def elu_grad(grads, activations, y, kernel_name="elu_grad"):
"""
do element-wise elu_grad operation
Parameters:
----------
grads: the dict of gradient input, only support float16, float32
activations: the dict of activation input, only support float16, float32
y : the dict of output
kernel_name : cce kernel name, default value is "cce_elu_grad"
Returns
-------
None
"""
shape_gradient = grads.get("shape")
shape_activation = activations.get("shape")
dtype_gradient = grads.get("dtype")
dtype_activation = activations.get("dtype")
check_shape(shape_gradient, param_name="grads")
check_shape(shape_activation, param_name="activations")
if not operator.eq(shape_gradient, shape_activation):
raise RuntimeError("all input shape must be equal")
shape_gradient, _ = refine_shape_axes(shape_gradient, [])
shape_activation, _ = refine_shape_axes(shape_activation, [])
check_list = ("float16", "float32")
check_dtype(dtype_gradient, check_list, param_name="grads")
check_dtype(dtype_activation, check_list, param_name="activations")
if dtype_gradient.lower() != dtype_activation.lower():
raise RuntimeError("all input dtype must be same")
dtype = dtype_gradient.lower()
data_gradient = tvm.placeholder(shape_gradient,
dtype=dtype,
name="data_gradient")
data_activation = tvm.placeholder(shape_activation,
dtype=dtype,
name="data_activation")
res = elu_grad_compute(data_gradient, data_activation, y, kernel_name)
with tvm.target.cce():
auto_sch = topi.generic.auto_schedule(res)
config = {
"name": kernel_name,
"print_ir": False,
"tensor_list": [data_gradient, data_activation, res]
}
te.lang.cce.cce_build_code(auto_sch, config)
def acos_grad(y, dy, z, kernel_name="acos_grad"):
"""
do element-wise acos_grad operation between two input tensors
Parameters:
----------
y : dict of y, include shape and dtype, dtype support float16, float32
dy : dict of dy, include shape and dtype, dtype support float16, float32
z : dict of z, include shape and dtype, dtype support float16, float32
kernel_name : cce kernel name, default value is "acos_grad"
-------
"""
# get the shape and dtype for input_1,input_2
shape_y = y.get("shape")
shape_dy = dy.get("shape")
dtype = y.get("dtype")
dtype1 = dy.get("dtype")
check_shape(shape_y, param_name="y")
check_shape(shape_dy, param_name="dy")
shape_y, _ = refine_shape_axes(shape_y, [])
shape_dy, _ = refine_shape_axes(shape_dy, [])
# raise runtimeerror if the input paras are invalid
check_list = ("float16", "float32")
check_dtype(dtype, check_list, param_name="y")
check_dtype(dtype1, check_list, param_name="dy")
dtype = dtype.lower()
dtype1 = dtype1.lower()
if not operator.eq(shape_y, shape_dy):
raise RuntimeError(
"acos_grad only support input shape while input_shape1 equals"
" to input_shape2")
if dtype != dtype1:
raise RuntimeError(
"acos_grad only support dtype while input_dtype1 equals"
" to input_dtype2")
shape_y, _ = refine_shape_axes(shape_y, [])
shape_dy, _ = refine_shape_axes(shape_dy, [])
data_y = tvm.placeholder(shape_y, dtype=dtype, name="data1")
data_dy = tvm.placeholder(shape_dy, dtype=dtype, name="data2")
res = acos_grad_compute(data_y, data_dy, z, kernel_name)
with tvm.target.cce():
sch = generic.auto_schedule(res)
config = {"name": kernel_name, "tensor_list": (data_y, data_dy, res)}
te.lang.cce.cce_build_code(sch, config)
def asin_grad(y, dy, z, kernel_name="asin_grad"):
"""
do element-wise asin_grad operation between two input tensors
Parameters:
----------
y : dict of y, include shape and dtype, dtype support float16, float32
dy : dict of dy, include shape and dtype, dtype support float16, float32
