1.下載

https://github.com/ARM-software/ComputeLibrary

2.由於我交叉編譯器已經加入環境變數，修改SConstruct檔案下

3.編譯

opencl一起編譯進去embed_kernels=1

scons Werror=0 debug=0 asserts=1 neon=1 opencl=1 embed_kernels=1 os=linux arch=armv7a

如果只用到neon加速

scons Werror=0 debug=0 asserts=0 neon=1 opencl=1 os=linux arch=armv7a

debug和asserts用於除錯，會增加執行時間；
 


4.連結
編譯成功後會在根目錄下生成build資料夾，我只用到下面兩個so檔案。

資料夾

#include "arm_compute/core/Types.h"

#include "tests/Utils.h"

#include "arm_compute/runtime/NEON/NEScheduler.h"

#include "arm_compute/runtime/NEON/functions/NEGEMM.h"

5.執行，矩陣乘法

neon加速

TensorShape AShape(K,M);

    TensorShape BShape(N,K);

    TensorShape OShape(N,M);

    



    Tensor ATensor, BTensor, OTensor , CTensor;



    ATensor.allocator()
 
->init(TensorInfo(AShape, Format::F32));

    BTensor.allocator()->init(TensorInfo(BShape, Format::F32));

    OTensor.allocator()->init(TensorInfo(OShape, Format::F32));



    NEGEMM armGemm;

    armGemm.configure(&ATensor, &BTensor,nullptr, &OTensor,ALPHA, 0.0);



    ATensor.allocator()
 
->allocate();

    BTensor.allocator()->allocate();

    OTensor.allocator()->allocate();





    Window A_window;

    A_window.use_tensor_dimensions(ATensor.info()->tensor_shape());

    Iterator A_it(&ATensor, A_window);

    execute_window_loop(A_window, [&](const Coordinates & id)

    {

        *reinterpret_cast<float *>(A_it.ptr()) = A[id.z() * (M * K) + id.y() * K + id.x()];

    },

    A_it);



    Window B_window;

    B_window.use_tensor_dimensions(BTensor.info()->tensor_shape());

    Iterator B_it(&BTensor, B_window);

    execute_window_loop(B_window, [&](const Coordinates & id)

    {

        *reinterpret_cast<float *>(B_it.ptr()) = B[id.z() * (K * N) + id.y() * N + id.x()];

    },

    B_it);



    //warmup run 

    armGemm.run();



    

    for(int h = 0; h < M; h++)

    {

        for(int w = 0; w < N; w++)

        {

            C[h*N + w] = *reinterpret_cast<float*>( OTensor.buffer() + 

                OTensor.info()->offset_element_in_bytes(Coordinates(w,h,0)));

        }

    }

opencl加速需要opencl2.0以上或者支援-cl-arm-non-uniform-work-group-size