GLES: Implement fast HQ scalers (#24611)

7 files changed, 297 insertions, 10 deletions

diff --git a/system/shaders/GLES/3.1/gles310_yuv2rgb.vert b/system/shaders/GLES/3.1/gles310_yuv2rgb.vert
new file mode 100644
index 0000000000..7f4dcdf649
--- /dev/null
+++ b/system/shaders/GLES/3.1/gles310_yuv2rgb.vert
@@ -0,0 +1,28 @@
+/*
+ *  Copyright (C) 2024 Team Kodi
+ *  This file is part of Kodi - https://kodi.tv
+ *
+ *  SPDX-License-Identifier: GPL-2.0-or-later
+ *  See LICENSES/README.md for more information.
+ */
+
+#version 310 es
+
+in vec4 m_attrpos;
+in vec2 m_attrcordY;
+in vec2 m_attrcordU;
+in vec2 m_attrcordV;
+out vec2 m_cordY;
+out vec2 m_cordU;
+out vec2 m_cordV;
+uniform mat4 m_proj;
+uniform mat4 m_model;
+
+void main()
+{
+  mat4 mvp = m_proj * m_model;
+  gl_Position = mvp * m_attrpos;
+  m_cordY = m_attrcordY;
+  m_cordU = m_attrcordU;
+  m_cordV = m_attrcordV;
+}
diff --git a/system/shaders/GLES/3.1/gles310_yuv2rgb_filter.frag b/system/shaders/GLES/3.1/gles310_yuv2rgb_filter.frag
new file mode 100644
index 0000000000..cabd56e3a2
--- /dev/null
+++ b/system/shaders/GLES/3.1/gles310_yuv2rgb_filter.frag
@@ -0,0 +1,129 @@
+/*
+ *  Copyright (C) 2024 Team Kodi
+ *  This file is part of Kodi - https://kodi.tv
+ *
+ *  SPDX-License-Identifier: GPL-2.0-or-later
+ *  See LICENSES/README.md for more information.
+ */
+
+#version 310 es
+
+precision highp float;
+
+uniform sampler2D m_sampY;
+uniform sampler2D m_sampU;
+uniform sampler2D m_sampV;
+uniform vec2 m_step;
+uniform mat4 m_yuvmat;
+uniform float m_stretch;
+uniform float m_alpha;
+uniform sampler2D m_kernelTex;
+uniform mat3 m_primMat;
+uniform float m_gammaDstInv;
+uniform float m_gammaSrc;
+uniform float m_toneP1;
+uniform float m_luminance;
+uniform vec3 m_coefsDst;
+in vec2 m_cordY;
+in vec2 m_cordU;
+in vec2 m_cordV;
+out vec4 fragColor;
+
+vec4[4] load4x4_0(sampler2D sampler, vec2 pos)
+{
+  vec4[4] tex4x4;
+  vec4 tex2x2 = textureGather(sampler, pos, 0);
+  tex4x4[0].xy = tex2x2.wz;
+  tex4x4[1].xy = tex2x2.xy;
+  tex2x2 = textureGatherOffset(sampler, pos, ivec2(2,0), 0);
+  tex4x4[0].zw = tex2x2.wz;
+  tex4x4[1].zw = tex2x2.xy;
+  tex2x2 = textureGatherOffset(sampler, pos, ivec2(0,2), 0);
+  tex4x4[2].xy = tex2x2.wz;
+  tex4x4[3].xy = tex2x2.xy;
+  tex2x2 = textureGatherOffset(sampler, pos, ivec2(2,2), 0);
+  tex4x4[2].zw = tex2x2.wz;
+  tex4x4[3].zw = tex2x2.xy;
+  return tex4x4;
+}
+
+float filter_0(sampler2D sampler, vec2 coord)
