decl_opr('Convolution',
         pyname='convolution_v0',
         inputs=[Doc('src',
                     'input image in (batch, channel, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(out channel, in channel, kern row, kern col) format')],
         params=[('param', 'ConvolutionV0'),
                 ('execution_polity', 'ExecutionPolicyV0')],
         desc='batched convolution on channeled 2D images')

decl_opr('Convolution',
        inputs=[Doc('src',
                     'input image in (batch, channel, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(out channel, in channel, kern row, kern col) format')],
         params=[('param', 'Convolution'),
                 ('execution_polity', 'ExecutionPolicy')],
         desc='batched convolution on channeled 2D images',
         version=2, has_out_dtype=True)

decl_opr('ConvolutionBackwardData',
         pyname='deconvolution_v0',
         inputs=[Doc('src',
                     'input image in (batch, channel, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(out channel, in channel, kern row, kern col) format')],
         params=[('param', 'ConvolutionV0'),
                 ('execution_polity', 'ExecutionPolicyV0')],
         body=[
             'a, b = all_inputs',
             'all_inputs = [b, a]'
         ],
         desc='batched deconvolution on channeled 2D images; the underlying '
         'computation is in fact gradient of convolution w.r.t. data')

decl_opr('ConvolutionBackwardData',
         pyname='deconvolution',
         inputs=[Doc('src',
                     'input image in (batch, channel, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(out channel, in channel, kern row, kern col) format')],
         params=[('param', 'Convolution'),
                 ('execution_polity', 'ExecutionPolicy')],
         body=[
             'a, b = all_inputs',
             'all_inputs = [b, a]'
         ],
         desc='batched deconvolution on channeled 2D images; the underlying '
         'computation is in fact gradient of convolution w.r.t. data',
         version=2, has_out_dtype=True)

decl_opr('MaskConvolution',
         inputs=[Doc('src',
                     'input image in (batch, channel, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(out channel, in channel, kern row, kern col) format'),
                 Doc('mask',
                     'the 0/1 matrix, each element at (i, j) indicates the '
                     '*output(i, j)* of DefaultConvolution should be zero')],
         params=[('param', 'Convolution')],
         desc=('batched mask conv on channeled 2D images, mask is correspoding '
               'to output'),
         version=2)

decl_opr('MaskPropagate',
         inputs=[Doc('src',
                     '0/1 matrix for MaskConvolution\'s input')],
         params=[('param', 'MaskPropagate')],
         desc=('calculates the mask for output by given kernel, stride and '
               'padding'))

decl_opr('Images2Neibs',
         inputs=[Doc('src',
                     'input image in (batch, channel, row, col) format')],
         params='Images2Neibs',
         desc=Doc(None,
r"""
    Apply a sliding window to input tensor and copy content in the window to
    corresponding output location. Assume input shape is :math:`(N, C, IH, IW)`,
    then output shape would be :math:`(N, C, OH, OW, window_h, window_w)` where
    :math:`(OH, OW)` would be computed from padding, stride, window and
    :math:`(IH, IW)`, as in convolution. For each output location, we have;

    .. math::

        out_{n, c, oh, ow, wh, ww} &= src_{n, c, ih+wh, iw+ww} \\\\
        \\text{where } & ih=-pad_h+oh \\times stride_h \\\\
                       & iw=-pad_w+ow \\times
        stride_w
"""))

decl_opr('Local',
         pyname='local',
         inputs=[Doc('src',
                     'input image in (batch, channel, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(out row, out col, in channel, '
                     'kern row, kern col, out channel) format')],
         params='Convolution',
         desc='batched convolution on channeled 2D images, but kernels are '
         'not shared across different output positions', version=2)

"""
decl_opr('Local',
         pyname='local_v1',
         inputs=[Doc('src',
                     'input image in (batch, channel, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(out row, out col, in channel, '
                     'kern row, kern col, out channel) format')],
         params='Convolution',
         desc='batched convolution on channeled 2D images, but kernels are '
         'not shared across different output positions',
         version=1)
"""

decl_opr('GroupLocal',
         inputs=[Doc('src',
                     'input image in (batch, channel, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(group, out row, out col, in channel / group, '
                     'kern row, kern col, out channel / group) format')],
         params=[('param', 'Convolution')],
         desc='batched convolution on groupped channeled 2D images, but '
         'kernels are not shared across different output positions',
         version=2)

decl_opr('LRN',
         inputs=['src'],
         params='LRN',
         desc='local response normalization')

decl_opr('Pooling',
         inputs=['src'],
         params='Pooling',version=1)

decl_opr('AdaptivePooling',
        inputs=[Doc('src', 'input image, shape (n, c, ih, iw)'),
                Doc('out_shape', 'output image shape, containing two elements specifying output height and width.')],
        params='AdaptivePooling',
        desc='Adaptive Pooling.'
        'The output shape is (n, c, oh, ow), where (oh, ow) is given by *out_shape*.',version=1)

