Compositor: CPU vs GPU Differences #118548

New Issue

Omar Emara · 2024-02-21T09:55:35+01:00

Omar Emara commented

2024-02-21 09:55:35 +01:00

Each of the following sections describe one difference between the CPU and GPU compositors, its
problem, its potential solution, the nodes it affects, and the failed regression tests where it
manifests.

Anisotropic Filtering

Problem

GPU uses the hardware anisotropic filtering capabilities of the GPU. Since each vendor, GPU, and
driver might have a different implementation, we can't really unify the implementations.

Solution

Reimplement our own anisotropic filter.

Affected Nodes

This affects the following nodes:

Corner Pin.
Displace.
Map UV.
Plane Track.

Failed Tests

node corner pin.
node displace.
node map u v.
node plane track image.
node plane track motion.

Jitter Anti-Aliasing (blender/blender!118853)

Problem

CPU uses an 8-sample jitter multi-sample anti-aliasing algorithm to anti-alias some masks. GPU
compositor uses bilinear interpolation with zero boundary to achieve anti-aliasing.

Solution

Use SMAA in all compositor operations that require anti-aliasing.

Affected Nodes

Corner Pin.
Plane Track.

Failed Tests

node corner pin plane.
node plane track motion plane.
node plane track plane.

SMAA (blender/blender!119414)

Problem

SMAA operation produce different results between CPU and GPU. GPU uses the original SMAA library,
CPU uses a C++ port of the library.

Solution

No solution known yet. Needs investigation to figure out where the difference originates.

Affected Nodes

Anti-Alias.
ID Mask.
Z Combine.
Dilate.

Failed Tests

node dilate threshold.

Fast Gaussian

Problem

GPU does not implement the Fast Gaussian blur mode and falls back to the slow convolution algorithm.
But, CPU implementation seems broken and could better match standard Gaussian for low dynamic range
images. Further, CPU sometimes uses Fast Gaussian as a utility in other operations, like the Glare
node, while a normal blur should be used instead since the blur radius is very small.

Solution

Implement Fast Gaussian for GPU, port the same implementation to CPU.

Affected Nodes

Blur.

Failed Tests

node blur fast gaussian.
compositor-nodes-desintegrate-wipe-01.

Curve Maps (Indirectly fixed by blender/blender!118624)

Problem

CPU and GPU evaluates curve maps slightly differently, which can produce a tiny difference in the
0.001 magnitude. Further, curve maps are stored as half floats in the GPU, so we lose some
precision, making the difference worse.

Solution

Don't use the hardware sampler and interpolate the curve maps manually. Use full float for curve
maps.

Affected Nodes

RGB Curves.
Vector Curves.
Color Ramp.

Failed Tests

node curve vec.

Clipped Transformations (blender/blender!119278)

Problem

The result of some transformations is clipped in CPU, but only in background rendering, so this is a
bug in the CPU implementation rather than a difference.

Solution

Unknown, still needs investigation.

Affected Nodes

Flip.
Scale.
Rotate.
Transform.
Translate.

Failed Tests

flip scalerotate.

Linear Space Color Conversion (blender/blender!118624)

Problem

CPU produces slightly different values after sRGB to Linear color transformations, for instance,
pixels of value 1.0 becomes 1.00002408. According to Sergey:

The sRGB transform uses ExponentWithLinearTransform, which internally does SSE SIMD for pow, and that might not be as precise as per-component powf from libc. #110895

Solution

Do the color space conversion on the CPU using the same OIIO method. Or move to a more accurate method on the CPU.

Affected Nodes

Image.
Movie Clip.

Failed Tests

node keying matte.
node chroma matte.
distorted bw key.

Legacy Cryptomatte (blender/blender!118570)

Problem

GPU doesn't have an implementation.

Solution

Unclear what this node represents or if it will be removed. Just do an implementation regardless.

Affected Nodes

Legacy Cryptomatte.

Failed Tests

node cryptomatte legacy.

Fog Glare

Problem

CPU convolves the image with an arbitrary point spread function. GPU uses a Bloom algorithm.

Solution

Move Bloom as to separate option, port implementation to CPU. Add an FFT implementation for GPU and
use that to do the convolution, getting rid of the arbitrary PSF for both CPU and GPU.

Affected Nodes

Glare.

Failed Tests

node glare fog glow.
Fire2.

Denoise (blender/blender!118553)

Problem

CPU always adds the normal and albedo passes even if they are not connected, and inflates their
values to full buffers.

