OLD | NEW |
1 /* | 1 /* |
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. | 2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license | 4 * Use of this source code is governed by a BSD-style license |
5 * that can be found in the LICENSE file in the root of the source | 5 * that can be found in the LICENSE file in the root of the source |
6 * tree. An additional intellectual property rights grant can be found | 6 * tree. An additional intellectual property rights grant can be found |
7 * in the file PATENTS. All contributing project authors may | 7 * in the file PATENTS. All contributing project authors may |
8 * be found in the AUTHORS file in the root of the source tree. | 8 * be found in the AUTHORS file in the root of the source tree. |
9 */ | 9 */ |
10 | 10 |
(...skipping 28 matching lines...) Expand all Loading... |
39 return (var - (((int64_t)avg * avg) >> 11)); | 39 return (var - (((int64_t)avg * avg) >> 11)); |
40 } | 40 } |
41 | 41 |
42 unsigned int vp9_sub_pixel_variance64x32_c(const uint8_t *src_ptr, | 42 unsigned int vp9_sub_pixel_variance64x32_c(const uint8_t *src_ptr, |
43 int src_pixels_per_line, | 43 int src_pixels_per_line, |
44 int xoffset, | 44 int xoffset, |
45 int yoffset, | 45 int yoffset, |
46 const uint8_t *dst_ptr, | 46 const uint8_t *dst_ptr, |
47 int dst_pixels_per_line, | 47 int dst_pixels_per_line, |
48 unsigned int *sse) { | 48 unsigned int *sse) { |
49 uint16_t fdata3[65 * 64]; // Temp data bufffer used in filtering | 49 uint16_t fdata3[65 * 64]; // Temp data buffer used in filtering |
50 uint8_t temp2[68 * 64]; | 50 uint8_t temp2[68 * 64]; |
51 const int16_t *hfilter, *vfilter; | 51 const int16_t *hfilter, *vfilter; |
52 | 52 |
53 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 53 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
54 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 54 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
55 | 55 |
56 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 56 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
57 1, 33, 64, hfilter); | 57 1, 33, 64, hfilter); |
58 var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 32, 64, vfilter); | 58 var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 32, 64, vfilter); |
59 | 59 |
60 return vp9_variance64x32(temp2, 64, dst_ptr, dst_pixels_per_line, sse); | 60 return vp9_variance64x32(temp2, 64, dst_ptr, dst_pixels_per_line, sse); |
61 } | 61 } |
62 | 62 |
63 unsigned int vp9_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, | 63 unsigned int vp9_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, |
64 int src_pixels_per_line, | 64 int src_pixels_per_line, |
65 int xoffset, | 65 int xoffset, |
66 int yoffset, | 66 int yoffset, |
67 const uint8_t *dst_ptr, | 67 const uint8_t *dst_ptr, |
68 int dst_pixels_per_line, | 68 int dst_pixels_per_line, |
69 unsigned int *sse, | 69 unsigned int *sse, |
70 const uint8_t *second_pred) { | 70 const uint8_t *second_pred) { |
71 uint16_t fdata3[65 * 64]; // Temp data bufffer used in filtering | 71 uint16_t fdata3[65 * 64]; // Temp data buffer used in filtering |
72 uint8_t temp2[68 * 64]; | 72 uint8_t temp2[68 * 64]; |
73 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 64 * 64); // compound pred buffer | 73 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 64 * 64); // compound pred buffer |
74 const int16_t *hfilter, *vfilter; | 74 const int16_t *hfilter, *vfilter; |
75 | 75 |
76 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 76 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
77 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 77 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
78 | 78 |
79 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 79 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
80 1, 33, 64, hfilter); | 80 1, 33, 64, hfilter); |
81 var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 32, 64, vfilter); | 81 var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 32, 64, vfilter); |
82 comp_avg_pred(temp3, second_pred, 64, 32, temp2, 64); | 82 comp_avg_pred(temp3, second_pred, 64, 32, temp2, 64); |
83 return vp9_variance64x32(temp3, 64, dst_ptr, dst_pixels_per_line, sse); | 83 return vp9_variance64x32(temp3, 64, dst_ptr, dst_pixels_per_line, sse); |
84 } | 84 } |
85 | 85 |
86 unsigned int vp9_variance32x64_c(const uint8_t *src_ptr, | 86 unsigned int vp9_variance32x64_c(const uint8_t *src_ptr, |
87 int source_stride, | 87 int source_stride, |
88 const uint8_t *ref_ptr, | 88 const uint8_t *ref_ptr, |
89 int recon_stride, | 89 int recon_stride, |
90 unsigned int *sse) { | 90 unsigned int *sse) { |
91 unsigned int var; | 91 unsigned int var; |
92 int avg; | 92 int avg; |
93 | 93 |
94 variance(src_ptr, source_stride, ref_ptr, recon_stride, 32, 64, &var, &avg); | 94 variance(src_ptr, source_stride, ref_ptr, recon_stride, 32, 64, &var, &avg); |
95 *sse = var; | 95 *sse = var; |
96 return (var - (((int64_t)avg * avg) >> 11)); | 96 return (var - (((int64_t)avg * avg) >> 11)); |
97 } | 97 } |
98 | 98 |
99 unsigned int vp9_sub_pixel_variance32x64_c(const uint8_t *src_ptr, | 99 unsigned int vp9_sub_pixel_variance32x64_c(const uint8_t *src_ptr, |
100 int src_pixels_per_line, | 100 int src_pixels_per_line, |
101 int xoffset, | 101 int xoffset, |
102 int yoffset, | 102 int yoffset, |
103 const uint8_t *dst_ptr, | 103 const uint8_t *dst_ptr, |
104 int dst_pixels_per_line, | 104 int dst_pixels_per_line, |
105 unsigned int *sse) { | 105 unsigned int *sse) { |
106 uint16_t fdata3[65 * 64]; // Temp data bufffer used in filtering | 106 uint16_t fdata3[65 * 64]; // Temp data buffer used in filtering |
107 uint8_t temp2[68 * 64]; | 107 uint8_t temp2[68 * 64]; |
108 const int16_t *hfilter, *vfilter; | 108 const int16_t *hfilter, *vfilter; |
109 | 109 |
110 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 110 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
111 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 111 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
112 | 112 |
