International Business Machines Corporation

Motohiro Kawahito

Hideaki Komatsu

Toshiaki Yasue

Toshihiko Koju

Xin Tong

Takeshi Ogasawara

Akira Koseki

Takuya Nakaike

Takao Moriyama

Ravi Nair

Hong Min

Petr Novotny

Vijay Sundaresan

Shyam Ramji

Qi Zhang

Ali Ijaz Sheikh

Lei Yu

Hiroshi Inoue

Hiroyuki Momose

Kideaki Komatsu

Toshio Nakatani

A computer-implemented method, system and computer program product for effectively searching for values in a multi-byte array of elements using an n-byte search instruction. Multiple values to be searched in an N-byte array of elements in a loop are received. The loop is optimized by searching the received search values at the starting address of the N-byte array of elements using the n-byte search instruction. A successful search is performed if the received return address points to an address found in the lowest n-bytes of the N-byte array of elements and an element of the address corresponds to a search value. Otherwise, a subsequent search for the search values at the address of the next element in the N-byte array of elements is performed if there are additional elements in the N-byte array of elements to be searched.

Searching an array of multi-byte elements using an n-byte search instruction

A method creates a table of keys and values. Each key is an element of an input array which is an input of a machine-learning pre-processing pipeline, and each value is an output of the pipeline. The method measures (1) a hit rate H to the memo table, (2) an average time Tto look up the table, (3) an average time Tto execute the pipeline, and (4) a threshold Ton a number of elements of the input array. The method looks up the value in the table by using an element of the input array as a key when T×H>Tand the number of elements in the input array is less than T. The method calls the pipeline in place of the lookup for all of the remaining elements in the input array when the value is not in the table.

Memoizing machine-learning pre-processing and feature engineering

Techniques for refinement of data pipelines are provided. An original file of serialized objects is received, and an original pipeline comprising a plurality of transformations is identified based on the original file. A first computing cost is determined for a first transformation of the plurality of transformations. The first transformation is modified using a predefined optimization, and a second cost of the modified first transformation is determined. Upon determining that the second cost is lower than the first cost, the first transformation is replaced, in the original pipeline, with the optimized first transformation.

Streamlining data processing optimizations for machine learning workloads

A method is provided for validating a compiler-generated program portion that forms an optimized runnable code relative to an input runnable code. The method computes respective data sets used by the compiler-generated program portion. The respective data sets include (i) memory inputs UM, (ii) constant data memory areas UM_CONST, (iii) output memory areas DM, and (iv) output registers DR. The method copies the compiler-generated program portion to another memory area from a current memory area and appends a return instruction back to the current memory area at each exit point of the compiler-generated program portion. The method computes minimum and maximum base register offsets for base registers, from a union formed from a subset of the respective data sets. The method computes an allocation size for each of the base registers and an address assigned to each of the base registers. The method executes the copied compiler-generated program portion using (a) the minimum and maximum base register offsets, (b) the allocation size, and (c) the address assigned to each of the base registers, executes the input runnable code, examines results of the input runnable code in the output memory areas DM and of the compiler-generated program portion in the output registers DR, and indicates that the compiler-generated program portion is valid responsive to the results being identical.

Executing a part of a compiler generated code as a function

A computer provides a first cache area and a second cache area on a cache memory in a data processing system comprising a database repository and the cache memory for the database repository. The computer receives, in a first time window, streaming data and writing the streaming data on the first cache area. The computer prohibits, in the first time window, carrying out synchronization between the first cache area and the database repository. The computer enables, in at least part of the first time window, synchronization between the database repository and the second cache area. The computer enables, upon elapse of the first time window, synchronization between the first cache area and the database repository.

Multiple cache processing of streaming data

Aspects of the present invention provide devices that generate an array in a first memory of a first programming language that includes an element of each object of a plurality of objects, wherein the array includes a plurality of items, and each item of the array includes the element of a corresponding object, copy the array in the first memory of the first programming language to a second memory of a second programming language, and generate each of the objects in the second programming language from the copied array.

Copying objects between programming languages

An apparatus for controlling an execution of a binary code by multiple threads includes a detection unit configured to detect an occurrence of modification of a first part that is a part of a first binary code by a self-modifying code; a specifying unit configured to specify a second part that is a part corresponding to the first part in a second binary code acquired by converting the first binary code, in response to detection of the occurrence of modification of the first part by the self-modifying code; and a correction unit configured to correct the second part such that a specific thread that executes the second part of the second binary code among the multiple threads causes an exception.

Controlling execution of binary code

Method, program and system for code optimization. A sign assignment instruction with input and output operands identical in size, which assigns a value of zero to a packed decimal data value that has a value of negative zero, is detected in source code or an executable program. A first instruction that checks for a possibility that the input operand is negative zero is generated based on that detection. The first instruction skips the sign assignment instruction when the input and output operand have identical addresses when and there is no possibility that the input value is negative zero; and executes a copy instruction copying the input value in the input operand to the output operand when the input and output operands have different addresses and are not overlapping. The first instruction is inserted into one of the source code or the binary executable program that is output.

Method, program, and system for code optimization

A computer-implemented method for controlling data transfer for data processing includes: receiving a data set by a first processor; storing the data set in a buffer by the first processor; and transferring, from the buffer to a queue by the first processor, a batch data set including all data sets stored in the buffer during a time interval, under a condition that the queue is not accessed by a second processor while the transferring is being performed, wherein the batch data set is to be processed by the second processor. In the method, the time interval is adjusted depending on a state of the queue.

Controlling data transfer for data processing

Method, program and system for code optimization. A sign assignment instruction with identically sized packed decimal format input and output operands is detected where the sign assignment instruction assigns a value of zero to a packed decimal data value input operand having a value of negative zero. If the input operand to the sign assignment instruction does not result from an add or subtract operation, or the value of the input operand is not greater than a value prior to that operation, the possibility that the value of the input operand of the sign assignment instruction is negative zero is checked when the input operand and the output operand have identical addresses. An instruction is generated and inserted for executing the sign assignment instruction only when there is the possibility that the operand value is negative zero.

A method for optimizing if statements in a program includes obtaining, by a processing device, for each of conditional expressions of a plurality of if statements in the program, a set of conditional expressions having an inclusion relation; computing, for each of the set, a position with low execution frequency in the program as a move destination of a conditional expression having an inclusion relation, using information of the set in which the conditional expression is included; and moving the conditional expression to the computed move destination of the conditional expression.

Growing community of inventors

Sagamihara, Japan

Motohiro Kawahito