Merge pull request #14014 from MaxKellermann/matrix_optimizations

Various CMatrixGL optimizations

19 files changed, 126 insertions, 154 deletions

diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGL.cpp b/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGL.cpp
index 1b48204499..3187f0095b 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGL.cpp
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGL.cpp
@@ -36,7 +36,7 @@
 #include "windowing/WinSystem.h"
 #include "guilib/Texture.h"
 #include "guilib/LocalizeStrings.h"
-#include "guilib/MatrixGLES.h"
+#include "rendering/MatrixGL.h"
 #include "rendering/gl/RenderSystemGL.h"
 #include "threads/SingleLock.h"
 #include "utils/log.h"
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp b/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp
index 8620e01257..cb97de8a04 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp
@@ -21,7 +21,6 @@
 #include "system_gl.h"
 
 #include <locale.h>
-#include "guilib/MatrixGLES.h"
 #include "LinuxRendererGLES.h"
 #include "ServiceBroker.h"
 #include "utils/MathUtils.h"
@@ -33,6 +32,7 @@
 #include "settings/Settings.h"
 #include "VideoShaders/YUV2RGBShaderGLES.h"
 #include "VideoShaders/VideoFilterShaderGLES.h"
+#include "rendering/MatrixGL.h"
 #include "rendering/gles/RenderSystemGLES.h"
 #include "guilib/Texture.h"
 #include "threads/SingleLock.h"
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/OverlayRendererGL.cpp b/xbmc/cores/VideoPlayer/VideoRenderers/OverlayRendererGL.cpp
index a8913c2bea..87c5e59c7c 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/OverlayRendererGL.cpp
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/OverlayRendererGL.cpp
@@ -29,7 +29,7 @@
 #include "LinuxRendererGLES.h"
 #include "rendering/gles/RenderSystemGLES.h"
 #endif
-#include "guilib/MatrixGLES.h"
+#include "rendering/MatrixGL.h"
 #include "RenderManager.h"
 #include "ServiceBroker.h"
 #include "cores/VideoPlayer/DVDCodecs/Overlay/DVDOverlayImage.h"
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/VideoFilterShaderGL.h b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/VideoFilterShaderGL.h
index fa43da8c50..2e6ad551c7 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/VideoFilterShaderGL.h
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/VideoFilterShaderGL.h
@@ -43,7 +43,7 @@ namespace Shaders {
     GLint GetVertexLoc() { return m_hVertex; }
     GLint GetCoordLoc() { return m_hCoord; }
 
-    void SetMatrices(GLfloat *p, GLfloat *m) { m_proj = p; m_model = m; }
+    void SetMatrices(const GLfloat *p, const GLfloat *m) { m_proj = p; m_model = m; }
 
   protected:
     int m_width;
@@ -53,8 +53,8 @@ namespace Shaders {
     float m_stretch;
     GLfloat m_alpha;
     GLint m_sourceTexUnit;
-    GLfloat *m_proj = nullptr;
-    GLfloat *m_model = nullptr;
+    const GLfloat *m_proj = nullptr;
+    const GLfloat *m_model = nullptr;
 
     // shader attribute handles
     GLint m_hSourceTex;
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/VideoFilterShaderGLES.h b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/VideoFilterShaderGLES.h
index 2a0a2c4f66..41355246ce 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/VideoFilterShaderGLES.h
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/VideoFilterShaderGLES.h
@@ -40,7 +40,7 @@ namespace Shaders {
     virtual bool GetTextureFilter(GLint& filter) { return false; }
     virtual GLint GetVertexLoc() { return m_hVertex; }
     virtual GLint GetcoordLoc() { return m_hcoord; }
-    virtual void SetMatrices(GLfloat *p, GLfloat *m) { m_proj = p; m_model = m; }
+    virtual void SetMatrices(const GLfloat *p, const GLfloat *m) { m_proj = p; m_model = m; }
     virtual void SetAlpha(GLfloat alpha)             { m_alpha = alpha; }
 
   protected:
@@ -60,8 +60,8 @@ namespace Shaders {
     GLint m_hModel;
     GLint m_hAlpha;
 
-    GLfloat *m_proj;
-    GLfloat *m_model;
+    const GLfloat *m_proj;
+    const GLfloat *m_model;
     GLfloat  m_alpha;
   };
 
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGL.h b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGL.h
index c237057d6f..f1e1d9c53e 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGL.h
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGL.h
@@ -64,7 +64,7 @@ public:
   GLint GetUcoordLoc() { return m_hUcoord; }
   GLint GetVcoordLoc() { return m_hVcoord; }
 
-  void SetMatrices(GLfloat *p, GLfloat *m) { m_proj = p; m_model = m; }
+  void SetMatrices(const GLfloat *p, const GLfloat *m) { m_proj = p; m_model = m; }
   void SetAlpha(GLfloat alpha)  { m_alpha = alpha; }
 
 protected:
@@ -90,8 +90,8 @@ protected:
   float m_contrast;
   float m_stretch;
 
-  GLfloat *m_proj = nullptr;
-  GLfloat *m_model = nullptr;
+  const GLfloat *m_proj = nullptr;
+  const GLfloat *m_model = nullptr;
   GLfloat m_alpha = 1.0f;
 
   std::string m_defines;
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.h b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.h
index edc34827cb..5841839dbf 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.h
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/YUV2RGBShaderGLES.h
@@ -50,7 +50,7 @@ namespace Shaders {
     virtual GLint GetUcoordLoc() { return 0; };
     virtual GLint GetVcoordLoc() { return 0; };
 