+{
+  vec2 pos = coord + m_step * 0.5;
+  vec2 f = fract(pos / m_step);
+
+  vec4 linetaps = texture(m_kernelTex, vec2(1.0 - f.x, 0.));
+  vec4 coltaps = texture(m_kernelTex, vec2(1.0 - f.y, 0.));
+  linetaps /= linetaps.r + linetaps.g + linetaps.b + linetaps.a;
+  coltaps /= coltaps.r + coltaps.g + coltaps.b + coltaps.a;
+  mat4 conv;
+  conv[0] = linetaps * coltaps.x;
+  conv[1] = linetaps * coltaps.y;
+  conv[2] = linetaps * coltaps.z;
+  conv[3] = linetaps * coltaps.w;
+
+  vec2 startPos = (-1.0 - f) * m_step + pos;
+  vec4[4] tex4x4 = load4x4_0(sampler, startPos);
+  vec4 imageLine0 = tex4x4[0];
+  vec4 imageLine1 = tex4x4[1];
+  vec4 imageLine2 = tex4x4[2];
+  vec4 imageLine3 = tex4x4[3];
+
+  return dot(imageLine0, conv[0]) +
+         dot(imageLine1, conv[1]) +
+         dot(imageLine2, conv[2]) +
+         dot(imageLine3, conv[3]);
+}
+
+void main()
+{
+  vec4 rgb;
+  vec4 yuv;
+
+#if defined(XBMC_YV12) || defined(XBMC_NV12)
+
+  yuv = vec4(filter_0(m_sampY, m_cordY),
+             texture2D(m_sampU, m_cordU).g,
+             texture2D(m_sampV, m_cordV).a,
+             1.0);
+
+#elif defined(XBMC_NV12_RRG)
+
+  yuv = vec4(filter_0(m_sampY, m_cordY),
+             texture2D(m_sampU, m_cordU).r,
+             texture2D(m_sampV, m_cordV).g,
+             1.0);
+
+#endif
+
+  rgb = m_yuvmat * yuv;
+  rgb.a = m_alpha;
+
+#if defined(XBMC_COL_CONVERSION)
+  rgb.rgb = pow(max(vec3(0), rgb.rgb), vec3(m_gammaSrc));
+  rgb.rgb = max(vec3(0), m_primMat * rgb.rgb);
+  rgb.rgb = pow(rgb.rgb, vec3(m_gammaDstInv));
+
+#if defined(KODI_TONE_MAPPING_REINHARD)
+  float luma = dot(rgb.rgb, m_coefsDst);
+  rgb.rgb *= reinhard(luma) / luma;
+
+#elif defined(KODI_TONE_MAPPING_ACES)
+  rgb.rgb = inversePQ(rgb.rgb);
+  rgb.rgb *= (10000.0 / m_luminance) * (2.0 / m_toneP1);
+  rgb.rgb = aces(rgb.rgb);
+  rgb.rgb *= (1.24 / m_toneP1);
+  rgb.rgb = pow(rgb.rgb, vec3(0.27));
+
+#elif defined(KODI_TONE_MAPPING_HABLE)
+  rgb.rgb = inversePQ(rgb.rgb);
+  rgb.rgb *= m_toneP1;
+  float wp = m_luminance / 100.0;
+  rgb.rgb = hable(rgb.rgb * wp) / hable(vec3(wp));
+  rgb.rgb = pow(rgb.rgb, vec3(1.0 / 2.2));
+#endif
+
+#endif
+
+  fragColor = rgb;
+}
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp b/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp
index 97663f9c09..67275f920b 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp
@@ -584,6 +584,8 @@ void CLinuxRendererGLES::UpdateVideoFilter()
   }
 
