OpenI
/
MegEngine
Not watched
1
0
Fork 0
Code
Releases 31 Wiki Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
Tree: dbd9483993
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'dbd9483993'
${ noResults }
MegEngine/dnn/scripts/opr_param_defs.py

1304 lines
61 kB
Raw Blame History 

  1. pdef('Empty')

  2. 

  3. pdef('Axis').add_fields('int32', 'axis', 0)

  4. 

  5. (pdef('Convolution', version=0, is_legacy=True).

  6.  add_enum('Mode', 'CROSS_CORRELATION = 0', 'CONVOLUTION = 1').

  7.  add_fields(

  8.      'uint32',

  9.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  10.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  11.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  12.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  13.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  14.          'on the second dimension'), 1,

  15.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  16.          'on the second dimension'), 1

  17.  ).

  18.  add_enum('DataType',

  19.           Doc('FLOAT = 0', 'input/output both float32/float16'),

  20.           'INT8x8x16 = 1',

  21.           'INT8x8x32 = 2',

  22.           Doc('FLOAT_IO16xC32 = 3', 'input/output both float16, the internal '

  23.               'compute is float32'),

  24.           Doc('QUINT8x8x32 = 4', 'input QuantizedAsymm8, output QuantizedS32'),

  25.           Doc('INT8x8xX = 5', 'input int8, output specified by tensor DType'),

  26.           Doc('QUINT4x4x32 = 6', 'input QuantizedAsymm4, output QuantizedS32'),

  27.           name_field='data_type').

  28.  add_enum('Sparse',

  29.           Doc('DENSE = 0', 'dense convolution: filter shape should be '

  30.               '[oc, ic, spatial...] if format is NCHW, '

  31.               '[oc, spatial..., ic] if format is NHWC'),

  32.           Doc('GROUP = 1', 'group convolution: filter shape should be '

  33.               '[group, oc_per_group, ic_per_group, spatial...] if format is NCHW, '

  34.               '[group, oc_per_group, spatial..., ic_per_group] if format is NHWC')

  35.           ).

  36.  add_enum(Doc('Format', 'convolution data/filter/output format; see '

  37.               ':class:`RelayoutFormat` for more details'),

  38.           'NCHW = 0', 'NHWC = 1', 'NHWCD4 = 2', 'NCHW4 = 3', 'NCHW8 = 4', 'NCHW32 = 5', 'NCHW88 = 6',

  39.           'NCHW44 = 7','NCHW44_DOT = 8',

  40.           Doc('NCHW_WINOGRAD = 9', 'NCHW layout with weights tranformed by winograd'),

  41.           Doc('NCHW88_WINOGRAD = 10', 'NCHW88 layout with weights tranformed by winograd'),

  42.           Doc('NCHW44_WINOGRAD = 11', 'NCHW44 layout with weights tranformed by winograd'),

  43.           Doc('NCHW4_NCHW32 = 12', 'NCHW4_NCHW32 means input tensors are nchw4 layout, output tensor is nchw32 layout'),

  44.           Doc('NCHW32_NCHW4 = 13', 'NCHW32_NCHW4 means input tensors are nchw32 layout, output tensor is nchw4 layout'),

  45.           Doc('NCHW4_NCHW = 14', 'NCHW4_NCHW means input tensors are nchw4 layout, output tensor is nchw layout'),

  46.           Doc('NHWC_NCHW = 15', 'NHWC_NCHW means input tensors are nhwc layout, '

  47.               'output tensor is nchw layout'),

  48.           Doc('NHWC_NCHW4_IC_SMALL = 16', 'NHWC_NCHW4_IC_SMALL means input tensors are nhwc(c < 4) layout, '

  49.               'output tensor is nchw4 layout, padding c=4'),

  50.           Doc('NCHW_NCHW4_IC_SMALL = 17', 'NCHW_NCHW4_IC_SMALL means input tensors are nchw(c < 4) layout, '

  51.               'output tensor is nchw4 layout, padding c=4'),

  52.           Doc('CHWN4 = 18', 'CHWN4 is currently only used on Nvidia platform for fast implementation '

  53.               'of convolution using CUDA/SASS. The channels are splitted to groups of 4 channels.'),

  54.           Doc('NCHW4_NHWC = 19', 'NCHW4_NHWC means input tensors are nchw4 layout, output tensor is nhwc layout'))

  55.  )

  56. 

  57. (pdef('Convolution', version=1, is_legacy=True).

  58.  add_enum_alias('Mode', 'ConvolutionV0').

  59.  add_fields(

  60.      'uint32',

  61.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  62.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  63.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  64.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  65.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  66.          'on the second dimension'), 1,

  67.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  68.          'on the second dimension'), 1

  69.  ).

  70.  add_enum_alias('Sparse', 'ConvolutionV0').

  71.  add_enum_alias('Format', 'ConvolutionV0').

  72.  add_enum(Doc('ComputeMode', 'Specifies special computation modes, e.g. '

  73.                              'different combinations of intermediate result '

  74.                              'data types.'),

  75.           Doc('DEFAULT = 0', 'No special requirements on the precision of '

  76.                          'intermediate results.'),

  77.           Doc('FLOAT32 = 1', 'Use Float32 accumulator and intermediate result. '

  78.                          'Only supported when input and output is Float16.'),

  79.           name_field='compute_mode')

  80.  )

  81. 

  82. (pdef('Convolution', version=2).

  83.  add_enum_alias('Mode', 'ConvolutionV0').

  84.  add_fields(

  85.      'uint32',

  86.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  87.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  88.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  89.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  90.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  91.          'on the second dimension'), 1,

  92.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  93.          'on the second dimension'), 1

  94.  ).

  95.  add_enum_alias('Sparse', 'ConvolutionV0').

  96.  add_enum(Doc('Format', 'convolution data/filter/output format; see '

  97.               ':class:`RelayoutFormat` for more details'),

  98.           'NCHW = 0', 'NHWC = 1', 'NHWCD4 = 2', 'NCHW4 = 3', 'NCHW8 = 4', 'NCHW32 = 5', 'NCHW88 = 6',

  99.           'NCHW44 = 7','NCHW44_DOT = 8',

  100.           Doc('NCHW4_NCHW32 = 9', 'NCHW4_NCHW32 means input tensors are nchw4 layout, output tensor is nchw32 layout'),

  101.           Doc('NCHW32_NCHW4 = 10', 'NCHW32_NCHW4 means input tensors are nchw32 layout, output tensor is nchw4 layout'),

  102.           Doc('NCHW4_NCHW = 11', 'NCHW4_NCHW means input tensors are nchw4 layout, output tensor is nchw layout'),

  103.           Doc('NHWC_NCHW = 12', 'NHWC_NCHW means input tensors are nhwc layout, '

  104.               'output tensor is nchw layout'),

  105.           Doc('NHWC_NCHW4_IC_SMALL = 13', 'NHWC_NCHW4_IC_SMALL means input tensors are nhwc(c < 4) layout, '

  106.               'output tensor is nchw4 layout, padding c=4'),

  107.           Doc('NCHW_NCHW4_IC_SMALL = 14', 'NCHW_NCHW4_IC_SMALL means input tensors are nchw(c < 4) layout, '

  108.               'output tensor is nchw4 layout, padding c=4'),

  109.           Doc('CHWN4 = 15', 'CHWN4 is currently only used on Nvidia platform for fast implementation '

  110.               'of convolution using CUDA/SASS. The channels are splitted to groups of 4 channels.'),

  111.           Doc('NCHW64 = 16', 'NCHW64 is designed for convolution implementation to utilizing TensorCore '

  112.               'instructions for 4-bit integers on Nvidia platforms'),

  113.           Doc('NCHW4_NHWC = 17', 'NCHW4_NHWC means input tensors are nchw4 layout, output tensor is nhwc layout')).

  114.  add_enum_alias('ComputeMode', 'ConvolutionV1',name_field='compute_mode')

  115.  )

  116. 

  117. 

  118. (pdef('MaskPropagate').

  119.  add_fields(

  120.      'uint32',

  121.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  122.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  123.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  124.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  125.      Doc('kernel_h', 'kernel height'), 1,

  126.      Doc('kernel_w', 'kernel width'), 1,

  127.      Doc('dilate_h', 'dilate height'), 1,

  128.      Doc('dilate_w', 'dilate width'), 1)

  129.  )

  130. 

  131. (pdef('ConvPooling').

  132.  add_enum('Method', 'WITH_TEXTURE_OBJ = 0', 'WITH_SHARED_MEM = 1').

  133.  add_enum_alias('ConvMode', 'ConvolutionV0', 'Mode').

  134.  add_enum('PoolMode', 'AVERAGE = 0', 'MAX = 1').

  135.  add_enum('NonlineMode', 'IDENTITY = 0', 'RELU = 1', 'SIGMOID = 2').

  136.  add_fields('uint32', 'pool_shape_h', 1, 'pool_shape_w', 1, 'pool_stride_h', 1, 'pool_stride_w', 1, \

  137.   'pool_pad_h', 0, 'pool_pad_w', 0, 'conv_stride_h', 1, 'conv_stride_w', 1, 'conv_pad_h', 0, 'conv_pad_w', 0))

  138. 

  139. (pdef('ConvBias', 'legacy conv_bias', version=0, is_legacy=True).

  140.  add_enum('NonlineMode', 'IDENTITY = 0', 'RELU = 1', 'SIGMOID = 2', 'H_SWISH = 3').

  141.  add_enum_alias('Mode', 'ConvolutionV0').

  142.  add_fields('uint32', 'pad_h', 0, 'pad_w', 0, 'stride_h', 1, 'stride_w', 1))

  143. 

  144. (pdef('ConvBias', 'active(conv(x, w) + bias)', version=1, is_legacy=True).

  145.  add_enum_alias('NonlineMode', 'ConvBiasV0').

  146.  add_enum_alias('Mode', 'ConvolutionV0').

  147.  add_enum_alias('DataType', 'ConvolutionV0', name_field='data_type').

  148.  add_enum_alias('Sparse', 'ConvolutionV0').

  149.  add_enum_alias('Format', 'ConvolutionV0').