-    virtual void SetMatrices(GLfloat *p, GLfloat *m) {};
+    virtual void SetMatrices(const GLfloat *p, const GLfloat *m) {};
     virtual void SetAlpha(GLfloat alpha) {};
 
     virtual void SetConvertFullColorRange(bool convertFullRange) {}
@@ -77,7 +77,7 @@ namespace Shaders {
     GLint GetUcoordLoc() override { return m_hUcoord; }
     GLint GetVcoordLoc() override { return m_hVcoord; }
 
-    void SetMatrices(GLfloat *p, GLfloat *m) override { m_proj = p; m_model = m; }
+    void SetMatrices(const GLfloat *p, const GLfloat *m) override { m_proj = p; m_model = m; }
     void SetAlpha(GLfloat alpha) override { m_alpha = alpha; }
 
   protected:
@@ -112,8 +112,8 @@ namespace Shaders {
     GLint m_hModel;
     GLint m_hAlpha;
 
-    GLfloat *m_proj;
-    GLfloat *m_model;
+    const GLfloat *m_proj;
+    const GLfloat *m_model;
     GLfloat  m_alpha;
 
     bool m_convertFullRange;
diff --git a/xbmc/guilib/CMakeLists.txt b/xbmc/guilib/CMakeLists.txt
index b76ec85d0f..803cbc638c 100644
--- a/xbmc/guilib/CMakeLists.txt
+++ b/xbmc/guilib/CMakeLists.txt
@@ -156,12 +156,10 @@ set(HEADERS DDSImage.h
 if(OPENGL_FOUND)
   list(APPEND SOURCES GUIFontTTFGL.cpp
                       GUITextureGL.cpp
-                      MatrixGLES.cpp
                       Shader.cpp
                       TextureGL.cpp)
   list(APPEND HEADERS GUIFontTTFGL.h
                       GUITextureGL.h
-                      MatrixGLES.h
                       Shader.h
                       TextureGL.h)
 endif()
@@ -169,12 +167,10 @@ endif()
 if(OPENGLES_FOUND)
   list(APPEND SOURCES GUIFontTTFGL.cpp
                       GUITextureGLES.cpp
-                      MatrixGLES.cpp
                       Shader.cpp
                       TextureGL.cpp)
   list(APPEND HEADERS GUIFontTTFGL.h
                       GUITextureGLES.h
-                      MatrixGLES.h
                       Shader.h
                       TextureGL.h)
 endif()
diff --git a/xbmc/guilib/GUIFontTTFGL.cpp b/xbmc/guilib/GUIFontTTFGL.cpp
index 26fc3dfb45..298c056952 100644
--- a/xbmc/guilib/GUIFontTTFGL.cpp
+++ b/xbmc/guilib/GUIFontTTFGL.cpp
@@ -33,7 +33,7 @@
 #elif HAS_GLES
 #include "rendering/gles/RenderSystemGLES.h"
 #endif
-#include "guilib/MatrixGLES.h"
+#include "rendering/MatrixGL.h"
 
 // stuff for freetype
 #include <ft2build.h>
diff --git a/xbmc/guilib/MatrixGLES.cpp b/xbmc/rendering/MatrixGL.cpp
index 16098da741..1f99f8169f 100644
--- a/xbmc/guilib/MatrixGLES.cpp
+++ b/xbmc/rendering/MatrixGL.cpp
@@ -18,21 +18,31 @@
  *
  */
 
+#include "MatrixGL.h"
+#include "utils/TransformMatrix.h"
 
-#include "system_gl.h"
-
-#include <cmath>
-#include "MatrixGLES.h"
-#include "utils/log.h"
 #if defined(HAS_NEON)
 #include "utils/CPUInfo.h"
 #endif
 
+#include <cmath>
 
 CMatrixGLStack glMatrixModview = CMatrixGLStack();
 CMatrixGLStack glMatrixProject = CMatrixGLStack();
 CMatrixGLStack glMatrixTexture = CMatrixGLStack();
 
+CMatrixGL::CMatrixGL(const TransformMatrix &src) noexcept
+{
+  for(int i = 0; i < 3; i++)
+    for(int j = 0; j < 4; j++)
+      m_pMatrix[j * 4 + i] = src.m[i][j];
+
+  m_pMatrix[3] = 0.0f;
+  m_pMatrix[7] = 0.0f;
+  m_pMatrix[11] = 0.0f;
+  m_pMatrix[15] = 1.0f;
+}
+
 void CMatrixGL::LoadIdentity()
 {
   m_pMatrix[0] = 1.0f;  m_pMatrix[4] = 0.0f;  m_pMatrix[8]  = 0.0f;  m_pMatrix[12] = 0.0f;
@@ -49,10 +59,10 @@ void CMatrixGL::Ortho(GLfloat l, GLfloat r, GLfloat b, GLfloat t, GLfloat n, GLf
   GLfloat x = - (r + l) / (r - l);
   GLfloat y = - (t + b) / (t - b);
   GLfloat z = - (f + n) / (f - n);
-  GLfloat matrix[16] = {   u, 0.0f, 0.0f, 0.0f,
-                        0.0f,    v, 0.0f, 0.0f,
-                        0.0f, 0.0f,    w, 0.0f,
-                           x,    y,    z, 1.0f};
+  const CMatrixGL matrix{   u, 0.0f, 0.0f, 0.0f,
+                         0.0f,    v, 0.0f, 0.0f,
+                         0.0f, 0.0f,    w, 0.0f,
+                            x,    y,    z, 1.0f};
   MultMatrixf(matrix);
 }
 
