feat(video): use unified color conversion matrix generator (#4387)

src/platform/linux/cuda.cu

@@ -330,7 +330,7 @@ namespace cuda {
   }
 
   void sws_t::apply_colorspace(const video::sunshine_colorspace_t &colorspace) {
-    auto color_p = video::color_vectors_from_colorspace(colorspace);
+    auto color_p = video::color_vectors_from_colorspace(colorspace, true);
     CU_CHECK_IGNORE(cudaMemcpy(color_matrix.get(), color_p, sizeof(video::color_t), cudaMemcpyHostToDevice), "Couldn't copy color matrix to cuda");
   }
 

src/platform/linux/graphics.cpp

@@ -698,7 +698,7 @@ namespace egl {
   }
 
   void sws_t::apply_colorspace(const video::sunshine_colorspace_t &colorspace) {
-    auto color_p = video::color_vectors_from_colorspace(colorspace);
+    auto color_p = video::color_vectors_from_colorspace(colorspace, true);
 
     std::string_view members[] {
       util::view(color_p->color_vec_y),
@@ -811,7 +811,7 @@ namespace egl {
     gl::ctx.UseProgram(sws.program[1].handle());
     gl::ctx.Uniform1fv(loc_width_i, 1, &width_i);
 
-    auto color_p = video::color_vectors_from_colorspace(video::colorspace_e::rec601, false);
+    auto color_p = video::color_vectors_from_colorspace({video::colorspace_e::rec601, false, 8}, true);
     std::pair<const char *, std::string_view> members[] {
       std::make_pair("color_vec_y", util::view(color_p->color_vec_y)),
       std::make_pair("color_vec_u", util::view(color_p->color_vec_u)),

src/platform/windows/display_vram.cpp

@@ -457,12 +457,12 @@ namespace platf::dxgi {
     }
 
     void apply_colorspace(const ::video::sunshine_colorspace_t &colorspace) {
-      auto color_vectors = ::video::color_vectors_from_colorspace(colorspace);
+      auto color_vectors = ::video::color_vectors_from_colorspace(colorspace, true);
 
       if (format == DXGI_FORMAT_AYUV ||
           format == DXGI_FORMAT_R16_UINT ||
           format == DXGI_FORMAT_Y410) {
-        color_vectors = ::video::new_color_vectors_from_colorspace(colorspace);
+        color_vectors = ::video::color_vectors_from_colorspace(colorspace, false);
       }
 
       if (!color_vectors) {
@@ -775,7 +775,7 @@ namespace platf::dxgi {
         BOOST_LOG(warning) << "Failed to increase encoding GPU thread priority. Please run application as administrator for optimal performance.";
       }
 
-      auto default_color_vectors = ::video::color_vectors_from_colorspace(::video::colorspace_e::rec601, false);
+      auto default_color_vectors = ::video::color_vectors_from_colorspace({::video::colorspace_e::rec601, false, 8}, true);
       if (!default_color_vectors) {
         BOOST_LOG(error) << "Missing color vectors for Rec. 601"sv;
         return -1;