   m_scalingMethodGui = m_videoSettings.m_ScalingMethod;
+  if (m_scalingMethod != m_scalingMethodGui)
+    m_reloadShaders = true;
   m_scalingMethod = m_scalingMethodGui;
   m_viewRect = viewRect;
 
@@ -616,6 +618,8 @@ void CLinuxRendererGLES::UpdateVideoFilter()
     return;
   }
   case VS_SCALINGMETHOD_LINEAR:
+  case VS_SCALINGMETHOD_LANCZOS3_FAST:
+  case VS_SCALINGMETHOD_SPLINE36_FAST:
   {
     CLog::Log(LOGINFO, "GLES: Selecting single pass rendering");
     SetTextureFilter(GL_LINEAR);
@@ -623,8 +627,6 @@ void CLinuxRendererGLES::UpdateVideoFilter()
     return;
   }
   case VS_SCALINGMETHOD_LANCZOS2:
-  case VS_SCALINGMETHOD_SPLINE36_FAST:
-  case VS_SCALINGMETHOD_LANCZOS3_FAST:
   case VS_SCALINGMETHOD_SPLINE36:
   case VS_SCALINGMETHOD_LANCZOS3:
   case VS_SCALINGMETHOD_CUBIC_B_SPLINE:
@@ -709,9 +711,19 @@ void CLinuxRendererGLES::LoadShaders(int field)
 
           EShaderFormat shaderFormat = GetShaderFormat();
           m_toneMapMethod = m_videoSettings.m_ToneMapMethod;
-          m_pYUVProgShader = new YUV2RGBProgressiveShader(
-              shaderFormat, m_passthroughHDR ? m_srcPrimaries : AVColorPrimaries::AVCOL_PRI_BT709,
-              m_srcPrimaries, m_toneMap, m_toneMapMethod);
+          if (m_scalingMethod == VS_SCALINGMETHOD_LANCZOS3_FAST ||
+              m_scalingMethod == VS_SCALINGMETHOD_SPLINE36_FAST)
+          {
+            m_pYUVProgShader = new YUV2RGBFilterShader(
+                shaderFormat, m_passthroughHDR ? m_srcPrimaries : AVColorPrimaries::AVCOL_PRI_BT709,
+                m_srcPrimaries, m_toneMap, m_toneMapMethod, m_scalingMethod);
+          }
+          else
+          {
+            m_pYUVProgShader = new YUV2RGBProgressiveShader(
+                shaderFormat, m_passthroughHDR ? m_srcPrimaries : AVColorPrimaries::AVCOL_PRI_BT709,
+                m_srcPrimaries, m_toneMap, m_toneMapMethod);
+          }
           m_pYUVProgShader->SetConvertFullColorRange(m_fullRange);
           m_pYUVBobShader = new YUV2RGBBobShader(
               shaderFormat, m_passthroughHDR ? m_srcPrimaries : AVColorPrimaries::AVCOL_PRI_BT709,
@@ -1790,6 +1802,18 @@ bool CLinuxRendererGLES::Supports(ESCALINGMETHOD method) const
       method == VS_SCALINGMETHOD_SPLINE36 ||
       method == VS_SCALINGMETHOD_LANCZOS3)
   {
+    if (method == VS_SCALINGMETHOD_SPLINE36_FAST || method == VS_SCALINGMETHOD_LANCZOS3_FAST)
+    {
+#if defined(GL_ES_VERSION_3_0)
+      // we need GLES 3.0 headers for GL_RGBA16f, but GLES 3.1 for the shader
+      uint32_t major, minor;
+      m_renderSystem->GetRenderVersion(major, minor);
+      if (major < 3 || minor == 0)
+        return false;
+#else
+      return false;
+#endif
+    }
     // if scaling is below level, avoid hq scaling
     float scaleX = fabs((static_cast<float>(m_sourceWidth) - m_destRect.Width()) / m_sourceWidth) * 100;
     float scaleY = fabs((static_cast<float>(m_sourceHeight) - m_destRect.Height()) / m_sourceHeight) * 100;
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.cpp b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.cpp
index 63e4d304a4..5f4e63090a 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.cpp
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.cpp
@@ -1,6 +1,6 @@
 /*
  *  Copyright (c) 2007 d4rk
- *  Copyright (C) 2007-2018 Team Kodi
+ *  Copyright (C) 2007-2024 Team Kodi
  *  This file is part of Kodi - https://kodi.tv
  *
  *  SPDX-License-Identifier: GPL-2.0-or-later
@@ -10,6 +10,7 @@
 #include "YUV2RGBShaderGLES.h"
 