  150.  add_fields(

  151.      'uint32',

  152.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  153.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  154.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  155.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  156.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  157.          'on the second dimension'), 1,

  158.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  159.          'on the second dimension'), 1)

  160.  )

  161. 

  162. (pdef('ConvBias', 'active(conv(x, w) + bias)', version=2, is_legacy=True).

  163.  add_enum_alias('NonlineMode', 'ConvBiasV0').

  164.  add_enum_alias('Mode', 'ConvolutionV0').

  165.  add_enum_alias('Sparse', 'ConvolutionV0').

  166.  add_enum_alias('Format', 'ConvolutionV0').

  167.  add_fields(

  168.      'uint32',

  169.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  170.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  171.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  172.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  173.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  174.          'on the second dimension'), 1,

  175.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  176.          'on the second dimension'), 1).

  177.  add_enum_alias('ComputeMode', 'ConvolutionV1', name_field='compute_mode')

  178.  )

  179. 

  180. (pdef('ConvBias', 'active(conv(x, w) + bias)', version=3, is_legacy=True).

  181.  add_enum_alias('NonlineMode', 'ConvBiasV0').

  182.  add_enum_alias('Mode', 'ConvolutionV0').

  183.  add_enum_alias('Sparse', 'ConvolutionV0').

  184.  add_enum_alias('Format', 'ConvolutionV0').

  185.  add_fields(

  186.      'uint32',

  187.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  188.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  189.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  190.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  191.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  192.          'on the second dimension'), 1,

  193.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  194.          'on the second dimension'), 1,

  195.      Doc('output_block_size', 'detail meaning \see winograd in conv bias'), 0).

  196.  add_enum_alias('ComputeMode', 'ConvolutionV1', name_field='compute_mode')

  197.  )

  198. 

  199. (pdef('ConvBias', 'active(conv(x, w) + bias)', version=4).

  200.  add_enum_alias('NonlineMode', 'ConvBiasV0').

  201.  add_enum_alias('Mode', 'ConvolutionV0').

  202.  add_enum_alias('Sparse', 'ConvolutionV0').

  203.  add_enum_alias('Format', 'Convolution').

  204.  add_fields(

  205.      'uint32',

  206.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  207.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  208.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  209.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  210.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  211.          'on the second dimension'), 1,

  212.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  213.          'on the second dimension'), 1).

  214.  add_enum_alias('ComputeMode', 'ConvolutionV1', name_field='compute_mode')

  215.  )

  216. (pdef('SeparableConv').

  217.  add_enum_alias('Mode', 'ConvolutionV0').

  218.  add_enum('BorderMode', 'BORDER_REPLICATE = 0', 'BORDER_REFLECT = 1',

  219.           'BORDER_REFLECT_101 = 2','BORDER_WRAP = 3',

  220.           'BORDER_CONSTANT = 4', 'BORDER_TRANSPARENT = 5','BORDER_ISOLATED = 6').

  221.  add_fields('bool', 'is_symm_kernel', 'true').

  222.  add_fields('uint32', 'pad_h', 0, 'pad_w', 0, 'stride_h', 1, 'stride_w', 1,

  223.             'ksize_h', 3, 'ksize_w', 3, 'anchor_h', 1, 'anchor_w', 1))

  224. 

  225. (pdef('Images2Neibs').

  226.  add_fields('uint32', 'pad_h', 0, 'pad_w', 0, 'stride_h', 1, 'stride_w', 1,

  227.             'dilate_h', 1, 'dilate_w', 1, 'window_h', 3, 'window_w', 3))

  228. 

  229. (pdef('SlidingWindowTranspose').

  230.  add_fields('uint32', 'out_h', 0, 'out_w', 0, 'pad_h', 0, 'pad_w', 0, 'stride_h', 1, 'stride_w', 1,

  231.             'dilate_h', 1, 'dilate_w', 1, 'window_h', 3, 'window_w', 3))

  232. 

  233. (pdef('Pooling', version=0, is_legacy=True).

  234.  add_enum(

  235.      'Mode',

  236.      Doc('MAX = 0', 'maximum value inside pooling window'),

  237.      Doc('AVERAGE = 1',

  238.          'arithmetic mean of all values inside pooling window. Padding values '

  239.          'are taken into account and are viewed as zero'),

  240.      Doc('AVERAGE_COUNT_EXCLUDE_PADDING = 2',

  241.          'arithmetic mean of all values inside pooling window. No padding is'

  242.          'used.')

  243.  ).

  244.  add_fields('uint32', 'pad_h', 0, 'pad_w', 0, 'stride_h', 2, 'stride_w', 2,

  245.             'window_h', 2, 'window_w', 2).

  246.  add_enum_alias('Format', 'ConvolutionV0')

  247.  )

  248. 

  249. (pdef('Pooling', version=1).

  250.  add_enum_alias('Mode','PoolingV0').

  251.  add_fields('uint32', 'pad_h', 0, 'pad_w', 0, 'stride_h', 2, 'stride_w', 2,

  252.             'window_h', 2, 'window_w', 2).

  253.  add_enum_alias('Format', 'Convolution')

  254.  )

  255. 

  256. (pdef('Softmax').

  257.  add_fields('int32', 'axis', -1)

  258. )

  259. 

  260. (pdef('AdaptivePooling', version=0, is_legacy=True).

  261.  add_enum_alias('Mode', 'PoolingV0').

  262.  add_enum_alias('Format', 'ConvolutionV0')

  263.  )

  264. 

  265. (pdef('AdaptivePooling', version=1).

  266.  add_enum_alias('Mode', 'PoolingV0').

  267.  add_enum_alias('Format', 'Convolution')

  268.  )

  269. 

  270. (pdef('LRN',

  271.       'see ImageNet Classification with Deep Convolutional Neural Networks for'

  272.       ' meaning of the fields').

  273.  add_fields('uint32', Doc('n', 'must be odd'), 5).

  274.  add_fields('float32', 'k', '2.f', 'alpha', '1e-4f', 'beta', '0.75f')

  275. )

  276. 

  277. (pdef('BN').

  278.  add_enum(

  279.      'ParamDim',

  280.      Doc('DIM_11HW = 0', 'Dim of params (Sigma, Mu) is 1 x 1 x H x W'),

  281.      Doc('DIM_1CHW = 1', 'Dim of params (Sigma, Mu) is 1 x C x H x W'),

  282.      Doc('DIM_1C11 = 2', 'Dim of params (Sigma, Mu) is 1 x C x 1 x 1'),

  283.      Doc('DIM_111C = 3', 'Dim of params (Sigma, Mu) is 1 x 1 x 1 x C'),

  284.      name_field='param_dim'

  285.  ).

  286.  add_enum(

  287.      'FwdMode',

  288.      Doc('TRAINING = 0', 'Training phase.'),

  289.      Doc('INFERENCE = 1', 'Inference phase.'),

  290.      name_field='fwd_mode'

  291.  ).

  292.  add_fields('float64', 'epsilon', '1e-4f').

  293.  add_fields('float64', 'avg_factor', '1.f').

  294.  add_fields('float32', 'scale', '1.f').

  295.  add_fields('float32', 'bias', '0.f')

  296. )

  297. 

  298. (pdef('ROIPooling').

  299.  add_enum(

  300.      'Mode',

  301.      Doc('MAX = 0', 'maximum value inside pooling window; pooling result would '

  302.          'be 0 if pooling window is empty'),

  303.      Doc('AVERAGE = 1',

  304.          'arithmetic mean of all values inside pooling window; pooling result '

  305.          'would be 0 if pooling window is empty')

  306.  ).

  307.  add_fields('float32', 'scale', '1.f'))

  308. 

  309. INTERP_MODES = ['NEAREST = 0', 'LINEAR = 1', 'AREA = 2', 'CUBIC = 3', 'LANCZOS4 = 4']

  310. BORDER_MODES = [Doc('REPLICATE = 0', 'aaaaaa|abcdefgh|hhhhhhh'),

  311.                 Doc('REFLECT = 1', 'fedcba|abcdefgh|hgfedcb'),

  312.                 Doc('REFLECT_101 = 2', 'gfedcb|abcdefgh|gfedcba'),

  313.                 Doc('WRAP = 3', 'cdefgh|abcdefgh|abcdefg'),

  314.                 Doc('CONSTANT = 4', 'iiiiii|abcdefgh|iiiiiii'),

  315.                 Doc('TRANSPARENT = 5', ''),

  316.                 Doc('ISOLATED = 6', '')]

  317. (pdef('WarpPerspective', version=1, is_legacy=True).

  318.  add_enum('InterpolationMode', *INTERP_MODES,

  319.           name_field='imode', default=1,

  320.           member_alias=[(i, 'INTER_{}'.format(i)) for i in INTERP_MODES]

  321.           ).

  322.  add_enum('BorderMode', *BORDER_MODES,

  323.           name_field='bmode',

  324.           member_alias=[(i, 'BORDER_{}'.format(i)) for i in BORDER_MODES]

  325.           ).

  326.  add_enum_alias('Format', 'ConvolutionV0').

  327.  add_fields('float32', Doc('border_val', 'used for CONSTANT bmode'), '.0f'))

  328. 

  329. (pdef('WarpPerspective', version=2).

  330.  add_enum_alias('InterpolationMode','WarpPerspectiveV1',name_field="imode").

  331.  add_enum_alias('BorderMode','WarpPerspectiveV1',name_field="bmode").

  332.  add_enum_alias('Format', 'Convolution').

  333.  add_fields('float32', Doc('border_val', 'used for CONSTANT bmode'), '.0f'))

  334. 

  335. 

  336. pdef('SpatialTfGridGenerator').add_enum('Mode', 'AFFINE = 0')

  337. pdef('SpatialTfSampler').add_enum('Mode', 'BILINEAR = 0')

  338. 

  339. pdef('AddUpdate').add_fields(

  340.     'float32', 'alpha', '1.f', 'beta', '1.f', 'bias', '0.f')

  341. 

