// Copyright 2022 Google Inc. All Rights Reserved.

//

// Use of this source code is governed by a BSD-style license

// that can be found in the COPYING 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.

// -----------------------------------------------------------------------------

//

// Sharp RGB to YUV conversion.

//

// Author: Skal (pascal.massimino@gmail.com)

#include "./sharpyuv.h"

#include <assert.h>

#include <limits.h>

#include <stddef.h>

#include <stdlib.h>

#include <string.h>

#include "./sharpyuv_cpu.h"

#include "./sharpyuv_dsp.h"

#include "./sharpyuv_gamma.h"

#include "webp/types.h"

//------------------------------------------------------------------------------

int SharpYuvGetVersion(void) { return SHARPYUV_VERSION; }

//------------------------------------------------------------------------------

// Sharp RGB->YUV conversion

static const int kNumIterations = 4;

// Max bit depth so that intermediate calculations fit in 16 bits.

static const int kMaxBitDepth = 14;

// Returns the precision shift to use based on the input rgb_bit_depth.

static int GetPrecisionShift(int rgb_bit_depth) {

  // Try to add 2 bits of precision if it fits in kMaxBitDepth. Otherwise remove

  // bits if needed.

  return ((rgb_bit_depth + 2) <= kMaxBitDepth) ? 2

                                               : (kMaxBitDepth - rgb_bit_depth);

}

typedef int16_t fixed_t;     // signed type with extra precision for UV

typedef uint16_t fixed_y_t;  // unsigned type with extra precision for W

//------------------------------------------------------------------------------

static fixed_y_t clip_bit_depth(int y, int bit_depth) {

  const int max = (1 << bit_depth) - 1;

  return (!(y & ~max)) ? (fixed_y_t)y : (y < 0) ? 0 : max;

}

//------------------------------------------------------------------------------

static uint32_t ScaleDown(uint16_t a, uint16_t b, uint16_t c, uint16_t d,

                          int bit_depth,

                          SharpYuvTransferFunctionType transfer_type) {

  const uint32_t A = SharpYuvGammaToLinear(a, bit_depth, transfer_type);

  const uint32_t B = SharpYuvGammaToLinear(b, bit_depth, transfer_type);

  const uint32_t C = SharpYuvGammaToLinear(c, bit_depth, transfer_type);

  const uint32_t D = SharpYuvGammaToLinear(d, bit_depth, transfer_type);

  return SharpYuvLinearToGamma((A + B + C + D + 2) >> 2, bit_depth,

                               transfer_type);

}

static WEBP_INLINE void UpdateW(const fixed_y_t* src, fixed_y_t* dst, int w,

                                int bit_depth,

                                SharpYuvTransferFunctionType transfer_type) {

  int i = 0;

  if (transfer_type == kSharpYuvTransferFunctionSrgb) {

    SharpYuvUpdateWSrgb(src, dst, w, bit_depth);

    return;

  }

  do {

    const uint32_t R =

        SharpYuvGammaToLinear(src[0 * w + i], bit_depth, transfer_type);

    const uint32_t G =

        SharpYuvGammaToLinear(src[1 * w + i], bit_depth, transfer_type);

    const uint32_t B =

        SharpYuvGammaToLinear(src[2 * w + i], bit_depth, transfer_type);

    const uint32_t Y = SharpYuvRGBToGray(R, G, B);

    dst[i] = (fixed_y_t)SharpYuvLinearToGamma(Y, bit_depth, transfer_type);

  } while (++i < w);

}

static void UpdateChroma(const fixed_y_t* src1, const fixed_y_t* src2,

                         fixed_t* dst, int uv_w, int bit_depth,

                         SharpYuvTransferFunctionType transfer_type) {

  int i = 0;

  if (transfer_type == kSharpYuvTransferFunctionSrgb) {

    SharpYuvUpdateChromaSrgb(src1, src2, dst, uv_w, bit_depth);

    return;

