chromium/third_party/libaom/source/libaom/av1/common/resize.c

/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved.
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include <assert.h>
#include <limits.h>
#include <math.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include "config/aom_config.h"
#include "config/av1_rtcd.h"

#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/flow_estimation/corner_detect.h"
#include "aom_ports/mem.h"
#include "av1/common/common.h"
#include "av1/common/resize.h"

#include "config/aom_dsp_rtcd.h"
#include "config/aom_scale_rtcd.h"

// Filters for interpolation (0.5-band) - note this also filters integer pels.
static const InterpKernel filteredinterp_filters500[(1 << RS_SUBPEL_BITS)] = …;

// Filters for interpolation (0.625-band) - note this also filters integer pels.
static const InterpKernel filteredinterp_filters625[(1 << RS_SUBPEL_BITS)] = …;

// Filters for interpolation (0.75-band) - note this also filters integer pels.
static const InterpKernel filteredinterp_filters750[(1 << RS_SUBPEL_BITS)] = …;

// Filters for interpolation (0.875-band) - note this also filters integer pels.
static const InterpKernel filteredinterp_filters875[(1 << RS_SUBPEL_BITS)] = …;

const int16_t av1_resize_filter_normative[(
    1 << RS_SUBPEL_BITS)][UPSCALE_NORMATIVE_TAPS] = …;

// Filters for interpolation (full-band) - no filtering for integer pixels
#define filteredinterp_filters1000 …

static const InterpKernel *choose_interp_filter(int in_length, int out_length) { … }

static void interpolate_core(const uint8_t *const input, int in_length,
                             uint8_t *output, int out_length,
                             const int16_t *interp_filters, int interp_taps) { … }

static void interpolate(const uint8_t *const input, int in_length,
                        uint8_t *output, int out_length) { … }

int32_t av1_get_upscale_convolve_step(int in_length, int out_length) { … }

static int32_t get_upscale_convolve_x0(int in_length, int out_length,
                                       int32_t x_step_qn) { … }

void down2_symeven(const uint8_t *const input, int length, uint8_t *output,
                   int start_offset) { … }

static void down2_symodd(const uint8_t *const input, int length,
                         uint8_t *output) { … }

static int get_down2_length(int length, int steps) { … }

static int get_down2_steps(int in_length, int out_length) { … }

static void resize_multistep(const uint8_t *const input, int length,
                             uint8_t *output, int olength, uint8_t *otmp) { … }

static void fill_col_to_arr(uint8_t *img, int stride, int len, uint8_t *arr) { … }

static void fill_arr_to_col(uint8_t *img, int stride, int len, uint8_t *arr) { … }

bool av1_resize_vert_dir_c(uint8_t *intbuf, uint8_t *output, int out_stride,
                           int height, int height2, int width2, int start_col) { … }

void av1_resize_horz_dir_c(const uint8_t *const input, int in_stride,
                           uint8_t *intbuf, int height, int filtered_length,
                           int width2) { … }

bool av1_resize_plane_to_half(const uint8_t *const input, int height, int width,
                              int in_stride, uint8_t *output, int height2,
                              int width2, int out_stride) { … }

// Check if both the output width and height are half of input width and
// height respectively.
bool should_resize_by_half(int height, int width, int height2, int width2) { … }

bool av1_resize_plane(const uint8_t *input, int height, int width,
                      int in_stride, uint8_t *output, int height2, int width2,
                      int out_stride) { … }

static bool upscale_normative_rect(const uint8_t *const input, int height,
                                   int width, int in_stride, uint8_t *output,
                                   int height2, int width2, int out_stride,
                                   int x_step_qn, int x0_qn, int pad_left,
                                   int pad_right) { … }