src/video_colorspace.cpp

@@ -120,66 +120,9 @@ namespace video {
     return avcodec_colorspace;
   }
 
-  const color_t *color_vectors_from_colorspace(const sunshine_colorspace_t &colorspace) {
+  const color_t *color_vectors_from_colorspace(const sunshine_colorspace_t &colorspace, bool unorm_output) {
-    return color_vectors_from_colorspace(colorspace.colorspace, colorspace.full_range);
+    constexpr auto generate_color_vectors = [](const sunshine_colorspace_t &colorspace, bool unorm_output) -> color_t {
-  }
+      // "Table 4 – Interpretation of matrix coefficients (MatrixCoefficients) value" section of ITU-T H.273
-
-  const color_t *color_vectors_from_colorspace(colorspace_e colorspace, bool full_range) {
-    using float2 = float[2];
-    auto make_color_matrix = [](float Cr, float Cb, const float2 &range_Y, const float2 &range_UV) -> color_t {
-      float Cg = 1.0f - Cr - Cb;
-
-      float Cr_i = 1.0f - Cr;
-      float Cb_i = 1.0f - Cb;
-
-      float shift_y = range_Y[0] / 255.0f;
-      float shift_uv = range_UV[0] / 255.0f;
-
-      float scale_y = (range_Y[1] - range_Y[0]) / 255.0f;
-      float scale_uv = (range_UV[1] - range_UV[0]) / 255.0f;
-      return {
-        {Cr, Cg, Cb, 0.0f},
-        {-(Cr * 0.5f / Cb_i), -(Cg * 0.5f / Cb_i), 0.5f, 0.5f},
-        {0.5f, -(Cg * 0.5f / Cr_i), -(Cb * 0.5f / Cr_i), 0.5f},
-        {scale_y, shift_y},
-        {scale_uv, shift_uv},
-      };
-    };
-
-    static const color_t colors[] {
-      make_color_matrix(0.299f, 0.114f, {16.0f, 235.0f}, {16.0f, 240.0f}),  // BT601 MPEG
-      make_color_matrix(0.299f, 0.114f, {0.0f, 255.0f}, {0.0f, 255.0f}),  // BT601 JPEG
-      make_color_matrix(0.2126f, 0.0722f, {16.0f, 235.0f}, {16.0f, 240.0f}),  // BT709 MPEG
-      make_color_matrix(0.2126f, 0.0722f, {0.0f, 255.0f}, {0.0f, 255.0f}),  // BT709 JPEG
-      make_color_matrix(0.2627f, 0.0593f, {16.0f, 235.0f}, {16.0f, 240.0f}),  // BT2020 MPEG
-      make_color_matrix(0.2627f, 0.0593f, {0.0f, 255.0f}, {0.0f, 255.0f}),  // BT2020 JPEG
-    };
-
-    const color_t *result = nullptr;
-
-    switch (colorspace) {
-      case colorspace_e::rec601:
-      default:
-        result = &colors[0];
-        break;
-      case colorspace_e::rec709:
-        result = &colors[2];
-        break;
-      case colorspace_e::bt2020:
-      case colorspace_e::bt2020sdr:
-        result = &colors[4];
-        break;
-    };
-
-    if (full_range) {
-      result++;
-    }
-
-    return result;
-  }
-
-  const color_t *new_color_vectors_from_colorspace(const sunshine_colorspace_t &colorspace) {
-    constexpr auto generate_color_vectors = [](const sunshine_colorspace_t &colorspace) -> color_t {
       double Kr, Kb;
       switch (colorspace.colorspace) {
         case colorspace_e::rec601:
@@ -202,7 +145,7 @@ namespace video {
       double y_mult, y_add;
       double uv_mult, uv_add;
 
-      // "Matrix coefficients" section of ITU-T H.273
+      // "8.3 Matrix coefficients" section of ITU-T H.273
       if (colorspace.full_range) {
         y_mult = (1 << colorspace.bit_depth) - 1;
         y_add = 0;
@@ -215,9 +158,17 @@ namespace video {
         uv_add = (1 << (colorspace.bit_depth - 8)) * 128;
       }
 
-      // For rounding
+      if (unorm_output) {
-      y_add += 0.5;
+        const double unorm_range = (1 << colorspace.bit_depth) - 1;
-      uv_add += 0.5;
+        y_mult /= unorm_range;
+        y_add /= unorm_range;
+        uv_mult /= unorm_range;
+        uv_add /= unorm_range;
+      } else {
+        // For rounding
+        y_add += 0.5f;
+        uv_add += 0.5f;
+      }
 
       color_t color_vectors;
 
