The key difference between NEON and SIMD instructions in Cortex-M7 is that NEON is a single instruction multiple data (SIMD) engine specialized for media processing while SIMD instructions are more general purpose parallel processing instructions. NEON provides acceleration for digital signal processing, image processing, and machine learning workloads in Cortex-A series CPUs. SIMD instructions in Cortex-M CPUs like M7 enable parallel processing of simple arithmetic and logic operations on multiple data values.
Overview of NEON
NEON is ARM’s advanced SIMD architecture extension for the Cortex-A series processors. It provides acceleration for workloads like:
- Digital signal processing (DSP)
- 2D/3D graphics
- Image processing
- Video encoding/decoding
- Speech recognition
- Computer vision
- Machine learning
NEON implements the SIMD concept by providing instructions that can perform the same operation on multiple data values concurrently. This allows parallel processing of data using a single instruction, which improves performance for suitable workloads.
Key features of NEON include:
- 128-bit wide SIMD registers – Allow parallel operations on multiple data values
- SIMD instructions – Allow same operation to be performed on multiple data values
- Saturated arithmetic – Prevent overflow/underflow for audio/image processing
- Floating point support – Accelerate math-intensive algorithms
- Advanced memory access – Improve data transfer performance
NEON provides instructions for various data types including integer, single precision float, double precision float, and polynomials. This flexibility allows tuning for optimal performance across different workloads.
Overview of SIMD in Cortex-M7
While NEON is designed for high performance media processing, the Cortex-M series focuses more on real-time embedded applications. Still, Cortex-M CPUs like M7 provide SIMD capabilities through general purpose instructions.
Key SIMD features in Cortex-M7 include:
- Most arithmetic and logical instructions work on operands twice the register width
- 32-bit registers allow 64-bit SIMD operation
- Dual 16-bit instructions allow 32-bit SIMD operation
- Saturation support to avoid overflow
- Packing and unpacking between SIMD and scalar registers
This allows simple parallel operations on data sets using the CPU’s existing registers and ALUs. While less flexible than NEON, SIMD support in M7 can still provide good speedups for suitable workloads with regular data parallelism.
Key Differences
The key differences between NEON and SIMD in Cortex-M7 are:
- Target workloads – NEON for media processing, SIMD for general purpose
- Vector size – NEON 128-bit, SIMD 64-bit
- Registers – NEON has 16x 128-bit registers, SIMD uses core 32-bit registers
- Instructions – NEON has 100+ specific SIMD instructions, SIMD uses existing arithmetic/logic instructions
- Data types – NEON supports wider variety of integer/float data types
- Features – NEON has more advanced capabilities like polynomials, fused multiply-add etc.
In summary, NEON is a dedicated high performance SIMD engine, while SIMD support in Cortex-M7 provides more basic parallel processing capabilities using existing CPU resources.
NEON Architecture
The NEON architecture is designed as a coprocessor that works alongside the main ARM CPU core to provide acceleration for SIMD workloads. The key architectural components of NEON are:
- NEON Register Bank – 16 128-bit registers for SIMD operations
- NEON Execution Unit – Hardware for executing NEON instructions
- NEON Load/Store Units – For efficient memory access
- NEON Instruction Set – 100+ instructions for SIMD processing
NEON instructions can perform parallel integer, single precision float, double precision float, and polynomial ops. Instructions are provided for data processing, memory access, conversion between data types, permutation, packing/unpacking etc.
NEON is integrated with the CPU so that scalar ARM code can set up data, then invoke NEON SIMD operations as needed, and continue with scalar processing of the results. This allows efficiently accelerating suitable portions of applications.
SIMD Implementation in Cortex-M7
Unlike NEON, Cortex-M7 does not have dedicated SIMD execution units. Instead, it exploits the existing CPU registers and arithmetic/logic units to perform parallel operations on data sets.
Key implementation aspects include:
- 32-bit registers used as 64-bit SIMD registers
- ALU supports 64-bit SIMD arithmetic/logic instructions
- Barrel shifter supports 64-bit shifts
- Dual 16-bit instructions allow 32-bit SIMD ops
- Saturation support avoids overflow issues
- Packing/unpacking between SIMD and scalar registers
So SIMD support is provided by enhancing the existing CPU datapth to perform parallel 64-bit operations on register pairs. This provides decent speedups for workloads with regular parallelism using standard code and registers.
Use Cases and Performance
While both NEON and SIMD in Cortex-M7 aim to accelerate suitable workloads using parallel processing, their different capabilities make them suited for different use cases.
NEON Use Cases
- Digital signal processing – audio/video codecs, filters, FFTs etc.
- Image processing – Convolutional neural networks, filtering, transformations etc.
- Computer vision – Object detection, image recognition etc.
- Speech recognition – Neural networks, voice encoding etc.
Typical performance improvements from NEON are 2-3X for suitable algorithms.
Cortex-M7 SIMD Use Cases
- Digital signal processing – FIR filters, IIR filters, FFT
- Image processing – Matrix operations, convolutions
- Data analysis – Statistics, regression
- Control systems – Sensor fusion, controls code
Typical Cortex-M7 SIMD speedups are around 2X for appropriate code segments.
So in summary, NEON provides much higher throughput optimized for media workloads, while Cortex-M7 SIMD allows more modest but useful acceleration in embedded applications.
Programming Considerations
Extracting maximum performance from NEON and SIMD requires adopting suitable programming practices.
NEON Programming
- Understand NEON architecture and instruction set
- Identify hotspots suitable for NEON acceleration
- Maximize use of wide NEON registers
- Optimize memory access patterns to use NEON loads/stores
- Align data structures and addresses for memory operations
- Minimize type conversions and movement between NEON and ARM cores
Cortex-M7 SIMD Programming
- Identify independent operations that can be parallelized
- Use dual 16-bit instructions where possible
- Combine operations using parallel arithmetic/logic instructions
- Pack and unpack between SIMD and scalar registers efficiently
- Ensure memory accesses and data alignment support SIMD widths
Efficiently using these capabilities requires adopting a parallel processing mindset during programming.
Conclusion
In conclusion, the key difference between NEON and SIMD in Cortex-M7 is:
- NEON is a dedicated high performance SIMD engine for accelerating media processing workloads like imaging, computer vision, speech recognition etc. in Cortex-A series processors.
- SIMD in Cortex-M7 provides more basic parallel processing capabilities using existing CPU resources, suitable for modest acceleration of DSP and embedded control applications.
So NEON targets specialized high throughput workloads with extensive SIMD capabilities, while Cortex-M7 SIMD focuses on straightforward acceleration of common embedded algorithms. Both can provide significant speedups but for different application domains.
