def _taylor_compute(data_x, x_square=None):
"""
do arcsinx compute use the 15th order taylor expansion when 0 <= x <= BOUNDARY_1
asin(x) = x + 1/6*x^3 + 3/40*x^5 + 5/112*x^7 + ... + 13!!/(14!!*15)*x^15
Parameters:
----------
data_x : the placeholder of data input
x_square : the placeholder of the square of data_x
Returns : A Tensor. Has the same type as data.
-------
"""
if x_square is None:
x_square = te.lang.cce.vmul(data_x, data_x)
res = te.lang.cce.vmuls(x_square,
tvm.const(COEF[TAYLOR_COUNT], x_square.dtype))
for temp in reversed(range(TAYLOR_COUNT)):
res = te.lang.cce.vadds(res, tvm.const(COEF[temp], x_square.dtype))
if temp == 0:
res = te.lang.cce.vmul(res, data_x)
else:
res = te.lang.cce.vmul(x_square, res)
return res
def _newton_taylor_iter(input_x, input_y, input_z):
"""
do element-wise Newton compute
z(n+1) = z(n) - (e^(z(n)*x(n)^-1) - y(n))/x(n)^-1*e^(z(n)*x(n)^-1)
Parameters:
----------
input_x: TVM tensor, the placeholder of input_x
input_y: TVM tensor, the placeholder of input_y
input_z: start value of Newton iteration
Returns : A Tensor. Has the same type as input_z.
-------
"""
#Newton begin:z(n+1) = z(n) - x(n) + x(n)*y(n)*e^(-z(n)*x(n)^-1)
input_x_mul = te.lang.cce.vmuls(input_x,
tvm.const(SCALAR_NEG_ONE, "float32"))
newton_taylor = te.lang.cce.vadd(input_x_mul, input_z)
input_xy = te.lang.cce.vmul(input_x, input_y)
input_x_rec = te.lang.cce.vrec(input_x)
input_x_res = te.lang.cce.vmuls(input_x_rec,
tvm.const(SCALAR_NEG_ONE, "float32"))
input_z_mul = te.lang.cce.vmul(input_x_res, input_z)
input_z_taylor = _exp_taylor_compute(input_z_mul)
input_z_res = te.lang.cce.vmul(input_z_taylor, input_xy)
newton_taylor = te.lang.cce.vadd(newton_taylor, input_z_res)
return newton_taylor
def _get_pd_x_front_nz(data, param_nz, cast_dtype):
"""
compute front part of pd_x according to data, params and shape_x
"""
pd_xl = _get_pd_xl_nz(data, param_nz)
pd_var, var_elta_2, sub_x_mean = _get_pd_var_nz(data, param_nz, pd_xl,
cast_dtype)
pd_mean = _get_pd_mean_nz(param_nz, pd_xl, pd_var, var_elta_2,
sub_x_mean, cast_dtype)
var_elta_2_cast = _broadcast_nz(var_elta_2, param_nz.get("shape_x_nz"))
pd_x_1 = te.lang.cce.vmul(var_elta_2_cast, pd_xl)
pdx2_broad = _broadcast_nz(pd_var, param_nz.get("shape_x_nz"))
pdx2_mul = te.lang.cce.vmul(pdx2_broad, sub_x_mean)
pd_x_2 = \
te.lang.cce.vmuls(pdx2_mul,
tvm.const((2*(param_nz.get("mean_nz_num")**(-1))),
dtype=cast_dtype))
pd_x_3 = \
te.lang.cce.vmuls(pd_mean,
tvm.const((param_nz.get("mean_nz_num")**(-1)),
dtype=cast_dtype))
return pd_x_1, pd_x_2, pd_x_3
def l1_loss_grad_compute(grad_out,
predict,
target,
y,
reduction="mean",
kernel_name="l1_loss_grad"):
predict_dtype = predict.dtype.lower()
zero_tensor = te.lang.cce.vmuls(predict, tvm.const(0, dtype=predict_dtype))
one_tensor = te.lang.cce.vadds(zero_tensor,
tvm.const(1, dtype=predict_dtype))
neg_one_tensor = te.lang.cce.vadds(zero_tensor,
tvm.const(-1, dtype=predict_dtype))
# if predict is equal or bigger than target, the sign will be given 1; else -1
sign = te.lang.cce.vcmpsel(predict, target, "gt", one_tensor,
neg_one_tensor)
# rectify sign to 0 when predict equal to target
sign = te.lang.cce.vcmpsel(predict, target, "eq", zero_tensor, sign)
grad_shape = te.lang.cce.util.shape_to_list(grad_out.shape)
n = reduce(lambda x, y: x * y, grad_shape)
norm = grad_out
# if choose "mean", grad_out should divide over n
if reduction == "mean":
norm = te.lang.cce.vmuls(norm, tvm.const(1 / n, dtype=predict_dtype))
# chain multiplication to get the gradient of L1 with respect to weights(grad_out)
res = te.lang.cce.vmul(sign, norm)
return res
def fake_quant_perchannel_compute(x, min_val, max_val, y, quant_min, quant_max,
kernel_name="fake_quant_perchannel"):
"""FakeQuantPerChannel"""
x_shape = te.lang.cce.util.shape_to_list(x.shape)
minmax_shape = te.lang.cce.util.shape_to_list(min_val.shape)
quant_min = tvm.const(quant_min, x.dtype)
quant_max = tvm.const(quant_max, x.dtype)
quant_min = te.lang.cce.broadcast(quant_min, minmax_shape, x.dtype)
quant_max = te.lang.cce.broadcast(quant_max, minmax_shape, x.dtype)
# CalNudge(NudgeMinMax)
scale = te.lang.cce.vdiv(te.lang.cce.vsub(
max_val, min_val), te.lang.cce.vsub(quant_max, quant_min))
