Temporally amortized supersampling using a kernel splatting network

A method of generating pixel data via at least one neural network,

• the method comprising:
• receiving a set of input data at a matrix accelerator of a graphics processor, the set of input data including first pixel data associated with a first rendered frame and motion vector data to map pixels from the first rendered frame to a second rendered frame, wherein the matrix accelerator is included in a streaming multiprocessor of a plurality of streaming multiprocessors of the graphics processor;
• executing first matrix multiply operations via the matrix accelerator, the first matrix multiply operations having input including the set of input data, wherein the first matrix multiply operations are associated with the at least one neural network and are executed to upsample the first pixel data from a first resolution to a second resolution;
• and
• executing second matrix multiply operations having input including upsampled first pixel data associated with the first rendered frame and second pixel data associated with the second rendered frame, wherein the second matrix multiply operations are associated with the at least one neural network and are executed to generate third pixel data based on the upsampled first pixel data and the second pixel data, wherein the second pixel data includes anti-aliased pixel data.

A non-transitory machine-readable medium having instructions stored thereon, the instructions, when executed cause one or more processors to perform operations to generate pixel data via at least one neural network, wherein the one or more processors include a graphics processor and the operations comprise:

• receiving a set of input data at a matrix accelerator of the graphics processor, the set of input data including first pixel data associated with a first rendered frame and motion vector data to map pixels from the first rendered frame to a second rendered frame, wherein the matrix accelerator is included in a streaming multiprocessor of a plurality of streaming multiprocessors of the graphics processor;
• executing first matrix multiply operations via the matrix accelerator, the first matrix multiply operations having input including the set of input data, wherein the first matrix multiply operations are associated with the at least one neural network and are executed to upsample the first pixel data from a first resolution to a second resolution;
• and
• executing second matrix multiply operations having input including upsampled first pixel data associated with the first rendered frame and second pixel data associated with the second rendered frame, wherein the second matrix multiply operations are associated with the at least one neural network and are executed to generate third pixel data based on the upsampled first pixel data and the second pixel data, wherein the second pixel data includes anti-aliased pixel data.

A graphics processing system comprising:

• a memory device;
• and
• a graphics processor comprising a level-two (L2) cache and a processing cluster array including a plurality of processing clusters coupled with the L2 cache, each processing cluster of the plurality of processing clusters including a plurality of streaming multiprocessors, wherein a streaming multiprocessor of the plurality of streaming multiprocessors includes multiple groups of execution resources, each group of execution resources includes a matrix accelerator configured to perform mixed precision matrix multiply and accumulate operations, and the matrix accelerator is configurable to:
• receive a set of input data, the set of input data including first color data associated with a first rendered frame and motion vector data to map pixels from the first rendered frame to a second rendered frame;
• execute first matrix multiply operations having input including the set of input data, the first matrix multiply operations associated with at least one neural network, to upsample the first color data from a first resolution to a second resolution;
• and
• execute second matrix multiply operations having input including upsampled first color data associated with the first rendered frame and second color data associated with the second rendered frame, the second matrix multiply operations associated with the at least one neural network, to generate third color data based on the upsampled first color data and the second color data, wherein the second color data includes anti-aliased pixel data.