)) | 1745 DEINT_L5(%%mm1, %%mm0, (%%REGa) , (%%REGa, %1) , (%%REGa, %1, 2)) |
1784 DEINT_L5(%%mm0, %%mm1, (%0, %1, 4) , (%%REGd) , (%%REGd, %1) ) | 1748 DEINT_L5(%%mm0, %%mm1, (%0, %1, 4) , (%%REGd) , (%%REGd, %1) ) |
1785 DEINT_L5(%%mm1, %%mm0, (%%REGd) , (%%REGd, %1) , (%%REGd, %1, 2)) | 1749 DEINT_L5(%%mm1, %%mm0, (%%REGd) , (%%REGd, %1) , (%%REGd, %1, 2)) |
1786 DEINT_L5(%%mm0, %%mm1, (%%REGd, %1) , (%%REGd, %1, 2), (%0, %1, 8) ) | 1750 DEINT_L5(%%mm0, %%mm1, (%%REGd, %1) , (%%REGd, %1, 2), (%0, %1, 8) ) |
1787 DEINT_L5(%%mm1, %%mm0, (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4)) | 1751 DEINT_L5(%%mm1, %%mm0, (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4)) |
1788 | 1752 |
1789 "movq %%mm0, (%2) \n\t" | 1753 "movq %%mm0, (%2) \n\t" |
1790 "movq %%mm1, (%3) \n\t" | 1754 "movq %%mm1, (%3) \n\t" |
1791 : : "r" (src), "r" ((long)stride), "r"(tmp), "r"(tmp2) | 1755 : : "r" (src), "r" ((long)stride), "r"(tmp), "r"(tmp2) |
1792 : "%"REG_a, "%"REG_d | 1756 : "%"REG_a, "%"REG_d |
1793 ); | 1757 ); |
1794 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1758 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1795 int x; | 1759 int x; |
1796 src+= stride*4; | 1760 src+= stride*4; |
1797 for(x=0; x<8; x++) | 1761 for(x=0; x<8; x++){ |
1798 { | 1762 int t1= tmp[x]; |
1799 int t1= tmp[x]; | 1763 int t2= tmp2[x]; |
1800 int t2= tmp2[x]; | 1764 int t3= src[0]; |
1801 int t3= src[0]; | 1765 |
1802 | 1766 src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3); |
1803 src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3); | 1767 t1= src[stride*1]; |
1804 t1= src[stride*1]; | 1768 src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3); |
1805 src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3); | 1769 t2= src[stride*2]; |
1806 t2= src[stride*2]; | 1770 src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3); |
1807 src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3); | 1771 t3= src[stride*3]; |
1808 t3= src[stride*3]; | 1772 src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3); |
1809 src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3); | 1773 t1= src[stride*4]; |
1810 t1= src[stride*4]; | 1774 src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3); |
1811 src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3); | 1775 t2= src[stride*5]; |
1812 t2= src[stride*5]; | 1776 src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3); |
1813 src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3); | 1777 t3= src[stride*6]; |
1814 t3= src[stride*6]; | 1778 src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3); |
1815 src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3); | 1779 t1= src[stride*7]; |
1816 t1= src[stride*7]; | 1780 src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3); |
1817 src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3); | 1781 |
1818 | 1782 tmp[x]= t3; |
1819 tmp[x]= t3; | 1783 tmp2[x]= t1; |
1820 tmp2[x]= t1; | 1784 |
1821 | 1785 src++; |
1822 src++; | 1786 } |
1823 } | |
1824 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1787 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1825 } | 1788 } |
1826 | 1789 |
1827 /** | 1790 /** |
1828 * Deinterlaces the given block by filtering all lines with a (1 2 1) filter. | 1791 * Deinterlaces the given block by filtering all lines with a (1 2 1) filter. |
1832 * this filter will read lines 4-13 and write 4-11 | 1795 * this filter will read lines 4-13 and write 4-11 |
1833 */ | 1796 */ |
1834 static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride, uint8_t *tmp) | 1797 static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride, uint8_t *tmp) |
1835 { | 1798 { |
1836 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1799 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1837 src+= 4*stride; | 1800 src+= 4*stride; |
1838 asm volatile( | 1801 asm volatile( |
1839 "lea (%0, %1), %%"REG_a" \n\t" | 1802 "lea (%0, %1), %%"REG_a" \n\t" |
1840 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" | 1803 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" |
1841 // 0 1 2 3 4 5 6 7 8 9 | 1804 // 0 1 2 3 4 5 6 7 8 9 |
1842 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 | 1805 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 |
1843 | 1806 |
1844 "movq (%2), %%mm0 \n\t" // L0 | 1807 "movq (%2), %%mm0 \n\t" // L0 |
1845 "movq (%%"REG_a"), %%mm1 \n\t" // L2 | 1808 "movq (%%"REG_a"), %%mm1 \n\t" // L2 |
1846 PAVGB(%%mm1, %%mm0) // L0+L2 | 1809 PAVGB(%%mm1, %%mm0) // L0+L2 |
1847 "movq (%0), %%mm2 \n\t" // L1 | 1810 "movq (%0), %%mm2 \n\t" // L1 |
1848 PAVGB(%%mm2, %%mm0) | 1811 PAVGB(%%mm2, %%mm0) |
1849 "movq %%mm0, (%0) \n\t" | 1812 "movq %%mm0, (%0) \n\t" |
1850 "movq (%%"REG_a", %1), %%mm0 \n\t" // L3 | 1813 "movq (%%"REG_a", %1), %%mm0 \n\t" // L3 |
1851 PAVGB(%%mm0, %%mm2) // L1+L3 | 1814 PAVGB(%%mm0, %%mm2) // L1+L3 |
1852 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3 | 1815 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3 |
1853 "movq %%mm2, (%%"REG_a") \n\t" | 1816 "movq %%mm2, (%%"REG_a") \n\t" |
1854 "movq (%%"REG_a", %1, 2), %%mm2 \n\t" // L4 | 1817 "movq (%%"REG_a", %1, 2), %%mm2 \n\t" // L4 |
1855 PAVGB(%%mm2, %%mm1) // L2+L4 | 1818 PAVGB(%%mm2, %%mm1) // L2+L4 |
1856 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4 | 1819 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4 |
1857 "movq %%mm1, (%%"REG_a", %1) \n\t" | 1820 "movq %%mm1, (%%"REG_a", %1) \n\t" |
1858 "movq (%0, %1, 4), %%mm1 \n\t" // L5 | 1821 "movq (%0, %1, 4), %%mm1 \n\t" // L5 |
1859 PAVGB(%%mm1, %%mm0) // L3+L5 | 1822 PAVGB(%%mm1, %%mm0) // L3+L5 |
1860 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5 | 1823 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5 |
1861 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" | 1824 "movq %%mm0, (%%"REG_a", %1, 2) \n\t" |
1862 "movq (%%"REG_d"), %%mm0 \n\t" // L6 | 1825 "movq (%%"REG_d"), %%mm0 \n\t" // L6 |
1863 PAVGB(%%mm0, %%mm2) // L4+L6 | 1826 PAVGB(%%mm0, %%mm2) // L4+L6 |
1864 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6 | 1827 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6 |
1865 "movq %%mm2, (%0, %1, 4) \n\t" | 1828 "movq %%mm2, (%0, %1, 4) \n\t" |
1866 "movq (%%"REG_d", %1), %%mm2 \n\t" // L7 | 1829 "movq (%%"REG_d", %1), %%mm2 \n\t" // L7 |
1867 PAVGB(%%mm2, %%mm1) // L5+L7 | 1830 PAVGB(%%mm2, %%mm1) // L5+L7 |
1868 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7 | 1831 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7 |
1869 "movq %%mm1, (%%"REG_d") \n\t" | 1832 "movq %%mm1, (%%"REG_d") \n\t" |
1870 "movq (%%"REG_d", %1, 2), %%mm1 \n\t" // L8 | 1833 "movq (%%"REG_d", %1, 2), %%mm1 \n\t" // L8 |
1871 PAVGB(%%mm1, %%mm0) // L6+L8 | 1834 PAVGB(%%mm1, %%mm0) // L6+L8 |
1872 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8 | 1835 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8 |
1873 "movq %%mm0, (%%"REG_d", %1) \n\t" | 1836 "movq %%mm0, (%%"REG_d", %1) \n\t" |
1874 "movq (%0, %1, 8), %%mm0 \n\t" // L9 | 1837 "movq (%0, %1, 8), %%mm0 \n\t" // L9 |
1875 PAVGB(%%mm0, %%mm2) // L7+L9 | 1838 PAVGB(%%mm0, %%mm2) // L7+L9 |
1876 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9 | 1839 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9 |
1877 "movq %%mm2, (%%"REG_d", %1, 2) \n\t" | 1840 "movq %%mm2, (%%"REG_d", %1, 2) \n\t" |
1878 "movq %%mm1, (%2) \n\t" | 1841 "movq %%mm1, (%2) \n\t" |
1879 | 1842 |
1880 : : "r" (src), "r" ((long)stride), "r" (tmp) | 1843 : : "r" (src), "r" ((long)stride), "r" (tmp) |
1881 : "%"REG_a, "%"REG_d | 1844 : "%"REG_a, "%"REG_d |
1882 ); | 1845 ); |
1883 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1846 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1884 int a, b, c, x; | 1847 int a, b, c, x; |
1885 src+= 4*stride; | 1848 src+= 4*stride; |
1886 | 1849 |
1887 for(x=0; x<2; x++){ | 1850 for(x=0; x<2; x++){ |
1888 a= *(uint32_t*)&tmp[stride*0]; | 1851 a= *(uint32_t*)&tmp[stride*0]; |
1889 b= *(uint32_t*)&src[stride*0]; | 1852 b= *(uint32_t*)&src[stride*0]; |
1890 c= *(uint32_t*)&src[stride*1]; | 1853 c= *(uint32_t*)&src[stride*1]; |
1891 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1); | 1854 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1); |
1892 *(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); | 1855 *(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); |
1893 | 1856 |
1894 a= *(uint32_t*)&src[stride*2]; | 1857 a= *(uint32_t*)&src[stride*2]; |
1895 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1); | 1858 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1); |
1896 *(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1); | 1859 *(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1); |
1897 | 1860 |
1898 b= *(uint32_t*)&src[stride*3]; | 1861 b= *(uint32_t*)&src[stride*3]; |
1899 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1); | 1862 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1); |
1900 *(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1); | 1863 *(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1); |
1901 | 1864 |
1902 c= *(uint32_t*)&src[stride*4]; | 1865 c= *(uint32_t*)&src[stride*4]; |
1903 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1); | 1866 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1); |
1904 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); | 1867 *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); |
1905 | 1868 |
1906 a= *(uint32_t*)&src[stride*5]; | 1869 a= *(uint32_t*)&src[stride*5]; |
1907 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1); | 1870 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1); |
1908 *(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1); | 1871 *(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1); |
1909 | 1872 |
1910 b= *(uint32_t*)&src[stride*6]; | 1873 b= *(uint32_t*)&src[stride*6]; |
1911 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1); | 1874 c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1); |
1912 *(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1); | 1875 *(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1); |
1913 | 1876 |
1914 c= *(uint32_t*)&src[stride*7]; | 1877 c= *(uint32_t*)&src[stride*7]; |
1915 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1); | 1878 a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1); |
1916 *(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); | 1879 *(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); |
1917 | 1880 |
1918 a= *(uint32_t*)&src[stride*8]; | 1881 a= *(uint32_t*)&src[stride*8]; |
1919 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1); | 1882 b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1); |
1920 *(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1); | 1883 *(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1); |
1921 | 1884 |
1922 *(uint32_t*)&tmp[stride*0]= c; | 1885 *(uint32_t*)&tmp[stride*0]= c; |
1923 src += 4; | 1886 src += 4; |
1924 tmp += 4; | 1887 tmp += 4; |
1925 } | 1888 } |
1926 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 1889 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1927 } | 1890 } |
1928 | 1891 |
1929 /** | 1892 /** |
1930 * Deinterlaces the given block by applying a median filter to every second line. | 1893 * Deinterlaces the given block by applying a median filter to every second line. |
1933 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | 1896 * lines 4-12 will be read into the deblocking filter and should be deinterlaced |
1934 */ | 1897 */ |
1935 static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride) | 1898 static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride) |
1936 { | 1899 { |
1937 #ifdef HAVE_MMX | 1900 #ifdef HAVE_MMX |
1938 src+= 4*stride; | 1901 src+= 4*stride; |
1939 #ifdef HAVE_MMX2 | 1902 #ifdef HAVE_MMX2 |
1940 asm volatile( | 1903 asm volatile( |
1941 "lea (%0, %1), %%"REG_a" \n\t" | 1904 "lea (%0, %1), %%"REG_a" \n\t" |
1942 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" | 1905 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" |
1943 // 0 1 2 3 4 5 6 7 8 9 | 1906 // 0 1 2 3 4 5 6 7 8 9 |
1944 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 | 1907 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 |
1945 | 1908 |
1946 "movq (%0), %%mm0 \n\t" // | 1909 "movq (%0), %%mm0 \n\t" // |
1947 "movq (%%"REG_a", %1), %%mm2 \n\t" // | 1910 "movq (%%"REG_a", %1), %%mm2 \n\t" // |
1948 "movq (%%"REG_a"), %%mm1 \n\t" // | 1911 "movq (%%"REG_a"), %%mm1 \n\t" // |
1949 "movq %%mm0, %%mm3 \n\t" | 1912 "movq %%mm0, %%mm3 \n\t" |
1950 "pmaxub %%mm1, %%mm0 \n\t" // | 1913 "pmaxub %%mm1, %%mm0 \n\t" // |
1951 "pminub %%mm3, %%mm1 \n\t" // | 1914 "pminub %%mm3, %%mm1 \n\t" // |
1952 "pmaxub %%mm2, %%mm1 \n\t" // | 1915 "pmaxub %%mm2, %%mm1 \n\t" // |
1953 "pminub %%mm1, %%mm0 \n\t" | 1916 "pminub %%mm1, %%mm0 \n\t" |
1954 "movq %%mm0, (%%"REG_a") \n\t" | 1917 "movq %%mm0, (%%"REG_a") \n\t" |
1955 | 1918 |
1956 "movq (%0, %1, 4), %%mm0 \n\t" // | 1919 "movq (%0, %1, 4), %%mm0 \n\t" // |
1957 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // | 1920 "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // |
1958 "movq %%mm2, %%mm3 \n\t" | 1921 "movq %%mm2, %%mm3 \n\t" |
1959 "pmaxub %%mm1, %%mm2 \n\t" // | 1922 "pmaxub %%mm1, %%mm2 \n\t" // |
1960 "pminub %%mm3, %%mm1 \n\t" // | 1923 "pminub %%mm3, %%mm1 \n\t" // |
1961 "pmaxub %%mm0, %%mm1 \n\t" // | 1924 "pmaxub %%mm0, %%mm1 \n\t" // |
1962 "pminub %%mm1, %%mm2 \n\t" | 1925 "pminub %%mm1, %%mm2 \n\t" |
1963 "movq %%mm2, (%%"REG_a", %1, 2) \n\t" | 1926 "movq %%mm2, (%%"REG_a", %1, 2) \n\t" |
1964 | 1927 |
1965 "movq (%%"REG_d"), %%mm2 \n\t" // | 1928 "movq (%%"REG_d"), %%mm2 \n\t" // |
1966 "movq (%%"REG_d", %1), %%mm1 \n\t" // | 1929 "movq (%%"REG_d", %1), %%mm1 \n\t" // |
1967 "movq %%mm2, %%mm3 \n\t" | 1930 "movq %%mm2, %%mm3 \n\t" |
1968 "pmaxub %%mm0, %%mm2 \n\t" // | 1931 "pmaxub %%mm0, %%mm2 \n\t" // |
1969 "pminub %%mm3, %%mm0 \n\t" // | 1932 "pminub %%mm3, %%mm0 \n\t" // |
1970 "pmaxub %%mm1, %%mm0 \n\t" // | 1933 "pmaxub %%mm1, %%mm0 \n\t" // |
1971 "pminub %%mm0, %%mm2 \n\t" | 1934 "pminub %%mm0, %%mm2 \n\t" |
1972 "movq %%mm2, (%%"REG_d") \n\t" | 1935 "movq %%mm2, (%%"REG_d") \n\t" |
1973 | 1936 |
1974 "movq (%%"REG_d", %1, 2), %%mm2 \n\t" // | 1937 "movq (%%"REG_d", %1, 2), %%mm2 \n\t" // |
1975 "movq (%0, %1, 8), %%mm0 \n\t" // | 1938 "movq (%0, %1, 8), %%mm0 \n\t" // |
1976 "movq %%mm2, %%mm3 \n\t" | 1939 "movq %%mm2, %%mm3 \n\t" |
1977 "pmaxub %%mm0, %%mm2 \n\t" // | 1940 "pmaxub %%mm0, %%mm2 \n\t" // |
1978 "pminub %%mm3, %%mm0 \n\t" // | 1941 "pminub %%mm3, %%mm0 \n\t" // |
1979 "pmaxub %%mm1, %%mm0 \n\t" // | 1942 "pmaxub %%mm1, %%mm0 \n\t" // |
1980 "pminub %%mm0, %%mm2 \n\t" | 1943 "pminub %%mm0, %%mm2 \n\t" |
1981 "movq %%mm2, (%%"REG_d", %1, 2) \n\t" | 1944 "movq %%mm2, (%%"REG_d", %1, 2) \n\t" |
1982 | 1945 |
1983 | 1946 |
1984 : : "r" (src), "r" ((long)stride) | 1947 : : "r" (src), "r" ((long)stride) |
1985 : "%"REG_a, "%"REG_d | 1948 : "%"REG_a, "%"REG_d |
1986 ); | 1949 ); |
1987 | 1950 |
1988 #else // MMX without MMX2 | 1951 #else // MMX without MMX2 |
1989 asm volatile( | 1952 asm volatile( |
1990 "lea (%0, %1), %%"REG_a" \n\t" | 1953 "lea (%0, %1), %%"REG_a" \n\t" |
1991 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" | 1954 "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" |
1992 // 0 1 2 3 4 5 6 7 8 9 | 1955 // 0 1 2 3 4 5 6 7 8 9 |
1993 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 | 1956 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 |
1994 "pxor %%mm7, %%mm7 \n\t" | 1957 "pxor %%mm7, %%mm7 \n\t" |
1995 | 1958 |
1996 #define REAL_MEDIAN(a,b,c)\ | 1959 #define REAL_MEDIAN(a,b,c)\ |
1997 "movq " #a ", %%mm0 \n\t"\ | 1960 "movq " #a ", %%mm0 \n\t"\ |
1998 "movq " #b ", %%mm2 \n\t"\ | 1961 "movq " #b ", %%mm2 \n\t"\ |
1999 "movq " #c ", %%mm1 \n\t"\ | 1962 "movq " #c ", %%mm1 \n\t"\ |
2000 "movq %%mm0, %%mm3 \n\t"\ | 1963 "movq %%mm0, %%mm3 \n\t"\ |
2001 "movq %%mm1, %%mm4 \n\t"\ | 1964 "movq %%mm1, %%mm4 \n\t"\ |
2002 "movq %%mm2, %%mm5 \n\t"\ | 1965 "movq %%mm2, %%mm5 \n\t"\ |
2003 "psubusb %%mm1, %%mm3 \n\t"\ | 1966 "psubusb %%mm1, %%mm3 \n\t"\ |
2004 "psubusb %%mm2, %%mm4 \n\t"\ | 1967 "psubusb %%mm2, %%mm4 \n\t"\ |
2005 "psubusb %%mm0, %%mm5 \n\t"\ | 1968 "psubusb %%mm0, %%mm5 \n\t"\ |
2006 "pcmpeqb %%mm7, %%mm3 \n\t"\ | 1969 "pcmpeqb %%mm7, %%mm3 \n\t"\ |
2007 "pcmpeqb %%mm7, %%mm4 \n\t"\ | 1970 "pcmpeqb %%mm7, %%mm4 \n\t"\ |
2008 "pcmpeqb %%mm7, %%mm5 \n\t"\ | 1971 "pcmpeqb %%mm7, %%mm5 \n\t"\ |
2009 "movq %%mm3, %%mm6 \n\t"\ | 1972 "movq %%mm3, %%mm6 \n\t"\ |
2010 "pxor %%mm4, %%mm3 \n\t"\ | 1973 "pxor %%mm4, %%mm3 \n\t"\ |
2011 "pxor %%mm5, %%mm4 \n\t"\ | 1974 "pxor %%mm5, %%mm4 \n\t"\ |
2012 "pxor %%mm6, %%mm5 \n\t"\ | 1975 "pxor %%mm6, %%mm5 \n\t"\ |
2013 "por %%mm3, %%mm1 \n\t"\ | 1976 "por %%mm3, %%mm1 \n\t"\ |
2014 "por %%mm4, %%mm2 \n\t"\ | 1977 "por %%mm4, %%mm2 \n\t"\ |
2015 "por %%mm5, %%mm0 \n\t"\ | 1978 "por %%mm5, %%mm0 \n\t"\ |
2016 "pand %%mm2, %%mm0 \n\t"\ | 1979 "pand %%mm2, %%mm0 \n\t"\ |
2017 "pand %%mm1, %%mm0 \n\t"\ | 1980 "pand %%mm1, %%mm0 \n\t"\ |
2018 "movq %%mm0, " #b " \n\t" | 1981 "movq %%mm0, " #b " \n\t" |
2019 #define MEDIAN(a,b,c) REAL_MEDIAN(a,b,c) | 1982 #define MEDIAN(a,b,c) REAL_MEDIAN(a,b,c) |
2020 | 1983 |
2021 MEDIAN((%0) , (%%REGa) , (%%REGa, %1)) | 1984 MEDIAN((%0) , (%%REGa) , (%%REGa, %1)) |
