International Business Machines Corporation

Puneeth A H Bhat

Satish Kumar Sadasivam

Shruti Saxena

Sangram Alapati

Biplob Mishra

Venkatesh Kr

A branch prediction system includes a neuron cache and logic coupled to the neuron cache. The neuron cache includes one or more weights of a neural network model trained for one or more selected code sections, and the logic is to be used with the neuron cache to predict a target address for a branch instruction of the one or more selected code sections.

Neuron cache-based hardware branch prediction

Aspects of the invention include a compiler detecting an expression in a loop that includes elements of mixed data types. The compiler then promotes elements of a sub-expression of the expression to a same intermediate data type. The compiler then calculates the sub-expression using the elements of the same intermediate data type.

Partial data type promotion to exploit efficient vectorization in microprocessors

A processor including a processor pipeline having one or more execution units configured to execute branch instructions, a branch predictor coupled to the processor pipeline and configured to predict a branch instruction outcome, and a branch classification unit coupled to the processor pipeline and the branch prediction unit. The branch classification unit is configured to, in response to detecting a branch instruction, classify the branch instruction as at least one of the following: static taken branch, static not-taken branch, simple easy-to-predict branch, flip flop hard-to-predict (HTP) branch, dynamic HTP branch, biased positive HTP branch, biased negative HTP branch, and other HTP branch.

System and method for multi-level classification of branches

Three new software instructions assist a processor in performing indirect prefetching, and managing a next-to-prefetch address list. The software instructions populate hardware register locations according to a hardware register description comprising a data structure of at least seven fields. Multiple instances of the data structure, shared across multiple respectively corresponding threads running concurrently, comprise an indirect-prefetch-tracker table. The indirect-prefetch-tracker table assists the processor to efficiently perform indirect prefetching, from random (not necessarily contiguous) memory locations, and reduces processor core real estate dedicated to control and management of data prefetch and loading operations.

Software assisted data address prefetch

A processor comprising a processor pipeline comprising one or more execution units configured to execute branch instructions, a branch predictor associated with the processor pipeline and configured to predict a branch instruction outcome, a branch classification unit associated with the processor pipeline and the branch prediction unit. The branch classification unit is configured to, in response to detecting a branch instruction, classify the branch instruction as at least one of the following: a simple branch or a hard-to-predict (HTP) branch, wherein a threshold used for the classification is dynamically adjusted based on a workload of the processor.

System and method for dynamic accuracy and threshold control for branch classification

A processor including a processor pipeline having one or more execution units configured to execute branch instructions, a branch predictor coupled to the processor pipeline and configured to predict a branch instruction outcome, and a branch classification unit coupled to the processor pipeline and the branch predictor. The branch classification unit is configured to, in response to detecting a branch instruction, classify the branch instruction as at least one of the following: a simple branch and a hard-to-predict branch. The branch classification unit includes a direct mapped branch type table (BTT) and a branch classification table (BCT).

Predicting a branch instruction classified as simple or hard to predict based on a confidence counter in a branch type table

Microprocessor with multiple issue queues in a microprocessor, where at least one of the issue queues is an adjustable mode queue that can be set to act as either of a priority queue, or a regular queue, with respect to intake of new instructions and/or outflow of old instructions. A set of summary bit value(s) can be set to control whether the adjustable mode queue has instruction intake priority and/or instruction outflow priority relative to the other issue queue(s).

Priority instruction handling with optimized issue queue design

Provided is a method for predicting a target address using a set of Indirect Target TAgged GEometric (ITTAGE) tables and a target address pattern table. A branch instruction that is to be executed may be identified. A first tag for the branch instruction may be determined. The first tag may be a unique identifier that corresponds to the branch instruction. Using the tag, the branch instruction may be determined to be in a target address pattern table, and an index may be generated. A predicted target address for the branch instruction may be determined using the generated index and the largest ITTAGE table. Instructions associated with the predicted target address may be fetched.

Indirect target tagged geometric branch prediction using a set of target address pattern data

Aspects of the present disclosure relate to control of speculative demand loads. In some embodiments, the method includes receiving instructions for a branch in a program, detecting the branch load is in the cache, monitoring a number of completed loads for the program, determining a cache pollution ratio of executed loads to completed loads, providing the cache pollution ratio to a branch prediction unit, and altering load instructions for the branch based on the cache pollution ratio.

Control of speculative demand loads

Disclosed aspects relate to branch predictor selection management in a pipelined microprocessor architecture. A set of selection factor data may be collected in the pipelined microprocessor architecture. The set of selection factor data may be analyzed using a perceptron-based learning technique with respect to a set of candidate branch predictors. A chosen branch predictor may be selected from the set of candidate branch predictors based on analyzing the set of selection factor data with respect to the set of candidate branch predictors using the perceptron-based learning technique. The chosen branch predictor may be invoked in the pipelined microprocessor architecture.

Growing community of inventors

Bengaluru, India

Puneeth A H Bhat