z : dict of output
kernel_name : cce kernel name, default value is "asin_grad"
Returns
-------
None
"""
# get the shape and dtype
shape_y = y.get("shape")
shape_dy = dy.get("shape")
dtype_y = y.get("dtype")
dtype_dy = dy.get("dtype")
# kernel name check: should be unique
# check whether the shape is right
check_shape(shape_y, param_name="y")
check_shape(shape_dy, param_name="dy")
if not operator.eq(shape_y, shape_dy):
raise RuntimeError("all input shape must be the same")
shape_y, _ = refine_shape_axes(shape_y, [])
shape_dy, _ = refine_shape_axes(shape_dy, [])
# check whether dtypes are fp16,fp32 and whether they are the same
check_list = ("float16", "float32")
check_dtype(dtype_y, check_list, param_name="y")
check_dtype(dtype_dy, check_list, param_name="dy")
dtype_y = dtype_y.lower()
if dtype_y != dtype_dy.lower():
raise RuntimeError("all input dtype must be same")
# get 2 input tensors: data_y, data_dy
data_y = tvm.placeholder(shape_y, name="data_y", dtype=dtype_y)
data_dy = tvm.placeholder(shape_y, name="data_dy", dtype=dtype_y)
res = asin_grad_compute(data_y, data_dy, z, kernel_name)
with tvm.target.cce():
sch = generic.auto_schedule(res)
config = {"name": kernel_name, "tensor_list": [data_y, data_dy, res]}
te.lang.cce.cce_build_code(sch, config)
def atan_grad(y, dy, z, kernel_name="atan_grad"):
"""
Gradient calculation for atan(x)
Parameters:
----------
y : dict of y, include shape and dtype, dtype support float16, float32
dy : dict of dy, include shape and dtype, dtype support float16, float32
z : dict of output, include shape and dtype
kernel_name : cce kernel name, default value is atan_grad
Algorithm :
----------
forward :
y = atan(x)
backward gradient :
de/dx = dy/dx*de/dy = 1/(1+x^2)*grad
Returns
----------
None
"""
# get the shape and dtype
shape = y.get("shape")
shape_grad = dy.get("shape")
dtype = y.get("dtype")
dtype_grad = dy.get("dtype")
# check whether kernel name is unique
# check whether the shape is right
check_shape(shape, param_name="y")
check_shape(shape_grad, param_name="dy")
if not operator.eq(shape, shape_grad):
raise RuntimeError("all input shape must be the same")
shape, _ = refine_shape_axes(shape, [])
# check whether dtypes are fp16,fp32 and whether they are the same
check_list = ("float16", "float32")
check_dtype(dtype, check_list, param_name="y")
check_dtype(dtype_grad, check_list, param_name="dy")
dtype = dtype.lower()
if dtype != dtype_grad.lower():
raise RuntimeError("all input dtype must be same")
# get 2 input placeholders: data_input, grad
data_input = tvm.placeholder(shape, name="input_data", dtype=dtype)
grad = tvm.placeholder(shape, name="input_grad", dtype=dtype)
# compute the backward gradient
res = atan_grad_compute(data_input, grad, z, kernel_name)
with tvm.target.cce():
sch = generic.auto_schedule(res)
config = {"name": kernel_name,
"tensor_list": [data_input, grad, res]}
te.lang.cce.cce_build_code(sch, config)
def data_format_dim_map(x,
y,
src_format="NHWC",
dst_format="NCHW",
kernel_name="data_format_dim_map"):
"""
Returns the dimension index in the destination data format given the one in.
Parameters
----------
x : A Tensor with each element as a dimension index in source data format.
Must be the following types: `int32`. Must be in the range [-4, 4).
y : Shape and dtype of y, reserved parameter, not used now.
src_format : An optional `string`. Defaults to `"NHWC"`. source data format.
dst_format : An optional `string`. Defaults to `"NCHW"`. destination data format.
kernel_name : CCE kernel name, default value is "data_format_dim_map" (optional).
