aom_convolve.c 23.8 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1 2 3 4 5 6 7 8 9
/*
 *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */
Christian Duvivier's avatar
Christian Duvivier committed
10

John Koleszar's avatar
John Koleszar committed
11
#include <assert.h>
Zoe Liu's avatar
Zoe Liu committed
12
#include <string.h>
John Koleszar's avatar
John Koleszar committed
13

Yaowu Xu's avatar
Yaowu Xu committed
14 15 16 17 18 19
#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
#include "aom/aom_integer.h"
#include "aom_dsp/aom_convolve.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/aom_filter.h"
20
#include "aom_ports/mem.h"
John Koleszar's avatar
John Koleszar committed
21

Dmitry Kovalev's avatar
Dmitry Kovalev committed
22 23
static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
                           uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
24 25
                           const InterpKernel *x_filters, int x0_q4,
                           int x_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
26 27
  int x, y;
  src -= SUBPEL_TAPS / 2 - 1;
John Koleszar's avatar
John Koleszar committed
28
  for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
29
    int x_q4 = x0_q4;
John Koleszar's avatar
John Koleszar committed
30
    for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
31 32 33
      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
clang-format's avatar
clang-format committed
34
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
35
      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
John Koleszar's avatar
John Koleszar committed
36 37 38 39 40 41 42
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
43 44
static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
                               uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
45 46
                               const InterpKernel *x_filters, int x0_q4,
                               int x_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
47 48
  int x, y;
  src -= SUBPEL_TAPS / 2 - 1;
John Koleszar's avatar
John Koleszar committed
49
  for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
50
    int x_q4 = x0_q4;
John Koleszar's avatar
John Koleszar committed
51
    for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
52 53 54
      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
clang-format's avatar
clang-format committed
55 56 57
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
      dst[x] = ROUND_POWER_OF_TWO(
          dst[x] + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
John Koleszar's avatar
John Koleszar committed
58 59 60 61 62 63 64
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
65 66
static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
                          uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
67 68
                          const InterpKernel *y_filters, int y0_q4,
                          int y_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
69 70
  int x, y;
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
John Koleszar's avatar
John Koleszar committed
71

72
  for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
73
    int y_q4 = y0_q4;
John Koleszar's avatar
John Koleszar committed
74
    for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
75 76 77 78 79 80
      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
John Koleszar's avatar
John Koleszar committed
81 82 83 84 85 86 87
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
88 89
static void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
                              uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
90 91
                              const InterpKernel *y_filters, int y0_q4,
                              int y_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
92 93
  int x, y;
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
John Koleszar's avatar
John Koleszar committed
94

95
  for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
96
    int y_q4 = y0_q4;
John Koleszar's avatar
John Koleszar committed
97
    for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
98 99 100 101 102
      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
clang-format's avatar
clang-format committed
103 104 105 106
      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
          dst[y * dst_stride] +
              clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)),
          1);
John Koleszar's avatar
John Koleszar committed
107 108 109 110 111 112 113
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

clang-format's avatar
clang-format committed
114 115
static void convolve(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
                     ptrdiff_t dst_stride, const InterpKernel *const x_filters,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
116
                     int x0_q4, int x_step_q4,
clang-format's avatar
clang-format committed
117 118
                     const InterpKernel *const y_filters, int y0_q4,
                     int y_step_q4, int w, int h) {
119 120 121 122 123 124 125 126 127 128 129 130
  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
  // 2d filtering proceeds in 2 steps:
  //   (1) Interpolate horizontally into an intermediate buffer, temp.
  //   (2) Interpolate temp vertically to derive the sub-pixel result.
  // Deriving the maximum number of rows in the temp buffer (135):
  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
  // --Largest block size is 64x64 pixels.
  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
  //   original frame (in 1/16th pixel units).
  // --Must round-up because block may be located at sub-pixel position.
  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
131
  uint8_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
132
  int intermediate_height =
clang-format's avatar
clang-format committed
133
      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
134