2022 MEDIAN((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4)) | 1985 MEDIAN((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4)) |
2023 MEDIAN((%0, %1, 4) , (%%REGd) , (%%REGd, %1)) | 1986 MEDIAN((%0, %1, 4) , (%%REGd) , (%%REGd, %1)) |
2024 MEDIAN((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8)) | 1987 MEDIAN((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8)) |
2025 | 1988 |
2026 : : "r" (src), "r" ((long)stride) | 1989 : : "r" (src), "r" ((long)stride) |
2027 : "%"REG_a, "%"REG_d | 1990 : "%"REG_a, "%"REG_d |
2028 ); | 1991 ); |
2029 #endif //HAVE_MMX2 | 1992 #endif //HAVE_MMX2 |
2030 #else //HAVE_MMX | 1993 #else //HAVE_MMX |
2031 int x, y; | 1994 int x, y; |
2032 src+= 4*stride; | 1995 src+= 4*stride; |
2033 // FIXME - there should be a way to do a few columns in parallel like w/mmx | 1996 // FIXME - there should be a way to do a few columns in parallel like w/mmx |
2034 for(x=0; x<8; x++) | 1997 for(x=0; x<8; x++){ |
2035 { | 1998 uint8_t *colsrc = src; |
2036 uint8_t *colsrc = src; | 1999 for (y=0; y<4; y++){ |
2037 for (y=0; y<4; y++) | 2000 int a, b, c, d, e, f; |
2038 { | 2001 a = colsrc[0 ]; |
2039 int a, b, c, d, e, f; | 2002 b = colsrc[stride ]; |
2040 a = colsrc[0 ]; | 2003 c = colsrc[stride*2]; |
2041 b = colsrc[stride ]; | 2004 d = (a-b)>>31; |
2042 c = colsrc[stride*2]; | 2005 e = (b-c)>>31; |
2043 d = (a-b)>>31; | 2006 f = (c-a)>>31; |
2044 e = (b-c)>>31; | 2007 colsrc[stride ] = (a|(d^f)) & (b|(d^e)) & (c|(e^f)); |
2045 f = (c-a)>>31; | 2008 colsrc += stride*2; |
2046 colsrc[stride ] = (a|(d^f)) & (b|(d^e)) & (c|(e^f)); | |
2047 colsrc += stride*2; | |
2048 } | |
2049 src++; | |
2050 } | 2009 } |
2010 src++; | |
2011 } | |
2051 #endif //HAVE_MMX | 2012 #endif //HAVE_MMX |
2052 } | 2013 } |
2053 | 2014 |
2054 #ifdef HAVE_MMX | 2015 #ifdef HAVE_MMX |
2055 /** | 2016 /** |
2056 * transposes and shift the given 8x8 Block into dst1 and dst2 | 2017 * transposes and shift the given 8x8 Block into dst1 and dst2 |
2057 */ | 2018 */ |
2058 static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride) | 2019 static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride) |
2059 { | 2020 { |
2060 asm( | 2021 asm( |
2061 "lea (%0, %1), %%"REG_a" \n\t" | 2022 "lea (%0, %1), %%"REG_a" \n\t" |
2062 // 0 1 2 3 4 5 6 7 8 9 | 2023 // 0 1 2 3 4 5 6 7 8 9 |
2063 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 | 2024 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 |
2064 "movq (%0), %%mm0 \n\t" // 12345678 | 2025 "movq (%0), %%mm0 \n\t" // 12345678 |
2065 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh | 2026 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh |
2066 "movq %%mm0, %%mm2 \n\t" // 12345678 | 2027 "movq %%mm0, %%mm2 \n\t" // 12345678 |
2067 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d | 2028 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
2068 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h | 2029 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
2069 | 2030 |
2070 "movq (%%"REG_a", %1), %%mm1 \n\t" | 2031 "movq (%%"REG_a", %1), %%mm1 \n\t" |
2071 "movq (%%"REG_a", %1, 2), %%mm3 \n\t" | 2032 "movq (%%"REG_a", %1, 2), %%mm3 \n\t" |
2072 "movq %%mm1, %%mm4 \n\t" | 2033 "movq %%mm1, %%mm4 \n\t" |
2073 "punpcklbw %%mm3, %%mm1 \n\t" | 2034 "punpcklbw %%mm3, %%mm1 \n\t" |
2074 "punpckhbw %%mm3, %%mm4 \n\t" | 2035 "punpckhbw %%mm3, %%mm4 \n\t" |
2075 | 2036 |
2076 "movq %%mm0, %%mm3 \n\t" | 2037 "movq %%mm0, %%mm3 \n\t" |
2077 "punpcklwd %%mm1, %%mm0 \n\t" | 2038 "punpcklwd %%mm1, %%mm0 \n\t" |
2078 "punpckhwd %%mm1, %%mm3 \n\t" | 2039 "punpckhwd %%mm1, %%mm3 \n\t" |
2079 "movq %%mm2, %%mm1 \n\t" | 2040 "movq %%mm2, %%mm1 \n\t" |
2080 "punpcklwd %%mm4, %%mm2 \n\t" | 2041 "punpcklwd %%mm4, %%mm2 \n\t" |
2081 "punpckhwd %%mm4, %%mm1 \n\t" | 2042 "punpckhwd %%mm4, %%mm1 \n\t" |
2082 | 2043 |
2083 "movd %%mm0, 128(%2) \n\t" | 2044 "movd %%mm0, 128(%2) \n\t" |
2084 "psrlq $32, %%mm0 \n\t" | 2045 "psrlq $32, %%mm0 \n\t" |
2085 "movd %%mm0, 144(%2) \n\t" | 2046 "movd %%mm0, 144(%2) \n\t" |
2086 "movd %%mm3, 160(%2) \n\t" | 2047 "movd %%mm3, 160(%2) \n\t" |
2087 "psrlq $32, %%mm3 \n\t" | 2048 "psrlq $32, %%mm3 \n\t" |
2088 "movd %%mm3, 176(%2) \n\t" | 2049 "movd %%mm3, 176(%2) \n\t" |
2089 "movd %%mm3, 48(%3) \n\t" | 2050 "movd %%mm3, 48(%3) \n\t" |
2090 "movd %%mm2, 192(%2) \n\t" | 2051 "movd %%mm2, 192(%2) \n\t" |
2091 "movd %%mm2, 64(%3) \n\t" | 2052 "movd %%mm2, 64(%3) \n\t" |
2092 "psrlq $32, %%mm2 \n\t" | 2053 "psrlq $32, %%mm2 \n\t" |
2093 "movd %%mm2, 80(%3) \n\t" | 2054 "movd %%mm2, 80(%3) \n\t" |
2094 "movd %%mm1, 96(%3) \n\t" | 2055 "movd %%mm1, 96(%3) \n\t" |
2095 "psrlq $32, %%mm1 \n\t" | 2056 "psrlq $32, %%mm1 \n\t" |
2096 "movd %%mm1, 112(%3) \n\t" | 2057 "movd %%mm1, 112(%3) \n\t" |
2097 | 2058 |
2098 "lea (%%"REG_a", %1, 4), %%"REG_a" \n\t" | 2059 "lea (%%"REG_a", %1, 4), %%"REG_a" \n\t" |
2099 | 2060 |
2100 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678 | 2061 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678 |
2101 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh | 2062 "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh |
2102 "movq %%mm0, %%mm2 \n\t" // 12345678 | 2063 "movq %%mm0, %%mm2 \n\t" // 12345678 |
2103 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d | 2064 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
2104 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h | 2065 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
2105 | 2066 |
2106 "movq (%%"REG_a", %1), %%mm1 \n\t" | 2067 "movq (%%"REG_a", %1), %%mm1 \n\t" |
2107 "movq (%%"REG_a", %1, 2), %%mm3 \n\t" | 2068 "movq (%%"REG_a", %1, 2), %%mm3 \n\t" |
2108 "movq %%mm1, %%mm4 \n\t" | 2069 "movq %%mm1, %%mm4 \n\t" |
2109 "punpcklbw %%mm3, %%mm1 \n\t" | 2070 "punpcklbw %%mm3, %%mm1 \n\t" |
2110 "punpckhbw %%mm3, %%mm4 \n\t" | 2071 "punpckhbw %%mm3, %%mm4 \n\t" |
2111 | 2072 |
2112 "movq %%mm0, %%mm3 \n\t" | 2073 "movq %%mm0, %%mm3 \n\t" |
2113 "punpcklwd %%mm1, %%mm0 \n\t" | 2074 "punpcklwd %%mm1, %%mm0 \n\t" |
2114 "punpckhwd %%mm1, %%mm3 \n\t" | 2075 "punpckhwd %%mm1, %%mm3 \n\t" |
2115 "movq %%mm2, %%mm1 \n\t" | 2076 "movq %%mm2, %%mm1 \n\t" |
2116 "punpcklwd %%mm4, %%mm2 \n\t" | 2077 "punpcklwd %%mm4, %%mm2 \n\t" |
2117 "punpckhwd %%mm4, %%mm1 \n\t" | 2078 "punpckhwd %%mm4, %%mm1 \n\t" |
2118 | 2079 |
2119 "movd %%mm0, 132(%2) \n\t" | 2080 "movd %%mm0, 132(%2) \n\t" |
2120 "psrlq $32, %%mm0 \n\t" | 2081 "psrlq $32, %%mm0 \n\t" |
2121 "movd %%mm0, 148(%2) \n\t" | 2082 "movd %%mm0, 148(%2) \n\t" |
2122 "movd %%mm3, 164(%2) \n\t" | 2083 "movd %%mm3, 164(%2) \n\t" |
2123 "psrlq $32, %%mm3 \n\t" | 2084 "psrlq $32, %%mm3 \n\t" |
2124 "movd %%mm3, 180(%2) \n\t" | 2085 "movd %%mm3, 180(%2) \n\t" |
2125 "movd %%mm3, 52(%3) \n\t" | 2086 "movd %%mm3, 52(%3) \n\t" |
2126 "movd %%mm2, 196(%2) \n\t" | 2087 "movd %%mm2, 196(%2) \n\t" |
2127 "movd %%mm2, 68(%3) \n\t" | 2088 "movd %%mm2, 68(%3) \n\t" |
2128 "psrlq $32, %%mm2 \n\t" | 2089 "psrlq $32, %%mm2 \n\t" |
2129 "movd %%mm2, 84(%3) \n\t" | 2090 "movd %%mm2, 84(%3) \n\t" |
2130 "movd %%mm1, 100(%3) \n\t" | 2091 "movd %%mm1, 100(%3) \n\t" |
2131 "psrlq $32, %%mm1 \n\t" | 2092 "psrlq $32, %%mm1 \n\t" |
2132 "movd %%mm1, 116(%3) \n\t" | 2093 "movd %%mm1, 116(%3) \n\t" |
2133 | 2094 |
2134 | 2095 |
2135 :: "r" (src), "r" ((long)srcStride), "r" (dst1), "r" (dst2) | 2096 :: "r" (src), "r" ((long)srcStride), "r" (dst1), "r" (dst2) |
2136 : "%"REG_a | 2097 : "%"REG_a |
2137 ); | 2098 ); |
2138 } | 2099 } |
2139 | 2100 |
2140 /** | 2101 /** |
2141 * transposes the given 8x8 block | 2102 * transposes the given 8x8 block |
2142 */ | 2103 */ |
2143 static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src) | 2104 static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src) |
2144 { | 2105 { |
2145 asm( | 2106 asm( |
2146 "lea (%0, %1), %%"REG_a" \n\t" | 2107 "lea (%0, %1), %%"REG_a" \n\t" |
2147 "lea (%%"REG_a",%1,4), %%"REG_d" \n\t" | 2108 "lea (%%"REG_a",%1,4), %%"REG_d" \n\t" |
2148 // 0 1 2 3 4 5 6 7 8 9 | 2109 // 0 1 2 3 4 5 6 7 8 9 |
2149 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 | 2110 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 |
2150 "movq (%2), %%mm0 \n\t" // 12345678 | 2111 "movq (%2), %%mm0 \n\t" // 12345678 |
2151 "movq 16(%2), %%mm1 \n\t" // abcdefgh | 2112 "movq 16(%2), %%mm1 \n\t" // abcdefgh |
2152 "movq %%mm0, %%mm2 \n\t" // 12345678 | 2113 "movq %%mm0, %%mm2 \n\t" // 12345678 |
2153 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d | 2114 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
2154 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h | 2115 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
2155 | 2116 |
2156 "movq 32(%2), %%mm1 \n\t" | 2117 "movq 32(%2), %%mm1 \n\t" |
2157 "movq 48(%2), %%mm3 \n\t" | 2118 "movq 48(%2), %%mm3 \n\t" |
2158 "movq %%mm1, %%mm4 \n\t" | 2119 "movq %%mm1, %%mm4 \n\t" |
2159 "punpcklbw %%mm3, %%mm1 \n\t" | 2120 "punpcklbw %%mm3, %%mm1 \n\t" |
2160 "punpckhbw %%mm3, %%mm4 \n\t" | 2121 "punpckhbw %%mm3, %%mm4 \n\t" |
2161 | 2122 |
2162 "movq %%mm0, %%mm3 \n\t" | 2123 "movq %%mm0, %%mm3 \n\t" |
2163 "punpcklwd %%mm1, %%mm0 \n\t" | 2124 "punpcklwd %%mm1, %%mm0 \n\t" |
2164 "punpckhwd %%mm1, %%mm3 \n\t" | 2125 "punpckhwd %%mm1, %%mm3 \n\t" |
2165 "movq %%mm2, %%mm1 \n\t" | 2126 "movq %%mm2, %%mm1 \n\t" |
2166 "punpcklwd %%mm4, %%mm2 \n\t" | 2127 "punpcklwd %%mm4, %%mm2 \n\t" |
2167 "punpckhwd %%mm4, %%mm1 \n\t" | 2128 "punpckhwd %%mm4, %%mm1 \n\t" |
2168 | 2129 |
2169 "movd %%mm0, (%0) \n\t" | 2130 "movd %%mm0, (%0) \n\t" |
2170 "psrlq $32, %%mm0 \n\t" | 2131 "psrlq $32, %%mm0 \n\t" |
2171 "movd %%mm0, (%%"REG_a") \n\t" | 2132 "movd %%mm0, (%%"REG_a") \n\t" |
2172 "movd %%mm3, (%%"REG_a", %1) \n\t" | 2133 "movd %%mm3, (%%"REG_a", %1) \n\t" |
2173 "psrlq $32, %%mm3 \n\t" | 2134 "psrlq $32, %%mm3 \n\t" |
2174 "movd %%mm3, (%%"REG_a", %1, 2) \n\t" | 2135 "movd %%mm3, (%%"REG_a", %1, 2) \n\t" |
2175 "movd %%mm2, (%0, %1, 4) \n\t" | 2136 "movd %%mm2, (%0, %1, 4) \n\t" |
2176 "psrlq $32, %%mm2 \n\t" | 2137 "psrlq $32, %%mm2 \n\t" |
2177 "movd %%mm2, (%%"REG_d") \n\t" | 2138 "movd %%mm2, (%%"REG_d") \n\t" |
2178 "movd %%mm1, (%%"REG_d", %1) \n\t" | 2139 "movd %%mm1, (%%"REG_d", %1) \n\t" |
2179 "psrlq $32, %%mm1 \n\t" | 2140 "psrlq $32, %%mm1 \n\t" |
2180 "movd %%mm1, (%%"REG_d", %1, 2) \n\t" | 2141 "movd %%mm1, (%%"REG_d", %1, 2) \n\t" |
2181 | 2142 |
2182 | 2143 |
2183 "movq 64(%2), %%mm0 \n\t" // 12345678 | 2144 "movq 64(%2), %%mm0 \n\t" // 12345678 |
2184 "movq 80(%2), %%mm1 \n\t" // abcdefgh | 2145 "movq 80(%2), %%mm1 \n\t" // abcdefgh |
2185 "movq %%mm0, %%mm2 \n\t" // 12345678 | 2146 "movq %%mm0, %%mm2 \n\t" // 12345678 |
2186 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d | 2147 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
2187 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h | 2148 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
2188 | 2149 |
2189 "movq 96(%2), %%mm1 \n\t" | 2150 "movq 96(%2), %%mm1 \n\t" |
2190 "movq 112(%2), %%mm3 \n\t" | 2151 "movq 112(%2), %%mm3 \n\t" |
2191 "movq %%mm1, %%mm4 \n\t" | 2152 "movq %%mm1, %%mm4 \n\t" |
2192 "punpcklbw %%mm3, %%mm1 \n\t" | 2153 "punpcklbw %%mm3, %%mm1 \n\t" |
2193 "punpckhbw %%mm3, %%mm4 \n\t" | 2154 "punpckhbw %%mm3, %%mm4 \n\t" |
2194 | 2155 |
2195 "movq %%mm0, %%mm3 \n\t" | 2156 "movq %%mm0, %%mm3 \n\t" |
2196 "punpcklwd %%mm1, %%mm0 \n\t" | 2157 "punpcklwd %%mm1, %%mm0 \n\t" |
2197 "punpckhwd %%mm1, %%mm3 \n\t" | 2158 "punpckhwd %%mm1, %%mm3 \n\t" |
2198 "movq %%mm2, %%mm1 \n\t" | 2159 "movq %%mm2, %%mm1 \n\t" |
2199 "punpcklwd %%mm4, %%mm2 \n\t" | 2160 "punpcklwd %%mm4, %%mm2 \n\t" |
2200 "punpckhwd %%mm4, %%mm1 \n\t" | 2161 "punpckhwd %%mm4, %%mm1 \n\t" |
2201 | 2162 |
2202 "movd %%mm0, 4(%0) \n\t" | 2163 "movd %%mm0, 4(%0) \n\t" |
2203 "psrlq $32, %%mm0 \n\t" | 2164 "psrlq $32, %%mm0 \n\t" |
2204 "movd %%mm0, 4(%%"REG_a") \n\t" | 2165 "movd %%mm0, 4(%%"REG_a") \n\t" |
2205 "movd %%mm3, 4(%%"REG_a", %1) \n\t" | 2166 "movd %%mm3, 4(%%"REG_a", %1) \n\t" |
2206 "psrlq $32, %%mm3 \n\t" | 2167 "psrlq $32, %%mm3 \n\t" |
2207 "movd %%mm3, 4(%%"REG_a", %1, 2) \n\t" | 2168 "movd %%mm3, 4(%%"REG_a", %1, 2) \n\t" |
2208 "movd %%mm2, 4(%0, %1, 4) \n\t" | 2169 "movd %%mm2, 4(%0, %1, 4) \n\t" |
2209 "psrlq $32, %%mm2 \n\t" | 2170 "psrlq $32, %%mm2 \n\t" |
2210 "movd %%mm2, 4(%%"REG_d") \n\t" | 2171 "movd %%mm2, 4(%%"REG_d") \n\t" |
2211 "movd %%mm1, 4(%%"REG_d", %1) \n\t" | 2172 "movd %%mm1, 4(%%"REG_d", %1) \n\t" |
2212 "psrlq $32, %%mm1 \n\t" | 2173 "psrlq $32, %%mm1 \n\t" |
2213 "movd %%mm1, 4(%%"REG_d", %1, 2) \n\t" | 2174 "movd %%mm1, 4(%%"REG_d", %1, 2) \n\t" |
2214 | 2175 |
2215 :: "r" (dst), "r" ((long)dstStride), "r" (src) | 2176 :: "r" (dst), "r" ((long)dstStride), "r" (src) |
2216 : "%"REG_a, "%"REG_d | 2177 : "%"REG_a, "%"REG_d |
2217 ); | 2178 ); |
2218 } | 2179 } |
2219 #endif //HAVE_MMX | 2180 #endif //HAVE_MMX |
2220 //static long test=0; | 2181 //static long test=0; |
2221 | 2182 |
2222 #ifndef HAVE_ALTIVEC | 2183 #ifndef HAVE_ALTIVEC |
2223 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride, | 2184 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride, |
2224 uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise) | 2185 uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise) |
2225 { | 2186 { |
2226 // to save a register (FIXME do this outside of the loops) | 2187 // to save a register (FIXME do this outside of the loops) |
2227 tempBluredPast[127]= maxNoise[0]; | 2188 tempBluredPast[127]= maxNoise[0]; |
2228 tempBluredPast[128]= maxNoise[1]; | 2189 tempBluredPast[128]= maxNoise[1]; |
2229 tempBluredPast[129]= maxNoise[2]; | 2190 tempBluredPast[129]= maxNoise[2]; |
2230 | 2191 |
2231 #define FAST_L2_DIFF | 2192 #define FAST_L2_DIFF |
2232 //#define L1_DIFF //u should change the thresholds too if u try that one | 2193 //#define L1_DIFF //u should change the thresholds too if u try that one |
2233 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 2194 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2234 asm volatile( | 2195 asm volatile( |
2235 "lea (%2, %2, 2), %%"REG_a" \n\t" // 3*stride | 2196 "lea (%2, %2, 2), %%"REG_a" \n\t" // 3*stride |
2236 "lea (%2, %2, 4), %%"REG_d" \n\t" // 5*stride | 2197 "lea (%2, %2, 4), %%"REG_d" \n\t" // 5*stride |
2237 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride | 2198 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride |
2238 // 0 1 2 3 4 5 6 7 8 9 | 2199 // 0 1 2 3 4 5 6 7 8 9 |
2239 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2 | 2200 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2 |
2240 //FIXME reorder? | 2201 //FIXME reorder? |
2241 #ifdef L1_DIFF //needs mmx2 | 2202 #ifdef L1_DIFF //needs mmx2 |
2242 "movq (%0), %%mm0 \n\t" // L0 | 2203 "movq (%0), %%mm0 \n\t" // L0 |
2243 "psadbw (%1), %%mm0 \n\t" // |L0-R0| | 2204 "psadbw (%1), %%mm0 \n\t" // |L0-R0| |
2244 "movq (%0, %2), %%mm1 \n\t" // L1 | 2205 "movq (%0, %2), %%mm1 \n\t" // L1 |
2245 "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1| | 2206 "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1| |
2246 "movq (%0, %2, 2), %%mm2 \n\t" // L2 | 2207 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
2247 "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2| | 2208 "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2| |
2248 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 | 2209 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 |
2249 "psadbw (%1, %%"REG_a"), %%mm3 \n\t" // |L3-R3| | 2210 "psadbw (%1, %%"REG_a"), %%mm3 \n\t" // |L3-R3| |
2250 | 2211 |
2251 "movq (%0, %2, 4), %%mm4 \n\t" // L4 | 2212 "movq (%0, %2, 4), %%mm4 \n\t" // L4 |
2252 "paddw %%mm1, %%mm0 \n\t" | 2213 "paddw %%mm1, %%mm0 \n\t" |
2253 "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4| | 2214 "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4| |
2254 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5 | 2215 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5 |
2255 "paddw %%mm2, %%mm0 \n\t" | 2216 "paddw %%mm2, %%mm0 \n\t" |
2256 "psadbw (%1, %%"REG_d"), %%mm5 \n\t" // |L5-R5| | 2217 "psadbw (%1, %%"REG_d"), %%mm5 \n\t" // |L5-R5| |
2257 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6 | 2218 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6 |
2258 "paddw %%mm3, %%mm0 \n\t" | 2219 "paddw %%mm3, %%mm0 \n\t" |
2259 "psadbw (%1, %%"REG_a", 2), %%mm6 \n\t" // |L6-R6| | 2220 "psadbw (%1, %%"REG_a", 2), %%mm6 \n\t" // |L6-R6| |
2260 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7 | 2221 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7 |
2261 "paddw %%mm4, %%mm0 \n\t" | 2222 "paddw %%mm4, %%mm0 \n\t" |
2262 "psadbw (%1, %%"REG_c"), %%mm7 \n\t" // |L7-R7| | 2223 "psadbw (%1, %%"REG_c"), %%mm7 \n\t" // |L7-R7| |
2263 "paddw %%mm5, %%mm6 \n\t" | 2224 "paddw %%mm5, %%mm6 \n\t" |
2264 "paddw %%mm7, %%mm6 \n\t" | 2225 "paddw %%mm7, %%mm6 \n\t" |
2265 "paddw %%mm6, %%mm0 \n\t" | 2226 "paddw %%mm6, %%mm0 \n\t" |
2266 #else //L1_DIFF | 2227 #else //L1_DIFF |
2267 #if defined (FAST_L2_DIFF) | 2228 #if defined (FAST_L2_DIFF) |
2268 "pcmpeqb %%mm7, %%mm7 \n\t" | 2229 "pcmpeqb %%mm7, %%mm7 \n\t" |
2269 "movq "MANGLE(b80)", %%mm6 \n\t" | 2230 "movq "MANGLE(b80)", %%mm6 \n\t" |
2270 "pxor %%mm0, %%mm0 \n\t" | 2231 "pxor %%mm0, %%mm0 \n\t" |
2271 #define REAL_L2_DIFF_CORE(a, b)\ | 2232 #define REAL_L2_DIFF_CORE(a, b)\ |
2272 "movq " #a ", %%mm5 \n\t"\ | 2233 "movq " #a ", %%mm5 \n\t"\ |