@@ -62,10 +72,10 @@ void CMatrixGL::Ortho2D(GLfloat l, GLfloat r, GLfloat b, GLfloat t)
   GLfloat v =  2.0f / (t - b);
   GLfloat x = - (r + l) / (r - l);
   GLfloat y = - (t + b) / (t - b);
-  GLfloat matrix[16] = {   u, 0.0f, 0.0f, 0.0f,
-                        0.0f,    v, 0.0f, 0.0f,
-                        0.0f, 0.0f,-1.0f, 0.0f,
-                           x,    y, 0.0f, 1.0f};
+  const CMatrixGL matrix{   u, 0.0f, 0.0f, 0.0f,
+                         0.0f,    v, 0.0f, 0.0f,
+                         0.0f, 0.0f,-1.0f, 0.0f,
+                            x,    y, 0.0f, 1.0f};
   MultMatrixf(matrix);
 }
 
@@ -77,42 +87,42 @@ void CMatrixGL::Frustum(GLfloat l, GLfloat r, GLfloat b, GLfloat t, GLfloat n, G
   GLfloat x = (t + b) / (t - b);
   GLfloat y = - (f + n) / (f - n);
   GLfloat z = - (2.0f * f * n) / (f - n);
-  GLfloat matrix[16] = {   u, 0.0f, 0.0f, 0.0f,
-                        0.0f,    v, 0.0f, 0.0f,
-                           w,    x,    y,-1.0f,
-                        0.0f, 0.0f,    z, 0.0f};
+  const CMatrixGL matrix{   u, 0.0f, 0.0f, 0.0f,
+                         0.0f,    v, 0.0f, 0.0f,
+                            w,    x,    y,-1.0f,
+                         0.0f, 0.0f,    z, 0.0f};
   MultMatrixf(matrix);
 }
 
 void CMatrixGL::Translatef(GLfloat x, GLfloat y, GLfloat z)
 {
-  GLfloat matrix[16] = {1.0f, 0.0f, 0.0f, 0.0f,
-                        0.0f, 1.0f, 0.0f, 0.0f,
-                        0.0f, 0.0f, 1.0f, 0.0f,
-                           x,    y,    z, 1.0f};
+  const CMatrixGL matrix{1.0f, 0.0f, 0.0f, 0.0f,
+                         0.0f, 1.0f, 0.0f, 0.0f,
+                         0.0f, 0.0f, 1.0f, 0.0f,
+                            x,    y,    z, 1.0f};
   MultMatrixf(matrix);
 }
 
 void CMatrixGL::Scalef(GLfloat x, GLfloat y, GLfloat z)
 {
-  GLfloat matrix[16] = {   x, 0.0f, 0.0f, 0.0f,
-                        0.0f,    y, 0.0f, 0.0f,
-                        0.0f, 0.0f,    z, 0.0f,
-                        0.0f, 0.0f, 0.0f, 1.0f};
+  const CMatrixGL matrix{   x, 0.0f, 0.0f, 0.0f,
+                         0.0f,    y, 0.0f, 0.0f,
+                         0.0f, 0.0f,    z, 0.0f,
+                         0.0f, 0.0f, 0.0f, 1.0f};
   MultMatrixf(matrix);
 }
 
 void CMatrixGL::Rotatef(GLfloat angle, GLfloat x, GLfloat y, GLfloat z)
 {
-  GLfloat modulus = sqrt((x*x)+(y*y)+(z*z));
+  GLfloat modulus = std::sqrt((x*x)+(y*y)+(z*z));
   if (modulus != 0.0)
   {
     x /= modulus;
     y /= modulus;
     z /= modulus;
   }
-  GLfloat cosine = cos(angle);
-  GLfloat sine   = sin(angle);
+  GLfloat cosine = std::cos(angle);
+  GLfloat sine   = std::sin(angle);
   GLfloat cos1   = 1 - cosine;
   GLfloat a = (x*x*cos1) + cosine;
   GLfloat b = (x*y*cos1) - (z*sine);
@@ -123,21 +133,21 @@ void CMatrixGL::Rotatef(GLfloat angle, GLfloat x, GLfloat y, GLfloat z)
   GLfloat g = (z*x*cos1) - (y*sine);
   GLfloat h = (z*y*cos1) + (x*sine);
   GLfloat i = (z*z*cos1) + cosine;
-  GLfloat matrix[16] = {   a,    d,    g, 0.0f,
-                           b,    e,    h, 0.0f,
-                           c,    f,    i, 0.0f,
-                        0.0f, 0.0f, 0.0f, 1.0f};
+  const CMatrixGL matrix{   a,    d,    g, 0.0f,
+                            b,    e,    h, 0.0f,
+                            c,    f,    i, 0.0f,
+                         0.0f, 0.0f, 0.0f, 1.0f};
   MultMatrixf(matrix);
 }
 
 #if defined(HAS_NEON) && !defined(__LP64__)
   
