1/*
2 *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
3 *
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
6 *  tree. An additional intellectual property rights grant can be found
7 *  in the file PATENTS.  All contributing project authors may
8 *  be found in the AUTHORS file in the root of the source tree.
9 */
10
11#include <assert.h>
12
13#include "./vpx_config.h"
14#include "./vp9_rtcd.h"
15#include "vp9/common/vp9_common.h"
16#include "vp9/common/vp9_convolve.h"
17#include "vp9/common/vp9_filter.h"
18#include "vpx/vpx_integer.h"
19#include "vpx_ports/mem.h"
20
21static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
22                           uint8_t *dst, ptrdiff_t dst_stride,
23                           const InterpKernel *x_filters,
24                           int x0_q4, int x_step_q4, int w, int h) {
25  int x, y;
26  src -= SUBPEL_TAPS / 2 - 1;
27  for (y = 0; y < h; ++y) {
28    int x_q4 = x0_q4;
29    for (x = 0; x < w; ++x) {
30      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
31      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
32      int k, sum = 0;
33      for (k = 0; k < SUBPEL_TAPS; ++k)
34        sum += src_x[k] * x_filter[k];
35      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
36      x_q4 += x_step_q4;
37    }
38    src += src_stride;
39    dst += dst_stride;
40  }
41}
42
43static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
44                               uint8_t *dst, ptrdiff_t dst_stride,
45                               const InterpKernel *x_filters,
46                               int x0_q4, int x_step_q4, int w, int h) {
47  int x, y;
48  src -= SUBPEL_TAPS / 2 - 1;
49  for (y = 0; y < h; ++y) {
50    int x_q4 = x0_q4;
51    for (x = 0; x < w; ++x) {
52      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
53      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
54      int k, sum = 0;
55      for (k = 0; k < SUBPEL_TAPS; ++k)
56        sum += src_x[k] * x_filter[k];
57      dst[x] = ROUND_POWER_OF_TWO(dst[x] +
58          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
59      x_q4 += x_step_q4;
60    }
61    src += src_stride;
62    dst += dst_stride;
63  }
64}
65
66static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
67                          uint8_t *dst, ptrdiff_t dst_stride,
68                          const InterpKernel *y_filters,
69                          int y0_q4, int y_step_q4, int w, int h) {
70  int x, y;
71  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
72
73  for (x = 0; x < w; ++x) {
74    int y_q4 = y0_q4;
75    for (y = 0; y < h; ++y) {
76      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
77      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
78      int k, sum = 0;
79      for (k = 0; k < SUBPEL_TAPS; ++k)
80        sum += src_y[k * src_stride] * y_filter[k];
81      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
82      y_q4 += y_step_q4;
83    }
84    ++src;
85    ++dst;
86  }
87}
88
89static void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
90                              uint8_t *dst, ptrdiff_t dst_stride,
91                              const InterpKernel *y_filters,
92                              int y0_q4, int y_step_q4, int w, int h) {
93  int x, y;
94  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
95
96  for (x = 0; x < w; ++x) {
97    int y_q4 = y0_q4;
98    for (y = 0; y < h; ++y) {
99      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
100      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
101      int k, sum = 0;
102      for (k = 0; k < SUBPEL_TAPS; ++k)
103        sum += src_y[k * src_stride] * y_filter[k];
104      dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
105          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
106      y_q4 += y_step_q4;
107    }
108    ++src;
109    ++dst;
110  }
111}
112
113static void convolve(const uint8_t *src, ptrdiff_t src_stride,
114                     uint8_t *dst, ptrdiff_t dst_stride,
115                     const InterpKernel *const x_filters,
116                     int x0_q4, int x_step_q4,
117                     const InterpKernel *const y_filters,
118                     int y0_q4, int y_step_q4,
119                     int w, int h) {
120  // Fixed size intermediate buffer places limits on parameters.
121  // Maximum intermediate_height is 324, for y_step_q4 == 80,
122  // h == 64, taps == 8.
123  // y_step_q4 of 80 allows for 1/10 scale for 5 layer svc
124  uint8_t temp[64 * 324];
125  int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + SUBPEL_TAPS;
126
127  assert(w <= 64);
128  assert(h <= 64);
129  assert(y_step_q4 <= 80);
130  assert(x_step_q4 <= 80);
131
132  if (intermediate_height < h)
133    intermediate_height = h;
134
135  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64,
136                 x_filters, x0_q4, x_step_q4, w, intermediate_height);
137  convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride,
138                y_filters, y0_q4, y_step_q4, w, h);
139}
140
141static const InterpKernel *get_filter_base(const int16_t *filter) {
142  // NOTE: This assumes that the filter table is 256-byte aligned.
