src/cobalt/media/base/simd/convert_rgb_to_yuv_sse2.cc

// Copyright (c) 2012 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

#include "build/build_config.h"
#include "media/base/simd/convert_rgb_to_yuv.h"

#if defined(COMPILER_MSVC)
#include <intrin.h>
#else
#include <mmintrin.h>
#include <emmintrin.h>

#include "starboard/types.h"
#endif

#if defined(COMPILER_MSVC)
#define SIMD_ALIGNED(var) __declspec(align(16)) var
#else
#define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
#endif

namespace cobalt {
namespace media {

#define FIX_SHIFT 12
#define FIX(x) ((x) * (1 << FIX_SHIFT))

// Define a convenient macro to do static cast.
#define INT16_FIX(x) static_cast<int16_t>(FIX(x))

// Android's pixel layout is RGBA, while other platforms
// are BGRA.
#if defined(OS_ANDROID)
SIMD_ALIGNED(const int16_t ConvertRGBAToYUV_kTable[8 * 3]) = {
    INT16_FIX(0.257),  INT16_FIX(0.504),  INT16_FIX(0.098),  0,
    INT16_FIX(0.257),  INT16_FIX(0.504),  INT16_FIX(0.098),  0,
    -INT16_FIX(0.148), -INT16_FIX(0.291), INT16_FIX(0.439),  0,
    -INT16_FIX(0.148), -INT16_FIX(0.291), INT16_FIX(0.439),  0,
    INT16_FIX(0.439),  -INT16_FIX(0.368), -INT16_FIX(0.071), 0,
    INT16_FIX(0.439),  -INT16_FIX(0.368), -INT16_FIX(0.071), 0,
};
#else
SIMD_ALIGNED(const int16_t ConvertRGBAToYUV_kTable[8 * 3]) = {
    INT16_FIX(0.098),  INT16_FIX(0.504),  INT16_FIX(0.257),  0,
    INT16_FIX(0.098),  INT16_FIX(0.504),  INT16_FIX(0.257),  0,
    INT16_FIX(0.439),  -INT16_FIX(0.291), -INT16_FIX(0.148), 0,
    INT16_FIX(0.439),  -INT16_FIX(0.291), -INT16_FIX(0.148), 0,
    -INT16_FIX(0.071), -INT16_FIX(0.368), INT16_FIX(0.439),  0,
    -INT16_FIX(0.071), -INT16_FIX(0.368), INT16_FIX(0.439),  0,
};
#endif

#undef INT16_FIX

// This is the final offset for the conversion from signed yuv values to
// unsigned values. It is arranged so that offset of 16 is applied to Y
// components and 128 is added to UV components for 2 pixels.
SIMD_ALIGNED(const int32_t kYOffset[4]) = {16, 16, 16, 16};

static inline uint8_t Clamp(int value) {
  if (value < 0) return 0;
  if (value > 255) return 255;
  return static_cast<uint8_t>(value);
}

static inline uint8_t RGBToY(int r, int g, int b) {
  int y = ConvertRGBAToYUV_kTable[0] * b + ConvertRGBAToYUV_kTable[1] * g +
          ConvertRGBAToYUV_kTable[2] * r;
  y >>= FIX_SHIFT;
  return Clamp(y + 16);
}

static inline uint8_t RGBToU(int r, int g, int b, int shift) {
  int u = ConvertRGBAToYUV_kTable[8] * b + ConvertRGBAToYUV_kTable[9] * g +
          ConvertRGBAToYUV_kTable[10] * r;
  u >>= FIX_SHIFT + shift;
  return Clamp(u + 128);
}

static inline uint8_t RGBToV(int r, int g, int b, int shift) {
  int v = ConvertRGBAToYUV_kTable[16] * b + ConvertRGBAToYUV_kTable[17] * g +
          ConvertRGBAToYUV_kTable[18] * r;
  v >>= FIX_SHIFT + shift;
  return Clamp(v + 128);
}

#define CONVERT_Y(rgb_buf, y_buf) \
  b = *rgb_buf++;                 \
  g = *rgb_buf++;                 \
  r = *rgb_buf++;                 \
  ++rgb_buf;                      \
  sum_b += b;                     \
  sum_g += g;                     \
  sum_r += r;                     \
  *y_buf++ = RGBToY(r, g, b);

static inline void ConvertRGBToYUV_V2H2(const uint8_t* rgb_buf_1,
                                        const uint8_t* rgb_buf_2,
                                        uint8_t* y_buf_1, uint8_t* y_buf_2,
                                        uint8_t* u_buf, uint8_t* v_buf) {
  int sum_b = 0;
  int sum_g = 0;
  int sum_r = 0;
  int r, g, b;