113 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 113 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
114 1, 65, 32, hfilter); | 114 1, 65, 32, hfilter); |
115 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 64, 32, vfilter); | 115 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 64, 32, vfilter); |
116 | 116 |
117 return vp9_variance32x64(temp2, 32, dst_ptr, dst_pixels_per_line, sse); | 117 return vp9_variance32x64(temp2, 32, dst_ptr, dst_pixels_per_line, sse); |
118 } | 118 } |
119 | 119 |
120 unsigned int vp9_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, | 120 unsigned int vp9_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, |
121 int src_pixels_per_line, | 121 int src_pixels_per_line, |
122 int xoffset, | 122 int xoffset, |
123 int yoffset, | 123 int yoffset, |
124 const uint8_t *dst_ptr, | 124 const uint8_t *dst_ptr, |
125 int dst_pixels_per_line, | 125 int dst_pixels_per_line, |
126 unsigned int *sse, | 126 unsigned int *sse, |
127 const uint8_t *second_pred) { | 127 const uint8_t *second_pred) { |
128 uint16_t fdata3[65 * 64]; // Temp data bufffer used in filtering | 128 uint16_t fdata3[65 * 64]; // Temp data buffer used in filtering |
129 uint8_t temp2[68 * 64]; | 129 uint8_t temp2[68 * 64]; |
130 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 32 * 64); // compound pred buffer | 130 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 32 * 64); // compound pred buffer |
131 const int16_t *hfilter, *vfilter; | 131 const int16_t *hfilter, *vfilter; |
132 | 132 |
133 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 133 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
134 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 134 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
135 | 135 |
136 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 136 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
137 1, 65, 32, hfilter); | 137 1, 65, 32, hfilter); |
138 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 64, 32, vfilter); | 138 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 64, 32, vfilter); |
139 comp_avg_pred(temp3, second_pred, 32, 64, temp2, 32); | 139 comp_avg_pred(temp3, second_pred, 32, 64, temp2, 32); |
140 return vp9_variance32x64(temp3, 32, dst_ptr, dst_pixels_per_line, sse); | 140 return vp9_variance32x64(temp3, 32, dst_ptr, dst_pixels_per_line, sse); |
141 } | 141 } |
142 | 142 |
143 unsigned int vp9_variance32x16_c(const uint8_t *src_ptr, | 143 unsigned int vp9_variance32x16_c(const uint8_t *src_ptr, |
144 int source_stride, | 144 int source_stride, |
145 const uint8_t *ref_ptr, | 145 const uint8_t *ref_ptr, |
146 int recon_stride, | 146 int recon_stride, |
147 unsigned int *sse) { | 147 unsigned int *sse) { |
148 unsigned int var; | 148 unsigned int var; |
149 int avg; | 149 int avg; |
150 | 150 |
151 variance(src_ptr, source_stride, ref_ptr, recon_stride, 32, 16, &var, &avg); | 151 variance(src_ptr, source_stride, ref_ptr, recon_stride, 32, 16, &var, &avg); |
152 *sse = var; | 152 *sse = var; |
153 return (var - (((int64_t)avg * avg) >> 9)); | 153 return (var - (((int64_t)avg * avg) >> 9)); |
154 } | 154 } |
155 | 155 |
156 unsigned int vp9_sub_pixel_variance32x16_c(const uint8_t *src_ptr, | 156 unsigned int vp9_sub_pixel_variance32x16_c(const uint8_t *src_ptr, |
157 int src_pixels_per_line, | 157 int src_pixels_per_line, |
158 int xoffset, | 158 int xoffset, |
159 int yoffset, | 159 int yoffset, |
160 const uint8_t *dst_ptr, | 160 const uint8_t *dst_ptr, |
161 int dst_pixels_per_line, | 161 int dst_pixels_per_line, |
162 unsigned int *sse) { | 162 unsigned int *sse) { |
163 uint16_t fdata3[33 * 32]; // Temp data bufffer used in filtering | 163 uint16_t fdata3[33 * 32]; // Temp data buffer used in filtering |
164 uint8_t temp2[36 * 32]; | 164 uint8_t temp2[36 * 32]; |
165 const int16_t *hfilter, *vfilter; | 165 const int16_t *hfilter, *vfilter; |
166 | 166 |
167 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 167 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
168 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 168 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
169 | 169 |
170 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 170 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
171 1, 17, 32, hfilter); | 171 1, 17, 32, hfilter); |
172 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 16, 32, vfilter); | 172 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 16, 32, vfilter); |
173 | 173 |
174 return vp9_variance32x16(temp2, 32, dst_ptr, dst_pixels_per_line, sse); | 174 return vp9_variance32x16(temp2, 32, dst_ptr, dst_pixels_per_line, sse); |
175 } | 175 } |
176 | 176 |
177 unsigned int vp9_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, | 177 unsigned int vp9_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, |
178 int src_pixels_per_line, | 178 int src_pixels_per_line, |
179 int xoffset, | 179 int xoffset, |
180 int yoffset, | 180 int yoffset, |
181 const uint8_t *dst_ptr, | 181 const uint8_t *dst_ptr, |
182 int dst_pixels_per_line, | 182 int dst_pixels_per_line, |
183 unsigned int *sse, | 183 unsigned int *sse, |
184 const uint8_t *second_pred) { | 184 const uint8_t *second_pred) { |
185 uint16_t fdata3[33 * 32]; // Temp data bufffer used in filtering | 185 uint16_t fdata3[33 * 32]; // Temp data buffer used in filtering |
186 uint8_t temp2[36 * 32]; | 186 uint8_t temp2[36 * 32]; |
187 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 32 * 16); // compound pred buffer | 187 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 32 * 16); // compound pred buffer |
188 const int16_t *hfilter, *vfilter; | 188 const int16_t *hfilter, *vfilter; |
189 | 189 |
190 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 190 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
191 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 191 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
192 | 192 |
193 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 193 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
