Triangle Visibility buffer

The filtered and culled
Visibility Buffer
Wolfgang Engel
Confetti
October 13th, 2016

Table of contents
• The Visibility Buffer
• Cluster Culling / Triangle Filtering
• Re-using triangle filtered results for multiple rendering passes

Motivation
• Forward rendering shades all fragments in triangle-submission order
• Wastes rendering power on pixels that don’t contribute to the final image
• Deferred shading solves this problem in 2 steps:
• First, surface attributes are stored in screen buffers -> G-Buffer
• Second, shading is computed for visible fragments only
• However, deferred shading increases memory bandwidth
consumption:
• Screen buffers for: normal, depth, albedo, material ID,…
• G-Buffer size becomes challenging at high resolutions

High-Level View 2009
G-Buffer (taken from [Engel2009] Killzone 2 layout)
Depth Buffer
Deferred
Lighting
Forward
Rendering
Switch off depth write
Specular /
Motion Vec
Normals
Albedo /
Shadow
Render opaque objects Transparent objects
Sort Back-To-Front
8:8:8:8 8:8:8:8 8:8:8:8 32-bit

G-Buffer – Frostbite Engine [Lagarde]
10:10:10:2
8:8:8:8
8:8:8:8
11:11:10
Depth - R32G8X24_TYPELESS

G-Buffer – Frostbite Engine [Lagarde]
Depth Buffer
Deferred
Lighting
Forward
Rendering
Switch off depth write
BaseColor etc.Normals etc. MetalMask etc.
Render opaque objects Transparent objects
Sort Back-To-Front
10:10:10:2 8:8:8:8 8:8:8:8 32-bit
IBL / GI
11:11:10

High-Level View
Visibility Buffer (similar to [Burns][Schied])
Depth Buffer
Vertex Buffer
Visibility Vertex Buffer holds opaque
and transparent objects
Lighting
Forward
Rendering
Transparent objects
Sort Back-To-Front
Encodes
- 1 bit alpha-mask
- 8-bit drawID
- 23-bit triangleID / primID
8:8:8:8

Visibility Buffer PC 1080p – Memory
Memory Description NoMSAA 2xMSAA 4xMSAA
Visibility Buffer keeps address of each triangle in 32-bit per pixel
4 bytes * 1920 * 1080
7.9 MB 15.8 MB 31.6* MB
Depth Buffer 4 byte * 1920 * 1080 7.9 MB 15.8 MB 31.6* MB
Hierarchical Z 4 byte * 1920 * 1080 * 1/64 0.12 MB - -
Vertex Buffer 28 bytes per vertex + padding
float3 position;
uint normal;
uint tangent;
uint texCoord;
uint materialID;
uint pad; // to align to 128 bits – 32 byte for NVIDIA
Does not increase with resolution.
Textures 21* MB - -
Draw arguments, Uniform, Descriptors etc. 2* MB - -
Overall 38.80* MB 54.60* MB 86.20* MB
* rough estimate; driver might increase it

G-Buffer PC 1080p – Memory
Normals 10:10:10:2 - 4 bytes * 1920 * 1080 7.9 MB 15.8* MB 31.6* MB
PBR 8:8:8:8 - 4 bytes * 1920 * 1080 7.9 MB 15.8* MB 31.6* MB
Albedo 8:8:8:8 - 4 bytes * 1920 * 1080 7.9 MB 15.8* MB 31.6* MB
GI / IBL 11:11:10 - 4 bytes * 1920 * 1080 7.9 MB 15.8* MB 31.6* MB
Depth 8:8:8:8 - 4 bytes * 1920 * 1080 7.9 MB 15.8* MB 31.6* MB
Hierarchical Z 4 bytes * 1920 * 1080 * 1/64 0.49 MB - -
Draw arguments. Etc. 2* MB
Overall 41.99* MB 81.49* MB 160.49* MB