2273 "movq " #b ", %%mm2 \n\t"\ | 2234 "movq " #b ", %%mm2 \n\t"\ |
2274 "pxor %%mm7, %%mm2 \n\t"\ | 2235 "pxor %%mm7, %%mm2 \n\t"\ |
2275 PAVGB(%%mm2, %%mm5)\ | 2236 PAVGB(%%mm2, %%mm5)\ |
2276 "paddb %%mm6, %%mm5 \n\t"\ | 2237 "paddb %%mm6, %%mm5 \n\t"\ |
2277 "movq %%mm5, %%mm2 \n\t"\ | 2238 "movq %%mm5, %%mm2 \n\t"\ |
2278 "psllw $8, %%mm5 \n\t"\ | 2239 "psllw $8, %%mm5 \n\t"\ |
2279 "pmaddwd %%mm5, %%mm5 \n\t"\ | 2240 "pmaddwd %%mm5, %%mm5 \n\t"\ |
2280 "pmaddwd %%mm2, %%mm2 \n\t"\ | 2241 "pmaddwd %%mm2, %%mm2 \n\t"\ |
2281 "paddd %%mm2, %%mm5 \n\t"\ | 2242 "paddd %%mm2, %%mm5 \n\t"\ |
2282 "psrld $14, %%mm5 \n\t"\ | 2243 "psrld $14, %%mm5 \n\t"\ |
2283 "paddd %%mm5, %%mm0 \n\t" | 2244 "paddd %%mm5, %%mm0 \n\t" |
2284 | 2245 |
2285 #else //defined (FAST_L2_DIFF) | 2246 #else //defined (FAST_L2_DIFF) |
2286 "pxor %%mm7, %%mm7 \n\t" | 2247 "pxor %%mm7, %%mm7 \n\t" |
2287 "pxor %%mm0, %%mm0 \n\t" | 2248 "pxor %%mm0, %%mm0 \n\t" |
2288 #define REAL_L2_DIFF_CORE(a, b)\ | 2249 #define REAL_L2_DIFF_CORE(a, b)\ |
2289 "movq " #a ", %%mm5 \n\t"\ | 2250 "movq " #a ", %%mm5 \n\t"\ |
2290 "movq " #b ", %%mm2 \n\t"\ | 2251 "movq " #b ", %%mm2 \n\t"\ |
2291 "movq %%mm5, %%mm1 \n\t"\ | 2252 "movq %%mm5, %%mm1 \n\t"\ |
2292 "movq %%mm2, %%mm3 \n\t"\ | 2253 "movq %%mm2, %%mm3 \n\t"\ |
2293 "punpcklbw %%mm7, %%mm5 \n\t"\ | 2254 "punpcklbw %%mm7, %%mm5 \n\t"\ |
2294 "punpckhbw %%mm7, %%mm1 \n\t"\ | 2255 "punpckhbw %%mm7, %%mm1 \n\t"\ |
2295 "punpcklbw %%mm7, %%mm2 \n\t"\ | 2256 "punpcklbw %%mm7, %%mm2 \n\t"\ |
2296 "punpckhbw %%mm7, %%mm3 \n\t"\ | 2257 "punpckhbw %%mm7, %%mm3 \n\t"\ |
2297 "psubw %%mm2, %%mm5 \n\t"\ | 2258 "psubw %%mm2, %%mm5 \n\t"\ |
2298 "psubw %%mm3, %%mm1 \n\t"\ | 2259 "psubw %%mm3, %%mm1 \n\t"\ |
2299 "pmaddwd %%mm5, %%mm5 \n\t"\ | 2260 "pmaddwd %%mm5, %%mm5 \n\t"\ |
2300 "pmaddwd %%mm1, %%mm1 \n\t"\ | 2261 "pmaddwd %%mm1, %%mm1 \n\t"\ |
2301 "paddd %%mm1, %%mm5 \n\t"\ | 2262 "paddd %%mm1, %%mm5 \n\t"\ |
2302 "paddd %%mm5, %%mm0 \n\t" | 2263 "paddd %%mm5, %%mm0 \n\t" |
2303 | 2264 |
2304 #endif //defined (FAST_L2_DIFF) | 2265 #endif //defined (FAST_L2_DIFF) |
2305 | 2266 |
2306 #define L2_DIFF_CORE(a, b) REAL_L2_DIFF_CORE(a, b) | 2267 #define L2_DIFF_CORE(a, b) REAL_L2_DIFF_CORE(a, b) |
2307 | 2268 |
2314 L2_DIFF_CORE((%0, %%REGa,2), (%1, %%REGa,2)) | 2275 L2_DIFF_CORE((%0, %%REGa,2), (%1, %%REGa,2)) |
2315 L2_DIFF_CORE((%0, %%REGc) , (%1, %%REGc)) | 2276 L2_DIFF_CORE((%0, %%REGc) , (%1, %%REGc)) |
2316 | 2277 |
2317 #endif //L1_DIFF | 2278 #endif //L1_DIFF |
2318 | 2279 |
2319 "movq %%mm0, %%mm4 \n\t" | 2280 "movq %%mm0, %%mm4 \n\t" |
2320 "psrlq $32, %%mm0 \n\t" | 2281 "psrlq $32, %%mm0 \n\t" |
2321 "paddd %%mm0, %%mm4 \n\t" | 2282 "paddd %%mm0, %%mm4 \n\t" |
2322 "movd %%mm4, %%ecx \n\t" | 2283 "movd %%mm4, %%ecx \n\t" |
2323 "shll $2, %%ecx \n\t" | 2284 "shll $2, %%ecx \n\t" |
2324 "mov %3, %%"REG_d" \n\t" | 2285 "mov %3, %%"REG_d" \n\t" |
2325 "addl -4(%%"REG_d"), %%ecx \n\t" | 2286 "addl -4(%%"REG_d"), %%ecx \n\t" |
2326 "addl 4(%%"REG_d"), %%ecx \n\t" | 2287 "addl 4(%%"REG_d"), %%ecx \n\t" |
2327 "addl -1024(%%"REG_d"), %%ecx \n\t" | 2288 "addl -1024(%%"REG_d"), %%ecx \n\t" |
2328 "addl $4, %%ecx \n\t" | 2289 "addl $4, %%ecx \n\t" |
2329 "addl 1024(%%"REG_d"), %%ecx \n\t" | 2290 "addl 1024(%%"REG_d"), %%ecx \n\t" |
2330 "shrl $3, %%ecx \n\t" | 2291 "shrl $3, %%ecx \n\t" |
2331 "movl %%ecx, (%%"REG_d") \n\t" | 2292 "movl %%ecx, (%%"REG_d") \n\t" |
2332 | 2293 |
2333 // "mov %3, %%"REG_c" \n\t" | 2294 // "mov %3, %%"REG_c" \n\t" |
2334 // "mov %%"REG_c", test \n\t" | 2295 // "mov %%"REG_c", test \n\t" |
2335 // "jmp 4f \n\t" | 2296 // "jmp 4f \n\t" |
2336 "cmpl 512(%%"REG_d"), %%ecx \n\t" | 2297 "cmpl 512(%%"REG_d"), %%ecx \n\t" |
2337 " jb 2f \n\t" | 2298 " jb 2f \n\t" |
2338 "cmpl 516(%%"REG_d"), %%ecx \n\t" | 2299 "cmpl 516(%%"REG_d"), %%ecx \n\t" |
2339 " jb 1f \n\t" | 2300 " jb 1f \n\t" |
2340 | 2301 |
2341 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride | 2302 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride |
2342 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride | 2303 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride |
2343 "movq (%0), %%mm0 \n\t" // L0 | 2304 "movq (%0), %%mm0 \n\t" // L0 |
2344 "movq (%0, %2), %%mm1 \n\t" // L1 | 2305 "movq (%0, %2), %%mm1 \n\t" // L1 |
2345 "movq (%0, %2, 2), %%mm2 \n\t" // L2 | 2306 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
2346 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 | 2307 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 |
2347 "movq (%0, %2, 4), %%mm4 \n\t" // L4 | 2308 "movq (%0, %2, 4), %%mm4 \n\t" // L4 |
2348 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5 | 2309 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5 |
2349 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6 | 2310 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6 |
2350 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7 | 2311 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7 |
2351 "movq %%mm0, (%1) \n\t" // L0 | 2312 "movq %%mm0, (%1) \n\t" // L0 |
2352 "movq %%mm1, (%1, %2) \n\t" // L1 | 2313 "movq %%mm1, (%1, %2) \n\t" // L1 |
2353 "movq %%mm2, (%1, %2, 2) \n\t" // L2 | 2314 "movq %%mm2, (%1, %2, 2) \n\t" // L2 |
2354 "movq %%mm3, (%1, %%"REG_a") \n\t" // L3 | 2315 "movq %%mm3, (%1, %%"REG_a") \n\t" // L3 |
2355 "movq %%mm4, (%1, %2, 4) \n\t" // L4 | 2316 "movq %%mm4, (%1, %2, 4) \n\t" // L4 |
2356 "movq %%mm5, (%1, %%"REG_d") \n\t" // L5 | 2317 "movq %%mm5, (%1, %%"REG_d") \n\t" // L5 |
2357 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // L6 | 2318 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // L6 |
2358 "movq %%mm7, (%1, %%"REG_c") \n\t" // L7 | 2319 "movq %%mm7, (%1, %%"REG_c") \n\t" // L7 |
2359 "jmp 4f \n\t" | 2320 "jmp 4f \n\t" |
2360 | 2321 |
2361 "1: \n\t" | 2322 "1: \n\t" |
2362 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride | 2323 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride |
2363 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride | 2324 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride |
2364 "movq (%0), %%mm0 \n\t" // L0 | 2325 "movq (%0), %%mm0 \n\t" // L0 |
2365 PAVGB((%1), %%mm0) // L0 | 2326 PAVGB((%1), %%mm0) // L0 |
2366 "movq (%0, %2), %%mm1 \n\t" // L1 | 2327 "movq (%0, %2), %%mm1 \n\t" // L1 |
2367 PAVGB((%1, %2), %%mm1) // L1 | 2328 PAVGB((%1, %2), %%mm1) // L1 |
2368 "movq (%0, %2, 2), %%mm2 \n\t" // L2 | 2329 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
2369 PAVGB((%1, %2, 2), %%mm2) // L2 | 2330 PAVGB((%1, %2, 2), %%mm2) // L2 |
2370 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 | 2331 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 |
2371 PAVGB((%1, %%REGa), %%mm3) // L3 | 2332 PAVGB((%1, %%REGa), %%mm3) // L3 |
2372 "movq (%0, %2, 4), %%mm4 \n\t" // L4 | 2333 "movq (%0, %2, 4), %%mm4 \n\t" // L4 |
2373 PAVGB((%1, %2, 4), %%mm4) // L4 | 2334 PAVGB((%1, %2, 4), %%mm4) // L4 |
2374 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5 | 2335 "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5 |
2375 PAVGB((%1, %%REGd), %%mm5) // L5 | 2336 PAVGB((%1, %%REGd), %%mm5) // L5 |
2376 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6 | 2337 "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6 |
2377 PAVGB((%1, %%REGa, 2), %%mm6) // L6 | 2338 PAVGB((%1, %%REGa, 2), %%mm6) // L6 |
2378 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7 | 2339 "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7 |
2379 PAVGB((%1, %%REGc), %%mm7) // L7 | 2340 PAVGB((%1, %%REGc), %%mm7) // L7 |
2380 "movq %%mm0, (%1) \n\t" // R0 | 2341 "movq %%mm0, (%1) \n\t" // R0 |
2381 "movq %%mm1, (%1, %2) \n\t" // R1 | 2342 "movq %%mm1, (%1, %2) \n\t" // R1 |
2382 "movq %%mm2, (%1, %2, 2) \n\t" // R2 | 2343 "movq %%mm2, (%1, %2, 2) \n\t" // R2 |
2383 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3 | 2344 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3 |
2384 "movq %%mm4, (%1, %2, 4) \n\t" // R4 | 2345 "movq %%mm4, (%1, %2, 4) \n\t" // R4 |
2385 "movq %%mm5, (%1, %%"REG_d") \n\t" // R5 | 2346 "movq %%mm5, (%1, %%"REG_d") \n\t" // R5 |
2386 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // R6 | 2347 "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // R6 |
2387 "movq %%mm7, (%1, %%"REG_c") \n\t" // R7 | 2348 "movq %%mm7, (%1, %%"REG_c") \n\t" // R7 |
2388 "movq %%mm0, (%0) \n\t" // L0 | 2349 "movq %%mm0, (%0) \n\t" // L0 |
2389 "movq %%mm1, (%0, %2) \n\t" // L1 | 2350 "movq %%mm1, (%0, %2) \n\t" // L1 |
2390 "movq %%mm2, (%0, %2, 2) \n\t" // L2 | 2351 "movq %%mm2, (%0, %2, 2) \n\t" // L2 |
2391 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3 | 2352 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3 |
2392 "movq %%mm4, (%0, %2, 4) \n\t" // L4 | 2353 "movq %%mm4, (%0, %2, 4) \n\t" // L4 |
2393 "movq %%mm5, (%0, %%"REG_d") \n\t" // L5 | 2354 "movq %%mm5, (%0, %%"REG_d") \n\t" // L5 |
2394 "movq %%mm6, (%0, %%"REG_a", 2) \n\t" // L6 | 2355 "movq %%mm6, (%0, %%"REG_a", 2) \n\t" // L6 |
2395 "movq %%mm7, (%0, %%"REG_c") \n\t" // L7 | 2356 "movq %%mm7, (%0, %%"REG_c") \n\t" // L7 |
2396 "jmp 4f \n\t" | 2357 "jmp 4f \n\t" |
2397 | 2358 |
2398 "2: \n\t" | 2359 "2: \n\t" |
2399 "cmpl 508(%%"REG_d"), %%ecx \n\t" | 2360 "cmpl 508(%%"REG_d"), %%ecx \n\t" |
2400 " jb 3f \n\t" | 2361 " jb 3f \n\t" |
2401 | 2362 |
2402 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride | 2363 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride |
2403 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride | 2364 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride |
2404 "movq (%0), %%mm0 \n\t" // L0 | 2365 "movq (%0), %%mm0 \n\t" // L0 |
2405 "movq (%0, %2), %%mm1 \n\t" // L1 | 2366 "movq (%0, %2), %%mm1 \n\t" // L1 |
2406 "movq (%0, %2, 2), %%mm2 \n\t" // L2 | 2367 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
2407 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 | 2368 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 |
2408 "movq (%1), %%mm4 \n\t" // R0 | 2369 "movq (%1), %%mm4 \n\t" // R0 |
2409 "movq (%1, %2), %%mm5 \n\t" // R1 | 2370 "movq (%1, %2), %%mm5 \n\t" // R1 |
2410 "movq (%1, %2, 2), %%mm6 \n\t" // R2 | 2371 "movq (%1, %2, 2), %%mm6 \n\t" // R2 |
2411 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3 | 2372 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3 |
2412 PAVGB(%%mm4, %%mm0) | 2373 PAVGB(%%mm4, %%mm0) |
2413 PAVGB(%%mm5, %%mm1) | 2374 PAVGB(%%mm5, %%mm1) |
2414 PAVGB(%%mm6, %%mm2) | 2375 PAVGB(%%mm6, %%mm2) |
2415 PAVGB(%%mm7, %%mm3) | 2376 PAVGB(%%mm7, %%mm3) |
2416 PAVGB(%%mm4, %%mm0) | 2377 PAVGB(%%mm4, %%mm0) |
2417 PAVGB(%%mm5, %%mm1) | 2378 PAVGB(%%mm5, %%mm1) |
2418 PAVGB(%%mm6, %%mm2) | 2379 PAVGB(%%mm6, %%mm2) |
2419 PAVGB(%%mm7, %%mm3) | 2380 PAVGB(%%mm7, %%mm3) |
2420 "movq %%mm0, (%1) \n\t" // R0 | 2381 "movq %%mm0, (%1) \n\t" // R0 |
2421 "movq %%mm1, (%1, %2) \n\t" // R1 | 2382 "movq %%mm1, (%1, %2) \n\t" // R1 |
2422 "movq %%mm2, (%1, %2, 2) \n\t" // R2 | 2383 "movq %%mm2, (%1, %2, 2) \n\t" // R2 |
2423 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3 | 2384 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3 |
2424 "movq %%mm0, (%0) \n\t" // L0 | 2385 "movq %%mm0, (%0) \n\t" // L0 |
2425 "movq %%mm1, (%0, %2) \n\t" // L1 | 2386 "movq %%mm1, (%0, %2) \n\t" // L1 |
2426 "movq %%mm2, (%0, %2, 2) \n\t" // L2 | 2387 "movq %%mm2, (%0, %2, 2) \n\t" // L2 |
2427 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3 | 2388 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3 |
2428 | 2389 |
2429 "movq (%0, %2, 4), %%mm0 \n\t" // L4 | 2390 "movq (%0, %2, 4), %%mm0 \n\t" // L4 |
2430 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5 | 2391 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5 |
2431 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6 | 2392 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6 |
2432 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7 | 2393 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7 |
2433 "movq (%1, %2, 4), %%mm4 \n\t" // R4 | 2394 "movq (%1, %2, 4), %%mm4 \n\t" // R4 |
2434 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5 | 2395 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5 |
2435 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6 | 2396 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6 |
2436 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7 | 2397 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7 |
2437 PAVGB(%%mm4, %%mm0) | 2398 PAVGB(%%mm4, %%mm0) |
2438 PAVGB(%%mm5, %%mm1) | 2399 PAVGB(%%mm5, %%mm1) |
2439 PAVGB(%%mm6, %%mm2) | 2400 PAVGB(%%mm6, %%mm2) |
2440 PAVGB(%%mm7, %%mm3) | 2401 PAVGB(%%mm7, %%mm3) |
2441 PAVGB(%%mm4, %%mm0) | 2402 PAVGB(%%mm4, %%mm0) |
2442 PAVGB(%%mm5, %%mm1) | 2403 PAVGB(%%mm5, %%mm1) |
2443 PAVGB(%%mm6, %%mm2) | 2404 PAVGB(%%mm6, %%mm2) |
2444 PAVGB(%%mm7, %%mm3) | 2405 PAVGB(%%mm7, %%mm3) |
2445 "movq %%mm0, (%1, %2, 4) \n\t" // R4 | 2406 "movq %%mm0, (%1, %2, 4) \n\t" // R4 |
2446 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5 | 2407 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5 |
2447 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6 | 2408 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6 |
2448 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7 | 2409 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7 |
2449 "movq %%mm0, (%0, %2, 4) \n\t" // L4 | 2410 "movq %%mm0, (%0, %2, 4) \n\t" // L4 |
2450 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5 | 2411 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5 |
2451 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6 | 2412 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6 |
2452 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7 | 2413 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7 |
2453 "jmp 4f \n\t" | 2414 "jmp 4f \n\t" |
2454 | 2415 |
2455 "3: \n\t" | 2416 "3: \n\t" |
2456 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride | 2417 "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride |
2457 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride | 2418 "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride |
2458 "movq (%0), %%mm0 \n\t" // L0 | 2419 "movq (%0), %%mm0 \n\t" // L0 |
2459 "movq (%0, %2), %%mm1 \n\t" // L1 | 2420 "movq (%0, %2), %%mm1 \n\t" // L1 |
2460 "movq (%0, %2, 2), %%mm2 \n\t" // L2 | 2421 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
2461 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 | 2422 "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 |
2462 "movq (%1), %%mm4 \n\t" // R0 | 2423 "movq (%1), %%mm4 \n\t" // R0 |
2463 "movq (%1, %2), %%mm5 \n\t" // R1 | 2424 "movq (%1, %2), %%mm5 \n\t" // R1 |
2464 "movq (%1, %2, 2), %%mm6 \n\t" // R2 | 2425 "movq (%1, %2, 2), %%mm6 \n\t" // R2 |
2465 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3 | 2426 "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3 |
2466 PAVGB(%%mm4, %%mm0) | 2427 PAVGB(%%mm4, %%mm0) |
2467 PAVGB(%%mm5, %%mm1) | 2428 PAVGB(%%mm5, %%mm1) |
2468 PAVGB(%%mm6, %%mm2) | 2429 PAVGB(%%mm6, %%mm2) |
2469 PAVGB(%%mm7, %%mm3) | 2430 PAVGB(%%mm7, %%mm3) |
2470 PAVGB(%%mm4, %%mm0) | 2431 PAVGB(%%mm4, %%mm0) |
2471 PAVGB(%%mm5, %%mm1) | 2432 PAVGB(%%mm5, %%mm1) |
2472 PAVGB(%%mm6, %%mm2) | 2433 PAVGB(%%mm6, %%mm2) |
2473 PAVGB(%%mm7, %%mm3) | 2434 PAVGB(%%mm7, %%mm3) |
2474 PAVGB(%%mm4, %%mm0) | 2435 PAVGB(%%mm4, %%mm0) |
2475 PAVGB(%%mm5, %%mm1) | 2436 PAVGB(%%mm5, %%mm1) |
2476 PAVGB(%%mm6, %%mm2) | 2437 PAVGB(%%mm6, %%mm2) |
2477 PAVGB(%%mm7, %%mm3) | 2438 PAVGB(%%mm7, %%mm3) |
2478 "movq %%mm0, (%1) \n\t" // R0 | 2439 "movq %%mm0, (%1) \n\t" // R0 |
2479 "movq %%mm1, (%1, %2) \n\t" // R1 | 2440 "movq %%mm1, (%1, %2) \n\t" // R1 |
2480 "movq %%mm2, (%1, %2, 2) \n\t" // R2 | 2441 "movq %%mm2, (%1, %2, 2) \n\t" // R2 |
2481 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3 | 2442 "movq %%mm3, (%1, %%"REG_a") \n\t" // R3 |
2482 "movq %%mm0, (%0) \n\t" // L0 | 2443 "movq %%mm0, (%0) \n\t" // L0 |
2483 "movq %%mm1, (%0, %2) \n\t" // L1 | 2444 "movq %%mm1, (%0, %2) \n\t" // L1 |
2484 "movq %%mm2, (%0, %2, 2) \n\t" // L2 | 2445 "movq %%mm2, (%0, %2, 2) \n\t" // L2 |
2485 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3 | 2446 "movq %%mm3, (%0, %%"REG_a") \n\t" // L3 |
2486 | 2447 |
2487 "movq (%0, %2, 4), %%mm0 \n\t" // L4 | 2448 "movq (%0, %2, 4), %%mm0 \n\t" // L4 |
2488 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5 | 2449 "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5 |
2489 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6 | 2450 "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6 |
2490 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7 | 2451 "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7 |
2491 "movq (%1, %2, 4), %%mm4 \n\t" // R4 | 2452 "movq (%1, %2, 4), %%mm4 \n\t" // R4 |
2492 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5 | 2453 "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5 |
2493 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6 | 2454 "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6 |
2494 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7 | 2455 "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7 |
2495 PAVGB(%%mm4, %%mm0) | 2456 PAVGB(%%mm4, %%mm0) |
2496 PAVGB(%%mm5, %%mm1) | 2457 PAVGB(%%mm5, %%mm1) |
2497 PAVGB(%%mm6, %%mm2) | 2458 PAVGB(%%mm6, %%mm2) |
2498 PAVGB(%%mm7, %%mm3) | 2459 PAVGB(%%mm7, %%mm3) |