143  // TODO(agrange) Modify to make independent of table alignment.
144  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
145}
146
147static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
148  return (int)((const InterpKernel *)(intptr_t)f - base);
149}
150
151void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
152                           uint8_t *dst, ptrdiff_t dst_stride,
153                           const int16_t *filter_x, int x_step_q4,
154                           const int16_t *filter_y, int y_step_q4,
155                           int w, int h) {
156  const InterpKernel *const filters_x = get_filter_base(filter_x);
157  const int x0_q4 = get_filter_offset(filter_x, filters_x);
158
159  convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
160                 x0_q4, x_step_q4, w, h);
161}
162
163void vp9_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
164                               uint8_t *dst, ptrdiff_t dst_stride,
165                               const int16_t *filter_x, int x_step_q4,
166                               const int16_t *filter_y, int y_step_q4,
167                               int w, int h) {
168  const InterpKernel *const filters_x = get_filter_base(filter_x);
169  const int x0_q4 = get_filter_offset(filter_x, filters_x);
170
171  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
172                     x0_q4, x_step_q4, w, h);
173}
174
175void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
176                          uint8_t *dst, ptrdiff_t dst_stride,
177                          const int16_t *filter_x, int x_step_q4,
178                          const int16_t *filter_y, int y_step_q4,
179                          int w, int h) {
180  const InterpKernel *const filters_y = get_filter_base(filter_y);
181  const int y0_q4 = get_filter_offset(filter_y, filters_y);
182  convolve_vert(src, src_stride, dst, dst_stride, filters_y,
183                y0_q4, y_step_q4, w, h);
184}
185
186void vp9_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
187                              uint8_t *dst, ptrdiff_t dst_stride,
188                              const int16_t *filter_x, int x_step_q4,
189                              const int16_t *filter_y, int y_step_q4,
190                              int w, int h) {
191  const InterpKernel *const filters_y = get_filter_base(filter_y);
192  const int y0_q4 = get_filter_offset(filter_y, filters_y);
193  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
194                    y0_q4, y_step_q4, w, h);
195}
196
197void vp9_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
198                     uint8_t *dst, ptrdiff_t dst_stride,
199                     const int16_t *filter_x, int x_step_q4,
200                     const int16_t *filter_y, int y_step_q4,
201                     int w, int h) {
202  const InterpKernel *const filters_x = get_filter_base(filter_x);
203  const int x0_q4 = get_filter_offset(filter_x, filters_x);
204
205  const InterpKernel *const filters_y = get_filter_base(filter_y);
206  const int y0_q4 = get_filter_offset(filter_y, filters_y);
207
208  convolve(src, src_stride, dst, dst_stride,
209           filters_x, x0_q4, x_step_q4,
210           filters_y, y0_q4, y_step_q4, w, h);
211}
212
213void vp9_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
214                         uint8_t *dst, ptrdiff_t dst_stride,
215                         const int16_t *filter_x, int x_step_q4,
216                         const int16_t *filter_y, int y_step_q4,
217                         int w, int h) {
218  /* Fixed size intermediate buffer places limits on parameters. */
219  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp, 64 * 64);
220  assert(w <= 64);
221  assert(h <= 64);
222
223  vp9_convolve8_c(src, src_stride, temp, 64,
224                  filter_x, x_step_q4, filter_y, y_step_q4, w, h);
225  vp9_convolve_avg_c(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
226}
227
228void vp9_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride,
229                         uint8_t *dst, ptrdiff_t dst_stride,
230                         const int16_t *filter_x, int filter_x_stride,
231                         const int16_t *filter_y, int filter_y_stride,
232                         int w, int h) {
233  int r;
234
235  for (r = h; r > 0; --r) {
236    vpx_memcpy(dst, src, w);
237    src += src_stride;
238    dst += dst_stride;
239  }
240}
241
242void vp9_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride,
243                        uint8_t *dst, ptrdiff_t dst_stride,
244                        const int16_t *filter_x, int filter_x_stride,
245                        const int16_t *filter_y, int filter_y_stride,
246                        int w, int h) {
247  int x, y;
248
249  for (y = 0; y < h; ++y) {
250    for (x = 0; x < w; ++x)
251      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
252
253    src += src_stride;
254    dst += dst_stride;
255  }
256}
257