194 1, 17, 32, hfilter); | 194 1, 17, 32, hfilter); |
195 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 16, 32, vfilter); | 195 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 16, 32, vfilter); |
196 comp_avg_pred(temp3, second_pred, 32, 16, temp2, 32); | 196 comp_avg_pred(temp3, second_pred, 32, 16, temp2, 32); |
197 return vp9_variance32x16(temp3, 32, dst_ptr, dst_pixels_per_line, sse); | 197 return vp9_variance32x16(temp3, 32, dst_ptr, dst_pixels_per_line, sse); |
198 } | 198 } |
199 | 199 |
200 unsigned int vp9_variance16x32_c(const uint8_t *src_ptr, | 200 unsigned int vp9_variance16x32_c(const uint8_t *src_ptr, |
201 int source_stride, | 201 int source_stride, |
202 const uint8_t *ref_ptr, | 202 const uint8_t *ref_ptr, |
203 int recon_stride, | 203 int recon_stride, |
204 unsigned int *sse) { | 204 unsigned int *sse) { |
205 unsigned int var; | 205 unsigned int var; |
206 int avg; | 206 int avg; |
207 | 207 |
208 variance(src_ptr, source_stride, ref_ptr, recon_stride, 16, 32, &var, &avg); | 208 variance(src_ptr, source_stride, ref_ptr, recon_stride, 16, 32, &var, &avg); |
209 *sse = var; | 209 *sse = var; |
210 return (var - (((int64_t)avg * avg) >> 9)); | 210 return (var - (((int64_t)avg * avg) >> 9)); |
211 } | 211 } |
212 | 212 |
213 unsigned int vp9_sub_pixel_variance16x32_c(const uint8_t *src_ptr, | 213 unsigned int vp9_sub_pixel_variance16x32_c(const uint8_t *src_ptr, |
214 int src_pixels_per_line, | 214 int src_pixels_per_line, |
215 int xoffset, | 215 int xoffset, |
216 int yoffset, | 216 int yoffset, |
217 const uint8_t *dst_ptr, | 217 const uint8_t *dst_ptr, |
218 int dst_pixels_per_line, | 218 int dst_pixels_per_line, |
219 unsigned int *sse) { | 219 unsigned int *sse) { |
220 uint16_t fdata3[33 * 32]; // Temp data bufffer used in filtering | 220 uint16_t fdata3[33 * 32]; // Temp data buffer used in filtering |
221 uint8_t temp2[36 * 32]; | 221 uint8_t temp2[36 * 32]; |
222 const int16_t *hfilter, *vfilter; | 222 const int16_t *hfilter, *vfilter; |
223 | 223 |
224 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 224 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
225 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 225 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
226 | 226 |
227 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 227 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
228 1, 33, 16, hfilter); | 228 1, 33, 16, hfilter); |
229 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 32, 16, vfilter); | 229 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 32, 16, vfilter); |
230 | 230 |
231 return vp9_variance16x32(temp2, 16, dst_ptr, dst_pixels_per_line, sse); | 231 return vp9_variance16x32(temp2, 16, dst_ptr, dst_pixels_per_line, sse); |
232 } | 232 } |
233 | 233 |
234 unsigned int vp9_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, | 234 unsigned int vp9_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, |
235 int src_pixels_per_line, | 235 int src_pixels_per_line, |
236 int xoffset, | 236 int xoffset, |
237 int yoffset, | 237 int yoffset, |
238 const uint8_t *dst_ptr, | 238 const uint8_t *dst_ptr, |
239 int dst_pixels_per_line, | 239 int dst_pixels_per_line, |
240 unsigned int *sse, | 240 unsigned int *sse, |
241 const uint8_t *second_pred) { | 241 const uint8_t *second_pred) { |
242 uint16_t fdata3[33 * 32]; // Temp data bufffer used in filtering | 242 uint16_t fdata3[33 * 32]; // Temp data buffer used in filtering |
243 uint8_t temp2[36 * 32]; | 243 uint8_t temp2[36 * 32]; |
244 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 16 * 32); // compound pred buffer | 244 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 16 * 32); // compound pred buffer |
245 const int16_t *hfilter, *vfilter; | 245 const int16_t *hfilter, *vfilter; |
246 | 246 |
247 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 247 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
248 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 248 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
249 | 249 |
250 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 250 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
251 1, 33, 16, hfilter); | 251 1, 33, 16, hfilter); |
252 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 32, 16, vfilter); | 252 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 32, 16, vfilter); |
253 comp_avg_pred(temp3, second_pred, 16, 32, temp2, 16); | 253 comp_avg_pred(temp3, second_pred, 16, 32, temp2, 16); |
254 return vp9_variance16x32(temp3, 16, dst_ptr, dst_pixels_per_line, sse); | 254 return vp9_variance16x32(temp3, 16, dst_ptr, dst_pixels_per_line, sse); |
255 } | 255 } |
256 | 256 |
257 unsigned int vp9_variance64x64_c(const uint8_t *src_ptr, | 257 unsigned int vp9_variance64x64_c(const uint8_t *src_ptr, |
258 int source_stride, | 258 int source_stride, |
(...skipping 174 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
433 | 433 |
434 unsigned int vp9_sub_pixel_variance4x4_c(const uint8_t *src_ptr, | 434 unsigned int vp9_sub_pixel_variance4x4_c(const uint8_t *src_ptr, |
435 int src_pixels_per_line, | 435 int src_pixels_per_line, |
436 int xoffset, | 436 int xoffset, |
437 int yoffset, | 437 int yoffset, |
438 const uint8_t *dst_ptr, | 438 const uint8_t *dst_ptr, |
439 int dst_pixels_per_line, | 439 int dst_pixels_per_line, |
440 unsigned int *sse) { | 440 unsigned int *sse) { |
441 uint8_t temp2[20 * 16]; | 441 uint8_t temp2[20 * 16]; |
442 const int16_t *hfilter, *vfilter; | 442 const int16_t *hfilter, *vfilter; |
443 uint16_t fdata3[5 * 4]; // Temp data bufffer used in filtering | 443 uint16_t fdata3[5 * 4]; // Temp data buffer used in filtering |
444 | 444 |
445 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 445 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
446 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 446 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
447 | 447 |
448 // First filter 1d Horizontal | 448 // First filter 1d Horizontal |