  342. pdef('Elemwise').add_enum(

  343.     'Mode',

  344.     Doc('RELU = 0', 'unary: max(x, 0)'),

  345.     Doc('ABS = 1', 'unary: abs(x)'),

  346.     Doc('ACOS = 2', 'unary: acos(x)'),

  347.     Doc('ASIN = 3', 'unary: asin(x)'),

  348.     Doc('CEIL = 4', 'unary: ceil(x)'),

  349.     Doc('COS = 5', 'unary: cos(x)'),

  350.     Doc('EXP = 6', 'unary: exp(x)'),

  351.     Doc('EXPM1 = 7', 'unary: numerically stable exp(x)-1'),

  352.     Doc('FLOOR = 8', 'unary: floor(x)'),

  353.     Doc('LOG = 9', 'unary: natural logarithm, log(x)'),

  354.     Doc('LOG1P = 10', 'unary: numerically stable log(x+1)'),

  355.     Doc('NEGATE = 11', 'unary: -x'),

  356.     Doc('SIGMOID = 12', 'unary: 1/(1+exp(-x))'),

  357.     Doc('SIN = 13', 'unary: sin(x)'),

  358.     Doc('TANH = 14', 'unary: tanh(x)'),

  359. 

  360.     Doc('ABS_GRAD = 15', 'binary: x > 0 ? y : -y'),

  361.     Doc('ADD = 16', 'binary: x + y'),

  362.     Doc('FLOOR_DIV = 17', 'binary: floor(x / y)'),

  363.     Doc('MAX = 18', 'binary: max(x, y)'),

  364.     Doc('MIN = 19', 'binary: min(x, y)'),

  365.     Doc('MOD = 20', 'binary: x % y or fmodf(x, y)'),

  366.     Doc('MUL = 21', 'binary: x * y'),

  367.     Doc('POW = 22', 'binary: pow(x, y)'),

  368.     Doc('SIGMOID_GRAD = 23', 'binary: x * (1 - x) * y'),

  369.     Doc('SUB = 24', 'binary: x - y'),

  370.     Doc('SWITCH_GT0 = 25', 'binary: (x > 0) * y'),

  371.     Doc('TANH_GRAD = 26', 'binary: (1 - x * x) * y'),

  372.     Doc('TRUE_DIV = 27', 'binary: x / y'),

  373.     Doc('LOG_SUM_EXP = 28', 'binary: numerically stable log(exp(x) + exp(y))'),

  374. 

  375.     Doc('LT = 29', 'binary: x < y'),

  376.     Doc('LEQ = 30', 'binary: x <= y'),

  377.     Doc('EQ = 31', 'binary: x == y'),

  378. 

  379.     Doc('SHL = 32', 'bitwise binary: x << y. '

  380.         'Note that result is undefined if y < 0 or y >= bitwidth. Logical '

  381.         'shift is performed for unsigned intergers, and arithmetic shift for '

  382.         'signed ones.'),

  383.     Doc('SHR = 33', 'bitwise binary: x >> y; see SHL mode for more details'),

  384. 

  385.     Doc('COND_LEQ_MOV = 34', 'ternary: x <= y ? z : 0'),

  386.     Doc('FUSE_MUL_ADD3 = 35',

  387.         'compute ``a * b + c`` where c must either have same layout as '

  388.         'a or b, or be a scalar'),

  389. 

  390.     Doc('FUSE_MUL_ADD4 = 36',

  391.         'compute ``a * A + b * B`` where a and b must have equal layout, '

  392.         'and A and B must have equal layout. In the inputs ``b`` and ``B`` '

  393.         'can be swapped'),

  394. 

  395.     Doc('FUSE_ADD_RELU = 37', 'binary: max(x+y, 0)'),

  396.     Doc('FUSE_ADD_SIGMOID = 38', 'binary: 1/(1+exp(-(x+y)))'),

  397.     Doc('FUSE_ADD_TANH = 39', 'binary: tanh(x+y)'),

  398.     Doc('FAST_TANH = 40', 'unary: rational approximation of tanh(x)'),

  399.     Doc('FAST_TANH_GRAD = 41', 'binary: grad of the rational approximation of tanh(x)'),

  400. 

  401.     Doc('ROUND = 42', 'unary: round(x), the nearest integer value to x, rounding '

  402.                  'halfway cases away from zero. Float only.'),

  403.     Doc('RMULH = 43', 'binary: rounded higher l bits of x * y, where l is the bit '

  404.                 'length of x.'),

  405. 

  406.     Doc('ATAN2 = 44','binary: atan2(y,x)'),

  407.     Doc('ERF = 45', 'unary: erf(x)'),

  408.     Doc('ERFINV = 46', 'unary: inverse function of erf(x)'),

  409.     Doc('ERFC = 47', 'unary: erfc(x)'),

  410.     Doc('ERFCINV = 48', 'unary: inverse function of erfc(x)'),

  411.     Doc('H_SWISH = 49', 'unary: x * clip(x + 3, 0, 6) / 6'),

  412.     Doc('H_SWISH_GRAD = 50', 'binary: x < -3 ? 0 : (x > 3 ? y : (2 * x + 3) / 6 * y)'),

  413.     Doc('FUSE_ADD_H_SWISH = 51', 'binary: hswish(x+y)'),

  414. 

  415.     Doc('NOT = 52', 'unary: !x'),

  416.     Doc('AND = 53', 'binary: x && y'),

  417.     Doc('OR = 54', 'binary: x || y'),

  418.     Doc('XOR = 55', 'binary: x ^ y'),

  419.     Doc('SILU = 56', 'unary: x / (1 + exp(-x))'),

  420.     Doc('SILU_GRAD = 57', 'binary: grad(x / (1 + exp(-x))'),

  421.     Doc('GELU = 58', 'unary: x Phi(x)'),

  422.     Doc('GELU_GRAD = 59', 'binary: grad(x Phi(x))'),

  423.     Doc('COND_LT_MOV = 60', 'ternary: x < y ? z : 0'),

  424.     Doc('NEQ = 61', 'binary: x != y'),

  425.     Doc('ISNAN = 62', 'unary: isnan(x)'),

  426.     Doc('ISINF = 63', 'unary: isinf(x)'),

  427. )

  428. 

  429. pdef('ElemwiseMultiType').add_enum(

  430.     'Mode',

  431.     Doc('FUSE_MUL_ADD3_INT16x32x32x32 = 0',

  432.         'compute ``a * b + c`` requiring that ``a`` be int16 and ``b`` and '

  433.         '``c``  int32, and the result is int32. This mode is optimized for '

  434.         'the channel-broadacsted case, i.e. ``a`` has shape (A, B, C) and '

  435.         '``b`` and ``c`` have shape (1, C, 1)'),

  436.     Doc('FUSE_MUL_ADD3_IXxF32xF32xI8 = 1',

  437.         'compuate ``a * b + c`` where the inputs ``a`` is an integer type '

  438.         '``b`` and ``c`` are both ``float32``, the result is '

  439.         '``int8``. This is currently only optimized for ``(1, x)`` '

  440.         'broadcast for ``b`` and ``c``. Computation is carried in floating '

  441.         'points and results are rounded towards zero with saturated cast to '

  442.         'int.'),

  443.     Doc('ROUND_SHR_SATURATE_IXxI8xI8 = 2',

  444.         'Compute ``a >> b``, round the result according to lower ``b`` bits '

  445.         'of ``a``` and make a saturating conversion to int8. Where ``a`` should'

  446.         ' be an integer tensor and ``b`` should be an int8 scalar.'),

  447.     Doc('FUSE_ADD_RMULH_ROUND_SHR_SATURATE_INT16x16x16x8 = 3',

  448.         'Fused operation of an int16 elemwise add, an int16 rounding multiply '

  449.         'high and an int16 to int8 rounding right shift with saturation.'),

  450.     Doc('FUSE_ADD_RMULH_ROUND_SHR_SATURATE_INT32x32x32x8 = 4',

  451.         'Fused operation of an int32 elemwise add, an int32 rounding multiply '

  452.         'high and an int32 to int8 rounding right shift with saturation.'),

  453.     Doc('ROUND_SHR_SATURATE_IXxI8xI16 = 5',

  454.         'Compute ``a >> b``, round the result according to lower ``b`` bits of '

  455.         '``a``` and make a saturating conversion to int16. Where ``a`` should'