zp_from_min = te.lang.cce.vsub(quant_min, te.lang.cce.vdiv(min_val, scale))
# Nudge zero point
nudge_zp_ = te.lang.cce.vmin(
quant_max, te.lang.cce.vmax(quant_min, zp_from_min))
nudge_zp = te.lang.cce.floor(te.lang.cce.vadds(nudge_zp_, 0.5))
nudge_min = te.lang.cce.vmul(te.lang.cce.vsub(quant_min, nudge_zp), scale)
nudge_max = te.lang.cce.vmul(te.lang.cce.vsub(quant_max, nudge_zp), scale)
# FakeQuant
nudge_min_b = te.lang.cce.broadcast(nudge_min, x_shape)
nudge_max_b = te.lang.cce.broadcast(nudge_max, x_shape)
scale_b = te.lang.cce.broadcast(scale, x_shape)
input_x = te.lang.cce.vmin(nudge_max_b, te.lang.cce.vmax(nudge_min_b, x))
nudge_input_ = te.lang.cce.vdiv(
te.lang.cce.vsub(input_x, nudge_min_b), scale_b)
nudge_input = te.lang.cce.floor(te.lang.cce.vadds(nudge_input_, 0.5))
res = te.lang.cce.vadd(te.lang.cce.vmul(nudge_input, scale_b), nudge_min_b)
return res
def _trans_input_shape(self, axis):
"""
trans the input shape into three dimensions (left, mid, right) and
get the range of different dims of the input shape.
Returns:
-------
x_reshape: new input shape of format with (left, mid, right)
left_range:left dim range
right_range:right dim range
"""
real_axis = axis + len(self.dims_var) if axis < 0 else axis
left_dim = tvm.const(1)
left_upper = 1
for idx in range(real_axis):
left_dim *= self.dim_vars[idx]
left_upper *= self.dim_bounds[idx][1]
self.left_range = (1, left_upper)
right_dim = tvm.const(1)
right_upper = 1
for idx in range(real_axis + 1, len(self.dim_vars)):
right_dim *= self.dim_vars[idx]
right_upper *= self.dim_bounds[idx][1]
self.right_range = (1, right_upper)
self.x_reshape = (left_dim, self.dim_vars[real_axis], right_dim)
def _get_pd_var_front(data, cast_dtype):
"""
compute front part of pd_var according to data_variance
Parameters
----------
data: dict
placeholders after cast
Returns
-------
pd_var_1: tvm.tensor
np.power((data_variance + EPSLON), (-1.5))
var_elta_2: tvm.tensor
np.power((data_variance + EPSLON), (-0.5))
"""
var_elta = te.lang.cce.vadds(data.get("data_variance"),
tvm.const(EPSLON, dtype=cast_dtype))
var_elta_log = te.lang.cce.vlog(var_elta)
var_elta_mul = te.lang.cce.vmuls(var_elta_log,
tvm.const(-0.5, dtype=cast_dtype))
var_elta_2 = te.lang.cce.vexp(var_elta_mul)
pdvar1_mul = te.lang.cce.vmul(var_elta_2, var_elta_2)
pd_var_1 = te.lang.cce.vmul(pdvar1_mul, var_elta_2)
return pd_var_1, var_elta_2
def _less_compare_float32(data_x, data_y):
"""
Compare data_x and data_y to determine whether data_x is less than data_y.
If the element in data_x is less than in data_y, then return 1,
else return 0.
max num of float32 is 2**126
but cce can only support 2**62, so use 62/62/2 to adaptor 126
(2**(-126))*(2**(62))*(2**(62))*(2**(2)) = 1
so min_value*max_value*max_value*factor_value = 1
"""
shape_inputs = te.lang.cce.util.shape_to_list(data_x.shape)
min_value = tvm.const(2 ** (-126), dtype=D_TYPE)
max_value = tvm.const(2 ** 62, dtype=D_TYPE)
factor_value = tvm.const(2 ** 2, dtype=D_TYPE)
if api_check_support("te.lang.cce.vmaxs", data_x.dtype):
res_sub = te.lang.cce.vsub(data_y, data_x)
res_min = te.lang.cce.vmins(res_sub, min_value)
res_max = te.lang.cce.vmaxs(res_min, tvm.const(0, dtype=D_TYPE))
else:
data_zero = te.lang.cce.vmuls(data_x, 0)
min_value_tensor = te.lang.cce.vadds(data_zero, min_value)
res_sub = te.lang.cce.vsub(data_y, data_x)
res_min = te.lang.cce.vmin(res_sub, min_value_tensor)
res_max = te.lang.cce.vmax(res_min, data_zero)
res_max_mul = te.lang.cce.vmuls(res_max, max_value)
res_max_mul_max = te.lang.cce.vmuls(res_max_mul, max_value)
res = te.lang.cce.vmuls(res_max_mul_max, factor_value)
return res
def rsqrt_grad_compute(input_y, input_dy, output_z, kernel_name="rsqrt_grad"):
"""
compute for rsqrt_grad
Parameters
----------
input_y: TVM tensor
the placeholder of input_y
input_dy: TVM tensor
the placeholder of input_dy
output_z: dict
dict info of output_z
kernel_name: str
cce kernel name, default value is "rsqrt_grad"
Returns
-------
res: TVM tensor
the result of compute
"""
dtype_input_y = input_y.dtype
rsqrt_const = tvm.const(SCALAR, dtype=dtype_input_y)
if dtype_input_y in ("int8", "float16"):
rsqrt_const = tvm.const(SCALAR, dtype="float32")
