NVIDIA HBAO+ 4.0.

Overview

HBAO+ is a SSAO algorithm designed to achieve high GPU efficiency. The algorithm is based on HBAO [Bavoil and Sainz 2008], with the following differences:

1. To minimize cache trashing, HBAO+ does not use any randomization texture. Instead, the algorithm uses an Interleaved Rendering approach, generating the AO in multiple passes with a unique jitter value per pass [Bavoil and Jansen 2013].
2. To avoid over-occlusion artifacts, HBAO+ uses a simpler AO approximation than HBAO, similar to “Scalable Ambient Obscurance” [McGuire et al. 2012] [Bukowski et al. 2012].
3. To minimize flickering, the HBAO+ is always rendered in full resolution, from full-resolution depths.
4. To reduce halo artifacts behind foreground objects, HBAO+ can take as input a second depth layer, using the multi-layer SSAO approach from [Bavoil and Sainz 2009] and [McGuire et al. 2013].

Package

doc/—this documentation page.

lib/—header file, import libraries and DLLs, for Win32 and Win64.

samples/—source for sample applications demonstrating NVIDIA HBAO+.

Getting Started

1. INITIALIZE THE LIBRARY:

   GFSDK_SSAO_CustomHeap CustomHeap;
   CustomHeap.new_ = ::operator new;
   CustomHeap.delete_ = ::operator delete;
   
   GFSDK_SSAO_Status status;
   GFSDK_SSAO_Context_D3D11* pAOContext;
   status = GFSDK_SSAO_CreateContext_D3D11(pD3D11Device, &pAOContext, &CustomHeap);
   assert(status == GFSDK_SSAO_OK); // HBAO+ requires feature level 11_0 or above

2. SET INPUT DEPTHS:

   GFSDK_SSAO_InputData_D3D11 Input;
   Input.DepthData.DepthTextureType = GFSDK_SSAO_HARDWARE_DEPTHS;
   Input.DepthData.pFullResDepthTextureSRV = pDepthStencilTextureSRV;
   Input.DepthData.pFullResDepthTexture2ndLayerSRV = pDepthStencilTexture2ndLayerSRV; // Required only if Params.DualLayerAO=true
   Input.DepthData.ProjectionMatrix.Data = GFSDK_SSAO_Float4x4(pProjectionMatrix);
   Input.DepthData.ProjectionMatrix.Layout = GFSDK_SSAO_ROW_MAJOR_ORDER;
   Input.DepthData.MetersToViewSpaceUnits = SceneScale;

3. SET AO PARAMETERS:

   GFSDK_SSAO_Parameters Params;
   Params.Radius = 2.f;
   Params.Bias = 0.1f;
   Params.PowerExponent = 2.f;
   Params.Blur.Enable = true;
   Params.Blur.Radius = GFSDK_SSAO_BLUR_RADIUS_4;
   Params.Blur.Sharpness = 16.f;
   Params.DualLayerAO = true;

4. SET RENDER TARGET:

   GFSDK_SSAO_Output_D3D11 Output;
   Output.pRenderTargetView = pOutputColorRTV;
   Output.Blend.Mode = GFSDK_SSAO_OVERWRITE_RGB;

5. RENDER AO:

   status = pAOContext->RenderAO(pD3D11Context, Input, Params, Output);
   assert(status == GFSDK_SSAO_OK);

Data Flow

Input Requirements

• The library has entry points for D3D11, D3D12 and GL3.2+.
• Requires a depth texture to be provided as input, along with associated projection info.
• Optionally, can also take as input a GBuffer normal texture associated with the input depth texture:
  • Can add normal-mapping details to the AO.
  • Can be used to fix normal reconstruction artifacts with dithered LOD dissolves.
  • But makes the integration more complex. We recommend starting with input normals disabled.
• Optionally, can also take as input a viewport rectangle associated with the input textures:
  • Defines a sub-area of the input & output full-resolution textures to be sourced and rendered to.
  • The library re-allocates its internal render targets if the Viewport.Width or Viewport.Height changes for a given AO context.
• Optionally, can also take as input a second depth texture (second depth layer), to add information about the scene geometry that is hidden behind the first depth layer.
  • This enables removing halo artifacts behind foreground characters, and that with a single call to RenderAO.
  • The second depth layer can simply be a copy of the scene depth buffer before the dynamic geometry is rendered into it.