-inline void Matrix4Mul(const float* src_mat_1, const float* src_mat_2, float* dst_mat)
+static inline void Matrix4Mul(float* src_mat_1, const float* src_mat_2)
 {
   asm volatile (
     // Store A & B leaving room at top of registers for result (q0-q3)
-    "vldmia %1, { q4-q7 }  \n\t"
-    "vldmia %2, { q8-q11 } \n\t"
+    "vldmia %0, { q4-q7 }  \n\t"
+    "vldmia %1, { q8-q11 } \n\t"
 
     // result = first column of B x first row of A
     "vmul.f32 q0, q8, d8[0]\n\t"
@@ -164,39 +174,38 @@ inline void Matrix4Mul(const float* src_mat_1, const float* src_mat_2, float* ds
     "vmla.f32 q3, q11, d15[1]\n\t"
 
     // output = result registers
-    "vstmia %2, { q0-q3 }"
+    "vstmia %1, { q0-q3 }"
     : //no output 
-    : "r" (dst_mat), "r" (src_mat_2), "r" (src_mat_1)       // input - note *value* of pointer doesn't change
+    : "r" (src_mat_2), "r" (src_mat_1)       // input - note *value* of pointer doesn't change
     : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11" //clobber
     );
 }
 #endif
-void CMatrixGL::MultMatrixf(const GLfloat *matrix)
+void CMatrixGL::MultMatrixf(const CMatrixGL &matrix) noexcept
 {
 #if defined(HAS_NEON) && !defined(__LP64__)
     if ((g_cpuInfo.GetCPUFeatures() & CPU_FEATURE_NEON) == CPU_FEATURE_NEON)
     {
-      GLfloat m[16];
-      Matrix4Mul(m_pMatrix, matrix, m);
+      Matrix4Mul(m_pMatrix, matrix.m_pMatrix);
       return;
     }
 #endif
-    GLfloat a = (matrix[0]  * m_pMatrix[0]) + (matrix[1]  * m_pMatrix[4]) + (matrix[2]  * m_pMatrix[8])  + (matrix[3]  * m_pMatrix[12]);
-    GLfloat b = (matrix[0]  * m_pMatrix[1]) + (matrix[1]  * m_pMatrix[5]) + (matrix[2]  * m_pMatrix[9])  + (matrix[3]  * m_pMatrix[13]);
-    GLfloat c = (matrix[0]  * m_pMatrix[2]) + (matrix[1]  * m_pMatrix[6]) + (matrix[2]  * m_pMatrix[10]) + (matrix[3]  * m_pMatrix[14]);
-    GLfloat d = (matrix[0]  * m_pMatrix[3]) + (matrix[1]  * m_pMatrix[7]) + (matrix[2]  * m_pMatrix[11]) + (matrix[3]  * m_pMatrix[15]);
-    GLfloat e = (matrix[4]  * m_pMatrix[0]) + (matrix[5]  * m_pMatrix[4]) + (matrix[6]  * m_pMatrix[8])  + (matrix[7]  * m_pMatrix[12]);
-    GLfloat f = (matrix[4]  * m_pMatrix[1]) + (matrix[5]  * m_pMatrix[5]) + (matrix[6]  * m_pMatrix[9])  + (matrix[7]  * m_pMatrix[13]);
-    GLfloat g = (matrix[4]  * m_pMatrix[2]) + (matrix[5]  * m_pMatrix[6]) + (matrix[6]  * m_pMatrix[10]) + (matrix[7]  * m_pMatrix[14]);
-    GLfloat h = (matrix[4]  * m_pMatrix[3]) + (matrix[5]  * m_pMatrix[7]) + (matrix[6]  * m_pMatrix[11]) + (matrix[7]  * m_pMatrix[15]);
-    GLfloat i = (matrix[8]  * m_pMatrix[0]) + (matrix[9]  * m_pMatrix[4]) + (matrix[10] * m_pMatrix[8])  + (matrix[11] * m_pMatrix[12]);
-    GLfloat j = (matrix[8]  * m_pMatrix[1]) + (matrix[9]  * m_pMatrix[5]) + (matrix[10] * m_pMatrix[9])  + (matrix[11] * m_pMatrix[13]);
-    GLfloat k = (matrix[8]  * m_pMatrix[2]) + (matrix[9]  * m_pMatrix[6]) + (matrix[10] * m_pMatrix[10]) + (matrix[11] * m_pMatrix[14]);
-    GLfloat l = (matrix[8]  * m_pMatrix[3]) + (matrix[9]  * m_pMatrix[7]) + (matrix[10] * m_pMatrix[11]) + (matrix[11] * m_pMatrix[15]);
-    GLfloat m = (matrix[12] * m_pMatrix[0]) + (matrix[13] * m_pMatrix[4]) + (matrix[14] * m_pMatrix[8])  + (matrix[15] * m_pMatrix[12]);
-    GLfloat n = (matrix[12] * m_pMatrix[1]) + (matrix[13] * m_pMatrix[5]) + (matrix[14] * m_pMatrix[9])  + (matrix[15] * m_pMatrix[13]);
-    GLfloat o = (matrix[12] * m_pMatrix[2]) + (matrix[13] * m_pMatrix[6]) + (matrix[14] * m_pMatrix[10]) + (matrix[15] * m_pMatrix[14]);