135 136
  assert(w <= MAX_SB_SIZE);
  assert(h <= MAX_SB_SIZE);
137

138 139
  assert(y_step_q4 <= 32);
  assert(x_step_q4 <= 32);
140

clang-format's avatar
clang-format committed
141 142 143 144 145
  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp,
                 MAX_SB_SIZE, x_filters, x0_q4, x_step_q4, w,
                 intermediate_height);
  convolve_vert(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), MAX_SB_SIZE, dst,
                dst_stride, y_filters, y0_q4, y_step_q4, w, h);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
146 147
}

148
static const InterpKernel *get_filter_base(const int16_t *filter) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
149 150
  // NOTE: This assumes that the filter table is 256-byte aligned.
  // TODO(agrange) Modify to make independent of table alignment.
151
  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
Dmitry Kovalev's avatar
Dmitry Kovalev committed
152 153
}

154
static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
155
  return (int)((const InterpKernel *)(intptr_t)f - base);
John Koleszar's avatar
John Koleszar committed
156 157
}

Yaowu Xu's avatar
Yaowu Xu committed
158
void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
159
                           uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
160
                           const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
161 162
                           const int16_t *filter_y, int y_step_q4, int w,
                           int h) {
163
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
164 165
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

166 167 168
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
169 170
  convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
                 w, h);
John Koleszar's avatar
John Koleszar committed
171 172
}

Yaowu Xu's avatar
Yaowu Xu committed
173
void aom_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
174
                               uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
175
                               const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
176 177
                               const int16_t *filter_y, int y_step_q4, int w,
                               int h) {
178
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
179 180
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

181 182 183
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
184 185
  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                     x_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
186 187
}

Yaowu Xu's avatar
Yaowu Xu committed
188
void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
189
                          uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
190
                          const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
191 192
                          const int16_t *filter_y, int y_step_q4, int w,
                          int h) {
193
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
194
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
195 196 197 198

  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
199 200
  convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, y_step_q4,
                w, h);
John Koleszar's avatar
John Koleszar committed
201 202
}

Yaowu Xu's avatar
Yaowu Xu committed
203
void aom_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
204
                              uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
205
                              const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
206 207
                              const int16_t *filter_y, int y_step_q4, int w,
                              int h) {
208
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
209
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
210 211 212 213

  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
214 215
  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
                    y_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
216 217
}

Yaowu Xu's avatar
Yaowu Xu committed
218
void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
219 220
                     ptrdiff_t dst_stride, const int16_t *filter_x,
                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
John Koleszar's avatar
John Koleszar committed
221
                     int w, int h) {
222
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
223 224
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

225
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
226 227
  const int y0_q4 = get_filter_offset(filter_y, filters_y);

clang-format's avatar
clang-format committed
228
  convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
229
           filters_y, y0_q4, y_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
230 231
}

Yaowu Xu's avatar
Yaowu Xu committed
232
void aom_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
233 234
                         ptrdiff_t dst_stride, const int16_t *filter_x,
                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
John Koleszar's avatar
John Koleszar committed
235
                         int w, int h) {
Christian Duvivier's avatar
Christian Duvivier committed
236
  /* Fixed size intermediate buffer places limits on parameters. */
237 238 239
  DECLARE_ALIGNED(16, uint8_t, temp[MAX_SB_SIZE * MAX_SB_SIZE]);
  assert(w <= MAX_SB_SIZE);
  assert(h <= MAX_SB_SIZE);
Christian Duvivier's avatar
Christian Duvivier committed
240

Yaowu Xu's avatar
Yaowu Xu committed
241
  aom_convolve8_c(src, src_stride, temp, MAX_SB_SIZE, filter_x, x_step_q4,
clang-format's avatar
clang-format committed
242
                  filter_y, y_step_q4, w, h);
Yaowu Xu's avatar
Yaowu Xu committed
243
  aom_convolve_avg_c(temp, MAX_SB_SIZE, dst, dst_stride, NULL, 0, NULL, 0, w,
clang-format's avatar
clang-format committed
244
                     h);
John Koleszar's avatar
John Koleszar committed
245
}
246