#if CONFIG_AV1_HIGHBITDEPTH
static void highbd_interpolate_core(const uint16_t *const input, int in_length,
                                    uint16_t *output, int out_length, int bd,
                                    const int16_t *interp_filters,
                                    int interp_taps) {
  const int32_t delta =
      (((uint32_t)in_length << RS_SCALE_SUBPEL_BITS) + out_length / 2) /
      out_length;
  const int32_t offset =
      in_length > out_length
          ? (((int32_t)(in_length - out_length) << (RS_SCALE_SUBPEL_BITS - 1)) +
             out_length / 2) /
                out_length
          : -(((int32_t)(out_length - in_length)
               << (RS_SCALE_SUBPEL_BITS - 1)) +
              out_length / 2) /
                out_length;
  uint16_t *optr = output;
  int x, x1, x2, sum, k, int_pel, sub_pel;
  int32_t y;

  x = 0;
  y = offset + RS_SCALE_EXTRA_OFF;
  while ((y >> RS_SCALE_SUBPEL_BITS) < (interp_taps / 2 - 1)) {
    x++;
    y += delta;
  }
  x1 = x;
  x = out_length - 1;
  y = delta * x + offset + RS_SCALE_EXTRA_OFF;
  while ((y >> RS_SCALE_SUBPEL_BITS) + (int32_t)(interp_taps / 2) >=
         in_length) {
    x--;
    y -= delta;
  }
  x2 = x;
  if (x1 > x2) {
    for (x = 0, y = offset + RS_SCALE_EXTRA_OFF; x < out_length;
         ++x, y += delta) {
      int_pel = y >> RS_SCALE_SUBPEL_BITS;
      sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
      const int16_t *filter = &interp_filters[sub_pel * interp_taps];
      sum = 0;
      for (k = 0; k < interp_taps; ++k) {
        const int pk = int_pel - interp_taps / 2 + 1 + k;
        sum += filter[k] * input[AOMMAX(AOMMIN(pk, in_length - 1), 0)];
      }
      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
    }
  } else {
    // Initial part.
    for (x = 0, y = offset + RS_SCALE_EXTRA_OFF; x < x1; ++x, y += delta) {
      int_pel = y >> RS_SCALE_SUBPEL_BITS;
      sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
      const int16_t *filter = &interp_filters[sub_pel * interp_taps];
      sum = 0;
      for (k = 0; k < interp_taps; ++k)
        sum += filter[k] * input[AOMMAX(int_pel - interp_taps / 2 + 1 + k, 0)];
      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
    }
    // Middle part.
    for (; x <= x2; ++x, y += delta) {
      int_pel = y >> RS_SCALE_SUBPEL_BITS;
      sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
      const int16_t *filter = &interp_filters[sub_pel * interp_taps];
      sum = 0;
      for (k = 0; k < interp_taps; ++k)
        sum += filter[k] * input[int_pel - interp_taps / 2 + 1 + k];
      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
    }
    // End part.
    for (; x < out_length; ++x, y += delta) {
      int_pel = y >> RS_SCALE_SUBPEL_BITS;
      sub_pel = (y >> RS_SCALE_EXTRA_BITS) & RS_SUBPEL_MASK;
      const int16_t *filter = &interp_filters[sub_pel * interp_taps];
      sum = 0;
      for (k = 0; k < interp_taps; ++k)
        sum += filter[k] *
               input[AOMMIN(int_pel - interp_taps / 2 + 1 + k, in_length - 1)];
      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
    }
  }
}

static void highbd_interpolate(const uint16_t *const input, int in_length,
                               uint16_t *output, int out_length, int bd) {
  const InterpKernel *interp_filters =
      choose_interp_filter(in_length, out_length);

  highbd_interpolate_core(input, in_length, output, out_length, bd,
                          &interp_filters[0][0], SUBPEL_TAPS);
}

