Joint denoising and supersampling of graphics data
Joint denoising and supersampling of graphics data is described. An example of a graphics processor includes multiple processing resources, including a least a first processing resource including a pipeline to perform a supersampling operation; and the pipeline including circuitry to jointly perform denoising and supersampling of received ray tracing input data, the circuitry including first circuitry to receive input data associated with an input block for a neural network, second circuitry to perform operations associated with a feature extraction and kernel prediction network of the neural network, and third circuitry to perform operations associated with a filtering block of the neural network.
1. A graphics processor comprising:
a plurality of processing resources, including a least a first processing resource including a pipeline to perform a supersampling operation; and
the pipeline including circuitry to jointly perform denoising and supersampling of received ray tracing input data in a single neural network, the circuitry including:
first circuitry to receive input data associated with an input block for the neural network and to upsample the received input data from a first resolution to a second resolution,
second circuitry to perform operations associated with a feature extraction and kernel prediction network of the neural network, the feature extraction and kernel prediction network including operation of a single shared feature extractor for the neural network, and
third circuitry to perform operations associated with a filtering block of the neural network.
2. The graphics processor of claim 1 , wherein:
the feature extraction and kernel prediction network includes a plurality of convolution layers, the second circuitry configured to perform operations associated with the plurality of convolution layers; and
the plurality of convolution layers operate at a first precision and the filtering block operates at a second precision, the first precision being lower than the second precision.
3. The graphics processor of claim 2 , wherein the first precision includes either INT4 or INT8.
4. The graphics processor of claim 1 , wherein the first circuitry is further to:
warp a previous output from the third circuitry using motion data to generate warped data; and
provide the warped data to the second circuitry for processing.
5. The graphics processor of claim 1 , wherein the neural network is a mixed-precision convolutional neural network (CNN).
6. The graphics processor of claim 1 , wherein the input block and the filtering block operate in a first resolution and the feature extractor runs in a second resolution, the second resolution being lower than the first resolution.
7. The graphics processor of claim 1 , wherein the filtering block includes a plurality of separate filter paths for different components of graphics data.
8. The graphics processor of claim 7 , wherein the plurality of separate filtering paths includes:
an albedo filtering path;
a specular filtering path; and
a diffuse filtering path.
9. A method comprising:
receiving ray tracing input data at a processing resource and upsampling the received input data from a first resolution to a second resolution, the processing resource including a pipeline to perform a supersampling operation, the pipeline including circuitry associated with operations of a single neural network;
processing the received ray tracing input data at the neural network to perform both denoising and supersampling of the input data, the processing including feature extraction and kernel prediction, the feature extraction including operation of a single shared feature extractor for the neural network; and
filtering the processed data at a filtering block.
10. The method of claim 9 , wherein:
processing the received ray tracing input data at the neural network includes processing utilizing a plurality of convolution layers; and
the plurality of convolution layers operate on a first precision and the filtering block operates on a second precision, the first precision being lower than the second precision.
11. The method of claim 9 , further comprising:
warping a previous output from the pipeline using motion data to generate warped data; and
providing the warped data to the neural network for processing.
12. The method of claim 9 , wherein filtering the processed data at the filtering block includes filtering with a plurality of separate filter paths for different components of graphics data.
13. The method of claim 12 , wherein filtering the processed data includes:
filtering with an albedo filtering path;
filtering with a specular filtering path; and
filtering with a diffuse filtering path.
14. One or more non-transitory computer-readable storage mediums having stored thereon executable computer program instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
receiving ray tracing input data at a processing resource and upsampling the received input data from a first resolution to a second resolution, the processing resource including a pipeline to perform a supersampling operation, the pipeline including circuitry associated with operations of a single neural network;
processing the received ray tracing input data at the neural network to perform both denoising and supersampling of the input data, the processing including feature extraction and kernel prediction, the feature extraction including operation of a single shared feature extractor for the neural network; and
filtering the processed data at a filtering block.
15. The one or more non-transitory computer-readable storage mediums of claim 14 , wherein:
processing the received ray tracing input data at the neural network includes processing utilizing a plurality of convolution layers; and
the plurality of convolution layers operate on a first precision and the filtering block operates on a second precision, the first precision being lower than the second precision.
16. The one or more non-transitory computer-readable storage mediums of claim 14 , further comprising instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
warping a previous output from the pipeline using motion data to generate warped data; and
providing the warped data to the neural network for processing.
17. The one or more non-transitory computer-readable storage mediums of claim 14 , wherein filtering the processed data at the filtering block includes filtering with a plurality of separate filter paths for different components of graphics data.
18. The one or more non-transitory computer-readable storage mediums of claim 17 , wherein filtering the processed data includes:
filtering with an albedo filtering path;
filtering with a specular filtering path; and
filtering with a diffuse filtering path.