Yaowu Xu's avatar
Yaowu Xu committed
247
void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
248 249 250
                         ptrdiff_t dst_stride, const int16_t *filter_x,
                         int filter_x_stride, const int16_t *filter_y,
                         int filter_y_stride, int w, int h) {
251 252
  int r;

clang-format's avatar
clang-format committed
253 254 255 256
  (void)filter_x;
  (void)filter_x_stride;
  (void)filter_y;
  (void)filter_y_stride;
257

258
  for (r = h; r > 0; --r) {
James Zern's avatar
James Zern committed
259
    memcpy(dst, src, w);
260 261
    src += src_stride;
    dst += dst_stride;
262 263 264
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
265
void aom_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
266 267 268
                        ptrdiff_t dst_stride, const int16_t *filter_x,
                        int filter_x_stride, const int16_t *filter_y,
                        int filter_y_stride, int w, int h) {
269 270
  int x, y;

clang-format's avatar
clang-format committed
271 272 273 274
  (void)filter_x;
  (void)filter_x_stride;
  (void)filter_y;
  (void)filter_y_stride;
275

276
  for (y = 0; y < h; ++y) {
clang-format's avatar
clang-format committed
277
    for (x = 0; x < w; ++x) dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
278

279 280 281 282
    src += src_stride;
    dst += dst_stride;
  }
}
283

Yaowu Xu's avatar
Yaowu Xu committed
284
void aom_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
285 286
                        ptrdiff_t dst_stride, const int16_t *filter_x,
                        int x_step_q4, const int16_t *filter_y, int y_step_q4,
287
                        int w, int h) {
Yaowu Xu's avatar
Yaowu Xu committed
288
  aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
289 290 291
                        filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
292
void aom_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
293 294
                       ptrdiff_t dst_stride, const int16_t *filter_x,
                       int x_step_q4, const int16_t *filter_y, int y_step_q4,
295
                       int w, int h) {
Yaowu Xu's avatar
Yaowu Xu committed
296
  aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
297 298 299
                       filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
300
void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
301 302
                     ptrdiff_t dst_stride, const int16_t *filter_x,
                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
303
                     int w, int h) {
Yaowu Xu's avatar
Yaowu Xu committed
304
  aom_convolve8_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
305 306 307
                  filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
308
void aom_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
309 310
                            uint8_t *dst, ptrdiff_t dst_stride,
                            const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
311 312
                            const int16_t *filter_y, int y_step_q4, int w,
                            int h) {
Yaowu Xu's avatar
Yaowu Xu committed
313
  aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
314 315 316
                            x_step_q4, filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
317
void aom_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
318 319
                           uint8_t *dst, ptrdiff_t dst_stride,
                           const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
320 321
                           const int16_t *filter_y, int y_step_q4, int w,
                           int h) {
Yaowu Xu's avatar
Yaowu Xu committed
322
  aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
323 324 325
                           x_step_q4, filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
326
void aom_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
327 328 329
                         ptrdiff_t dst_stride, const int16_t *filter_x,
                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
                         int w, int h) {
Yaowu Xu's avatar
Yaowu Xu committed
330
  aom_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
331 332 333
                      filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
334
#if CONFIG_AOM_HIGHBITDEPTH
335 336
static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
                                  uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
337 338
                                  const InterpKernel *x_filters, int x0_q4,
                                  int x_step_q4, int w, int h, int bd) {
339 340 341 342 343 344 345 346 347 348
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= SUBPEL_TAPS / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_q4 = x0_q4;
    for (x = 0; x < w; ++x) {
      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
clang-format's avatar
clang-format committed
349
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
350
      dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
351 352 353 354 355 356 357
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

358 359
static void highbd_convolve_avg_horiz(const uint8_t *src8, ptrdiff_t src_stride,
                                      uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
360 361
                                      const InterpKernel *x_filters, int x0_q4,
                                      int x_step_q4, int w, int h, int bd) {
362 363 364 365 366 367 368 369 370 371
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= SUBPEL_TAPS / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_q4 = x0_q4;
    for (x = 0; x < w; ++x) {
      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
clang-format's avatar
clang-format committed
372 373 374 375
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
      dst[x] = ROUND_POWER_OF_TWO(
          dst[x] + clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
          1);
376 377 378 379 380 381 382
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