Returns
-------
None
"""
shape_input = x.get("shape")
dtype_input = x.get("dtype")
# check kernel name, shape, size, dtype
check_shape(shape_input, param_name="x")
shape_input, _ = refine_shape_axes(shape_input, [])
check_list = ("int32", )
dtype_input = dtype_input.lower()
check_dtype(dtype_input, check_list, param_name="x")
# check length of format
if len(src_format) != 4:
raise ValueError(
"source format must of length 4, received src_format = %s" %
src_format)
if len(dst_format) != 4:
raise ValueError(
"destination format must of length 4, received dst_format = %s" %
dst_format)
# get data and compute
data_input = tvm.placeholder(shape_input,
dtype=dtype_input,
name="data_input")
res = _data_format_dim_map_compute(data_input, y, src_format, dst_format,
kernel_name)
with tvm.target.cce():
sch = topi.generic.auto_schedule(res)
config = {
"name": kernel_name,
"print_ir": False,
"tensor_list": (data_input, res),
"bool_storage_as_1bit": False
}
te.lang.cce.cce_build_code(sch, config)
def tensor_equal(input_x, input_y, output_z, kernel_name="tensor_equal"):
'''
True if two tensors have the same size and elements, False otherwise
:param input_x: dict
input tenser x
:param input_y: dict
input tensor y
:param kernel_name: str
kernel name, default value is "tensor_equal"
:return: none
'''
shape_x = input_x.get("shape")
dtype_x = input_x.get("dtype")
shape_y = input_y.get("shape")
dtype_y = input_y.get("dtype")
check_shape(shape_x)
check_shape(shape_y)
check_list = ("float16", "float32", "int32", "int8", "uint8")
check_dtype(dtype_x, check_list)
check_dtype(dtype_y, check_list)
shape_x = list(shape_x)
shape_x, _ = refine_shape_axes(shape_x, [])
data_input_x = tvm.placeholder(shape_x, name="data_input_x", dtype=dtype_x)
shape_y, _ = refine_shape_axes(shape_y, [])
data_input_y = tvm.placeholder(shape_y, name="data_input_y", dtype=dtype_y)
# use vsub method compute equal result
res = tensor_equal_compute_use_sub(data_input_x, data_input_y, output_z,
kernel_name)
with tvm.target.cce():
schedule = generic.auto_schedule(res)
config = {
"name": kernel_name,
"tensor_list": [data_input_x, data_input_y, res],
"bool_storage_as_1bit": False
}
te.lang.cce.cce_build_code(schedule, config)
def assign_sub(var, value, out, kernel_name='assign_sub'):
"""
Update var by subtracting value from it.
Parameters:
----------
var : dict
dict of input_var, include shape and dtype,
dtype support int8, uint8, int32, float16, float32
value : dict
dict of input_value, include shape and dtype,
dtype support int8, uint8, int32, float16, float32.