-    GLfloat p = (matrix[12] * m_pMatrix[3]) + (matrix[13] * m_pMatrix[7]) + (matrix[14] * m_pMatrix[11]) + (matrix[15] * m_pMatrix[15]);
+    GLfloat a = (matrix.m_pMatrix[0]  * m_pMatrix[0]) + (matrix.m_pMatrix[1]  * m_pMatrix[4]) + (matrix.m_pMatrix[2]  * m_pMatrix[8])  + (matrix.m_pMatrix[3]  * m_pMatrix[12]);
+    GLfloat b = (matrix.m_pMatrix[0]  * m_pMatrix[1]) + (matrix.m_pMatrix[1]  * m_pMatrix[5]) + (matrix.m_pMatrix[2]  * m_pMatrix[9])  + (matrix.m_pMatrix[3]  * m_pMatrix[13]);
+    GLfloat c = (matrix.m_pMatrix[0]  * m_pMatrix[2]) + (matrix.m_pMatrix[1]  * m_pMatrix[6]) + (matrix.m_pMatrix[2]  * m_pMatrix[10]) + (matrix.m_pMatrix[3]  * m_pMatrix[14]);
+    GLfloat d = (matrix.m_pMatrix[0]  * m_pMatrix[3]) + (matrix.m_pMatrix[1]  * m_pMatrix[7]) + (matrix.m_pMatrix[2]  * m_pMatrix[11]) + (matrix.m_pMatrix[3]  * m_pMatrix[15]);
+    GLfloat e = (matrix.m_pMatrix[4]  * m_pMatrix[0]) + (matrix.m_pMatrix[5]  * m_pMatrix[4]) + (matrix.m_pMatrix[6]  * m_pMatrix[8])  + (matrix.m_pMatrix[7]  * m_pMatrix[12]);
+    GLfloat f = (matrix.m_pMatrix[4]  * m_pMatrix[1]) + (matrix.m_pMatrix[5]  * m_pMatrix[5]) + (matrix.m_pMatrix[6]  * m_pMatrix[9])  + (matrix.m_pMatrix[7]  * m_pMatrix[13]);
+    GLfloat g = (matrix.m_pMatrix[4]  * m_pMatrix[2]) + (matrix.m_pMatrix[5]  * m_pMatrix[6]) + (matrix.m_pMatrix[6]  * m_pMatrix[10]) + (matrix.m_pMatrix[7]  * m_pMatrix[14]);
+    GLfloat h = (matrix.m_pMatrix[4]  * m_pMatrix[3]) + (matrix.m_pMatrix[5]  * m_pMatrix[7]) + (matrix.m_pMatrix[6]  * m_pMatrix[11]) + (matrix.m_pMatrix[7]  * m_pMatrix[15]);
+    GLfloat i = (matrix.m_pMatrix[8]  * m_pMatrix[0]) + (matrix.m_pMatrix[9]  * m_pMatrix[4]) + (matrix.m_pMatrix[10] * m_pMatrix[8])  + (matrix.m_pMatrix[11] * m_pMatrix[12]);
+    GLfloat j = (matrix.m_pMatrix[8]  * m_pMatrix[1]) + (matrix.m_pMatrix[9]  * m_pMatrix[5]) + (matrix.m_pMatrix[10] * m_pMatrix[9])  + (matrix.m_pMatrix[11] * m_pMatrix[13]);
+    GLfloat k = (matrix.m_pMatrix[8]  * m_pMatrix[2]) + (matrix.m_pMatrix[9]  * m_pMatrix[6]) + (matrix.m_pMatrix[10] * m_pMatrix[10]) + (matrix.m_pMatrix[11] * m_pMatrix[14]);
+    GLfloat l = (matrix.m_pMatrix[8]  * m_pMatrix[3]) + (matrix.m_pMatrix[9]  * m_pMatrix[7]) + (matrix.m_pMatrix[10] * m_pMatrix[11]) + (matrix.m_pMatrix[11] * m_pMatrix[15]);
+    GLfloat m = (matrix.m_pMatrix[12] * m_pMatrix[0]) + (matrix.m_pMatrix[13] * m_pMatrix[4]) + (matrix.m_pMatrix[14] * m_pMatrix[8])  + (matrix.m_pMatrix[15] * m_pMatrix[12]);
+    GLfloat n = (matrix.m_pMatrix[12] * m_pMatrix[1]) + (matrix.m_pMatrix[13] * m_pMatrix[5]) + (matrix.m_pMatrix[14] * m_pMatrix[9])  + (matrix.m_pMatrix[15] * m_pMatrix[13]);
+    GLfloat o = (matrix.m_pMatrix[12] * m_pMatrix[2]) + (matrix.m_pMatrix[13] * m_pMatrix[6]) + (matrix.m_pMatrix[14] * m_pMatrix[10]) + (matrix.m_pMatrix[15] * m_pMatrix[14]);
+    GLfloat p = (matrix.m_pMatrix[12] * m_pMatrix[3]) + (matrix.m_pMatrix[13] * m_pMatrix[7]) + (matrix.m_pMatrix[14] * m_pMatrix[11]) + (matrix.m_pMatrix[15] * m_pMatrix[15]);
     m_pMatrix[0] = a;  m_pMatrix[4] = e;  m_pMatrix[8]  = i;  m_pMatrix[12] = m;
     m_pMatrix[1] = b;  m_pMatrix[5] = f;  m_pMatrix[9]  = j;  m_pMatrix[13] = n;
     m_pMatrix[2] = c;  m_pMatrix[6] = g;  m_pMatrix[10] = k;  m_pMatrix[14] = o;
@@ -207,7 +216,6 @@ void CMatrixGL::MultMatrixf(const GLfloat *matrix)
 void CMatrixGL::LookAt(GLfloat eyex, GLfloat eyey, GLfloat eyez, GLfloat centerx, GLfloat centery, GLfloat centerz, GLfloat upx, GLfloat upy, GLfloat upz)
 {
   GLfloat forward[3], side[3], up[3];
-  GLfloat m[4][4];
 