383 384
static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride,
                                 uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
385 386
                                 const InterpKernel *y_filters, int y0_q4,
                                 int y_step_q4, int w, int h, int bd) {
387 388 389 390 391 392 393 394 395 396 397 398
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
  for (x = 0; x < w; ++x) {
    int y_q4 = y0_q4;
    for (y = 0; y < h; ++y) {
      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
clang-format's avatar
clang-format committed
399 400
      dst[y * dst_stride] =
          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
401 402 403 404 405 406 407
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

408 409
static void highbd_convolve_avg_vert(const uint8_t *src8, ptrdiff_t src_stride,
                                     uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
410 411
                                     const InterpKernel *y_filters, int y0_q4,
                                     int y_step_q4, int w, int h, int bd) {
412 413 414 415 416 417 418 419 420 421 422 423
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
  for (x = 0; x < w; ++x) {
    int y_q4 = y0_q4;
    for (y = 0; y < h; ++y) {
      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
clang-format's avatar
clang-format committed
424 425 426 427
      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
          dst[y * dst_stride] +
              clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
          1);
428 429 430 431 432 433 434
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

435 436
static void highbd_convolve(const uint8_t *src, ptrdiff_t src_stride,
                            uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
437 438 439
                            const InterpKernel *const x_filters, int x0_q4,
                            int x_step_q4, const InterpKernel *const y_filters,
                            int y0_q4, int y_step_q4, int w, int h, int bd) {
440 441 442 443 444 445 446 447 448 449 450 451
  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
  // 2d filtering proceeds in 2 steps:
  //   (1) Interpolate horizontally into an intermediate buffer, temp.
  //   (2) Interpolate temp vertically to derive the sub-pixel result.
  // Deriving the maximum number of rows in the temp buffer (135):
  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
  // --Largest block size is 64x64 pixels.
  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
  //   original frame (in 1/16th pixel units).
  // --Must round-up because block may be located at sub-pixel position.
  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
452
  uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
453
  int intermediate_height =
clang-format's avatar
clang-format committed
454
      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
455

456 457
  assert(w <= MAX_SB_SIZE);
  assert(h <= MAX_SB_SIZE);
458 459 460
  assert(y_step_q4 <= 32);
  assert(x_step_q4 <= 32);

461
  highbd_convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride,
clang-format's avatar
clang-format committed
462 463
                        CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE, x_filters, x0_q4,
                        x_step_q4, w, intermediate_height, bd);
464
  highbd_convolve_vert(
clang-format's avatar
clang-format committed
465 466
      CONVERT_TO_BYTEPTR(temp) + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
      MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, y_step_q4, w, h, bd);
467 468
}

Yaowu Xu's avatar
Yaowu Xu committed
469
void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
470 471
                                  uint8_t *dst, ptrdiff_t dst_stride,
                                  const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
472 473
                                  const int16_t *filter_y, int y_step_q4, int w,
                                  int h, int bd) {
474 475 476 477 478
  const InterpKernel *const filters_x = get_filter_base(filter_x);
  const int x0_q4 = get_filter_offset(filter_x, filters_x);
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
479 480
  highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                        x_step_q4, w, h, bd);
481 482
}

Yaowu Xu's avatar
Yaowu Xu committed
483
void aom_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
484 485 486 487
                                      uint8_t *dst, ptrdiff_t dst_stride,
                                      const int16_t *filter_x, int x_step_q4,
                                      const int16_t *filter_y, int y_step_q4,
                                      int w, int h, int bd) {
488 489 490 491 492
  const InterpKernel *const filters_x = get_filter_base(filter_x);
  const int x0_q4 = get_filter_offset(filter_x, filters_x);
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
493 494
  highbd_convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                            x_step_q4, w, h, bd);
495 496
}