MSAA Support

• Requirements
  • The input depth & normal textures are required to have matching dimensions and MSAA sample count.
  • The output render target can have arbitrary dimensions and MSAA sample count.
• Per-Pixel AO Algorithm
  • If the input textures are MSAA, only sample 0 is used to render the AO.
  • If the output render target is MSAA, a per-pixel AO value is written to all samples.
  • In practice, we have found this strategy to not cause any objectionable artifacts, even when using HBAO+ with TXAA.

HBAO+ Pipeline

Occlusion Samples

• For each pixel, 32 occlusion samples are taken in a variable-radius disk, with a minimum of 4 full-resolution pixels between the center pixel and the sample coordinates.
• The 8 nearest samples to the center pixel are weighted by the NearAO parameter.
• The other samples are weighted by the FarAO parameter.

Parameters

AO Radius

Definition

For a given AO receiver point P and AO Radius R, sample point S is ignored if ||P-S|| > R

Impact on search area

The AO radius is a multiplier for the screen-space search radius, and the number of samples is fixed. So if the AO radius is too large, under-sampling artifacts may appear.

MetersToViewSpaceUnits

If you are not sure what to set this value to, you can:

• Set the AO Radius parameter to 1.0 and
• Increase MetersToViewSpaceUnits until the AO looks like it’s being cast up to 1 meter away

MetersToViewSpaceUnits is used internally

• To convert the AO radius parameter from meters to view-space units
• To adjust the blur sharpness parameter

Power Exponent

• The PowerExponent parameter controls the darkness of the final AO: FinalAO = pow(AO, PowerExponent).
• Typical PowerExponent values are in the range [2.0, 3.0].

AO Bias

• The AO Bias parameter can be used to hide low-tessellation artifacts
• Can also help reducing false-occlusion artifacts near screen borders
• It weights the AO contributions more strongly for samples towards the normal direction

Foreground AO

• When this feature is enabled, the screen-space AO kernel radius is:
  • inversely proportional to ViewDepth for ViewDepth > ForegroundViewDepth
  • uniform in screen-space for ViewDepth <= ForegroundViewDepth (instead of getting arbitrarily large)
• To tune the ForegroundViewDepth parameter
  • Start with ForegroundViewDepth = 0.f
  • Increase the ForegroundViewDepth to get less closer-scale occlusion on the foreground objects.

Background AO

• When this feature is enabled, the screen-space AO kernel radius is:
  • inversely proportional to ViewDepth for ViewDepth < BackgroundViewDepth
  • uniform in screen-space for ViewDepth >= BackgroundViewDepth (instead of falling off to zero).
• To tune the BackgroundViewDepth parameter
  • Start from a very large BackgroundViewDepth value (e.g. the camera’s ZFar)
  • Decrease the BackgroundViewDepth to get larger-scale occlusion in the background.

AO Blur

• Optional cross-bilateral filter
  • To remove jittered-sampling artifacts (noise)
  • To hide under-sampling artifacts (banding) & reduce jittering
• 3 Modes
  • No Blur
  • Blur Radius 2
  • Blur Radius 4

Blur Sharpness

The greater the sharpness parameter, the more the blur preserves edges

Dual Layer

The dual layer feature increases the overall quality of the AO, by removing halo artifacts behind foreground objects. To use this feature, the application needs to provide a second depth texture via Input.DepthData.pFullResDepthTexture2ndLayerSRV, and set Params.DualLayerAO = true.

Here are screenshots from our DX11 sample app, which is drawing

• the shader-ball object into a first depth buffer
• the shader-ball and background box into a second depth buffer

and feeding these 2 depth layers as input to RenderAO.

Integration Time Estimates

• Initial integration (for a rendering engineer)
  • <1 man-day with no input normals
  • 1-2 man-days with input normals
• Initial parameter tuning
  • <1 man-hour
  • Tuning the parameter should be quick once the input data are correctly fed into the library
  • The same HBAO+ parameters may be used globally across the whole game

Additional Links

[Bavoil et al. 2008] “Image-Space Horizon-Based Ambient Occlusion”

[Bavoil and Sainz 2009] “Multi-Layer Dual-Resolution Screen-Space Ambient Occlusion”

[McGuire et al. 2012] “Scalable Ambient Obscurance”

[Bukowski et al. 2012] “Scalable High-Quality Motion Blur and Ambient Occlusion”

[Bavoil and Jansen 2013] “Particle Shadows & Cache-Efficient Post-Processing”

[McGuire et al. 2013] “Lighting Deep G-Buffers: Single-Pass, Layered Depth Images with Minimum Separation Applied to Indirect Illumination”