.. _conv_depthwise: Depthwise Convolution Prototype and Function List ------------------------------------------------- Description ^^^^^^^^^^^ This kernel implements a 2D depthwise convolution operation applying each filter channel to each input channel separately. It applies each filter of weights tensor to each framed area of the size of the input tensor. The main difference with general 2D convolution is that to calculate one channel of output feature map, only one channel of input feature map is used. In contrast, for general 2D convolution all channels of input feature map in a framed area are used to calculate the value in one output channel. A depthwise convolution operation is shown in Figure :ref:`f_conv_depwise`. .. _f_conv_depwise: .. figure:: ../images/conv_depwise.png :align: center Depthwise Convolution .. For example, in a HWCN data layout, if in feature map is :math:`(Hi, Wi, Ci)` and weights is :math:`(Hk, Wk, 1, Co)`, the output feature map is :math:`(Ho, Wo, Co)` tensor where the spatial dimensions comply with the system of equations :eq:`eq_conv2d_shapes`. .. note:: For more details on calculations, see chapter 2 of `A guide to convolution arithmetic for deep learning `_. .. This kernel does not support channel multiplier logic that allows applying several filters for each channel of input. Such functionality refers to group convolution and can be obtained by the corresponding kernel (see :ref:`grp_conv`). Optionally, a saturating ReLU activation function can be applied to the result of the convolution during the function's execution. For more information on supported ReLU types and calculations, see :ref:`relu_prot`. This is a MAC-based kernel which implies accumulation. See :ref:`quant_accum_infl` for more information on related quantization aspects. The Number of accumulation series in terms of above-defined variables is equal to :math:`(Hk * Wk)`. Functions ^^^^^^^^^ Kernels which implement depthwise convolution have the following prototype: .. code:: c mli_status mli_krn_depthwise_conv2d_hwcn_( const mli_tensor *in, const mli_tensor *weights, const mli_tensor *bias, const mli_conv2d_cfg *cfg, mli_tensor *out); .. where ``data_format`` is one of the data formats listed in Table :ref:`mli_data_fmts` and the function parameters are shown in the following table: .. table:: Depthwise Convolution Function Parameters :align: center :widths: auto +---------------+----------------------+--------------------------------------------------+ | **Parameter** | **Type** | **Description** | +===============+======================+==================================================+ | ``in`` | ``mli_tensor *`` | [IN] Pointer to constant input tensor | +---------------+----------------------+--------------------------------------------------+ | ``weights`` | ``mli_tensor *`` | [IN] Pointer to constant weights tensor | +---------------+----------------------+--------------------------------------------------+ | ``bias`` | ``mli_tensor *`` | [IN] Pointer to constant bias tensor | +---------------+----------------------+--------------------------------------------------+ | ``cfg`` | ``mli_conv2d_cfg *`` | [IN] Pointer to convolution parameters structure | +---------------+----------------------+--------------------------------------------------+ | ``out`` | ``mli_tensor *`` | [IN | OUT] Pointer to output feature map tensor. | | | | Result is stored here | +---------------+----------------------+--------------------------------------------------+ .. Here is a list of all available Depth-Wise Convolution functions: .. tabularcolumns:: |\Y{0.65}|\Y{0.3}| .. table:: List of Available Depth-Wise Convolution Functions :align: center :class: longtable +-----------------------------------------------------+-------------------------------------+ | **Function Name** | **Details** | +=====================================================+=====================================+ | ``mli_krn_depthwise_conv2d_hwcn_sa8_sa8_sa32`` | In/out layout: **HWC** | | | | | | Weights layout: **HWCN** | | | | | | In/out/weights data format: **sa8** | | | | | | Bias data format: **sa32** | +-----------------------------------------------------+-------------------------------------+ | ``mli_krn_depthwise_conv2d_hwcn_fx16`` | In/out layout: **HWC** | | | | | | Weights layout: **HWCN** | | | | | | All tensors data format: **fx16** | +-----------------------------------------------------+-------------------------------------+ | ``mli_krn_depthwise_conv2d_hwcn_fx16_fx8_fx8`` | In/out layout: **HWC** | | | | | | Weights layout: **HWCN** | | | | | | In/out data format: **fx16** | | | | | | Weights/Bias data format: **fx8** | +-----------------------------------------------------+-------------------------------------+ | ``mli_krn_depthwise_conv2d_hwcn_fx16_k3x3`` | In/out layout: **HWC** | | | | | | Weights layout: **HWCN** | | | | | | All tensors data format: **fx16** | | | | | | Width of weights tensor: **3** | | | | | | Height of weights tensor: **3** | +-----------------------------------------------------+-------------------------------------+ | ``mli_krn_depthwise_conv2d_hwcn_sa8_sa8_sa32_k3x3`` | In/out layout: **HWC** | | | | | | Weights layout: **HWCN** | | | | | | In/out/weights data format: **sa8** | | | | | | Bias data format: **sa32** | | | | | | Width of weights tensor: **3** | | | | | | Height of weights tensor: **3** | +-----------------------------------------------------+-------------------------------------+ | ``mli_krn_depthwise_conv2d_hwcn_fx16_fx8_fx8_k3x3`` | In/out layout: **HWC** | | | | | | Weights layout: **HWCN** | | | | | | In/out data format: **fx16** | | | | | | Weights/Bias data format: **fx8** | | | | | | Width of weights tensor: **3** | | | | | | Height of weights tensor: **3** | +-----------------------------------------------------+-------------------------------------+ | ``mli_krn_depthwise_conv2d_hwcn_sa8_sa8_sa32_k5x5`` | In/out layout: **HWC** | | | | | | Weights layout: **HWCN** | | | | | | In/out/weights data format: **sa8** | | | | | | Bias data format: **sa32** | | | | | | Width of weights tensor: **5** | | | | | | Height of weights tensor: **5** | +-----------------------------------------------------+-------------------------------------+ | ``mli_krn_depthwise_conv2d_hwcn_fx16_k5x5`` | In/out layout: **HWC** | | | | | | Weights layout: **HWCN** | | | | | | All tensors data format: **fx16** | | | | | | Width of weights tensor: **5** | | | | | | Height of weights tensor: **5** | +-----------------------------------------------------+-------------------------------------+ | ``mli_krn_depthwise_conv2d_hwcn_fx16_fx8_fx8_k5x5`` | In/out layout: **HWC** | | | | | | Weights layout: **HWCN** | | | | | | In/out data format: **fx16** | | | | | | Weights/Bias data format: **fx8** | | | | | | Width of weights tensor: **5** | | | | | | Height of weights tensor: **5** | +-----------------------------------------------------+-------------------------------------+ .. Conditions ^^^^^^^^^^ Ensure that you satisfy the following general conditions before calling the function: - ``in``, ``out``, ``weights`` and ``bias`` tensors must be valid (see :ref:`mli_tnsr_struc`) and satisfy data requirements of the selected version of the kernel. - Shapes of ``in``, ``out``, ``weights`` and ``bias`` tensors must be compatible, which implies the following requirements: - ``in`` and ``out`` are 3-dimensional tensors (rank==3). Dimensions meaning, and order (layout) is aligned with the specific version of kernel. - ``weights`` is a 4-dimensional tensor (rank==4). Dimensions meaning, and order (layout) is aligned with the specifc kernel. - ``bias`` must be a one-dimensional tensor (rank==1). Its length must be equal to :math:`Co` (output channels OR number of filters). - Channel :math:`Ci` dimension of ``weights`` tensors must be 1. - Channel :math:`Ci` dimension of ``in`` and :math:`Co` (output channels OR number of filters) dimension of ``weights`` and ``out`` tensors must be equal. - Shapes of ``in``, ``out`` and ``weights`` tensors together with ``cfg`` structure must satisfy the equations :eq:`eq_conv2d_shapes` - Effective width and height of the ``weights`` tensor after applying dilation factor (see :eq:`eq_conv2d_shapes`) must not exceed appropriate dimensions of the ``in`` tensor. - ``in`` and ``out`` tensors must not point to overlapped memory regions. - ``mem_stride`` of the innermost dimension must be equal to 1 for all the tensors. - ``padding_top`` and ``padding_bottom`` parameters must be in the range of [0, :math:`\hat{Hk}`) where :math:`\hat{Hk}` is the effective kernel height (See :eq:`eq_conv2d_shapes`) - ``padding_left`` and ``padding_right`` parameters must be in the range of [0, :math:`\hat{Wk}`) where :math:`\hat{Wk}` is the effective kernel width (See :eq:`eq_conv2d_shapes`) - ``stride_width`` and ``stride_height`` parameters must not be equal to 0. - ``dilation_width`` and ``dilation_height`` parameters must not be equal to 0. For **fx16** and **fx16_fx8_fx8** versions of kernel, in addition to the general conditions, ensure that you satisfy the following quantization conditions before calling the function: - The number of ``frac_bits`` in the ``bias`` and ``out`` tensors must not exceed the sum of ``frac_bits`` in the ``in`` and ``weights`` tensors. For **sa8_sa8_sa32** versions of kernel, in addition to the general conditions, ensure that you satisfy the following quantization conditions before calling the function: - ``in`` and ``out`` tensor must be quantized on the tensor level. This implies that each tensor contains a single scale factor and a single zero offset. - Zero offset of ``in`` and ``out`` tensors must be within [-128, 127] range. - ``weights`` and ``bias`` tensors must be symmetric. Both must be quantized on the same level. Allowed Options: - Per Tensor level. This implies that each tensor contains a single scale factor and a single zero offset equal to 0. - Per :math:`Co` dimension level (number of filters). This implies that each tensor contains separate scale point for each sub-tensor. All tensors contain single zero offset equal to 0. - Scale factors of bias tensor must be equal to the multiplication of input scale factor broadcasted on weights array of scale factors. See the example for the similar condition in the :ref:`conv_2d`. Ensure that you satisfy the platform-specific conditions in addition to those listed above (see the :ref:`platform_spec_chptr` chapter). Result ^^^^^^ These functions only modify the memory pointed by ``out.data.mem`` field. It is assumed that all the other fields of ``out`` tensor are properly populated to be used in calculations and are not modified by the kernel. Depending on the debug level (see section :ref:`err_codes`) this function performs a parameter check and returns the result as an ``mli_status`` code as described in section :ref:`kernl_sp_conf`.