  456.         ' be an integer tensor and ``b`` should be an int8 scalar.'),

  457.     Doc('QADD = 6', 'Fused elemwise add two quantized int8 with specified'

  458.         'output quantized dtype'),

  459.     Doc('QFUSE_ADD_RELU = 7', 'Fused elemwise add two quantized int8 followed'

  460.          ' by ReLU and typecvt to specified dtype'),

  461.     Doc('QMUL = 8', 'Fused elemwise multiply two quantized int8 with specified'

  462.         'output quantized dtype'),

  463.     Doc('QMIN = 9', 'Fused elemwise min two quantized int8 with specified'

  464.         'output quantized dtype'),

  465.     Doc('QMAX = 10', 'quantized: max(x, y), with specified output quantized dtype'),

  466.     Doc('QSUB = 11', 'quantized: x - y'),

  467.     Doc('QTRUE_DIV = 12', 'quantized: x / y'),

  468.     Doc('QFUSE_ADD_SIGMOID = 13', 'quantized: sigmoid(x + y)'),

  469.     Doc('QFUSE_ADD_TANH = 14', 'quantized: tanh(x + y)'),

  470.     Doc('QRELU = 15', 'quantized: x > 0 ? x : 0'),

  471.     Doc('QABS = 16', 'quantized: x > 0 ? x : -x'),

  472.     Doc('QSIGMOID = 17', 'quantized: sigmoid(x)'),

  473.     Doc('QEXP = 18', 'quantized: exp(x)'),

  474.     Doc('QTANH = 19', 'quantized: tanh(x)'),

  475.     Doc('QFUSE_MUL_ADD3 = 20', 'quantized: x * y + z'),

  476.     Doc('QFAST_TANH = 21', 'quantized: fast_tanh(x)'),

  477.     Doc('QNEGATE = 22', 'quantized: -x'),

  478.     Doc('QACOS = 23', 'quantized: acos(x)'),

  479.     Doc('QASIN = 24', 'quantized: asin(x)'),

  480.     Doc('QCEIL = 25', 'quantized: ceil(x)'),

  481.     Doc('QCOS = 26', 'quantized: cos(x)'),

  482.     Doc('QEXPM1 = 27', 'quantized: expm1(x)'),

  483.     Doc('QFLOOR = 28', 'quantized: floor(x)'),

  484.     Doc('QLOG = 29', 'quantized: log(x)'),

  485.     Doc('QLOG1P = 30', 'quantized: log1p(x)'),

  486.     Doc('QSIN = 31', 'quantized: sin(x)'),

  487.     Doc('QROUND = 32', 'quantized: round(x)'),

  488.     Doc('QERF = 33', 'quantized: erf(x)'),

  489.     Doc('QERFINV = 34', 'quantized: erfinv(x)'),

  490.     Doc('QERFC = 35', 'quantized: erfc(x)'),

  491.     Doc('QERFCINV = 36', 'quantized: erfcinv(x)'),

  492.     Doc('QABS_GRAD = 37', 'quantized: abs_grad'),

  493.     Doc('QFLOOR_DIV = 38', 'quantized floor_div'),

  494.     Doc('QMOD = 39', 'quantized mod'),

  495.     Doc('QSIGMOID_GRAD = 40', 'quantized sigmoid_grad'),

  496.     Doc('QSWITCH_GT0 = 41', 'quantized switch_gt0'),

  497.     Doc('QTANH_GRAD = 42', 'quantized tanh_grad'),

  498.     Doc('QLT = 43', 'quantized lt'),

  499.     Doc('QLEQ = 44', 'quantized leq'),

  500.     Doc('QEQ = 45', 'quantized eq'),

  501.     Doc('QPOW = 46', 'quantized pow'),

  502.     Doc('QLOG_SUM_EXP = 47', 'quantized log_sum_exp'),

  503.     Doc('QFAST_TANH_GRAD = 48', 'quantized fast_tanh_grad'),

  504.     Doc('QATAN2 = 49', 'quantized atan2'),

  505.     Doc('QCOND_LEQ_MOV = 50', 'quantized cond_leq_mov'),

  506.     Doc('QH_SWISH = 51', 'quantized h_swish'),

  507.     Doc('QFUSE_ADD_H_SWISH = 52', 'quantized h_swish(x+y)'),

  508.     Doc('QH_SWISH_GRAD = 53', 'quantized h_swish_grad'),

  509.     Doc('FUSE_MUL_ADD3_INT16xF32xF32xF32 = 54',

  510.         'compute ``a * b + c`` requiring that ``a`` be int16 and ``b`` and '

  511.         '``c``  float32, and the result is float32.'),

  512.     Doc('MUL_INT16xF32xF32 = 55',

  513.         'compute ``a * b `` requiring that ``a`` be int16 and ``b`` float32, '

  514.         'and the result is float32.'),

  515.     Doc('FUSE_MUL_ADD3_UINT8xF32xF32xF32 = 56',

  516.         'compute ``a * b + c`` requiring that ``a`` be uint8 and ``b`` and '

  517.         '``c``  float32, and the result is float32.'),

  518.     Doc('QCOND_LT_MOV = 57', 'quantized cond_lt_mov'),

  519.     Doc('EQ = 58', 'eq'),

  520.     Doc('NEQ = 59', 'eq'),

  521.     Doc('LT = 60', 'lt'),

  522.     Doc('LEQ = 61', 'leq'),

  523.     Doc('ISNAN = 62', 'isnan'),

  524.     Doc('ISINF = 63', 'isinf')

  525. )

  526. 

  527. pdef('PowC', 'power with constant exponent').add_fields('float32', 'exp', 0)

  528. 

  529. (pdef('DctChannelSelect', '2d discrete cosine transform', version=0, is_legacy=True).add_enum_alias('Format', 'ConvolutionV0').

  530.  add_enum('FastImpl', 'NONE = 0', 'FIX_32_MASK = 1').add_fields('int32', 'dct_block_size', 8))

  531. 

  532. (pdef('DctChannelSelect', '2d discrete cosine transform', version=1).add_enum_alias('Format', 'Convolution').

  533.  add_enum_alias('FastImpl', 'DctChannelSelectV0').add_fields('int32', 'dct_block_size', 8))

  534. 

  535. (pdef('MatrixMul', version=0, is_legacy=True).

  536.  add_fields('bool', 'transposeA', 'false', 'transposeB', 'false').

  537.  add_enum('DataType',

  538.      Doc('FLOAT = 0', 'input/output both float32/float16'),

  539.      'INT8x8x16 = 1',

  540.      'INT8x8x32 = 2',

  541.      Doc('FLOAT_IO16xC32 = 3', 'input/output both float16, the internal compute is '

  542.          'float32'),

  543.      Doc('QUINT8x8x32 = 4', 'input QuantizedAsymm8, output QuantizedS32'),

  544.      Doc('QUINT4x4x32 = 5', 'input QuantizedAsymm4, output QuantizedS32'),

  545.      name_field='data_type'))

  546. 

  547. (pdef('MatrixMul', version=1, is_legacy=True).

  548.  add_fields('bool', 'transposeA', 'false', 'transposeB', 'false').

  549.  add_enum(Doc('ComputeMode', 'Specifies special computation modes, e.g. '

  550.                              'different combinations of intermediate result '

  551.                              'data types.'),

  552.           Doc('DEFAULT = 0', 'No special requirements on the precision of '

  553.                          'intermediate results.'),

  554.           Doc('FLOAT32 = 1', 'Use Float32 accumulator and intermediate result. '

  555.                          'Only supported when input and output is Float16.'),

  556.           name_field='compute_mode'))

  557. 

  558. (pdef('MatrixMul', version=2).

  559.  add_fields('bool', 'transposeA', 'false', 'transposeB', 'false').

  560.  add_enum_alias('ComputeMode', 'MatrixMulV1', name_field='compute_mode').

  561.  add_enum('Format',

  562.           Doc('DEFAULT = 0', 'Normal matrix mul: (M, K) x (K, N) = (M, N)'),

  563.           Doc('MK4 = 1', 'Split 4 from M and K, better for neon compute:'

  564.               '(M/4, K/4, 4(k), 4(m)) x (K/4, N, 4(k)). if transposeA the '

  565.               'layout is (K/4, M/4, 4(k), 4(m)) x (K/4, N, 4(k))'),

  566.           Doc('MK8 = 2', 'Split 8 from M and K, better for neon compute:'

  567.               '(M/8, K/8, 8(k), 8(m)) x (K/8, N, 8(k)). if transposeA the '

  568.               'layout is (K/8, M/8, 8(k), 8(m)) x (K/8, N, 8(k))'),

  569.           Doc('MK4_DOT = 3', 'Split 4 from M and K, better for neon dotprod:'

  570.               'M/4, K/4, 4(m), 4(k)) x (K/4, N, 4(k)). if transposeA the '

  571.               'layout is (K/4, M/4, 4(m), 4(k)) x (K/4, N, 4(k))'),

  572.           Doc('N32K4_DOT = 4', 'Split 32 from N and 4 from K, better for neon gevm dotprod:'

  573.               'N/32, K/4, 32(n), 4(k)')

  574.               )

  575.  )

  576. 

  577. (pdef('SVD').

  578.  add_fields('bool',

  579.             Doc('full_matrices',

  580.                 'Whether to compute the full-sized u and v or only the leading'

  581.                 ' min(m, n) singular vectors. Ignored if compute_uv is '

  582.                 'false.'),

  583.             'false',

  584.             Doc('compute_uv',

  585.                 'Whether the left (u) and right (v) singular vectors will be '

  586.                 'computed and outputted.'),

  587.             'true'))

  588. 

  589. (pdef('Reduce', 'legacy reduce', version=0, is_legacy=True).

  590.  add_enum('Mode',

  591.           'SUM = 0',

  592.           Doc('SUM_SQR = 1', 'sum of x * x for each element x'),

  593.           'PRODUCT = 2', 'MIN = 3', 'MAX = 4').

  594.  add_fields('int32',

  595.             Doc('axis',

  596.                 'axis along which reduction is performed; if -1 is given, '

  597.                 'reduce to given target shape (only used in megbrain)'),

  598.             -1))

  599. 

  600. (pdef('Reduce', 'reduce along given axis', version=1, is_legacy=True).

  601.  add_enum('Mode',

  602.           'SUM = 0',

  603.           Doc('SUM_SQR = 1', 'sum of x * x for each element x'),

  604.           'PRODUCT = 2', 'MIN = 3', 'MAX = 4', 'MEAN = 5').

  605.  add_fields('int32',

  606.             Doc('axis',

  607.                 'axis along which reduction is performed; if -1 is given, '

  608.                 'reduce to given target shape (only used in megbrain)'),

  609.             -1).

  610.  add_enum('DataType',

  611.           Doc('DEFAULT = 0',

  612. '''

  613. input/output are the same data type, and the internal computation type would be chosen by the input/output dtypes and the reduction mode.

  614. Currently, ```DEFAULT``` mode means:

  615. 

  616. +--------------------+-----------------------------------+-------------------+

  617. | Input/Output DType | Mode                              | Computation DType |

  618. +====================+===================================+===================+

  619. | FLOAT32            | MIN/MAX/MEAN/SUM/SUM_SQR/PRODUCT  | FLOAT32           |

  620. +--------------------+-----------------------------------+-------------------+

  621. | FLOAT16            | MIN/MAX/MEAN/SUM/SUM_SQR/PRODUCT  | FLOAT16           |

  622. +--------------------+-----------------------------------+-------------------+

  623. | INT32              | MIN/MAX/MEAN/SUM/SUM_SQR/PRODUCT  | INT32             |

  624. +--------------------+-----------------------------------+-------------------+

  625. | INT8               | MIN/MAX/MEAN/SUM/SUM_SQR/PRODUCT  | INT8              |

  626. +--------------------+-----------------------------------+-------------------+

  627. | QuantizedS8        | MIN/MAX                           | QuantizedS8       |

  628. +--------------------+-----------------------------------+-------------------+

  629. | QuantizedS8        | MEAN/SUM                          | QuantizedS32      |

  630. +--------------------+-----------------------------------+-------------------+

  631. | Quantized8Asymm    | MIN/MAX                           | Quantized8Asymm   |

  632. +--------------------+-----------------------------------+-------------------+

  633. | Quantized8Asymm    | MEAN/SUM                          | QuantizedS32      |

  634. +--------------------+-----------------------------------+-------------------+

  635. 

