Lossless tiling in convolution networks—section boundaries
Disclosed is a data processing system that includes compile time logic to section a graph into a sequence of sections, configure a first section to generate a first set of output tiles in a first target tiling configuration in response to processing a first set of input tiles in a first input tiling configuration, and configure a second section to generate a second set of output tiles in a second target tiling configuration in response to processing the first set of output tiles in a second input tiling configuration. Runtime logic is configured to pad a first input into a first padded input, read the first set of input tiles from the first padded input in the first input tiling configuration, and process the first set of input tiles through the first section to generate the first set of output tiles in the first target tiling configuration.
1 . A computer implemented method comprising:
processing, in a first section of a processing graph, a first set of input tiles in a first input tiling configuration, to generate a first set of output tiles in a first target tiling configuration;
generating, from the first set of output tiles in the first target tiling configuration, a second set of input tiles in a second input tiling configuration;
reading, by a second section of the processing graph, the second set of input tiles in the second input tiling configuration that is different from the first target tiling configuration; and
processing, in the second section of the processing graph, the second set of input tiles in the second input tiling configuration, to generate a second set of output tiles in a second target tiling configuration;
wherein individual output tiles in the first set of output tiles have a first size; and
individual input tiles in the second set of input tiles have a second size that is larger than the first size.
2 . The method of claim 1 , wherein generating, from the first set of output tiles in the first target tiling configuration, the second set of input tiles in the second input tiling configuration comprises:
writing the first set of output tiles in a memory, with zero-padding applied to the first set of output tiles, to create a zero-padded first set of output tiles; and
retiling the zero-padded first set of output tiles in the memory, to generate the second set of input tiles in the second input tiling configuration.
3 . The method of claim 2 , wherein:
writing the first set of output tiles in the memory comprises writing the first set of output tiles in the memory in a non-overlapping configuration; and
retiling the zero-padded first set of output tiles comprises retiling the zero-padded first set of output tiles in an overlapping configuration.
4 . The method of claim 2 , wherein writing the first set of output tiles in the memory comprises:
initializing an area of the memory with zeros; and
writing individual tiles of the first set of output tiles in the area of the memory that was initialized with the zeros.
5 . The method of claim 4 , wherein the area of the memory that was initialized with the zeros is larger than a combined size of individual tiles of the first set of output tiles, such that a periphery of the area of the memory that was initialized with the zeros forms the zero-padding applied around the first set of output tiles in the memory.
6 . The method of claim 1 , further comprising:
compiling the processing graph by sectioning the processing graph into a sequence of sections, the sequence of sections including at least the first section and the second section.
7 . The method of claim 1 , wherein:
the first target tiling configuration configures neighboring tiles in the first set of output tiles to be non-overlapping with each other.
8 . The method of claim 1 , wherein:
the second input tiling configuration configures at least a first tile and a neighboring second tile in the second set of input tiles to be overlapping with each other.
9 . The method of claim 1 , wherein:
the first input tiling configuration configures at least a first tile and a second tile of the first set of input tiles to be overlapping with each other.
10 . The method of claim 1 , wherein:
the second target tiling configuration configures neighboring tiles in the second set of output tiles to be non-overlapping with each other.
11 . A data processing system, comprising:
storage medium storing runtime logic; and
one or more processors coupled to the storage medium, the one or more processors executing the runtime logic,
wherein the runtime logic executed by the one or more processors is configured to:
cause a first section of a processing graph to process a first set of input tiles in a first input tiling configuration, to generate a first set of output tiles in a first target tiling configuration,
cause to generate, from the first set of output tiles in the first target tiling configuration, a second set of input tiles in a second input tiling configuration, the second input tiling configuration different from the first target tiling configuration,
cause a second section of the processing graph to read the second set of input tiles in the second input tiling configuration, and
cause the second section of the processing graph to process the second set of input tiles in the second input tiling configuration, to generate a second set of output tiles in a second target tiling configuration;
wherein individual output tiles in the first set of output tiles have a first size; and
individual input tiles in the second set of input tiles have a second size that is larger than the first size.
12 . The data processing system of claim 11 , wherein to generate the second set of input tiles from the first set of output tiles, the runtime logic is configured to:
write the first set of output tiles in a memory, with zero-padding applied to the first set of output tiles, to create a zero-padded first set of output tiles; and
retile the zero-padded first set of output tiles in the memory, to generate the second set of input tiles in the second input tiling configuration.
13 . The data processing system of claim 12 , wherein:
to write the first set of output tiles in the memory, the runtime logic is configured to write the first set of output tiles in the memory in a non-overlapping configuration; and
to retile the zero-padded first set of output tiles, the runtime logic is configured to retile the zero-padded first set of output tiles in an overlapping configuration, such that a tile in the second set of input tiles overlaps with one or more neighboring tiles in the second set of input tiles.
14 . The data processing system of claim 11 , wherein:
the first target tiling configuration configures neighboring tiles in the first set of output tiles to be non-overlapping with each other; and
the second input tiling configuration configures at least a first tile and a neighboring second tile in the second set of input tiles to be overlapping with each other.
15 . The data processing system of claim 11 , wherein the storage medium stores compile time logic, wherein the one or more processors execute the compile time logic, and wherein the compile time logic executed by the one or more processors is configured to:
compile the processing graph by sectioning the processing graph into a sequence of sections, the sequence of sections including at least the first section and the second section.
16 . A non-transitory computer readable storage medium impressed with computer program instructions, the instructions, when executed on a processor, implement a method comprising:
generating, by a first section of a processing graph, a first set of output tiles of an output tensor in a first target tiling configuration;
zero-padding and retiling the first set of output tiles, to generate a set of input tiles of an input tensor in an input tiling configuration, the input tiling configuration different from the first target tiling configuration;
reading, by a second section of the processing graph, the set of input tiles in the input tiling configuration; and
processing, in the second section of the processing graph, the set of input tiles in the input tiling configuration, to generate a second set of output tiles in a second target tiling configuration.
17 . The non-transitory computer readable storage medium of claim 16 , wherein:
the first target tiling configuration configures neighboring tiles in the first set of output tiles to be non-overlapping with each other; and
the input tiling configuration configures at least a first tile and a neighboring second tile in the set of input tiles to be overlapping with each other.
18 . The non-transitory computer readable storage medium of claim 16 , wherein:
individual tiles in the first set of output tiles have a first size; and
individual tiles in the set of input tiles have a second size that is larger than the first size.