  }

  do {

    const int r =

        ScaleDown(src1[0 * uv_w + 0], src1[0 * uv_w + 1], src2[0 * uv_w + 0],

                  src2[0 * uv_w + 1], bit_depth, transfer_type);

    const int g =

        ScaleDown(src1[2 * uv_w + 0], src1[2 * uv_w + 1], src2[2 * uv_w + 0],

                  src2[2 * uv_w + 1], bit_depth, transfer_type);

    const int b =

        ScaleDown(src1[4 * uv_w + 0], src1[4 * uv_w + 1], src2[4 * uv_w + 0],

                  src2[4 * uv_w + 1], bit_depth, transfer_type);

    const int W = SharpYuvRGBToGray(r, g, b);

    dst[0 * uv_w] = (fixed_t)(r - W);

    dst[1 * uv_w] = (fixed_t)(g - W);

    dst[2 * uv_w] = (fixed_t)(b - W);

    dst += 1;

    src1 += 2;

    src2 += 2;

  } while (++i < uv_w);

}

static void StoreGray(const fixed_y_t* rgb, fixed_y_t* y, int w) {

  int i = 0;

  assert(w > 0);

  do {

    y[i] = SharpYuvRGBToGray(rgb[0 * w + i], rgb[1 * w + i], rgb[2 * w + i]);

  } while (++i < w);

}

//------------------------------------------------------------------------------

static WEBP_INLINE fixed_y_t Filter2(int A, int B, int W0, int bit_depth) {

  const int v0 = (A * 3 + B + 2) >> 2;

  return clip_bit_depth(v0 + W0, bit_depth);

}

//------------------------------------------------------------------------------

static WEBP_INLINE int Shift(int v, int shift) {

  return (shift >= 0) ? (v << shift) : (v >> -shift);

}

static void ImportOneRow(const uint8_t* const r_ptr, const uint8_t* const g_ptr,

                         const uint8_t* const b_ptr, int rgb_step,

                         int rgb_bit_depth, int pic_width,

                         fixed_y_t* const dst) {

  // Convert the rgb_step from a number of bytes to a number of uint8_t or

  // uint16_t values depending the bit depth.

  const int step = (rgb_bit_depth > 8) ? rgb_step / 2 : rgb_step;

  const int w = (pic_width + 1) & ~1;

  const int shift = GetPrecisionShift(rgb_bit_depth);

  const int max_val = (1 << rgb_bit_depth) - 1;

  int i = 0;

  if (rgb_bit_depth == 8) {

    do {

      const int off = i * step;

      dst[i + 0 * w] = Shift(r_ptr[off], shift);

      dst[i + 1 * w] = Shift(g_ptr[off], shift);

      dst[i + 2 * w] = Shift(b_ptr[off], shift);

    } while (++i < pic_width);

  } else if (rgb_bit_depth < 16) {

    do {

      const int off = i * step;

      int r = ((const uint16_t*)r_ptr)[off];

      int g = ((const uint16_t*)g_ptr)[off];

      int b = ((const uint16_t*)b_ptr)[off];

      dst[i + 0 * w] = Shift(r > max_val ? max_val : r, shift);

      dst[i + 1 * w] = Shift(g > max_val ? max_val : g, shift);

      dst[i + 2 * w] = Shift(b > max_val ? max_val : b, shift);

    } while (++i < pic_width);

  } else {  // rgb_bit_depth == 16

    do {

      const int off = i * step;

      int r = ((const uint16_t*)r_ptr)[off];

      int g = ((const uint16_t*)g_ptr)[off];

      int b = ((const uint16_t*)b_ptr)[off];

      dst[i + 0 * w] = Shift(r, shift);

      dst[i + 1 * w] = Shift(g, shift);

      dst[i + 2 * w] = Shift(b, shift);

    } while (++i < pic_width);