 #include "../RenderFlags.h"
+#include "ConvolutionKernels.h"
 #include "ToneMappers.h"
 #include "settings/AdvancedSettings.h"
 #include "utils/GLUtils.h"
@@ -267,3 +268,71 @@ bool YUV2RGBBobShader::OnEnabled()
   VerifyGLState();
   return true;
 }
+
+//------------------------------------------------------------------------------
+// YUV2RGBFilterShader
+//------------------------------------------------------------------------------
+
+YUV2RGBFilterShader::YUV2RGBFilterShader(EShaderFormat format,
+                                         AVColorPrimaries dstPrimaries,
+                                         AVColorPrimaries srcPrimaries,
+                                         bool toneMap,
+                                         ETONEMAPMETHOD toneMapMethod,
+                                         ESCALINGMETHOD method)
+  : BaseYUV2RGBGLSLShader(format, dstPrimaries, srcPrimaries, toneMap, toneMapMethod)
+{
+  m_scaling = method;
+  PixelShader()->LoadSource("gles310_yuv2rgb_filter.frag", m_defines);
+  VertexShader()->LoadSource("gles310_yuv2rgb.vert");
+  PixelShader()->AppendSource("gl_output.glsl");
+
+  PixelShader()->InsertSource("gl_tonemap.glsl", "void main()");
+}
+
+YUV2RGBFilterShader::~YUV2RGBFilterShader()
+{
+  if (m_kernelTex)
+    glDeleteTextures(1, &m_kernelTex);
+  m_kernelTex = 0;
+}
+
+void YUV2RGBFilterShader::OnCompiledAndLinked()
+{
+  BaseYUV2RGBGLSLShader::OnCompiledAndLinked();
+  m_hKernTex = glGetUniformLocation(ProgramHandle(), "m_kernelTex");
+
+  if (m_scaling != VS_SCALINGMETHOD_LANCZOS3_FAST && m_scaling != VS_SCALINGMETHOD_SPLINE36_FAST)
+    m_scaling = VS_SCALINGMETHOD_LANCZOS3_FAST;
+
+  CConvolutionKernel kernel(m_scaling, 256);
+
+  if (m_kernelTex)
+  {
+    glDeleteTextures(1, &m_kernelTex);
+    m_kernelTex = 0;
+  }
+  glGenTextures(1, &m_kernelTex);
+
+  glActiveTexture(GL_TEXTURE3);
+  glBindTexture(GL_TEXTURE_2D, m_kernelTex);
+  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+  glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+
+  GLvoid* data = (GLvoid*)kernel.GetFloatPixels();
+#if defined(GL_ES_VERSION_3_0)
+  glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F, kernel.GetSize(), 1, 0, GL_RGBA, GL_FLOAT, data);
+#endif
+  glActiveTexture(GL_TEXTURE0);
+  VerifyGLState();
+}
+
+bool YUV2RGBFilterShader::OnEnabled()
+{
+  glActiveTexture(GL_TEXTURE3);
+  glBindTexture(GL_TEXTURE_2D, m_kernelTex);
+  glUniform1i(m_hKernTex, 3);
+  glActiveTexture(GL_TEXTURE0);
+
+  return BaseYUV2RGBGLSLShader::OnEnabled();
+}
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.h b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.h
index 917f0f35f4..75bf05f325 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.h
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.h
@@ -1,5 +1,5 @@
 /*
- *  Copyright (C) 2007-2018 Team Kodi
+ *  Copyright (C) 2007-2024 Team Kodi
  *  This file is part of Kodi - https://kodi.tv
  *
  *  SPDX-License-Identifier: GPL-2.0-or-later
@@ -135,5 +135,25 @@ class BaseYUV2RGBGLSLShader : public CGLSLShaderProgram
     GLint m_hField = -1;
   };
 
