RC1	NEON programming

This course explains how to use NEON SIMD instructions to boost multimedia algorithms

Objectives This course has been designed for programmers wanting to run multimedia algorithms on NEON Single Instruction Multiple Data execute units. Each instruction family is detailed, first at assembly level, and then at C level using macros developed present in arm_neon.h file. Several tricky usage of processing instructions are provided. Vector and vector element load / store instructions are studied and guidelines for organizing data in memory are provided to minimize the number of memory accesses. The underlying cache operation as well as preload mechanisms (instruction and hardware prefetch) are detailed to explain how a processing can be pipelined . The course shows how DSP typical algorithms such as FIR and FFT can be vectorized and then optimized to be executed on NEON unit. This course has already been delivered to companies developing audio applications for mobile phones. THIS COURSE IS PROPOSED EITHER AS AN INSTRUCTOR-LED COURSE OR AS E-LEARNING. ACSYS has developed an optimized NEON based FFT coded in assembler language performance for 1024 complex floating point single precision samples is: - 106_000 core clock cycles for Cortex-A9 - 90_000 core clock cycles for Cortex-A8 for any information contact guillaume.peron@ac6.fr Labs are run under RVDS A more detailed course description is available on request at info@ac6-training.com

Prerequisites
	Knowledge of ARM V4T / V5TE instruction set.

Outline

DAY 1 CORTEX-A8 AND CORTEX-A9(MP) ARCHITECTURE Data path, studying how data are loaded from external memory and copied into level 1 and possibly level 2 caches Programmer’s model Highlighting coherency issues when data are shared by several cores, purpose of the SCU implemented in Cortex-A9 Cortex-A8 and Cortex-A9 instruction pipeline, branch predictors INTRODUCTION TO NEON/VFPv3 Clarifying the resources shared by NEON and VFP Register bank, Q registers, D registers Data types Vector vs scalar Related system registers Alignment issues Enabling NEON/VFP NEON INSTRUCTION SYNTAX Instructions producing wider / narrower results Instructions modifiers Selecting the shape Selecting the operand / result type Syntax flexibility Declaring initialized vectors in C language Using unions with vectors and arrays of vectors to simplify the debug Casting vectors LOAD / STORE INSTRUCTIONS Addressing modes Vector load / store Vector load / store multiple Element and structure load / store instructions Multiple single elements Single element to 1 lane Single elements to all lanes Optimizing the ordering of data in memory to take benefit of 2-, 3- and 4- element structures Example: managing audio samples Processor acceleration mechanisms: store merging buffers Practical lab: using load with de-interleaving instructions to store all right lane samples into a vector and left lane samples into another vector DAY 2 DATA TRANSFER INSTRUCTIONS Move Swap Table lookup Vector transpose Vector zip / unzip Data transfer between NEON and integer unit Practical lab: clarifying narrow and long instructions, building a vector from bytes selected from a pair of vectors

LOGICAL AND BITFIELD INSTRUCTIONS Logical AND, Bit Clear, OR, XOR Operations with immediate values Bitwise insert instructions, avoiding branches Count Leading zeros, ones, signs Normalizing floating point numbers when VFP is not implemented Scalar duplicate Extract Shift with possible rounding and saturation Bitfield revers Practical lab: Transposing a matrix, shifting a large bitmap using vector instructions ARITHMERICAL INSTRUCTIONS Add, modulo vs saturated arithmetic Halving / Doubling the result Rounding Subtract Multiply Multiply accumulate / Multiply subtract Absolute value Min / Max Converting Floating Point numbers into Fixed point numbers Converting Fixed point numbers into Floating point numbers Reciprocal estimate, reciprocal square root estimate, Newton-raphson algorithm Pairwise instructions Element comparison Practical lab: implementing a complex multiply accumulate with NEON Practical lab: converting fixed-point elements into single precision floating point values and adding the resulting elements NEON CODING EXAMPLES FIR filter Converting the scalar algorithm into a vector algorithm Finding the NEON instructions to encode the vector algorithm Optimizing the code Using the performance monitor to tune the algorithm FFT (DFT) Converting the scalar algorithm into a vector algorithm, understanding how circle properties can be used to process 4 angles concurrently Finding the NEON instructions to encode the vector algorithm Optimizing the code Using the performance monitor to tune the algorithm

 

Duration	Cost	Calendar
2 days	1250 €