input_y = te.lang.cce.cast_to(input_y, "float32")
input_dy = te.lang.cce.cast_to(input_dy, "float32")
res_vmul = te.lang.cce.vmul(input_y, input_y)
res_vmul1 = te.lang.cce.vmul(res_vmul, input_y)
res_vmul2 = te.lang.cce.vmul(res_vmul1, input_dy)
res = te.lang.cce.vmuls(res_vmul2, rsqrt_const)
if dtype_input_y in ("int8", "int32", "float16"):
res = te.lang.cce.cast_to(res, dtype_input_y, f1628IntegerFlag=True)
return res
def sigmoid_compute(input_x):
"""
calculating sigmoid
"""
data_input = input_x
dtype = input_x.dtype
exp_support = cce.cce_conf.api_check_support(
"te.lang.cce.vexp", "float32")
mul_support = cce.cce_conf.api_check_support(
"te.lang.cce.vmuls", "float32")
if dtype == "float32" and not mul_support:
error_manager_vector.raise_err_specific_reson("DynamicLSTM",
"Input dtype only support float16 while input dtype is float32")
const_num_neg_one = tvm.const(-1, dtype=dtype)
const_num_one = tvm.const(1, dtype=dtype)
tmp_negative = te.lang.cce.vmuls(data_input, const_num_neg_one)
if dtype == "float32" and not exp_support:
tmp_negative = te.lang.cce.cast_to(tmp_negative, "float16")
tmp_exp = te.lang.cce.vexp(tmp_negative)
if dtype == "float32" and not exp_support:
tmp_exp = te.lang.cce.cast_to(tmp_exp, "float32")
tmp_sum = te.lang.cce.vadds(tmp_exp, const_num_one)
if dtype == "float32":
inp_shape = tmp_sum.shape
tensor_one = te.lang.cce.broadcast(tvm.const(1, dtype), inp_shape)
res = te.lang.cce.vdiv(tensor_one, tmp_sum)
else:
res = te.lang.cce.vrec(tmp_sum)
return res
def _negative_compute(input_x, input_y):
"""
compute result of pow when data_x is less than 0,
use [-2 * (|y| % 2) + 1] * exp(y * ln|x|)
"""
dtype = input_x.dtype
shape = input_x.shape
abs_value = te.lang.cce.vabs(input_y)
if not tbe_platform.cce_conf.api_check_support("te.lang.cce.vmod",
"float32"):
dtype = "float16"
abs_value = te.lang.cce.cast_to(abs_value, "float16")
data_two = te.lang.cce.broadcast(tvm.const(2, dtype), shape, dtype)
mod_value = te.lang.cce.vmod(abs_value, data_two)
mul_value = te.lang.cce.vmuls(mod_value, tvm.const(-2, dtype))
add_value = te.lang.cce.vadds(mul_value, tvm.const(1, dtype))
if tbe_platform.cce_conf.api_check_support("te.lang.cce.vexp", "float32"):
add_value = te.lang.cce.cast_to(add_value, "float32")
abs_data_x = te.lang.cce.vabs(input_x)
log_value = te.lang.cce.vlog(abs_data_x)
mul_value = te.lang.cce.vmul(input_y, log_value)
exp_value = te.lang.cce.vexp(mul_value)
res = te.lang.cce.vmul(add_value, exp_value)
return res
def _compute(data_input, shape):
"""
Algrithm: atanh(x) = 0.5*log((1+x)/(1-x))
Parameters
----------
data_input: the placeholder of data input
shape: the shape of data_input
Returns
-------
data_res : return of atanh
"""
data_1_sum_x = te.lang.cce.vadds(data_input, tvm.const(CONST_ONE,
data_input.dtype))
data_sub_x = te.lang.cce.vmuls(data_input, tvm.const(CONST_NEG_ONE,
data_input.dtype))
data_1_sub_x = te.lang.cce.vadds(data_sub_x, tvm.const(CONST_ONE,
data_input.dtype))
data_x_mul = te.lang.cce.vdiv(data_1_sum_x, data_1_sub_x)
data_x_log = te.lang.cce.vlog(data_x_mul, 1)
data_res = te.lang.cce.vmuls(data_x_log, tvm.const(CONST_HALF,
data_input.dtype))
return data_res
def _compute_process(var, m, lr_broad, alpha_broad, sign_decay_broad,
beta_broad, grad):
"""
calculate
m_t <- beta1 * m_{t-1} + (1 - beta1) * g
update <- (alpha + sign_decay * sign(g) *sign(m)) * g
variable <- variable - lr_t * update
Parameters:
----------
var: the dict of var, support float16, float32
m: the dict of m, support float16, float32
lr: the dict of lr, support float16, float32
alpha: the dict of alpha, support float16, float32
sign_decay: the dict of sign_decay, support float16, float32
beta: the dict of beta, support float16, float32
grad: the dict of grad, support float16, float32
Returns
-------
the new value of var and out
the output
"""
m_out = _update_m(m, beta_broad, grad)
sign_gm = te.lang.cce.vmul(_sign_compute(grad), _sign_compute(m_out))
decay_gm = te.lang.cce.vmul(sign_gm, sign_decay_broad)
var_out = _update_var(decay_gm, alpha_broad, lr_broad, grad, var)
output_data = te.lang.cce.vadds(var_out, tvm.const(CONST_ZERO, "float32"))
m_output_data = te.lang.cce.vadds(m_out, tvm.const(CONST_ZERO, "float32"))