Visibility Buffer PC 4k – Memory
4 bytes * 3840 * 2160
31.64 MB 63.28* MB 126.56* MB
Depth Buffer 4 bytes * 3840 * 2160 31.64 MB 63.28* MB 126.56* MB
Vertex Buffer 28 bytes per vertex + padding
float3 position;
uint normal;
uint tangent;
uint texCoord;
uint materialID;
uint pad; // to align to 128 bits – 32 byte for NVIDIA
Does not increase with resolution.
- -
Textures 21* MB
Draw arguments, Uniform, Descriptors etc. 2* MB
Overall 86.77 MB 150.05 MB 276.61 MB

G-Buffer PC 4k – Memory
Normals 10:10:10:2 - 4 bytes * 3840 * 2160 31.64 MB 63.28* MB 126.56* MB
PBR 8:8:8:8 - 4 bytes * 3840 * 2160 31.64 MB 63.28* MB 126.56* MB
Albedo 8:8:8:8 - 4 bytes * 3840 * 2160 31.64 MB 63.28* MB 126.56* MB
GI / IBL 11:11:10 - 4 bytes * 3840 * 2160 31.64 MB 63.28* MB 126.56* MB
Depth 8:8:8:8 - 4 bytes * 3840 * 2160 31.64 MB 63.28* MB 126.56* MB
Draw arguments. Etc. 2* MB - -
Overall 160.69* MB 318.89* MB 635.29* MB

Visibility Buffer XOne 1080p – Memory
ESRAM Description NoMSAA 2xMSAA
4 bytes * 1920 * 1080
7.9 MB 15.8 MB
Depth Buffer 4 byte * 1920 * 1080 7.9 MB 15.8 MB
Hierarchical Z 4 byte * 1920 * 1080 * 1/64 0.12 MB -

Visibility Buffer – Filling pseudocode
• Visibility Buffer generation step
• For each pixel on screen:
• Pack (alpha masked bit, drawID, primitiveID) into 1 32-bit UINT
• Write that into a screen-sized buffer
• The tuple (alpha masked bit, drawID, primitiveID) will allow a shader to access the
triangle data in the shading step

Visibility Buffer – Shader code
// in visibilityPass.shd
uint calculateOutputVBID(bool opaque, uint drawID, uint primitiveID){
uint drawID_primID = ((drawID << 23) & 0x7F800000) | (primitiveID & 0x007FFFFF);
if (opaque)
return drawID_primID;
else
return (1 << 31) | drawID_primID;
}
float4 unpackUnorm4x8(uint p){
return float4(float(p & 0x000000FF) / 255.0, float((p & 0x0000FF00) >> 8) / 255.0,
float((p & 0x00FF0000) >> 16) / 255.0, float((p & 0xFF000000) >> 24) / 255.0);
}
cbuffer RootConstants : register(b0, space0){
uint DrawID;
};
float4 main(uint primitiveID : SV_PrimitiveID) : SV_Target{
return unpackUnorm4x8(calculateOutputVBID(true,DrawID,primitiveID));
}

Visibility Buffer – Shading pseudocode
• For each pixel in screen-space we do:
1. Get drawID/triangleID at pixel pos
2. Load data for the 3 vertices from the VB
3. Compute the partial derivatives of the barycentric coordinates – triangle gradients
i. Compute x and y value of each vertex of a triangle in projected screen-space
ii. Compute Barycentric Coordinates
iii. Compute partial derivatives of the barycentric coordinates
4. Interpolate vertex attributes at pixelpos using gradients (could do triangle / object-
space lighting)
a) Attribs use w from position to compute perspective correct interpolation
b) MVP matrix is applied to position
5. We have all data ready: shade and calculate final color

Barycentric coordinates λ1, λ2, λ3 on an equilateral triangle
and on a right triangle

/* taken from computeDerivaties.shd
void computeBarycentricGradients(float2 v[3], out float3 db_dx, out float3 db_dy){
float d = 1.0 / determinant(float2x2(v[2] - v[1], v[0] - v[1]));
db_dx = float3(v[1].y - v[2].y, v[2].y - v[0].y, v[0].y - v[1].y) * d;
db_dy = float3(v[2].x - v[1].x, v[0].x - v[2].x, v[1].x - v[0].x) * d;
}*/
// in shadeSun.shd
float3 one_over_w = 1.0 / float3(position0.w, position1.w, position2.w);
// projected vertices
float2 pos_scr[3] = {position0.xy *one_over_w.x, position1.xy *one_over_w.y, position2.xy *one_over_w.z };
float3 db_dx, db_dy;
// gradient barycentric coords x/y
computeBarycentricGradients(pos_scr, db_dx, db_dy);
Following [Schied2015] Appendix A Equation (4):