   forward[0] = centerx - eyex;
   forward[1] = centery - eyey;
@@ -217,7 +225,7 @@ void CMatrixGL::LookAt(GLfloat eyex, GLfloat eyey, GLfloat eyez, GLfloat centerx
   up[1] = upy;
   up[2] = upz;
 
-  GLfloat tmp = sqrt(forward[0]*forward[0] + forward[1]*forward[1] + forward[2]*forward[2]);
+  GLfloat tmp = std::sqrt(forward[0]*forward[0] + forward[1]*forward[1] + forward[2]*forward[2]);
   if (tmp != 0.0)
   {
     forward[0] /= tmp;
@@ -229,7 +237,7 @@ void CMatrixGL::LookAt(GLfloat eyex, GLfloat eyey, GLfloat eyez, GLfloat centerx
   side[1] = forward[2]*up[0] - forward[0]*up[2];
   side[2] = forward[0]*up[1] - forward[1]*up[0];
 
-  tmp = sqrt(side[0]*side[0] + side[1]*side[1] + side[2]*side[2]);
+  tmp = std::sqrt(side[0]*side[0] + side[1]*side[1] + side[2]*side[2]);
   if (tmp != 0.0)
   {
     side[0] /= tmp;
@@ -241,24 +249,14 @@ void CMatrixGL::LookAt(GLfloat eyex, GLfloat eyey, GLfloat eyez, GLfloat centerx
   up[1] = side[2]*forward[0] - side[0]*forward[2];
   up[2] = side[0]*forward[1] - side[1]*forward[0];
 
-  m[0][0] = 1.0f; m[0][1] = 0.0f; m[0][2] = 0.0f; m[0][3] = 0.0f;
-  m[1][0] = 0.0f; m[1][1] = 1.0f; m[1][2] = 0.0f; m[1][3] = 0.0f;
-  m[2][0] = 0.0f; m[2][1] = 0.0f; m[2][2] = 1.0f; m[2][3] = 0.0f;
-  m[3][0] = 0.0f; m[3][1] = 0.0f; m[3][2] = 0.0f; m[3][3] = 1.0f;
-
-  m[0][0] = side[0];
-  m[1][0] = side[1];
-  m[2][0] = side[2];
+  const CMatrixGL matrix{
+    side[0], up[0], -forward[0], 0.0f,
+    side[1], up[1], -forward[1], 0.0f,
+    side[2], up[2], -forward[2], 0.0f,
+    0.0f, 0.0f, 0.0f, 1.0f,
+  };
 
-  m[0][1] = up[0];
-  m[1][1] = up[1];
-  m[2][1] = up[2];
-
-  m[0][2] = -forward[0];
-  m[1][2] = -forward[1];
-  m[2][2] = -forward[2];
-
-  MultMatrixf(&m[0][0]);
+  MultMatrixf(matrix);
   Translatef(-eyex, -eyey, -eyez);
 }
 
@@ -307,14 +305,6 @@ bool CMatrixGL::Project(GLfloat objx, GLfloat objy, GLfloat objz, const GLfloat
   return true;
 }
 
-void CMatrixGL::PrintMatrix(void)
-{
-  CLog::Log(LOGDEBUG, "%f %f %f %f", m_pMatrix[0], m_pMatrix[4], m_pMatrix[8],  m_pMatrix[12]);
-  CLog::Log(LOGDEBUG, "%f %f %f %f", m_pMatrix[1], m_pMatrix[5], m_pMatrix[9],  m_pMatrix[13]);
-  CLog::Log(LOGDEBUG, "%f %f %f %f", m_pMatrix[2], m_pMatrix[6], m_pMatrix[10], m_pMatrix[14]);
-  CLog::Log(LOGDEBUG, "%f %f %f %f", m_pMatrix[3], m_pMatrix[7], m_pMatrix[11], m_pMatrix[15]);
-}
-
 void CMatrixGLStack::Load()
 {
 
diff --git a/xbmc/guilib/MatrixGLES.h b/xbmc/rendering/MatrixGL.h
index 67dd6f309e..75a161728e 100644
--- a/xbmc/guilib/MatrixGLES.h
+++ b/xbmc/rendering/MatrixGL.h
@@ -19,20 +19,25 @@
  *
  */
 
+#include "system_gl.h"
 
-#include <cmath>
-#include <cstring>
 #include <stack>
 
+class TransformMatrix;
+
 class CMatrixGL
 {
 public:
+  CMatrixGL() = default;
+
+  constexpr CMatrixGL(GLfloat x0, GLfloat x1, GLfloat x2, GLfloat x3,
+                      GLfloat x4, GLfloat x5, GLfloat x6, GLfloat x7,
+                      GLfloat x8, GLfloat x9, GLfloat x10, GLfloat x11,
+                      GLfloat x12, GLfloat x13, GLfloat x14, GLfloat x15)
+    :m_pMatrix{x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15} {}
+
+  CMatrixGL(const TransformMatrix &src) noexcept;
 