Solution

Only add the normal and albedo passes if they are actually connected and are full buffers.

Affected Nodes

Denoise.

Failed tests

denoise.

Interpolation

Problem

CPU and GPU interpolate images differently, but it is likely that the GPU is more correct.

Solution

Unknown, needs more investigation.

Affected Nodes

Corner Pin.
Displace.
Lens Distortion.
Map UV.
Movie Distortion.
Plane Track.
Rotate.
Translate.
Scale.
Transform.
Stabilize 2D.
Directional Blur.
Sun Beams.

Failed Tests

node corner pin.
node displace.
node lens distortion negative.
node lens distortion positive.
node map u v.
node movie distortion distort.
node movie distortion undistort.
node plane track image.
node plane track motion image.
node rotate.
node scale.
node stabilize 2d.
node stabilize 2d invert.
node transform.
node translate.
node d blur.
node sun beams.
Fire2.
compositor-nodes-desintegrate-wipe-01.
distorted bw key.
flip scalerotate.

Vector Blur (blender/blender#120135)

Problem

GPU uses an implementation similar to EEVEE, while CPU uses a more compute intensive operation.

Solution

Formalize GPU implementation and port it to CPU.

Affected Nodes

Vector Blur.

Failed Tests

node vector blur.

Variable Transformations (blender/blender#120314)

Problem

CPU allow the inputs of transform nodes to be variable, GPU does not.

Solution

Allow variable transformations for GPU.

Affected Nodes

Rotate.
Translate.
Scale.
Transform.

Failed Tests

Fire2.