decl_opr('ROIPooling', outputs=[0],
         inputs=[Doc('src', 'input image, shape (n, c, ih, iw)'),
                 Doc('rois', 'regions of interest, shape (m, 5). '
                     'Note that rois[:, 0] denotes the input image index; we '
                     'store it as a float, but it should be an integral value.'
                     ' The rois[:, 1:5] are (x0, y0, x1, y1) for each ROI, '
                     'which would be multiplied by the scale value given in '
                     'param.'),
                 Doc('dest_shape', 'a var to describe output shape, should '
                     'contain exactly two elements')],
         params='ROIPooling',
         desc='ROI pooling, see '
         'https://github.com/rbgirshick/caffe-fast-rcnn. '
         'The output shape is (m, c, oh, ow), where (oh, ow) is given by '
         '*dest_shape*.')


decl_opr('Convolution3D',
         inputs=[Doc('src',
                     'input image in (batch, channel, depth, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(out channel, in channel, kern depth, kern row, kern col) format')],
         params=[('param', 'Convolution3D'),
                 ('execution_polity', 'ExecutionPolicy')],
         desc='batched convolution on channeled 3D images')


decl_opr('Convolution3DBackwardData',
         pyname='deconvolution3d',
         inputs=[Doc('src',
                     'input image in (batch, channel, depth, row, col) format'),
                 Doc('filter',
                     'convolution kernel in '
                     '(out channel, in channel, kern depth, kern row, kern col) format')],
         params=[('param', 'Convolution3D'),
                 ('execution_policy', 'ExecutionPolicy')],
         body=[
             'a, b = all_inputs',
             'all_inputs = [b, a]'
         ],
         desc='batched deconvolution on channeled 3D images; the underlying '
         'computation is in fact gradient of convolution w.r.t. data')

decl_opr('ConvBiasForward',
         pyname='conv_bias_activation_v1',
         inputs=[
             Doc('src', 'input image, allow NCHW, NHWC, NCHW4'),
             Doc('filter', 'filter'),
             Doc('bias', 'bias'),
         ],
         params=[('param', 'ConvBiasV1'),
                 ('execution_policy', 'ExecutionPolicyV0')],
         desc=('activation(convolution(src, filter) + bias) with specified '
               'dtype'),
         has_out_dtype=True)

decl_opr('ConvBiasForward',
         pyname='conv_bias_activation',
         inputs=[
             Doc('src', 'input image, allow NCHW, NHWC, NCHW4'),
             Doc('filter', 'filter'),
             Doc('bias', 'bias'),
         ],
         params=[('param', 'ConvBias'),
                 ('execution_policy', 'ExecutionPolicy')],
         desc=('activation(convolution(src, filter) + bias) with specified '
               'dtype'),
         version=4, has_out_dtype=True)

decl_opr('BatchNorm',
         pyname='batch_norm',
         inputs=['x', 'scale', 'bias', 'running_mean', 'running_variance'],
         desc=('batch normalization similar to cudnn, all params '
               'have the same definition with cudnnBatchNormalization. '
               'It has five outputs: running_mean, running_variance, '
               'save_mean, save_inv_variance, y.'),
         params='BN')

decl_opr('BatchNorm',
         pyname='batch_norm_no_statistic',
         inputs=['x', 'scale', 'bias'],
         desc=('batch noamlization and no need to update mean and variance. '
               'It has three outputs: save_mean, save_inv_variance, y.'),
         params='BN')

decl_opr('LocalShareForward',
         pyname='local_share',
         inputs=[Doc('src', 'input image in (batch, channel, row, col) format'),
                 Doc('filter', 'local share weights in '
                     '(spatial_groups_h, spatial_groups_w, in channel, kern row, kern col, out channel) format')],
         params=[('param', 'LocalShare'),
                 ('execution_policy', 'ExecutionPolicy')],
         desc=Doc(None,
r"""
    Apply a spatial group convolution of input tensor and filter tensor. The output tensor will be split into spatial_groups_hxspatial_groups_w groups. Output locations in the same spatial group share same weights. And weights corresponding to different spatial groups are different.
    Assume input shape is :math:`(N, IC, IH, IW)` and spatial groups in horizontal and vertical directions are :math:`(spatial_groups_h, spatial_groups_w)`,
    then filter shape would be :math:`(spatial_groups_h, spatial_groups_w, IC, FH, FW, OC)` and output shape would be :math:`(N, OC, OH, OW)` where :math:`(OH, OW)` would be computed from padding, stride, :math:`(FH, FW)` and :math:`(IH, IW)`, as in convolution.
    for each output location, we have;

