/src/aom/aom_dsp/aom_convolve.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * Copyright (c) 2016, Alliance for Open Media. All rights reserved. |
3 | | * |
4 | | * This source code is subject to the terms of the BSD 2 Clause License and |
5 | | * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
6 | | * was not distributed with this source code in the LICENSE file, you can |
7 | | * obtain it at www.aomedia.org/license/software. If the Alliance for Open |
8 | | * Media Patent License 1.0 was not distributed with this source code in the |
9 | | * PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
10 | | */ |
11 | | |
12 | | #include <assert.h> |
13 | | #include <string.h> |
14 | | |
15 | | #include "config/aom_config.h" |
16 | | #include "config/aom_dsp_rtcd.h" |
17 | | |
18 | | #include "aom/aom_integer.h" |
19 | | #include "aom_dsp/aom_dsp_common.h" |
20 | | #include "aom_dsp/aom_filter.h" |
21 | | #include "aom_ports/mem.h" |
22 | | |
23 | 0 | static inline int horz_scalar_product(const uint8_t *a, const int16_t *b) { |
24 | 0 | int sum = 0; |
25 | 0 | for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k]; |
26 | 0 | return sum; |
27 | 0 | } |
28 | | |
29 | | static inline int vert_scalar_product(const uint8_t *a, ptrdiff_t a_stride, |
30 | 0 | const int16_t *b) { |
31 | 0 | int sum = 0; |
32 | 0 | for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k * a_stride] * b[k]; |
33 | 0 | return sum; |
34 | 0 | } |
35 | | |
36 | | static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride, |
37 | | uint8_t *dst, ptrdiff_t dst_stride, |
38 | | const InterpKernel *x_filters, int x0_q4, |
39 | 0 | int x_step_q4, int w, int h) { |
40 | 0 | src -= SUBPEL_TAPS / 2 - 1; |
41 | 0 | for (int y = 0; y < h; ++y) { |
42 | 0 | int x_q4 = x0_q4; |
43 | 0 | for (int x = 0; x < w; ++x) { |
44 | 0 | const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; |
45 | 0 | const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; |
46 | 0 | const int sum = horz_scalar_product(src_x, x_filter); |
47 | 0 | dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)); |
48 | 0 | x_q4 += x_step_q4; |
49 | 0 | } |
50 | 0 | src += src_stride; |
51 | 0 | dst += dst_stride; |
52 | 0 | } |
53 | 0 | } |
54 | | |
55 | | static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride, |
56 | | uint8_t *dst, ptrdiff_t dst_stride, |
57 | | const InterpKernel *y_filters, int y0_q4, |
58 | 0 | int y_step_q4, int w, int h) { |
59 | 0 | src -= src_stride * (SUBPEL_TAPS / 2 - 1); |
60 | |
|
61 | 0 | for (int x = 0; x < w; ++x) { |
62 | 0 | int y_q4 = y0_q4; |
63 | 0 | for (int y = 0; y < h; ++y) { |
64 | 0 | const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; |
65 | 0 | const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; |
66 | 0 | const int sum = vert_scalar_product(src_y, src_stride, y_filter); |
67 | 0 | dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)); |
68 | 0 | y_q4 += y_step_q4; |
69 | 0 | } |
70 | 0 | ++src; |
71 | 0 | ++dst; |
72 | 0 | } |
73 | 0 | } |
74 | | |
75 | 0 | static const InterpKernel *get_filter_base(const int16_t *filter) { |
76 | | // NOTE: This assumes that the filter table is 256-byte aligned. |
77 | 0 | return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF)); |
78 | 0 | } |
79 | | |
80 | 0 | static int get_filter_offset(const int16_t *f, const InterpKernel *base) { |
81 | 0 | return (int)((const InterpKernel *)(intptr_t)f - base); |
82 | 0 | } |
83 | | |
84 | | void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, |
85 | | uint8_t *dst, ptrdiff_t dst_stride, |
86 | | const int16_t *filter_x, int x_step_q4, |