@@ -246,18 +197,31 @@ namespace video {
     };
 
     static constexpr color_t colors[] = {
-      generate_color_vectors({colorspace_e::rec601, false, 8}),
+      generate_color_vectors({colorspace_e::rec601, false, 8}, false),
-      generate_color_vectors({colorspace_e::rec601, true, 8}),
+      generate_color_vectors({colorspace_e::rec601, true, 8}, false),
-      generate_color_vectors({colorspace_e::rec601, false, 10}),
+      generate_color_vectors({colorspace_e::rec601, false, 10}, false),
-      generate_color_vectors({colorspace_e::rec601, true, 10}),
+      generate_color_vectors({colorspace_e::rec601, true, 10}, false),
-      generate_color_vectors({colorspace_e::rec709, false, 8}),
+      generate_color_vectors({colorspace_e::rec709, false, 8}, false),
-      generate_color_vectors({colorspace_e::rec709, true, 8}),
+      generate_color_vectors({colorspace_e::rec709, true, 8}, false),
-      generate_color_vectors({colorspace_e::rec709, false, 10}),
+      generate_color_vectors({colorspace_e::rec709, false, 10}, false),
-      generate_color_vectors({colorspace_e::rec709, true, 10}),
+      generate_color_vectors({colorspace_e::rec709, true, 10}, false),
-      generate_color_vectors({colorspace_e::bt2020, false, 8}),
+      generate_color_vectors({colorspace_e::bt2020, false, 8}, false),
-      generate_color_vectors({colorspace_e::bt2020, true, 8}),
+      generate_color_vectors({colorspace_e::bt2020, true, 8}, false),
-      generate_color_vectors({colorspace_e::bt2020, false, 10}),
+      generate_color_vectors({colorspace_e::bt2020, false, 10}, false),
-      generate_color_vectors({colorspace_e::bt2020, true, 10}),
+      generate_color_vectors({colorspace_e::bt2020, true, 10}, false),
+
+      generate_color_vectors({colorspace_e::rec601, false, 8}, true),
+      generate_color_vectors({colorspace_e::rec601, true, 8}, true),
+      generate_color_vectors({colorspace_e::rec601, false, 10}, true),
+      generate_color_vectors({colorspace_e::rec601, true, 10}, true),
+      generate_color_vectors({colorspace_e::rec709, false, 8}, true),
+      generate_color_vectors({colorspace_e::rec709, true, 8}, true),
+      generate_color_vectors({colorspace_e::rec709, false, 10}, true),
+      generate_color_vectors({colorspace_e::rec709, true, 10}, true),
+      generate_color_vectors({colorspace_e::bt2020, false, 8}, true),
+      generate_color_vectors({colorspace_e::bt2020, true, 8}, true),
+      generate_color_vectors({colorspace_e::bt2020, false, 10}, true),
+      generate_color_vectors({colorspace_e::bt2020, true, 10}, true),
     };
 
     const color_t *result = nullptr;
@@ -282,6 +246,9 @@ namespace video {
     if (colorspace.full_range) {
       result += 1;
     }
+    if (unorm_output) {
+      result += 12;
+    }
 
     return result;
   }

src/video_colorspace.h

@@ -48,20 +48,15 @@ namespace video {
     float range_uv[2];
   };
 
-  const color_t *color_vectors_from_colorspace(const sunshine_colorspace_t &colorspace);
-
-  const color_t *color_vectors_from_colorspace(colorspace_e colorspace, bool full_range);
-
   /**
-   * @brief New version of `color_vectors_from_colorspace()` function that better adheres to the standards.
+   * @brief Get static RGB->YUV color conversion matrix.
-   *        Returned vectors are used to perform RGB->YUV conversion.
+   *        This matrix expects RGB input in UNORM (0.0 to 1.0) range and doesn't perform any
-   *        Unlike its predecessor, color vectors will produce output in `UINT` range, not `UNORM` range.
+   *        gamut mapping or gamma correction.
-   *        Input is still in `UNORM` range. Returned vectors won't modify color primaries and color
-   *        transfer function.
    * @param colorspace Targeted YUV colorspace.
+   * @param unorm_output Whether the matrix should produce output in UNORM or UINT range.
    * @return `const color_t*` that contains RGB->YUV transformation vectors.
    *         Components `range_y` and `range_uv` are there for backwards compatibility
    *         and can be ignored in the computation.
    */
-  const color_t *new_color_vectors_from_colorspace(const sunshine_colorspace_t &colorspace);
+  const color_t *color_vectors_from_colorspace(const sunshine_colorspace_t &colorspace, bool unorm_output);
 }  // namespace video