Implement Gaussian blur using Compute Shader

The filters are ~3.5 time faster with CS.

data/shaders/gaussian.comp

@@ -0,0 +1,36 @@
+// From http://http.developer.nvidia.com/GPUGems3/gpugems3_ch40.html
+
+uniform layout(size1x16) restrict readonly image2D source;
+uniform layout(size1x16) volatile restrict writeonly image2D dest;
+uniform float sigma = 5.;
+
+layout (local_size_x = 8, local_size_y = 8) in;
+
+shared float local_src[8 + 2 * 8][8];
+
+void main()
+{
+    int x = int(gl_LocalInvocationID.x), y = int(gl_LocalInvocationID.y);
+    ivec2 uv = ivec2(gl_GlobalInvocationID.x, gl_GlobalInvocationID.y);
+    local_src[x][y] = imageLoad(source, ivec2(uv) - ivec2(8, 0)).x;
+    local_src[x + 8][y] = imageLoad(source, ivec2(uv)).x;
+    local_src[x + 16][y] = imageLoad(source, ivec2(uv) + ivec2(8, 0)).x;
+
+    barrier();
+
+    float g0, g1, g2;
+    g0 = 1.0 / (sqrt(2.0 * 3.14) * sigma);
+    g1 = exp(-0.5 / (sigma * sigma));
+    g2 = g1 * g1;
+    float sum = local_src[x + 8][y] * g0;
+    g0 *= g1;
+    g1 *= g2;
+    for (int j = 1; j < 8; j++) {
+        sum += local_src[8 + x - j][y] * g0;
+        sum += local_src[8 + x + j][y] * g0;
+        g0 *= g1;
+        g1 *= g2;
+    }
+    imageStore(dest, ivec2(uv), vec4(sum));
+}
+

data/shaders/gaussianv.comp

@@ -0,0 +1,36 @@
+// From http://http.developer.nvidia.com/GPUGems3/gpugems3_ch40.html
+
+uniform layout(size1x16) restrict readonly image2D source;
+uniform layout(size1x16) volatile restrict writeonly image2D dest;
+uniform float sigma = 5.;
+
+layout (local_size_x = 8, local_size_y = 8) in;
+
+shared float local_src[8][8 + 2 * 8];
+
+void main()
+{
+    int x = int(gl_LocalInvocationID.x), y = int(gl_LocalInvocationID.y);
+    ivec2 uv = ivec2(gl_GlobalInvocationID.x, gl_GlobalInvocationID.y);
+    local_src[x][y] = imageLoad(source, ivec2(uv) - ivec2(0, 8)).x;
+    local_src[x][y + 8] = imageLoad(source, ivec2(uv)).x;
+    local_src[x][y + 16] = imageLoad(source, ivec2(uv) + ivec2(0, 8)).x;
+
+    barrier();
+
+    float g0, g1, g2;
+    g0 = 1.0 / (sqrt(2.0 * 3.14) * sigma);
+    g1 = exp(-0.5 / (sigma * sigma));
+    g2 = g1 * g1;
+    float sum = local_src[x][y + 8] * g0;
+    g0 *= g1;
+    g1 *= g2;
+    for (int j = 1; j < 8; j++) {
+        sum += local_src[x][y + 8 + j] * g0;
+        sum += local_src[x][y + 8 - j] * g0;
+        g0 *= g1;
+        g1 *= g2;
+    }
+    imageStore(dest, ivec2(uv), vec4(sum));
+}
+

src/graphics/glwrap.cpp

@@ -74,6 +74,8 @@ PFNGLGETCOMPRESSEDTEXIMAGEPROC glGetCompressedTexImage;
 PFNGLTEXSTORAGE1DPROC glTexStorage1D;
 PFNGLTEXSTORAGE2DPROC glTexStorage2D;
 PFNGLTEXSTORAGE3DPROC glTexStorage3D;
+PFNGLBINDIMAGETEXTUREPROC glBindImageTexture;
+PFNGLDISPATCHCOMPUTEPROC glDispatchCompute;
 #endif
 