Must have the same shape and dtype as input_var
out : dict
dict of out
kernel_name : str
cce kernel name, default value is "assign_sub"
Returns
-------
None
"""
# get the shape and dtype
shape_var = var.get("shape")
shape_value = value.get("shape")
dtype_var = var.get("dtype")
dtype_value = value.get("dtype")
# kernel name check: should be unique
# check whether the shape is right
check_shape(shape_var, param_name="var")
check_shape(shape_value, param_name="value")
if not operator.eq(shape_var, shape_value):
raise RuntimeError("all input shape must be the equal")
# check whether dtypes are fp16, fp32, int8, uint8, int32
# and whether they are the same
check_list = ("float16", "float32", "int8", "uint8", "int32")
check_dtype(dtype_var, check_list, param_name="var")
check_dtype(dtype_value, check_list, param_name="value")
dtype_var = dtype_var.lower()
dtype_value = dtype_value.lower()
if dtype_var != dtype_value:
raise RuntimeError("all input dtype must be same")
shape, _ = refine_shape_axes(shape_var, [])
data_var = tvm.placeholder(shape, dtype=dtype_var, name='data_var')
data_value = tvm.placeholder(shape, dtype=dtype_value, name='data_value')
sch, res = _assign_sub_compute(data_var, data_value, out, kernel_name)
with set_bool_storage_config():
tvm.build(sch, [data_var, data_value, res], "cce", name=kernel_name)
def asinh(input_x, output_y, kernel_name="asinh"):
"""
algrithm: asinh(x) = log(x + sqrt(x^2 + 1))
Parameters
----------
input_x: the dict of input_x, only support float16, float32
output_y : the dict of output_y
kernel_name : cce kernel name, default value is "asinh"
Returns
-------
None
"""
shape_input = input_x.get("shape")
dtype_input = input_x.get("dtype")
check_shape(shape_input, param_name="input_x")
shape_input, _ = refine_shape_axes(shape_input, [])
check_list = ("float16", "float32")
check_dtype(dtype_input, check_list, param_name="input_x")
inp_dtype = dtype_input.lower()
shape_input = (functool_reduce(lambda x, y: x * y, shape_input), )
data_input = tvm.placeholder(shape_input,
dtype=inp_dtype,
name="data_input")
with tvm.target.cce():
if tbe_platform.cce_conf.api_check_support("te.lang.cce.vlog",
"float32") or not \
tbe_platform.cce_conf.api_check_support("te.lang.cce.vrec",
"float32"):
res = asinh_compute_cloud(data_input, output_y, kernel_name)
else:
res = asinh_compute_mini(data_input, output_y, kernel_name)
sch = generic.auto_schedule(res)
config = {
"name": kernel_name,
"tensor_list": [data_input, res],
"bool_storage_as_1bit": False
}
te.lang.cce.cce_build_code(sch, config)
def asin(x, y, kernel_name="asin"):
"""
----------
asin(x) = | arcsin(sqrt(1-x^2)) - HALF_PI, x belongs to (-1, 2^(-0.5))
| the 15th order taylor expansion, x belongs to (-2^(-0.5), 2^(-0.5))
| HALF_PI - arcsin(sqrt(1-x^2)), x belongs to (2^(-0.5), 1)
Parameters:
----------
x : the placeholder of data input
y : the dict of output
kernel_name : cce kernel name, default value is "asin"
Returns : None
-------
"""
shape_input = x.get("shape")
dtype_input = x.get("dtype")
check_shape(shape_input, param_name="x")
shape_input, _ = refine_shape_axes(shape_input, [])
check_list = ("float16", "float32")
check_dtype(dtype_input, check_list, param_name="x")
inp_dtype = dtype_input.lower()
data_input = tvm.placeholder(shape_input,
dtype=inp_dtype,
name="data_input")
res = asin_compute(data_input, y, kernel_name)
with tvm.target.cce():
auto_sch = topi.generic.auto_schedule(res)
config = {
"name": kernel_name,
"print_ir": False,
"tensor_list": [data_input, res],
"bool_storage_as_1bit": False
}
te.lang.cce.cce_build_code(auto_sch, config)
def atanh(x, y, kernel_name="atanh"):
"""
Algrithm: atanh(x) = atanh
Parameters
----------
Algorithm: atanh
Parameters:
x: the dict of input data, only support float16, float32.
y: the dict of output
kernel_name: cce kernel name, default value is "atanh".