return m_out, var_out, output_data, m_output_data
def _sigmoid_compute(input_x):
"""
calculating sigmoid
"""
data_input = input_x
dtype = input_x.dtype
exp_support = tbe_platform.cce_conf.api_check_support(
"te.lang.cce.vexp", "float32")
mul_support = tbe_platform.cce_conf.api_check_support(
"te.lang.cce.vmuls", "float32")
if dtype == "float32" and not mul_support:
error_manager_vector.raise_err_check_params_rules(
"DynamicGRU", 'vmuls should support float32', 'mul_support',
str(mul_support))
const_num_neg_one = tvm.const(-1, dtype=dtype)
const_num_one = tvm.const(1, dtype=dtype)
tmp_negative = tbe.vmuls(data_input, const_num_neg_one)
if dtype == "float32" and not exp_support:
tmp_negative = tbe.cast_to(tmp_negative, "float16")
tmp_exp = tbe.vexp(tmp_negative)
if dtype == "float32" and not exp_support:
tmp_exp = tbe.cast_to(tmp_exp, "float32")
tmp_sum = tbe.vadds(tmp_exp, const_num_one)
if dtype == "float32":
inp_shape = tmp_sum.shape
tensor_one = tbe.broadcast(tvm.const(1, dtype), inp_shape)
res = tbe.vdiv(tensor_one, tmp_sum)
else:
res = tbe.vrec(tmp_sum)
return res
def less_compute(input_x, input_y, output_z, kernel_name="less"):
"""
if x is less than y, then return 1, else return 0.
Parameters:
----------
input_x: TVM tensor
the placeholder of first input data
input_y: TVM tensor
the placeholder of second input data
output_x: dict
shape and dtype of output, should be broadcast shape and type as input
kernel_name: str
cce kernel name, default value is less
Returns
-------
the result
"""
shape_x = te.lang.cce.util.shape_to_list(input_x.shape)
shape_y = te.lang.cce.util.shape_to_list(input_y.shape)
shape_x, shape_y, shape = broadcast_shapes(shape_x,
shape_y,
param_name_input1="input_x",
param_name_input2="input_y")
cce_product = tbe_platform.cce_conf.get_soc_spec("SOC_VERSION")
dtype = input_x.dtype
if dtype in ("uint8", "int8"):
input_x = te.lang.cce.cast_to(input_x, "float16")
input_y = te.lang.cce.cast_to(input_y, "float16")
dtype = "float16"
if dtype == "float32":
# minimun num of float32 2**(-126)
data_min = te.lang.cce.broadcast(tvm.const(2**(-126), dtype=dtype),
shape, dtype)
elif dtype == "float16" and cce_product not in ("Ascend910", "Ascend710"):
# minimun num of float16 2**(-24)
data_min = te.lang.cce.broadcast(tvm.const(2**(-24), dtype=dtype),
shape, dtype)
elif dtype == "float16" and cce_product in ("Ascend910", "Ascend710"):
input_x = te.lang.cce.cast_to(input_x, "float32")
input_y = te.lang.cce.cast_to(input_y, "float32")
dtype = "float32"
data_min = te.lang.cce.broadcast(tvm.const(2**(-126), dtype=dtype),
shape, dtype)
elif dtype == "int32" and cce_product not in ("Ascend910", "Ascend710"):
data_min = te.lang.cce.broadcast(tvm.const(1, dtype=dtype), shape,
dtype)
else:
input_x = te.lang.cce.cast_to(input_x, "float32")
input_y = te.lang.cce.cast_to(input_y, "float32")
dtype = "float32"
data_min = te.lang.cce.broadcast(tvm.const(2**(-126), dtype=dtype),
shape, dtype)
input_x = te.lang.cce.broadcast(input_x, shape)
input_y = te.lang.cce.broadcast(input_y, shape)
return _less_compare((input_x, input_y), shape, dtype, data_min)
def _taylor_compute(data):
"""
algrithm: log(x) = ((((0.2x - 0.25)x + 0.33333)x - 0.5)x + 1)x
Parameters
----------
data: input tensor that we want to calculate log
Returns
-------
None
"""
# 0.2x - 0.25
taylor_five = te.lang.cce.vmuls(data, tvm.const(CONST_ONE_FIVE, "float32"))
taylor_four_1 = te.lang.cce.vadds(taylor_five,
tvm.const(CONST_ONE_FOUR_NEG, "float32"))
# (0.2x - 0.25)x + 0.33333
taylor_four_2 = te.lang.cce.vmul(taylor_four_1, data)
taylor_three_1 = te.lang.cce.vadds(taylor_four_2,
tvm.const(CONST_ONE_THREE, "float32"))
# ((0.2x - 0.25)x + 0.33333)x - 0.5
taylor_three_2 = te.lang.cce.vmul(taylor_three_1, data)
taylor_two_1 = te.lang.cce.vadds(
taylor_three_2, tvm.const(CONST_NEWTON_FACTOR_NEG, "float32"))
# (((0.2x - 0.25)x + 0.33333)x - 0.5)x+1
taylor_two_2 = te.lang.cce.vmul(taylor_two_1, data)
taylor_one = te.lang.cce.vadds(taylor_two_2,
tvm.const(CONST_ONE, "float32"))
# ((((0.2x - 0.25)x + 0.33333)x - 0.5)x + 1)x
taylor = te.lang.cce.vmul(taylor_one, data)
return taylor
def _cosh_taylor_compute(data):
"""
Calculate cosh = 1 + x^2( 1/2! + x^2( 1/4! + x^2/6!))
