def _atan_compute(data):
"""compute for atan"""
dtype = data.dtype
if dtype == "float16":
data = topi.cast(data, "float32")
abs_data = topi.abs(data)
tensor_one = dc.one_const(abs_data.dtype)
abs_data_sub_one = topi.subtract(abs_data, tensor_one)
abs_data_add_one = topi.add(abs_data, tensor_one)
abs_data2 = topi.abs(topi.divide(abs_data_sub_one, abs_data_add_one))
# calucate data less than one
res = _do_atan_taylor(abs_data)
# calucate data more than one
res_mt_one = topi.add(_do_atan_taylor(abs_data2),
tvm.const(CONST_PI_BY_FOUR, abs_data2.dtype))
res = topi.minimum(res, res_mt_one)
if utils.product_is_mini() and data.dtype == "float32":
sign_mask = topi.cast(topi.sign(topi.cast(data, "float16")), "float32")
else:
sign_mask = topi.sign(data)
res = topi.multiply(res, sign_mask)
if dtype == "float16":
res = topi.cast(res, "float16")
return res
def selu_compute(input_data):
"""selu compute implemention"""
# if input_dtype is float16,convert it to float32
dtype = input_data.dtype
if dtype == "float16" or dtype == "float32":
input_data = topi.cast(input_data, "float32")
type_tmp = "float32"
else:
input_data = topi.cast(input_data, "float16")
type_tmp = "float16"
# generate tensor_zero to be compared
tensor_zero = topi.multiply(input_data, tvm.const(0, dtype=type_tmp))
# generate negative_res and positive_res to compute
# When the element value is greater than 0 and less than 0
negative_res = topi.minimum(input_data, tensor_zero)
positive_res = topi.maximum(input_data, tensor_zero)
exp_res = exp(negative_res)
sub_res = topi.add(exp_res, tvm.const(SCALAR_NEGATIVE_ONE, dtype=type_tmp))
negative_muls_res = topi.multiply(sub_res, tvm.const(SCALE_ALPHA_PRODUCT, dtype=type_tmp))
if dtype == "int8":
negative_muls_res = akg.lang.cce.ceil(negative_muls_res)
positive_muls_res = topi.multiply(positive_res, tvm.const(SCALE, dtype=type_tmp))
res = topi.add(negative_muls_res, positive_muls_res)
# cast to ori_dtype
if dtype == "float16" or dtype == "int8" or dtype == "int32":
res = topi.cast(res, dtype)
return res
def fake_quant_with_min_max_args(input_data,
min_=-6,
max_=6,
num_bits=8,
narrow_range=False):
"""
Computes Fake-quantize the 'input_data' tensor,
type float32 to 'output_data' tensor of same type
output_data = (floor(clamped_shifted * inv_nudged_scale + 0.5f))) * scale
+ nudged_min
scale = (max-min) / (quant_max-quant_min)
Args:
data_x1 (tvm.tensor.Tensor): Tensor of dtype "float32"
min ([float, int]): scalar, defaults to -6
max ([float, int]): scalar, defaults to 6. [min; max] define the
clamping range for the input_data data
num_bits ([float, int]): Defaults to 8. num_bits is the bitwidth
of the quantization,between 2 and 16
narrow_range ([bool]):
True, quantized into the quantization range [1; 2^num_bits - 1]
False,quantized into the quantization range [0; 2^num_bits - 1]
Returns:
tvm.tensor.Tensor
"""
shape = get_shape(input_data)
utils.check_shape(shape)
dtype = input_data.dtype
utils.ops_dtype_check(dtype, utils.DtypeForDavinci.FLOAT32)
nudged_min, nudged_max, scale = nudge_min_max(min_, max_, num_bits,
narrow_range)
zero_tensor = tvm.compute(input_data.shape,
lambda *i: tvm.const(0, dtype="float32"),
name="zero_tensor")
nudged_max_tensor = topi.add(zero_tensor, nudged_max)
nudged_min_tensor = topi.add(zero_tensor, nudged_min)
inv_nudged_scale = 1.00 / scale
# Transform the input between nudged_max and nudged_min
clamped_vmin = topi.minimum(input_data, nudged_max_tensor)
clamped = topi.maximum(clamped_vmin, nudged_min_tensor)
# Calculate the quantized and dequantized results
clamped_shifted = topi.subtract(clamped, nudged_min_tensor)
vmul_shifted = topi.multiply(clamped_shifted, inv_nudged_scale)
vadds_shifted = topi.add(vmul_shifted, 0.5)
floor_vadds_shifted = floor(vadds_shifted)
floor_cast = akg.lang.ascend.cast_to(floor_vadds_shifted, dtype)
res_scale = topi.multiply(floor_cast, scale)
res = topi.add(res_scale, nudged_min_tensor)