3. Compute the partial derivates of the barycentric coordinates – triangle
gradidents
4. Interpolate vertex attributes at pixelpos using gradients (could do triangle /
object-space lighting)

// in shadeSun.shd
float3 interpolateAttribute(float3x3 attributes, float3 db_dx, float3 db_dy, float2 d){
float3 attribute_x = mul(db_dx, attributes);
float3 attribute_y = mul(db_dy, attributes);
float3 attribute_s = attributes[0];
return (attribute_s + d.x * attribute_x + d.y * attribute_y);
}
float2 d = In.screenPos + -pos_scr[0];
float w = 1.0 / interpolateAttribute(one_over_w, db_dx, db_dy, d);
float z = w * projection[2][2] + projection[2][3];
float3 position = mul(invVP, float4(In.screenPos * w, z, w)).xyz;

3. Compute the partial derivates of the barycentric coordinates – triangle
gradidents
4. Interpolate vertex attributes at pixelpos using gradients (could do triangle /
object-space lighting)
-> we skip the shading source code here … just Blinn-Phong

Visibility Buffer - Benefits
• Better decouples visibility from shading
• Calculating derivatives can be done separate from the shading phase (we
include it in the moment)
• We can shade then with different frequency or quality
• Improves memory efficiency
• Improves cache utilization
• Memory accesses are highly coherent
 high cache hit rates
• A G-Buffer needs to store data per screen-space pixel. Compared to a vertex / index buffer some
of this data is redundant
-> we can see 99% L2 cache hits for the Visibility Buffer for textures, vertex and index buffers

Visibility Buffer - Benefits
• Stores less data for complex lighting models like e.g. PBR compared to a
G-Buffer
• PBR data for Visibility Buffer is a struct in constant memory indexed by material id in
vertex structure
• This struct holds indices into a texture array for various PBR textures
• It also holds per-material descriptions of what is necessary to drive BRDF
• Any data that changes per-pixel is stored in textures that are referenced by the struct
• Decouples G-buffer footprint from screen resolution
• Improves performance at high resolutions: 2K, 4K, MSAA …
• Improves performance on bandwidth-limited platforms

Visibility Buffer – Triangle Counts
Triangles rendered after culling
• 1.87 Million triangles main view
• 2.40 Million triangles shadow map
San Miguel Scene
• 8 Million Triangles
• 5 Million Vertices
Modern Game in 2016
1.8 Million triangles combined in Ultra

How do you do lighting?
• We calculate for each pixel the lighting based on the triangle
data and therefore replace the whole rendering pipeline incl.
the vertex shader, primitive assembly and rasterizer in the
pixel shader
• Because all the triangle data is available, Object-Space lighting
should be possible
-> the idea is that the whole triangle is lit in object-space and
the result is cached

What about Tessellation?
•Following [Wihlidal][Brainerd] this is a pre-step before
or in parallel to Triangle filtering

Why didn’t we implement this earlier?
•Two recent developments made the Visibility Buffer
more attractive compared to a G-Buffer
• DirectX 12 / Vulkan with ExecuteIndirect or multi draw indirect
• Triangle culling / filtering [EDGE][Chajdas] [Wihlidal]
… see the next slides …

Cluster Culling / Triangle filtering

Motivation
• Polygonal complexity of games increases every year
• Efficient triangle removal is an important aspect
• 2 culling stages
• Cluster culling : cull groups of triangles before sending them to the GPU
(following [Chajdas] on the CPU; [Wihlidal] on GPU)
• Triangle Filtering: cull individual triangles after being sent to the GPU

Cluster culling
• Groups triangles in small chunks of
256 triangles with similar
orientations
• Chunks have a model matrix
associated (they can be moved
around)
• Each chunk must pass a quick
visibility test before being sent to
the GPU:
• Cone test