static void highbd_down2_symeven(const uint16_t *const input, int length,
                                 uint16_t *output, int bd) {
  // Actual filter len = 2 * filter_len_half.
  static const int16_t *filter = av1_down2_symeven_half_filter;
  const int filter_len_half = sizeof(av1_down2_symeven_half_filter) / 2;
  int i, j;
  uint16_t *optr = output;
  int l1 = filter_len_half;
  int l2 = (length - filter_len_half);
  l1 += (l1 & 1);
  l2 += (l2 & 1);
  if (l1 > l2) {
    // Short input length.
    for (i = 0; i < length; i += 2) {
      int sum = (1 << (FILTER_BITS - 1));
      for (j = 0; j < filter_len_half; ++j) {
        sum +=
            (input[AOMMAX(0, i - j)] + input[AOMMIN(i + 1 + j, length - 1)]) *
            filter[j];
      }
      sum >>= FILTER_BITS;
      *optr++ = clip_pixel_highbd(sum, bd);
    }
  } else {
    // Initial part.
    for (i = 0; i < l1; i += 2) {
      int sum = (1 << (FILTER_BITS - 1));
      for (j = 0; j < filter_len_half; ++j) {
        sum += (input[AOMMAX(0, i - j)] + input[i + 1 + j]) * filter[j];
      }
      sum >>= FILTER_BITS;
      *optr++ = clip_pixel_highbd(sum, bd);
    }
    // Middle part.
    for (; i < l2; i += 2) {
      int sum = (1 << (FILTER_BITS - 1));
      for (j = 0; j < filter_len_half; ++j) {
        sum += (input[i - j] + input[i + 1 + j]) * filter[j];
      }
      sum >>= FILTER_BITS;
      *optr++ = clip_pixel_highbd(sum, bd);
    }
    // End part.
    for (; i < length; i += 2) {
      int sum = (1 << (FILTER_BITS - 1));
      for (j = 0; j < filter_len_half; ++j) {
        sum +=
            (input[i - j] + input[AOMMIN(i + 1 + j, length - 1)]) * filter[j];
      }
      sum >>= FILTER_BITS;
      *optr++ = clip_pixel_highbd(sum, bd);
    }
  }
}

static void highbd_down2_symodd(const uint16_t *const input, int length,
                                uint16_t *output, int bd) {
  // Actual filter len = 2 * filter_len_half - 1.
  static const int16_t *filter = av1_down2_symodd_half_filter;
  const int filter_len_half = sizeof(av1_down2_symodd_half_filter) / 2;
  int i, j;
  uint16_t *optr = output;
  int l1 = filter_len_half - 1;
  int l2 = (length - filter_len_half + 1);
  l1 += (l1 & 1);
  l2 += (l2 & 1);
  if (l1 > l2) {
    // Short input length.
    for (i = 0; i < length; i += 2) {
      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
      for (j = 1; j < filter_len_half; ++j) {
        sum += (input[AOMMAX(i - j, 0)] + input[AOMMIN(i + j, length - 1)]) *
               filter[j];
      }
      sum >>= FILTER_BITS;
      *optr++ = clip_pixel_highbd(sum, bd);
    }
  } else {
    // Initial part.
    for (i = 0; i < l1; i += 2) {
      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
      for (j = 1; j < filter_len_half; ++j) {
        sum += (input[AOMMAX(i - j, 0)] + input[i + j]) * filter[j];
      }
      sum >>= FILTER_BITS;
      *optr++ = clip_pixel_highbd(sum, bd);
    }
    // Middle part.
    for (; i < l2; i += 2) {
      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
      for (j = 1; j < filter_len_half; ++j) {
        sum += (input[i - j] + input[i + j]) * filter[j];
      }
      sum >>= FILTER_BITS;
      *optr++ = clip_pixel_highbd(sum, bd);
    }
    // End part.
    for (; i < length; i += 2) {
      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
      for (j = 1; j < filter_len_half; ++j) {
        sum += (input[i - j] + input[AOMMIN(i + j, length - 1)]) * filter[j];
      }
      sum >>= FILTER_BITS;
      *optr++ = clip_pixel_highbd(sum, bd);
    }
  }
}

static void highbd_resize_multistep(const uint16_t *const input, int length,
                                    uint16_t *output, int olength,
                                    uint16_t *otmp, int bd) {
  if (length == olength) {
    memcpy(output, input, sizeof(output[0]) * length);
    return;
  }
  const int steps = get_down2_steps(length, olength);

  if (steps > 0) {
    uint16_t *out = NULL;
    int filteredlength = length;

