def check_and_config_para(filter, out_backprop, y, input_size, strides,
pads, dilations, data_format, kernel_name):
ori_shape_filters = filter.get("ori_shape")
ori_shape_out_backprop = out_backprop.get("ori_shape")
ori_range_filters = filter.get("range")
ori_range_out_backprop = out_backprop.get("range")
ori_range_y = y.get("range")
filter_dtype = filter.get("dtype")
out_backprop_dtype = out_backprop.get("dtype")
res_dtype = y.get("dtype")
ori_format_filters = filter.get("ori_format")
ori_format_out_backprop = out_backprop.get("ori_format")
if len(strides) == 4:
h_index = data_format.find('H')
w_index = data_format.find('W')
strides = [strides[h_index], strides[w_index]]
shape_filters = comm.get_filter_shape(
ori_format_filters, ori_shape_filters
)
shape_out_backprop = comm.get_shape_out_backprop(
ori_format_out_backprop, ori_shape_out_backprop)
dilations = comm.get_shape_dilation(data_format, dilations)
input_size = get_input_size(y)
dynamic_mode = config_dynamic_para(shape_out_backprop)
dx_shape, dedy, filter_frac, input_size, shape_out_backprop, var_map = \
config_placeholder(
shape_out_backprop, shape_filters, input_size,
filter_dtype, out_backprop_dtype,
ori_range_filters, ori_range_out_backprop, ori_range_y,
dynamic_mode)
shape_filter, shape_out_backprop, input_sizes, strides, pads, \
dilations, filter_dtype, out_backprop_dtype, res_dtype, kernel_name \
= comm.check_conv2dbp_input_params(shape_filters, shape_out_backprop,
input_size, strides, pads,
dilations, filter_dtype,
out_backprop_dtype, res_dtype,
kernel_name, None, dynamic_mode)
return dx_shape, dedy, filter_frac, input_size, shape_filter, shape_out_backprop, \
strides, pads, dilations, res_dtype, var_map, dynamic_mode, \
ori_range_out_backprop, ori_range_y
def conv2d_backprop_input_d(
filter, # pylint: disable=W0622,C0103,R0913,R0914
out_backprop,
y,
input_size,
strides,
pads,
dilations=(1, 1, 1, 1),
groups=None,
data_format="NHWC",
kernel_name="conv2d_backprop_input"):
"""
algorithm: conv2d_backprop_input
Parameters
----------
filter: dict with keys(shape and dtype)
input weight tensor
out_backprop: dict with keys(shape and dtype)
The shape of gradients.
y: dict with keys(shape and dtype)
conv2d_backprop_input output tensor, dtype must be assigned
input_size: The shape of feature map.
4-D with shape [batch, channels, height, weight].
strides: tuple/list of 4 integers
filter move stride
pads: tuple/list of 4 integers
[pad_top, pad_bottom, pad_left, pad_right]
dilations: tuple/list of 4 integers
filter expand size of dilated conv2d_backprop_input
groups: int
param for group conv2d_backprop_input
data_format: str
An optional string from: "NHWC", "NCHW". Defaults to "NHWC".
Specify the data format of the input and output data.