Cone test for fast cluster culling
Triangles
Normals
If the eye is in the safe area then
we can NOT see any triangle
because they are back-facing
Exclusion volume
Triangle cluster

Exclusion volume
Triangles
Normals
First, locate the center of
the cluster

Exclusion volume Normals
Then, negatively accumulate
normal starting at cluster
center

Negatively accumulate
second normal after the first
one

Accumulate next
one

After accumulating the last
one we have the starting
point of the exclusion
volume
and the direction

Exclusion volume
This is the calculated
exclusion volume If the eye is in this area then
the eye can NOT see any
triangle
Most restrictive triangle
planes used to calculate cone
open angle
Cone angle

Cluster Culling
• Effectivity depends on the orientation of the faces
in the cluster
• The more similar the orientations, the bigger the exclusion /
culling volume
• Depending on the triangles the exclusion volume can
not be calculated
• Invalid cluster for cluster culling  just pass it!
Invalid cluster  cluster culling not
possible

Cluster Culling
• Code can be found in TriangleFiltering.h
• CreateClusters() // creates triangle clusters
• AddRequest() // tests the triangle clusters

Compute-based triangle filtering
• Motivation:
• Filter triangles before they go into the graphics pipeline
• Use the unused compute units during graphics pipeline execution with async
compute
• Compute-based filter  one triangle per thread
• Degenerate triangle culling
• Back-face culling
• Frustum culling
• Small primitives culling
• Depth culling (requires coarse depth buffer)(not in this demo)
• Triangle indices that pass these tests are appended to index buffer

• Degenerate triangle culling
• Allows to cull invisible zero-area triangles
• Cost: quick test (discard if at least two triangle indices are equal)
• Effectiveness: low
// in filterTriangles.shd
#if ENABLE_CULL_INDEX
if ( indices[0] == indices[1]
|| indices[1] == indices[2]
|| indices[0] == indices[2])
{
cull = true;
}
#endif

• Back-face culling
• Allows to cull triangles that face away from the viewer
• If tessellation is used must take into account max patch height
• Cost: calculate the determinant of a 3x3 matrix [Olano] (homogeneous 2D
coordinates)
• Effectiveness: high (potentially cull 50% of the geometry)

#if ENABLE_CULL_BACKFACE
// Culling in homogenous coordinates
// Read: "Triangle Scan Conversion using 2D Homogeneous Coordinates"
// by Marc Olano, Trey Greer
// http://www.cs.unc.edu/~olano/papers/2dh-tri/2dh-tri.pdf
float3x3 m = float3x3(vertices[0].xyw, vertices[1].xyw, vertices[2].xyw);
if (cullBackFace)
cull = cull || (determinant(m) > 0);
#endif

• Near Plane clipping -> triangles behind the near clipping
plane of the frustum need to be clipped
#if ENABLE_CULL_FRUSTUM || ENABLE_CULL_SMALL_PRIMITIVES
int verticesInFrontOfNearPlane = 0;
// Transform vertices[i].xy into normalized 0..1 screen space
for (uint i = 0; i < 3; ++i) {
if (vertices[i].w < 0) { // check for behind near clipping plane
++verticesInFrontOfNearPlane;
// flip the w so that any triangle that stradles the plane wont be projected onto
// two sides of the frustrum
vertices[i].w *= (-1.f);
}
vertices[i].xy /= vertices[i].w * 2;
vertices[i].xy += float2(0.5, 0.5);
}
#endif

if (verticesInFrontOfNearPlane == 3)
{
cull = true;
}
• Part 2 source code – triangle behind near clipping plane

• Frustum culling
• Allows to cull triangles that are projected
outside the clipping cube
• Takes into account near and far planes
• Cost: check if all vertices lie in the negative
side of the clip-space cube
• Effectiveness: medium-high (depends on
the size of the scene and eye pos)