2499 PAVGB(%%mm4, %%mm0) | 2460 PAVGB(%%mm4, %%mm0) |
2500 PAVGB(%%mm5, %%mm1) | 2461 PAVGB(%%mm5, %%mm1) |
2501 PAVGB(%%mm6, %%mm2) | 2462 PAVGB(%%mm6, %%mm2) |
2502 PAVGB(%%mm7, %%mm3) | 2463 PAVGB(%%mm7, %%mm3) |
2503 PAVGB(%%mm4, %%mm0) | 2464 PAVGB(%%mm4, %%mm0) |
2504 PAVGB(%%mm5, %%mm1) | 2465 PAVGB(%%mm5, %%mm1) |
2505 PAVGB(%%mm6, %%mm2) | 2466 PAVGB(%%mm6, %%mm2) |
2506 PAVGB(%%mm7, %%mm3) | 2467 PAVGB(%%mm7, %%mm3) |
2507 "movq %%mm0, (%1, %2, 4) \n\t" // R4 | 2468 "movq %%mm0, (%1, %2, 4) \n\t" // R4 |
2508 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5 | 2469 "movq %%mm1, (%1, %%"REG_d") \n\t" // R5 |
2509 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6 | 2470 "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6 |
2510 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7 | 2471 "movq %%mm3, (%1, %%"REG_c") \n\t" // R7 |
2511 "movq %%mm0, (%0, %2, 4) \n\t" // L4 | 2472 "movq %%mm0, (%0, %2, 4) \n\t" // L4 |
2512 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5 | 2473 "movq %%mm1, (%0, %%"REG_d") \n\t" // L5 |
2513 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6 | 2474 "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6 |
2514 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7 | 2475 "movq %%mm3, (%0, %%"REG_c") \n\t" // L7 |
2515 | 2476 |
2516 "4: \n\t" | 2477 "4: \n\t" |
2517 | 2478 |
2518 :: "r" (src), "r" (tempBlured), "r"((long)stride), "m" (tempBluredPast) | 2479 :: "r" (src), "r" (tempBlured), "r"((long)stride), "m" (tempBluredPast) |
2519 : "%"REG_a, "%"REG_d, "%"REG_c, "memory" | 2480 : "%"REG_a, "%"REG_d, "%"REG_c, "memory" |
2520 ); | 2481 ); |
2521 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 2482 #else //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2522 { | 2483 { |
2523 int y; | 2484 int y; |
2524 int d=0; | 2485 int d=0; |
2525 // int sysd=0; | 2486 // int sysd=0; |
2526 int i; | 2487 int i; |
2527 | 2488 |
2528 for(y=0; y<8; y++) | 2489 for(y=0; y<8; y++){ |
2529 { | 2490 int x; |
2530 int x; | 2491 for(x=0; x<8; x++){ |
2531 for(x=0; x<8; x++) | 2492 int ref= tempBlured[ x + y*stride ]; |
2532 { | 2493 int cur= src[ x + y*stride ]; |
2533 int ref= tempBlured[ x + y*stride ]; | 2494 int d1=ref - cur; |
2534 int cur= src[ x + y*stride ]; | 2495 // if(x==0 || x==7) d1+= d1>>1; |
2535 int d1=ref - cur; | 2496 // if(y==0 || y==7) d1+= d1>>1; |
2536 // if(x==0 || x==7) d1+= d1>>1; | 2497 // d+= FFABS(d1); |
2537 // if(y==0 || y==7) d1+= d1>>1; | 2498 d+= d1*d1; |
2538 // d+= FFABS(d1); | 2499 // sysd+= d1; |
2539 d+= d1*d1; | |
2540 // sysd+= d1; | |
2541 } | |
2542 } | 2500 } |
2543 i=d; | 2501 } |
2544 d= ( | 2502 i=d; |
2545 4*d | 2503 d= ( |
2546 +(*(tempBluredPast-256)) | 2504 4*d |
2547 +(*(tempBluredPast-1))+ (*(tempBluredPast+1)) | 2505 +(*(tempBluredPast-256)) |
2548 +(*(tempBluredPast+256)) | 2506 +(*(tempBluredPast-1))+ (*(tempBluredPast+1)) |
2549 +4)>>3; | 2507 +(*(tempBluredPast+256)) |
2550 *tempBluredPast=i; | 2508 +4)>>3; |
2551 // ((*tempBluredPast)*3 + d + 2)>>2; | 2509 *tempBluredPast=i; |
2510 // ((*tempBluredPast)*3 + d + 2)>>2; | |
2552 | 2511 |
2553 /* | 2512 /* |
2554 Switch between | 2513 Switch between |
2555 1 0 0 0 0 0 0 (0) | 2514 1 0 0 0 0 0 0 (0) |
2556 64 32 16 8 4 2 1 (1) | 2515 64 32 16 8 4 2 1 (1) |
2557 64 48 36 27 20 15 11 (33) (approx) | 2516 64 48 36 27 20 15 11 (33) (approx) |
2558 64 56 49 43 37 33 29 (200) (approx) | 2517 64 56 49 43 37 33 29 (200) (approx) |
2559 */ | 2518 */ |
2560 if(d > maxNoise[1]) | 2519 if(d > maxNoise[1]){ |
2561 { | 2520 if(d < maxNoise[2]){ |
2562 if(d < maxNoise[2]) | 2521 for(y=0; y<8; y++){ |
2563 { | 2522 int x; |
2564 for(y=0; y<8; y++) | 2523 for(x=0; x<8; x++){ |
2565 { | 2524 int ref= tempBlured[ x + y*stride ]; |
2566 int x; | 2525 int cur= src[ x + y*stride ]; |
2567 for(x=0; x<8; x++) | 2526 tempBlured[ x + y*stride ]= |
2568 { | 2527 src[ x + y*stride ]= |
2569 int ref= tempBlured[ x + y*stride ]; | 2528 (ref + cur + 1)>>1; |
2570 int cur= src[ x + y*stride ]; | |
2571 tempBlured[ x + y*stride ]= | |
2572 src[ x + y*stride ]= | |
2573 (ref + cur + 1)>>1; | |
2574 } | |
2575 } | |
2576 } | 2529 } |
2577 else | 2530 } |
2578 { | 2531 }else{ |
2579 for(y=0; y<8; y++) | 2532 for(y=0; y<8; y++){ |
2580 { | 2533 int x; |
2581 int x; | 2534 for(x=0; x<8; x++){ |
2582 for(x=0; x<8; x++) | 2535 tempBlured[ x + y*stride ]= src[ x + y*stride ]; |
2583 { | |
2584 tempBlured[ x + y*stride ]= src[ x + y*stride ]; | |
2585 } | |
2586 } | |
2587 } | 2536 } |
2537 } | |
2588 } | 2538 } |
2589 else | 2539 }else{ |
2590 { | 2540 if(d < maxNoise[0]){ |
2591 if(d < maxNoise[0]) | 2541 for(y=0; y<8; y++){ |
2592 { | 2542 int x; |
2593 for(y=0; y<8; y++) | 2543 for(x=0; x<8; x++){ |
2594 { | 2544 int ref= tempBlured[ x + y*stride ]; |
2595 int x; | 2545 int cur= src[ x + y*stride ]; |
2596 for(x=0; x<8; x++) | 2546 tempBlured[ x + y*stride ]= |
2597 { | 2547 src[ x + y*stride ]= |
2598 int ref= tempBlured[ x + y*stride ]; | 2548 (ref*7 + cur + 4)>>3; |
2599 int cur= src[ x + y*stride ]; | |
2600 tempBlured[ x + y*stride ]= | |
2601 src[ x + y*stride ]= | |
2602 (ref*7 + cur + 4)>>3; | |
2603 } | |
2604 } | |
2605 } | 2549 } |
2606 else | 2550 } |
2607 { | 2551 }else{ |
2608 for(y=0; y<8; y++) | 2552 for(y=0; y<8; y++){ |
2609 { | 2553 int x; |
2610 int x; | 2554 for(x=0; x<8; x++){ |
2611 for(x=0; x<8; x++) | 2555 int ref= tempBlured[ x + y*stride ]; |
2612 { | 2556 int cur= src[ x + y*stride ]; |
2613 int ref= tempBlured[ x + y*stride ]; | 2557 tempBlured[ x + y*stride ]= |
2614 int cur= src[ x + y*stride ]; | 2558 src[ x + y*stride ]= |
2615 tempBlured[ x + y*stride ]= | 2559 (ref*3 + cur + 2)>>2; |
2616 src[ x + y*stride ]= | |
2617 (ref*3 + cur + 2)>>2; | |
2618 } | |
2619 } | |
2620 } | 2560 } |
2561 } | |
2621 } | 2562 } |
2563 } | |
2622 } | 2564 } |
2623 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) | 2565 #endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2624 } | 2566 } |
2625 #endif //HAVE_ALTIVEC | 2567 #endif //HAVE_ALTIVEC |
2626 | 2568 |
2627 #ifdef HAVE_MMX | 2569 #ifdef HAVE_MMX |
2628 /** | 2570 /** |
2629 * accurate deblock filter | 2571 * accurate deblock filter |
2630 */ | 2572 */ |
2631 static av_always_inline void RENAME(do_a_deblock)(uint8_t *src, int step, int stride, PPContext *c){ | 2573 static av_always_inline void RENAME(do_a_deblock)(uint8_t *src, int step, int stride, PPContext *c){ |
2632 int64_t dc_mask, eq_mask, both_masks; | 2574 int64_t dc_mask, eq_mask, both_masks; |
2633 int64_t sums[10*8*2]; | 2575 int64_t sums[10*8*2]; |
2634 src+= step*3; // src points to begin of the 8x8 Block | 2576 src+= step*3; // src points to begin of the 8x8 Block |
2635 //START_TIMER | 2577 //START_TIMER |
2636 asm volatile( | 2578 asm volatile( |
2637 "movq %0, %%mm7 \n\t" | 2579 "movq %0, %%mm7 \n\t" |
2638 "movq %1, %%mm6 \n\t" | 2580 "movq %1, %%mm6 \n\t" |
2639 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP]) | 2581 : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP]) |
2640 ); | 2582 ); |
2641 | 2583 |
2642 asm volatile( | 2584 asm volatile( |
2643 "lea (%2, %3), %%"REG_a" \n\t" | 2585 "lea (%2, %3), %%"REG_a" \n\t" |
2644 // 0 1 2 3 4 5 6 7 8 9 | 2586 // 0 1 2 3 4 5 6 7 8 9 |
2645 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2 | 2587 // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2 |
2646 | 2588 |
2647 "movq (%2), %%mm0 \n\t" | 2589 "movq (%2), %%mm0 \n\t" |
2648 "movq (%%"REG_a"), %%mm1 \n\t" | 2590 "movq (%%"REG_a"), %%mm1 \n\t" |
2649 "movq %%mm1, %%mm3 \n\t" | 2591 "movq %%mm1, %%mm3 \n\t" |
2650 "movq %%mm1, %%mm4 \n\t" | 2592 "movq %%mm1, %%mm4 \n\t" |
2651 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece | 2593 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
2652 "paddb %%mm7, %%mm0 \n\t" | 2594 "paddb %%mm7, %%mm0 \n\t" |
2653 "pcmpgtb %%mm6, %%mm0 \n\t" | 2595 "pcmpgtb %%mm6, %%mm0 \n\t" |
2654 | 2596 |
2655 "movq (%%"REG_a",%3), %%mm2 \n\t" | 2597 "movq (%%"REG_a",%3), %%mm2 \n\t" |
2656 PMAXUB(%%mm2, %%mm4) | 2598 PMAXUB(%%mm2, %%mm4) |
2657 PMINUB(%%mm2, %%mm3, %%mm5) | 2599 PMINUB(%%mm2, %%mm3, %%mm5) |
2658 "psubb %%mm2, %%mm1 \n\t" | 2600 "psubb %%mm2, %%mm1 \n\t" |
2659 "paddb %%mm7, %%mm1 \n\t" | 2601 "paddb %%mm7, %%mm1 \n\t" |
2660 "pcmpgtb %%mm6, %%mm1 \n\t" | 2602 "pcmpgtb %%mm6, %%mm1 \n\t" |
2661 "paddb %%mm1, %%mm0 \n\t" | 2603 "paddb %%mm1, %%mm0 \n\t" |
2662 | 2604 |
2663 "movq (%%"REG_a", %3, 2), %%mm1 \n\t" | 2605 "movq (%%"REG_a", %3, 2), %%mm1 \n\t" |
2664 PMAXUB(%%mm1, %%mm4) | 2606 PMAXUB(%%mm1, %%mm4) |
2665 PMINUB(%%mm1, %%mm3, %%mm5) | 2607 PMINUB(%%mm1, %%mm3, %%mm5) |
2666 "psubb %%mm1, %%mm2 \n\t" | 2608 "psubb %%mm1, %%mm2 \n\t" |
2667 "paddb %%mm7, %%mm2 \n\t" | 2609 "paddb %%mm7, %%mm2 \n\t" |
2668 "pcmpgtb %%mm6, %%mm2 \n\t" | 2610 "pcmpgtb %%mm6, %%mm2 \n\t" |
2669 "paddb %%mm2, %%mm0 \n\t" | 2611 "paddb %%mm2, %%mm0 \n\t" |
2670 | 2612 |
2671 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t" | 2613 "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t" |
2672 | 2614 |
2673 "movq (%2, %3, 4), %%mm2 \n\t" | 2615 "movq (%2, %3, 4), %%mm2 \n\t" |
2674 PMAXUB(%%mm2, %%mm4) | 2616 PMAXUB(%%mm2, %%mm4) |
2675 PMINUB(%%mm2, %%mm3, %%mm5) | 2617 PMINUB(%%mm2, %%mm3, %%mm5) |
2676 "psubb %%mm2, %%mm1 \n\t" | 2618 "psubb %%mm2, %%mm1 \n\t" |
2677 "paddb %%mm7, %%mm1 \n\t" | 2619 "paddb %%mm7, %%mm1 \n\t" |
2678 "pcmpgtb %%mm6, %%mm1 \n\t" | 2620 "pcmpgtb %%mm6, %%mm1 \n\t" |
2679 "paddb %%mm1, %%mm0 \n\t" | 2621 "paddb %%mm1, %%mm0 \n\t" |
2680 | 2622 |
2681 "movq (%%"REG_a"), %%mm1 \n\t" | 2623 "movq (%%"REG_a"), %%mm1 \n\t" |
2682 PMAXUB(%%mm1, %%mm4) | 2624 PMAXUB(%%mm1, %%mm4) |
2683 PMINUB(%%mm1, %%mm3, %%mm5) | 2625 PMINUB(%%mm1, %%mm3, %%mm5) |
2684 "psubb %%mm1, %%mm2 \n\t" | 2626 "psubb %%mm1, %%mm2 \n\t" |
2685 "paddb %%mm7, %%mm2 \n\t" | 2627 "paddb %%mm7, %%mm2 \n\t" |
2686 "pcmpgtb %%mm6, %%mm2 \n\t" | 2628 "pcmpgtb %%mm6, %%mm2 \n\t" |
2687 "paddb %%mm2, %%mm0 \n\t" | 2629 "paddb %%mm2, %%mm0 \n\t" |
2688 | 2630 |
2689 "movq (%%"REG_a", %3), %%mm2 \n\t" | 2631 "movq (%%"REG_a", %3), %%mm2 \n\t" |
2690 PMAXUB(%%mm2, %%mm4) | 2632 PMAXUB(%%mm2, %%mm4) |
2691 PMINUB(%%mm2, %%mm3, %%mm5) | 2633 PMINUB(%%mm2, %%mm3, %%mm5) |
2692 "psubb %%mm2, %%mm1 \n\t" | 2634 "psubb %%mm2, %%mm1 \n\t" |
2693 "paddb %%mm7, %%mm1 \n\t" | 2635 "paddb %%mm7, %%mm1 \n\t" |
2694 "pcmpgtb %%mm6, %%mm1 \n\t" | 2636 "pcmpgtb %%mm6, %%mm1 \n\t" |
2695 "paddb %%mm1, %%mm0 \n\t" | 2637 "paddb %%mm1, %%mm0 \n\t" |
2696 | 2638 |
2697 "movq (%%"REG_a", %3, 2), %%mm1 \n\t" | 2639 "movq (%%"REG_a", %3, 2), %%mm1 \n\t" |
2698 PMAXUB(%%mm1, %%mm4) | 2640 PMAXUB(%%mm1, %%mm4) |
2699 PMINUB(%%mm1, %%mm3, %%mm5) | 2641 PMINUB(%%mm1, %%mm3, %%mm5) |
2700 "psubb %%mm1, %%mm2 \n\t" | 2642 "psubb %%mm1, %%mm2 \n\t" |
2701 "paddb %%mm7, %%mm2 \n\t" | 2643 "paddb %%mm7, %%mm2 \n\t" |
2702 "pcmpgtb %%mm6, %%mm2 \n\t" | 2644 "pcmpgtb %%mm6, %%mm2 \n\t" |
2703 "paddb %%mm2, %%mm0 \n\t" | 2645 "paddb %%mm2, %%mm0 \n\t" |
2704 | 2646 |
2705 "movq (%2, %3, 8), %%mm2 \n\t" | 2647 "movq (%2, %3, 8), %%mm2 \n\t" |
2706 PMAXUB(%%mm2, %%mm4) | 2648 PMAXUB(%%mm2, %%mm4) |
2707 PMINUB(%%mm2, %%mm3, %%mm5) | 2649 PMINUB(%%mm2, %%mm3, %%mm5) |
2708 "psubb %%mm2, %%mm1 \n\t" | 2650 "psubb %%mm2, %%mm1 \n\t" |
2709 "paddb %%mm7, %%mm1 \n\t" | 2651 "paddb %%mm7, %%mm1 \n\t" |
2710 "pcmpgtb %%mm6, %%mm1 \n\t" | 2652 "pcmpgtb %%mm6, %%mm1 \n\t" |
2711 "paddb %%mm1, %%mm0 \n\t" | 2653 "paddb %%mm1, %%mm0 \n\t" |
2712 | 2654 |
2713 "movq (%%"REG_a", %3, 4), %%mm1 \n\t" | 2655 "movq (%%"REG_a", %3, 4), %%mm1 \n\t" |
2714 "psubb %%mm1, %%mm2 \n\t" | 2656 "psubb %%mm1, %%mm2 \n\t" |
2715 "paddb %%mm7, %%mm2 \n\t" | 2657 "paddb %%mm7, %%mm2 \n\t" |
2716 "pcmpgtb %%mm6, %%mm2 \n\t" | 2658 "pcmpgtb %%mm6, %%mm2 \n\t" |
2717 "paddb %%mm2, %%mm0 \n\t" | 2659 "paddb %%mm2, %%mm0 \n\t" |
2718 "psubusb %%mm3, %%mm4 \n\t" | 2660 "psubusb %%mm3, %%mm4 \n\t" |
2719 | 2661 |
2720 "pxor %%mm6, %%mm6 \n\t" | 2662 "pxor %%mm6, %%mm6 \n\t" |
2721 "movq %4, %%mm7 \n\t" // QP,..., QP | 2663 "movq %4, %%mm7 \n\t" // QP,..., QP |
2722 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP | 2664 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
2723 "psubusb %%mm4, %%mm7 \n\t" // Diff >=2QP -> 0 | 2665 "psubusb %%mm4, %%mm7 \n\t" // Diff >=2QP -> 0 |
2724 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0 | 2666 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0 |
2725 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0 | 2667 "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0 |
2726 "movq %%mm7, %1 \n\t" | 2668 "movq %%mm7, %1 \n\t" |
2727 | 2669 |
2728 "movq %5, %%mm7 \n\t" | 2670 "movq %5, %%mm7 \n\t" |
2729 "punpcklbw %%mm7, %%mm7 \n\t" | 2671 "punpcklbw %%mm7, %%mm7 \n\t" |
2730 "punpcklbw %%mm7, %%mm7 \n\t" | 2672 "punpcklbw %%mm7, %%mm7 \n\t" |
2731 "punpcklbw %%mm7, %%mm7 \n\t" | 2673 "punpcklbw %%mm7, %%mm7 \n\t" |
2732 "psubb %%mm0, %%mm6 \n\t" | 2674 "psubb %%mm0, %%mm6 \n\t" |
2733 "pcmpgtb %%mm7, %%mm6 \n\t" | 2675 "pcmpgtb %%mm7, %%mm6 \n\t" |
2734 "movq %%mm6, %0 \n\t" | 2676 "movq %%mm6, %0 \n\t" |
2735 | 2677 |
2736 : "=m" (eq_mask), "=m" (dc_mask) | 2678 : "=m" (eq_mask), "=m" (dc_mask) |
2737 : "r" (src), "r" ((long)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold) | 2679 : "r" (src), "r" ((long)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold) |
2738 : "%"REG_a | 2680 : "%"REG_a |
2739 ); | 2681 ); |
2740 | 2682 |
2741 both_masks = dc_mask & eq_mask; | 2683 both_masks = dc_mask & eq_mask; |
2742 | 2684 |
2743 if(both_masks){ | 2685 if(both_masks){ |
2744 long offset= -8*step; | 2686 long offset= -8*step; |
2745 int64_t *temp_sums= sums; | 2687 int64_t *temp_sums= sums; |
2746 | 2688 |
2747 asm volatile( | 2689 asm volatile( |
2748 "movq %2, %%mm0 \n\t" // QP,..., QP | 2690 "movq %2, %%mm0 \n\t" // QP,..., QP |
2749 "pxor %%mm4, %%mm4 \n\t" | 2691 "pxor %%mm4, %%mm4 \n\t" |
2750 | 2692 |
2751 "movq (%0), %%mm6 \n\t" | 2693 "movq (%0), %%mm6 \n\t" |
2752 "movq (%0, %1), %%mm5 \n\t" | 2694 "movq (%0, %1), %%mm5 \n\t" |
2753 "movq %%mm5, %%mm1 \n\t" | 2695 "movq %%mm5, %%mm1 \n\t" |
2754 "movq %%mm6, %%mm2 \n\t" | 2696 "movq %%mm6, %%mm2 \n\t" |
2755 "psubusb %%mm6, %%mm5 \n\t" | 2697 "psubusb %%mm6, %%mm5 \n\t" |
2756 "psubusb %%mm1, %%mm2 \n\t" | 2698 "psubusb %%mm1, %%mm2 \n\t" |
2757 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines | 2699 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
2758 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0 | 2700 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0 |
2759 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF | 2701 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF |
2760 | 2702 |
2761 "pxor %%mm6, %%mm1 \n\t" | 2703 "pxor %%mm6, %%mm1 \n\t" |
2762 "pand %%mm0, %%mm1 \n\t" | 2704 "pand %%mm0, %%mm1 \n\t" |
2763 "pxor %%mm1, %%mm6 \n\t" | 2705 "pxor %%mm1, %%mm6 \n\t" |
2764 // 0:QP 6:First | 2706 // 0:QP 6:First |
2765 | 2707 |
2766 "movq (%0, %1, 8), %%mm5 \n\t" | 2708 "movq (%0, %1, 8), %%mm5 \n\t" |
2767 "add %1, %0 \n\t" // %0 points to line 1 not 0 | 2709 "add %1, %0 \n\t" // %0 points to line 1 not 0 |
2768 "movq (%0, %1, 8), %%mm7 \n\t" | 2710 "movq (%0, %1, 8), %%mm7 \n\t" |
2769 "movq %%mm5, %%mm1 \n\t" | 2711 "movq %%mm5, %%mm1 \n\t" |
2770 "movq %%mm7, %%mm2 \n\t" | 2712 "movq %%mm7, %%mm2 \n\t" |
2771 "psubusb %%mm7, %%mm5 \n\t" | 2713 "psubusb %%mm7, %%mm5 \n\t" |
2772 "psubusb %%mm1, %%mm2 \n\t" | 2714 "psubusb %%mm1, %%mm2 \n\t" |
2773 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines | 2715 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
2774 "movq %2, %%mm0 \n\t" // QP,..., QP | 2716 "movq %2, %%mm0 \n\t" // QP,..., QP |
2775 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0 | 2717 "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0 |
2776 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF | 2718 "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF |
2777 | 2719 |
2778 "pxor %%mm7, %%mm1 \n\t" | 2720 "pxor %%mm7, %%mm1 \n\t" |
2779 "pand %%mm0, %%mm1 \n\t" | 2721 "pand %%mm0, %%mm1 \n\t" |
2780 "pxor %%mm1, %%mm7 \n\t" | 2722 "pxor %%mm1, %%mm7 \n\t" |
2781 | 2723 |
2782 "movq %%mm6, %%mm5 \n\t" | 2724 "movq %%mm6, %%mm5 \n\t" |
2783 "punpckhbw %%mm4, %%mm6 \n\t" | 2725 "punpckhbw %%mm4, %%mm6 \n\t" |
2784 "punpcklbw %%mm4, %%mm5 \n\t" | 2726 "punpcklbw %%mm4, %%mm5 \n\t" |
2785 // 4:0 5/6:First 7:Last | 2727 // 4:0 5/6:First 7:Last |
2786 | 2728 |
2787 "movq %%mm5, %%mm0 \n\t" | 2729 "movq %%mm5, %%mm0 \n\t" |
2788 "movq %%mm6, %%mm1 \n\t" | 2730 "movq %%mm6, %%mm1 \n\t" |
2789 "psllw $2, %%mm0 \n\t" | 2731 "psllw $2, %%mm0 \n\t" |
2790 "psllw $2, %%mm1 \n\t" | 2732 "psllw $2, %%mm1 \n\t" |
2791 "paddw "MANGLE(w04)", %%mm0 \n\t" | 2733 "paddw "MANGLE(w04)", %%mm0 \n\t" |
2792 "paddw "MANGLE(w04)", %%mm1 \n\t" | 2734 "paddw "MANGLE(w04)", %%mm1 \n\t" |
2793 | 2735 |
2794 #define NEXT\ | 2736 #define NEXT\ |
2795 "movq (%0), %%mm2 \n\t"\ | 2737 "movq (%0), %%mm2 \n\t"\ |
2796 "movq (%0), %%mm3 \n\t"\ | 2738 "movq (%0), %%mm3 \n\t"\ |
2797 "add %1, %0 \n\t"\ | 2739 "add %1, %0 \n\t"\ |
2798 "punpcklbw %%mm4, %%mm2 \n\t"\ | 2740 "punpcklbw %%mm4, %%mm2 \n\t"\ |
2799 "punpckhbw %%mm4, %%mm3 \n\t"\ | 2741 "punpckhbw %%mm4, %%mm3 \n\t"\ |
2800 "paddw %%mm2, %%mm0 \n\t"\ | 2742 "paddw %%mm2, %%mm0 \n\t"\ |
2801 "paddw %%mm3, %%mm1 \n\t" | 2743 "paddw %%mm3, %%mm1 \n\t" |
