Efficient data layout and alignment for wide-vector accelerator systems
View Patent ↗Efficient data layout and alignment techniques for effectively executing AI workloads in wide-vector accelerator systems are provided. In one aspect, a method for processing AI workloads includes: logically dividing a data vector into a hierarchy of segments and sub-segments with each of the segments including more than one of the sub-segments, wherein each of the sub-segments includes words, and each of the words includes data-bits; and physically mapping the data-bits such that the words belonging to a same given one of the sub-segments are mapped contiguously across all of the segments. An AI accelerator system is also provided.
1. A method for processing artificial intelligence (AI) workloads, the method comprising:
logically dividing a data vector into a hierarchy of segments and sub-segments with each of the segments comprising more than one of the sub-segments, wherein the data vector is a row from memory, each of the sub-segments comprises words, and each of the words comprises data-bits;
physically mapping the data-bits such that the words belonging to a same given one of the sub-segments are mapped contiguously across all of the segments;
pulling the row from the memory; and
performing alignment operations on the segments, the sub-segments, or a combination thereof to create an aligned data-vector.
2. The method of claim 1 , wherein the data-bits are physically mapped such that a distance between the words of a particular one of the sub-segments is minimized across all of the segments.
3. The method of claim 1 , wherein the data-bits are physically mapped such that a distance between instances of a same given one of the words in each of the sub-segments is minimized across all of the segments.
4. The method of claim 1 , wherein the alignment operations comprise one or more of rotating the segments either clockwise or counter-clockwise by at least one of the segments, and shifting the sub-segments either left or right by at least one of the sub-segments.
5. The method of claim 4 , wherein the performing of the alignment operations comprises rotating the segments by at least one of the segments, and wherein at least one of the segments is rotated from one end of the row to an opposite end of the row.
6. The method of claim 1 , further comprising:
copying a portion of the aligned data-vector into a final output-vector, while selectively masking a remainder of the aligned data-vector.
7. The method of claim 6 , further comprising:
repeating the logically dividing, the mapping, the copying, the pulling and the performing iteratively over multiple other rows in the memory.
8. The method of claim 6 , further comprising:
transmitting the final output-vector to a compute engine.
9. The method of claim 8 , wherein the compute engine comprises an analog AI crossbar array, and wherein the method further comprises:
rearranging weights across rows or columns of the analog AI crossbar array using valid-zero interleaving mapping.
10. An artificial intelligence (AI) accelerator system, comprising:
a memory;
a compute engine connected to the memory; and
a hardware logic unit configured to:
pull a row from the memory, wherein the row is logically divided into a hierarchy of segments and sub-segments with each of the segments comprising more than one of the sub-segments, each of the sub-segments comprising words, and each of the words comprising data-bits, and wherein the data-bits are physically mapped such that the words belonging to a same given one of the sub-segments are mapped contiguously across all of the segments;
perform alignment operations on the segments, sub-segments or combinations thereof to create an aligned data vector;
copy a portion of the aligned data-vector into a final output-vector, while selectively masking a remainder of the aligned data-vector; and
transmit the final output-vector to the compute engine.
11. The AI accelerator system of claim 10 , wherein the compute engine comprises an analog AI crossbar array.
12. The AI accelerator system of claim 10 , wherein the hardware logic unit when performing the alignment operations is configured to perform one or more of: rotating the segments either clockwise or counter-clockwise by at least one of the segments, and shifting the sub-segments either left or right by at least one of the sub-segments.
13. The AI accelerator system of claim 12 , wherein the hardware logic unit comprises:
at least one set of first multiplexers configured to perform the rotating of the segments by at least one of the segments; and
a set of second multiplexers configured to perform the shifting of the sub-segments by at least one of the sub-segments.
14. The AI accelerator system of claim 10 , wherein the hardware logic unit comprises:
segment mask-registers configured to mask bits in the aligned data vector from the alignment operations performed on the segments; and
sub-segment mask-registers configured to mask bits in the aligned data vector from the alignment operations performed on the sub-segments.
15. A non-transitory computer program product for processing artificial intelligence (AI) workloads, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to:
logically divide a data vector into a hierarchy of segments and sub-segments with each of the segments comprising more than one of the sub-segments, wherein the data vector is a row from memory each of the sub-segments comprises words, and each of the words comprises data-bits;
physically map the data-bits such that the words belonging to a same given one of the sub-segments are mapped contiguously across all of the segments;
pulling the row from the memory; and
performing alignment operations on the segments, the sub-segments, or a combination thereof to create an aligned data-vector.
16. The non-transitory computer program product of claim 15 , wherein the data-bits are physically mapped such that a distance between the words of a particular one of the sub-segments is minimized across all of the segments, and wherein the data-bits are physically mapped such that a distance between instances of a same given one of the words in each of the sub-segments is minimized across all of the segments.
17. The non-transitory computer program product of claim 15 , wherein the alignment operations comprise one or more of rotating the segments either clockwise or counter-clockwise by at least one of the segments, and shifting the sub-segments either left or right by at least one of the sub-segments.
18. The non-transitory computer program product of claim 15 , wherein the program instructions further cause the computer to:
copy a portion of the aligned data-vector into a final output-vector, while selectively masking a remainder of the aligned data-vector; and
transmit the final output-vector to a compute engine.
19. The method of claim 1 , wherein the alignment operations are not performed on data at a granularity finer than the sub-segments.
20. The method of claim 8 , wherein the compute engine comprises an analog AI crossbar array that includes an arrangement of uniquely addressable resistive processing units that embody a neural network.