Each of the following sections describe one difference between the CPU and GPU compositors, its problem, its potential solution, the nodes it affects, and the failed regression tests where it manifests. ## Anisotropic Filtering #### Problem GPU uses the hardware anisotropic filtering capabilities of the GPU. Since each vendor, GPU, and driver might have a different implementation, we can't really unify the implementations. #### Solution Reimplement our own anisotropic filter. #### Affected Nodes This affects the following nodes: - Corner Pin. - Displace. - Map UV. - Plane Track. #### Failed Tests - node corner pin. - node displace. - node map u v. - node plane track image. - node plane track motion. ## Jitter Anti-Aliasing (blender/blender!118853) #### Problem CPU uses an 8-sample jitter multi-sample anti-aliasing algorithm to anti-alias some masks. GPU compositor uses bilinear interpolation with zero boundary to achieve anti-aliasing. #### Solution Use SMAA in all compositor operations that require anti-aliasing. #### Affected Nodes - Corner Pin. - Plane Track. #### Failed Tests - node corner pin plane. - node plane track motion plane. - node plane track plane. ## SMAA (blender/blender!119414) #### Problem SMAA operation produce different results between CPU and GPU. GPU uses the original SMAA library, CPU uses a C++ port of the library. #### Solution No solution known yet. Needs investigation to figure out where the difference originates. #### Affected Nodes - Anti-Alias. - ID Mask. - Z Combine. - Dilate. #### Failed Tests - node dilate threshold. ## Fast Gaussian #### Problem GPU does not implement the Fast Gaussian blur mode and falls back to the slow convolution algorithm. But, CPU implementation seems broken and could better match standard Gaussian for low dynamic range images. Further, CPU sometimes uses Fast Gaussian as a utility in other operations, like the Glare node, while a normal blur should be used instead since the blur radius is very small. #### Solution Implement Fast Gaussian for GPU, port the same implementation to CPU. #### Affected Nodes - Blur. #### Failed Tests - node blur fast gaussian. - compositor-nodes-desintegrate-wipe-01. ## Curve Maps (Indirectly fixed by blender/blender!118624) #### Problem CPU and GPU evaluates curve maps slightly differently, which can produce a tiny difference in the 0.001 magnitude. Further, curve maps are stored as half floats in the GPU, so we lose some precision, making the difference worse. #### Solution Don't use the hardware sampler and interpolate the curve maps manually. Use full float for curve maps. #### Affected Nodes - RGB Curves. - Vector Curves. - Color Ramp. #### Failed Tests - node curve vec. ## Clipped Transformations (blender/blender!119278) #### Problem The result of some transformations is clipped in CPU, but only in background rendering, so this is a bug in the CPU implementation rather than a difference. #### Solution Unknown, still needs investigation. #### Affected Nodes - Flip. - Scale. - Rotate. - Transform. - Translate. #### Failed Tests - flip scalerotate. ## Linear Space Color Conversion (blender/blender!118624) #### Problem CPU produces slightly different values after sRGB to Linear color transformations, for instance, pixels of value 1.0 becomes 1.00002408. According to Sergey: > The sRGB transform uses `ExponentWithLinearTransform`, which internally does SSE SIMD for `pow`, and that might not be as precise as per-component `powf` from libc. https://projects.blender.org/blender/blender/pulls/110895 #### Solution Do the color space conversion on the CPU using the same OIIO method. Or move to a more accurate method on the CPU. #### Affected Nodes - Image. - Movie Clip. #### Failed Tests - node keying matte. - node chroma matte. - distorted bw key. ## Legacy Cryptomatte (blender/blender!118570) #### Problem GPU doesn't have an implementation. #### Solution Unclear what this node represents or if it will be removed. Just do an implementation regardless. #### Affected Nodes - Legacy Cryptomatte. #### Failed Tests - node cryptomatte legacy. ## Fog Glare #### Problem CPU convolves the image with an arbitrary point spread function. GPU uses a Bloom algorithm. #### Solution Move Bloom as to separate option, port implementation to CPU. Add an FFT implementation for GPU and use that to do the convolution, getting rid of the arbitrary PSF for both CPU and GPU. #### Affected Nodes - Glare. #### Failed Tests - node glare fog glow. - Fire2. ## Denoise (blender/blender!118553) #### Problem CPU always adds the normal and albedo passes even if they are not connected, and inflates their values to full buffers. #### Solution Only add the normal and albedo passes if they are actually connected and are full buffers. #### Affected Nodes - Denoise. #### Failed tests - denoise. ## Interpolation #### Problem CPU and GPU interpolate images differently, but it is likely that the GPU is more correct. #### Solution Unknown, needs more investigation. #### Affected Nodes - Corner Pin. - Displace. - Lens Distortion. - Map UV. - Movie Distortion. - Plane Track. - Rotate. - Translate. - Scale. - Transform. - Stabilize 2D. - Directional Blur. - Sun Beams. #### Failed Tests - node corner pin. - node displace. - node lens distortion negative. - node lens distortion positive. - node map u v. - node movie distortion distort. - node movie distortion undistort. - node plane track image. - node plane track motion image. - node rotate. - node scale. - node stabilize 2d. - node stabilize 2d invert. - node transform. - node translate. - node d blur. - node sun beams. - Fire2. - compositor-nodes-desintegrate-wipe-01. - distorted bw key. - flip scalerotate. ## Vector Blur (blender/blender#120135) #### Problem GPU uses an implementation similar to EEVEE, while CPU uses a more compute intensive operation. #### Solution Formalize GPU implementation and port it to CPU. #### Affected Nodes - Vector Blur. #### Failed Tests - node vector blur. ## Variable Transformations (blender/blender#120314) #### Problem CPU allow the inputs of transform nodes to be variable, GPU does not. #### Solution Allow variable transformations for GPU. #### Affected Nodes - Rotate. - Translate. - Scale. - Transform. #### Failed Tests - Fire2.

👍 1

Omar Emara added the

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Compositor: CPU vs GPU Differences #118548

Anisotropic Filtering

Problem

Solution

Affected Nodes

Failed Tests

Jitter Anti-Aliasing (blender/blender!118853)

Problem

Solution

Affected Nodes

Failed Tests

SMAA (blender/blender!119414)

Problem

Solution

Affected Nodes

Failed Tests

Fast Gaussian

Problem

Solution

Affected Nodes

Failed Tests

Curve Maps (Indirectly fixed by blender/blender!118624)

Problem

Solution

Affected Nodes

Failed Tests

Clipped Transformations (blender/blender!119278)

Problem

Solution

Affected Nodes

Failed Tests

Linear Space Color Conversion (blender/blender!118624)

Problem

Solution

Affected Nodes

Failed Tests

Legacy Cryptomatte (blender/blender!118570)

Problem

Solution

Affected Nodes

Failed Tests

Fog Glare

Problem

Solution

Affected Nodes

Failed Tests

Denoise (blender/blender!118553)

Problem

Solution

Affected Nodes

Failed tests

Interpolation

Problem

Solution

Affected Nodes

Failed Tests

Vector Blur (blender/blender#120135)

Problem

Solution

Affected Nodes

Failed Tests

Variable Transformations (blender/blender#120314)

Problem

Solution

Affected Nodes

Failed Tests