  636. '''

  637. ),

  638.           Doc('FLOAT_IO16xC32 = 1', 'Deprecated. This was replaced by '

  639.               'FLOAT_O16xC32, and input\'s dtype decided by actual input tensor.'),

  640.           Doc('FLOAT_O32xC32 = 2', 'compute/output both are float32'),

  641.           Doc('FLOAT_O16xC32 = 3', 'compute are float32, output float16'),

  642.           Doc('QUINT_I8xO32 = 4', 'input quint8, compute and output are qint32'),

  643.           Doc('QINT_I8xO32 = 5', 'input qint8, compute and output are qint32'),

  644.      name_field='data_type'))

  645. 

  646. (pdef('Reduce', 'reduce along given axis', version=2).

  647.  add_enum('Mode',

  648.           'SUM = 0',

  649.           Doc('SUM_SQR = 1', 'sum of x * x for each element x'),

  650.           'PRODUCT = 2', 'MIN = 3', 'MAX = 4', 'MEAN = 5').

  651.  add_fields('int32',

  652.             Doc('axis',

  653.                 'axis along which reduction is performed; if INT_MAX is given, '

  654.                 'reduce to given target shape (only used in megbrain)'),

  655.             (1<<31)-1).

  656.  add_enum('DataType',

  657.           Doc('DEFAULT = 0',

  658. '''

  659. input/output are the same data type, and the internal computation type would be chosen by the input/output dtypes and the reduction mode.

  660. Currently, ```DEFAULT``` mode means:

  661. 

  662. +--------------------+-----------------------------------+-------------------+

  663. | Input/Output DType | Mode                              | Computation DType |

  664. +====================+===================================+===================+

  665. | FLOAT32            | MIN/MAX/MEAN/SUM/SUM_SQR/PRODUCT  | FLOAT32           |

  666. +--------------------+-----------------------------------+-------------------+

  667. | FLOAT16            | MIN/MAX/MEAN/SUM/SUM_SQR/PRODUCT  | FLOAT16           |

  668. +--------------------+-----------------------------------+-------------------+

  669. | INT32              | MIN/MAX/MEAN/SUM/SUM_SQR/PRODUCT  | INT32             |

  670. +--------------------+-----------------------------------+-------------------+

  671. | INT8               | MIN/MAX/MEAN/SUM/SUM_SQR/PRODUCT  | INT8              |

  672. +--------------------+-----------------------------------+-------------------+

  673. | QuantizedS8        | MIN/MAX                           | QuantizedS8       |

  674. +--------------------+-----------------------------------+-------------------+

  675. | QuantizedS8        | MEAN/SUM                          | QuantizedS32      |

  676. +--------------------+-----------------------------------+-------------------+

  677. | Quantized8Asymm    | MIN/MAX                           | Quantized8Asymm   |

  678. +--------------------+-----------------------------------+-------------------+

  679. | Quantized8Asymm    | MEAN/SUM                          | QuantizedS32      |

  680. +--------------------+-----------------------------------+-------------------+

  681. 

  682. '''

  683. ),

  684.           Doc('FLOAT_IO16xC32 = 1', 'Deprecated. This was replaced by '

  685.               'FLOAT_O16xC32, and input\'s dtype decided by actual input tensor.'),

  686.           Doc('FLOAT_O32xC32 = 2', 'compute/output both are float32'),

  687.           Doc('FLOAT_O16xC32 = 3', 'compute are float32, output float16'),

  688.           Doc('QUINT_I8xO32 = 4', 'input quint8, compute and output are qint32'),

  689.           Doc('QINT_I8xO32 = 5', 'input qint8, compute and output are qint32'),

  690.      name_field='data_type'))

  691. 

  692. (pdef('Cumsum', 'calculate accumulated sum along given axis', version=0, is_legacy=True).

  693.  add_fields('int32',

  694.           Doc('axis',

  695.               'axis along which cumsum is performed'),

  696.           -1).

  697.  add_fields('bool',

  698.           Doc('exclusive',

  699.               'whether the current element is taken into account'),

  700.           'true').

  701.  add_fields('bool',

  702.           Doc('reverse',

  703.               'whether the cumsum is forward or backward'),

  704.           'false'))

  705. 

  706. (pdef('Cumsum', 'calculate accumulated sum along given axis', version=1).

  707.  add_fields('int32',

  708.           Doc('axis',

  709.               'axis along which cumsum is performed, default with INT_MAX'),

  710.           (1<<31)-1).

  711.  add_fields('bool',

  712.           Doc('exclusive',

  713.               'whether the current element is taken into account'),

  714.           'true').

  715.  add_fields('bool',

  716.           Doc('reverse',

  717.               'whether the cumsum is forward or backward'),

  718.           'false'))

  719. 

  720. (pdef('CondTake').

  721.  add_enum('Mode',

  722.           Doc('EQ = 0', 'take if ``abs(data-val)<eps``'),

  723.           Doc('NEQ = 1', 'take if ``abs(data-val)>=eps``'),

  724.           Doc('LT = 2', 'take if ``data<val``'),

  725.           Doc('LEQ = 3', 'take if ``data<=val``'),

  726.           Doc('GT = 4', 'take if ``data>val``'),

  727.           Doc('GEQ = 5', 'take if ``data>=val``')).

  728.  add_fields('float32',

  729.             Doc('val', 'the value to be compared with; note that for integer '

  730.                 'data, val is also converted to int'), 0).

  731.  add_fields('float32', Doc('eps', 'used for float equality comparison'),

  732.             1e-6))

  733. 

  734. 

  735. pdef('Argsort').add_enum('Order', 'ASCENDING = 0', 'DESCENDING = 1')

  736. 

  737. (pdef('IndexingRemap').

  738.  add_fields('bool',

  739.             Doc('is_non_overlapping',

  740.                 'Whether no two dst element maps to the same src element. '

  741.                 'Enabling this option can accelerate gradient operator since'

  742.                 ' atomic adding operations could be avoided.'),

  743.             'false'))

  744. 

  745. pdef('Sleep').add_fields('float32', Doc('time', 'time to sleep in seconds'), 0)

  746. 

  747. (pdef('Linspace').

  748.  add_fields('bool',

  749.             Doc('endpoint',

  750.                 'Whether stop is included in the generated tensor'),

  751.             'true'))

  752. 

  753. (pdef('LinspaceFull').

  754.  add_fields('float64',

  755.             Doc('start', 'The first val.'),

  756.             0).

  757.  add_fields('float64',

  758.             Doc('stop', 'The last val.'),

  759.             1).

  760.  add_fields('bool',

  761.             Doc('endpoint',

  762.                 'Whether stop is included in the generated tensor'),

  763.             'true'))

  764. 

  765. (pdef('Eye').

  766.  add_fields(

  767.      'int32',

  768.      Doc('k', 'Index of the diagonal: 0 (the default) refers to the main '

  769.          'diagonal, a positive value refers to an upper diagonal, and a '

  770.          'negative value to a lower diagonal.'),

  771.      0).

  772.  add_fields(

  773.      'dtype', Doc('dtype', 'data type of output value'),

  774.      'DTypeEnum::Float32'))

  775. 

  776. (pdef('Diag').

  777.  add_fields(

  778.      'int32',

  779.      Doc('k', 'Index of the diagonal: 0 (the default) refers to the main '

  780.          'diagonal, a positive value refers to an upper diagonal, and a '

  781.          'negative value to a lower diagonal.'),

  782.      0))

  783. 

  784. (pdef('UniformRNG', version=0, is_legacy=True).

  785.  add_fields('uint64', 'seed', 0))

  786. 

  787. (pdef('UniformRNG', version=1).

  788.  add_fields('uint64', 'seed', 0).

  789.  add_fields(

  790.      'dtype', Doc('dtype', 'The dtype of output Tensor. Only support Float32.'),

  791.      'DTypeEnum::Float32'))

  792. 

  793. (pdef('GaussianRNG', version=0, is_legacy=True).

  794.  add_fields('uint64', 'seed', 0).

  795.  add_fields('float32', 'mean', 0, 'std', 1))

  796. 

  797. (pdef('GaussianRNG', version=1).

  798.  add_fields('uint64', 'seed', 0).

  799.  add_fields('float32', 'mean', 0, 'std', 1).

  800.  add_fields(

  801.      'dtype', Doc('dtype', 'The dtype of output Tensor. Only support Float32.'),

  802.      'DTypeEnum::Float32'))

  803. 

  804. (pdef('GammaRNG').

  805.  add_fields('uint64', 'seed', 0))

  806. 

  807. (pdef('BetaRNG').

  808.  add_fields('uint64', 'seed', 0))

  809. 

  810. (pdef('PoissonRNG').

  811.  add_fields('uint64', 'seed', 0))

  812. 

  813. (pdef('PermutationRNG').

  814.  add_fields('uint64', 'seed', 0).

  815.  add_fields(

  816.      'dtype', Doc('dtype', 'The dtype of output Tensor. Int32, Int16 and '

  817.                   'Float32 are supported.'),

  818.      'DTypeEnum::Int32'))

  819. 

  820. (pdef('ShuffleRNG').

  821.  add_fields('uint64', 'seed', 0))

  822. 

  823. (pdef('Flip').

  824.  add_fields('bool', 'vertical', 'false', 'horizontal', 'false'))

  825. 

  826. (pdef('Rotate')

  827.  .add_fields('bool', 'clockwise', 'true'))

  828. 

  829. (pdef('ROICopy')

  830.  .add_fields('uint32', 'row_from', 0, 'row_to', 0, 'col_from', 0, 'col_to', 0))

  831. 