return res
def my_dsl(dtype, kernel_name, attrs):
m = tvm.var("M")
n = tvm.var("N")
A = tvm.placeholder((m, ), name="A", dtype=dtype)
B = tvm.placeholder((m, ), name="B", dtype=dtype)
if insn == "add":
C = topi.add(A, B)
elif insn == "sub":
C = topi.subtract(A, B)
if insn == "mul":
C = topi.multiply(A, B)
elif insn == "div":
C = topi.divide(A, B)
elif insn == "max":
C = topi.maximum(A, B)
elif insn == "min":
C = topi.minimum(A, B)
elif insn == "abs":
C = tvm.compute(A.shape, lambda *index: tvm.abs(A(*index)), name='C')
elif insn == "exp":
C = topi.exp(A)
elif insn == "log":
C = topi.log(A)
elif insn == "sqrt":
C = topi.sqrt(A)
C = topi.log(A)
elif insn == "sqrt":
C = topi.sqrt(A)
elif insn == "adds":
C = A + tvm.const(2, dtype)
elif insn == "muls":
C = A * tvm.const(2, dtype)
# C = tvm.compute((m, ), lambda i: A[i] + B[i], name="C")
s = tvm.create_schedule([C.op])
with akg.build_config(add_lower_pass=cce.debug_mode(0), dump_pass_ir=True):
if insnType == "binary":
mod = akg.build(s, [A, B, C],
"cce",
name=kernel_name,
attrs=attrs,
polyhedral=True)
else:
mod = akg.build(s, [A, C],
"cce",
name=kernel_name,
attrs=attrs,
polyhedral=True)
return mod
def _asin_compute(data_input):
"""Compute asin"""
dtype = data_input.dtype
boundary = tvm.const(BOUNDARY, "float32")
# Change dtype to float32
if dtype == "float16":
data_input = topi.cast(data_input, "float32")
# Sign mask
data_sign = sign(data_input)
# All positive
data1 = topi.multiply(data_input, data_sign)
# x belongs to (0, 2^(-0.5))
choice_1 = topi.minimum(data1, boundary)
choice_1 = topi.subtract(choice_1, boundary)
choice_1_floor = akg.lang.cce.floor(choice_1)
# the dtype of choice_1_floor is int32, need to be cast to fp32.
if utils.product_is_mini():
choice_1_floor = topi.cast(choice_1_floor, "float16")
choice_1_floor = topi.cast(choice_1_floor, "float32")
else:
choice_1_floor = topi.cast(choice_1_floor, "float32")
choice_1 = topi.multiply(choice_1_floor, neg_one_const("float32"))
taylor1 = _taylor_compute(data1)
res_1 = topi.multiply(taylor1, choice_1)
# x belongs to (2^(-0.5), 1)
choice_2 = topi.subtract(one_const("float32"), choice_1)
data2 = topi.subtract(one_const("float32"), topi.multiply(data1, data1))
data2_sqrt = _sqrt(data2)
taylor2 = _taylor_compute(data2_sqrt, data2)
res_2 = topi.multiply(taylor2, neg_one_const("float32"))
res_2 = topi.add(res_2, tvm.const(HALF_PI, "float32"))
res_2 = topi.multiply(res_2, choice_2)
# Restore sign
res_1 = topi.add(res_1, res_2)
res_1 = topi.multiply(res_1, data_sign)
# Restore dtype
if dtype == "float16":
res_1 = topi.cast(res_1, "float16")
return res_1
def truncate_div_compute(input_x1, input_x2):
"""compute for truncate_div"""
int_list = ("int32", "int8", "uint8")
if input_x1.dtype in int_list:
data_zero = dc.zero_const("float32")
data_x_broad = cast(input_x1, "float32")
data_y_broad = cast(input_x2, "float32")
res_div = topi.divide(data_x_broad, data_y_broad)
res_min_int = ceil(topi.minimum(res_div, data_zero))
res_max_int = floor(topi.maximum(res_div, data_zero))
res_trunc = topi.add(res_min_int, res_max_int)
res_trunc = cast(res_trunc, "float32")
else:
res_trunc = topi.divide(input_x1, input_x2)
return cast(res_trunc, input_x1.dtype)
def less_compare_float32(data_x, data_y):
"""if x is less than y, then return 1, else return 0"""
shape_inputs = get_shape(data_x)
# minimun num of float32 2**(-126)
data_min = akg.lang.ascend.broadcast(tvm.const(2**(-126), dtype="float32"),
shape_inputs, "float32")
data_zero = akg.lang.ascend.broadcast(dc.zero_const("float32"),
shape_inputs, "float32")
res_sub = topi.subtract(data_y, data_x)
res_min = topi.minimum(res_sub, data_min)
res_max = topi.maximum(res_min, data_zero)
# max num of float32 is 2**126
# but cce can only support 2**62, so use 62 * 62 * 2 to adaptor 126
res_mul_fierst = topi.multiply(res_max, tvm.const(2**62, dtype="float32"))