-  CMatrixGL()                                  { memset(&m_pMatrix, 0, sizeof(m_pMatrix)); };
-  explicit CMatrixGL(const float matrix[16])   { memcpy(m_pMatrix, matrix, sizeof(m_pMatrix)); }
-  CMatrixGL(const CMatrixGL &rhs )             { memcpy(m_pMatrix, rhs.m_pMatrix, sizeof(m_pMatrix)); }
-  CMatrixGL &operator=( const CMatrixGL &rhs ) { memcpy(m_pMatrix, rhs.m_pMatrix, sizeof(m_pMatrix)); return *this;}
-  operator float*()                            { return m_pMatrix; }
   operator const float*() const                { return m_pMatrix; }
 
   void LoadIdentity();
@@ -42,21 +47,20 @@ public:
   void Translatef(GLfloat x, GLfloat y, GLfloat z);
   void Scalef(GLfloat x, GLfloat y, GLfloat z);
   void Rotatef(GLfloat angle, GLfloat x, GLfloat y, GLfloat z);
-  void MultMatrixf(const GLfloat *matrix);
+  void MultMatrixf(const CMatrixGL &matrix) noexcept;
   void LookAt(GLfloat eyex, GLfloat eyey, GLfloat eyez, GLfloat centerx, GLfloat centery, GLfloat centerz, GLfloat upx, GLfloat upy, GLfloat upz);
 
   static bool Project(GLfloat objx, GLfloat objy, GLfloat objz, const GLfloat modelMatrix[16], const GLfloat projMatrix[16], const GLint viewport[4], GLfloat* winx, GLfloat* winy, GLfloat* winz);
 
-  void PrintMatrix(void);
-
-  GLfloat m_pMatrix[16];
+private:
+  /* alignas(16) allows better SIMD optimizations (e.g. SSE2 benefits
+     a lot from this) */
+  alignas(16) GLfloat m_pMatrix[16];
 };
 
 class CMatrixGLStack
 {
 public:
-  explicit CMatrixGLStack() {}
-
   void Push()
   {
     m_stack.push(m_current);
diff --git a/xbmc/rendering/gl/CMakeLists.txt b/xbmc/rendering/gl/CMakeLists.txt
index 5d14fbbf7a..4c4a245fbe 100644
--- a/xbmc/rendering/gl/CMakeLists.txt
+++ b/xbmc/rendering/gl/CMakeLists.txt
@@ -1,9 +1,11 @@
 set(SOURCES GUIWindowTestPatternGL.cpp
             RenderSystemGL.cpp
+            ../MatrixGL.cpp
             GLShader.cpp)
 
 set(HEADERS GUIWindowTestPatternGL.h
             RenderSystemGL.h
+            ../MatrixGL.h
             GLShader.h)
 
 core_add_library(rendering_gl)
diff --git a/xbmc/rendering/gl/GLShader.cpp b/xbmc/rendering/gl/GLShader.cpp
index 07e4cbacf8..74b538985c 100644
--- a/xbmc/rendering/gl/GLShader.cpp
+++ b/xbmc/rendering/gl/GLShader.cpp
@@ -24,8 +24,8 @@
 #include "ServiceBroker.h"
 #include "utils/log.h"
 #include "rendering/RenderSystem.h"
+#include "rendering/MatrixGL.h"
 #include "windowing/GraphicContext.h"
-#include "guilib/MatrixGLES.h"
 
 using namespace Shaders;
 
@@ -81,8 +81,8 @@ bool CGLShader::OnEnabled()
 {
   // This is called after glUseProgram()
 
-  GLfloat *projMatrix = glMatrixProject.Get().m_pMatrix;
-  GLfloat *modelMatrix = glMatrixModview.Get().m_pMatrix;
+  const GLfloat *projMatrix = glMatrixProject.Get();
+  const GLfloat *modelMatrix = glMatrixModview.Get();
   glUniformMatrix4fv(m_hProj,  1, GL_FALSE, projMatrix);
   glUniformMatrix4fv(m_hModel, 1, GL_FALSE, modelMatrix);
 
diff --git a/xbmc/rendering/gl/GLShader.h b/xbmc/rendering/gl/GLShader.h
index ff7d1b6dc2..005b905375 100644
--- a/xbmc/rendering/gl/GLShader.h
+++ b/xbmc/rendering/gl/GLShader.h
@@ -55,8 +55,8 @@ protected:
   GLint m_hCord0 = 0;
   GLint m_hCord1 = 0;
 
-  GLfloat *m_proj = nullptr;
-  GLfloat *m_model = nullptr;
+  const GLfloat *m_proj = nullptr;
+  const GLfloat *m_model = nullptr;
 
   bool m_clipPossible = false;
   GLfloat m_clipXFactor;
diff --git a/xbmc/rendering/gl/RenderSystemGL.cpp b/xbmc/rendering/gl/RenderSystemGL.cpp
index df946c079c..c66b7db2ab 100644
--- a/xbmc/rendering/gl/RenderSystemGL.cpp
+++ b/xbmc/rendering/gl/RenderSystemGL.cpp
@@ -20,9 +20,9 @@
 