  }

  if (pic_width & 1) {  // replicate rightmost pixel

    dst[pic_width + 0 * w] = dst[pic_width + 0 * w - 1];

    dst[pic_width + 1 * w] = dst[pic_width + 1 * w - 1];

    dst[pic_width + 2 * w] = dst[pic_width + 2 * w - 1];

  }

}

static void InterpolateTwoRows(const fixed_y_t* const best_y,

                               const fixed_t* prev_uv, const fixed_t* cur_uv,

                               const fixed_t* next_uv, int w, fixed_y_t* out1,

                               fixed_y_t* out2, int bit_depth) {

  const int uv_w = w >> 1;

  const int len = (w - 1) >> 1;  // length to filter

  int k = 3;

  while (k-- > 0) {  // process each R/G/B segments in turn

    // special boundary case for i==0

    out1[0] = Filter2(cur_uv[0], prev_uv[0], best_y[0], bit_depth);

    out2[0] = Filter2(cur_uv[0], next_uv[0], best_y[w], bit_depth);

    SharpYuvFilterRow(cur_uv, prev_uv, len, best_y + 0 + 1, out1 + 1,

                      bit_depth);

    SharpYuvFilterRow(cur_uv, next_uv, len, best_y + w + 1, out2 + 1,

                      bit_depth);

    // special boundary case for i == w - 1 when w is even

    if (!(w & 1)) {

      out1[w - 1] = Filter2(cur_uv[uv_w - 1], prev_uv[uv_w - 1],

                            best_y[w - 1 + 0], bit_depth);

      out2[w - 1] = Filter2(cur_uv[uv_w - 1], next_uv[uv_w - 1],

                            best_y[w - 1 + w], bit_depth);

    }

    out1 += w;

    out2 += w;

    prev_uv += uv_w;

    cur_uv += uv_w;

    next_uv += uv_w;

  }

}

static int ConvertWRGBToYUV(const fixed_y_t* best_y, const fixed_t* best_uv,

                            uint8_t* y_ptr, int y_stride, uint8_t* u_ptr,

                            int u_stride, uint8_t* v_ptr, int v_stride,

                            int rgb_bit_depth, int yuv_bit_depth, int width,

                            int height,

                            const SharpYuvConversionMatrix* yuv_matrix) {

  int j;

  const fixed_t* const best_uv_base = best_uv;

  const int w = (width + 1) & ~1;

  const int h = (height + 1) & ~1;

  const int uv_w = w >> 1;

  const int uv_h = h >> 1;

  const int sfix = GetPrecisionShift(rgb_bit_depth);

  best_uv = best_uv_base;

  j = 0;

  do {

    SharpYuvConvertRowY(best_y, best_uv, width, uv_w, yuv_matrix->rgb_to_y,

                        sfix, yuv_bit_depth, y_ptr);

    best_y += w;

    best_uv += (j & 1) * 3 * uv_w;

    y_ptr += y_stride;

  } while (++j < height);

  best_uv = best_uv_base;

  j = 0;

  do {

    // Note r, g and b values here are off by W, but a constant offset on all

    // 3 components doesn't change the value of u and v with a YCbCr matrix.

    SharpYuvConvertRowUV(best_uv, uv_w, yuv_matrix->rgb_to_u,

                         yuv_matrix->rgb_to_v, sfix, yuv_bit_depth, u_ptr,

                         v_ptr);

    best_uv += 3 * uv_w;

    u_ptr += u_stride;

    v_ptr += v_stride;

  } while (++j < uv_h);

  return 1;

}

//------------------------------------------------------------------------------

// Main function

static void* SafeMalloc(uint64_t nmemb, size_t size) {

  const uint64_t total_size = nmemb * (uint64_t)size;

  if (total_size != (size_t)total_size) return NULL;

  return malloc((size_t)total_size);