    .. math::

        out_{n, oc, oh, ow} &= \sum_{ic=0}^{IC}\sum_{kh=0}^{FH}\sum_{kw=0}^{FW}src_{n, ic, ih+kh, iw+kw}
                            * filter_{grp_h, grp_w, ic, kh, kw, oc} \\\\
        \\text{where} & ih=-pad_h+oh \\times stride_h \\\\
                      & iw=-pad_w+ow \\times stride_w \\\\
                      & grp_h = oh / (OH / spatial_groups_h) \\\\
                      & grp_w = ow / (OW / spatial_groups_w)
"""), version=1,
         has_out_dtype=True)

decl_opr('ROIAlign', outputs=[0],
         inputs=[Doc('src', 'input image, shape (n, c, ih, iw)'),
                 Doc('rois', 'regions of interest, shape (m, 5). '
                     'Note that rois[:, 0] denotes the input image index; we '
                     'store it as a float, but it should be an integral value.'
                     ' The rois[:, 1:5] are (x0, y0, x1, y1) for each ROI, '
                     'which would be multiplied by the scale value given in '
                     'param.')],
         params='ROIAlign',
         desc='ROI Align, see '
         'Mask-RCNN: https://arxiv.org/pdf/1703.06870.pdf, '
         'The output shape is (m, c, pooled_height, pooled_width), where (pooled_height, pooled_width) is given by '
         '*Param*.',version=1)

decl_opr('DeformableConvForward',
         pyname='deformable_conv',
         inputs=[Doc('im', 'input feature map in (batch, channel, row, col) format'),
                 Doc('filter', 'weights in (output channel, input channel, filter row, filter col) or (group, output channel per group, input channel per group, filter row, filter col) format'),
                 Doc('offset', 'deformable offset in (batch, deformable group * filter row * filter col * 2, output row, output col) format'),
                 Doc('mask', 'deformable mask in (batch, deformable group * filter row * filter col, output row, output col) format')],
         params=[('param', 'Convolution'), ('execution_policy', 'ExecutionPolicy')],
         desc=Doc(None, r""" Apply a deformable convolution to input tensor and filter tensor. The offset tensor will adjust the position of each grid of a convolution filter. The mask tensor will be applied to the deformed input tensor. """),
         has_out_dtype=True, version=1)

decl_opr('DeformablePSROIPoolingForward',
         pyname='deformable_psroi_pooling',
         inputs=[Doc('data', 'input feature map in (batch, channel, row, col) format'),
                 Doc('rois', 'region of interest in (bbox count, 5) format'),
                 Doc('trans', 'bbox position transform parameter in (bbox count, 2, pooled_h, pooled_w) format')],
         params=[('param', 'DeformablePSROIPooling')],
         desc=Doc(None, r""" PSROIPooling with a bbox deformation. """),
         has_out_dtype=True)

decl_opr('BatchConvBiasForward',
         pyname='batch_conv_bias_activation',
         inputs=[Doc('src', 'input image in (batch, channel//4, row, col, 4) format'),
                 Doc('filter', 'weights unshared in batch dimension'
                     '(batch, out_channel, in_channel//4, kern row, kern col, 4) format'),
                 Doc('bias', 'bias'),
         ],
         params=[('param', 'BatchConvBias'),
                 ('execution_policy', 'ExecutionPolicy')],
         desc=Doc(None,
r"""
    Apply a convolution of input tensor and filter tensor whose weights are not shared in batch dimensions. Outputs with batch index use the same weight.
    Assume input shape is :math:`(N, IC, IH, IW)` and filter shape is :math:`(batch, OC, IC, FH, FW)`, the output shape will be :math:`(N, OC, OH, OW)` where :math:`(OH, OW)` would be computed from padding, stride, :math:`(FH, FW)` and :math:`(IH, IW)`, as in convolution.
    for each output location, we have;

    .. math::

        out_{n, oc, oh, ow} &= \sum_{ic=0}^{IC}\sum_{kh=0}^{FH}\sum_{kw=0}^{FW}src_{n, ic, ih+kh, iw+kw}
                            * filter_{n, oc, ic, kh, kw} \\\\
        \\text{where} & ih=-pad_h+oh \\times stride_h \\\\
                      & iw=-pad_w+ow \\times stride_w
"""), version=1,
         has_out_dtype=True)

decl_opr('FakeQuant',
         inputs=[Doc('src','input tenosr'),Doc('scale','scale tensor'),Doc('zero_point','zero point tensor')],
         params='FakeQuant')

decl_opr('TQT',
         inputs=[Doc('src','input tensor'),Doc('scale','scale tensor')],
         params='TQT')
decl_opr('LSQ',
         inputs=[Doc('src','input tensor'),Doc('scale','scale tensor'),Doc('zero_point','zero point tensor'),Doc('grad_scale','grad scale tensor')],
         params='LSQ')
# vim: ft=python