DirectCompute Accelerated Separable Filtering

DirectCompute Accelerated Separable Filtering

28th February 2011 2AMD‘s Favorite Effects

Separable Filters• Much faster than executing a box filter• Classically performed by the Pixel Shader• Consists of a horizontal and vertical pass • Source image over-sampling increases with

kernel size– Shader is usually TEX instruction limited

28th February 2011 AMD‘s Favorite Effects 3

Separable? – Who Cares • In many cases developers use this technique

even though the filter may not actually be separable– Results are often still acceptable– Much faster than performing a real box filter– Accelerates many bilateral cases


Typical Pipeline Steps


SourceRT

IntermediateRT

Destination RT

Horizontal Pass Vertical Pass

Use Bilinear HW filtering?• Bilinear filter HW can halve the number of

ALU and TEX instructions– Just need to compute the correct sampling offsets

• Not possible with more advanced filters– Usually because weighting is a dynamic operation– Think about bilateral cases...


Where to start with DirectCompute

• Is the Pixel Shader version TEX or ALU limited?– You need to know what to optimize for!– Use IHV tools to establish this

• Achieving peak performance is not easy – so write a highly configurable kernel– Will allow you to easily experiment and fine tune


Thread Group Shared Memory (TGSM)• TGSM can be used to reduce TEX ops• TGSM can also be used to cache results

– Thus saving ALU ops too

• Load a sensible run length – base this on HW wavefront/warp size (AMD = 64, NVIDIA = 32) – Choose a good common factor (multiples of 64)


Kernel #1

• Redundant compute threads 28th February 2011 AMD‘s Favorite Effects 9

...........

128 threads load 128 texels

128 – ( Kernel Radius * 2 ) threads compute results

Kernel Radius

Avoid Redundant Threads• Should ensure that all threads in a group have

useful work to do – wherever possible• Redundant threads will not be reassigned

work from another group• This would involve alot of redundancy for a

large kernel diameter28th February 2011 AMD‘s Favorite Effects 10

Kernel #2


...........


128 threads compute results

Kernel Radius

• No redundant compute threads

Kernel Radius * 2 threadsload 1 extra texel each

Multiple Pixels per Thread• Allows for natural vectorization

– 4 works well on AMD HW– Doesn‘t hurt performance on scalar HW

• Possible to cache TGSM reads on General Purpose Registers (GPRs)– Quartering TGSM reads - absolute winner!!


Kernel #3

• Compute threads not a multiple of 64 28th February 2011 AMD‘s Favorite Effects 13

...........

32 threads compute 128 results

Kernel Radius



Multiple Lines per Thread Group• Process multiple lines per thread group

– Better than one long line– 2 or 4 works well

• Improved texture cache efficiency• Compute threads back to a multiple of 64


Kernel #4


...........

...........

Kernel Radius

64 threads compute 256 results



Kernel Diameter• Kernel diameter needs to be > 7 to see a

DirectCompute win– Otherwise the overhead cancels out the

advantage

• The larger the kernel diameter the greater the win


Use Packing in TGSM• Use packing to reduce storage space required in

TGSM– Only have 32k per SIMD

• Reduces reads/writes from TGSM• Often a uint is sufficient for color filtering• Use SM5.0 instructions f32tof16(), f16tof32()28th February 2011 AMD‘s Favorite Effects 17

High Definition Ambient Occlusion


Depth + Normals

HDAO buffer

* =

Original Scene Final Scene

Perform at Half Resolution• HDAO at full resolution is expensive• Running at half resolution captures more

occlusion – and is obviously much faster• Problem: Artifacts are introduced when

combined with the full resolution scene


Bilateral Dilate & Blur


HDAO buffer doesn‘t match with scene

A bilateral dilate & blur fixes the issue

New Pipeline...


Bilinear Upsample Intermediate UAV Dilated & Blurred


½ Res Still much faster than performing at full res!

Pixel Shader vs DirectCompute


*Tested on a range of AMD and NVIDIA DX11 HW, DirectCompute is between ~2.53x to ~3.17x faster than the Pixel Shader

Depth of Field• Many techniques exist to solve this problem• A common technique is to figure out how

blurry a pixel should be– Often called the Cirle of Confusion (CoC)

• A Gaussian blur weighted by CoC is a pretty efficient way to implement this effect


The Pipeline...


Intermediate UAV

CoC



Shogun 2: DoF OFF


Shogun 2: DoF ON

Pixel Shader vs DirectCompute


*Tested on a range of AMD and NVIDIA DX11 HW, DirectCompute is between ~1.48x to ~1.86x faster than the Pixel Shader

Summary• DirectCompute greatly accelerates larger kernel diameter

filters• Allows for filtering at full resolution• For access to source code:

– HDAO11: [email protected]– DoF11: [email protected]


[email protected]

[email protected]@amd.com

Please fill in the feedback forms!28th February 2011 29AMD‘s Favorite Effects

Documents

DirectCompute Accelerated Separable Filtering