449 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 449 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
450 1, 5, 4, hfilter); | 450 1, 5, 4, hfilter); |
451 | 451 |
452 // Now filter Verticaly | 452 // Now filter Verticaly |
453 var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, vfilter); | 453 var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, vfilter); |
454 | 454 |
455 return vp9_variance4x4(temp2, 4, dst_ptr, dst_pixels_per_line, sse); | 455 return vp9_variance4x4(temp2, 4, dst_ptr, dst_pixels_per_line, sse); |
456 } | 456 } |
457 | 457 |
458 unsigned int vp9_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, | 458 unsigned int vp9_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, |
459 int src_pixels_per_line, | 459 int src_pixels_per_line, |
460 int xoffset, | 460 int xoffset, |
461 int yoffset, | 461 int yoffset, |
462 const uint8_t *dst_ptr, | 462 const uint8_t *dst_ptr, |
463 int dst_pixels_per_line, | 463 int dst_pixels_per_line, |
464 unsigned int *sse, | 464 unsigned int *sse, |
465 const uint8_t *second_pred) { | 465 const uint8_t *second_pred) { |
466 uint8_t temp2[20 * 16]; | 466 uint8_t temp2[20 * 16]; |
467 const int16_t *hfilter, *vfilter; | 467 const int16_t *hfilter, *vfilter; |
468 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 4 * 4); // compound pred buffer | 468 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 4 * 4); // compound pred buffer |
469 uint16_t fdata3[5 * 4]; // Temp data bufffer used in filtering | 469 uint16_t fdata3[5 * 4]; // Temp data buffer used in filtering |
470 | 470 |
471 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 471 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
472 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 472 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
473 | 473 |
474 // First filter 1d Horizontal | 474 // First filter 1d Horizontal |
475 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 475 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
476 1, 5, 4, hfilter); | 476 1, 5, 4, hfilter); |
477 | 477 |
478 // Now filter Verticaly | 478 // Now filter Verticaly |
479 var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, vfilter); | 479 var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, vfilter); |
480 comp_avg_pred(temp3, second_pred, 4, 4, temp2, 4); | 480 comp_avg_pred(temp3, second_pred, 4, 4, temp2, 4); |
481 return vp9_variance4x4(temp3, 4, dst_ptr, dst_pixels_per_line, sse); | 481 return vp9_variance4x4(temp3, 4, dst_ptr, dst_pixels_per_line, sse); |
482 } | 482 } |
483 | 483 |
484 unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, | 484 unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, |
485 int src_pixels_per_line, | 485 int src_pixels_per_line, |
486 int xoffset, | 486 int xoffset, |
487 int yoffset, | 487 int yoffset, |
488 const uint8_t *dst_ptr, | 488 const uint8_t *dst_ptr, |
489 int dst_pixels_per_line, | 489 int dst_pixels_per_line, |
490 unsigned int *sse) { | 490 unsigned int *sse) { |
491 uint16_t fdata3[9 * 8]; // Temp data bufffer used in filtering | 491 uint16_t fdata3[9 * 8]; // Temp data buffer used in filtering |
492 uint8_t temp2[20 * 16]; | 492 uint8_t temp2[20 * 16]; |
493 const int16_t *hfilter, *vfilter; | 493 const int16_t *hfilter, *vfilter; |
494 | 494 |
495 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 495 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
496 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 496 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
497 | 497 |
498 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 498 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
499 1, 9, 8, hfilter); | 499 1, 9, 8, hfilter); |
500 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 8, 8, vfilter); | 500 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 8, 8, vfilter); |
501 | 501 |
502 return vp9_variance8x8(temp2, 8, dst_ptr, dst_pixels_per_line, sse); | 502 return vp9_variance8x8(temp2, 8, dst_ptr, dst_pixels_per_line, sse); |
503 } | 503 } |
504 | 504 |
505 unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, | 505 unsigned int vp9_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, |
506 int src_pixels_per_line, | 506 int src_pixels_per_line, |
507 int xoffset, | 507 int xoffset, |
508 int yoffset, | 508 int yoffset, |
509 const uint8_t *dst_ptr, | 509 const uint8_t *dst_ptr, |
510 int dst_pixels_per_line, | 510 int dst_pixels_per_line, |
511 unsigned int *sse, | 511 unsigned int *sse, |
512 const uint8_t *second_pred) { | 512 const uint8_t *second_pred) { |
513 uint16_t fdata3[9 * 8]; // Temp data bufffer used in filtering | 513 uint16_t fdata3[9 * 8]; // Temp data buffer used in filtering |
514 uint8_t temp2[20 * 16]; | 514 uint8_t temp2[20 * 16]; |
515 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 8); // compound pred buffer | 515 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 8); // compound pred buffer |
516 const int16_t *hfilter, *vfilter; | 516 const int16_t *hfilter, *vfilter; |
517 | 517 |
518 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 518 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
519 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 519 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
520 | 520 |
521 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 521 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
522 1, 9, 8, hfilter); | 522 1, 9, 8, hfilter); |
523 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 8, 8, vfilter); | 523 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 8, 8, vfilter); |
524 comp_avg_pred(temp3, second_pred, 8, 8, temp2, 8); | 524 comp_avg_pred(temp3, second_pred, 8, 8, temp2, 8); |
525 return vp9_variance8x8(temp3, 8, dst_ptr, dst_pixels_per_line, sse); | 525 return vp9_variance8x8(temp3, 8, dst_ptr, dst_pixels_per_line, sse); |
526 } | 526 } |
527 | 527 |
528 unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, | 528 unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, |
529 int src_pixels_per_line, | 529 int src_pixels_per_line, |
530 int xoffset, | 530 int xoffset, |
531 int yoffset, | 531 int yoffset, |
532 const uint8_t *dst_ptr, | 532 const uint8_t *dst_ptr, |
533 int dst_pixels_per_line, | 533 int dst_pixels_per_line, |
534 unsigned int *sse) { | 534 unsigned int *sse) { |
535 uint16_t fdata3[17 * 16]; // Temp data bufffer used in filtering | 535 uint16_t fdata3[17 * 16]; // Temp data buffer used in filtering |
536 uint8_t temp2[20 * 16]; | 536 uint8_t temp2[20 * 16]; |
537 const int16_t *hfilter, *vfilter; | 537 const int16_t *hfilter, *vfilter; |
538 | 538 |
539 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 539 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
540 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 540 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
541 | 541 |
542 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 542 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
543 1, 17, 16, hfilter); | 543 1, 17, 16, hfilter); |
544 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 16, 16, vfilter); | 544 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 16, 16, vfilter); |
545 | 545 |
546 return vp9_variance16x16(temp2, 16, dst_ptr, dst_pixels_per_line, sse); | 546 return vp9_variance16x16(temp2, 16, dst_ptr, dst_pixels_per_line, sse); |
547 } | 547 } |
548 | 548 |
549 unsigned int vp9_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, | 549 unsigned int vp9_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, |
550 int src_pixels_per_line, | 550 int src_pixels_per_line, |
551 int xoffset, | 551 int xoffset, |
552 int yoffset, | 552 int yoffset, |
553 const uint8_t *dst_ptr, | 553 const uint8_t *dst_ptr, |
554 int dst_pixels_per_line, | 554 int dst_pixels_per_line, |
555 unsigned int *sse, | 555 unsigned int *sse, |
556 const uint8_t *second_pred) { | 556 const uint8_t *second_pred) { |
557 uint16_t fdata3[17 * 16]; | 557 uint16_t fdata3[17 * 16]; |
558 uint8_t temp2[20 * 16]; | 558 uint8_t temp2[20 * 16]; |
559 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 16 * 16); // compound pred buffer | 559 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 16 * 16); // compound pred buffer |
560 const int16_t *hfilter, *vfilter; | 560 const int16_t *hfilter, *vfilter; |
561 | 561 |
562 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 562 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
563 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 563 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
564 | 564 |
565 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 565 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
566 1, 17, 16, hfilter); | 566 1, 17, 16, hfilter); |
567 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 16, 16, vfilter); | 567 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 16, 16, vfilter); |
568 | 568 |
569 comp_avg_pred(temp3, second_pred, 16, 16, temp2, 16); | 569 comp_avg_pred(temp3, second_pred, 16, 16, temp2, 16); |
570 return vp9_variance16x16(temp3, 16, dst_ptr, dst_pixels_per_line, sse); | 570 return vp9_variance16x16(temp3, 16, dst_ptr, dst_pixels_per_line, sse); |
571 } | 571 } |
572 | 572 |
573 unsigned int vp9_sub_pixel_variance64x64_c(const uint8_t *src_ptr, | 573 unsigned int vp9_sub_pixel_variance64x64_c(const uint8_t *src_ptr, |
574 int src_pixels_per_line, | 574 int src_pixels_per_line, |
575 int xoffset, | 575 int xoffset, |
576 int yoffset, | 576 int yoffset, |
577 const uint8_t *dst_ptr, | 577 const uint8_t *dst_ptr, |
578 int dst_pixels_per_line, | 578 int dst_pixels_per_line, |
579 unsigned int *sse) { | 579 unsigned int *sse) { |
580 uint16_t fdata3[65 * 64]; // Temp data bufffer used in filtering | 580 uint16_t fdata3[65 * 64]; // Temp data buffer used in filtering |
581 uint8_t temp2[68 * 64]; | 581 uint8_t temp2[68 * 64]; |
582 const int16_t *hfilter, *vfilter; | 582 const int16_t *hfilter, *vfilter; |
583 | 583 |
584 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 584 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
585 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 585 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
586 | 586 |
587 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 587 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
588 1, 65, 64, hfilter); | 588 1, 65, 64, hfilter); |
589 var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 64, 64, vfilter); | 589 var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 64, 64, vfilter); |
590 | 590 |
591 return vp9_variance64x64(temp2, 64, dst_ptr, dst_pixels_per_line, sse); | 591 return vp9_variance64x64(temp2, 64, dst_ptr, dst_pixels_per_line, sse); |
592 } | 592 } |
593 | 593 |
594 unsigned int vp9_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, | 594 unsigned int vp9_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, |
595 int src_pixels_per_line, | 595 int src_pixels_per_line, |
596 int xoffset, | 596 int xoffset, |
597 int yoffset, | 597 int yoffset, |
598 const uint8_t *dst_ptr, | 598 const uint8_t *dst_ptr, |
599 int dst_pixels_per_line, | 599 int dst_pixels_per_line, |
600 unsigned int *sse, | 600 unsigned int *sse, |
601 const uint8_t *second_pred) { | 601 const uint8_t *second_pred) { |
602 uint16_t fdata3[65 * 64]; // Temp data bufffer used in filtering | 602 uint16_t fdata3[65 * 64]; // Temp data buffer used in filtering |
603 uint8_t temp2[68 * 64]; | 603 uint8_t temp2[68 * 64]; |
604 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 64 * 64); // compound pred buffer | 604 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 64 * 64); // compound pred buffer |
605 const int16_t *hfilter, *vfilter; | 605 const int16_t *hfilter, *vfilter; |
606 | 606 |
607 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 607 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
608 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 608 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
609 | 609 |