Yaowu Xu's avatar
Yaowu Xu committed
497
void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
498 499
                                 uint8_t *dst, ptrdiff_t dst_stride,
                                 const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
500 501
                                 const int16_t *filter_y, int y_step_q4, int w,
                                 int h, int bd) {
502 503 504 505 506
  const InterpKernel *const filters_y = get_filter_base(filter_y);
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
507 508
  highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
                       y_step_q4, w, h, bd);
509 510
}

Yaowu Xu's avatar
Yaowu Xu committed
511
void aom_highbd_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
512 513 514 515
                                     uint8_t *dst, ptrdiff_t dst_stride,
                                     const int16_t *filter_x, int x_step_q4,
                                     const int16_t *filter_y, int y_step_q4,
                                     int w, int h, int bd) {
516 517 518 519 520
  const InterpKernel *const filters_y = get_filter_base(filter_y);
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
521 522
  highbd_convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
                           y_step_q4, w, h, bd);
523 524
}

Yaowu Xu's avatar
Yaowu Xu committed
525
void aom_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
526 527
                            uint8_t *dst, ptrdiff_t dst_stride,
                            const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
528 529
                            const int16_t *filter_y, int y_step_q4, int w,
                            int h, int bd) {
530 531 532 533 534 535
  const InterpKernel *const filters_x = get_filter_base(filter_x);
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

  const InterpKernel *const filters_y = get_filter_base(filter_y);
  const int y0_q4 = get_filter_offset(filter_y, filters_y);

clang-format's avatar
clang-format committed
536
  highbd_convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
537
                  filters_y, y0_q4, y_step_q4, w, h, bd);
538 539
}

Yaowu Xu's avatar
Yaowu Xu committed
540
void aom_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
541 542
                                uint8_t *dst, ptrdiff_t dst_stride,
                                const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
543 544
                                const int16_t *filter_y, int y_step_q4, int w,
                                int h, int bd) {
545
  // Fixed size intermediate buffer places limits on parameters.
546 547 548
  DECLARE_ALIGNED(16, uint16_t, temp[MAX_SB_SIZE * MAX_SB_SIZE]);
  assert(w <= MAX_SB_SIZE);
  assert(h <= MAX_SB_SIZE);
549

Yaowu Xu's avatar
Yaowu Xu committed
550
  aom_highbd_convolve8_c(src, src_stride, CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE,
551
                         filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);
Yaowu Xu's avatar
Yaowu Xu committed
552
  aom_highbd_convolve_avg_c(CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE, dst,
clang-format's avatar
clang-format committed
553
                            dst_stride, NULL, 0, NULL, 0, w, h, bd);
554 555
}

Yaowu Xu's avatar
Yaowu Xu committed
556
void aom_highbd_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride,
557 558 559 560
                                uint8_t *dst8, ptrdiff_t dst_stride,
                                const int16_t *filter_x, int filter_x_stride,
                                const int16_t *filter_y, int filter_y_stride,
                                int w, int h, int bd) {
561 562 563 564 565 566 567 568 569 570
  int r;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  (void)filter_x;
  (void)filter_y;
  (void)filter_x_stride;
  (void)filter_y_stride;
  (void)bd;

  for (r = h; r > 0; --r) {
James Zern's avatar
James Zern committed
571
    memcpy(dst, src, w * sizeof(uint16_t));
572 573 574 575 576
    src += src_stride;
    dst += dst_stride;
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
577
void aom_highbd_convolve_avg_c(const uint8_t *src8, ptrdiff_t src_stride,
578 579 580 581
                               uint8_t *dst8, ptrdiff_t dst_stride,
                               const int16_t *filter_x, int filter_x_stride,
                               const int16_t *filter_y, int filter_y_stride,
                               int w, int h, int bd) {
582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  (void)filter_x;
  (void)filter_y;
  (void)filter_x_stride;
  (void)filter_y_stride;
  (void)bd;

  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
    }
    src += src_stride;
    dst += dst_stride;
  }
}
#endif