}

static int DoSharpArgbToYuv(const uint8_t* r_ptr, const uint8_t* g_ptr,

                            const uint8_t* b_ptr, int rgb_step, int rgb_stride,

                            int rgb_bit_depth, uint8_t* y_ptr, int y_stride,

                            uint8_t* u_ptr, int u_stride, uint8_t* v_ptr,

                            int v_stride, int yuv_bit_depth, int width,

                            int height,

                            const SharpYuvConversionMatrix* yuv_matrix,

                            SharpYuvTransferFunctionType transfer_type) {

  // we expand the right/bottom border if needed

  const int w = (width + 1) & ~1;

  const int h = (height + 1) & ~1;

  const int uv_w = w >> 1;

  const int uv_h = h >> 1;

  const int y_bit_depth = rgb_bit_depth + GetPrecisionShift(rgb_bit_depth);

  uint64_t prev_diff_y_sum = ~0;

  int j, iter;

  const uint64_t tmp_buffer_size = (uint64_t)w * 3 * 2;

  const uint64_t best_y_base_size = (uint64_t)w * h;

  const uint64_t target_y_base_size = (uint64_t)w * h;

  const uint64_t best_rgb_y_size = (uint64_t)w * 2;

  const uint64_t best_uv_base_size = (uint64_t)uv_w * 3 * uv_h;

  const uint64_t target_uv_base_size = (uint64_t)uv_w * 3 * uv_h;

  const uint64_t best_rgb_uv_size = (uint64_t)uv_w * 3;

  fixed_y_t* const tmp_buffer = (fixed_y_t*)SafeMalloc(

      (tmp_buffer_size + best_y_base_size + target_y_base_size +

       best_rgb_y_size) +

          (best_uv_base_size + target_uv_base_size + best_rgb_uv_size),

      sizeof(*tmp_buffer));

  fixed_y_t *best_y_base, *target_y_base, *best_rgb_y;

  fixed_t *best_uv_base, *target_uv_base, *best_rgb_uv;

  fixed_y_t *best_y, *target_y;

  fixed_t *best_uv, *target_uv;

  const uint64_t diff_y_threshold = (uint64_t)(3.0 * w * h);

  int ok;

  assert(w > 0);

  assert(h > 0);

  assert(sizeof(fixed_y_t) == sizeof(fixed_t));

  if (tmp_buffer == NULL) {

    ok = 0;

    goto End;

  }

  best_y_base = tmp_buffer + tmp_buffer_size;

  target_y_base = best_y_base + best_y_base_size;

  best_rgb_y = target_y_base + target_y_base_size;

  best_uv_base = (fixed_t*)(best_rgb_y + best_rgb_y_size);

  target_uv_base = best_uv_base + best_uv_base_size;

  best_rgb_uv = target_uv_base + target_uv_base_size;

  best_y = best_y_base;

  target_y = target_y_base;

  best_uv = best_uv_base;

  target_uv = target_uv_base;

  // Import RGB samples to W/RGB representation.

  for (j = 0; j < height; j += 2) {

    const int is_last_row = (j == height - 1);

    fixed_y_t* const src1 = tmp_buffer + 0 * w;

    fixed_y_t* const src2 = tmp_buffer + 3 * w;

    // prepare two rows of input

    ImportOneRow(r_ptr, g_ptr, b_ptr, rgb_step, rgb_bit_depth, width, src1);

    if (!is_last_row) {

      ImportOneRow(r_ptr + rgb_stride, g_ptr + rgb_stride, b_ptr + rgb_stride,

                   rgb_step, rgb_bit_depth, width, src2);

    } else {

      memcpy(src2, src1, 3 * w * sizeof(*src2));

    }

    StoreGray(src1, best_y + 0, w);

    StoreGray(src2, best_y + w, w);

    UpdateW(src1, target_y, w, y_bit_depth, transfer_type);

    UpdateW(src2, target_y + w, w, y_bit_depth, transfer_type);