#if ENABLE_CULL_FRUSTUM
{
if (verticesInFrontOfNearPlane == 3)
{
cull = true;
}
float minx = min (min (vertices[0].x, vertices[1].x), vertices[2].x);
float miny = min (min (vertices[0].y, vertices[1].y), vertices[2].y);
float maxx = max (max (vertices[0].x, vertices[1].x), vertices[2].x);
float maxy = max (max (vertices[0].y, vertices[1].y), vertices[2].y);
cull = cull || (maxx < 0) || (maxy < 0) || (minx > 1) || (miny > 1);
}
#endif

• Small-primitives culling
• Allows to cull triangles that are too small to be seen
• Triangles that do not touch any sample point after projection
• Long and thin triangles that do not touch any sample are culled as well
• More efficient use of hardware resources
• Cost: triangle touches any subpixel samples
• Effectiveness: medium (depends on the size of the triangles and
screen res)
Primitive-rate bound
• only one primitive per cycle per tile can be
scanned (see [Wihlidal])
• Very inefficient use of rasterization units

• Implementation is a two-step approach
• Test if one of the vertices is outside the Guardband
-> can’t be a small triangle
• If all vertices of a triangle are inside the Guardband
-> do the actual small size test

• Guard band clipping
Usually part of rasterization: there is a guard band around the
viewport
• Medium gray triangles -> outside viewport -> rejected
• Light gray triangles -> completely in the guard band -> rasterized and the pixels outside of the
viewport are ignored by rasterizer
• Dark triangle needs to be clipped

• Test if vertex outside the Guardband
int2 minBB = int2(1 << 30, 1 << 30);
int2 maxBB = int2(-(1 << 30), -(1 << 30));
bool insideGuardBand = true;
for (uint i = 0; i < 3; ++i) {
float2 screenSpacePositionFP = vertices[i].xy * windowSize;
// Check if we would overflow after conversion
if (screenSpacePositionFP.x < -(1 << 23)
|| screenSpacePositionFP.x > (1 << 23)
|| screenSpacePositionFP.y < -(1 << 23)
|| screenSpacePositionFP.y > (1 << 23))
{ insideGuardBand = false;
}
// scale based on distance to from center to msaa sample point
int2 screenSpacePosition = int2(screenSpacePositionFP * (SUBPIXEL_SAMPLES * samples));
minBB = min (screenSpacePosition, minBB);
maxBB = max (screenSpacePosition, maxBB);
}

• Is the triangle of “small” size?
if (insideGuardBand)
{
const uint SUBPIXEL_SAMPLE_CENTER = SUBPIXEL_SAMPLES / 2;
const uint SUBPIXEL_SAMPLE_SIZE = SUBPIXEL_SAMPLES - 1;
/*
Test is:
Is the minimum of the bounding box right or above the sample
point and is the width less than the pixel width in samples in
one direction.
This will also cull very long triangles which fall between
multiple samples.
*/
cull = cull || any( ((minBB & SUBPIXEL_MASK) > SUBPIXEL_SAMPLE_CENTER)
&& ((maxBB - ((minBB & ~SUBPIXEL_MASK) + SUBPIXEL_SAMPLE_CENTER))
< (SUBPIXEL_SAMPLE_SIZE)));
}

• Depth culling (not used in this
demo)
• Allows to cull triangles that are occluded
by the scene
• This test requires a coarse depth buffer
• Cost: load depth values from map and
check triangle/BB intersection
• Effectiveness: medium-high (depends
on scene complexity and the size of the
triangles)
Can be generated by
• Downsampling previous z-buffer and
reprojecting depths
• Rendering selected LOD geometry at
low res

Degenerated triangles
Backfaces
Frustum
Small primitives
Depth
Culling tests Culled results Compaction
• Triangle filtering is executed on groups of 256 triangles (one batch) ->
empty draws
• Draw batch compaction to the rescue
• Can be run in parallel in a compute shader
• Eliminates empty draws from the multi indirect draw buffer
Draw this using
multi draw indirect / ExecuteIndirect

Compute-based triangle filtering - Frame
pseudocode
1. [CPU] Early discard invisible geometry using triangle cluster culling
2. [CS] Generate unculled indices and multi draw indirect buffers using
triangle filtering (one triangle per thread)
3. Like before

