Description

Dispatch to vectorized function based on the capability available at runtime.

AVX2 Instruction Set was introduced by Intel in 2014, which provided significant performance boost over Intel's AVX and SIMD instruction sets. According to the official release, AVX2 provides following benefits over other existing instruction sets during that period:

1. Doubling the number of FLOPS per clock cycle.
2. 256-bit integer instructions.
3. Floating-point fused multiply-add instructions and gather operations.

This Dispatcher aims to dispatch the operation to the capability depending on your CPU during the runtime itself. If AVX2 is supported and demanded (use_cpu is not set to true), then the operator will be dispatched to AVX2 instruction set (see supported operators section).

Build Instructions

Update the submodule: git submodule update --init

Build the source code:

$ mkdir build && cd build
$ cmake ..
$ make
$ ./test_bin  # Run tests

Usage

As of Release 0.0.1, the Dispatcher supports arrays of size 4 and these operators for both AVX2 and SIMT instruction sets. Here is a sample code to start with: (also check main.cpp for complete template)

Let's start with the necessary header files.

#include <array>  // for using std::array
#include "Dispatcher.h"  // for accessing Dispatcher struct
#include "avx/interface.h"  // for accessing avx2::<op>
#include "cpu/interface.h"  // for accessing cpu::<op>

Initialize 2 arrays x, y using std::array<double, 4> x = {1.0, 2.0, 3.0, 4.0} and std::array<double, 4> y = {4.0, 3.0, 2.0, 1.0} and initialize the Dispatcher.

Dispatcher op_dispatcher = Dispatcher{};
Dispatcher cpu_dispatcher = Dispatcher{};

In order to allow cpu_dispatcher to use CPU, you need to set use_cpu flag (default is false): cpu_dispatcher.use_cpu = true. Now since the arrays and dispatchers are initialized, go ahead and dispatch the op (mul for example, see supported operators section):

op_dispatcher.binary_dispatch(avx2::mul, cpu::mul)(x, y);
cpu_dispatcher.binary_dispatch(avx2::mul, cpu::mul)(x, y);

If the dispatcher found AVX2 capability available, the op_dispatcher.binary_dispatch call should print AVX IMPL while cpu_dispatcher.binary_dispatch should print NORMAL IMPL.

Supported Operators

As of Release 0.0.1, the Dispatcher supports following ops:

• mul