2802 | 2744 |
2803 #define PREV\ | 2745 #define PREV\ |
2804 "movq (%0), %%mm2 \n\t"\ | 2746 "movq (%0), %%mm2 \n\t"\ |
2805 "movq (%0), %%mm3 \n\t"\ | 2747 "movq (%0), %%mm3 \n\t"\ |
2806 "add %1, %0 \n\t"\ | 2748 "add %1, %0 \n\t"\ |
2807 "punpcklbw %%mm4, %%mm2 \n\t"\ | 2749 "punpcklbw %%mm4, %%mm2 \n\t"\ |
2808 "punpckhbw %%mm4, %%mm3 \n\t"\ | 2750 "punpckhbw %%mm4, %%mm3 \n\t"\ |
2809 "psubw %%mm2, %%mm0 \n\t"\ | 2751 "psubw %%mm2, %%mm0 \n\t"\ |
2810 "psubw %%mm3, %%mm1 \n\t" | 2752 "psubw %%mm3, %%mm1 \n\t" |
2811 | 2753 |
2812 | 2754 |
2813 NEXT //0 | 2755 NEXT //0 |
2814 NEXT //1 | 2756 NEXT //1 |
2815 NEXT //2 | 2757 NEXT //2 |
2816 "movq %%mm0, (%3) \n\t" | 2758 "movq %%mm0, (%3) \n\t" |
2817 "movq %%mm1, 8(%3) \n\t" | 2759 "movq %%mm1, 8(%3) \n\t" |
2818 | 2760 |
2819 NEXT //3 | 2761 NEXT //3 |
2820 "psubw %%mm5, %%mm0 \n\t" | 2762 "psubw %%mm5, %%mm0 \n\t" |
2821 "psubw %%mm6, %%mm1 \n\t" | 2763 "psubw %%mm6, %%mm1 \n\t" |
2822 "movq %%mm0, 16(%3) \n\t" | 2764 "movq %%mm0, 16(%3) \n\t" |
2823 "movq %%mm1, 24(%3) \n\t" | 2765 "movq %%mm1, 24(%3) \n\t" |
2824 | 2766 |
2825 NEXT //4 | 2767 NEXT //4 |
2826 "psubw %%mm5, %%mm0 \n\t" | 2768 "psubw %%mm5, %%mm0 \n\t" |
2827 "psubw %%mm6, %%mm1 \n\t" | 2769 "psubw %%mm6, %%mm1 \n\t" |
2828 "movq %%mm0, 32(%3) \n\t" | 2770 "movq %%mm0, 32(%3) \n\t" |
2829 "movq %%mm1, 40(%3) \n\t" | 2771 "movq %%mm1, 40(%3) \n\t" |
2830 | 2772 |
2831 NEXT //5 | 2773 NEXT //5 |
2832 "psubw %%mm5, %%mm0 \n\t" | 2774 "psubw %%mm5, %%mm0 \n\t" |
2833 "psubw %%mm6, %%mm1 \n\t" | 2775 "psubw %%mm6, %%mm1 \n\t" |
2834 "movq %%mm0, 48(%3) \n\t" | 2776 "movq %%mm0, 48(%3) \n\t" |
2835 "movq %%mm1, 56(%3) \n\t" | 2777 "movq %%mm1, 56(%3) \n\t" |
2836 | 2778 |
2837 NEXT //6 | 2779 NEXT //6 |
2838 "psubw %%mm5, %%mm0 \n\t" | 2780 "psubw %%mm5, %%mm0 \n\t" |
2839 "psubw %%mm6, %%mm1 \n\t" | 2781 "psubw %%mm6, %%mm1 \n\t" |
2840 "movq %%mm0, 64(%3) \n\t" | 2782 "movq %%mm0, 64(%3) \n\t" |
2841 "movq %%mm1, 72(%3) \n\t" | 2783 "movq %%mm1, 72(%3) \n\t" |
2842 | 2784 |
2843 "movq %%mm7, %%mm6 \n\t" | 2785 "movq %%mm7, %%mm6 \n\t" |
2844 "punpckhbw %%mm4, %%mm7 \n\t" | 2786 "punpckhbw %%mm4, %%mm7 \n\t" |
2845 "punpcklbw %%mm4, %%mm6 \n\t" | 2787 "punpcklbw %%mm4, %%mm6 \n\t" |
2846 | 2788 |
2847 NEXT //7 | 2789 NEXT //7 |
2848 "mov %4, %0 \n\t" | 2790 "mov %4, %0 \n\t" |
2849 "add %1, %0 \n\t" | 2791 "add %1, %0 \n\t" |
2850 PREV //0 | 2792 PREV //0 |
2851 "movq %%mm0, 80(%3) \n\t" | 2793 "movq %%mm0, 80(%3) \n\t" |
2852 "movq %%mm1, 88(%3) \n\t" | 2794 "movq %%mm1, 88(%3) \n\t" |
2853 | 2795 |
2854 PREV //1 | 2796 PREV //1 |
2855 "paddw %%mm6, %%mm0 \n\t" | 2797 "paddw %%mm6, %%mm0 \n\t" |
2856 "paddw %%mm7, %%mm1 \n\t" | 2798 "paddw %%mm7, %%mm1 \n\t" |
2857 "movq %%mm0, 96(%3) \n\t" | 2799 "movq %%mm0, 96(%3) \n\t" |
2858 "movq %%mm1, 104(%3) \n\t" | 2800 "movq %%mm1, 104(%3) \n\t" |
2859 | 2801 |
2860 PREV //2 | 2802 PREV //2 |
2861 "paddw %%mm6, %%mm0 \n\t" | 2803 "paddw %%mm6, %%mm0 \n\t" |
2862 "paddw %%mm7, %%mm1 \n\t" | 2804 "paddw %%mm7, %%mm1 \n\t" |
2863 "movq %%mm0, 112(%3) \n\t" | 2805 "movq %%mm0, 112(%3) \n\t" |
2864 "movq %%mm1, 120(%3) \n\t" | 2806 "movq %%mm1, 120(%3) \n\t" |
2865 | 2807 |
2866 PREV //3 | 2808 PREV //3 |
2867 "paddw %%mm6, %%mm0 \n\t" | 2809 "paddw %%mm6, %%mm0 \n\t" |
2868 "paddw %%mm7, %%mm1 \n\t" | 2810 "paddw %%mm7, %%mm1 \n\t" |
2869 "movq %%mm0, 128(%3) \n\t" | 2811 "movq %%mm0, 128(%3) \n\t" |
2870 "movq %%mm1, 136(%3) \n\t" | 2812 "movq %%mm1, 136(%3) \n\t" |
2871 | 2813 |
2872 PREV //4 | 2814 PREV //4 |
2873 "paddw %%mm6, %%mm0 \n\t" | 2815 "paddw %%mm6, %%mm0 \n\t" |
2874 "paddw %%mm7, %%mm1 \n\t" | 2816 "paddw %%mm7, %%mm1 \n\t" |
2875 "movq %%mm0, 144(%3) \n\t" | 2817 "movq %%mm0, 144(%3) \n\t" |
2876 "movq %%mm1, 152(%3) \n\t" | 2818 "movq %%mm1, 152(%3) \n\t" |
2877 | 2819 |
2878 "mov %4, %0 \n\t" //FIXME | 2820 "mov %4, %0 \n\t" //FIXME |
2879 | 2821 |
2880 : "+&r"(src) | 2822 : "+&r"(src) |
2881 : "r" ((long)step), "m" (c->pQPb), "r"(sums), "g"(src) | 2823 : "r" ((long)step), "m" (c->pQPb), "r"(sums), "g"(src) |
2882 ); | 2824 ); |
2883 | 2825 |
2884 src+= step; // src points to begin of the 8x8 Block | 2826 src+= step; // src points to begin of the 8x8 Block |
2885 | 2827 |
2886 asm volatile( | 2828 asm volatile( |
2887 "movq %4, %%mm6 \n\t" | 2829 "movq %4, %%mm6 \n\t" |
2888 "pcmpeqb %%mm5, %%mm5 \n\t" | 2830 "pcmpeqb %%mm5, %%mm5 \n\t" |
2889 "pxor %%mm6, %%mm5 \n\t" | 2831 "pxor %%mm6, %%mm5 \n\t" |
2890 "pxor %%mm7, %%mm7 \n\t" | 2832 "pxor %%mm7, %%mm7 \n\t" |
2891 | 2833 |
2892 "1: \n\t" | 2834 "1: \n\t" |
2893 "movq (%1), %%mm0 \n\t" | 2835 "movq (%1), %%mm0 \n\t" |
2894 "movq 8(%1), %%mm1 \n\t" | 2836 "movq 8(%1), %%mm1 \n\t" |
2895 "paddw 32(%1), %%mm0 \n\t" | 2837 "paddw 32(%1), %%mm0 \n\t" |
2896 "paddw 40(%1), %%mm1 \n\t" | 2838 "paddw 40(%1), %%mm1 \n\t" |
2897 "movq (%0, %3), %%mm2 \n\t" | 2839 "movq (%0, %3), %%mm2 \n\t" |
2898 "movq %%mm2, %%mm3 \n\t" | 2840 "movq %%mm2, %%mm3 \n\t" |
2899 "movq %%mm2, %%mm4 \n\t" | 2841 "movq %%mm2, %%mm4 \n\t" |
2900 "punpcklbw %%mm7, %%mm2 \n\t" | 2842 "punpcklbw %%mm7, %%mm2 \n\t" |
2901 "punpckhbw %%mm7, %%mm3 \n\t" | 2843 "punpckhbw %%mm7, %%mm3 \n\t" |
2902 "paddw %%mm2, %%mm0 \n\t" | 2844 "paddw %%mm2, %%mm0 \n\t" |
2903 "paddw %%mm3, %%mm1 \n\t" | 2845 "paddw %%mm3, %%mm1 \n\t" |
2904 "paddw %%mm2, %%mm0 \n\t" | 2846 "paddw %%mm2, %%mm0 \n\t" |
2905 "paddw %%mm3, %%mm1 \n\t" | 2847 "paddw %%mm3, %%mm1 \n\t" |
2906 "psrlw $4, %%mm0 \n\t" | 2848 "psrlw $4, %%mm0 \n\t" |
2907 "psrlw $4, %%mm1 \n\t" | 2849 "psrlw $4, %%mm1 \n\t" |
2908 "packuswb %%mm1, %%mm0 \n\t" | 2850 "packuswb %%mm1, %%mm0 \n\t" |
2909 "pand %%mm6, %%mm0 \n\t" | 2851 "pand %%mm6, %%mm0 \n\t" |
2910 "pand %%mm5, %%mm4 \n\t" | 2852 "pand %%mm5, %%mm4 \n\t" |
2911 "por %%mm4, %%mm0 \n\t" | 2853 "por %%mm4, %%mm0 \n\t" |
2912 "movq %%mm0, (%0, %3) \n\t" | 2854 "movq %%mm0, (%0, %3) \n\t" |
2913 "add $16, %1 \n\t" | 2855 "add $16, %1 \n\t" |
2914 "add %2, %0 \n\t" | 2856 "add %2, %0 \n\t" |
2915 " js 1b \n\t" | 2857 " js 1b \n\t" |
2916 | 2858 |
2917 : "+r"(offset), "+r"(temp_sums) | 2859 : "+r"(offset), "+r"(temp_sums) |
2918 : "r" ((long)step), "r"(src - offset), "m"(both_masks) | 2860 : "r" ((long)step), "r"(src - offset), "m"(both_masks) |
2919 ); | 2861 ); |
2920 }else | 2862 }else |
2921 src+= step; // src points to begin of the 8x8 Block | 2863 src+= step; // src points to begin of the 8x8 Block |
2922 | 2864 |
2923 if(eq_mask != -1LL){ | 2865 if(eq_mask != -1LL){ |
2924 uint8_t *temp_src= src; | 2866 uint8_t *temp_src= src; |
2925 asm volatile( | 2867 asm volatile( |
2926 "pxor %%mm7, %%mm7 \n\t" | 2868 "pxor %%mm7, %%mm7 \n\t" |
2927 "lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars | 2869 "lea -40(%%"REG_SP"), %%"REG_c" \n\t" // make space for 4 8-byte vars |
2928 "and "ALIGN_MASK", %%"REG_c" \n\t" // align | 2870 "and "ALIGN_MASK", %%"REG_c" \n\t" // align |
2929 // 0 1 2 3 4 5 6 7 8 9 | 2871 // 0 1 2 3 4 5 6 7 8 9 |
2930 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %1+8%1 ecx+4%1 | 2872 // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %1+8%1 ecx+4%1 |
2931 | 2873 |
2932 "movq (%0), %%mm0 \n\t" | 2874 "movq (%0), %%mm0 \n\t" |
2933 "movq %%mm0, %%mm1 \n\t" | 2875 "movq %%mm0, %%mm1 \n\t" |
2934 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 | 2876 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
2935 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 | 2877 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
2936 | 2878 |
2937 "movq (%0, %1), %%mm2 \n\t" | 2879 "movq (%0, %1), %%mm2 \n\t" |
2938 "lea (%0, %1, 2), %%"REG_a" \n\t" | 2880 "lea (%0, %1, 2), %%"REG_a" \n\t" |
2939 "movq %%mm2, %%mm3 \n\t" | 2881 "movq %%mm2, %%mm3 \n\t" |
2940 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 | 2882 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
2941 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 | 2883 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
2942 | 2884 |
2943 "movq (%%"REG_a"), %%mm4 \n\t" | 2885 "movq (%%"REG_a"), %%mm4 \n\t" |
2944 "movq %%mm4, %%mm5 \n\t" | 2886 "movq %%mm4, %%mm5 \n\t" |
2945 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 | 2887 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
2946 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 | 2888 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
2947 | 2889 |
2948 "paddw %%mm0, %%mm0 \n\t" // 2L0 | 2890 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
2949 "paddw %%mm1, %%mm1 \n\t" // 2H0 | 2891 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
2950 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 | 2892 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
2951 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 | 2893 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
2952 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 | 2894 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
2953 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 | 2895 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
2954 | 2896 |
2955 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 | 2897 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
2956 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 | 2898 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
2957 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 | 2899 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
2958 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 | 2900 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
2959 | 2901 |
2960 "movq (%%"REG_a", %1), %%mm2 \n\t" | 2902 "movq (%%"REG_a", %1), %%mm2 \n\t" |
2961 "movq %%mm2, %%mm3 \n\t" | 2903 "movq %%mm2, %%mm3 \n\t" |
2962 "punpcklbw %%mm7, %%mm2 \n\t" // L3 | 2904 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
2963 "punpckhbw %%mm7, %%mm3 \n\t" // H3 | 2905 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
2964 | 2906 |
2965 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 | 2907 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
2966 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 | 2908 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
2967 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 | 2909 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
2968 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 | 2910 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
2969 "movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3 | 2911 "movq %%mm0, (%%"REG_c") \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
2970 "movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3 | 2912 "movq %%mm1, 8(%%"REG_c") \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
2971 | 2913 |
2972 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" | 2914 "movq (%%"REG_a", %1, 2), %%mm0 \n\t" |
2973 "movq %%mm0, %%mm1 \n\t" | 2915 "movq %%mm0, %%mm1 \n\t" |
2974 "punpcklbw %%mm7, %%mm0 \n\t" // L4 | 2916 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
2975 "punpckhbw %%mm7, %%mm1 \n\t" // H4 | 2917 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
2976 | 2918 |
2977 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 | 2919 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
2978 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 | 2920 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
2979 "movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4 | 2921 "movq %%mm2, 16(%%"REG_c") \n\t" // L3 - L4 |
2980 "movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4 | 2922 "movq %%mm3, 24(%%"REG_c") \n\t" // H3 - H4 |
2981 "paddw %%mm4, %%mm4 \n\t" // 2L2 | 2923 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
2982 "paddw %%mm5, %%mm5 \n\t" // 2H2 | 2924 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
2983 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 | 2925 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
2984 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 | 2926 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
2985 | 2927 |
2986 "lea (%%"REG_a", %1), %0 \n\t" | 2928 "lea (%%"REG_a", %1), %0 \n\t" |
2987 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 | 2929 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
2988 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 | 2930 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
2989 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 | 2931 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
2990 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 | 2932 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
2991 //50 opcodes so far | 2933 //50 opcodes so far |
2992 "movq (%0, %1, 2), %%mm2 \n\t" | 2934 "movq (%0, %1, 2), %%mm2 \n\t" |
2993 "movq %%mm2, %%mm3 \n\t" | 2935 "movq %%mm2, %%mm3 \n\t" |
2994 "punpcklbw %%mm7, %%mm2 \n\t" // L5 | 2936 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
2995 "punpckhbw %%mm7, %%mm3 \n\t" // H5 | 2937 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
2996 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 | 2938 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
2997 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 | 2939 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
2998 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 | 2940 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
2999 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 | 2941 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
3000 | 2942 |
3001 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" | 2943 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" |
3002 "punpcklbw %%mm7, %%mm6 \n\t" // L6 | 2944 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
3003 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 | 2945 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
3004 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" | 2946 "movq (%%"REG_a", %1, 4), %%mm6 \n\t" |
3005 "punpckhbw %%mm7, %%mm6 \n\t" // H6 | 2947 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
3006 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 | 2948 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
3007 | 2949 |
3008 "paddw %%mm0, %%mm0 \n\t" // 2L4 | 2950 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
3009 "paddw %%mm1, %%mm1 \n\t" // 2H4 | 2951 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
3010 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 | 2952 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
3011 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 | 2953 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
3012 | 2954 |
3013 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 | 2955 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
3014 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 | 2956 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
3015 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 | 2957 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
3016 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 | 2958 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
3017 | 2959 |
3018 "movq (%0, %1, 4), %%mm2 \n\t" | 2960 "movq (%0, %1, 4), %%mm2 \n\t" |
3019 "movq %%mm2, %%mm3 \n\t" | 2961 "movq %%mm2, %%mm3 \n\t" |
3020 "punpcklbw %%mm7, %%mm2 \n\t" // L7 | 2962 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
3021 "punpckhbw %%mm7, %%mm3 \n\t" // H7 | 2963 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
3022 | 2964 |
3023 "paddw %%mm2, %%mm2 \n\t" // 2L7 | 2965 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
3024 "paddw %%mm3, %%mm3 \n\t" // 2H7 | 2966 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
3025 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 | 2967 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
3026 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 | 2968 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
3027 | 2969 |
3028 "movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 | 2970 "movq (%%"REG_c"), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
3029 "movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 | 2971 "movq 8(%%"REG_c"), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
3030 | 2972 |
3031 #ifdef HAVE_MMX2 | 2973 #ifdef HAVE_MMX2 |
3032 "movq %%mm7, %%mm6 \n\t" // 0 | 2974 "movq %%mm7, %%mm6 \n\t" // 0 |
3033 "psubw %%mm0, %%mm6 \n\t" | 2975 "psubw %%mm0, %%mm6 \n\t" |
3034 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| | 2976 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
3035 "movq %%mm7, %%mm6 \n\t" // 0 | 2977 "movq %%mm7, %%mm6 \n\t" // 0 |
3036 "psubw %%mm1, %%mm6 \n\t" | 2978 "psubw %%mm1, %%mm6 \n\t" |
3037 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| | 2979 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
3038 "movq %%mm7, %%mm6 \n\t" // 0 | 2980 "movq %%mm7, %%mm6 \n\t" // 0 |
3039 "psubw %%mm2, %%mm6 \n\t" | 2981 "psubw %%mm2, %%mm6 \n\t" |
3040 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| | 2982 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
3041 "movq %%mm7, %%mm6 \n\t" // 0 | 2983 "movq %%mm7, %%mm6 \n\t" // 0 |
3042 "psubw %%mm3, %%mm6 \n\t" | 2984 "psubw %%mm3, %%mm6 \n\t" |
3043 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| | 2985 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
3044 #else | 2986 #else |