Returns
-------
None
"""
shape = x.get("shape")
dtype = x.get("dtype")
check_shape(shape, param_name="x")
shape, _ = refine_shape_axes(shape, [])
check_list = ("float16", "float32")
check_dtype(dtype, check_list, param_name="x")
dtype = dtype.lower()
input_data = tvm.placeholder(shape, dtype, "input_data")
with tvm.target.cce():
res = atanh_compute(input_data, y, kernel_name)
sch = generic.auto_schedule(res)
config = {"name": kernel_name,
"tensor_list": [input_data, res],
"print_ir": False,
"bool_storage_as_1bit": False
}
te.lang.cce.cce_build_code(sch, config)
def atan(x, y, kernel_name="atan"):
"""
Algorithm: atan
----------------------------------
Parameters:
x: the dict of input data, only support float16, float32.
y: the dict of output
kernel_name: cce kernel name, default value is "atan".
----------------------------------
Returns:
None
"""
shape = x.get("shape")
dtype = x.get("dtype")
check_shape(shape, param_name="x")
shape, _ = refine_shape_axes(shape, [])
check_list = ("float16", "float32")
check_dtype(dtype, check_list, param_name="x")
with tvm.target.cce():
dtype = dtype.lower()
input_data = tvm.placeholder(shape, dtype=dtype, name="input_data")
res = atan_compute(input_data, y, kernel_name)
res = te.lang.cce.cast_to(res, dtype)
auto_sch = topi.generic.auto_schedule(res)
config = {
"name": kernel_name,
"print_ir": False,
"tensor_list": (input_data, res)
}
te.lang.cce.cce_build_code(auto_sch, config)
def acosh(input_data, output_res, kernel_name="acosh"):
"""
calculating data's acosh,y= log(x+sqrt(x^(2)-1))
Parameters
----------
input_data: the dict of input, only support float16, float32
output_res : the dict of output
kernel_name : cce kernel name, default value is "cce_acosh"
Returns
-------
None
"""
shape_input = input_data.get("shape")
dtype_input = input_data.get("dtype")
check_shape(shape_input, param_name="input_data")
check_list = ("float16", "float32")
check_dtype(dtype_input, check_list, param_name="input_data")
shape_input, _ = refine_shape_axes(shape_input, [])
input_dtype = dtype_input.lower()
data = tvm.placeholder(shape_input, dtype=input_dtype, name="data_input")
res = acosh_compute(data, output_res, kernel_name)
with tvm.target.cce():
sch = topi.generic.auto_schedule(res)
config = {"name": kernel_name,
"print_ir": False,
"tensor_list": (data, res),
"bool_storage_as_1bit": False}
te.lang.cce.cce_build_code(sch, config)
def bessel_i1e(x, y, kernel_name="bessel_i1e"):
"""
Algrithm: calculating data's bessel
Parameters
----------
x: the dict of input, only support float16, float32
y : the dict of output
kernel_name : cce kernel name, default value is "bessel_i1e"
Returns
-------
None
"""
shape_input = x.get("shape")
dtype_input = x.get("dtype")
check_shape(shape_input, param_name="x")
shape_input, _ = refine_shape_axes(shape_input, [])
check_list = ("float16", "float32")
check_dtype(dtype_input, check_list, param_name="x")
input_dtype = dtype_input.lower()
data = tvm.placeholder(shape_input, dtype=input_dtype, name="data_input")
res = bessel_i1e_compute(data, y, kernel_name)
with tvm.target.cce():
sch = topi.generic.auto_schedule(res)
config = {"name": kernel_name,
"print_ir": False,
"tensor_list": (data, res),
"bool_storage_as_1bit": False}
te.lang.cce.cce_build_code(sch, config)
def mse_loss(predict, label, y, reduction='mean', kernel_name="mse_loss"):
'''
calculating data
sum = (predict_n - label_n)^2
if reduction == sum: res = sum output a scalal
if reduction == mean: res == sum/total_number_of_tensor output a scalar
if reduction == none: res == (predict_n - label_n)^2 output a tensor
:param predict: dict
shape and dtype of tensor predict
:param label: dict
shape and dtype of tensor real label,
should be same shape and dtype as predict
:param y: dict
shape and dtype of output, loss result after compute
:param reduction: str
Specifies the reduction to apply to the output:'none' | 'mean' | 'sum'
Default: 'mean'
'none': no reduction will be applied,
'mean': the sum of the output will be divided by the number
of elements in the output
'sum': the output will be summed. Note: size_average and reduce
are in the process of being deprecated
and in the meantime, specifying either of those
two args will override reduction.