Parameters:
----------
data : the placeholder of data input
Returns
-------
A Tensor represents cosh(data). Has the same type as data.
"""
# x^2 / 6!
pow_2 = te.lang.cce.vmul(data, data)
pow_2_div = te.lang.cce.vmuls(pow_2, tvm.const(TAYLOR_SIXTH, data.dtype))
# 1/4! + x^2 / 6!
pow_2_plus = te.lang.cce.vadds(pow_2_div,
tvm.const(TAYLOR_FOURTH, data.dtype))
# 1/2! + x^2( 1/4! + x^2/6!)
pow_4 = te.lang.cce.vmul(pow_2_plus, pow_2)
pow_4_plus = te.lang.cce.vadds(pow_4, tvm.const(TAYLOR_SECOND, data.dtype))
# 1 + x^2( 1/2! + x^2( 1/4! + x^2/6!))
pow_6 = te.lang.cce.vmul(pow_4_plus, pow_2)
res = te.lang.cce.vadds(pow_6, tvm.const(NUM_ONE, data.dtype))
return res
def _compare_value_float(x_data, y_data):
"""
The input data type of the function only support float;
The return value of the function: if x_data >= y_data return 1; else return 0.
"""
# The smallest positive subnormal number of float32 is 2**(-126)
min_value = tvm.const(2**(-126), dtype="float32")
# (2**(-126))*(2**(62))*(2**(62))*(2**(2)) = 1
# so min_value*max_value*max_value*max_value_1 = 1
max_value = tvm.const(2**(62), dtype="float32")
max_value_1 = tvm.const(2**(2), dtype="float32")
data_zero = te.lang.cce.vmuls(x_data, 0)
min_value_tensor = te.lang.cce.vadds(data_zero, min_value)
max_value_tensor = te.lang.cce.vadds(data_zero, max_value)
max_value_1_tensor = te.lang.cce.vadds(data_zero, max_value_1)
sub_xy = te.lang.cce.vsub(x_data, y_data)
add_min_value = te.lang.cce.vadds(sub_xy, min_value)
vmax_zero = te.lang.cce.vmax(add_min_value, data_zero)
vmin_min_value = te.lang.cce.vmin(vmax_zero, min_value_tensor)
vmul_max_value = te.lang.cce.vmul(vmin_min_value, max_value_tensor)
vmul_max_value_1 = te.lang.cce.vmul(vmul_max_value, max_value_tensor)
result = te.lang.cce.vmul(vmul_max_value_1, max_value_1_tensor)
return result
def _log1p_mini_compute(mini_res, input_x, shape):
"""
do element-wise log(x + 1) compute in mini scene
f(y) = e^y(n), y(n) <= TAYLOR_NEGATIVE_THRESHOLD or y(n) >= TAYLOR_POSITIVE_THRESHOLD
f(y) = seventh taylor computer, TAYLOR_NEGATIVE_THRESHOLD < y(n) < TAYLOR_POSITIVE_THRESHOLD
Parameters:
----------
mini_res: TVM tensor, the tensor of log(x + 1)
input_x : TVM tensor, the placeholder of input_x
shape : tuple, the shape of input_x
Returns : A Tensor. Has the same type as mini_res.
-------
"""
input_y = mini_res
newton_taylor_res = _newton_taylor_log1p(input_x, input_y)
newton_exp_res = _newton_exp_log1p(input_x, input_y)
input_left_border = tvm.const(TAYLOR_NEGATIVE_THRESHOLD, input_y.dtype)
tensor_input_left_border = te.lang.dynamic.broadcast(
input_left_border, shape)
input_right_border = tvm.const(TAYLOR_POSITIVE_THRESHOLD, input_y.dtype)
tensor_input_right_border = te.lang.dynamic.broadcast(
input_right_border, shape)
exp_taylor_neg = te.lang.dynamic.vcmpsel(input_y, tensor_input_left_border,
'gt', newton_taylor_res,
newton_exp_res)
exp_taylor_neg = te.lang.dynamic.vcmpsel(input_y,
tensor_input_right_border, 'lt',
exp_taylor_neg, newton_exp_res)
return mini_res
def prelu_compute(input_x, weight_input, output_y, kernel_name="prelu"):
"""
calculating data
Parameters
----------
input_x : TVM tensor
the placeholder of input_x
output_y : dict
dict of output_y, include keys(shape and dtype)
weight_input : TVM tensor
the placeholder of weight_input
kernel_name : str
kernel name, default value is "prelu"
Returns
-------
output tensor
"""
shape_x = te.lang.cce.util.shape_to_list(input_x.shape)
if input_x.dtype == "float16":
scalar_zero = tvm.const(0, dtype="float16")
else:
scalar_zero = tvm.const(0, dtype="float32")
val_max = te.lang.cce.vmaxs(input_x, scalar_zero)
val_min = te.lang.cce.vmins(input_x, scalar_zero)
weight_input = te.lang.cce.broadcast(weight_input, shape_x)
val_prod = te.lang.cce.vmul(val_min, weight_input)
res = te.lang.cce.vadd(val_max, val_prod)
return res
def abs_grad_compute(y, dy, z, kernel_name="abs_grad"):
"""
do abs_grad compute
Parameters:
----------------
y: input tensor y
dy: input tensor dy
z: output dict