    assert(otmp != NULL);
    uint16_t *otmp2 = otmp + get_down2_length(length, 1);
    for (int s = 0; s < steps; ++s) {
      const int proj_filteredlength = get_down2_length(filteredlength, 1);
      const uint16_t *const in = (s == 0 ? input : out);
      if (s == steps - 1 && proj_filteredlength == olength)
        out = output;
      else
        out = (s & 1 ? otmp2 : otmp);
      if (filteredlength & 1)
        highbd_down2_symodd(in, filteredlength, out, bd);
      else
        highbd_down2_symeven(in, filteredlength, out, bd);
      filteredlength = proj_filteredlength;
    }
    if (filteredlength != olength) {
      highbd_interpolate(out, filteredlength, output, olength, bd);
    }
  } else {
    highbd_interpolate(input, length, output, olength, bd);
  }
}

static void highbd_fill_col_to_arr(uint16_t *img, int stride, int len,
                                   uint16_t *arr) {
  int i;
  uint16_t *iptr = img;
  uint16_t *aptr = arr;
  for (i = 0; i < len; ++i, iptr += stride) {
    *aptr++ = *iptr;
  }
}

static void highbd_fill_arr_to_col(uint16_t *img, int stride, int len,
                                   uint16_t *arr) {
  int i;
  uint16_t *iptr = img;
  uint16_t *aptr = arr;
  for (i = 0; i < len; ++i, iptr += stride) {
    *iptr = *aptr++;
  }
}

void av1_highbd_resize_plane(const uint8_t *input, int height, int width,
                             int in_stride, uint8_t *output, int height2,
                             int width2, int out_stride, int bd) {
  int i;
  uint16_t *intbuf = (uint16_t *)aom_malloc(sizeof(uint16_t) * width2 * height);
  uint16_t *tmpbuf =
      (uint16_t *)aom_malloc(sizeof(uint16_t) * AOMMAX(width, height));
  uint16_t *arrbuf = (uint16_t *)aom_malloc(sizeof(uint16_t) * height);
  uint16_t *arrbuf2 = (uint16_t *)aom_malloc(sizeof(uint16_t) * height2);
  if (intbuf == NULL || tmpbuf == NULL || arrbuf == NULL || arrbuf2 == NULL)
    goto Error;
  for (i = 0; i < height; ++i) {
    highbd_resize_multistep(CONVERT_TO_SHORTPTR(input + in_stride * i), width,
                            intbuf + width2 * i, width2, tmpbuf, bd);
  }
  for (i = 0; i < width2; ++i) {
    highbd_fill_col_to_arr(intbuf + i, width2, height, arrbuf);
    highbd_resize_multistep(arrbuf, height, arrbuf2, height2, tmpbuf, bd);
    highbd_fill_arr_to_col(CONVERT_TO_SHORTPTR(output + i), out_stride, height2,
                           arrbuf2);
  }

Error:
  aom_free(intbuf);
  aom_free(tmpbuf);
  aom_free(arrbuf);
  aom_free(arrbuf2);
}

static bool highbd_upscale_normative_rect(const uint8_t *const input,
                                          int height, int width, int in_stride,
                                          uint8_t *output, int height2,
                                          int width2, int out_stride,
                                          int x_step_qn, int x0_qn,
                                          int pad_left, int pad_right, int bd) {
  assert(width > 0);
  assert(height > 0);
  assert(width2 > 0);
  assert(height2 > 0);
  assert(height2 == height);