 static bool is_gl_init = false;
@@ -226,6 +228,8 @@ void initGL()
     glTexStorage1D = (PFNGLTEXSTORAGE1DPROC)IRR_OGL_LOAD_EXTENSION("glTexStorage1D");
     glTexStorage2D = (PFNGLTEXSTORAGE2DPROC)IRR_OGL_LOAD_EXTENSION("glTexStorage2D");
     glTexStorage3D = (PFNGLTEXSTORAGE3DPROC)IRR_OGL_LOAD_EXTENSION("glTexStorage3D");
+    glBindImageTexture = (PFNGLBINDIMAGETEXTUREPROC)IRR_OGL_LOAD_EXTENSION("glBindImageTexture");
+    glDispatchCompute = (PFNGLDISPATCHCOMPUTEPROC)IRR_OGL_LOAD_EXTENSION("glDispatchCompute");
 #ifdef DEBUG
     glDebugMessageCallbackARB = (PFNGLDEBUGMESSAGECALLBACKARBPROC)IRR_OGL_LOAD_EXTENSION("glDebugMessageCallbackARB");
 #endif

src/graphics/glwrap.hpp

@@ -97,6 +97,8 @@ extern PFNGLGETCOMPRESSEDTEXIMAGEPROC glGetCompressedTexImage;
 extern PFNGLTEXSTORAGE1DPROC glTexStorage1D;
 extern PFNGLTEXSTORAGE2DPROC glTexStorage2D;
 extern PFNGLTEXSTORAGE3DPROC glTexStorage3D;
+extern PFNGLBINDIMAGETEXTUREPROC glBindImageTexture;
+extern PFNGLDISPATCHCOMPUTEPROC glDispatchCompute;
 #ifdef DEBUG
 extern PFNGLDEBUGMESSAGECALLBACKARBPROC glDebugMessageCallbackARB;
 #endif

src/graphics/post_processing.cpp

@@ -396,6 +396,8 @@ void PostProcessing::renderGaussian17TapBlur(FrameBuffer &in_fbo, FrameBuffer &a
 {
     assert(in_fbo.getWidth() == auxiliary.getWidth() && in_fbo.getHeight() == auxiliary.getHeight());
     float inv_width = 1.0f / in_fbo.getWidth(), inv_height = 1.0f / in_fbo.getHeight();
+    {
+        if (irr_driver->getGLSLVersion() < 430)
         {
             auxiliary.Bind();
             glUseProgram(FullScreenShader::Gaussian17TapHShader::Program);
@@ -410,6 +412,18 @@ void PostProcessing::renderGaussian17TapBlur(FrameBuffer &in_fbo, FrameBuffer &a
 
             glDrawArrays(GL_TRIANGLES, 0, 3);
         }
+        else
+        {
+            glUseProgram(FullScreenShader::ComputeGaussian17TapHShader::Program);
+            glBindImageTexture(0, in_fbo.getRTT()[0], 0, false, 0, GL_READ_ONLY, GL_R16F);
+            glBindImageTexture(1, auxiliary.getRTT()[0], 0, false, 0, GL_WRITE_ONLY, GL_R16F);
+            glUniform1i(FullScreenShader::ComputeGaussian17TapHShader::uniform_source, 0);
+            glUniform1i(FullScreenShader::ComputeGaussian17TapHShader::uniform_dest, 1);
+            glDispatchCompute(in_fbo.getWidth() / 8, in_fbo.getHeight() / 8, 1);
+        }
+    }
+    {
+        if (irr_driver->getGLSLVersion() < 430)
         {
             in_fbo.Bind();
             glUseProgram(FullScreenShader::Gaussian17TapVShader::Program);
@@ -424,6 +438,16 @@ void PostProcessing::renderGaussian17TapBlur(FrameBuffer &in_fbo, FrameBuffer &a
 