kernel_name: cce kernel name, default value is "abs_grad"
return: data_dy * sign(data_y)
----------------
"""
dtype = dy.dtype
if dtype == "float16":
fp_max = tvm.const(2**15, dtype)
fp_min = tvm.const(2**(-15), dtype)
else:
fp_max = tvm.const(2**62, dtype)
fp_min = tvm.const(2**(-127), dtype)
new_data = te.lang.cce.vmuls(y, fp_max)
abs_data = te.lang.cce.vabs(new_data)
denominator = te.lang.cce.vadds(abs_data, fp_min)
res = te.lang.cce.vdiv(new_data, denominator)
res = te.lang.cce.round(res)
data1_res = te.lang.cce.vmul(res, dy)
return data1_res
def tanh_grad_compute(y, dy, z, kernel_name="tanh_grad"):
"""
do element-wise tanh_grad operation between two input tensors
Parameters
----------
y: TVM tensor
the placeholder of y input data
dy: TVM tensor
the placeholder of dy input data
z: dict
shape and dtype of output, should be same shape and type as input
kernel_name: str
cce kernel name, default value is tanh_grad
Returns
-------
res : tvm.tensor
the result of tanh_grad
"""
dtype = y.dtype
if dtype == "float16":
y = te.lang.cce.cast_to(y, "float32")
dy = te.lang.cce.cast_to(dy, "float32")
data1_square = te.lang.cce.vmul(y, y)
data_mul = te.lang.cce.vmuls(data1_square, tvm.const(-1, dtype=dtype))
anuminate = te.lang.cce.vadds(data_mul, tvm.const(1, dtype=dtype))
res = te.lang.cce.vmul(anuminate, dy)
if dtype == "float16":
res = te.lang.cce.cast_to(res, "float16")
return res
def mish_compute(input_x, output_y, kernel_name="mish"):
"""
algorithm: mish
calculating data's mish,y= x*(1 - 2/(1+(1+exp(x))^2))
Parameters
----------
input_x: TVM tensor
the placeholder of input data
output_y: dict
shape and dtype of output, should be same shape and type as input
kernel_name: str
cce kernel name, default value is mish
Returns
-------
res : tvm.tensor
the result of mish
"""
dtype = input_x.dtype
exp_val = te.lang.cce.vexp(input_x)
add_exp_val = te.lang.cce.vadds(exp_val, tvm.const(1, dtype))
pow_var = te.lang.cce.vmul(add_exp_val, add_exp_val)
add_val = te.lang.cce.vadds(pow_var, tvm.const(1, dtype))
rec_val = te.lang.cce.vrec(add_val)
mul_val = te.lang.cce.vmuls(rec_val, tvm.const(-2, dtype=dtype))
add_val2 = te.lang.cce.vadds(mul_val, tvm.const(1, dtype=dtype))
res = te.lang.cce.vmul(input_x, add_val2)
return res
def acos_grad_compute(y, dy, z, kernel_name="acos_grad"):
"""
do acos_grad compute with sqrt and div
Parameters:
----------------
y: input tensor y
dy: input tensor dy
z: output dict
kernel_name: cce kernel name, default value is "acos_grad"
return: dy * (- 1 / (1 - data_y^2)^1/2)
----------------
"""
dtype = y.dtype
dtype_1 = dtype
if dtype == "float16" and \
api_check_support("te.lang.cce.vadd", "float32"):
y = te.lang.cce.cast_to(y, "float32")
dy = te.lang.cce.cast_to(dy, "float32")
dtype = "float32"
data1_square = te.lang.cce.vmul(y, y)
data1_square = te.lang.cce.vmuls(data1_square,
tvm.const(NUM_MINUS_ONE, dtype=dtype))
data1_square = te.lang.cce.vadds(data1_square,
tvm.const(NUM_ONE, dtype=dtype))
data1_reciprocal = te.lang.cce.vsqrt(data1_square, 1)
data1_reciprocal = te.lang.cce.vdiv(dy, data1_reciprocal)
res = te.lang.cce.vmuls(data1_reciprocal,
tvm.const(NUM_MINUS_ONE, dtype=dtype))
if dtype_1 == "float16":
res = te.lang.cce.cast_to(res, "float16")
return res
def _compute_positive(prox_v, alpha_broad, l1_broad, l2_broad):
"""
the operator's compute
var = sign(prox_v)/(1+alpha*l2) * max{|prox_v|-alpha*l1,0}
Parameters:
----------
prox_v: the value of prox_v
alpha_broad: the value of alpha_broad
l1_broad: the value of l1_broad
l2_broad: the value of l2_broad
Returns
the value of var_res
"""
prox_v_abs = te.lang.cce.vabs(prox_v)
prox_v_sign = sign(prox_v)
# 1+alpha*l2
alpha_l2 = te.lang.cce.vmul(alpha_broad, l2_broad)
alpha_l2_1 = te.lang.cce.vadds(alpha_l2, tvm.const(CONST_ONE, "float32"))
# max{|prox_v|-alpha*l1,0}
alpha_l1 = te.lang.cce.vmul(alpha_broad, l1_broad)
alpha_l1_neg = te.lang.cce.vmuls(alpha_l1,
tvm.const(CONST_ONE_NEG, "float32"))
prox_v_l1 = te.lang.cce.vadd(prox_v_abs, alpha_l1_neg)
max_value = te.lang.cce.vmax(
prox_v_l1,