3045 "movq %%mm7, %%mm6 \n\t" // 0 | 2987 "movq %%mm7, %%mm6 \n\t" // 0 |
3046 "pcmpgtw %%mm0, %%mm6 \n\t" | 2988 "pcmpgtw %%mm0, %%mm6 \n\t" |
3047 "pxor %%mm6, %%mm0 \n\t" | 2989 "pxor %%mm6, %%mm0 \n\t" |
3048 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| | 2990 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
3049 "movq %%mm7, %%mm6 \n\t" // 0 | 2991 "movq %%mm7, %%mm6 \n\t" // 0 |
3050 "pcmpgtw %%mm1, %%mm6 \n\t" | 2992 "pcmpgtw %%mm1, %%mm6 \n\t" |
3051 "pxor %%mm6, %%mm1 \n\t" | 2993 "pxor %%mm6, %%mm1 \n\t" |
3052 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| | 2994 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
3053 "movq %%mm7, %%mm6 \n\t" // 0 | 2995 "movq %%mm7, %%mm6 \n\t" // 0 |
3054 "pcmpgtw %%mm2, %%mm6 \n\t" | 2996 "pcmpgtw %%mm2, %%mm6 \n\t" |
3055 "pxor %%mm6, %%mm2 \n\t" | 2997 "pxor %%mm6, %%mm2 \n\t" |
3056 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| | 2998 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
3057 "movq %%mm7, %%mm6 \n\t" // 0 | 2999 "movq %%mm7, %%mm6 \n\t" // 0 |
3058 "pcmpgtw %%mm3, %%mm6 \n\t" | 3000 "pcmpgtw %%mm3, %%mm6 \n\t" |
3059 "pxor %%mm6, %%mm3 \n\t" | 3001 "pxor %%mm6, %%mm3 \n\t" |
3060 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| | 3002 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
3061 #endif | 3003 #endif |
3062 | 3004 |
3063 #ifdef HAVE_MMX2 | 3005 #ifdef HAVE_MMX2 |
3064 "pminsw %%mm2, %%mm0 \n\t" | 3006 "pminsw %%mm2, %%mm0 \n\t" |
3065 "pminsw %%mm3, %%mm1 \n\t" | 3007 "pminsw %%mm3, %%mm1 \n\t" |
3066 #else | 3008 #else |
3067 "movq %%mm0, %%mm6 \n\t" | 3009 "movq %%mm0, %%mm6 \n\t" |
3068 "psubusw %%mm2, %%mm6 \n\t" | 3010 "psubusw %%mm2, %%mm6 \n\t" |
3069 "psubw %%mm6, %%mm0 \n\t" | 3011 "psubw %%mm6, %%mm0 \n\t" |
3070 "movq %%mm1, %%mm6 \n\t" | 3012 "movq %%mm1, %%mm6 \n\t" |
3071 "psubusw %%mm3, %%mm6 \n\t" | 3013 "psubusw %%mm3, %%mm6 \n\t" |
3072 "psubw %%mm6, %%mm1 \n\t" | 3014 "psubw %%mm6, %%mm1 \n\t" |
3073 #endif | 3015 #endif |
3074 | 3016 |
3075 "movd %2, %%mm2 \n\t" // QP | 3017 "movd %2, %%mm2 \n\t" // QP |
3076 "punpcklbw %%mm7, %%mm2 \n\t" | 3018 "punpcklbw %%mm7, %%mm2 \n\t" |
3077 | 3019 |
3078 "movq %%mm7, %%mm6 \n\t" // 0 | 3020 "movq %%mm7, %%mm6 \n\t" // 0 |
3079 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) | 3021 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
3080 "pxor %%mm6, %%mm4 \n\t" | 3022 "pxor %%mm6, %%mm4 \n\t" |
3081 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| | 3023 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
3082 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) | 3024 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
3083 "pxor %%mm7, %%mm5 \n\t" | 3025 "pxor %%mm7, %%mm5 \n\t" |
3084 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| | 3026 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
3085 // 100 opcodes | 3027 // 100 opcodes |
3086 "psllw $3, %%mm2 \n\t" // 8QP | 3028 "psllw $3, %%mm2 \n\t" // 8QP |
3087 "movq %%mm2, %%mm3 \n\t" // 8QP | 3029 "movq %%mm2, %%mm3 \n\t" // 8QP |
3088 "pcmpgtw %%mm4, %%mm2 \n\t" | 3030 "pcmpgtw %%mm4, %%mm2 \n\t" |
3089 "pcmpgtw %%mm5, %%mm3 \n\t" | 3031 "pcmpgtw %%mm5, %%mm3 \n\t" |
3090 "pand %%mm2, %%mm4 \n\t" | 3032 "pand %%mm2, %%mm4 \n\t" |
3091 "pand %%mm3, %%mm5 \n\t" | 3033 "pand %%mm3, %%mm5 \n\t" |
3092 | 3034 |
3093 | 3035 |
3094 "psubusw %%mm0, %%mm4 \n\t" // hd | 3036 "psubusw %%mm0, %%mm4 \n\t" // hd |
3095 "psubusw %%mm1, %%mm5 \n\t" // ld | 3037 "psubusw %%mm1, %%mm5 \n\t" // ld |
3096 | 3038 |
3097 | 3039 |
3098 "movq "MANGLE(w05)", %%mm2 \n\t" // 5 | 3040 "movq "MANGLE(w05)", %%mm2 \n\t" // 5 |
3099 "pmullw %%mm2, %%mm4 \n\t" | 3041 "pmullw %%mm2, %%mm4 \n\t" |
3100 "pmullw %%mm2, %%mm5 \n\t" | 3042 "pmullw %%mm2, %%mm5 \n\t" |
3101 "movq "MANGLE(w20)", %%mm2 \n\t" // 32 | 3043 "movq "MANGLE(w20)", %%mm2 \n\t" // 32 |
3102 "paddw %%mm2, %%mm4 \n\t" | 3044 "paddw %%mm2, %%mm4 \n\t" |
3103 "paddw %%mm2, %%mm5 \n\t" | 3045 "paddw %%mm2, %%mm5 \n\t" |
3104 "psrlw $6, %%mm4 \n\t" | 3046 "psrlw $6, %%mm4 \n\t" |
3105 "psrlw $6, %%mm5 \n\t" | 3047 "psrlw $6, %%mm5 \n\t" |
3106 | 3048 |
3107 "movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4 | 3049 "movq 16(%%"REG_c"), %%mm0 \n\t" // L3 - L4 |
3108 "movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4 | 3050 "movq 24(%%"REG_c"), %%mm1 \n\t" // H3 - H4 |
3109 | 3051 |
3110 "pxor %%mm2, %%mm2 \n\t" | 3052 "pxor %%mm2, %%mm2 \n\t" |
3111 "pxor %%mm3, %%mm3 \n\t" | 3053 "pxor %%mm3, %%mm3 \n\t" |
3112 | 3054 |
3113 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) | 3055 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
3114 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) | 3056 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
3115 "pxor %%mm2, %%mm0 \n\t" | 3057 "pxor %%mm2, %%mm0 \n\t" |
3116 "pxor %%mm3, %%mm1 \n\t" | 3058 "pxor %%mm3, %%mm1 \n\t" |
3117 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| | 3059 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
3118 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| | 3060 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| |
3119 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 | 3061 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 |
3120 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 | 3062 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
3121 | 3063 |
3122 "pxor %%mm6, %%mm2 \n\t" | 3064 "pxor %%mm6, %%mm2 \n\t" |
3123 "pxor %%mm7, %%mm3 \n\t" | 3065 "pxor %%mm7, %%mm3 \n\t" |
3124 "pand %%mm2, %%mm4 \n\t" | 3066 "pand %%mm2, %%mm4 \n\t" |
3125 "pand %%mm3, %%mm5 \n\t" | 3067 "pand %%mm3, %%mm5 \n\t" |
3126 | 3068 |
3127 #ifdef HAVE_MMX2 | 3069 #ifdef HAVE_MMX2 |
3128 "pminsw %%mm0, %%mm4 \n\t" | 3070 "pminsw %%mm0, %%mm4 \n\t" |
3129 "pminsw %%mm1, %%mm5 \n\t" | 3071 "pminsw %%mm1, %%mm5 \n\t" |
3130 #else | 3072 #else |
3131 "movq %%mm4, %%mm2 \n\t" | 3073 "movq %%mm4, %%mm2 \n\t" |
3132 "psubusw %%mm0, %%mm2 \n\t" | 3074 "psubusw %%mm0, %%mm2 \n\t" |
3133 "psubw %%mm2, %%mm4 \n\t" | 3075 "psubw %%mm2, %%mm4 \n\t" |
3134 "movq %%mm5, %%mm2 \n\t" | 3076 "movq %%mm5, %%mm2 \n\t" |
3135 "psubusw %%mm1, %%mm2 \n\t" | 3077 "psubusw %%mm1, %%mm2 \n\t" |
3136 "psubw %%mm2, %%mm5 \n\t" | 3078 "psubw %%mm2, %%mm5 \n\t" |
3137 #endif | 3079 #endif |
3138 "pxor %%mm6, %%mm4 \n\t" | 3080 "pxor %%mm6, %%mm4 \n\t" |
3139 "pxor %%mm7, %%mm5 \n\t" | 3081 "pxor %%mm7, %%mm5 \n\t" |
3140 "psubw %%mm6, %%mm4 \n\t" | 3082 "psubw %%mm6, %%mm4 \n\t" |
3141 "psubw %%mm7, %%mm5 \n\t" | 3083 "psubw %%mm7, %%mm5 \n\t" |
3142 "packsswb %%mm5, %%mm4 \n\t" | 3084 "packsswb %%mm5, %%mm4 \n\t" |
3143 "movq %3, %%mm1 \n\t" | 3085 "movq %3, %%mm1 \n\t" |
3144 "pandn %%mm4, %%mm1 \n\t" | 3086 "pandn %%mm4, %%mm1 \n\t" |
3145 "movq (%0), %%mm0 \n\t" | 3087 "movq (%0), %%mm0 \n\t" |
3146 "paddb %%mm1, %%mm0 \n\t" | 3088 "paddb %%mm1, %%mm0 \n\t" |
3147 "movq %%mm0, (%0) \n\t" | 3089 "movq %%mm0, (%0) \n\t" |
3148 "movq (%0, %1), %%mm0 \n\t" | 3090 "movq (%0, %1), %%mm0 \n\t" |
3149 "psubb %%mm1, %%mm0 \n\t" | 3091 "psubb %%mm1, %%mm0 \n\t" |
3150 "movq %%mm0, (%0, %1) \n\t" | 3092 "movq %%mm0, (%0, %1) \n\t" |
3151 | 3093 |
3152 : "+r" (temp_src) | 3094 : "+r" (temp_src) |
3153 : "r" ((long)step), "m" (c->pQPb), "m"(eq_mask) | 3095 : "r" ((long)step), "m" (c->pQPb), "m"(eq_mask) |
3154 : "%"REG_a, "%"REG_c | 3096 : "%"REG_a, "%"REG_c |
3155 ); | 3097 ); |
3156 } | 3098 } |
3157 /*if(step==16){ | 3099 /*if(step==16){ |
3158 STOP_TIMER("step16") | 3100 STOP_TIMER("step16") |
3159 }else{ | 3101 }else{ |
3160 STOP_TIMER("stepX") | 3102 STOP_TIMER("stepX") |
3161 }*/ | 3103 }*/ |
3162 } | 3104 } |
3163 #endif //HAVE_MMX | 3105 #endif //HAVE_MMX |
3164 | 3106 |
3165 static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, | 3107 static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
3166 const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c); | 3108 const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c); |
3167 | 3109 |
3168 /** | 3110 /** |
3169 * Copies a block from src to dst and fixes the blacklevel. | 3111 * Copies a block from src to dst and fixes the blacklevel. |
3170 * levelFix == 0 -> do not touch the brighness & contrast | 3112 * levelFix == 0 -> do not touch the brighness & contrast |
3171 */ | 3113 */ |
3172 #undef SCALED_CPY | 3114 #undef SCALED_CPY |
3173 | 3115 |
3174 static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, const uint8_t src[], int srcStride, | 3116 static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, const uint8_t src[], int srcStride, |
3175 int levelFix, int64_t *packedOffsetAndScale) | 3117 int levelFix, int64_t *packedOffsetAndScale) |
3176 { | 3118 { |
3177 #ifndef HAVE_MMX | 3119 #ifndef HAVE_MMX |
3178 int i; | 3120 int i; |
3179 #endif | 3121 #endif |
3180 if(levelFix) | 3122 if(levelFix){ |
3181 { | |
3182 #ifdef HAVE_MMX | 3123 #ifdef HAVE_MMX |
3183 asm volatile( | 3124 asm volatile( |
3184 "movq (%%"REG_a"), %%mm2 \n\t" // packedYOffset | 3125 "movq (%%"REG_a"), %%mm2 \n\t" // packedYOffset |
3185 "movq 8(%%"REG_a"), %%mm3 \n\t" // packedYScale | 3126 "movq 8(%%"REG_a"), %%mm3 \n\t" // packedYScale |
3186 "lea (%2,%4), %%"REG_a" \n\t" | 3127 "lea (%2,%4), %%"REG_a" \n\t" |
3187 "lea (%3,%5), %%"REG_d" \n\t" | 3128 "lea (%3,%5), %%"REG_d" \n\t" |
3188 "pxor %%mm4, %%mm4 \n\t" | 3129 "pxor %%mm4, %%mm4 \n\t" |
3189 #ifdef HAVE_MMX2 | 3130 #ifdef HAVE_MMX2 |
3190 #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \ | 3131 #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \ |
3191 "movq " #src1 ", %%mm0 \n\t"\ | 3132 "movq " #src1 ", %%mm0 \n\t"\ |
3192 "movq " #src1 ", %%mm5 \n\t"\ | 3133 "movq " #src1 ", %%mm5 \n\t"\ |
3193 "movq " #src2 ", %%mm1 \n\t"\ | 3134 "movq " #src2 ", %%mm1 \n\t"\ |
3194 "movq " #src2 ", %%mm6 \n\t"\ | 3135 "movq " #src2 ", %%mm6 \n\t"\ |
3195 "punpcklbw %%mm0, %%mm0 \n\t"\ | 3136 "punpcklbw %%mm0, %%mm0 \n\t"\ |
3196 "punpckhbw %%mm5, %%mm5 \n\t"\ | 3137 "punpckhbw %%mm5, %%mm5 \n\t"\ |
3197 "punpcklbw %%mm1, %%mm1 \n\t"\ | 3138 "punpcklbw %%mm1, %%mm1 \n\t"\ |
3198 "punpckhbw %%mm6, %%mm6 \n\t"\ | 3139 "punpckhbw %%mm6, %%mm6 \n\t"\ |
3199 "pmulhuw %%mm3, %%mm0 \n\t"\ | 3140 "pmulhuw %%mm3, %%mm0 \n\t"\ |
3200 "pmulhuw %%mm3, %%mm5 \n\t"\ | 3141 "pmulhuw %%mm3, %%mm5 \n\t"\ |
3201 "pmulhuw %%mm3, %%mm1 \n\t"\ | 3142 "pmulhuw %%mm3, %%mm1 \n\t"\ |
3202 "pmulhuw %%mm3, %%mm6 \n\t"\ | 3143 "pmulhuw %%mm3, %%mm6 \n\t"\ |
3203 "psubw %%mm2, %%mm0 \n\t"\ | 3144 "psubw %%mm2, %%mm0 \n\t"\ |
3204 "psubw %%mm2, %%mm5 \n\t"\ | 3145 "psubw %%mm2, %%mm5 \n\t"\ |
3205 "psubw %%mm2, %%mm1 \n\t"\ | 3146 "psubw %%mm2, %%mm1 \n\t"\ |
3206 "psubw %%mm2, %%mm6 \n\t"\ | 3147 "psubw %%mm2, %%mm6 \n\t"\ |
3207 "packuswb %%mm5, %%mm0 \n\t"\ | 3148 "packuswb %%mm5, %%mm0 \n\t"\ |
3208 "packuswb %%mm6, %%mm1 \n\t"\ | 3149 "packuswb %%mm6, %%mm1 \n\t"\ |
3209 "movq %%mm0, " #dst1 " \n\t"\ | 3150 "movq %%mm0, " #dst1 " \n\t"\ |
3210 "movq %%mm1, " #dst2 " \n\t"\ | 3151 "movq %%mm1, " #dst2 " \n\t"\ |
3211 | 3152 |
3212 #else //HAVE_MMX2 | 3153 #else //HAVE_MMX2 |
3213 #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \ | 3154 #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \ |
3214 "movq " #src1 ", %%mm0 \n\t"\ | 3155 "movq " #src1 ", %%mm0 \n\t"\ |
3215 "movq " #src1 ", %%mm5 \n\t"\ | 3156 "movq " #src1 ", %%mm5 \n\t"\ |
3216 "punpcklbw %%mm4, %%mm0 \n\t"\ | 3157 "punpcklbw %%mm4, %%mm0 \n\t"\ |
3217 "punpckhbw %%mm4, %%mm5 \n\t"\ | 3158 "punpckhbw %%mm4, %%mm5 \n\t"\ |
3218 "psubw %%mm2, %%mm0 \n\t"\ | 3159 "psubw %%mm2, %%mm0 \n\t"\ |
3219 "psubw %%mm2, %%mm5 \n\t"\ | 3160 "psubw %%mm2, %%mm5 \n\t"\ |
3220 "movq " #src2 ", %%mm1 \n\t"\ | 3161 "movq " #src2 ", %%mm1 \n\t"\ |
3221 "psllw $6, %%mm0 \n\t"\ | 3162 "psllw $6, %%mm0 \n\t"\ |
3222 "psllw $6, %%mm5 \n\t"\ | 3163 "psllw $6, %%mm5 \n\t"\ |
3223 "pmulhw %%mm3, %%mm0 \n\t"\ | 3164 "pmulhw %%mm3, %%mm0 \n\t"\ |
3224 "movq " #src2 ", %%mm6 \n\t"\ | 3165 "movq " #src2 ", %%mm6 \n\t"\ |
3225 "pmulhw %%mm3, %%mm5 \n\t"\ | 3166 "pmulhw %%mm3, %%mm5 \n\t"\ |
3226 "punpcklbw %%mm4, %%mm1 \n\t"\ | 3167 "punpcklbw %%mm4, %%mm1 \n\t"\ |
3227 "punpckhbw %%mm4, %%mm6 \n\t"\ | 3168 "punpckhbw %%mm4, %%mm6 \n\t"\ |
3228 "psubw %%mm2, %%mm1 \n\t"\ | 3169 "psubw %%mm2, %%mm1 \n\t"\ |
3229 "psubw %%mm2, %%mm6 \n\t"\ | 3170 "psubw %%mm2, %%mm6 \n\t"\ |
3230 "psllw $6, %%mm1 \n\t"\ | 3171 "psllw $6, %%mm1 \n\t"\ |
3231 "psllw $6, %%mm6 \n\t"\ | 3172 "psllw $6, %%mm6 \n\t"\ |
3232 "pmulhw %%mm3, %%mm1 \n\t"\ | 3173 "pmulhw %%mm3, %%mm1 \n\t"\ |
3233 "pmulhw %%mm3, %%mm6 \n\t"\ | 3174 "pmulhw %%mm3, %%mm6 \n\t"\ |
3234 "packuswb %%mm5, %%mm0 \n\t"\ | 3175 "packuswb %%mm5, %%mm0 \n\t"\ |
3235 "packuswb %%mm6, %%mm1 \n\t"\ | 3176 "packuswb %%mm6, %%mm1 \n\t"\ |
3236 "movq %%mm0, " #dst1 " \n\t"\ | 3177 "movq %%mm0, " #dst1 " \n\t"\ |
3237 "movq %%mm1, " #dst2 " \n\t"\ | 3178 "movq %%mm1, " #dst2 " \n\t"\ |
3238 | 3179 |
3239 #endif //HAVE_MMX2 | 3180 #endif //HAVE_MMX2 |
3240 #define SCALED_CPY(src1, src2, dst1, dst2)\ | 3181 #define SCALED_CPY(src1, src2, dst1, dst2)\ |
3241 REAL_SCALED_CPY(src1, src2, dst1, dst2) | 3182 REAL_SCALED_CPY(src1, src2, dst1, dst2) |
3242 | 3183 |
3243 SCALED_CPY((%2) , (%2, %4) , (%3) , (%3, %5)) | 3184 SCALED_CPY((%2) , (%2, %4) , (%3) , (%3, %5)) |
3244 SCALED_CPY((%2, %4, 2), (%%REGa, %4, 2), (%3, %5, 2), (%%REGd, %5, 2)) | 3185 SCALED_CPY((%2, %4, 2), (%%REGa, %4, 2), (%3, %5, 2), (%%REGd, %5, 2)) |
3245 SCALED_CPY((%2, %4, 4), (%%REGa, %4, 4), (%3, %5, 4), (%%REGd, %5, 4)) | 3186 SCALED_CPY((%2, %4, 4), (%%REGa, %4, 4), (%3, %5, 4), (%%REGd, %5, 4)) |
3246 "lea (%%"REG_a",%4,4), %%"REG_a" \n\t" | 3187 "lea (%%"REG_a",%4,4), %%"REG_a" \n\t" |
3247 "lea (%%"REG_d",%5,4), %%"REG_d" \n\t" | 3188 "lea (%%"REG_d",%5,4), %%"REG_d" \n\t" |
3248 SCALED_CPY((%%REGa, %4), (%%REGa, %4, 2), (%%REGd, %5), (%%REGd, %5, 2)) | 3189 SCALED_CPY((%%REGa, %4), (%%REGa, %4, 2), (%%REGd, %5), (%%REGd, %5, 2)) |
3249 | 3190 |
3250 | 3191 |
3251 : "=&a" (packedOffsetAndScale) | 3192 : "=&a" (packedOffsetAndScale) |
3252 : "0" (packedOffsetAndScale), | 3193 : "0" (packedOffsetAndScale), |
3253 "r"(src), | 3194 "r"(src), |
3254 "r"(dst), | 3195 "r"(dst), |
3255 "r" ((long)srcStride), | 3196 "r" ((long)srcStride), |
3256 "r" ((long)dstStride) | 3197 "r" ((long)dstStride) |
3257 : "%"REG_d | 3198 : "%"REG_d |
3258 ); | 3199 ); |
3259 #else //HAVE_MMX | 3200 #else //HAVE_MMX |
3260 for(i=0; i<8; i++) | 3201 for(i=0; i<8; i++) |
3261 memcpy( &(dst[dstStride*i]), | 3202 memcpy( &(dst[dstStride*i]), |
3262 &(src[srcStride*i]), BLOCK_SIZE); | 3203 &(src[srcStride*i]), BLOCK_SIZE); |
3263 #endif //HAVE_MMX | 3204 #endif //HAVE_MMX |
3264 } | 3205 }else{ |
3265 else | |
3266 { | |
3267 #ifdef HAVE_MMX | 3206 #ifdef HAVE_MMX |
3268 asm volatile( | 3207 asm volatile( |
3269 "lea (%0,%2), %%"REG_a" \n\t" | 3208 "lea (%0,%2), %%"REG_a" \n\t" |
3270 "lea (%1,%3), %%"REG_d" \n\t" | 3209 "lea (%1,%3), %%"REG_d" \n\t" |
3271 | 3210 |
3272 #define REAL_SIMPLE_CPY(src1, src2, dst1, dst2) \ | 3211 #define REAL_SIMPLE_CPY(src1, src2, dst1, dst2) \ |
3273 "movq " #src1 ", %%mm0 \n\t"\ | 3212 "movq " #src1 ", %%mm0 \n\t"\ |
3274 "movq " #src2 ", %%mm1 \n\t"\ | 3213 "movq " #src2 ", %%mm1 \n\t"\ |
3275 "movq %%mm0, " #dst1 " \n\t"\ | 3214 "movq %%mm0, " #dst1 " \n\t"\ |
3276 "movq %%mm1, " #dst2 " \n\t"\ | 3215 "movq %%mm1, " #dst2 " \n\t"\ |
3277 | 3216 |
3278 #define SIMPLE_CPY(src1, src2, dst1, dst2)\ | 3217 #define SIMPLE_CPY(src1, src2, dst1, dst2)\ |
3279 REAL_SIMPLE_CPY(src1, src2, dst1, dst2) | 3218 REAL_SIMPLE_CPY(src1, src2, dst1, dst2) |
3280 | 3219 |
3281 SIMPLE_CPY((%0) , (%0, %2) , (%1) , (%1, %3)) | 3220 SIMPLE_CPY((%0) , (%0, %2) , (%1) , (%1, %3)) |
3282 SIMPLE_CPY((%0, %2, 2), (%%REGa, %2, 2), (%1, %3, 2), (%%REGd, %3, 2)) | 3221 SIMPLE_CPY((%0, %2, 2), (%%REGa, %2, 2), (%1, %3, 2), (%%REGd, %3, 2)) |
3283 SIMPLE_CPY((%0, %2, 4), (%%REGa, %2, 4), (%1, %3, 4), (%%REGd, %3, 4)) | 3222 SIMPLE_CPY((%0, %2, 4), (%%REGa, %2, 4), (%1, %3, 4), (%%REGd, %3, 4)) |
3284 "lea (%%"REG_a",%2,4), %%"REG_a" \n\t" | 3223 "lea (%%"REG_a",%2,4), %%"REG_a" \n\t" |
3285 "lea (%%"REG_d",%3,4), %%"REG_d" \n\t" | 3224 "lea (%%"REG_d",%3,4), %%"REG_d" \n\t" |
3286 SIMPLE_CPY((%%REGa, %2), (%%REGa, %2, 2), (%%REGd, %3), (%%REGd, %3, 2)) | 3225 SIMPLE_CPY((%%REGa, %2), (%%REGa, %2, 2), (%%REGd, %3), (%%REGd, %3, 2)) |
3287 | 3226 |
3288 : : "r" (src), | 3227 : : "r" (src), |
3289 "r" (dst), | 3228 "r" (dst), |
3290 "r" ((long)srcStride), | 3229 "r" ((long)srcStride), |
3291 "r" ((long)dstStride) | 3230 "r" ((long)dstStride) |
3292 : "%"REG_a, "%"REG_d | 3231 : "%"REG_a, "%"REG_d |
3293 ); | 3232 ); |
3294 #else //HAVE_MMX | 3233 #else //HAVE_MMX |
3295 for(i=0; i<8; i++) | 3234 for(i=0; i<8; i++) |
3296 memcpy( &(dst[dstStride*i]), | 3235 memcpy( &(dst[dstStride*i]), |
3297 &(src[srcStride*i]), BLOCK_SIZE); | 3236 &(src[srcStride*i]), BLOCK_SIZE); |
3298 #endif //HAVE_MMX | 3237 #endif //HAVE_MMX |
3299 } | 3238 } |
3300 } | 3239 } |
3301 | 3240 |
3302 /** | 3241 /** |
3303 * Duplicates the given 8 src pixels ? times upward | 3242 * Duplicates the given 8 src pixels ? times upward |
3304 */ | 3243 */ |
3305 static inline void RENAME(duplicate)(uint8_t src[], int stride) | 3244 static inline void RENAME(duplicate)(uint8_t src[], int stride) |
3306 { | 3245 { |
3307 #ifdef HAVE_MMX | 3246 #ifdef HAVE_MMX |
3308 asm volatile( | 3247 asm volatile( |
