Method run

src/gpu/cl/operators/ClGemmConv2d.cpp:582–640

}

void ClGemmConv2d::run(ITensorPack &tensors)
{
    prepare(tensors);

    auto src                = tensors.get_const_tensor(ACL_SRC_0);
    auto biases             = tensors.get_const_tensor(ACL_SRC_2);
    auto dst                = tensors.get_tensor(ACL_DST);
    auto gemm_input_to_use  = src;
    auto gemm_output_to_use = dst;

    CLAuxTensorHandler im2col_output(offset_int_vec(Im2ColOutput), _im2col_output, tensors, false);
    CLAuxTensorHandler gemm_output(offset_int_vec(GemmOutput), _gemm_output, tensors, false);
    CLAuxTensorHandler weights_reshaped(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors, false);

    // Run im2col
    if (!_skip_im2col)
    {
        ITensorPack pack = {{TensorType::ACL_SRC, src}, {TensorType::ACL_DST, im2col_output.get()}};
        CLScheduler::get().enqueue_op(*_im2col_kernel, pack, false);
        gemm_input_to_use = im2col_output.get();
    }
    if (!_skip_col2im)
    {
        gemm_output_to_use = gemm_output.get();
    }
    ITensorPack pack_mm = tensors;
    pack_mm.add_const_tensor(TensorType::ACL_SRC_0, gemm_input_to_use);
    pack_mm.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get());
    if (!_append_bias)
    {
        pack_mm.add_const_tensor(TensorType::ACL_SRC_2, biases);
    }
    pack_mm.add_tensor(TensorType::ACL_DST, gemm_output_to_use);
    // Runs ClGemm or ClGemmLowpMatrixMultiplyCore functions
    if (_is_quantized)
    {
        // Run gemmlowp
        _mm_gemmlowp->run(pack_mm);
    }
    else
    {
        // Run gemm
        _mm_gemm->run(pack_mm);
    }

    // Reshape output matrix
    if (!_skip_col2im)
    {
        ITensorPack pack = {{TensorType::ACL_SRC, gemm_output_to_use}, {TensorType::ACL_DST, dst}};
        CLScheduler::get().enqueue_op(*_col2im_kernel.get(), pack, false);
    }

    //Run Activation Layer if we cannot fuse in GEMM
    if (!_fuse_activation)
    {
        ITensorPack pack = {{TensorType::ACL_SRC, dst}, {TensorType::ACL_DST, dst}};
        CLScheduler::get().enqueue_op(*_activation_kernel.get(), pack, false);
    }