Compression of floating-point data by identifying a previous loss of precision

Patent metadata

Field	Value
Publication number	US-9935650-B2
Application number	US-201514638728-A
Country	US
Kind code	B2
Filing date	Mar 4, 2015
Priority date	Apr 7, 2014
Publication date	Apr 3, 2018
Grant date	Apr 3, 2018

Abstract

Each binary floating-point value in a set of binary floating-point values is converted to a decimal floating-point value. Data are determined including an exponent, a mantissa and a quantity of decimal digits of the mantissa for each decimal floating-point value. The exponents, the mantissas and the quantity of decimal digits are individually compressed to produce compressed floating-point values based on the individual compressions.

First claim

What is claimed is:

1. A computer-implemented method of compressing floating-point data comprising:
converting each binary floating-point value in a set of binary floating-point values to a decimal floating-point value;
determining data for each decimal floating-point value including an exponent, a mantissa, and a quantity of decimal digits of the mantissa; and
individually compressing the exponents, the mantissas and the quantity of decimal digits of the decimal floating-point values and producing compressed floating-point values based on the individual compressions.

2. The computer-implemented method of claim 1 , wherein the converting includes generating a mantissa with a minimum number of digits.

3. The computer-implemented method of claim 2 , wherein the determining data includes determining whether one or more of the exponent and the quantity of decimal digits are constant for the decimal floating-point values.

4. The computer-implemented method of claim 1 , wherein the individually compressing includes:
analyzing the determined data to identify reduced precision within the decimal floating-point values and producing the compressed floating-point values based on the reduced precision.

5. The computer-implemented method of claim 1 , wherein the converting includes at least one from the group of:
normalizing a position of a decimal point; and
normalizing a length of a mantissa.

6. The computer-implemented method of claim 1 , wherein the individually compressing includes:
compressing the exponents, the mantissas and the quantity of decimal digits of the decimal floating-point values using a plurality of compression formats.

7. The computer-implemented method of claim 6 , wherein the individually compressing includes:
selecting a compressed exponent, a compressed mantissa and a compressed quantity of decimal digits with a shortest bit length from results of the compressing using the plurality of compression formats.

8. The computer-implemented method of claim 1 , wherein each converted binary floating-point value comprises a corresponding set of the exponent, the mantissa and the quantity of decimal digits to form a data triplet for the individually compressing the exponents, the mantissas and the quantity of decimal digits, the computer-implemented method further comprising:
processing each data triplet among plural parallel processing paths, wherein each processing path includes one or more from the group of:
compressing the values in each data triplet;
normalizing the exponent in each data triplet and compressing the values in each data triplet;
normalizing the mantissa length in each data triplet and compressing the values in each data triplet; and
normalizing the exponent and the mantissa length in each data triplet and compressing the values in each data triplet; and

selecting a compressed data triplet comprising a shortest bit length from among the compressed data triplets processed among the plural parallel processing paths.

Assignees

• IBM

Inventors

• Dickie Garth A

Classifications

• H03M7/24 Primary

  Conversion to or from floating-point codes · CPC title

• G06F7/483

  Computations with numbers represented by a non-linear combination of denominational numbers, e.g. rational numbers, logarithmic number system or floating-point numbers {(G06F7/4806, G06F7/4824, G06F7/49, G06F7/491, G06F7/544 take precedence)} · CPC title

• H03M7/00 Primary

  Conversion of a code where information is represented by a given sequence or number of digits to a code where the same {, similar or subset of} information is represented by a different sequence or number of digits · CPC title