ilikecats/stk-code_catmod
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Implemented bilateral for compute path.

Browse Source
This commit is contained in:
Vincent Lejeune 2014-06-21 00:27:23 +02:00 committed by vlj
parent d8edf64caf
commit 3219e564d0
5 changed files with 45 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
data/shaders/gaussian.comp → data/shaders/bilateralH.comp
View File

@ -1,20 +1,25 @@
// From http://http.developer.nvidia.com/GPUGems3/gpugems3_ch40.html
uniform layout(size1x16) restrict readonly image2D source;
uniform layout(size1x32) restrict readonly image2D depth;
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];
shared float local_depth[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_depth[x][y] = imageLoad(depth, ivec2(uv) - ivec2(8, 0)).x;
local_src[x + 8][y] = imageLoad(source, ivec2(uv)).x;
local_depth[x + 8][y] = imageLoad(depth, ivec2(uv)).x;
local_src[x + 16][y] = imageLoad(source, ivec2(uv) + ivec2(8, 0)).x;
local_depth[x + 16][y] = imageLoad(depth, ivec2(uv) + ivec2(8, 0)).x;
barrier();
@ -23,14 +28,20 @@ void main()
g1 = exp(-0.5 / (sigma * sigma));
g2 = g1 * g1;
float sum = local_src[x + 8][y] * g0;
float pixel_depth = local_depth[x + 8][y];
g0 *= g1;
g1 *= g2;
float tmp_weight, total_weight = g0;
for (int j = 1; j < 8; j++) {
sum += local_src[8 + x - j][y] * g0;
sum += local_src[8 + x + j][y] * g0;
tmp_weight = max(0.0, 1.0 - .001 * abs(local_depth[8 + x - j][y] - pixel_depth));
total_weight += g0 * tmp_weight;
sum += local_src[8 + x - j][y] * g0 * tmp_weight;
tmp_weight = max(0.0, 1.0 - .001 * abs(local_depth[8 + x + j][y] - pixel_depth));
total_weight += g0 * tmp_weight;
sum += local_src[8 + x + j][y] * g0 * tmp_weight;
g0 *= g1;
g1 *= g2;
}
imageStore(dest, ivec2(uv), vec4(sum));
imageStore(dest, ivec2(uv), vec4(sum / total_weight));
}

18
data/shaders/gaussianv.comp → data/shaders/bilateralV.comp
View File

@ -1,20 +1,25 @@
// From http://http.developer.nvidia.com/GPUGems3/gpugems3_ch40.html
uniform layout(size1x16) restrict readonly image2D source;
uniform layout(size1x32) restrict readonly image2D depth;
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];
shared float local_depth[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_depth[x][y] = imageLoad(depth, ivec2(uv) - ivec2(0, 8)).x;
local_src[x][y + 8] = imageLoad(source, ivec2(uv)).x;
local_depth[x][y + 8] = imageLoad(depth, ivec2(uv)).x;
local_src[x][y + 16] = imageLoad(source, ivec2(uv) + ivec2(0, 8)).x;
local_depth[x][y + 16] = imageLoad(depth, ivec2(uv) + ivec2(0, 8)).x;
barrier();
@ -23,14 +28,21 @@ void main()
g1 = exp(-0.5 / (sigma * sigma));
g2 = g1 * g1;
float sum = local_src[x][y + 8] * g0;
float pixel_depth = local_depth[x][y + 8];
g0 *= g1;
g1 *= g2;
float tmp_weight, total_weight = g0;
for (int j = 1; j < 8; j++) {
sum += local_src[x][y + 8 + j] * g0;
sum += local_src[x][y + 8 - j] * g0;
tmp_weight = max(0.0, 1.0 - .001 * abs(local_depth[x][y + 8 + j] - pixel_depth));
sum += local_src[x][y + 8 + j] * g0 * tmp_weight;
total_weight += g0 * tmp_weight;
tmp_weight = max(0.0, 1.0 - .001 * abs(local_depth[x][y + 8 - j] - pixel_depth));
sum += local_src[x][y + 8 - j] * g0 * tmp_weight;
total_weight += g0 * tmp_weight;
g0 *= g1;
g1 *= g2;
}
imageStore(dest, ivec2(uv), vec4(sum));
imageStore(dest, ivec2(uv), vec4(sum / total_weight));
}