610 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 610 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
611 1, 65, 64, hfilter); | 611 1, 65, 64, hfilter); |
612 var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 64, 64, vfilter); | 612 var_filter_block2d_bil_second_pass(fdata3, temp2, 64, 64, 64, 64, vfilter); |
613 comp_avg_pred(temp3, second_pred, 64, 64, temp2, 64); | 613 comp_avg_pred(temp3, second_pred, 64, 64, temp2, 64); |
614 return vp9_variance64x64(temp3, 64, dst_ptr, dst_pixels_per_line, sse); | 614 return vp9_variance64x64(temp3, 64, dst_ptr, dst_pixels_per_line, sse); |
615 } | 615 } |
616 | 616 |
617 unsigned int vp9_sub_pixel_variance32x32_c(const uint8_t *src_ptr, | 617 unsigned int vp9_sub_pixel_variance32x32_c(const uint8_t *src_ptr, |
618 int src_pixels_per_line, | 618 int src_pixels_per_line, |
619 int xoffset, | 619 int xoffset, |
620 int yoffset, | 620 int yoffset, |
621 const uint8_t *dst_ptr, | 621 const uint8_t *dst_ptr, |
622 int dst_pixels_per_line, | 622 int dst_pixels_per_line, |
623 unsigned int *sse) { | 623 unsigned int *sse) { |
624 uint16_t fdata3[33 * 32]; // Temp data bufffer used in filtering | 624 uint16_t fdata3[33 * 32]; // Temp data buffer used in filtering |
625 uint8_t temp2[36 * 32]; | 625 uint8_t temp2[36 * 32]; |
626 const int16_t *hfilter, *vfilter; | 626 const int16_t *hfilter, *vfilter; |
627 | 627 |
628 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 628 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
629 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 629 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
630 | 630 |
631 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 631 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
632 1, 33, 32, hfilter); | 632 1, 33, 32, hfilter); |
633 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 32, 32, vfilter); | 633 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 32, 32, vfilter); |
634 | 634 |
635 return vp9_variance32x32(temp2, 32, dst_ptr, dst_pixels_per_line, sse); | 635 return vp9_variance32x32(temp2, 32, dst_ptr, dst_pixels_per_line, sse); |
636 } | 636 } |
637 | 637 |
638 unsigned int vp9_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, | 638 unsigned int vp9_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, |
639 int src_pixels_per_line, | 639 int src_pixels_per_line, |
640 int xoffset, | 640 int xoffset, |
641 int yoffset, | 641 int yoffset, |
642 const uint8_t *dst_ptr, | 642 const uint8_t *dst_ptr, |
643 int dst_pixels_per_line, | 643 int dst_pixels_per_line, |
644 unsigned int *sse, | 644 unsigned int *sse, |
645 const uint8_t *second_pred) { | 645 const uint8_t *second_pred) { |
646 uint16_t fdata3[33 * 32]; // Temp data bufffer used in filtering | 646 uint16_t fdata3[33 * 32]; // Temp data buffer used in filtering |
647 uint8_t temp2[36 * 32]; | 647 uint8_t temp2[36 * 32]; |
648 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 32 * 32); // compound pred buffer | 648 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 32 * 32); // compound pred buffer |
649 const int16_t *hfilter, *vfilter; | 649 const int16_t *hfilter, *vfilter; |
650 | 650 |
651 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 651 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
652 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 652 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
653 | 653 |
654 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 654 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
655 1, 33, 32, hfilter); | 655 1, 33, 32, hfilter); |
656 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 32, 32, vfilter); | 656 var_filter_block2d_bil_second_pass(fdata3, temp2, 32, 32, 32, 32, vfilter); |
657 comp_avg_pred(temp3, second_pred, 32, 32, temp2, 32); | 657 comp_avg_pred(temp3, second_pred, 32, 32, temp2, 32); |
658 return vp9_variance32x32(temp3, 32, dst_ptr, dst_pixels_per_line, sse); | 658 return vp9_variance32x32(temp3, 32, dst_ptr, dst_pixels_per_line, sse); |
659 } | 659 } |
660 | 660 |
661 unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr, | 661 unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr, |
662 int source_stride, | 662 int source_stride, |
(...skipping 115 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
778 return *sse; | 778 return *sse; |
779 } | 779 } |
780 | 780 |
781 unsigned int vp9_sub_pixel_variance16x8_c(const uint8_t *src_ptr, | 781 unsigned int vp9_sub_pixel_variance16x8_c(const uint8_t *src_ptr, |
782 int src_pixels_per_line, | 782 int src_pixels_per_line, |
783 int xoffset, | 783 int xoffset, |
784 int yoffset, | 784 int yoffset, |
785 const uint8_t *dst_ptr, | 785 const uint8_t *dst_ptr, |
786 int dst_pixels_per_line, | 786 int dst_pixels_per_line, |
787 unsigned int *sse) { | 787 unsigned int *sse) { |
788 uint16_t fdata3[16 * 9]; // Temp data bufffer used in filtering | 788 uint16_t fdata3[16 * 9]; // Temp data buffer used in filtering |
789 uint8_t temp2[20 * 16]; | 789 uint8_t temp2[20 * 16]; |
790 const int16_t *hfilter, *vfilter; | 790 const int16_t *hfilter, *vfilter; |
791 | 791 |
792 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 792 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
793 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 793 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
794 | 794 |
795 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 795 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
796 1, 9, 16, hfilter); | 796 1, 9, 16, hfilter); |
797 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 8, 16, vfilter); | 797 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 8, 16, vfilter); |
798 | 798 |
799 return vp9_variance16x8(temp2, 16, dst_ptr, dst_pixels_per_line, sse); | 799 return vp9_variance16x8(temp2, 16, dst_ptr, dst_pixels_per_line, sse); |
800 } | 800 } |
801 | 801 |
802 unsigned int vp9_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, | 802 unsigned int vp9_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, |
803 int src_pixels_per_line, | 803 int src_pixels_per_line, |