+  class YUV2RGBFilterShader : public BaseYUV2RGBGLSLShader
+  {
+  public:
+    YUV2RGBFilterShader(EShaderFormat format,
+                        AVColorPrimaries dstPrimaries,
+                        AVColorPrimaries srcPrimaries,
+                        bool toneMap,
+                        ETONEMAPMETHOD toneMapMethod,
+                        ESCALINGMETHOD method);
+    ~YUV2RGBFilterShader() override;
+
+  protected:
+    void OnCompiledAndLinked() override;
+    bool OnEnabled() override;
+
+    GLuint m_kernelTex = 0;
+    GLint m_hKernTex = -1;
+    ESCALINGMETHOD m_scaling = VS_SCALINGMETHOD_LANCZOS3_FAST;
+  };
+
   } // namespace GLES
 } // end namespace
diff --git a/xbmc/rendering/gles/RenderSystemGLES.cpp b/xbmc/rendering/gles/RenderSystemGLES.cpp
index 471938a714..ef2261ff18 100644
--- a/xbmc/rendering/gles/RenderSystemGLES.cpp
+++ b/xbmc/rendering/gles/RenderSystemGLES.cpp
@@ -1,5 +1,5 @@
 /*
- *  Copyright (C) 2005-2018 Team Kodi
+ *  Copyright (C) 2005-2024 Team Kodi
  *  This file is part of Kodi - https://kodi.tv
  *
  *  SPDX-License-Identifier: GPL-2.0-or-later
@@ -8,11 +8,13 @@
 
 #include "RenderSystemGLES.h"
 
+#include "URL.h"
 #include "guilib/DirtyRegion.h"
 #include "guilib/GUITextureGLES.h"
 #include "rendering/MatrixGL.h"
 #include "settings/AdvancedSettings.h"
 #include "settings/SettingsComponent.h"
+#include "utils/FileUtils.h"
 #include "utils/GLUtils.h"
 #include "utils/MathUtils.h"
 #include "utils/SystemInfo.h"
@@ -774,3 +776,18 @@ GLint CRenderSystemGLES::GUIShaderGetCoordStep()
 
   return -1;
 }
+
+std::string CRenderSystemGLES::GetShaderPath(const std::string& filename)
+{
+  std::string path = "GLES/2.0/";
+
+  if (m_RenderVersionMajor >= 3 && m_RenderVersionMinor >= 1)
+  {
+    std::string file = "special://xbmc/system/shaders/GLES/3.1/" + filename;
+    const CURL pathToUrl(file);
+    if (CFileUtils::Exists(pathToUrl.Get()))
+      return "GLES/3.1/";
+  }
+
+  return path;
+}
diff --git a/xbmc/rendering/gles/RenderSystemGLES.h b/xbmc/rendering/gles/RenderSystemGLES.h
index 9c19bf6c28..1ea0ea60a5 100644
--- a/xbmc/rendering/gles/RenderSystemGLES.h
+++ b/xbmc/rendering/gles/RenderSystemGLES.h
@@ -1,5 +1,5 @@
 /*
- *  Copyright (C) 2005-2018 Team Kodi
+ *  Copyright (C) 2005-2024 Team Kodi
  *  This file is part of Kodi - https://kodi.tv
  *
  *  SPDX-License-Identifier: GPL-2.0-or-later
@@ -111,7 +111,7 @@ public:
 
   void Project(float &x, float &y, float &z) override;
 
-  std::string GetShaderPath(const std::string &filename) override { return "GLES/2.0/"; }
+  std::string GetShaderPath(const std::string& filename) override;
 
   void InitialiseShaders();
   void ReleaseShaders();