87 | | const int16_t *filter_y, int y_step_q4, int w, |
88 | 0 | int h) { |
89 | 0 | const InterpKernel *const filters_x = get_filter_base(filter_x); |
90 | 0 | const int x0_q4 = get_filter_offset(filter_x, filters_x); |
91 | |
|
92 | 0 | (void)filter_y; |
93 | 0 | (void)y_step_q4; |
94 | |
|
95 | 0 | convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4, |
96 | 0 | w, h); |
97 | 0 | } |
98 | | |
99 | | void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, |
100 | | uint8_t *dst, ptrdiff_t dst_stride, |
101 | | const int16_t *filter_x, int x_step_q4, |
102 | | const int16_t *filter_y, int y_step_q4, int w, |
103 | 0 | int h) { |
104 | 0 | const InterpKernel *const filters_y = get_filter_base(filter_y); |
105 | 0 | const int y0_q4 = get_filter_offset(filter_y, filters_y); |
106 | |
|
107 | 0 | (void)filter_x; |
108 | 0 | (void)x_step_q4; |
109 | |
|
110 | 0 | convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, y_step_q4, |
111 | 0 | w, h); |
112 | 0 | } |
113 | | |
114 | | void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, |
115 | | ptrdiff_t dst_stride, const InterpKernel *filter, |
116 | | int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, |
117 | 0 | int h) { |
118 | | // Note: Fixed size intermediate buffer, temp, places limits on parameters. |
119 | | // 2d filtering proceeds in 2 steps: |
120 | | // (1) Interpolate horizontally into an intermediate buffer, temp. |
121 | | // (2) Interpolate temp vertically to derive the sub-pixel result. |
122 | | // Deriving the maximum number of rows in the temp buffer (135): |
123 | | // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative). |
124 | | // --Largest block size is 64x64 pixels. |
125 | | // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the |
126 | | // original frame (in 1/16th pixel units). |
127 | | // --Must round-up because block may be located at sub-pixel position. |
128 | | // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails. |
129 | | // --((64 - 1) * 32 + 15) >> 4 + 8 = 135. |
130 | | // When calling in frame scaling function, the smallest scaling factor is x1/4 |
131 | | // ==> y_step_q4 = 64. Since w and h are at most 16, the temp buffer is still |
132 | | // big enough. |
133 | 0 | uint8_t temp[64 * 135]; |
134 | 0 | const int intermediate_height = |
135 | 0 | (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS; |
136 | |
|
137 | 0 | assert(w <= 64); |
138 | 0 | assert(h <= 64); |
139 | 0 | assert(y_step_q4 <= 32 || (y_step_q4 <= 64 && h <= 32)); |
140 | 0 | assert(x_step_q4 <= 64); |
141 | | |
142 | 0 | convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64, |
143 | 0 | filter, x0_q4, x_step_q4, w, intermediate_height); |
144 | 0 | convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride, filter, |
145 | 0 | y0_q4, y_step_q4, w, h); |
146 | 0 | } |
147 | | |
148 | | void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, |
149 | 0 | ptrdiff_t dst_stride, int w, int h) { |
150 | 0 | for (int r = h; r > 0; --r) { |
151 | 0 | memmove(dst, src, w); |
152 | 0 | src += src_stride; |
153 | 0 | dst += dst_stride; |
154 | 0 | } |
155 | 0 | } |
156 | | |
157 | | #if CONFIG_AV1_HIGHBITDEPTH |
158 | | static inline int highbd_vert_scalar_product(const uint16_t *a, |
159 | | ptrdiff_t a_stride, |
160 | 0 | const int16_t *b) { |
161 | 0 | int sum = 0; |
162 | 0 | for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k * a_stride] * b[k]; |
163 | 0 | return sum; |
164 | 0 | } |
165 | | |
166 | | static inline int highbd_horz_scalar_product(const uint16_t *a, |
167 | 0 | const int16_t *b) { |