  CONVERT_Y(rgb_buf_1, y_buf_1);
  CONVERT_Y(rgb_buf_1, y_buf_1);
  CONVERT_Y(rgb_buf_2, y_buf_2);
  CONVERT_Y(rgb_buf_2, y_buf_2);
  *u_buf++ = RGBToU(sum_r, sum_g, sum_b, 2);
  *v_buf++ = RGBToV(sum_r, sum_g, sum_b, 2);
}

static inline void ConvertRGBToYUV_V2H1(const uint8_t* rgb_buf_1,
                                        const uint8_t* rgb_buf_2,
                                        uint8_t* y_buf_1, uint8_t* y_buf_2,
                                        uint8_t* u_buf, uint8_t* v_buf) {
  int sum_b = 0;
  int sum_g = 0;
  int sum_r = 0;
  int r, g, b;

  CONVERT_Y(rgb_buf_1, y_buf_1);
  CONVERT_Y(rgb_buf_2, y_buf_2);
  *u_buf++ = RGBToU(sum_r, sum_g, sum_b, 1);
  *v_buf++ = RGBToV(sum_r, sum_g, sum_b, 1);
}

static inline void ConvertRGBToYUV_V1H2(const uint8_t* rgb_buf, uint8_t* y_buf,
                                        uint8_t* u_buf, uint8_t* v_buf) {
  int sum_b = 0;
  int sum_g = 0;
  int sum_r = 0;
  int r, g, b;

  CONVERT_Y(rgb_buf, y_buf);
  CONVERT_Y(rgb_buf, y_buf);
  *u_buf++ = RGBToU(sum_r, sum_g, sum_b, 1);
  *v_buf++ = RGBToV(sum_r, sum_g, sum_b, 1);
}

static inline void ConvertRGBToYUV_V1H1(const uint8_t* rgb_buf, uint8_t* y_buf,
                                        uint8_t* u_buf, uint8_t* v_buf) {
  int sum_b = 0;
  int sum_g = 0;
  int sum_r = 0;
  int r, g, b;

  CONVERT_Y(rgb_buf, y_buf);
  *u_buf++ = RGBToU(r, g, b, 0);
  *v_buf++ = RGBToV(r, g, b, 0);
}

static void ConvertRGB32ToYUVRow_SSE2(const uint8_t* rgb_buf_1,
                                      const uint8_t* rgb_buf_2,
                                      uint8_t* y_buf_1, uint8_t* y_buf_2,
                                      uint8_t* u_buf, uint8_t* v_buf,
                                      int width) {
  while (width >= 4) {
    // Name for the Y pixels:
    // Row 1: a b c d
    // Row 2: e f g h
    //
    // First row 4 pixels.
    __m128i rgb_row_1 =
        _mm_loadu_si128(reinterpret_cast<const __m128i*>(rgb_buf_1));
    __m128i zero_1 = _mm_xor_si128(rgb_row_1, rgb_row_1);

    __m128i y_table = _mm_load_si128(
        reinterpret_cast<const __m128i*>(ConvertRGBAToYUV_kTable));

    __m128i rgb_a_b = _mm_unpackhi_epi8(rgb_row_1, zero_1);
    rgb_a_b = _mm_madd_epi16(rgb_a_b, y_table);

    __m128i rgb_c_d = _mm_unpacklo_epi8(rgb_row_1, zero_1);
    rgb_c_d = _mm_madd_epi16(rgb_c_d, y_table);

    // Do a crazh shuffle so that we get:
    //  v------------ Multiply Add
    // BG: a b c d
    // A0: a b c d
    __m128i bg_abcd = _mm_castps_si128(
        _mm_shuffle_ps(_mm_castsi128_ps(rgb_c_d), _mm_castsi128_ps(rgb_a_b),
                       (3 << 6) | (1 << 4) | (3 << 2) | 1));
    __m128i r_abcd = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(rgb_c_d),
                                                     _mm_castsi128_ps(rgb_a_b),
                                                     (2 << 6) | (2 << 2)));
    __m128i y_abcd = _mm_add_epi32(bg_abcd, r_abcd);