Triangle cluster culling / filtering – Draw Calls
• For this static scene one large vertex buffer and an index buffer
generated by triangle culling and filtering is used
• Draw batches that hold a block of geometry each for one material
-> Only two “materials” opaque and alpha masked, transparent objects
and other materials would go into the same buffer
• For dynamic objects we would use a dedicated VB/IB pair for each; this
is optional

• ExecuteIndirect makes it possible to completely defer the workload to
the GPU, so that the actual work items can be batched in a more
coherent way
• Grouping multiple draw calls into a single one
• reduces CPU overhead -> only a single Graphics API call
• reduces GPU overhead (alleviates pressure on VS, CP and triangle assembly) by
letting the user provide an optimal set of workload through
• culling indices == invisible trianges through async compute
• culling draw calls -> the resulting indirect argument buffer only includes valid draw calls

San Miguel Scene Number of Draw calls
(ExecuteIndirect)
• Shadow opaque 214
• Shadow alpha masked 59
• Main view opaque 200
• Main view alpha masked 60
• Dispatch calls for filtering 81

Triangle cluster culling / filtering – Benefits
• Allows to cull triangles before sending them to the graphics
pipeline
• Avoid overwhelming parts of the graphics pipeline (rasterizer)
• Graphics pipeline is better utilized with the visible triangles
(rasterizer efficiency, command processor,…)
• Can make use of async compute to potentially overlap with
the graphics pipeline

Re-using triangle filtered results
for multiple Views / Rendering
Passes

Motivation
• Compute-based triangle filtering comes at a cost
• For every triangle: load indices and vertices, transform vertices,
append (lock) triangle data to index buffer
• We came up with the idea to generate filtered data for several rendering
passes like main view, shadows etc.
• Use filtered data to cull the same triangle set from different views
• Load indices/vertices, transform vertices only once for all views

Motivation
• Reduces the effectivity of cluster culling / triangle filtering
• harder to cull cluster for N views
• however, it was worth it:

Use filtered data to cull triangles from different views
• The algorithm is generalized to test against different N-views
• Load indices / vertices once, transform vertices for every view
Algorithm
unculledClusters  ClusterCulling(sceneObjs, views)
filteredIndicesArray  TriangleCulling(unculledClusters, views)
ShadowPass( filteredIndices[shadowView] )
MainPass1( filteredIndices[mainView] )
CPU
GPU
GPU
GPU
Compute
Graphics
Graphics
MainPass2( filteredIndices[mainView] )GPU Graphics

Adding re-usage of triangle filtered data - Frame pseudocode
1. [CPU] Early discard geometry not visible from any view using cluster
culling
2. [CS] Generate N index and N multi draw indirect / ExecuteIndirect
buffers using triangle filtering testing against the N views (one
triangle per thread)
3. For each i view use (ith index buffer and ith ExecuteIndirect buffer):
1. [Gfx] Clear visibility and depth buffers
2. [VS,PS] Visibility buffer pass
[PS] Output triangle / instance IDs
3. [PS] Interpolate attributes from gradients and shade pixel * Using a dedicated Visibility
Buffer for shadow pass is
overkill, but you can still use the
filtered data for it.

Results San Miguel Scene average for main view
 8 Million triangles
 5 Million vertices
Total triangles Rendered Culled
8,010,146 851,517 (10.6%) 7,158,629 (89.4%)
3,185,203 (39.8%) Back-face
5,244,787 (65.5%) Frustum
1,950,030 (24.4%) Small primitives

GPU Culling Shadow
Map
Fill VB HDAO Shade VB Resolve
MSAA
UI Overall
Visibility Buffer 4k – No MSAA
AMD RADEON R9 380 2.66 1.42 5.07 2.51 3.43 - 0.02 15.19
NVIDIA GeForce GTX 970 3.46 1.12 3.36 1.82 3.29 - 0.02 13.52
Visibility Buffer 4k – No MSAA No Culling
AMD RADEON R9 380 - 5.18 9.25 2.51 3.43 - 0.02 20.45
NVIDIA GeForce GTX 970 - 3.74 5.31 1.79 3.25 - 0.02 14.39
Visibility Buffer 4k – 2x MSAA
AMD RADEON R9 380 2.72 1.42 8.27 3.65 8.27 2.29 0.03 25.87
NVIDIA GeForce GTX 970 3.34 1.09 6.17 3.58 6.17 0.34 0.02 19.87
Visibility Buffer 4k – 4x MSAA
AMD RADEON R9 380 2.70 1.43 8.73 5.70 15.58 3.61 0.03 37.86
Visibility Buffer 3840 x2160