kernel_name: str
kernel name, default value is "conv2d_backprop_input"
Returns
-------
None
"""
ori_shape_filters = filter.get("ori_shape")
ori_shape_out_backprop = out_backprop.get("ori_shape")
ori_shape_res = y.get("ori_shape")
filters_dtype = filter.get("dtype")
out_backprop_dtype = out_backprop.get("dtype")
res_dtype = y.get("dtype")
ori_format_filters = filter.get("ori_format")
ori_format_out_backprop = out_backprop.get("ori_format")
ori_format_res = y.get("ori_format")
if list(input_size) != list(ori_shape_res):
dict_args = {}
dict_args['errCode'] = "E65007"
dict_args['param1'] = "input_size"
dict_args['param2'] = "ori_shape of y"
dict_args['actual_value'] = "{}, {}". \
format(input_size, ori_shape_res)
raise RuntimeError(dict_args, err_man.get_error_message(dict_args))
util.check_kernel_name(kernel_name)
util.check_shape_rule(ori_shape_filters, CONV_BACKPROP_SHAPE_DIM,
CONV_BACKPROP_SHAPE_DIM, DEFAULT_MAX_SHAPE_NUM)
util.check_shape_rule(ori_shape_out_backprop, CONV_BACKPROP_SHAPE_DIM,
CONV_BACKPROP_SHAPE_DIM, DEFAULT_MAX_SHAPE_NUM)
util.check_shape_rule(input_size, CONV_BACKPROP_SHAPE_DIM,
CONV_BACKPROP_SHAPE_DIM, DEFAULT_MAX_SHAPE_NUM)
util.check_shape_rule(ori_shape_res, CONV_BACKPROP_SHAPE_DIM,
CONV_BACKPROP_SHAPE_DIM, DEFAULT_MAX_SHAPE_NUM)
util.check_shape_rule(dilations, CONV_BACKPROP_SHAPE_DIM,
CONV_BACKPROP_SHAPE_DIM, DEFAULT_MAX_SHAPE_NUM)
if len(strides) == 4:
h_index = data_format.find('H')
w_index = data_format.find('W')
strides = [strides[h_index], strides[w_index]]
shape_filters = comm.get_filter_shape(ori_format_filters,
ori_shape_filters)
shape_out_backprop = comm.get_shape_out_backprop(ori_format_out_backprop,
ori_shape_out_backprop)
shape_res = comm.get_shape_res(ori_format_res, ori_shape_res)
dilations = comm.get_shape_dilation(data_format, dilations)
conv2d_backprop_input_cce(shape_filters, shape_out_backprop, shape_res,
strides, pads, dilations, filters_dtype,
out_backprop_dtype, res_dtype, kernel_name)
def conv2d_backprop_input_d_compute( # pylint: disable=C0103,W0622,R0913,R0914
filter,
out_backprop,
y,
input_size,
strides,
pads,
dilations=(1, 1, 1, 1),
groups=None,
data_format="NHWC", # pylint: disable=W0613
kernel_name="conv2d_backprop_input"):
"""
used for fusion
Parameters
----------
filter: Tensor
input weight tensor
out_backprop: Tensor
conv2d output gradients tenosr.
y: dict with keys(shape and dtype)
conv2d_backprop_input output tensor, dtype must be assigned
input_size: The shape of feature map.
4-D with shape [batch, channels, height, weight].
strides: tuple/list of 4 integers
filter move stride
pads: tuple/list of 4 integers
[pad_top, pad_bottom, pad_left, pad_right]
dilations: tuple/list of 4 integers
filter expand size of dilated conv2d_backprop_input
groups: int
param for group conv2d_backprop_input
data_format: str
An optional string from: "NHWC", "NCHW". Defaults to "NHWC".
Specify the data format of the input and output data.