 #include "RenderSystemGL.h"
 #include "filesystem/File.h"
+#include "rendering/MatrixGL.h"
 #include "windowing/GraphicContext.h"
 #include "settings/AdvancedSettings.h"
-#include "guilib/MatrixGLES.h"
 #include "settings/DisplaySettings.h"
 #include "utils/log.h"
 #include "utils/GLUtils.h"
@@ -402,18 +402,7 @@ void CRenderSystemGL::ApplyHardwareTransform(const TransformMatrix &finalMatrix)
     return;
 
   glMatrixModview.Push();
-  GLfloat matrix[4][4];
-
-  for(int i = 0; i < 3; i++)
-    for(int j = 0; j < 4; j++)
-      matrix[j][i] = finalMatrix.m[i][j];
-
-  matrix[0][3] = 0.0f;
-  matrix[1][3] = 0.0f;
-  matrix[2][3] = 0.0f;
-  matrix[3][3] = 1.0f;
-
-  glMatrixModview->MultMatrixf(&matrix[0][0]);
+  glMatrixModview->MultMatrixf(finalMatrix);
   glMatrixModview.Load();
 }
 
diff --git a/xbmc/rendering/gles/CMakeLists.txt b/xbmc/rendering/gles/CMakeLists.txt
index cac8ef9ebe..4df58e3c68 100644
--- a/xbmc/rendering/gles/CMakeLists.txt
+++ b/xbmc/rendering/gles/CMakeLists.txt
@@ -1,8 +1,10 @@
 if(OPENGLES_FOUND)
   set(SOURCES RenderSystemGLES.cpp
+              ../MatrixGL.cpp
               GLESShader.cpp)
 
   set(HEADERS RenderSystemGLES.h
+              ../MatrixGL.h
               GLESShader.h)
 
   core_add_library(rendering_gles)
diff --git a/xbmc/rendering/gles/GLESShader.cpp b/xbmc/rendering/gles/GLESShader.cpp
index bf23ef7fe9..4600c317e5 100644
--- a/xbmc/rendering/gles/GLESShader.cpp
+++ b/xbmc/rendering/gles/GLESShader.cpp
@@ -19,10 +19,10 @@
  */
 
 #include "GLESShader.h"
-#include "xbmc/guilib/MatrixGLES.h"
 #include "ServiceBroker.h"
 #include "utils/log.h"
 #include "rendering/RenderSystem.h"
+#include "rendering/MatrixGL.h"
 #include "windowing/GraphicContext.h"
 
 using namespace Shaders;
@@ -92,8 +92,8 @@ bool CGLESShader::OnEnabled()
 {
   // This is called after glUseProgram()
 
-  GLfloat *projMatrix = glMatrixProject.Get().m_pMatrix;
-  GLfloat *modelMatrix = glMatrixModview.Get().m_pMatrix;
+  const GLfloat *projMatrix = glMatrixProject.Get();
+  const GLfloat *modelMatrix = glMatrixModview.Get();
   glUniformMatrix4fv(m_hProj,  1, GL_FALSE, projMatrix);
   glUniformMatrix4fv(m_hModel, 1, GL_FALSE, modelMatrix);
 
diff --git a/xbmc/rendering/gles/GLESShader.h b/xbmc/rendering/gles/GLESShader.h
index fdb688bd49..2456eac6c7 100644
--- a/xbmc/rendering/gles/GLESShader.h
+++ b/xbmc/rendering/gles/GLESShader.h
@@ -65,8 +65,8 @@ protected:
   GLint m_hContrast = 0;
   GLint m_hBrightness = 0;
 
-  GLfloat *m_proj;
-  GLfloat *m_model;
+  const GLfloat *m_proj;
+  const GLfloat *m_model;
 
   bool m_clipPossible;
   GLfloat m_clipXFactor;
diff --git a/xbmc/rendering/gles/RenderSystemGLES.cpp b/xbmc/rendering/gles/RenderSystemGLES.cpp
index e1634f61f3..7805b16017 100644
--- a/xbmc/rendering/gles/RenderSystemGLES.cpp
+++ b/xbmc/rendering/gles/RenderSystemGLES.cpp
@@ -21,7 +21,7 @@
 #include "windowing/GraphicContext.h"
 #include "settings/AdvancedSettings.h"
 #include "RenderSystemGLES.h"
-#include "guilib/MatrixGLES.h"
+#include "rendering/MatrixGL.h"
 #include "utils/log.h"
 #include "utils/GLUtils.h"
 #include "utils/TimeUtils.h"
@@ -372,18 +372,7 @@ void CRenderSystemGLES::ApplyHardwareTransform(const TransformMatrix &finalMatri
     return;
 
   glMatrixModview.Push();
-  GLfloat matrix[4][4];
-
-  for(int i = 0; i < 3; i++)
-    for(int j = 0; j < 4; j++)
-      matrix[j][i] = finalMatrix.m[i][j];
-
-  matrix[0][3] = 0.0f;
-  matrix[1][3] = 0.0f;
-  matrix[2][3] = 0.0f;
-  matrix[3][3] = 1.0f;
-
-  glMatrixModview->MultMatrixf(&matrix[0][0]);
+  glMatrixModview->MultMatrixf(finalMatrix);
   glMatrixModview.Load();
 }