te.lang.cce.broadcast(tvm.const(CONST_ZERO, "float32"), prox_v.shape))
# sign(prox_v)/(1+alpha*l2) * max{|prox_v|-alpha*l1,0}
res = te.lang.cce.vdiv(prox_v_sign, alpha_l2_1)
var_res = te.lang.cce.vmul(res, max_value)
return var_res
def fake_quant_with_min_max_grad_compute(dout, x, min_val, max_val, quant_min, quant_max,
kernel_name="fake_quant_with_min_max_grad"):
"""FakeQuantWithMinMaxGrad"""
shape = te.lang.cce.util.shape_to_list(x.shape)
shape_min = te.lang.cce.util.shape_to_list(min_val.shape)
quant_min = tvm.const(quant_min, x.dtype)
quant_max = tvm.const(quant_max, x.dtype)
quant_min = te.lang.cce.broadcast(quant_min, shape_min)
quant_max = te.lang.cce.broadcast(quant_max, shape_min)
# CalNudge(NudgeMinMax)
scale = te.lang.cce.vdiv(te.lang.cce.vsub(max_val, min_val), te.lang.cce.vsub(quant_max, quant_min))
zp_from_min = te.lang.cce.vsub(quant_min, te.lang.cce.vdiv(min_val, scale))
# Nudge zero point
nudge_zp = te.lang.cce.round(te.lang.cce.vmin(quant_max, te.lang.cce.vmax(quant_min, zp_from_min)))
nudge_min = te.lang.cce.vmul(te.lang.cce.vsub(quant_min, nudge_zp), scale)
nudge_max = te.lang.cce.vmul(te.lang.cce.vsub(quant_max, nudge_zp), scale)
nudge_min = te.lang.cce.broadcast(nudge_min, shape)
nudge_max = te.lang.cce.broadcast(nudge_max, shape)
bool_over_min = _less_compare_float32(nudge_min, x)
bool_less_max = _less_compare_float32(x, nudge_max)
bool_between = te.lang.cce.vmul(bool_over_min, bool_less_max)
res = te.lang.cce.vmul(dout, bool_between)
return res
def eltwise_compute(x, y, mode=1, coeff=[], kernel_name="eltwise"):
'''
Compute elementwise operation
'''
tensor_num = len(x)
inp_dtype = x[0].dtype
data0_tmp = x[0]
tmp_y = {}
tmp_y["addr_type"] = 0
tmp_y["valid_shape"] = []
tmp_y["slice_offset"] = []
fuse_y = tmp_y if y is None else y
fusion_params = get_fusion_params(x, fuse_y, tensor_num)
if mode == 1:
if len(coeff) != 0 and len(coeff) != tensor_num:
errorInfo = {}
errorInfo['errCode'] = "E81002"
errorInfo['op_name'] = 'eltwise'
errorInfo['coeff_length'] = str(len(coeff))
errorInfo['input_num'] = str(tensor_num)
raise RuntimeError(errorInfo, "In op[%s], the parameter[coeff]'s length[%s] should be "
"equal to inputs'num[%s]." %
(errorInfo['op_name'], errorInfo['coeff_length'],
errorInfo['input_num']))
if len(coeff) == tensor_num:
if type(coeff[0]) != int and type(coeff[0]) != float:
raise RuntimeError("ele of coeff must be a number.")
if coeff[0] != 1:
coeff1 = tvm.const(coeff[0], dtype=inp_dtype)
data0_tmp = te.lang.cce.vmuls(data0_tmp, coeff1)
res = None
if tensor_num == 1:
const_val_0 = tvm.const(0, dtype=inp_dtype)
data0_tmp = te.lang.cce.vadds(data0_tmp, const_val_0)
res = data0_tmp
elif tensor_num > 1:
for i in range(1, tensor_num):
datan_tmp = x[i]
if mode == 0:
data0_tmp = te.lang.cce.vmul(data0_tmp, datan_tmp)
elif mode == 2:
data0_tmp = te.lang.cce.vmax(data0_tmp, datan_tmp)
else:
if len(coeff) == 0:
data0_tmp = te.lang.cce.vadd(data0_tmp, datan_tmp)
elif coeff[i] == 1:
data0_tmp = te.lang.cce.vadd(data0_tmp, datan_tmp)
else:
coeff2 = tvm.const(coeff[i], dtype=inp_dtype)
datan_tmp = te.lang.cce.vmuls(datan_tmp, coeff2)
data0_tmp = te.lang.cce.vadd(data0_tmp, datan_tmp)
res = data0_tmp
res.op.attrs["ele_fusion_params"] = fusion_params
return res
def newton_iteration(shape, tensor_x_rec, tensor_x, symbol, iter_num):
"""
the function of newton_iteration
Parameters
----------
shape: tensor shape
tensor_x_rec: tensor
tensor_x: tensor
symbol: tensor symbol
Returns
-------
tensor_list: dict
scope_list: dict
emit_list: dict
"""
dtype_c = tensor_x_rec.dtype
num_two = tvm.const(2, dtype=dtype_c)
neg_one = tvm.const(-1, dtype=dtype_c)
tmp = tensor_x_rec
tensor_list = {}
scope_list = {}
emit_list = {}
tmp_mul = None
tmp_neg = None
tmp_add = None
for index in range(0, iter_num):
key = "tmp_mul_" + symbol + str(index)
tmp_mul = tvm.compute(shape,
lambda *i: tensor_x(*i) * tmp(*i),
name=key)
tensor_list[key] = tmp_mul
scope_list[key] = cce.scope_ubuf
emit_list[key] = "vector_mul"
key = "tmp_neg_" + symbol + str(index)
tmp_neg = tvm.compute(shape,