GPU Culling Shadow
Map
Fill Buffer HDAO Shade Buffer Resolve
MSAA
UI Overall
Deferred Shading 4k – No MSAA
AMD RADEON R9 380 2.67 1.42 12.19 2.51 1.29 - 0.03 20.19
NVIDIA GeForce GTX 970 3.36 1.20 9.04 1.79 1.21 - 0.02 16.82
Deferred Shading 4k – No MSAA No Culling
AMD RADEON R9 380 - 5.18 15.00 2.51 1.29 - 0.03 24.06
NVIDIA GeForce GTX 970 - 3.75 10.44 1.79 1.22 - 0.02 17.49
Deferred Shading 4k – 2x MSAA
AMD RADEON R9 380 2.70 1.42 21.65 3.65 10.86 2.29 0.03 42.68
Deferred Shading 4k – 4x MSAA
AMD RADEON R9 380 2.72 1.44 35.88 5.74 20.13 3.60 0.02 69.64
Deferred Shading 3840 x2160

GPU Culling Shadow
Map
Fill VB HDAO Shade VB Resolve
MSAA
UI Overall
Visibility Buffer 1080p – No MSAA
Xbox One 7.19 3.32 3.90 1.73 3.90 - 0.02 19.78
Visibility Buffer 1080p – No MSAA No Culling
Xbox One - 9.16 9.09 1.73 3.89 - 0.02 23.98
Visibility Buffer 1080p – 2x MSAA
Xbox One 7.28 3.20 7.57 5.17 7.57 0.46 0.02 28.00
XBOX One - Visibility Buffer 1080p

GPU Culling Shadow
Map
Fill Buffer HDAO Shade Buffer Resolve
MSAA
UI Overall
Deferred Shading 1080p – No MSAA
Xbox One 7.19 3.14 11.18 1.74 1.38 - 0.02 24.77
Deferred Shading 1080p – No MSAA No Culling
Xbox One - 9.09 15.07 1.73 1.39 - 0.02 27.45
Deferred Shading 1080p – 2x MSAA
Xbox One 7.21 3.18 21.85 5.18 8.46 0.47 0.02 46.41
XBOX One - Deferred Shading 1080p

Deferred Shading
GPU
AMD RADEON R9 380
1080p 1440p 2160p
No MSAA 9.75 12.30 20.19
No MSAA – No Culling 14.16 16.66 24.06
2x MSAA 16.16 23.09 42.68
4x MSAA 24.90 36.37 69.64
NVIDIA GeForce GTX
970
1080p 1440p 2160p
No MSAA 8.72 10.67 16.82
2x MSAA 11.58 15.23 26.27
4x MSAA 17.00 24.76 48.12
GPU
AMD RADEON R9 380
1080p 1440p 2160p
No MSAA 8.57 10.72 15.19
2x MSAA 11.44 16.38 25.87
4x MSAA 15.27 20.82 37.86
Visibility Buffer
Summary
Xbox One 1080p 1440p 2160p
No MSAA 24.77 - -
No MSAA – No Culling 27.45 - -
2x MSAA 46.41 - -
Xbox One 1080p 1440p 2160p
No MSAA 19.78 - -
No MSAA – No Culling 23.98 - -
2x MSAA 28.00 - -
NVIDIA GeForce GTX
970
1080p 1440p 2160p
No MSAA 7.68 8.83 13.52
2x MSAA 9.63 12.21 19.87
4x MSAA 12.47 17.47 32.68

How about VR?
•We are working on the StarVR SDK. StarVR uses a large
field of view with very high resolution
•The Visibility Buffer helps substantially with
performance here
• We cull and prepare the data for all views and the shadow map views in
one go