804 int xoffset, | 804 int xoffset, |
805 int yoffset, | 805 int yoffset, |
806 const uint8_t *dst_ptr, | 806 const uint8_t *dst_ptr, |
807 int dst_pixels_per_line, | 807 int dst_pixels_per_line, |
808 unsigned int *sse, | 808 unsigned int *sse, |
809 const uint8_t *second_pred) { | 809 const uint8_t *second_pred) { |
810 uint16_t fdata3[16 * 9]; // Temp data bufffer used in filtering | 810 uint16_t fdata3[16 * 9]; // Temp data buffer used in filtering |
811 uint8_t temp2[20 * 16]; | 811 uint8_t temp2[20 * 16]; |
812 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 16 * 8); // compound pred buffer | 812 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 16 * 8); // compound pred buffer |
813 const int16_t *hfilter, *vfilter; | 813 const int16_t *hfilter, *vfilter; |
814 | 814 |
815 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 815 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
816 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 816 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
817 | 817 |
818 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 818 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
819 1, 9, 16, hfilter); | 819 1, 9, 16, hfilter); |
820 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 8, 16, vfilter); | 820 var_filter_block2d_bil_second_pass(fdata3, temp2, 16, 16, 8, 16, vfilter); |
821 comp_avg_pred(temp3, second_pred, 16, 8, temp2, 16); | 821 comp_avg_pred(temp3, second_pred, 16, 8, temp2, 16); |
822 return vp9_variance16x8(temp3, 16, dst_ptr, dst_pixels_per_line, sse); | 822 return vp9_variance16x8(temp3, 16, dst_ptr, dst_pixels_per_line, sse); |
823 } | 823 } |
824 | 824 |
825 unsigned int vp9_sub_pixel_variance8x16_c(const uint8_t *src_ptr, | 825 unsigned int vp9_sub_pixel_variance8x16_c(const uint8_t *src_ptr, |
826 int src_pixels_per_line, | 826 int src_pixels_per_line, |
827 int xoffset, | 827 int xoffset, |
828 int yoffset, | 828 int yoffset, |
829 const uint8_t *dst_ptr, | 829 const uint8_t *dst_ptr, |
830 int dst_pixels_per_line, | 830 int dst_pixels_per_line, |
831 unsigned int *sse) { | 831 unsigned int *sse) { |
832 uint16_t fdata3[9 * 16]; // Temp data bufffer used in filtering | 832 uint16_t fdata3[9 * 16]; // Temp data buffer used in filtering |
833 uint8_t temp2[20 * 16]; | 833 uint8_t temp2[20 * 16]; |
834 const int16_t *hfilter, *vfilter; | 834 const int16_t *hfilter, *vfilter; |
835 | 835 |
836 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 836 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
837 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 837 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
838 | 838 |
839 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 839 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
840 1, 17, 8, hfilter); | 840 1, 17, 8, hfilter); |
841 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 16, 8, vfilter); | 841 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 16, 8, vfilter); |
842 | 842 |
843 return vp9_variance8x16(temp2, 8, dst_ptr, dst_pixels_per_line, sse); | 843 return vp9_variance8x16(temp2, 8, dst_ptr, dst_pixels_per_line, sse); |
844 } | 844 } |
845 | 845 |
846 unsigned int vp9_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, | 846 unsigned int vp9_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, |
847 int src_pixels_per_line, | 847 int src_pixels_per_line, |
848 int xoffset, | 848 int xoffset, |
849 int yoffset, | 849 int yoffset, |
850 const uint8_t *dst_ptr, | 850 const uint8_t *dst_ptr, |
851 int dst_pixels_per_line, | 851 int dst_pixels_per_line, |
852 unsigned int *sse, | 852 unsigned int *sse, |
853 const uint8_t *second_pred) { | 853 const uint8_t *second_pred) { |
854 uint16_t fdata3[9 * 16]; // Temp data bufffer used in filtering | 854 uint16_t fdata3[9 * 16]; // Temp data buffer used in filtering |
855 uint8_t temp2[20 * 16]; | 855 uint8_t temp2[20 * 16]; |
856 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 16); // compound pred buffer | 856 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 16); // compound pred buffer |
857 const int16_t *hfilter, *vfilter; | 857 const int16_t *hfilter, *vfilter; |
858 | 858 |
859 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 859 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
860 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 860 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
861 | 861 |
862 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 862 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
863 1, 17, 8, hfilter); | 863 1, 17, 8, hfilter); |
864 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 16, 8, vfilter); | 864 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 16, 8, vfilter); |
865 comp_avg_pred(temp3, second_pred, 8, 16, temp2, 8); | 865 comp_avg_pred(temp3, second_pred, 8, 16, temp2, 8); |
866 return vp9_variance8x16(temp3, 8, dst_ptr, dst_pixels_per_line, sse); | 866 return vp9_variance8x16(temp3, 8, dst_ptr, dst_pixels_per_line, sse); |
867 } | 867 } |
868 | 868 |
869 unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr, | 869 unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr, |
870 int src_pixels_per_line, | 870 int src_pixels_per_line, |
871 int xoffset, | 871 int xoffset, |
872 int yoffset, | 872 int yoffset, |
873 const uint8_t *dst_ptr, | 873 const uint8_t *dst_ptr, |
874 int dst_pixels_per_line, | 874 int dst_pixels_per_line, |
875 unsigned int *sse) { | 875 unsigned int *sse) { |
876 uint16_t fdata3[8 * 5]; // Temp data bufffer used in filtering | 876 uint16_t fdata3[8 * 5]; // Temp data buffer used in filtering |
877 uint8_t temp2[20 * 16]; | 877 uint8_t temp2[20 * 16]; |
878 const int16_t *hfilter, *vfilter; | 878 const int16_t *hfilter, *vfilter; |
879 | 879 |
880 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 880 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
881 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 881 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
882 | 882 |