:param kernel_name: str
kernel name, default value is "mse_loss"
:return: none
'''
predict_shape = predict.get("shape")
predict_dtype = predict.get("dtype")
predict_dtype_lower = predict_dtype.lower()
label_shape = label.get("shape")
label_dtype = label.get("dtype")
label_dtype_lower = label_dtype.lower()
# check dtype
dtype_list = ("float16", "float32")
op_utils.check_dtype(predict_dtype, dtype_list)
op_utils.check_dtype(predict_dtype, dtype_list)
# check shape
op_utils.check_shape(predict_shape)
op_utils.check_shape(label_shape)
# check kernel_name
util.check_kernel_name(kernel_name)
predict_size, _ = op_utils.refine_shape_axes(predict_shape, [])
data_predict = tvm.placeholder(predict_size, dtype=predict_dtype_lower, name="data_predict")
label_size, _ = op_utils.refine_shape_axes(label_shape, [])
data_label = tvm.placeholder(label_size, dtype=label_dtype_lower, name="data_label")
if predict_size != label_size:
raise RuntimeError("predict tensor size don't match label tensor")
if reduction not in ("mean", "sum", "none"):
raise RuntimeError("reduction type should in mean/sum/none")
res = mse_loss_compute(data_predict, data_label, reduction, kernel_name)
with tvm.target.cce():
schedule = generic.auto_schedule(res)
config = {"print_ir": False,
"name": kernel_name,
"tensor_list": [data_predict, data_label, res]}
te.lang.cce.cce_build_code(schedule, config)
def approximate_equal(input_x,
input_y,
output_z,
tolerance=1e-5,
kernel_name="approximate_equal"):
"""
abs(x-y) <= tolerance
Parameters
----------
input_x : dict, include shape and dtype, support fp16 and fp32
shape of tensors, assume src_shape equals dst_shape
input_y : dict, include shape and dtype, support fp16 and fp32
shape of tensors, assume src_shape equals dst_shape
output_z : dict, include shape and dtype, reserve
tolerance: default 1e-5
kernel_name : str
cce kernel name, default value is "approximate_equal"
Returns
------
None
"""
shape_x = input_x.get("shape")
shape_y = input_y.get("shape")
in_dtype = input_x.get("dtype")
in_y_dtype = input_y.get("dtype")
if tolerance < 0:
raise RuntimeError("tolerance should >= 0")
# check shape
if not operator.eq(shape_x, shape_y):
raise RuntimeError("all input shape must same")
check_shape(shape_x, param_name="input_x")
shape_x, _ = refine_shape_axes(shape_x, [])
shape_y, _ = refine_shape_axes(shape_y, [])
# check input tensor data_type
check_list = ("float16", "float32")
check_dtype(in_dtype, check_list, param_name="input_x")
check_dtype(in_y_dtype, check_list, param_name="input_y")
in_dtype = input_x.get("dtype").lower()
in_y_dtype = input_y.get("dtype").lower()
if not operator.eq(in_dtype, in_y_dtype):
raise RuntimeError("all input type must same.")
in_data_x = tvm.placeholder(shape_x, name="shape_x", dtype=in_dtype)
in_data_y = tvm.placeholder(shape_y, name="shape_y", dtype=in_dtype)
res = approximate_equal_compute(in_data_x, in_data_y, output_z, tolerance,
kernel_name)
with tvm.target.cce():
auto_sch = topi.generic.auto_schedule(res)
config = {
"name": kernel_name,
"tensor_list": [in_data_x, in_data_y, res],
"bool_storage_as_1bit": False
}
te.lang.cce.cce_build_code(auto_sch, config)
