FT_DSP.gitlink.net/master/html/_sources/functionlib/dsplib/maxpoolfusion.rst.txt

MaxPoolFusion
=================


    对输入张量执行最大池化操作

    .. math::

        \text{output}_{b, h_o, w_o, c} =
            \operatorname{clip}\Bigg(
            \max_{h_i, w_i \in \mathcal{W}(h_o, w_o)}
            \Big(
                \text{input}_{b,\; h_i,\; w_i,\; c}
            \Big),\;
            \text{minf},\; \text{maxf}
            \Bigg)

    其中，窗口区域 :math:`\mathcal{W}(h_o, w_o)` 的定义如下：

    .. math::

        h_i = h_o \cdot \text{stride}_h - \text{pad}_u + \Delta h

        w_i = w_o \cdot \text{stride}_w - \text{pad}_l + \Delta w

        \Delta h \in [0,\ \text{win}_h - 1],\quad
        \Delta w \in [0,\ \text{win}_w - 1]

    有效窗口点满足：

    .. math::

        0 \le h_i < \text{in}_h,\qquad
        0 \le w_i < \text{in}_w

    原始窗口起点定义为：

    .. math::

        h_{\text{start}} = h_o \cdot \text{stride}_h - \text{pad}_u

    .. math::

        w_{\text{start}} = w_o \cdot \text{stride}_w - \text{pad}_l

    合法采样范围为：

    .. math::

        \Delta h \in
        \Big[
            \max(0,\ -h_{\text{start}}),\;
            \min(\text{win}_h,\ \text{in}_h - h_{\text{start}})
        \Big)

    .. math::

        \Delta w \in
        \Big[
            \max(0,\ -w_{\text{start}}),\;
            \min(\text{win}_w,\ \text{in}_w - w_{\text{start}})
        \Big)

    最大池化计算：

    .. math::

        v_{\max} =
            \max_{\Delta h,\ \Delta w}\;
            \text{input}_{b,\;
                h_{\text{start}} + \Delta h,\;
                w_{\text{start}} + \Delta w,\;
                c}

    最终输出：

    .. math::

        \text{output}_{b, h_o, w_o, c} =
            \min\big(\max(v_{\max},\ \text{minf}),\ \text{maxf}\big)



    输入：
        - **input** - 输入张量指针，采用 **NHWC 格式**，形状为   :math:`[batch,\ in\_h,\ in\_w,\ channel]`
        - **params** - 参数数组，包含以下元素：
            - **in_w** - 输入张量的宽度 (W)
            - **in_h** - 输入张量的高度 (H)
            - **win_w** - 池化窗口的宽度，即窗口在 W 方向的大小
            - **win_h** - 池化窗口的高度，即窗口在 H 方向的大小
            - **output_w** - 输出特征图的宽度
            - **output_h** - 输出特征图的高度
            - **batch** - 批次大小，即输入中的 batch 数
            - **channel** - 通道数 C ，每个池化位置都分别对 C 个通道独立执行最大池化与裁剪
            - **stride_w** - 池化窗口在 W 方向的步长
            - **stride_h** - 池化窗口在 H 方向的步长
            - **pad_l** - 输入特征图左侧的填充大小
            - **pad_u** - 输入特征图上侧的填充大小
            - **minf** - 输出结果的下界值，传指针
            - **maxf** - 输出结果的上界值，传指针
        - **core_mask** - 核心掩码，指定使用的计算核心

    输出：
        - **output** - 输出张量指针，采用 **NHWC 格式**，形状为  :math:`[batch,\ output\_h,\ output\_w,\ channel]`。


    支持平台：
        ``FT78NE``
        ``MT7004``

    .. note::
        - FT78NE 支持fp32, fp64
        - MT7004 支持fp16, fp32
        - 调用时将除 core_mask 外的参数打包通过 long long params 数组传入，顺序为：
          input, output, in_w, in_h, win_w, win_h, output_w, output_h, batch, channel,
          stride_w, stride_h, pad_l, pad_u, minf, maxf

