17527545. SHAPE AND DATA FORMAT CONVERSION FOR ACCELERATORS simplified abstract (International Business Machines Corporation)

From WikiPatents
Revision as of 01:42, 4 January 2024 by Wikipatents (talk | contribs) (Creating a new page)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

SHAPE AND DATA FORMAT CONVERSION FOR ACCELERATORS

Organization Name

International Business Machines Corporation

Inventor(s)

YASUSHI Negishi of Machida-shi (JP)

Tung D. Le of Ichikawa (JP)

HARUKI Imai of Yokohama-shi (JP)

KIYOKUNI Kawachiya of Yokohama (JP)

SHAPE AND DATA FORMAT CONVERSION FOR ACCELERATORS - A simplified explanation of the abstract

This abstract first appeared for US patent application 17527545 titled 'SHAPE AND DATA FORMAT CONVERSION FOR ACCELERATORS

Simplified Explanation

The patent application describes a method for converting tensor data to meet the specific data format of a hardware accelerator. The method takes input tensors with a certain shape and format and modifies them to match the required format. Here are the key points:

  • Input tensors L and L are received, each with a data format of <X x Y x Z> and an n-dimension input tensor shape of <X x X x X x ... x X>.
  • The input tensor shape is stored for reference.
  • The method calculates a modified shape for the input tensors by finding the largest divisor of (X x X x ... x X) that is less than or equal to L, denoted as S.
  • It then finds the largest divisor of ((X x X x ... x X) / S) that is less than or equal to L, also denoted as S.
  • Next, it sets (((X x X x ... x X) / (S x S)) to S.
  • Finally, the method returns the modified shape as .

Potential applications of this technology:

  • This method can be used in hardware accelerators that require a specific data format for tensor data.
  • It can be applied in various fields where tensor data processing is involved, such as machine learning, computer vision, and natural language processing.

Problems solved by this technology:

  • The method solves the problem of converting tensor data to meet the specific format required by a hardware accelerator.
  • It provides a systematic approach to modify the shape of input tensors to fit the hardware accelerator's data format.

Benefits of this technology:

  • The method allows for efficient utilization of hardware accelerators by converting tensor data to the required format.
  • It simplifies the process of adapting tensor data to different hardware accelerators, saving time and effort.
  • By optimizing the shape of input tensors, it can potentially improve the performance and speed of tensor data processing.


Original Abstract Submitted

A method for converting a shape and a format of tensor data to meet a specific data format of a hardware accelerator is provided. The method receives input tensors L and L, each being constants having a data format of < X x Y x Z >, and each further having an n-dimension input tensor shape as <X x X x X x ... x X >. The method stores input tensor shape. The method calculates an n-dimension modified shape of the input tensors by (a) setting a largest divisor of (X x X x...x X ) ≤ L to S, (b) setting a largest divisor of ((X x X x...x X ) / S) ≤ L to S, (c) setting (((X x X x... x X ) / (S x S)) to S, and (d) returning the n-dimension modified shape as < S x S x S >.

Retrieved from "http://wikipatents.org/index.php?title=17527545._SHAPE_AND_DATA_FORMAT_CONVERSION_FOR_ACCELERATORS_simplified_abstract_(International_Business_Machines_Corporation)&oldid=23528"