  832. (pdef('CvtColor')

  833.  .add_enum('Mode', 'RGB2GRAY = 0', 'RGB2YUV = 1', 'YUV2RGB = 2', 'GRAY2RGB = 3', 'RGBA2RGB = 4',

  834.     'RGBA2BGR = 5', 'RGBA2GRAY = 6', 'RGB2BGR = 7', 'BGR2GRAY = 8', 'BGR2RGB = 9',

  835.     Doc('YUV2GRAY_NV21 = 10', 'For historical reasons, referred to as YCC by opencv'),

  836.     'YUV2RGB_NV21 = 11', 'YUV2BGR_NV21 = 12', 'YUV2GRAY_NV12 = 13', 'YUV2RGB_NV12 = 14',

  837.     'YUV2BGR_NV12 = 15', 'YUV2GRAY_YV12 = 16', 'YUV2RGB_YV12 = 17', 'YUV2BGR_YV12 = 18',

  838.     'YUV2GRAY_YU12 = 19', 'YUV2RGB_YU12 = 20', 'YUV2BGR_YU12 = 21',

  839.     'YCrCb2RGB = 22', 'YCrCb2BGR = 23',

  840.     Doc('BT601_YUV2RGB_NV21 = 24', 'BT601 yuv format, referred to as YUV by opencv'),

  841.     'BT601_YUV2BGR_NV21 = 25', 'BT601_YUV2RGB_NV12 = 26', 'BT601_YUV2BGR_NV12 = 27',

  842.     'BT601_YUV2RGB_YV12 = 28', 'BT601_YUV2BGR_YV12 = 29', 'BT601_YUV2RGB_YU12 = 30',

  843.     'BT601_YUV2BGR_YU12 = 31',

  844.     member_alias=[('YUV2GRAY_NV21', 'BT601_YUV2GRAY_NV21'),

  845.                   ('YUV2GRAY_NV12', 'BT601_YUV2GRAY_NV12'),

  846.                   ('YUV2GRAY_YV12', 'BT601_YUV2GRAY_YV12'),

  847.                   ('YUV2GRAY_YU12', 'BT601_YUV2GRAY_YU12')],

  848.     name_field = 'mode'))

  849. 

  850. (pdef('WarpAffine', version=0, is_legacy=True)

  851.  .add_enum_alias('InterpolationMode', 'WarpPerspectiveV1', name_field='imode')

  852.  .add_enum_alias('BorderMode', 'WarpPerspectiveV1', name_field='border_mode')

  853.  .add_fields('float32', Doc('border_val', 'used for CONSTANT bmode'), '.0f'))

  854. 

  855. (pdef('WarpAffine', version=1, is_legacy=True)

  856.  .add_enum_alias('InterpolationMode', 'WarpPerspectiveV1', name_field='imode')

  857.  .add_enum_alias('BorderMode', 'WarpPerspectiveV1', name_field='border_mode')

  858.  .add_fields('float32', Doc('border_val', 'used for CONSTANT bmode'), '.0f')

  859.  .add_enum_alias('Format', 'ConvolutionV0', default=1))

  860. 

  861. (pdef('WarpAffine', version=2)

  862.  .add_enum_alias('InterpolationMode', 'WarpPerspectiveV1', name_field='imode')

  863.  .add_enum_alias('BorderMode', 'WarpPerspectiveV1', name_field='border_mode')

  864.  .add_fields('float32', Doc('border_val', 'used for CONSTANT bmode'), '.0f')

  865.  .add_enum_alias('Format', 'Convolution', default=1))

  866. 

  867. 

  868. (pdef('GaussianBlur')

  869.  .add_enum_alias('BorderMode', 'WarpPerspectiveV1', name_field='border_mode')

  870.  .add_fields('uint32', 'kernel_height', 0, 'kernel_width', 0)

  871.  .add_fields('float32','sigma_x', '0.f', 'sigma_y', '0.f'))

  872. 

  873. (pdef('Resize', version=0, is_legacy=True)

  874.  .add_enum_alias('InterpolationMode', 'WarpPerspectiveV1', name_field='imode'))

  875. 

  876. (pdef('Resize', version=1, is_legacy=True)

  877.  .add_enum_alias('InterpolationMode', 'WarpPerspectiveV1', name_field='imode')

  878.  .add_enum_alias('Format', 'ConvolutionV0', default=1))

  879. 

  880. (pdef('Resize', version=2)

  881.  .add_enum_alias('InterpolationMode', 'WarpPerspectiveV1', name_field='imode')

  882.  .add_enum_alias('Format', 'Convolution', default=1))

  883. 

  884. (pdef('Remap', version=0,is_legacy=True)

  885.  .add_enum_alias('InterpolationMode', 'WarpPerspectiveV1', name_field='imode')

  886.  .add_enum_alias('BorderMode', 'WarpPerspectiveV1', name_field='border_type')

  887.  .add_enum_alias('Format', 'ConvolutionV0', default=1)

  888.  .add_fields('float32', 'scalar', '0.f'))

  889. 

  890. (pdef('Remap', version=1)

  891.  .add_enum_alias('InterpolationMode', 'WarpPerspectiveV1', name_field='imode')

  892.  .add_enum_alias('BorderMode', 'WarpPerspectiveV1', name_field='border_type')

  893.  .add_enum_alias('Format', 'Convolution', default=1)

  894.  .add_fields('float32', 'scalar', '0.f'))

  895. 

  896. (pdef('Convolution3D').

  897.  add_enum('Mode', 'CROSS_CORRELATION = 0', 'CONVOLUTION = 1').

  898.  add_fields(

  899.      'uint32',

  900.      Doc('pad_d', 'padding on one side on the first dimension'), 0,

  901.      Doc('pad_h', 'padding on one side on the second dimension'), 0,

  902.      Doc('pad_w', 'padding on one side on the third dimension'), 0,

  903.      Doc('stride_d', 'kernel stride on the first dimension'), 1,

  904.      Doc('stride_h', 'kernel stride on the second dimension'), 1,

  905.      Doc('stride_w', 'kernel stride on the third dimension'), 1,

  906.      Doc('dilate_d', 'dilation (i.e. size of each zero-padded kernel block) '

  907.          'on the first dimension'), 1,

  908.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  909.          'on the second dimension'), 1,

  910.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  911.          'on the third dimension'), 1

  912.  ).

  913.  add_enum('Sparse',

  914.           Doc('DENSE = 0', 'dense convolution: filter shape should be '

  915.               '[oc, ic, spatial...] if format is NCDHW, '

  916.               '[oc, spatial..., ic] if format is NDHWC'),

  917.           Doc('GROUP = 1', 'group convolution: filter shape should be '

  918.               '[group, oc_per_group, ic_per_group, spatial...] if format is NCDHW, '

  919.               '[group, oc_per_group, spatial..., ic_per_group] if format is NDHWC')

  920.           ).

  921.  add_enum('DataType',

  922.           Doc('FLOAT = 0', 'input/output both float32/float16'),

  923.           Doc('FLOAT_IO16xC32 = 1', 'input/output both float16, the internal '

  924.               'compute is float32'),

  925.           name_field='data_type').

  926.  add_enum('Format', 'NCDHW = 0', 'NDHWC = 1')

  927.  )

  928. 

  929. (pdef('Conv3DBias').

  930.  add_enum('NonlineMode', 'IDENTITY = 0', 'RELU = 1', 'SIGMOID = 2').

  931.  add_enum_alias('Mode', 'Convolution3D').

  932.  add_fields('uint32', 'pad_d', 0, 'pad_h', 0, 'pad_w', 0,

  933.                 'stride_d', 1, 'stride_h', 1, 'stride_w', 0))

  934. 

  935. (pdef('SeparableConv3D').

  936.  add_enum_alias('Mode', 'Convolution3D').

  937.  add_enum('BorderMode', 'BORDER_REPLICATE = 0', 'BORDER_REFLECT = 1',

  938.           'BORDER_REFLECT_101 = 2','BORDER_WRAP = 3',

  939.           'BORDER_CONSTANT = 4', 'BORDER_TRANSPARENT = 5','BORDER_ISOLATED = 6').

  940.  add_fields('bool', 'is_symm_kernel', 'true').

  941.  add_fields('uint32', 'pad_d', 0, 'pad_h', 0, 'pad_w', 0,

  942.             'stride_d', 0, 'stride_h', 1, 'stride_w', 1,

  943.             'ksize_d', 0, 'ksize_h', 3, 'ksize_w', 3,

  944.             'anchor_d', 0, 'anchor_h', 1, 'anchor_w', 1))

  945. 

  946. (pdef('TopK').

  947.  add_enum(

  948.      'Mode',

  949.      Doc('KTH_ONLY = 0', "only the value of the k'th element would be computed"),

  950.      Doc('VALUE_IDX_NOSORT = 1',

  951.          'all the top-k values and corresponding indices would be computed; '

  952.          'no order is guaranteed'),

  953.      Doc('VALUE_IDX_SORTED = 2',

  954.          'all the top-k values and corresponding indices sorted'))

  955.  )

  956. 

  957. RELAYOUT_FORMAT_MODE_DOC = """

  958. Relayout mode.

  959. 

  960. **Naming conventions**

  961. 

  962. 1. ``A_B`` means change from layout format ``A`` to ``B``.

  963. 2. ``INTER_WEIGHT_xx`` means relayout the weight for faster processing by

  964.    :attr:`Convolution.Format.NHWCD4` convolutions.

  965. 3. A suffix of ``I`` means ``Image2DPack4TensorFormat`` tensor format is used

  966.    for faster processing on GPUs.

  967. 

  968. **Layout definitions**

  969. 

  970. * ``NCHW`` layout: ``{N, C, H, W}``

  971. * ``NHWC`` layout: ``{N, H, W, C}``

  972. * ``NHWCD4`` layout: ``{N, H, (C + 3) / 4, W, 4}``

  973. * ``NHWCD4I`` layout: with ``align_axis = 2``

  974. * ``NCHW4`` layout: ``{N, C/4, H, W, 4}``

  975. * ``NCHW88`` layout: ``{N, C/8, H, W, 8}``

  976. * ``CHWN4`` layout: ``{C/4, H, W, N, 4}``

  977. * ``NCHW64`` layout: ``{N, C/64, H, W, 64}``

  978. 

