SIMD playground

experiments with simd instructions on x86 and ARM platforms

This Repo is a collection of various code examples and implementations using SIMD paralleism.

Introduction to SIMD:

SIMD instructions provide powerful data crunching capabilities by allowing operations on Multiple Data using a Single Instruction call.

The power of parallelism dominates against faster single pipeline data processing methods.

This technology aids in places such as Improved HPC, Signal Processing, etc.

SIMD instruction types:

each architecture has its own implmentation of it.

x86:

• For x86 it has been SSE (Streaming SIMD Extensions) and AVX (Advanced Vector Extensions) instructions.

• The x86 simplified library will mostly focus on implmentations using AVX2 (256 bit wide operations) but may also use AVX (128 bit wide operations)

ARMv8

• For the Arm implmentation it will be using NEON instructions.

Testing:

setup up the compile environment with:

git clone https://github.com/ZephyrLabs/simd-playground
cd simd-playground
make init
make -s all

run an example with:

make <example name>

example: make vec_add

note: you will need to specify the correct arch makefile with:

-f avx-Makefile/neon-Makefile accordingly.

Examples:

• vec_add
• vec_sub
• vec_mul
• vec_div
• tensor_add
• tensor_sub
• tensor_mul
• convolution

sample run:

Compiler: g++ Optimizations: none

-------------------AVX-VECTOR-ADD------------------
Time taken by normal function: 5819 microseconds
Time taken by AVX function: 3437 microseconds
Speed Uplift: 169.305 %
---------------------------------------------------
-------------------AVX-TENSOR-SUB------------------
Time taken by normal function: 6056 microseconds
Time taken by AVX function: 3539 microseconds
Speed Uplift: 171.122 %
---------------------------------------------------
-------------------AVX-TENSOR-MUL------------------
Time taken by normal function: 5886 microseconds
Time taken by AVX function: 3348 microseconds
Speed Uplift: 175.806 %
---------------------------------------------------
-------------------AVX-VECTOR-DIV------------------
Time taken by normal function: 5867 microseconds
Time taken by AVX function: 3416 microseconds
Speed Uplift: 171.751 %
---------------------------------------------------
-------------------AVX-TENSOR-ADD------------------
Time taken by normal function: 168 nanoseconds
Time taken by AVX function: 85 nanoseconds
Speed Uplift: 197.647 %
---------------------------------------------------
-------------------AVX-TENSOR-SUB------------------
Time taken by normal function: 186 nanoseconds
Time taken by AVX function: 75 nanoseconds
Speed Uplift: 248 %
-------------------AVX-TENSOR-MUL------------------
Time taken by normal function: 306 nanoseconds
Time taken by AVX function: 184 nanoseconds
Speed Uplift: 166.304 %
---------------------------------------------------
------------------AVX-CONVOLUTION------------------
Time taken by normal function: 2673 nanoseconds
Time taken by AVX function: 1476 nanoseconds
Speed Uplift: 181.098 %
---------------------------------------------------