    UpdateChroma(src1, src2, target_uv, uv_w, y_bit_depth, transfer_type);

    memcpy(best_uv, target_uv, 3 * uv_w * sizeof(*best_uv));

    best_y += 2 * w;

    best_uv += 3 * uv_w;

    target_y += 2 * w;

    target_uv += 3 * uv_w;

    r_ptr += 2 * rgb_stride;

    g_ptr += 2 * rgb_stride;

    b_ptr += 2 * rgb_stride;

  }

  // Iterate and resolve clipping conflicts.

  for (iter = 0; iter < kNumIterations; ++iter) {

    const fixed_t* cur_uv = best_uv_base;

    const fixed_t* prev_uv = best_uv_base;

    uint64_t diff_y_sum = 0;

    best_y = best_y_base;

    best_uv = best_uv_base;

    target_y = target_y_base;

    target_uv = target_uv_base;

    j = 0;

    do {

      fixed_y_t* const src1 = tmp_buffer + 0 * w;

      fixed_y_t* const src2 = tmp_buffer + 3 * w;

      {

        const fixed_t* const next_uv = cur_uv + ((j < h - 2) ? 3 * uv_w : 0);

        InterpolateTwoRows(best_y, prev_uv, cur_uv, next_uv, w, src1, src2,

                           y_bit_depth);

        prev_uv = cur_uv;

        cur_uv = next_uv;

      }

      UpdateW(src1, best_rgb_y + 0 * w, w, y_bit_depth, transfer_type);

      UpdateW(src2, best_rgb_y + 1 * w, w, y_bit_depth, transfer_type);

      UpdateChroma(src1, src2, best_rgb_uv, uv_w, y_bit_depth, transfer_type);

      // update two rows of Y and one row of RGB

      diff_y_sum +=

          SharpYuvUpdateY(target_y, best_rgb_y, best_y, 2 * w, y_bit_depth);

      SharpYuvUpdateRGB(target_uv, best_rgb_uv, best_uv, 3 * uv_w);

      best_y += 2 * w;

      best_uv += 3 * uv_w;

      target_y += 2 * w;

      target_uv += 3 * uv_w;

      j += 2;

    } while (j < h);

    // test exit condition

    if (diff_y_sum < diff_y_threshold) break;

    if (iter > 0 && diff_y_sum > prev_diff_y_sum) break;

    prev_diff_y_sum = diff_y_sum;

  }

  // final reconstruction

  ok = ConvertWRGBToYUV(best_y_base, best_uv_base, y_ptr, y_stride, u_ptr,

                        u_stride, v_ptr, v_stride, rgb_bit_depth, yuv_bit_depth,

                        width, height, yuv_matrix);

End:

  free(tmp_buffer);

  return ok;

}

#if defined(WEBP_USE_THREAD) && !defined(_WIN32)

#include <pthread.h>  // NOLINT

#define LOCK_ACCESS                                                 \

  static pthread_mutex_t sharpyuv_lock = PTHREAD_MUTEX_INITIALIZER; \

  if (pthread_mutex_lock(&sharpyuv_lock)) return

#define UNLOCK_ACCESS_AND_RETURN                \

  do {                                          \

    (void)pthread_mutex_unlock(&sharpyuv_lock); \

    return;                                     \

  } while (0)

#else  // !(defined(WEBP_USE_THREAD) && !defined(_WIN32))

#define LOCK_ACCESS \

  do {              \

  } while (0)

#define UNLOCK_ACCESS_AND_RETURN return

#endif  // defined(WEBP_USE_THREAD) && !defined(_WIN32)

// Hidden exported init function.