kernel_name: str
kernel name, default value is "conv2d_backprop_input"
Returns
-------
Tensor of conv2d_backprop_input
"""
ori_shape_filters = [i.value for i in filter.op.attrs["ori_shape"]]
ori_shape_out_backprop = \
[i.value for i in out_backprop.op.attrs["ori_shape"]]
ori_shape_res = [i for i in y["ori_shape"]]
filters_dtype = filter.dtype
out_backprop_dtype = out_backprop.dtype
res_dtype = y["dtype"]
ori_format_filters = filter.op.attrs["ori_format"]
ori_format_out_backprop = out_backprop.op.attrs["ori_format"]
ori_format_res = y["ori_format"]
if list(input_size) != list(ori_shape_res):
dict_args = {}
dict_args['errCode'] = "E65007"
dict_args['param1'] = "input_size"
dict_args['param2'] = "ori_shape of y"
dict_args['actual_value'] = "{}, {}". \
format(input_size, ori_shape_res)
raise RuntimeError(dict_args, err_man.get_error_message(dict_args))
if len(strides) == 4:
h_index = data_format.find('H')
w_index = data_format.find('W')
strides = [strides[h_index], strides[w_index]]
shape_filters = comm.get_filter_shape(ori_format_filters,
ori_shape_filters)
shape_out_backprop = comm.get_shape_out_backprop(ori_format_out_backprop,
ori_shape_out_backprop)
shape_res = comm.get_shape_res(ori_format_res, ori_shape_res)
dilations = comm.get_shape_dilation(ori_format_out_backprop, dilations)
comm.check_conv2dbp_input_params(shape_filters, shape_out_backprop,
shape_res, strides, pads, dilations,
filters_dtype, out_backprop_dtype,
res_dtype, kernel_name)
pads = comm.get_padlist(pads, shape_res, strides, shape_filters, dilations)
res = te.lang.cce.conv2d_backprop_input_compute(filter,
out_backprop,
shape_filters,
shape_res,
strides,
pads,
dilations,
res_dtype=res_dtype,
kernel_name=kernel_name)
return res
def conv2d_transpose_d( # pylint: disable=R0913,R0914,W0613,W0622,C0103
x,
filter,
bias,
offset_w,
y,
input_size,
strides,
pads,
dilations=(1, 1, 1, 1),
groups=None,
data_format="NHWC",
output_padding=(0, 0, 0, 0),
offset_x=0,
kernel_name="conv2d_transpose_d"):
"""
algorithm: conv2d_transpose_d
Parameters
----------
x: dict with keys(shape and dtype)
The shape of gradients.
filter: dict with keys(shape and dtype)
input filter tensor
offset_w: the offset for filter
bias: dict with keys(shape and dtype)
The shape of bias.
y: dict with keys(shape and dtype)
conv2d_transpose_d output tensor, dtype must be assigned
input_size: The shape of feature map.
4-D with shape [batch, channels, height, filter].
strides: tuple/list of 4 integers
filter move stride
pads: tuple/list of 4 integers
[pad_top, pad_bottom, pad_left, pad_right]
dilations: tuple/list of 4 integers
filter expand size of dilated conv2d_transpose_d
groups: int
param for group conv2d_transpose_d
data_format: str
An optional string from: "NHWC", "NCHW". Defaults to "NHWC".
Specify the data format of the input and output data.
kernel_name: str
kernel name, default value is "conv2d_transpose_d"
output_padding: The size will be added in the output shape.
offset_x: the offset for x
Returns
-------
None
"""
ori_shape_x = x.get("ori_shape")
ori_shape_filters = filter.get("ori_shape")
ori_shape_res = y.get("ori_shape")
x_dtype = x.get("dtype")
filters_dtype = filter.get("dtype")
res_dtype = y.get("dtype")
ori_format_x = x.get("ori_format")
ori_format_filters = filter.get("ori_format")
ori_format_res = y.get("ori_format")
util.check_kernel_name(kernel_name)
util.check_shape_rule(ori_shape_filters, CONV_BACKPROP_SHAPE_DIM,
CONV_BACKPROP_SHAPE_DIM, DEFAULT_MAX_SHAPE_NUM)
util.check_shape_rule(ori_shape_x, CONV_BACKPROP_SHAPE_DIM,
CONV_BACKPROP_SHAPE_DIM, DEFAULT_MAX_SHAPE_NUM)
util.check_shape_rule(ori_shape_res, CONV_BACKPROP_SHAPE_DIM,
CONV_BACKPROP_SHAPE_DIM, DEFAULT_MAX_SHAPE_NUM)
util.check_shape_rule(dilations, CONV_BACKPROP_SHAPE_DIM,
CONV_BACKPROP_SHAPE_DIM, DEFAULT_MAX_SHAPE_NUM)
if list(input_size) != list(ori_shape_res):
dict_args = {}
dict_args['errCode'] = "E65007"
dict_args['param1'] = "input_size"
dict_args['param2'] = "ori_shape of y"
dict_args['actual_value'] = "{}, {}". \
format(input_size, ori_shape_res)
raise RuntimeError(dict_args, err_man.get_error_message(dict_args))
if len(strides) == 4:
h_index = data_format.find('H')
w_index = data_format.find('W')
strides = [strides[h_index], strides[w_index]]
shape_filters = comm.get_filter_shape(ori_format_filters,
ori_shape_filters)
shape_x = comm.get_shape_out_backprop(ori_format_x, ori_shape_x)
shape_res = comm.get_shape_res(ori_format_res, ori_shape_res)
dilations = comm.get_shape_dilation(ori_format_x, dilations)
bias_flag = bias is not None
conv2d_transpose_cce(shape_filters,
shape_x,
shape_res,
strides,
pads,
dilations=dilations,
filter_dtype=filters_dtype,
x_dtype=x_dtype,
res_dtype=res_dtype,
bias=bias_flag,
offset_x=offset_x,
kernel_name=kernel_name)
def conv2d_transpose_d_compute( # pylint: disable=R0913,R0914,W0613,C0103,W0622
x,
filter,
bias,
offset_w,
y,
input_size,
strides,
pads,
dilations=(1, 1, 1, 1),
groups=None,
data_format="NHWC",
output_padding=(0, 0, 0, 0),
offset_x=0,
kernel_name="conv2d_transpose_d"):
"""
used for fusion
Parameters
----------
x: dict with keys(shape and dtype)
The shape of gradients.