  // Extend the left/right pixels of the tile column if needed
  // (either because we can't sample from other tiles, or because we're at
  // a frame edge).
  // Save the overwritten pixels into tmp_left and tmp_right.
  // Note: Because we pass input-1 to av1_convolve_horiz_rs, we need one extra
  // column of border pixels compared to what we'd naively think.
  const int border_cols = UPSCALE_NORMATIVE_TAPS / 2 + 1;
  const int border_size = border_cols * sizeof(uint16_t);
  uint16_t *tmp_left =
      NULL;  // Silence spurious "may be used uninitialized" warnings
  uint16_t *tmp_right = NULL;
  uint16_t *const input16 = CONVERT_TO_SHORTPTR(input);
  uint16_t *const in_tl = input16 - border_cols;
  uint16_t *const in_tr = input16 + width;
  if (pad_left) {
    tmp_left = (uint16_t *)aom_malloc(sizeof(*tmp_left) * border_cols * height);
    if (!tmp_left) return false;
    for (int i = 0; i < height; i++) {
      memcpy(tmp_left + i * border_cols, in_tl + i * in_stride, border_size);
      aom_memset16(in_tl + i * in_stride, input16[i * in_stride], border_cols);
    }
  }
  if (pad_right) {
    tmp_right =
        (uint16_t *)aom_malloc(sizeof(*tmp_right) * border_cols * height);
    if (!tmp_right) {
      aom_free(tmp_left);
      return false;
    }
    for (int i = 0; i < height; i++) {
      memcpy(tmp_right + i * border_cols, in_tr + i * in_stride, border_size);
      aom_memset16(in_tr + i * in_stride, input16[i * in_stride + width - 1],
                   border_cols);
    }
  }

  av1_highbd_convolve_horiz_rs(CONVERT_TO_SHORTPTR(input - 1), in_stride,
                               CONVERT_TO_SHORTPTR(output), out_stride, width2,
                               height2, &av1_resize_filter_normative[0][0],
                               x0_qn, x_step_qn, bd);

  // Restore the left/right border pixels
  if (pad_left) {
    for (int i = 0; i < height; i++) {
      memcpy(in_tl + i * in_stride, tmp_left + i * border_cols, border_size);
    }
    aom_free(tmp_left);
  }
  if (pad_right) {
    for (int i = 0; i < height; i++) {
      memcpy(in_tr + i * in_stride, tmp_right + i * border_cols, border_size);
    }
    aom_free(tmp_right);
  }
  return true;
}
#endif  // CONFIG_AV1_HIGHBITDEPTH

void av1_resize_frame420(const uint8_t *y, int y_stride, const uint8_t *u,
                         const uint8_t *v, int uv_stride, int height, int width,
                         uint8_t *oy, int oy_stride, uint8_t *ou, uint8_t *ov,
                         int ouv_stride, int oheight, int owidth) { … }

void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
                                   YV12_BUFFER_CONFIG *dst,
                                   const InterpFilter filter,
                                   const int phase_scaler,
                                   const int num_planes) { … }

bool av1_resize_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
                                              YV12_BUFFER_CONFIG *dst, int bd,
                                              int num_planes) { … }

void av1_upscale_normative_rows(const AV1_COMMON *cm, const uint8_t *src,
                                int src_stride, uint8_t *dst, int dst_stride,
                                int plane, int rows) { … }

void av1_upscale_normative_and_extend_frame(const AV1_COMMON *cm,
                                            const YV12_BUFFER_CONFIG *src,
                                            YV12_BUFFER_CONFIG *dst) { … }

YV12_BUFFER_CONFIG *av1_realloc_and_scale_if_required(
    AV1_COMMON *cm, YV12_BUFFER_CONFIG *unscaled, YV12_BUFFER_CONFIG *scaled,
    const InterpFilter filter, const int phase, const bool use_optimized_scaler,
    const bool for_psnr, const int border_in_pixels, const bool alloc_pyramid) { … }

// Calculates the scaled dimension given the original dimension and the scale
// denominator.
static void calculate_scaled_size_helper(int *dim, int denom) { … }

void av1_calculate_scaled_size(int *width, int *height, int resize_denom) { … }

void av1_calculate_scaled_superres_size(int *width, int *height,
                                        int superres_denom) { … }

void av1_calculate_unscaled_superres_size(int *width, int *height, int denom) { … }

// Copy only the config data from 'src' to 'dst'.
static void copy_buffer_config(const YV12_BUFFER_CONFIG *const src,
                               YV12_BUFFER_CONFIG *const dst) { … }

// TODO(afergs): Look for in-place upscaling
// TODO(afergs): aom_ vs av1_ functions? Which can I use?
// Upscale decoded image.
void av1_superres_upscale(AV1_COMMON *cm, BufferPool *const pool,
                          bool alloc_pyramid) { … }