12
src/graphics/post_processing.cpp
View File

@ -425,9 +425,11 @@ void PostProcessing::renderGaussian17TapBlur(FrameBuffer &in_fbo, FrameBuffer &a
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);
glBindImageTexture(1, irr_driver->getFBO(FBO_LINEAR_DEPTH).getRTT()[0], 1, false, 0, GL_READ_ONLY, GL_R32F);
glBindImageTexture(2, auxiliary.getRTT()[0], 0, false, 0, GL_WRITE_ONLY, GL_R16F);
glUniform1i(FullScreenShader::ComputeGaussian17TapHShader::uniform_source, 0);
glUniform1i(FullScreenShader::ComputeGaussian17TapHShader::uniform_dest, 1);
glUniform1i(FullScreenShader::ComputeGaussian17TapHShader::uniform_depth, 1);
glUniform1i(FullScreenShader::ComputeGaussian17TapHShader::uniform_dest, 2);
glDispatchCompute(in_fbo.getWidth() / 8, in_fbo.getHeight() / 8, 1);
}
#endif
@ -455,9 +457,11 @@ void PostProcessing::renderGaussian17TapBlur(FrameBuffer &in_fbo, FrameBuffer &a
{
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);
glBindImageTexture(1, irr_driver->getFBO(FBO_LINEAR_DEPTH).getRTT()[0], 1, false, 0, GL_READ_ONLY, GL_R32F);
glBindImageTexture(2, in_fbo.getRTT()[0], 0, false, 0, GL_WRITE_ONLY, GL_R16F);
glUniform1i(FullScreenShader::ComputeGaussian17TapVShader::uniform_source, 0);
glUniform1i(FullScreenShader::ComputeGaussian17TapVShader::uniform_dest, 1);
glUniform1i(FullScreenShader::ComputeGaussian17TapVShader::uniform_depth, 1);
glUniform1i(FullScreenShader::ComputeGaussian17TapVShader::uniform_dest, 2);
glDispatchCompute(in_fbo.getWidth() / 8, in_fbo.getHeight() / 8, 1);
}
#endif

8
src/graphics/shaders.cpp
View File

@ -2509,13 +2509,15 @@ namespace FullScreenShader
GLuint ComputeGaussian17TapHShader::Program;
GLuint ComputeGaussian17TapHShader::uniform_source;
GLuint ComputeGaussian17TapHShader::uniform_depth;
GLuint ComputeGaussian17TapHShader::uniform_dest;
void ComputeGaussian17TapHShader::init()
{
#if WIN32
Program = LoadProgram(
GL_COMPUTE_SHADER, file_manager->getAsset("shaders/gaussian.comp").c_str());
GL_COMPUTE_SHADER, file_manager->getAsset("shaders/bilateralH.comp").c_str());
uniform_source = glGetUniformLocation(Program, "source");
uniform_depth = glGetUniformLocation(Program, "depth");
uniform_dest = glGetUniformLocation(Program, "dest");
#endif
}
@ -2564,13 +2566,15 @@ namespace FullScreenShader
GLuint ComputeGaussian17TapVShader::Program;
GLuint ComputeGaussian17TapVShader::uniform_source;
GLuint ComputeGaussian17TapVShader::uniform_depth;
GLuint ComputeGaussian17TapVShader::uniform_dest;
void ComputeGaussian17TapVShader::init()
{
#if WIN32
Program = LoadProgram(
GL_COMPUTE_SHADER, file_manager->getAsset("shaders/gaussianv.comp").c_str());
GL_COMPUTE_SHADER, file_manager->getAsset("shaders/bilateralV.comp").c_str());
uniform_source = glGetUniformLocation(Program, "source");
uniform_depth = glGetUniformLocation(Program, "depth");
uniform_dest = glGetUniformLocation(Program, "dest");
#endif
}

4
src/graphics/shaders.hpp
View File

@ -657,7 +657,7 @@ class ComputeGaussian17TapHShader
{
public:
static GLuint Program;
static GLuint uniform_source, uniform_dest;
static GLuint uniform_source, uniform_depth, uniform_dest;
static void init();
};
@ -696,7 +696,7 @@ class ComputeGaussian17TapVShader
{
public:
static GLuint Program;
static GLuint uniform_source, uniform_dest;
static GLuint uniform_source, uniform_depth, uniform_dest;
static void init();
};