    // Down shift back to 8bits range.
    __m128i y_offset =
        _mm_load_si128(reinterpret_cast<const __m128i*>(kYOffset));
    y_abcd = _mm_srai_epi32(y_abcd, FIX_SHIFT);
    y_abcd = _mm_add_epi32(y_abcd, y_offset);
    y_abcd = _mm_packs_epi32(y_abcd, y_abcd);
    y_abcd = _mm_packus_epi16(y_abcd, y_abcd);
    *reinterpret_cast<uint32_t*>(y_buf_1) = _mm_cvtsi128_si32(y_abcd);
    y_buf_1 += 4;

    // Second row 4 pixels.
    __m128i rgb_row_2 =
        _mm_loadu_si128(reinterpret_cast<const __m128i*>(rgb_buf_2));
    __m128i zero_2 = _mm_xor_si128(rgb_row_2, rgb_row_2);
    __m128i rgb_e_f = _mm_unpackhi_epi8(rgb_row_2, zero_2);
    __m128i rgb_g_h = _mm_unpacklo_epi8(rgb_row_2, zero_2);

    // Add two rows together.
    __m128i rgb_ae_bf =
        _mm_add_epi16(_mm_unpackhi_epi8(rgb_row_1, zero_2), rgb_e_f);
    __m128i rgb_cg_dh =
        _mm_add_epi16(_mm_unpacklo_epi8(rgb_row_1, zero_2), rgb_g_h);

    // Multiply add like the previous row.
    rgb_e_f = _mm_madd_epi16(rgb_e_f, y_table);
    rgb_g_h = _mm_madd_epi16(rgb_g_h, y_table);

    __m128i bg_efgh = _mm_castps_si128(
        _mm_shuffle_ps(_mm_castsi128_ps(rgb_g_h), _mm_castsi128_ps(rgb_e_f),
                       (3 << 6) | (1 << 4) | (3 << 2) | 1));
    __m128i r_efgh = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(rgb_g_h),
                                                     _mm_castsi128_ps(rgb_e_f),
                                                     (2 << 6) | (2 << 2)));
    __m128i y_efgh = _mm_add_epi32(bg_efgh, r_efgh);
    y_efgh = _mm_srai_epi32(y_efgh, FIX_SHIFT);
    y_efgh = _mm_add_epi32(y_efgh, y_offset);
    y_efgh = _mm_packs_epi32(y_efgh, y_efgh);
    y_efgh = _mm_packus_epi16(y_efgh, y_efgh);
    *reinterpret_cast<uint32_t*>(y_buf_2) = _mm_cvtsi128_si32(y_efgh);
    y_buf_2 += 4;

    __m128i rgb_ae_cg = _mm_castps_si128(
        _mm_shuffle_ps(_mm_castsi128_ps(rgb_cg_dh), _mm_castsi128_ps(rgb_ae_bf),
                       (3 << 6) | (2 << 4) | (3 << 2) | 2));
    __m128i rgb_bf_dh = _mm_castps_si128(
        _mm_shuffle_ps(_mm_castsi128_ps(rgb_cg_dh), _mm_castsi128_ps(rgb_ae_bf),
                       (1 << 6) | (1 << 2)));

    // This is a 2x2 subsampling for 2 pixels.
    __m128i rgb_abef_cdgh = _mm_add_epi16(rgb_ae_cg, rgb_bf_dh);

    // Do a multiply add with U table.
    __m128i u_a_b = _mm_madd_epi16(
        rgb_abef_cdgh, _mm_load_si128(reinterpret_cast<const __m128i*>(
                           ConvertRGBAToYUV_kTable + 8)));
    u_a_b = _mm_add_epi32(_mm_shuffle_epi32(u_a_b, ((3 << 2) | 1)),
                          _mm_shuffle_epi32(u_a_b, (2 << 2)));
    // Right shift 14 because of 12 from fixed point and 2 from subsampling.
    u_a_b = _mm_srai_epi32(u_a_b, FIX_SHIFT + 2);
    __m128i uv_offset = _mm_slli_epi32(y_offset, 3);
    u_a_b = _mm_add_epi32(u_a_b, uv_offset);
    u_a_b = _mm_packs_epi32(u_a_b, u_a_b);
    u_a_b = _mm_packus_epi16(u_a_b, u_a_b);
    *reinterpret_cast<uint16_t*>(u_buf) =
        static_cast<uint16_t>(_mm_extract_epi16(u_a_b, 0));
    u_buf += 2;

    __m128i v_a_b = _mm_madd_epi16(
        rgb_abef_cdgh, _mm_load_si128(reinterpret_cast<const __m128i*>(
                           ConvertRGBAToYUV_kTable + 16)));
    v_a_b = _mm_add_epi32(_mm_shuffle_epi32(v_a_b, ((3 << 2) | 1)),
                          _mm_shuffle_epi32(v_a_b, (2 << 2)));
    v_a_b = _mm_srai_epi32(v_a_b, FIX_SHIFT + 2);
    v_a_b = _mm_add_epi32(v_a_b, uv_offset);
    v_a_b = _mm_packs_epi32(v_a_b, v_a_b);
    v_a_b = _mm_packus_epi16(v_a_b, v_a_b);
    *reinterpret_cast<uint16_t*>(v_buf) =
        static_cast<uint16_t>(_mm_extract_epi16(v_a_b, 0));
    v_buf += 2;

    rgb_buf_1 += 16;
    rgb_buf_2 += 16;

    // Move forward by 4 pixels.
    width -= 4;
  }

  // Just use C code to convert the remaining pixels.
  if (width >= 2) {
    ConvertRGBToYUV_V2H2(rgb_buf_1, rgb_buf_2, y_buf_1, y_buf_2, u_buf, v_buf);
    rgb_buf_1 += 8;
    rgb_buf_2 += 8;
    y_buf_1 += 2;
    y_buf_2 += 2;
    ++u_buf;
    ++v_buf;
    width -= 2;
  }

  if (width)
    ConvertRGBToYUV_V2H1(rgb_buf_1, rgb_buf_2, y_buf_1, y_buf_2, u_buf, v_buf);
}

extern void ConvertRGB32ToYUV_SSE2(const uint8_t* rgbframe, uint8_t* yplane,
                                   uint8_t* uplane, uint8_t* vplane, int width,
                                   int height, int rgbstride, int ystride,
                                   int uvstride) {
  while (height >= 2) {
    ConvertRGB32ToYUVRow_SSE2(rgbframe, rgbframe + rgbstride, yplane,
                              yplane + ystride, uplane, vplane, width);
    rgbframe += 2 * rgbstride;
    yplane += 2 * ystride;
    uplane += uvstride;
    vplane += uvstride;
    height -= 2;
  }

  if (!height) return;

  // Handle the last row.
  while (width >= 2) {
    ConvertRGBToYUV_V1H2(rgbframe, yplane, uplane, vplane);
    rgbframe += 8;
    yplane += 2;
    ++uplane;
    ++vplane;
    width -= 2;
  }

  if (width) ConvertRGBToYUV_V1H1(rgbframe, yplane, uplane, vplane);
}

void ConvertRGB32ToYUV_SSE2_Reference(const uint8_t* rgbframe, uint8_t* yplane,
                                      uint8_t* uplane, uint8_t* vplane,
                                      int width, int height, int rgbstride,
                                      int ystride, int uvstride) {
  while (height >= 2) {
    int i = 0;

    // Convert a 2x2 block.
    while (i + 2 <= width) {
      ConvertRGBToYUV_V2H2(rgbframe + i * 4, rgbframe + rgbstride + i * 4,
                           yplane + i, yplane + ystride + i, uplane + i / 2,
                           vplane + i / 2);
      i += 2;
    }

    // Convert the last pixel of two rows.
    if (i < width) {
      ConvertRGBToYUV_V2H1(rgbframe + i * 4, rgbframe + rgbstride + i * 4,
                           yplane + i, yplane + ystride + i, uplane + i / 2,
                           vplane + i / 2);
    }

    rgbframe += 2 * rgbstride;
    yplane += 2 * ystride;
    uplane += uvstride;
    vplane += uvstride;
    height -= 2;
  }

  if (!height) return;

  // Handle the last row.
  while (width >= 2) {
    ConvertRGBToYUV_V1H2(rgbframe, yplane, uplane, vplane);
    rgbframe += 8;
    yplane += 2;
    ++uplane;
    ++vplane;
    width -= 2;
  }

  // Handle the last pixel in the last row.
  if (width) ConvertRGBToYUV_V1H1(rgbframe, yplane, uplane, vplane);
}

}  // namespace media
}  // namespace cobalt