Intel's IPP Library (Integrated Performance Primitives)

        IPP The library contains many low-level optimized functions . If already installed IPP library ,opencv It will also be called automatically at runtime .NVIDIA The company aims at GPU Developed GPU Accelerated images 、 video 、 Signal processing library NPP（NVIDIA Performance Primitives）.NPP and IPP Library functions are basically one-to-one correspondence , The difference is one in CPU Calculate one on GPU Count up . Highly optimized NPP Computing unit , The execution speed is faster than using only CPU Be quick 80 times .

        CUDA It covers a wide range of functions , Hardware functions almost reflect NV Home GPGPU All the possibilities of , In addition, the driver and software upper layer packaging （ Various libraries , such as cuBLAS,cuFFT,cuDNN And so on ）, And a complete suite of development tools （ compiler 、 The debugger 、profiler wait ）.CUDA Many functions are deeply coupled with hardware , Inconsistent hardware design , Software encapsulation to maintain consistency , The workload is really beyond the average company's digestion .

        CUDA Our programming model is largely reused C Language , So there is cudaMalloc、cudaFree、cudaMemset This custom version “C API”.CUDA Most of the mathematical functions in cmath In the form of , such as exp() yes double function ,expf() yes float Version of exp.cuBLAS Not to mention , Most of the API All from LAPACK（ although BLAS This set of API Now it is almost a standard interface ）.cuFFT Of API And FFTW Although there is no small difference , But the inheritance relationship between the two is also obvious . these API Basically from the once very popular