168 | 0 | int sum = 0; |
169 | 0 | for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k]; |
170 | 0 | return sum; |
171 | 0 | } |
172 | | |
173 | | static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride, |
174 | | uint8_t *dst8, ptrdiff_t dst_stride, |
175 | | const InterpKernel *x_filters, int x0_q4, |
176 | 0 | int x_step_q4, int w, int h, int bd) { |
177 | 0 | uint16_t *src = CONVERT_TO_SHORTPTR(src8); |
178 | 0 | uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); |
179 | 0 | src -= SUBPEL_TAPS / 2 - 1; |
180 | 0 | for (int y = 0; y < h; ++y) { |
181 | 0 | int x_q4 = x0_q4; |
182 | 0 | for (int x = 0; x < w; ++x) { |
183 | 0 | const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; |
184 | 0 | const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; |
185 | 0 | const int sum = highbd_horz_scalar_product(src_x, x_filter); |
186 | 0 | dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd); |
187 | 0 | x_q4 += x_step_q4; |
188 | 0 | } |
189 | 0 | src += src_stride; |
190 | 0 | dst += dst_stride; |
191 | 0 | } |
192 | 0 | } |
193 | | |
194 | | static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride, |
195 | | uint8_t *dst8, ptrdiff_t dst_stride, |
196 | | const InterpKernel *y_filters, int y0_q4, |
197 | 0 | int y_step_q4, int w, int h, int bd) { |
198 | 0 | uint16_t *src = CONVERT_TO_SHORTPTR(src8); |
199 | 0 | uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); |
200 | 0 | src -= src_stride * (SUBPEL_TAPS / 2 - 1); |
201 | 0 | for (int x = 0; x < w; ++x) { |
202 | 0 | int y_q4 = y0_q4; |
203 | 0 | for (int y = 0; y < h; ++y) { |
204 | 0 | const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; |
205 | 0 | const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; |
206 | 0 | const int sum = highbd_vert_scalar_product(src_y, src_stride, y_filter); |
207 | 0 | dst[y * dst_stride] = |
208 | 0 | clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd); |
209 | 0 | y_q4 += y_step_q4; |
210 | 0 | } |
211 | 0 | ++src; |
212 | 0 | ++dst; |
213 | 0 | } |
214 | 0 | } |
215 | | |
216 | | void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, |
217 | | uint8_t *dst, ptrdiff_t dst_stride, |
218 | | const int16_t *filter_x, int x_step_q4, |
219 | | const int16_t *filter_y, int y_step_q4, int w, |
220 | 0 | int h, int bd) { |
221 | 0 | const InterpKernel *const filters_x = get_filter_base(filter_x); |
222 | 0 | const int x0_q4 = get_filter_offset(filter_x, filters_x); |
223 | 0 | (void)filter_y; |
224 | 0 | (void)y_step_q4; |
225 | |
|
226 | 0 | highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, |
227 | 0 | x_step_q4, w, h, bd); |
228 | 0 | } |
229 | | |
230 | | void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, |
231 | | uint8_t *dst, ptrdiff_t dst_stride, |
232 | | const int16_t *filter_x, int x_step_q4, |
233 | | const int16_t *filter_y, int y_step_q4, int w, |
234 | 0 | int h, int bd) { |
235 | 0 | const InterpKernel *const filters_y = get_filter_base(filter_y); |
236 | 0 | const int y0_q4 = get_filter_offset(filter_y, filters_y); |
237 | 0 | (void)filter_x; |
238 | 0 | (void)x_step_q4; |
239 | |
|
240 | 0 | highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, |
241 | 0 | y_step_q4, w, h, bd); |
242 | 0 | } |
243 | | |
244 | | void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, |
245 | | uint16_t *dst, ptrdiff_t dst_stride, int w, |
246 | 0 | int h) { |
247 | 0 | for (int y = 0; y < h; ++y) { |
248 | 0 | memmove(dst, src, w * sizeof(src[0])); |
249 | 0 | src += src_stride; |
250 | 0 | dst += dst_stride; |
251 | 0 | } |
252 | 0 | } |
253 | | #endif // CONFIG_AV1_HIGHBITDEPTH |