Function main

examples/neon_gemm_qasymm8.cpp:96–273

}

int main(int argc, char **argv)
{
    Tensor src1;
    Tensor src2;
    Tensor dst0;
    Tensor q_src1;
    Tensor q_src2;
    Tensor q_dst0;
    Tensor q_res;
    Tensor q_res_output;
    size_t M             = 4;
    size_t N             = 4;
    size_t K             = 4;
    bool   default_input = true;

    // Parse args
    if (argc < 3) /* case default matrix sizes */
    {
        // Print help
        std::cout << "Usage: ./build/neon_gemm_qasymm8 M N K\n";
        std::cout << "Too few or no inputs provided. Using default M=4, N=4, K=4\n\n";
    }
    else /* case M N K arguments provided */
    {
        M             = strtol(argv[1], nullptr, 10);
        N             = strtol(argv[2], nullptr, 10);
        K             = strtol(argv[3], nullptr, 10);
        default_input = false;
    }

    /*** Floating point matrix multiplication ***/

    // Initialise input matrices
    NEGEMM fgemm{};

    src1.allocator()->init(TensorInfo(TensorShape(K, M), 1, DataType::F32));
    src2.allocator()->init(TensorInfo(TensorShape(N, K), 1, DataType::F32));
    dst0.allocator()->init(TensorInfo(TensorShape(N, M), 1, DataType::F32));
    fgemm.configure(&src1, &src2, nullptr, &dst0, 1, 0);

    // Allocate matrices
    src1.allocator()->allocate();
    src2.allocator()->allocate();
    dst0.allocator()->allocate();

    // Fill in tensors, by default fill in with known data - for easy testing
    auto *src1_ptr = reinterpret_cast<float *>(src1.buffer());
    auto *src2_ptr = reinterpret_cast<float *>(src2.buffer());
    auto *dst0_ptr = reinterpret_cast<float *>(dst0.buffer());

    // Fill in: one is the identity matrix, other is sequential values
    // src1: Identity matrix
    for (size_t i = 0; i < M * K; i++)
    {
        src1_ptr[i] = 0;
    }
    for (size_t i = 0; i < M; i++)
    {