// By default SharpYuvConvert calls it with SharpYuvGetCPUInfo. If needed,

// users can declare it as extern and call it with an alternate VP8CPUInfo

// function.

extern VP8CPUInfo SharpYuvGetCPUInfo;

SHARPYUV_EXTERN void SharpYuvInit(VP8CPUInfo cpu_info_func);

void SharpYuvInit(VP8CPUInfo cpu_info_func) {

  static volatile VP8CPUInfo sharpyuv_last_cpuinfo_used =

      (VP8CPUInfo)&sharpyuv_last_cpuinfo_used;

  LOCK_ACCESS;

  // Only update SharpYuvGetCPUInfo when called from external code to avoid a

  // race on reading the value in SharpYuvConvert().

  if (cpu_info_func != (VP8CPUInfo)&SharpYuvGetCPUInfo) {

    SharpYuvGetCPUInfo = cpu_info_func;

  }

  if (sharpyuv_last_cpuinfo_used == SharpYuvGetCPUInfo) {

    UNLOCK_ACCESS_AND_RETURN;

  }

  SharpYuvInitDsp();

  SharpYuvInitGammaTables();

  sharpyuv_last_cpuinfo_used = SharpYuvGetCPUInfo;

  UNLOCK_ACCESS_AND_RETURN;

}

int SharpYuvConvert(const void* r_ptr, const void* g_ptr, const void* b_ptr,

                    int rgb_step, int rgb_stride, int rgb_bit_depth,

                    void* y_ptr, int y_stride, void* u_ptr, int u_stride,

                    void* v_ptr, int v_stride, int yuv_bit_depth, int width,

                    int height, const SharpYuvConversionMatrix* yuv_matrix) {

  SharpYuvOptions options;

  options.yuv_matrix = yuv_matrix;

  options.transfer_type = kSharpYuvTransferFunctionSrgb;

  return SharpYuvConvertWithOptions(

      r_ptr, g_ptr, b_ptr, rgb_step, rgb_stride, rgb_bit_depth, y_ptr, y_stride,

      u_ptr, u_stride, v_ptr, v_stride, yuv_bit_depth, width, height, &options);

}

int SharpYuvOptionsInitInternal(const SharpYuvConversionMatrix* yuv_matrix,

                                SharpYuvOptions* options, int version) {

  const int major = (version >> 24);

  const int minor = (version >> 16) & 0xff;

  if (options == NULL || yuv_matrix == NULL ||

      (major == SHARPYUV_VERSION_MAJOR && major == 0 &&

       minor != SHARPYUV_VERSION_MINOR) ||

      (major != SHARPYUV_VERSION_MAJOR)) {

    return 0;

  }

  options->yuv_matrix = yuv_matrix;

  options->transfer_type = kSharpYuvTransferFunctionSrgb;

  return 1;

}

int SharpYuvConvertWithOptions(const void* r_ptr, const void* g_ptr,

                               const void* b_ptr, int rgb_step, int rgb_stride,

                               int rgb_bit_depth, void* y_ptr, int y_stride,

                               void* u_ptr, int u_stride, void* v_ptr,

                               int v_stride, int yuv_bit_depth, int width,

                               int height, const SharpYuvOptions* options) {

  const SharpYuvConversionMatrix* yuv_matrix = options->yuv_matrix;

  SharpYuvTransferFunctionType transfer_type = options->transfer_type;

  SharpYuvConversionMatrix scaled_matrix;

  const int rgb_max = (1 << rgb_bit_depth) - 1;

  const int rgb_round = 1 << (rgb_bit_depth - 1);

  const int yuv_max = (1 << yuv_bit_depth) - 1;

  const int sfix = GetPrecisionShift(rgb_bit_depth);

  if (width < 1 || height < 1 || width == INT_MAX || height == INT_MAX ||

      r_ptr == NULL || g_ptr == NULL || b_ptr == NULL || y_ptr == NULL ||

      u_ptr == NULL || v_ptr == NULL) {

    return 0;

  }

  if (rgb_bit_depth != 8 && rgb_bit_depth != 10 && rgb_bit_depth != 12 &&

      rgb_bit_depth != 16) {

    return 0;

