Lesson 4 – Deferred shading
PV227 – GPU Rendering
Jiˇrí Chmelík, Jan ˇCejka
Fakulta informatiky Masarykovy univerzity
10. 10. 2016
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 1 / 23
Motivation: A loooot of lights
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 2 / 23
Forward and deferred shading
Forward shading
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 3 / 23
Forward and deferred shading
Deferred shading
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 4 / 23
Task: Create and display G-buffer
Task 1
Complete notexture_deferred_fragment.glsl and
texture_deferred_fragment.glsl, output position, normal, and
albedo.
Look at the result. Note that positions are scaled to 0.02 (our scene
is large).
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 5 / 23
Task: Create and display G-buffer
Positions Normals Albedo
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 6 / 23
Task: Evaluate the lighting
Task 2
Evaluate the lighting in evaluate_quad_fragment.glsl.
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 7 / 23
Task: Evaluate the lighting
Result
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 8 / 23
Reducing the number of evaluated lights
View-frustum culling
Light volume
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 9 / 23
Light volume implementation
Render one sphere per light, use instancing :-)
In vertex shader, place the sphere at the position of the light, and
scale it to the range of the light
In fragment shader, evaluate the light
Add all lights together using blending
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 10 / 23
Light volume implementation
Use depth test
Don’t evaluate pixels if the light is behind something.
We need the depth texture of the scene, in addition to the color
buffer.
Only compare the depth with the scene, don’t render the spheres
into the depth buffer.
Render front faces only
Use blending, configure the blend function properly
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 11 / 23
Task: Implement light volume algorithm
Task 3
Configurate the depth test, culling, and blending in C++ code
In evaluate_sphere_vertex.glsl, place the sphere to the position of
the light. Also, its radius is one, so enlarge it to the range of the light
(light_range constant).
In evaluate_sphere_fragment.glsl, do not evaluate the light yet,
output only the diffuse color of the light.
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 12 / 23
Task: Implement light volume algorithm
Result
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 13 / 23
Task: Implement light volume algorithm
Task 4
Evaluate the lighting in evaluate_sphere_fragment.glsl.
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 14 / 23
Task: Implement light volume algorithm
Result
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 15 / 23
Task: Implement light volume algorithm
Task 5
In evaluate_sphere_fragment.glsl, discard fragments that are too
far from the light.
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 16 / 23
Task: Implement light volume algorithm
Result (when displaying only the diffuse color of the light)
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 17 / 23
Pros and cons
Pros:
Lighting (one of the most expensive operations) is evaluated only
once per pixel.
Lighting is processed independently, reducing the number of
combinations of materials and lighting.
G-buffer is good for other postprocessing effects.
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 18 / 23
Pros and cons
Cons:
Does not work with multisampling.
Transparency
G-buffer needs a lot of memory (ours: 3x16B/pixel,
100 MB FullHD)
Materials must not be too much complicated (the space of
G-buffer is limited)
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 19 / 23
Things we used
OpenGL queries, very briefly:
glGenQueries generates a query object
glBeginQuery(GL_TIME_ELAPSED, obj) starts measuring of the
time, only one measuing may run at the same time.
glEndQuery(GL_TIME_ELAPSED) stops measuring.
GPU time is measured, not the CPU time.
glGetQueryObjectiv(obj, GL_QUERY_RESULT_AVAILABLE,
result) returns 1 if the result of the query is available.
glGetQueryObjectui64v(obj, GL_QUERY_RESULT, result) returns
the result of the query.
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 20 / 23
Things we used
Interpolation of variables between vertex shader and fragment shader
Usage: flat int light_idx, in vertex and fragment shader.
smooth: default interpolation, interpolates the values between
vertices with perspective correction.
noperspective: linear interpolation, i.e. without perspective
correction.
flat: no interpolation, the value of the “provoking” vertex is used.
Necessary for integers.
Interpolation without and with perspective correction. Source
Wikipedia.
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 21 / 23
Things we used
Shader storage buffer objects (SSBO):
Since OpenGL 4.3
Similar to uniform buffers, similar usage:
Use GL_SHADER_STORAGE_BUFFER instead of
GL_UNIFORM_BUFFER
Use buffer instead of uniform in shaders
Much larger than uniform buffers (at least 128 MB, but usually
limited by the memory of the GPU)
Shaders can write to them (preferably using atomic operations)
Try changing LIGHTS_MAX_COUNT and LIGHTS_STORAGE in the
project (6 places) to use shader storage buffers and 1000 lights. Don’t
use forward shading for this :-)
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 22 / 23
Next lecture
Using G-buffer for:
Screen-space ambient occlusion
Depth of field
Screen-space ambient occlusion
PV227 – GPU Rendering (FI MUNI) Lesson 4 – Deferred shading 10. 10. 2016 23 / 23