  979. **Float weight transformation definitions**

  980. 

  981. +---------------+---------------------------------+--------------------+--------------------------------------+------+

  982. | Sparsity Type | Input Layout                    | Input Req          | Output Layout                        | Axis |

  983. +===============+=================================+====================+======================================+======+

  984. | DENSE         | ``{OC, IC, FH, FW}``            | ``OC % 4 == 0``    | ``{OC/4, FH, FW, IC, 4}``            | 3    |

  985. +---------------+---------------------------------+--------------------+--------------------------------------+------+

  986. | GROUP         | ``{GROUP, OCPG, ICPG, FH, FW}`` | ``OCPG % 4 == 0``  | ``{GROUP, OCPG/4, FH, FW, ICPG, 4}`` | 4    |

  987. |               |                                 | ``ICPG % 4 == 0``  |                                      |      |

  988. +---------------+---------------------------------+--------------------+--------------------------------------+------+

  989. | CHAN          | ``{GROUP, 1, 1, FH, FW}``       | ``GROUP % 4 == 0`` | ``{GROUP / 4, 1, FH ,FW, 4}``        | 1    |

  990. +---------------+---------------------------------+--------------------+--------------------------------------+------+

  991. 

  992. **Float weight transformation nchw88 definitions**

  993. 

  994. +---------------+---------------------------------+--------------------+--------------------------------------+

  995. | Sparsity Type | Input Layout                    | Input Req          | Output Layout                        |

  996. +===============+=================================+====================+======================================+

  997. | DENSE         | ``{OC, IC, FH, FW}``            | ``OC % 8 == 0``    |``{OC/8, IC/8 ,FH, FW, 8(IC), 8(OC)}``|

  998. |               |                                 | ``IC % 8 == 0``    |                                      |

  999. +---------------+---------------------------------+--------------------+--------------------------------------+

  1000. | GROUP         | ``{GROUP, OCPG, ICPG, FH, FW}`` | ``OCPG % 8 == 0``  | ``{GROUP, OCPG/8, ICPG/8 FH, FW,     |

  1001. |               |                                 | ``ICPG % 8 == 0``  |  8(ICPG), 8(OCPG)} ``                |

  1002. +---------------+---------------------------------+--------------------+--------------------------------------+

  1003. | CHAN          | ``{GROUP, 1, 1, FH, FW}``       | ``GROUP % 8 == 0`` | ``{GROUP / 8, 1, FH ,FW, 8}``        |

  1004. +---------------+---------------------------------+--------------------+--------------------------------------+

  1005. 

  1006. **Int8(DOT) weight transformation definitions**

  1007. 

  1008. +---------------+---------------------------------+--------------------+------------------------------------------+------+

  1009. | Sparsity Type | Input Layout                    | Input Req          | Output Layout                            | Axis |

  1010. +===============+=================================+====================+==========================================+======+

  1011. | DENSE         | ``{OC, IC, FH, FW}``            | ``OC % 4 == 0``    | ``{OC/4, FH, FW, IC/4, 4, 4}`            | 3    |

  1012. +---------------+---------------------------------+--------------------+------------------------------------------+------+

  1013. | GROUP         | ``{GROUP, OCPG, ICPG, FH, FW}`` | ``OCPG % 4 == 0``  | ``{GROUP, OCPG/4, FH, FW, ICPG/4, 4, 4}``| 4    |

  1014. |               |                                 | ``ICPG % 4 == 0``  |                                          |      |

  1015. +---------------+---------------------------------+--------------------+------------------------------------------+------+

  1016. 

  1017. Note: the axis column means the corresponding ``align_axis`` for image format

  1018. when the ``I`` suffix is present.

  1019. 

  1020. Note: NCHW_NCHW4_WEIGHT will auto pad oc and ic, you should remove oc in later opr by seting group and oc param with NCHW4_NCHW

  1021. """

  1022. (pdef('RelayoutFormat', 'Change the tensor layout format', version=0, is_legacy=True).

  1023.  add_enum(

  1024.      Doc('Mode', RELAYOUT_FORMAT_MODE_DOC),

  1025.      'NHWC_NHWCD4 = 0',

  1026.      'NHWCD4_NHWC = 1',

  1027.      'NHWC_NHWCD4I = 2',

  1028.      'NCHW_NHWCD4 = 3',

  1029.      'NCHW_NHWCD4I = 4',

  1030.      'NHWCD4I_NCHW = 5',

  1031.      'NHWCD4_NCHW = 6',

  1032.      'INTER_WEIGHT_DENSE = 7',

  1033.      'INTER_WEIGHT_DENSEI = 8',

  1034.      'INTER_WEIGHT_GROUP = 9',

  1035.      'INTER_WEIGHT_GROUPI = 10',

  1036.      'INTER_WEIGHT_CHAN = 11',

  1037.      'INTER_WEIGHT_CHANI = 12',

  1038.      'INTER_WEIGHT_DENSEI_DOT = 13',

  1039.      'INTER_WEIGHT_GROUPI_DOT = 14',

  1040.      'NCHW4_CHWN4 = 15',

  1041.      'CHWN4_NCHW4 = 16',

  1042.      'NCHW_NCHW88_CONV_DENSE_WEIGHT = 17',

  1043.      'NCHW_NCHW88_CONV_CHAN_WEIGHT = 18',

  1044.      'NCHW_NCHW88_CONV_GROUP_WEIGHT = 19',

  1045.      'NCHW_NCHW88 = 20',

  1046.      'NCHW88_NCHW = 21',

  1047.      'NCHW_NCHW4_IC_SMALL = 22',

  1048.      'NCHW_NCHW4_IC_SMALL_CONV_DENSE_WEIGHT = 23',

  1049.      'NCHW_NCHW4 = 24',

  1050.      'NCHW4_NCHW = 25',

  1051.      'NCHW_NCHW4_WEIGHT = 26',

  1052.      'NCHW_NCHW64 = 27',

  1053.      'NCHW64_NCHW = 28',

  1054.      'NCHW_NHWC = 29',

  1055.      'NHWC_NCHW = 30',

  1056.      'NHWCD4I_NHWC = 31',

  1057.     )

  1058.  )

  1059. 

  1060. (pdef('RelayoutFormat', 'Change the tensor layout format', version=1).

  1061.     add_enum_alias('Mode', 'RelayoutFormatV0').

  1062.     add_fields('uint32', 'oc', '0').

  1063.     add_fields('uint32', 'group', '1')

  1064. )

  1065. 

  1066. (pdef('SeparableFilter', version=0, is_legacy=True).

  1067.  add_enum_alias('Format', 'ConvolutionV0').

  1068.  add_enum_alias('BorderMode', 'WarpPerspectiveV1').

  1069.  add_fields('bool', 'is_symm_kernel', 'true').

  1070.  add_fields('uint32', 'ksize_h', 3, 'ksize_w', 3, 'anchor_h', 1, 'anchor_w', 1))

  1071. 

  1072. (pdef('SeparableFilter', version=1).

  1073.  add_enum_alias('Format', 'Convolution').

  1074.  add_enum_alias('BorderMode', 'WarpPerspectiveV1').

  1075.  add_fields('bool', 'is_symm_kernel', 'true').

  1076.  add_fields('uint32', 'ksize_h', 3, 'ksize_w', 3, 'anchor_h', 1, 'anchor_w', 1))

  1077. 

  1078. (pdef('LocalShare', 'Local share convolution',version=0, is_legacy=True).

  1079.  add_enum_alias('Mode', 'ConvolutionV0').

  1080.  add_fields(

  1081.      'uint32',

  1082.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  1083.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  1084.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  1085.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  1086.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  1087.          'on the second dimension'), 1,

  1088.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  1089.          'on the second dimension'), 1,

  1090.      Doc('spatial_groups_h', 'spatial groups on the first dimension'), 1,

  1091.      Doc('spatial_groups_w', 'spatial groups on the second dimension'), 1

  1092.  ).

  1093.  add_enum_alias('Sparse', 'ConvolutionV0').

  1094.  add_enum_alias('Format', 'ConvolutionV0').

  1095.  add_enum_alias('ComputeMode', 'ConvolutionV1')

  1096.  )

  1097. 

  1098. (pdef('LocalShare', 'Local share convolution', version=1).

  1099.  add_enum_alias('Mode', 'ConvolutionV0').

  1100.  add_fields(

  1101.      'uint32',

  1102.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  1103.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  1104.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  1105.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  1106.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  1107.          'on the second dimension'), 1,

  1108.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  1109.          'on the second dimension'), 1,

  1110.      Doc('spatial_groups_h', 'spatial groups on the first dimension'), 1,

  1111.      Doc('spatial_groups_w', 'spatial groups on the second dimension'), 1

  1112.  ).

  1113.  add_enum_alias('Sparse', 'ConvolutionV0').

  1114.  add_enum_alias('Format', 'Convolution').

  1115.  add_enum_alias('ComputeMode', 'ConvolutionV1')

  1116.  )

  1117. 

  1118. 

  1119. (pdef('ROIAlign',version=0,is_legacy=True).

  1120.  add_enum('Mode', 'MAX = 0', 'AVERAGE = 1', name_field='mode').

  1121.  add_enum_alias('Format', 'ConvolutionV0').

  1122.  add_fields('float32', 'spatial_scale', '1.0').

  1123.  add_fields('float32', 'offset', '0.0').

  1124.  add_fields('uint32',

  1125.             'pooled_height', '1',

  1126.             'pooled_width', '1',

  1127.             'sample_height', '2',

  1128.             'sample_width', '2')

  1129.  )

  1130. 

  1131. (pdef('ROIAlign', version=1).

  1132.  add_enum_alias('Mode', 'ROIAlignV0', name_field='mode').

  1133.  add_enum_alias('Format', 'Convolution').

  1134.  add_fields('float32', 'spatial_scale', '1.0').

  1135.  add_fields('float32', 'offset', '0.0').

  1136.  add_fields('uint32',

  1137.             'pooled_height', '1',

  1138.             'pooled_width', '1',

  1139.             'sample_height', '2',

  1140.             'sample_width', '2')

  1141.  )

  1142. 