3309 "movq (%0), %%mm0 \n\t" | 3248 "movq (%0), %%mm0 \n\t" |
3310 "add %1, %0 \n\t" | 3249 "add %1, %0 \n\t" |
3311 "movq %%mm0, (%0) \n\t" | 3250 "movq %%mm0, (%0) \n\t" |
3312 "movq %%mm0, (%0, %1) \n\t" | 3251 "movq %%mm0, (%0, %1) \n\t" |
3313 "movq %%mm0, (%0, %1, 2) \n\t" | 3252 "movq %%mm0, (%0, %1, 2) \n\t" |
3314 : "+r" (src) | 3253 : "+r" (src) |
3315 : "r" ((long)-stride) | 3254 : "r" ((long)-stride) |
3316 ); | 3255 ); |
3317 #else | 3256 #else |
3318 int i; | 3257 int i; |
3319 uint8_t *p=src; | 3258 uint8_t *p=src; |
3320 for(i=0; i<3; i++) | 3259 for(i=0; i<3; i++){ |
3321 { | 3260 p-= stride; |
3322 p-= stride; | 3261 memcpy(p, src, 8); |
3323 memcpy(p, src, 8); | 3262 } |
3324 } | |
3325 #endif | 3263 #endif |
3326 } | 3264 } |
3327 | 3265 |
3328 /** | 3266 /** |
3329 * Filters array of bytes (Y or U or V values) | 3267 * Filters array of bytes (Y or U or V values) |
3330 */ | 3268 */ |
3331 static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, | 3269 static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
3332 const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2) | 3270 const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2) |
3333 { | 3271 { |
3334 DECLARE_ALIGNED(8, PPContext, c)= *c2; //copy to stack for faster access | 3272 DECLARE_ALIGNED(8, PPContext, c)= *c2; //copy to stack for faster access |
3335 int x,y; | 3273 int x,y; |
3336 #ifdef COMPILE_TIME_MODE | 3274 #ifdef COMPILE_TIME_MODE |
3337 const int mode= COMPILE_TIME_MODE; | 3275 const int mode= COMPILE_TIME_MODE; |
3338 #else | 3276 #else |
3339 const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode; | 3277 const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode; |
3340 #endif | 3278 #endif |
3341 int black=0, white=255; // blackest black and whitest white in the picture | 3279 int black=0, white=255; // blackest black and whitest white in the picture |
3342 int QPCorrecture= 256*256; | 3280 int QPCorrecture= 256*256; |
3343 | 3281 |
3344 int copyAhead; | 3282 int copyAhead; |
3345 #ifdef HAVE_MMX | 3283 #ifdef HAVE_MMX |
3284 int i; | |
3285 #endif | |
3286 | |
3287 const int qpHShift= isColor ? 4-c.hChromaSubSample : 4; | |
3288 const int qpVShift= isColor ? 4-c.vChromaSubSample : 4; | |
3289 | |
3290 //FIXME remove | |
3291 uint64_t * const yHistogram= c.yHistogram; | |
3292 uint8_t * const tempSrc= srcStride > 0 ? c.tempSrc : c.tempSrc - 23*srcStride; | |
3293 uint8_t * const tempDst= dstStride > 0 ? c.tempDst : c.tempDst - 23*dstStride; | |
3294 //const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4; | |
3295 | |
3296 #ifdef HAVE_MMX | |
3297 for(i=0; i<57; i++){ | |
3298 int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1; | |
3299 int threshold= offset*2 + 1; | |
3300 c.mmxDcOffset[i]= 0x7F - offset; | |
3301 c.mmxDcThreshold[i]= 0x7F - threshold; | |
3302 c.mmxDcOffset[i]*= 0x0101010101010101LL; | |
3303 c.mmxDcThreshold[i]*= 0x0101010101010101LL; | |
3304 } | |
3305 #endif | |
3306 | |
3307 if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16; | |
3308 else if( (mode & LINEAR_BLEND_DEINT_FILTER) | |
3309 || (mode & FFMPEG_DEINT_FILTER) | |
3310 || (mode & LOWPASS5_DEINT_FILTER)) copyAhead=14; | |
3311 else if( (mode & V_DEBLOCK) | |
3312 || (mode & LINEAR_IPOL_DEINT_FILTER) | |
3313 || (mode & MEDIAN_DEINT_FILTER) | |
3314 || (mode & V_A_DEBLOCK)) copyAhead=13; | |
3315 else if(mode & V_X1_FILTER) copyAhead=11; | |
3316 // else if(mode & V_RK1_FILTER) copyAhead=10; | |
3317 else if(mode & DERING) copyAhead=9; | |
3318 else copyAhead=8; | |
3319 | |
3320 copyAhead-= 8; | |
3321 | |
3322 if(!isColor){ | |
3323 uint64_t sum= 0; | |
3346 int i; | 3324 int i; |
3325 uint64_t maxClipped; | |
3326 uint64_t clipped; | |
3327 double scale; | |
3328 | |
3329 c.frameNum++; | |
3330 // first frame is fscked so we ignore it | |
3331 if(c.frameNum == 1) yHistogram[0]= width*height/64*15/256; | |
3332 | |
3333 for(i=0; i<256; i++){ | |
3334 sum+= yHistogram[i]; | |
3335 } | |
3336 | |
3337 /* We always get a completely black picture first. */ | |
3338 maxClipped= (uint64_t)(sum * c.ppMode.maxClippedThreshold); | |
3339 | |
3340 clipped= sum; | |
3341 for(black=255; black>0; black--){ | |
3342 if(clipped < maxClipped) break; | |
3343 clipped-= yHistogram[black]; | |
3344 } | |
3345 | |
3346 clipped= sum; | |
3347 for(white=0; white<256; white++){ | |
3348 if(clipped < maxClipped) break; | |
3349 clipped-= yHistogram[white]; | |
3350 } | |
3351 | |
3352 scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black); | |
3353 | |
3354 #ifdef HAVE_MMX2 | |
3355 c.packedYScale= (uint16_t)(scale*256.0 + 0.5); | |
3356 c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF; | |
3357 #else | |
3358 c.packedYScale= (uint16_t)(scale*1024.0 + 0.5); | |
3359 c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF; | |
3347 #endif | 3360 #endif |
3348 | 3361 |
3349 const int qpHShift= isColor ? 4-c.hChromaSubSample : 4; | 3362 c.packedYOffset|= c.packedYOffset<<32; |
3350 const int qpVShift= isColor ? 4-c.vChromaSubSample : 4; | 3363 c.packedYOffset|= c.packedYOffset<<16; |
3351 | 3364 |
3352 //FIXME remove | 3365 c.packedYScale|= c.packedYScale<<32; |
3353 uint64_t * const yHistogram= c.yHistogram; | 3366 c.packedYScale|= c.packedYScale<<16; |
3354 uint8_t * const tempSrc= srcStride > 0 ? c.tempSrc : c.tempSrc - 23*srcStride; | 3367 |
3355 uint8_t * const tempDst= dstStride > 0 ? c.tempDst : c.tempDst - 23*dstStride; | 3368 if(mode & LEVEL_FIX) QPCorrecture= (int)(scale*256*256 + 0.5); |
3356 //const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4; | 3369 else QPCorrecture= 256*256; |
3357 | 3370 }else{ |
3358 #ifdef HAVE_MMX | 3371 c.packedYScale= 0x0100010001000100LL; |
3359 for(i=0; i<57; i++){ | 3372 c.packedYOffset= 0; |
3360 int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1; | 3373 QPCorrecture= 256*256; |
3361 int threshold= offset*2 + 1; | 3374 } |
3362 c.mmxDcOffset[i]= 0x7F - offset; | 3375 |
3363 c.mmxDcThreshold[i]= 0x7F - threshold; | 3376 /* copy & deinterlace first row of blocks */ |
3364 c.mmxDcOffset[i]*= 0x0101010101010101LL; | 3377 y=-BLOCK_SIZE; |
3365 c.mmxDcThreshold[i]*= 0x0101010101010101LL; | 3378 { |
3366 } | 3379 const uint8_t *srcBlock= &(src[y*srcStride]); |
3367 #endif | 3380 uint8_t *dstBlock= tempDst + dstStride; |
3368 | 3381 |
3369 if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16; | 3382 // From this point on it is guaranteed that we can read and write 16 lines downward |
3370 else if( (mode & LINEAR_BLEND_DEINT_FILTER) | 3383 // finish 1 block before the next otherwise we might have a problem |
3371 || (mode & FFMPEG_DEINT_FILTER) | 3384 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing |
3372 || (mode & LOWPASS5_DEINT_FILTER)) copyAhead=14; | 3385 for(x=0; x<width; x+=BLOCK_SIZE){ |
3373 else if( (mode & V_DEBLOCK) | |
3374 || (mode & LINEAR_IPOL_DEINT_FILTER) | |
3375 || (mode & MEDIAN_DEINT_FILTER) | |
3376 || (mode & V_A_DEBLOCK)) copyAhead=13; | |
3377 else if(mode & V_X1_FILTER) copyAhead=11; | |
3378 // else if(mode & V_RK1_FILTER) copyAhead=10; | |
3379 else if(mode & DERING) copyAhead=9; | |
3380 else copyAhead=8; | |
3381 | |
3382 copyAhead-= 8; | |
3383 | |
3384 if(!isColor) | |
3385 { | |
3386 uint64_t sum= 0; | |
3387 int i; | |
3388 uint64_t maxClipped; | |
3389 uint64_t clipped; | |
3390 double scale; | |
3391 | |
3392 c.frameNum++; | |
3393 // first frame is fscked so we ignore it | |
3394 if(c.frameNum == 1) yHistogram[0]= width*height/64*15/256; | |
3395 | |
3396 for(i=0; i<256; i++) | |
3397 { | |
3398 sum+= yHistogram[i]; | |
3399 } | |
3400 | |
3401 /* We always get a completely black picture first. */ | |
3402 maxClipped= (uint64_t)(sum * c.ppMode.maxClippedThreshold); | |
3403 | |
3404 clipped= sum; | |
3405 for(black=255; black>0; black--) | |
3406 { | |
3407 if(clipped < maxClipped) break; | |
3408 clipped-= yHistogram[black]; | |
3409 } | |
3410 | |
3411 clipped= sum; | |
3412 for(white=0; white<256; white++) | |
3413 { | |
3414 if(clipped < maxClipped) break; | |
3415 clipped-= yHistogram[white]; | |
3416 } | |
3417 | |
3418 scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black); | |
3419 | |
3420 #ifdef HAVE_MMX2 | |
3421 c.packedYScale= (uint16_t)(scale*256.0 + 0.5); | |
3422 c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF; | |
3423 #else | |
3424 c.packedYScale= (uint16_t)(scale*1024.0 + 0.5); | |
3425 c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF; | |
3426 #endif | |
3427 | |
3428 c.packedYOffset|= c.packedYOffset<<32; | |
3429 c.packedYOffset|= c.packedYOffset<<16; | |
3430 | |
3431 c.packedYScale|= c.packedYScale<<32; | |
3432 c.packedYScale|= c.packedYScale<<16; | |
3433 | |
3434 if(mode & LEVEL_FIX) QPCorrecture= (int)(scale*256*256 + 0.5); | |
3435 else QPCorrecture= 256*256; | |
3436 } | |
3437 else | |
3438 { | |
3439 c.packedYScale= 0x0100010001000100LL; | |
3440 c.packedYOffset= 0; | |
3441 QPCorrecture= 256*256; | |
3442 } | |
3443 | |
3444 /* copy & deinterlace first row of blocks */ | |
3445 y=-BLOCK_SIZE; | |
3446 { | |
3447 const uint8_t *srcBlock= &(src[y*srcStride]); | |
3448 uint8_t *dstBlock= tempDst + dstStride; | |
3449 | |
3450 // From this point on it is guaranteed that we can read and write 16 lines downward | |
3451 // finish 1 block before the next otherwise we might have a problem | |
3452 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing | |
3453 for(x=0; x<width; x+=BLOCK_SIZE) | |
3454 { | |
3455 | 3386 |
3456 #ifdef HAVE_MMX2 | 3387 #ifdef HAVE_MMX2 |
3457 /* | 3388 /* |
3458 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); | 3389 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); |
3459 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); | 3390 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); |
3460 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); | 3391 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); |
3461 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); | 3392 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); |
3462 */ | 3393 */ |
3463 | 3394 |
3464 asm( | 3395 asm( |
3465 "mov %4, %%"REG_a" \n\t" | 3396 "mov %4, %%"REG_a" \n\t" |
3466 "shr $2, %%"REG_a" \n\t" | 3397 "shr $2, %%"REG_a" \n\t" |
3467 "and $6, %%"REG_a" \n\t" | 3398 "and $6, %%"REG_a" \n\t" |
3468 "add %5, %%"REG_a" \n\t" | 3399 "add %5, %%"REG_a" \n\t" |
3469 "mov %%"REG_a", %%"REG_d" \n\t" | 3400 "mov %%"REG_a", %%"REG_d" \n\t" |
3470 "imul %1, %%"REG_a" \n\t" | 3401 "imul %1, %%"REG_a" \n\t" |
3471 "imul %3, %%"REG_d" \n\t" | 3402 "imul %3, %%"REG_d" \n\t" |
3472 "prefetchnta 32(%%"REG_a", %0) \n\t" | 3403 "prefetchnta 32(%%"REG_a", %0) \n\t" |
3473 "prefetcht0 32(%%"REG_d", %2) \n\t" | 3404 "prefetcht0 32(%%"REG_d", %2) \n\t" |
3474 "add %1, %%"REG_a" \n\t" | 3405 "add %1, %%"REG_a" \n\t" |
3475 "add %3, %%"REG_d" \n\t" | 3406 "add %3, %%"REG_d" \n\t" |
3476 "prefetchnta 32(%%"REG_a", %0) \n\t" | 3407 "prefetchnta 32(%%"REG_a", %0) \n\t" |
3477 "prefetcht0 32(%%"REG_d", %2) \n\t" | 3408 "prefetcht0 32(%%"REG_d", %2) \n\t" |
3478 :: "r" (srcBlock), "r" ((long)srcStride), "r" (dstBlock), "r" ((long)dstStride), | 3409 :: "r" (srcBlock), "r" ((long)srcStride), "r" (dstBlock), "r" ((long)dstStride), |
3479 "g" ((long)x), "g" ((long)copyAhead) | 3410 "g" ((long)x), "g" ((long)copyAhead) |
3480 : "%"REG_a, "%"REG_d | 3411 : "%"REG_a, "%"REG_d |
3481 ); | 3412 ); |
3482 | 3413 |
3483 #elif defined(HAVE_3DNOW) | 3414 #elif defined(HAVE_3DNOW) |
3484 //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ... | 3415 //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ... |
3485 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); | 3416 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
3486 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | 3417 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); |
3487 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | 3418 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); |
3488 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | 3419 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); |
3489 */ | 3420 */ |
3490 #endif | 3421 #endif |
3491 | 3422 |
3492 RENAME(blockCopy)(dstBlock + dstStride*8, dstStride, | 3423 RENAME(blockCopy)(dstBlock + dstStride*8, dstStride, |
3493 srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset); | 3424 srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset); |
3494 | 3425 |
3495 RENAME(duplicate)(dstBlock + dstStride*8, dstStride); | 3426 RENAME(duplicate)(dstBlock + dstStride*8, dstStride); |
3496 | 3427 |
3497 if(mode & LINEAR_IPOL_DEINT_FILTER) | 3428 if(mode & LINEAR_IPOL_DEINT_FILTER) |
3498 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride); | 3429 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride); |
3499 else if(mode & LINEAR_BLEND_DEINT_FILTER) | 3430 else if(mode & LINEAR_BLEND_DEINT_FILTER) |
3500 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x); | 3431 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x); |
3501 else if(mode & MEDIAN_DEINT_FILTER) | 3432 else if(mode & MEDIAN_DEINT_FILTER) |
3502 RENAME(deInterlaceMedian)(dstBlock, dstStride); | 3433 RENAME(deInterlaceMedian)(dstBlock, dstStride); |
3503 else if(mode & CUBIC_IPOL_DEINT_FILTER) | 3434 else if(mode & CUBIC_IPOL_DEINT_FILTER) |
3504 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride); | 3435 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride); |
3505 else if(mode & FFMPEG_DEINT_FILTER) | 3436 else if(mode & FFMPEG_DEINT_FILTER) |
3506 RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x); | 3437 RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x); |
3507 else if(mode & LOWPASS5_DEINT_FILTER) | 3438 else if(mode & LOWPASS5_DEINT_FILTER) |
3508 RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x); | 3439 RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x); |
3509 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | 3440 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) |
3510 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride); | 3441 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride); |
3511 */ | 3442 */ |
3512 dstBlock+=8; | 3443 dstBlock+=8; |
3513 srcBlock+=8; | 3444 srcBlock+=8; |
3514 } | |
3515 if(width==FFABS(dstStride)) | |
3516 linecpy(dst, tempDst + 9*dstStride, copyAhead, dstStride); | |
3517 else | |
3518 { | |
3519 int i; | |
3520 for(i=0; i<copyAhead; i++) | |
3521 { | |
3522 memcpy(dst + i*dstStride, tempDst + (9+i)*dstStride, width); | |
3523 } | |
3524 } | |
3525 } | 3445 } |
3526 | 3446 if(width==FFABS(dstStride)) |
3527 for(y=0; y<height; y+=BLOCK_SIZE) | 3447 linecpy(dst, tempDst + 9*dstStride, copyAhead, dstStride); |
3528 { | 3448 else{ |
3529 //1% speedup if these are here instead of the inner loop | 3449 int i; |
3530 const uint8_t *srcBlock= &(src[y*srcStride]); | 3450 for(i=0; i<copyAhead; i++){ |
3531 uint8_t *dstBlock= &(dst[y*dstStride]); | 3451 memcpy(dst + i*dstStride, tempDst + (9+i)*dstStride, width); |
3452 } | |
3453 } | |
3454 } | |
3455 | |
3456 for(y=0; y<height; y+=BLOCK_SIZE){ | |
3457 //1% speedup if these are here instead of the inner loop | |
3458 const uint8_t *srcBlock= &(src[y*srcStride]); | |
3459 uint8_t *dstBlock= &(dst[y*dstStride]); | |
3532 #ifdef HAVE_MMX | 3460 #ifdef HAVE_MMX |
3533 uint8_t *tempBlock1= c.tempBlocks; | 3461 uint8_t *tempBlock1= c.tempBlocks; |
3534 uint8_t *tempBlock2= c.tempBlocks + 8; | 3462 uint8_t *tempBlock2= c.tempBlocks + 8; |
3535 #endif | 3463 #endif |
3536 const int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride]; | 3464 const int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride]; |
3537 int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*FFABS(QPStride)]; | 3465 int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*FFABS(QPStride)]; |
3538 int QP=0; | 3466 int QP=0; |
3539 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards | 3467 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards |
3540 if not than use a temporary buffer */ | 3468 if not than use a temporary buffer */ |
3541 if(y+15 >= height) | 3469 if(y+15 >= height){ |
3542 { | 3470 int i; |
3543 int i; | 3471 /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with |
3544 /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with | 3472 blockcopy to dst later */ |
3545 blockcopy to dst later */ | 3473 linecpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead, |
3546 linecpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead, | 3474 FFMAX(height-y-copyAhead, 0), srcStride); |
3547 FFMAX(height-y-copyAhead, 0), srcStride); | 3475 |
3548 | 3476 /* duplicate last line of src to fill the void upto line (copyAhead+7) */ |
3549 /* duplicate last line of src to fill the void upto line (copyAhead+7) */ | 3477 for(i=FFMAX(height-y, 8); i<copyAhead+8; i++) |
3550 for(i=FFMAX(height-y, 8); i<copyAhead+8; i++) | 3478 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), FFABS(srcStride)); |
3551 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), FFABS(srcStride)); | 3479 |
3552 | 3480 /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/ |
3553 /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/ | 3481 linecpy(tempDst, dstBlock - dstStride, FFMIN(height-y+1, copyAhead+1), dstStride); |
3554 linecpy(tempDst, dstBlock - dstStride, FFMIN(height-y+1, copyAhead+1), dstStride); | 3482 |
3555 | 3483 /* duplicate last line of dst to fill the void upto line (copyAhead) */ |