filter: dict with keys(shape and dtype)
input filter tensor
offset_w: the offset for filter
bias: dict with keys(shape and dtype)
The shape of bias.
y: dict with keys(shape and dtype)
conv2d_transpose_d output tensor, dtype must be assigned
strides: tuple/list of 4 integers
filter move stride
pads: tuple/list of 4 integers
[pad_top, pad_bottom, pad_left, pad_right]
dilations: tuple/list of 4 integers
filter expand size of dilated conv2d_transpose_d
groups: int
param for group conv2d_transpose_d
data_format: str
An optional string from: "NHWC", "NCHW". Defaults to "NHWC".
Specify the data format of the input and output data.
kernel_name: str
kernel name, default value is "conv2d_transpose_d"
output_padding: The size will be added in the output shape.
offset_x: the offset for x
Returns
-------
None
"""
ori_shape_filter = [i.value for i in filter.op.attrs["ori_shape"]]
ori_shape_x = [i.value for i in x.op.attrs["ori_shape"]]
ori_shape_res = [i for i in y["ori_shape"]]
filter_dtype = filter.dtype
x_dtype = x.dtype
res_dtype = y["dtype"]
ori_format_filter = filter.op.attrs["ori_format"]
ori_format_x = x.op.attrs["ori_format"]
ori_format_res = y["ori_format"]
if list(input_size) != list(ori_shape_res):
dict_args = {}
dict_args['errCode'] = "E65007"
dict_args['param1'] = "input_size"
dict_args['param2'] = "ori_shape of y"
dict_args['actual_value'] = "{}, {}". \
format(input_size, ori_shape_res)
raise RuntimeError(dict_args, err_man.get_error_message(dict_args))
if len(strides) == 4:
h_index = data_format.find('H')
w_index = data_format.find('W')
strides = [strides[h_index], strides[w_index]]
if filter_dtype == 'int8':
ori_shape_filter = comm.exchange_filter_nc_axis(
ori_format_filter, ori_shape_filter)
shape_filter = comm.get_filter_shape(ori_format_filter, ori_shape_filter)
shape_x = comm.get_shape_out_backprop(ori_format_x, ori_shape_x)
shape_res = comm.get_shape_res(ori_format_res, ori_shape_res)
dilations = comm.get_shape_dilation(ori_format_x, dilations)
comm.check_conv2dbp_input_params(shape_filter, shape_x, shape_res, strides,
pads, dilations, filter_dtype, x_dtype,
res_dtype, kernel_name)
pads = comm.get_padlist(pads, shape_res, strides, shape_filter, dilations)
res = te.lang.cce.conv2d_backprop_input_compute(filter,
x,
shape_filter,
shape_res,
strides,
pads,
dilations,
res_dtype=res_dtype,
tensor_bias=bias,
offset_x=offset_x,
kernel_name=kernel_name)
return res