Executive Summary
We built a rendering system that
• Cluster culls and filters triangles for different views like main
view, shadow view, reflection view, GI view etc. at the same time
• The optimized triangles are used to fill a screen-space Visibility
Buffer or more Visibility Buffers for more views
• We then render lights, shadows, bounce lights with the
optimized geometry based on visibility
• We can differ between visibility of geometry and shading
frequency
• We can light per triangle or in so called object space

Future work
• Re-use culled triangles over several frames
• Use intrinsics for several parts of the pipeline [Chajdas
Compaction]
• Better improve asynchronous scheduling
• Async compute is powerful

Source Code
• Source code: https://we.tl/MOdVkljptr
• Probably later on github if we can fit everything on there  (1 GB
limit)

Credits
• Christoph Schied – wrote implementation of his paper with the OpenGL 4.5 run-time at our office
• Confetti People
• Marijn Tamis – wrote the initial OpenGL 4.5 run-time
• Leroy Sikkes – wrote the initial DirectX 12 run-time and added hardware performance
counters
• Max Oomen (intern) added linear lighting and fixed many bugs
• Jesús Gumbau – added triangle filtering, came up with the idea and implemented re-usage of
filtered triangle data and made it cross-platform running on NVIDIA and AMD GPUs and then
brought it to DirectX 12
• Jordan Logan – brought it to XBOX One and optimized for this console

Credits
• Most of the code for triangle culling and filtering is based on
[Chajdas]. We added three features:
• MSAA support for small triangle removal
• Triangle in front of Near Plane culling
• Multi-Viewport support

Acknowledgements
• Graham Wihlidal DICE Frostbite
• Nicolas Thibieroz, Gareth Thomas, Matthaeus Chajdas, Steven Tovey AMD
• Mike Acton Insomniac
• James McLaren, Q-Games
• Remi Arnaud, Starbreeze / StarVR
• Kev Gee Microsoft

References
• [Burns] Christopher A. Burns, Warren A. Hunt “The Visibility Buffer: A Cache-Friendly Approach to Deferred Shading” Journal of Computer
Graphics Techniques (JCGT) 2:2 (2013), 55- 69. Available online at http://jcgt.org/published/0002/02/04
• [Chajdas] Matthaeus Chajdas “GeometryFX” http://gpuopen.com/gaming-product/geometryfx/
• [Chajdas Compaction] Matthaeus Chajdas “Fast compaction with mbcnt”, http://gpuopen.com/fast-compaction-with-mbcnt/
• [Edge] Edge Library, PS3 SDK
• [Engel2009] Wolfgang Engel, “Light Pre-Pass”, “Advances in Real-Time Rendering in 3D Graphics and Games”, SIGGRAPH 2009,
http://halo.bungie.net/news/content.aspx?link=Siggraph_09
• [Lagarde] Sebastien Lagarde, Charles de Rousiers, “Moving Frostbite to Physically Based Rendering”, Course notes SIGGRAPH 2014
• [Olano] Marc Olano, http://www.cs.unc.edu/~olano/papers/2dh-tri/2dh-tri.pdf
• [Schied2015] Christoph Schied, Carsten Dachsbacher “Deferred Attribute Interpolation for Memory-Efficient Deferred Shading”,
http://cg.ivd.kit.edu/publications/2015/dais/DAIS.pdf
• [Schied2016] Christoph Schied, Carten Dachsbacher “Deferred Attribute Interpolation Shading”, GPU Pro 7, CRC Press / shorter and free
version: http://cg.ivd.kit.edu/publications/2015/dais/DAIS.pdf
• [Wihlidal] Graham Wihlidal, “Optimizing the Graphics Pipeline with Compute”, GDC 2016, http://www.frostbite.com/2016/03/optimizing-
the-graphics-pipeline-with-compute/

Triangle Visibility buffer

More Related Content

What's hot (20)

Similar to Triangle Visibility buffer (20)

More from Wolfgang Engel (13)

Recently uploaded (20)

Triangle Visibility buffer

Editor's Notes