883 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 883 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
884 1, 5, 8, hfilter); | 884 1, 5, 8, hfilter); |
885 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 4, 8, vfilter); | 885 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 4, 8, vfilter); |
886 | 886 |
887 return vp9_variance8x4(temp2, 8, dst_ptr, dst_pixels_per_line, sse); | 887 return vp9_variance8x4(temp2, 8, dst_ptr, dst_pixels_per_line, sse); |
888 } | 888 } |
889 | 889 |
890 unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, | 890 unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, |
891 int src_pixels_per_line, | 891 int src_pixels_per_line, |
892 int xoffset, | 892 int xoffset, |
893 int yoffset, | 893 int yoffset, |
894 const uint8_t *dst_ptr, | 894 const uint8_t *dst_ptr, |
895 int dst_pixels_per_line, | 895 int dst_pixels_per_line, |
896 unsigned int *sse, | 896 unsigned int *sse, |
897 const uint8_t *second_pred) { | 897 const uint8_t *second_pred) { |
898 uint16_t fdata3[8 * 5]; // Temp data bufffer used in filtering | 898 uint16_t fdata3[8 * 5]; // Temp data buffer used in filtering |
899 uint8_t temp2[20 * 16]; | 899 uint8_t temp2[20 * 16]; |
900 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 4); // compound pred buffer | 900 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 4); // compound pred buffer |
901 const int16_t *hfilter, *vfilter; | 901 const int16_t *hfilter, *vfilter; |
902 | 902 |
903 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 903 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
904 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 904 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
905 | 905 |
906 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 906 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
907 1, 5, 8, hfilter); | 907 1, 5, 8, hfilter); |
908 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 4, 8, vfilter); | 908 var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 4, 8, vfilter); |
909 comp_avg_pred(temp3, second_pred, 8, 4, temp2, 8); | 909 comp_avg_pred(temp3, second_pred, 8, 4, temp2, 8); |
910 return vp9_variance8x4(temp3, 8, dst_ptr, dst_pixels_per_line, sse); | 910 return vp9_variance8x4(temp3, 8, dst_ptr, dst_pixels_per_line, sse); |
911 } | 911 } |
912 | 912 |
913 unsigned int vp9_sub_pixel_variance4x8_c(const uint8_t *src_ptr, | 913 unsigned int vp9_sub_pixel_variance4x8_c(const uint8_t *src_ptr, |
914 int src_pixels_per_line, | 914 int src_pixels_per_line, |
915 int xoffset, | 915 int xoffset, |
916 int yoffset, | 916 int yoffset, |
917 const uint8_t *dst_ptr, | 917 const uint8_t *dst_ptr, |
918 int dst_pixels_per_line, | 918 int dst_pixels_per_line, |
919 unsigned int *sse) { | 919 unsigned int *sse) { |
920 uint16_t fdata3[5 * 8]; // Temp data bufffer used in filtering | 920 uint16_t fdata3[5 * 8]; // Temp data buffer used in filtering |
921 // FIXME(jingning,rbultje): this temp2 buffer probably doesn't need to be | 921 // FIXME(jingning,rbultje): this temp2 buffer probably doesn't need to be |
922 // of this big? same issue appears in all other block size settings. | 922 // of this big? same issue appears in all other block size settings. |
923 uint8_t temp2[20 * 16]; | 923 uint8_t temp2[20 * 16]; |
924 const int16_t *hfilter, *vfilter; | 924 const int16_t *hfilter, *vfilter; |
925 | 925 |
926 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 926 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
927 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 927 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
928 | 928 |
929 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 929 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
930 1, 9, 4, hfilter); | 930 1, 9, 4, hfilter); |
931 var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 8, 4, vfilter); | 931 var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 8, 4, vfilter); |
932 | 932 |
933 return vp9_variance4x8(temp2, 4, dst_ptr, dst_pixels_per_line, sse); | 933 return vp9_variance4x8(temp2, 4, dst_ptr, dst_pixels_per_line, sse); |
934 } | 934 } |
935 | 935 |
936 unsigned int vp9_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, | 936 unsigned int vp9_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, |
937 int src_pixels_per_line, | 937 int src_pixels_per_line, |
938 int xoffset, | 938 int xoffset, |
939 int yoffset, | 939 int yoffset, |
940 const uint8_t *dst_ptr, | 940 const uint8_t *dst_ptr, |
941 int dst_pixels_per_line, | 941 int dst_pixels_per_line, |
942 unsigned int *sse, | 942 unsigned int *sse, |
943 const uint8_t *second_pred) { | 943 const uint8_t *second_pred) { |
944 uint16_t fdata3[5 * 8]; // Temp data bufffer used in filtering | 944 uint16_t fdata3[5 * 8]; // Temp data buffer used in filtering |
945 uint8_t temp2[20 * 16]; | 945 uint8_t temp2[20 * 16]; |
946 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 4 * 8); // compound pred buffer | 946 DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 4 * 8); // compound pred buffer |
947 const int16_t *hfilter, *vfilter; | 947 const int16_t *hfilter, *vfilter; |
948 | 948 |
949 hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset); | 949 hfilter = BILINEAR_FILTERS_2TAP(xoffset); |
950 vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset); | 950 vfilter = BILINEAR_FILTERS_2TAP(yoffset); |
951 | 951 |
952 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, | 952 var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line, |
953 1, 9, 4, hfilter); | 953 1, 9, 4, hfilter); |
954 var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 8, 4, vfilter); | 954 var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 8, 4, vfilter); |
955 comp_avg_pred(temp3, second_pred, 4, 8, temp2, 4); | 955 comp_avg_pred(temp3, second_pred, 4, 8, temp2, 4); |
956 return vp9_variance4x8(temp3, 4, dst_ptr, dst_pixels_per_line, sse); | 956 return vp9_variance4x8(temp3, 4, dst_ptr, dst_pixels_per_line, sse); |
957 } | 957 } |
OLD | NEW |