**共享存储版本:**

.. c:function:: void hp_maxpool_fusion_s(half* input, half* output, long long *params, int core_mask)
.. c:function:: void fp_maxpool_fusion_s(float* input, float* output, long long *params, int core_mask)
.. c:function:: void dp_maxpool_fusion_s(double* input, double* output, long long *params, int core_mask)

    **C调用示例：**

    .. code-block:: c
        :linenos:
        :emphasize-lines: 45

        //FT78NE示例
        #include <stdio.h>
        
        int main(int argc, char* argv[]) {
            float* input_ptr = (float*)0xA0000000;
            float* output_ptr = (float*)0xB0000000;
            float* check_ptr = (float*)0xC0000000;
            int in_w = 32;
            int in_h = 32;
            int win_w = 6;
            int win_h = 6;
            int batch = 4;
            int channel = 2;
            int stride_w = 4;
            int stride_h = 4;
            int pad_l = 0;
            int pad_u = 0;
            float minf = 0.0f;
            float maxf = 50.0f;

            // 根据标准公式计算输出尺寸
            int dividor = in_w + pad_l * 2 - win_w;
            int output_w = (dividor + stride_w - 1) / stride_w + 1;
            int dividor2 = in_h + pad_u * 2 - win_h;
            int output_h = (dividor2 + stride_h - 1) / stride_h + 1;

            long long params[16];
            params[0] = (long long)in_w;
            params[1] = (long long)in_h;
            params[2] = (long long)win_w;
            params[3] = (long long)win_h;
            params[4] = (long long)output_w;
            params[5] = (long long)output_h;
            params[6] = (long long)batch;
            params[7] = (long long)channel;
            params[8] = (long long)stride_w;
            params[9] = (long long)stride_h;
            params[10] = (long long)pad_l;
            params[11] = (long long)pad_u;
            params[12] = (long long)&minf;
            params[13] = (long long)&maxf;
            int core_mask = 0x0f;
            
            fp_maxpool_fusion_s(input_ptr, output_ptr, params, core_mask);
            return 0;
        }


**私有存储版本:**

.. c:function:: void hp_maxpool_fusion_p(half* input, half* output, long long *params)
.. c:function:: void fp_maxpool_fusion_p(float* input, float* output, long long *params)
.. c:function:: void dp_maxpool_fusion_p(double* input, double* output, long long *params)

    **C调用示例：**

    .. code-block:: c
        :linenos:
        :emphasize-lines: 44

        //FT78NE示例
        #include <stdio.h>
        
        int main(int argc, char* argv[]) {
            float* input_ptr = (float*)0xA0000000;
            float* output_ptr = (float*)0xB0000000;
            float* check_ptr = (float*)0xC0000000;
            int in_w = 32;
            int in_h = 32;
            int win_w = 6;
            int win_h = 6;
            int batch = 4;
            int channel = 2;
            int stride_w = 4;
            int stride_h = 4;
            int pad_l = 0;
            int pad_u = 0;
            float minf = 0.0f;
            float maxf = 50.0f;

            // 根据标准公式计算输出尺寸
            int dividor = in_w + pad_l * 2 - win_w;
            int output_w = (dividor + stride_w - 1) / stride_w + 1;
            int dividor2 = in_h + pad_u * 2 - win_h;
            int output_h = (dividor2 + stride_h - 1) / stride_h + 1;

            long long params[16];
            params[0] = (long long)in_w;
            params[1] = (long long)in_h;
            params[2] = (long long)win_w;
            params[3] = (long long)win_h;
            params[4] = (long long)output_w;
            params[5] = (long long)output_h;
            params[6] = (long long)batch;
            params[7] = (long long)channel;
            params[8] = (long long)stride_w;
            params[9] = (long long)stride_h;
            params[10] = (long long)pad_l;
            params[11] = (long long)pad_u;
            params[12] = (long long)&minf;
            params[13] = (long long)&maxf;
            
            fp_maxpool_fusion_p(input_ptr, output_ptr, params);
            return 0;
        }