RC2 | NEON-v8 programming |
This course explains how to use ARMv8 NEON SIMD instructions to boost multimedia algorithms
Objectives
|
- Knowledge of ARMv7 instruction sets.
- Theoretical course
- PDF course material (in English) supplemented by a printed version.
- The trainer answers trainees' questions during the training and provide technical and pedagogical assistance.
- Practical activities
- Practical activities represent from 40% to 50% of course duration.
- Code examples, exercises and solutions
- One PC (Linux ou Windows) for the practical activities with, if appropriate, a target board.
- One PC for two trainees when there are more than 6 trainees.
- For onsite trainings:
- An installation and test manual is provided to allow preinstallation of the needed software.
- The trainer come with target boards if needed during the practical activities (and bring them back at the end of the course).
- Downloadable preconfigured virtual machine for post-course practical activities
- At the start of each session the trainer will interact with the trainees to ensure the course fits their expectations and correct if needed
- Any embedded systems engineer or technician with the above prerequisites.
- The prerequisites indicated above are assessed before the training by the technical supervision of the traineein his company, or by the trainee himself in the exceptional case of an individual trainee.
- Trainee progress is assessed in two different ways, depending on the course:
- For courses lending themselves to practical exercises, the results of the exercises are checked by the trainer while, if necessary, helping trainees to carry them out by providing additional details.
- Quizzes are offered at the end of sections that do not include practical exercises to verifythat the trainees have assimilated the points presented
- At the end of the training, each trainee receives a certificate attesting that they have successfully completed the course.
- In the event of a problem, discovered during the course, due to a lack of prerequisites by the trainee a different or additional training is offered to them, generally to reinforce their prerequisites,in agreement with their company manager if applicable.
Course Outline
- Clarifying the resources shared by NEON and the scalar floating point engine
- Explaining the AArch32 and AArch64 differences
- NEON Register banks
- S, D and Q registers (AArch32)
- B, H, S, D and V registers (AArch64)
- Data types
- Vector vs scalar
- Related system registers
- Alignment issues
- Enabling NEON
- Differences between NEONv7 and NEONv8
- 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
- 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
Exercise: | Example: managing audio samples | |
Exercise: | Using load with de-interleaving instructions to store all right lane samples into a vector and left lane samples into another vector | |
Exercise: | Clarifying narrow and long instructions, building a vector from bytes selected from a pair of vectors |
- Arithmetic instructions
- Add, modulo vs saturated arithmetic
- Halving / Doubling the result
- Rounding
- Subtract
- Multiply
- Multiply accumulate / Multiply subtract
- Absolute value
- Min / Max
Exercise: | Implementing a complex multiply accumulate with NEON |
- Conversion instructions
- Converting Floating Point numbers into Fixed point numbers
- Converting Fixed point numbers into Floating point numbers
Exercise: | Converting fixed-point elements into single precision floating point values and adding the resulting elements |
- Advanced arithmetic instructions
- Reciprocal estimate, reciprocal square root estimate, Newton-raphson algorithm
- Pairwise instructions
- Element comparison
- Logic instructions
- Logical AND, Bit Clear, OR, XOR
- Operations with immediate values
- Bitfield instructions
- Count Leading zeros, ones, signs
- Bitwise insert instructions
- Conditional bitwise insert instructions, avoiding branches
- Shifts with possible rounding and saturation
- Bitfield reverse
Exercise: | Transposing a matrix, shifting a large bitmap using vector instructions |
- The Cryptography extension
- Algorithms supported
- AES
- SHA1
- SHA256
- Automatic vectorization
- Tuning loops for optimal results
- Avoid loop feedbacks
- Avoid loop-dependent conditionals
- Avoid early termination
- Padding loops
Exercise: | Experimenting with loop auto-vecorization |
- Pointers and arrays
- indirect addressing
- pointer aliasing and restrict
Exercise: | Using restrict to eliminate dependencies |
- Function calls and inlining
- promises
Exercise: | Making promises to help the compiler optimize |
- Avoiding data dependencies
- 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