             glDrawArrays(GL_TRIANGLES, 0, 3);
         }
+        else
+        {
+            glUseProgram(FullScreenShader::ComputeGaussian17TapVShader::Program);
+            glBindImageTexture(0, auxiliary.getRTT()[0], 0, false, 0, GL_READ_ONLY, GL_R16F);
+            glBindImageTexture(1, in_fbo.getRTT()[0], 0, false, 0, GL_WRITE_ONLY, GL_R16F);
+            glUniform1i(FullScreenShader::ComputeGaussian17TapVShader::uniform_source, 0);
+            glUniform1i(FullScreenShader::ComputeGaussian17TapVShader::uniform_dest, 1);
+            glDispatchCompute(in_fbo.getWidth() / 8, in_fbo.getHeight() / 8, 1);
+        }
+    }
 }
 
 void PostProcessing::renderPassThrough(GLuint tex)

src/graphics/shaders.cpp

@@ -277,9 +277,11 @@ void Shaders::loadShaders()
     FullScreenShader::DepthOfFieldShader::init();
     FullScreenShader::FogShader::init();
     FullScreenShader::Gaussian17TapHShader::init();
+    FullScreenShader::ComputeGaussian17TapHShader::init();
     FullScreenShader::Gaussian3HBlurShader::init();
     FullScreenShader::Gaussian3VBlurShader::init();
     FullScreenShader::Gaussian17TapVShader::init();
+    FullScreenShader::ComputeGaussian17TapVShader::init();
     FullScreenShader::Gaussian6HBlurShader::init();
     FullScreenShader::Gaussian6VBlurShader::init();
     FullScreenShader::GlowShader::init();
@@ -2439,6 +2441,17 @@ namespace FullScreenShader
         vao = createFullScreenVAO(Program);
     }
 
+    GLuint ComputeGaussian17TapHShader::Program;
+    GLuint ComputeGaussian17TapHShader::uniform_source;
+    GLuint ComputeGaussian17TapHShader::uniform_dest;
+    void ComputeGaussian17TapHShader::init()
+    {
+        Program = LoadProgram(
+            GL_COMPUTE_SHADER, file_manager->getAsset("shaders/gaussian.comp").c_str());
+        uniform_source = glGetUniformLocation(Program, "source");
+        uniform_dest = glGetUniformLocation(Program, "dest");
+    }
+
     GLuint Gaussian6HBlurShader::Program;
     GLuint Gaussian6HBlurShader::uniform_tex;
     GLuint Gaussian6HBlurShader::uniform_pixel;
@@ -2481,6 +2494,17 @@ namespace FullScreenShader
         vao = createFullScreenVAO(Program);
     }
 
+    GLuint ComputeGaussian17TapVShader::Program;
+    GLuint ComputeGaussian17TapVShader::uniform_source;
+    GLuint ComputeGaussian17TapVShader::uniform_dest;
+    void ComputeGaussian17TapVShader::init()
+    {
+        Program = LoadProgram(
+            GL_COMPUTE_SHADER, file_manager->getAsset("shaders/gaussianv.comp").c_str());
+        uniform_source = glGetUniformLocation(Program, "source");
+        uniform_dest = glGetUniformLocation(Program, "dest");
+    }
+
     GLuint Gaussian6VBlurShader::Program;
     GLuint Gaussian6VBlurShader::uniform_tex;
     GLuint Gaussian6VBlurShader::uniform_pixel;

src/graphics/shaders.hpp

@@ -641,6 +641,15 @@ public:
     static void init();
 };
 
+class ComputeGaussian17TapHShader
+{
+public:
+    static GLuint Program;
+    static GLuint uniform_source, uniform_dest;
+
+    static void init();
+};
+
 class Gaussian6HBlurShader
 {
 public:
@@ -671,6 +680,16 @@ public:
     static void init();
 };
 
+class ComputeGaussian17TapVShader
+{
+public:
+    static GLuint Program;
+    static GLuint uniform_source, uniform_dest;
+
+    static void init();
+};
+
+
 class Gaussian6VBlurShader
 {
 public: