Multidimensional address generation for direct memory access
Techniques for data manipulation using processor cluster address generation are disclosed. One or more processor clusters capable of executing software-initiated work requests are accessed. A direct memory access (DMA) engine, coupled to the one or more processor clusters, is configured, wherein the DMA engine employs address generation across a plurality of tensor dimensions. A work request address field is parsed, where the address field contains unique address space descriptors for each of the plurality of dimensions, along with a common address space descriptor. DMA addresses are generated based on the unique address space descriptors and the common address space descriptor. Memory using two or more of the DMA addresses that were generated is accessed, where the two or more DMA addresses enable processing within the one or more processor clusters.
1. A processor-implemented method for data manipulation comprising:
accessing one or more processor clusters capable of executing software-initiated work requests;
configuring a direct memory access (DMA) engine coupled to the one or more processor clusters, wherein the DMA engine employs address generation across a plurality of tensor dimensions;
parsing a work request address field, wherein the address field contains unique address space descriptors for each of the plurality of dimensions, along with a common address space descriptor;
generating DMA addresses, based on the unique address space descriptors and the common address space descriptor; and
accessing memory, using two or more of the DMA addresses that were generated, wherein the two or more DMA addresses enable processing within the one or more processor clusters.
2. The method of claim 1 further comprising summing across the plurality of dimensions to generate a single address.
3. The method of claim 1 wherein the plurality of dimensions includes four dimensions.
4. The method of claim 3 wherein the plurality of dimensions does not include channels.
5. The method of claim 4 further comprising summing across channels as part of a convolutional operation.
6. The method of claim 1 wherein the generating comprises establishing five programming loops to accomplish five-dimensional (5-D) address generation.
7. The method of claim 6 wherein the 5-D address is a portion of a larger dimensional address.
8. The method of claim 6 wherein the innermost dimension is defined by hardware.
9. The method of claim 8 wherein the dimensions are ordered by dimension size in the tensor.
10. The method of claim 1 wherein the unique address space descriptors and the common address space descriptor comprise a 256-bit field.
11. The method of claim 1 wherein the common address space descriptor includes a base address field, a dimensions field, and an element size field.
12. The method of claim 11 wherein the common address space descriptor comprises a 69-bit field.
13. The method of claim 1 wherein a first address space descriptor of the unique address space descriptors includes a count field and a stride field.
14. The method of claim 13 wherein the first address space descriptor comprises a 35-bit field.
15. The method of claim 13 wherein second, third, fourth, and fifth address space descriptors of the unique address space descriptors each includes an offset field and a count field.
16. The method of claim 15 wherein the second, third, fourth, and fifth address space descriptors each comprise a 38-bit field.
17. The method of claim 1 wherein DMA memory source and destination addresses comprise a 5-dimensional descriptor.
18. The method of claim 17 wherein the 5-dimensional descriptor enables processing of a flattened tensor.
19. The method of claim 17 wherein the 5-dimensional descriptor includes tensor strides and offsets.
20. The method of claim 1 further comprising flattening a plurality of dimensions from a tensor into a single dimension.
21. The method of claim 20 wherein the DMA engine is further configured to employ address generation across the plurality of dimensions that were flattened into a single dimension.
22. A computer program product embodied in a non-transitory computer readable medium for data manipulation, the computer program product comprising code which causes one or more processors to perform operations of:
accessing one or more processor clusters capable of executing software-initiated work requests;
configuring a direct memory access (DMA) engine coupled to the one or more processor clusters, wherein the DMA engine employs address generation across a plurality of tensor dimensions;
parsing a work request address field, wherein the address field contains unique address space descriptors for each of the plurality of dimensions, along with a common address space descriptor;
generating DMA addresses, based on the unique address space descriptors and the common address space descriptor; and
accessing memory, using two or more of the DMA addresses that were generated, wherein the two or more DMA addresses enable processing within the one or more processor clusters.
23. A computer system for data manipulation comprising:
a memory which stores instructions;
one or more processors coupled to the memory wherein the one or more processors, when executing the instructions which are stored, are configured to:
access one or more processor clusters capable of executing software-initiated work requests;
configure a direct memory access (DMA) engine coupled to the one or more processor clusters, wherein the DMA engine employs address generation across a plurality of tensor dimensions;
parse a work request address field, wherein the address field contains unique address space descriptors for each of the plurality of dimensions, along with a common address space descriptor;
generate DMA addresses, based on the unique address space descriptors and the common address space descriptor; and
access memory, using two or more of the DMA addresses that were generated, wherein the two or more DMA addresses enable processing within the one or more processor clusters.