  }

  if (yuv_bit_depth != 8 && yuv_bit_depth != 10 && yuv_bit_depth != 12) {

    return 0;

  }

  if (rgb_bit_depth > 8 && (rgb_step % 2 != 0 || rgb_stride % 2 != 0)) {

    // Step/stride should be even for uint16_t buffers.

    return 0;

  }

  {

    const uint64_t yuv_bytes = (yuv_bit_depth > 8) ? 2 : 1;

    const uint64_t uv_width = (width + 1) / 2;

    const uint64_t abs_step =

        (uint64_t)((rgb_step < 0) ? -(int64_t)rgb_step : (int64_t)rgb_step);

    const uint64_t abs_stride =

        (uint64_t)((rgb_stride < 0) ? -(int64_t)rgb_stride

                                    : (int64_t)rgb_stride);

    const uint64_t total_rgb_size = (uint64_t)height * abs_stride;

    const uint64_t uv_height = (height + 1) / 2;

    const uint64_t total_y_size = (uint64_t)height * y_stride;

    const uint64_t total_u_size = uv_height * u_stride;

    const uint64_t total_v_size = uv_height * v_stride;

    if (y_stride < 0 || (uint64_t)y_stride < (uint64_t)width * yuv_bytes ||

        u_stride < 0 || (uint64_t)u_stride < uv_width * yuv_bytes ||

        v_stride < 0 || (uint64_t)v_stride < uv_width * yuv_bytes) {

      return 0;

    }

    if (abs_step == 0 || abs_stride < (uint64_t)width * abs_step) {

      return 0;

    }

    if (total_rgb_size != (size_t)total_rgb_size ||

        total_y_size != (size_t)total_y_size ||

        total_u_size != (size_t)total_u_size ||

        total_v_size != (size_t)total_v_size) {

      return 0;

    }

  }

  if (yuv_bit_depth > 8 &&

      (y_stride % 2 != 0 || u_stride % 2 != 0 || v_stride % 2 != 0)) {

    // Stride should be even for uint16_t buffers.

    return 0;

  }

  // The address of the function pointer is used to avoid a read race.

  SharpYuvInit((VP8CPUInfo)&SharpYuvGetCPUInfo);

  // Add scaling factor to go from rgb_bit_depth to yuv_bit_depth, to the

  // rgb->yuv conversion matrix.

  if (rgb_bit_depth == yuv_bit_depth) {

    memcpy(&scaled_matrix, yuv_matrix, sizeof(scaled_matrix));

  } else {

    int i;

    for (i = 0; i < 3; ++i) {

      scaled_matrix.rgb_to_y[i] =

          (yuv_matrix->rgb_to_y[i] * yuv_max + rgb_round) / rgb_max;

      scaled_matrix.rgb_to_u[i] =

          (yuv_matrix->rgb_to_u[i] * yuv_max + rgb_round) / rgb_max;

      scaled_matrix.rgb_to_v[i] =

          (yuv_matrix->rgb_to_v[i] * yuv_max + rgb_round) / rgb_max;

    }

  }

  // Also incorporate precision change scaling.

  scaled_matrix.rgb_to_y[3] = Shift(yuv_matrix->rgb_to_y[3], sfix);

  scaled_matrix.rgb_to_u[3] = Shift(yuv_matrix->rgb_to_u[3], sfix);

  scaled_matrix.rgb_to_v[3] = Shift(yuv_matrix->rgb_to_v[3], sfix);

  return DoSharpArgbToYuv(

      (const uint8_t*)r_ptr, (const uint8_t*)g_ptr, (const uint8_t*)b_ptr,

      rgb_step, rgb_stride, rgb_bit_depth, (uint8_t*)y_ptr, y_stride,

      (uint8_t*)u_ptr, u_stride, (uint8_t*)v_ptr, v_stride, yuv_bit_depth,

      width, height, &scaled_matrix, transfer_type);

}

//------------------------------------------------------------------------------