lambda *i: tmp_mul(*i) * neg_one,
name=key)
tensor_list[key] = tmp_neg
scope_list[key] = cce.scope_ubuf
emit_list[key] = "vector_muls"
key = "tmp_add_" + symbol + str(index)
tmp_add = tvm.compute(shape,
lambda *i: tmp_neg(*i) + num_two,
name=key)
tensor_list[key] = tmp_add
scope_list[key] = cce.scope_ubuf
emit_list[key] = "vector_adds"
key = "tmp_" + symbol + str(index)
tmp = tvm.compute(shape, lambda *i: tmp_add(*i) * tmp(*i), name=key)
tensor_list[key] = tmp
scope_list[key] = cce.scope_ubuf
emit_list[key] = "vector_mul"
return tensor_list, scope_list, emit_list
def select_compute(condition, x1, x2, kernel_name="select"):
"""
compute for select
Parameters
----------
condition: TVM tensor
the placeholder of input condition
x1: TVM tensor
the placeholder of first input data
x2: TVM tensor
the placeholder of second input data
kernel_name: str
cce kernel name, default value is "select"
Returns
-------
res : output of the result of select compute
"""
shape = te.lang.cce.util.shape_to_list(x1.shape)
x1_dtype = x1.dtype
con_shape = te.lang.cce.util.shape_to_list(condition.shape)
bool_dtype = condition.dtype
if x1_dtype in ("int8", "uint8"):
x1_dtype = "float32"
ones = te.lang.cce.broadcast(tvm.const(1, dtype=x1_dtype),
shape,
output_dtype=x1_dtype)
x1 = te.lang.cce.cast_to(x1, "float32")
x2 = te.lang.cce.cast_to(x2, "float32")
else:
ones = te.lang.cce.broadcast(tvm.const(1, dtype=x1_dtype),
shape,
output_dtype=x1_dtype)
if bool_dtype == "int8":
if x1_dtype == "int32":
condition_dtype = te.lang.cce.ceil(condition)
else:
condition_dtype = te.lang.cce.cast_to(condition, x1_dtype)
else:
if x1_dtype == "int32":
condition_dtype = condition
else:
condition_dtype = te.lang.cce.cast_to(condition, x1_dtype)
if list(con_shape) != list(shape):
condition_dtype = te.lang.cce.broadcast(condition_dtype, shape)
condition_opp = te.lang.cce.vsub(ones, condition_dtype)
temp_x = te.lang.cce.vmul(x1, condition_dtype)
temp_y = te.lang.cce.vmul(x2, condition_opp)
res = te.lang.cce.vadd(temp_x, temp_y)
if x1_dtype in ("int8", "uint8"):
res = te.lang.cce.cast_to(res, x1_dtype)
return res
def xlogy_grad_compute(placeholders, shape_max, dtype, rx, ry):
"""
do element-wise xlogy_grad compute
Parameters
----------
placeholders : the placeholder of data input
shape_max : the shape of broadcast
dtype : the type of data input
rx : the reduction indices of data input with broadcast
ry : the reduction indices for data input with broadcast
Returns
-------
output_y1 : result of xlogy_grad
output_y2 : result of xlogy_grad
None
"""
x1_ori = placeholders[0]
x2_ori = placeholders[1]
grad_ori = placeholders[2]
fp32_support = tbe_platform.cce_conf.api_check_support(
"te.lang.cce.vdiv", "float32")
if dtype == "float32" and not fp32_support:
raise RuntimeError("Don't support float32 in the platform.")
if dtype == "float16" and fp32_support:
x1 = te.lang.cce.cast_to(x1_ori, "float32")
x2 = te.lang.cce.cast_to(x2_ori, "float32")
grad = te.lang.cce.cast_to(grad_ori, "float32")
x1 = te.lang.cce.broadcast(x1, shape_max)
x2 = te.lang.cce.broadcast(x2, shape_max)
grad = te.lang.cce.broadcast(grad, shape_max)
else:
x1 = te.lang.cce.broadcast(x1_ori, shape_max)
x2 = te.lang.cce.broadcast(x2_ori, shape_max)
grad = te.lang.cce.broadcast(grad_ori, shape_max)
if dtype == "float16" and not fp32_support:
esp_min = tvm.const(1.18e-7, dtype="float16")
else:
esp_min = tvm.const(1.18e-38, dtype="float32")
x1_addespmin = te.lang.cce.vadds(x1, esp_min)
not_zero_x1 = te.lang.cce.vdiv(x1, x1_addespmin)
log_x2 = te.lang.cce.vlog(x2)
partial_x1 = te.lang.cce.vmul(not_zero_x1, log_x2)
partial_x1g = te.lang.cce.vmul(partial_x1, grad)
partial_x2 = te.lang.cce.vdiv(x1, x2)
partial_x2g = te.lang.cce.vmul(partial_x2, grad)
output_y1 = te.lang.cce.sum(partial_x1g, rx, keepdims=True)
output_y2 = te.lang.cce.sum(partial_x2g, ry, keepdims=True)
if dtype == "float16" and fp32_support:
output_y1 = te.lang.cce.cast_to(output_y1, "float16")
output_y2 = te.lang.cce.cast_to(output_y2, "float16")
return output_y1, output_y2