3556 /* duplicate last line of dst to fill the void upto line (copyAhead) */ | 3484 for(i=height-y+1; i<=copyAhead; i++) |
3557 for(i=height-y+1; i<=copyAhead; i++) | 3485 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), FFABS(dstStride)); |
3558 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), FFABS(dstStride)); | 3486 |
3559 | 3487 dstBlock= tempDst + dstStride; |
3560 dstBlock= tempDst + dstStride; | 3488 srcBlock= tempSrc; |
3561 srcBlock= tempSrc; | 3489 } |
3562 } | 3490 |
3563 | 3491 // From this point on it is guaranteed that we can read and write 16 lines downward |
3564 // From this point on it is guaranteed that we can read and write 16 lines downward | 3492 // finish 1 block before the next otherwise we might have a problem |
3565 // finish 1 block before the next otherwise we might have a problem | 3493 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing |
3566 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing | 3494 for(x=0; x<width; x+=BLOCK_SIZE){ |
3567 for(x=0; x<width; x+=BLOCK_SIZE) | 3495 const int stride= dstStride; |
3568 { | |
3569 const int stride= dstStride; | |
3570 #ifdef HAVE_MMX | 3496 #ifdef HAVE_MMX |
3571 uint8_t *tmpXchg; | 3497 uint8_t *tmpXchg; |
3572 #endif | 3498 #endif |
3573 if(isColor) | 3499 if(isColor){ |
3574 { | 3500 QP= QPptr[x>>qpHShift]; |
3575 QP= QPptr[x>>qpHShift]; | 3501 c.nonBQP= nonBQPptr[x>>qpHShift]; |
3576 c.nonBQP= nonBQPptr[x>>qpHShift]; | 3502 }else{ |
3577 } | 3503 QP= QPptr[x>>4]; |
3578 else | 3504 QP= (QP* QPCorrecture + 256*128)>>16; |
3579 { | 3505 c.nonBQP= nonBQPptr[x>>4]; |
3580 QP= QPptr[x>>4]; | 3506 c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16; |
3581 QP= (QP* QPCorrecture + 256*128)>>16; | 3507 yHistogram[ srcBlock[srcStride*12 + 4] ]++; |
3582 c.nonBQP= nonBQPptr[x>>4]; | 3508 } |
3583 c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16; | 3509 c.QP= QP; |
3584 yHistogram[ srcBlock[srcStride*12 + 4] ]++; | |
3585 } | |
3586 c.QP= QP; | |
3587 #ifdef HAVE_MMX | 3510 #ifdef HAVE_MMX |
3588 asm volatile( | 3511 asm volatile( |
3589 "movd %1, %%mm7 \n\t" | 3512 "movd %1, %%mm7 \n\t" |
3590 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP | 3513 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP |
3591 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP | 3514 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP |
3592 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP | 3515 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP |
3593 "movq %%mm7, %0 \n\t" | 3516 "movq %%mm7, %0 \n\t" |
3594 : "=m" (c.pQPb) | 3517 : "=m" (c.pQPb) |
3595 : "r" (QP) | 3518 : "r" (QP) |
3596 ); | 3519 ); |
3597 #endif | 3520 #endif |
3598 | 3521 |
3599 | 3522 |
3600 #ifdef HAVE_MMX2 | 3523 #ifdef HAVE_MMX2 |
3601 /* | 3524 /* |
3602 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); | 3525 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); |
3603 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); | 3526 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); |
3604 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); | 3527 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); |
3605 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); | 3528 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); |
3606 */ | 3529 */ |
3607 | 3530 |
3608 asm( | 3531 asm( |
3609 "mov %4, %%"REG_a" \n\t" | 3532 "mov %4, %%"REG_a" \n\t" |
3610 "shr $2, %%"REG_a" \n\t" | 3533 "shr $2, %%"REG_a" \n\t" |
3611 "and $6, %%"REG_a" \n\t" | 3534 "and $6, %%"REG_a" \n\t" |
3612 "add %5, %%"REG_a" \n\t" | 3535 "add %5, %%"REG_a" \n\t" |
3613 "mov %%"REG_a", %%"REG_d" \n\t" | 3536 "mov %%"REG_a", %%"REG_d" \n\t" |
3614 "imul %1, %%"REG_a" \n\t" | 3537 "imul %1, %%"REG_a" \n\t" |
3615 "imul %3, %%"REG_d" \n\t" | 3538 "imul %3, %%"REG_d" \n\t" |
3616 "prefetchnta 32(%%"REG_a", %0) \n\t" | 3539 "prefetchnta 32(%%"REG_a", %0) \n\t" |
3617 "prefetcht0 32(%%"REG_d", %2) \n\t" | 3540 "prefetcht0 32(%%"REG_d", %2) \n\t" |
3618 "add %1, %%"REG_a" \n\t" | 3541 "add %1, %%"REG_a" \n\t" |
3619 "add %3, %%"REG_d" \n\t" | 3542 "add %3, %%"REG_d" \n\t" |
3620 "prefetchnta 32(%%"REG_a", %0) \n\t" | 3543 "prefetchnta 32(%%"REG_a", %0) \n\t" |
3621 "prefetcht0 32(%%"REG_d", %2) \n\t" | 3544 "prefetcht0 32(%%"REG_d", %2) \n\t" |
3622 :: "r" (srcBlock), "r" ((long)srcStride), "r" (dstBlock), "r" ((long)dstStride), | 3545 :: "r" (srcBlock), "r" ((long)srcStride), "r" (dstBlock), "r" ((long)dstStride), |
3623 "g" ((long)x), "g" ((long)copyAhead) | 3546 "g" ((long)x), "g" ((long)copyAhead) |
3624 : "%"REG_a, "%"REG_d | 3547 : "%"REG_a, "%"REG_d |
3625 ); | 3548 ); |
3626 | 3549 |
3627 #elif defined(HAVE_3DNOW) | 3550 #elif defined(HAVE_3DNOW) |
3628 //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ... | 3551 //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ... |
3629 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); | 3552 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
3630 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | 3553 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); |
3631 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | 3554 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); |
3632 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | 3555 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); |
3633 */ | 3556 */ |
3634 #endif | 3557 #endif |
3635 | 3558 |
3636 RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride, | 3559 RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride, |
3637 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset); | 3560 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset); |
3638 | 3561 |
3639 if(mode & LINEAR_IPOL_DEINT_FILTER) | 3562 if(mode & LINEAR_IPOL_DEINT_FILTER) |
3640 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride); | 3563 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride); |
3641 else if(mode & LINEAR_BLEND_DEINT_FILTER) | 3564 else if(mode & LINEAR_BLEND_DEINT_FILTER) |
3642 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x); | 3565 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x); |
3643 else if(mode & MEDIAN_DEINT_FILTER) | 3566 else if(mode & MEDIAN_DEINT_FILTER) |
3644 RENAME(deInterlaceMedian)(dstBlock, dstStride); | 3567 RENAME(deInterlaceMedian)(dstBlock, dstStride); |
3645 else if(mode & CUBIC_IPOL_DEINT_FILTER) | 3568 else if(mode & CUBIC_IPOL_DEINT_FILTER) |
3646 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride); | 3569 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride); |
3647 else if(mode & FFMPEG_DEINT_FILTER) | 3570 else if(mode & FFMPEG_DEINT_FILTER) |
3648 RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x); | 3571 RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x); |
3649 else if(mode & LOWPASS5_DEINT_FILTER) | 3572 else if(mode & LOWPASS5_DEINT_FILTER) |
3650 RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x); | 3573 RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x); |
3651 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | 3574 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) |
3652 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride); | 3575 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride); |
3653 */ | 3576 */ |
3654 | 3577 |
3655 /* only deblock if we have 2 blocks */ | 3578 /* only deblock if we have 2 blocks */ |
3656 if(y + 8 < height) | 3579 if(y + 8 < height){ |
3657 { | 3580 if(mode & V_X1_FILTER) |
3658 if(mode & V_X1_FILTER) | 3581 RENAME(vertX1Filter)(dstBlock, stride, &c); |
3659 RENAME(vertX1Filter)(dstBlock, stride, &c); | 3582 else if(mode & V_DEBLOCK){ |
3660 else if(mode & V_DEBLOCK) | 3583 const int t= RENAME(vertClassify)(dstBlock, stride, &c); |
3661 { | 3584 |
3662 const int t= RENAME(vertClassify)(dstBlock, stride, &c); | 3585 if(t==1) |
3663 | 3586 RENAME(doVertLowPass)(dstBlock, stride, &c); |
3664 if(t==1) | 3587 else if(t==2) |
3665 RENAME(doVertLowPass)(dstBlock, stride, &c); | 3588 RENAME(doVertDefFilter)(dstBlock, stride, &c); |
3666 else if(t==2) | 3589 }else if(mode & V_A_DEBLOCK){ |
3667 RENAME(doVertDefFilter)(dstBlock, stride, &c); | 3590 RENAME(do_a_deblock)(dstBlock, stride, 1, &c); |
3668 }else if(mode & V_A_DEBLOCK){ | 3591 } |
3669 RENAME(do_a_deblock)(dstBlock, stride, 1, &c); | 3592 } |
3670 } | |
3671 } | |
3672 | 3593 |
3673 #ifdef HAVE_MMX | 3594 #ifdef HAVE_MMX |
3674 RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride); | 3595 RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride); |
3675 #endif | 3596 #endif |
3676 /* check if we have a previous block to deblock it with dstBlock */ | 3597 /* check if we have a previous block to deblock it with dstBlock */ |
3677 if(x - 8 >= 0) | 3598 if(x - 8 >= 0){ |
3678 { | |
3679 #ifdef HAVE_MMX | 3599 #ifdef HAVE_MMX |
3680 if(mode & H_X1_FILTER) | 3600 if(mode & H_X1_FILTER) |
3681 RENAME(vertX1Filter)(tempBlock1, 16, &c); | 3601 RENAME(vertX1Filter)(tempBlock1, 16, &c); |
3682 else if(mode & H_DEBLOCK) | 3602 else if(mode & H_DEBLOCK){ |
3683 { | |
3684 //START_TIMER | 3603 //START_TIMER |
3685 const int t= RENAME(vertClassify)(tempBlock1, 16, &c); | 3604 const int t= RENAME(vertClassify)(tempBlock1, 16, &c); |
3686 //STOP_TIMER("dc & minmax") | 3605 //STOP_TIMER("dc & minmax") |
3687 if(t==1) | 3606 if(t==1) |
3688 RENAME(doVertLowPass)(tempBlock1, 16, &c); | 3607 RENAME(doVertLowPass)(tempBlock1, 16, &c); |
3689 else if(t==2) | 3608 else if(t==2) |
3690 RENAME(doVertDefFilter)(tempBlock1, 16, &c); | 3609 RENAME(doVertDefFilter)(tempBlock1, 16, &c); |
3691 }else if(mode & H_A_DEBLOCK){ | 3610 }else if(mode & H_A_DEBLOCK){ |
3692 RENAME(do_a_deblock)(tempBlock1, 16, 1, &c); | 3611 RENAME(do_a_deblock)(tempBlock1, 16, 1, &c); |
3693 } | 3612 } |
3694 | 3613 |
3695 RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16); | 3614 RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16); |
3696 | 3615 |
3697 #else | 3616 #else |
3698 if(mode & H_X1_FILTER) | 3617 if(mode & H_X1_FILTER) |
3699 horizX1Filter(dstBlock-4, stride, QP); | 3618 horizX1Filter(dstBlock-4, stride, QP); |
3700 else if(mode & H_DEBLOCK) | 3619 else if(mode & H_DEBLOCK){ |
3701 { | |
3702 #ifdef HAVE_ALTIVEC | 3620 #ifdef HAVE_ALTIVEC |
3703 DECLARE_ALIGNED(16, unsigned char, tempBlock[272]); | 3621 DECLARE_ALIGNED(16, unsigned char, tempBlock[272]); |
3704 transpose_16x8_char_toPackedAlign_altivec(tempBlock, dstBlock - (4 + 1), stride); | 3622 transpose_16x8_char_toPackedAlign_altivec(tempBlock, dstBlock - (4 + 1), stride); |
3705 | 3623 |
3706 const int t=vertClassify_altivec(tempBlock-48, 16, &c); | 3624 const int t=vertClassify_altivec(tempBlock-48, 16, &c); |
3707 if(t==1) { | 3625 if(t==1) { |
3708 doVertLowPass_altivec(tempBlock-48, 16, &c); | 3626 doVertLowPass_altivec(tempBlock-48, 16, &c); |
3709 transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride); | 3627 transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride); |
3710 } | 3628 } |
3711 else if(t==2) { | 3629 else if(t==2) { |
3712 doVertDefFilter_altivec(tempBlock-48, 16, &c); | 3630 doVertDefFilter_altivec(tempBlock-48, 16, &c); |
3713 transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride); | 3631 transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride); |
3714 } | 3632 } |
3715 #else | 3633 #else |
3716 const int t= RENAME(horizClassify)(dstBlock-4, stride, &c); | 3634 const int t= RENAME(horizClassify)(dstBlock-4, stride, &c); |
3717 | 3635 |
3718 if(t==1) | 3636 if(t==1) |
3719 RENAME(doHorizLowPass)(dstBlock-4, stride, &c); | 3637 RENAME(doHorizLowPass)(dstBlock-4, stride, &c); |
3720 else if(t==2) | 3638 else if(t==2) |
3721 RENAME(doHorizDefFilter)(dstBlock-4, stride, &c); | 3639 RENAME(doHorizDefFilter)(dstBlock-4, stride, &c); |
3722 #endif | 3640 #endif |
3723 }else if(mode & H_A_DEBLOCK){ | 3641 }else if(mode & H_A_DEBLOCK){ |
3724 RENAME(do_a_deblock)(dstBlock-8, 1, stride, &c); | 3642 RENAME(do_a_deblock)(dstBlock-8, 1, stride, &c); |
3725 } | 3643 } |
3726 #endif //HAVE_MMX | 3644 #endif //HAVE_MMX |
3727 if(mode & DERING) | 3645 if(mode & DERING){ |
3728 { | 3646 //FIXME filter first line |
3729 //FIXME filter first line | 3647 if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c); |
3730 if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c); | 3648 } |
3731 } | 3649 |
3732 | 3650 if(mode & TEMP_NOISE_FILTER) |
3733 if(mode & TEMP_NOISE_FILTER) | 3651 { |
3734 { | 3652 RENAME(tempNoiseReducer)(dstBlock-8, stride, |
3735 RENAME(tempNoiseReducer)(dstBlock-8, stride, | 3653 c.tempBlured[isColor] + y*dstStride + x, |
3736 c.tempBlured[isColor] + y*dstStride + x, | 3654 c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3), |
3737 c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3), | 3655 c.ppMode.maxTmpNoise); |
3738 c.ppMode.maxTmpNoise); | 3656 } |
3739 } | 3657 } |
3740 } | 3658 |
3741 | 3659 dstBlock+=8; |
3742 dstBlock+=8; | 3660 srcBlock+=8; |
3743 srcBlock+=8; | |
3744 | 3661 |
3745 #ifdef HAVE_MMX | 3662 #ifdef HAVE_MMX |
3746 tmpXchg= tempBlock1; | 3663 tmpXchg= tempBlock1; |
3747 tempBlock1= tempBlock2; | 3664 tempBlock1= tempBlock2; |
3748 tempBlock2 = tmpXchg; | 3665 tempBlock2 = tmpXchg; |
3749 #endif | 3666 #endif |
3667 } | |
3668 | |
3669 if(mode & DERING){ | |
3670 if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c); | |
3671 } | |
3672 | |
3673 if((mode & TEMP_NOISE_FILTER)){ | |
3674 RENAME(tempNoiseReducer)(dstBlock-8, dstStride, | |
3675 c.tempBlured[isColor] + y*dstStride + x, | |
3676 c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3), | |
3677 c.ppMode.maxTmpNoise); | |
3678 } | |
3679 | |
3680 /* did we use a tmp buffer for the last lines*/ | |
3681 if(y+15 >= height){ | |
3682 uint8_t *dstBlock= &(dst[y*dstStride]); | |
3683 if(width==FFABS(dstStride)) | |
3684 linecpy(dstBlock, tempDst + dstStride, height-y, dstStride); | |
3685 else{ | |
3686 int i; | |
3687 for(i=0; i<height-y; i++){ | |
3688 memcpy(dstBlock + i*dstStride, tempDst + (i+1)*dstStride, width); | |
3750 } | 3689 } |
3751 | 3690 } |
3752 if(mode & DERING) | 3691 } |
3753 { | |
3754 if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c); | |
3755 } | |
3756 | |
3757 if((mode & TEMP_NOISE_FILTER)) | |
3758 { | |
3759 RENAME(tempNoiseReducer)(dstBlock-8, dstStride, | |
3760 c.tempBlured[isColor] + y*dstStride + x, | |
3761 c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3), | |
3762 c.ppMode.maxTmpNoise); | |
3763 } | |
3764 | |
3765 /* did we use a tmp buffer for the last lines*/ | |
3766 if(y+15 >= height) | |
3767 { | |
3768 uint8_t *dstBlock= &(dst[y*dstStride]); | |
3769 if(width==FFABS(dstStride)) | |
3770 linecpy(dstBlock, tempDst + dstStride, height-y, dstStride); | |
3771 else | |
3772 { | |
3773 int i; | |
3774 for(i=0; i<height-y; i++) | |
3775 { | |
3776 memcpy(dstBlock + i*dstStride, tempDst + (i+1)*dstStride, width); | |
3777 } | |
3778 } | |
3779 } | |
3780 /* | 3692 /* |
3781 for(x=0; x<width; x+=32) | 3693 for(x=0; x<width; x+=32){ |
3782 { | 3694 volatile int i; |
3783 volatile int i; | 3695 i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride] |
3784 i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride] | 3696 + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride] |
3785 + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride] | 3697 + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride]; |
3786 + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride]; | 3698 + dstBlock[x +13*dstStride] |
3787 // + dstBlock[x +13*dstStride] | 3699 + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride]; |
3788 // + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride]; | 3700 }*/ |
3789 }*/ | 3701 } |
3702 #ifdef HAVE_3DNOW | |
3703 asm volatile("femms"); | |
3704 #elif defined (HAVE_MMX) | |
3705 asm volatile("emms"); | |
3706 #endif | |
3707 | |
3708 #ifdef DEBUG_BRIGHTNESS | |
3709 if(!isColor){ | |
3710 int max=1; | |
3711 int i; | |
3712 for(i=0; i<256; i++) | |
3713 if(yHistogram[i] > max) max=yHistogram[i]; | |
3714 | |
3715 for(i=1; i<256; i++){ | |
3716 int x; | |
3717 int start=yHistogram[i-1]/(max/256+1); | |
3718 int end=yHistogram[i]/(max/256+1); | |
3719 int inc= end > start ? 1 : -1; | |
3720 for(x=start; x!=end+inc; x+=inc) | |
3721 dst[ i*dstStride + x]+=128; | |
3790 } | 3722 } |
3791 #ifdef HAVE_3DNOW | 3723 |
3792 asm volatile("femms"); | 3724 for(i=0; i<100; i+=2){ |
3793 #elif defined (HAVE_MMX) | 3725 dst[ (white)*dstStride + i]+=128; |
3794 asm volatile("emms"); | 3726 dst[ (black)*dstStride + i]+=128; |
3727 } | |
3728 } | |
3795 #endif | 3729 #endif |
3796 | 3730 |
3797 #ifdef DEBUG_BRIGHTNESS | 3731 *c2= c; //copy local context back |
3798 if(!isColor) | |
3799 { | |
3800 int max=1; | |
3801 int i; | |
3802 for(i=0; i<256; i++) | |
3803 if(yHistogram[i] > max) max=yHistogram[i]; | |
3804 | |
3805 for(i=1; i<256; i++) | |
3806 { | |
3807 int x; | |
3808 int start=yHistogram[i-1]/(max/256+1); | |
3809 int end=yHistogram[i]/(max/256+1); | |
3810 int inc= end > start ? 1 : -1; | |
3811 for(x=start; x!=end+inc; x+=inc) | |
3812 dst[ i*dstStride + x]+=128; | |
3813 } | |
3814 | |
3815 for(i=0; i<100; i+=2) | |
3816 { | |
3817 dst[ (white)*dstStride + i]+=128; | |
3818 dst[ (black)*dstStride + i]+=128; | |
3819 } | |
3820 | |
3821 } | |
3822 #endif | |
3823 | |
3824 *c2= c; //copy local context back | |
3825 | 3732 |
3826 } | 3733 } |