  1143. (pdef('Correlation').

  1144.  add_enum_alias('Format', 'ConvolutionV0').

  1145.  add_fields('uint32', 'kernel_size', '1').

  1146.  add_fields('uint32', 'max_displacement', '1').

  1147.  add_fields('uint32', 'stride1', '1').

  1148.  add_fields('uint32', 'stride2', '1').

  1149.  add_fields('uint32', 'pad_size', '0').

  1150.  add_fields('bool', 'is_multiply', 'true')

  1151.  )

  1152. 

  1153. (pdef('DeformablePSROIPooling').

  1154.  add_fields('bool', 'no_trans', 'true').

  1155.  add_fields('float32', 'spatial_scale', 1,

  1156.      'trans_std', 1).

  1157.  add_fields('uint32',

  1158.     Doc('pooled_h', 'height of pooling output'), 1,

  1159.     Doc('pooled_w', 'width of pooling output'), 1,

  1160.     Doc('part_size', 'size of each deformable part'), 1,

  1161.     Doc('sample_per_part', 'sample count of each bbox'), 1))

  1162. 

  1163. (pdef('BatchConvBias', 'Batch convolution (unshare weights on the batch dimension)',version=0,is_legacy=True).

  1164.  add_enum_alias('NonlineMode', 'ConvBiasV0').

  1165.  add_enum_alias('Mode', 'ConvolutionV0').

  1166.  add_fields(

  1167.      'uint32',

  1168.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  1169.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  1170.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  1171.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  1172.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  1173.          'on the second dimension'), 1,

  1174.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  1175.          'on the second dimension'), 1,

  1176.  ).

  1177.  add_enum_alias('Sparse', 'ConvolutionV0').

  1178.  add_enum_alias('Format', 'ConvolutionV0').

  1179.  add_enum_alias('ComputeMode', 'ConvolutionV1', name_field="compute_mode")

  1180.  )

  1181. 

  1182. (pdef('BatchConvBias', 'Batch convolution (unshare weights on the batch dimension)',version=1).

  1183.  add_enum_alias('NonlineMode', 'ConvBiasV0').

  1184.  add_enum_alias('Mode', 'ConvolutionV0').

  1185.  add_fields(

  1186.      'uint32',

  1187.      Doc('pad_h', 'padding on one side on the first dimension'), 0,

  1188.      Doc('pad_w', 'padding on one side on the second dimension'), 0,

  1189.      Doc('stride_h', 'kernel stride on the first dimension'), 1,

  1190.      Doc('stride_w', 'kernel stride on the second dimension'), 1,

  1191.      Doc('dilate_h', 'dilation (i.e. size of each zero-padded kernel block) '

  1192.          'on the second dimension'), 1,

  1193.      Doc('dilate_w', 'dilation (i.e. size of each zero-padded kernel block) '

  1194.          'on the second dimension'), 1,

  1195.  ).

  1196.  add_enum_alias('Sparse', 'ConvolutionV0').

  1197.  add_enum_alias('Format', 'Convolution').

  1198.  add_enum_alias('ComputeMode', 'ConvolutionV1', name_field="compute_mode")

  1199.  )

  1200. 

  1201. (pdef('FakeQuant').

  1202.  add_fields('int32','qmin','-2147483648').

  1203.  add_fields('int32','qmax','2147483647')

  1204.  )

  1205. (pdef('TQT').

  1206.  add_fields('int32', 'qmin', '-2147483648').

  1207.  add_fields('int32', 'qmax', '2147483647')

  1208.  )

  1209. (pdef('LSQ').

  1210.  add_fields('int32', 'qmin', '-2147483648').

  1211.  add_fields('int32', 'qmax', '2147483647')

  1212.  )

  1213. pdef('Fill').add_fields('float32', 'value', '0')

  1214. 

  1215. pdef('CheckNonFinite').add_fields('float32', 'scale', '1.0')

  1216. 

  1217. 

  1218. PADDING_MODES = [Doc('REPLICATE = 0', 'aaaaaa|abcdefgh|hhhhhhh'),

  1219.                 Doc('REFLECT = 1', 'fedcba|abcdefgh|hgfedcb'),

  1220.                 Doc('CONSTANT = 2', 'iiiiii|abcdefgh|iiiiiii')]

  1221. (pdef('Padding').

  1222.  add_fields('uint32', Doc('front_offset_dim0','offset in dim 0'), 0).

  1223.  add_fields('uint32', Doc('front_offset_dim1','offset in dim 1'), 0).

  1224.  add_fields('uint32', Doc('front_offset_dim2','offset in dim 2'), 0).

  1225.  add_fields('uint32', Doc('front_offset_dim3','offset in dim 3'), 0).

  1226.  add_fields('uint32', Doc('front_offset_dim4','offset in dim 4'), 0).

  1227.  add_fields('uint32', Doc('front_offset_dim5','offset in dim 5'), 0).

  1228.  add_fields('uint32', Doc('front_offset_dim6','offset in dim 6'), 0).

  1229.  add_fields('uint32', Doc('back_offset_dim0', 'back offset in dim0'), 0).

  1230.  add_fields('uint32', Doc('back_offset_dim1', 'back offset in dim1'), 0).

  1231.  add_fields('uint32', Doc('back_offset_dim2', 'back offset in dim2'), 0).

  1232.  add_fields('uint32', Doc('back_offset_dim3', 'back offset in dim3'), 0).

  1233.  add_fields('uint32', Doc('back_offset_dim4', 'back offset in dim4'), 0).

  1234.  add_fields('uint32', Doc('back_offset_dim5', 'back offset in dim5'), 0).

  1235.  add_fields('uint32', Doc('back_offset_dim6', 'back offset in dim6'), 0).

  1236.  add_fields('float32', Doc('padding_val','param of padding opr'), 0).

  1237.  add_enum('PaddingMode', *PADDING_MODES,

  1238.           name_field='padding_mode', default=2,

  1239.           member_alias=[(i, 'PADDING_{}'.format(i)) for i in PADDING_MODES]

  1240.           )

  1241. )

  1242. 

  1243. (pdef('LayerNorm')

  1244.  .add_fields('bool', 'affine', 'true')

  1245.  .add_fields('float32', 'eps', '1e-5f')

  1246.  .add_fields('uint64', 'normalized_dim', '1')

  1247.  .add_fields('uint64', 'normalized_size', '1')

  1248. )

  1249. 

  1250. (pdef('GroupNorm')

  1251.  .add_fields('bool', 'affine', 'true')

  1252.  .add_fields('float32', 'eps', '1e-5f')

  1253.  .add_fields('uint32', 'group', '1')

  1254.  .add_enum_alias('Format', 'Convolution')

  1255. )

  1256. 

  1257. (pdef('Dropout')

  1258.  .add_fields('float32', 'drop_prob', '0')

  1259.  .add_fields('uint64', 'seed', '0')

  1260. )

  1261. 

  1262. (pdef('RNNCell').

  1263.  add_enum('NonlineMode', 'IDENTITY = 0', 'RELU = 1', 'TANH = 2')

  1264.  )

  1265. 

  1266. (pdef('RNN').

  1267.  add_fields('uint32', Doc('num_layers', 'Number of recurrent layers'), '1').

  1268.  add_fields('bool', Doc('bidirectional', 'If becomes a bidirectional RNN'), 'false').

  1269.  add_fields('bool', Doc('bias', 'If the layer use bias weights b_ih and b_hh'), 'true').

  1270.  add_fields('uint32', Doc('hidden_size', 'The number of features in the hidden state'), '128').

  1271.  add_fields('float32', Doc('dropout', 'If introduce a Dropout layer on the outputs of each RNN layer'), '0.f').

  1272.  add_enum_alias('NonlineMode', 'RNNCell').

  1273.  add_enum_alias('FwdMode', 'BN', name_field='fwd_mode')

  1274.  )

  1275. 

  1276. (pdef('LSTM').

  1277.  add_fields('uint32', Doc('num_layers', 'Number of recurrent layers'), '1').

  1278.  add_fields('bool', Doc('bidirectional', 'If becomes a bidirectional LSTM'), 'false').

  1279.  add_fields('bool', Doc('bias', 'If the layer use bias weights b_ih and b_hh'), 'true').

  1280.  add_fields('uint32', Doc('hidden_size', 'The number of features in the hidden state'), '128').

  1281.  add_fields('uint32', Doc('proj_size', 'If use LSTM with projections of corresponding size'), '0').

  1282.  add_fields('float32', Doc('dropout', 'If introduce a Dropout layer on the outputs of each LSTM layer'), '0.f').

  1283.  add_enum_alias('FwdMode', 'BN', name_field='fwd_mode')

  1284.  )

  1285. 

  1286. (pdef('LAMBUpdate').

  1287.  add_fields('float32', Doc('beta_1', 'beta_1 paramter of lamb'), '1.f').

  1288.  add_fields('float32', Doc('beta_2', 'beta_2 paramter of lamb'), '1.f').

  1289.  add_fields('float32', Doc('step', 'training step'), '1.f').

  1290.  add_fields('float32', Doc('lr', 'learning rate'), '1.f').

  1291.  add_fields('float32', Doc('weight_decay', 'weight decay to adjust learning rate'), '1.f').

  1292.  add_fields('float32', Doc('eps', 'eps to multi'), '1.f').

  1293.  add_fields('bool', Doc('bias_correction', 'whether correct bias'), 'true').

  1294.  add_fields('bool', Doc('always_adapt', 'apply adaptive lr to 0.0'), 'false')

  1295. )

  1296. (pdef("Norm").

  1297.  add_enum('Mode',

  1298.             Doc('P_NORM=0', 'calculate p-norm, parameter p would be ignored in other mode'),

  1299.             Doc('INF_NORM=1', 'infinite norm'),

  1300.             Doc('NEG_INF_NORM=2', 'negative infinite norm'), name_field="mode").

  1301.  add_fields('float32', Doc('p', 'the order of norm'), '2').

  1302.  add_fields('int32', Doc('dim', 'which dim the norm performed along'), '-1'),

  1303.  )