res_mul_second = topi.multiply(res_mul_fierst,
tvm.const(2**62, dtype="float32"))
res = topi.multiply(res_mul_second, tvm.const(2**2, dtype="float32"))
return res
def fake_quant_with_min_max_vars_per_channel_compute(input_data,
input_min,
input_max,
num_bits=8,
narrow_range=False):
"""fake_quant_with_min_max_vars_per_channel compute implemention"""
shape = get_shape(input_data.shape)
dtype = input_data.dtype
min_broadcast = akg.lang.ascend.broadcast(input_min, shape, dtype)
max_broadcast = akg.lang.ascend.broadcast(input_max, shape, dtype)
# get nudged_min and nudged_max by nudged_min_max_compute function
nudged_min_nudged_max = nudged_min_max_compute(min_broadcast,
max_broadcast, num_bits,
narrow_range)
# transform the input between nudged_max and nudged_min
clamped_tmp = topi.minimum(input_data, nudged_min_nudged_max[1])
clamped = topi.maximum(clamped_tmp, nudged_min_nudged_max[0])
# calculate the quantized and dequantized results
clamped_shifted = topi.subtract(clamped, nudged_min_nudged_max[0])
if product_is_mini():
clamped_shifted_div_scale = mul(clamped_shifted,
reciprocal(nudged_min_nudged_max[2]),
target=utils.CCE)
else:
clamped_shifted_div_scale = Divide(clamped_shifted,
nudged_min_nudged_max[2],
target=utils.CCE)
result_tmp = topi.add(clamped_shifted_div_scale, dc.half_const(dtype))
floor_result_tmp = akg.lang.ascend.floor(result_tmp)
if product_is_mini():
floor_result_tmp = topi.cast(floor_result_tmp, "float16")
floor_result_tmp = topi.cast(floor_result_tmp, "float32")
scale_product = topi.multiply(floor_result_tmp, nudged_min_nudged_max[2])
tmp_res = topi.add(scale_product, nudged_min_nudged_max[0])
# get bool_both_zero_value by bool_both_zero_compute function
bool_both_zero_value = bool_both_zero_compute(min_broadcast, max_broadcast)
res = topi.multiply(tmp_res, bool_both_zero_value)
return res
def _cmpare_value(input_data, nudged_min, nudged_max):
"""
where((input_data<=nudged_max)&(x>=nudged_min),1,0)
Args:
input_data (tvm.tensor.Tensor): Input data
nudged_min (tvm.tensor.Tensor): Minimum value of comparison
nudged_max (tvm.tensor.Tensor): Maximum value of comparison
Returns:
tvm.tensor.Tensor
"""
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_one = 1
max_value = tvm.const(2**(62), dtype="float32")
max_value_one = tvm.const(2**(2), dtype="float32")
data_zero = topi.multiply(input_data, 0)
max_value_tensor = topi.add(data_zero, max_value)
min_value_tensor = topi.add(data_zero, min_value)
max_value_one_tensor = topi.add(data_zero, max_value_one)
sub_tmp = topi.subtract(input_data, nudged_min)
sub_min = topi.add(sub_tmp, min_value)
vmax_tmp = topi.maximum(sub_min, data_zero)
sub_tmp_max = topi.subtract(nudged_max, input_data)
sub_max = topi.add(sub_tmp_max, min_value)
vmin_tmp = topi.maximum(sub_max, data_zero)
one_tmp = topi.multiply(vmax_tmp, vmin_tmp)
one_min = topi.minimum(one_tmp, min_value_tensor)
vmul_max_value = topi.multiply(one_min, max_value_tensor)
vmul_max_value_one = topi.multiply(vmul_max_value, max_value_tensor)
between_nudged_min_max = topi.multiply(vmul_max_value_one,
max_value_one_tensor)
return between_nudged_min_max
def _do_atan_taylor(data):
"""
Taylor algorithm for atan.
if x > 0 and x < tan(pi/8):
atan(x) = x - x^3/3 + x^5/5 - x^7/7 ...
elif x > tan(pi/8) and x < tan(pi/4):
atan(x) = atan(y) + atan((x-y)/(1+xy))
Args:
data (tvm.tensor.Tensor): Input data.
Returns:
A tvm.tensor.Tensor of atan(x).
"""
dtype = data.dtype
tensor_offset = tvm.const(TAN_PI_BY_EIGHT, dtype)
deno = topi.multiply(data, tvm.const(TAN_PI_BY_EIGHT, dtype))
deno = topi.add(deno, dc.one_const(dtype))
molecule = topi.subtract(data, tensor_offset)
ddata = topi.divide(molecule, deno)
ddata = topi.abs(ddata)
square_ddata = topi.multiply(ddata, ddata)
res = tvm.const(ATAN_TAYLOR_COEF[CONST_ITERTOR], dtype)
for i in reversed(range(CONST_ITERTOR)):
res = topi.multiply(res, square_ddata)
res = topi.add(res, tvm.const(ATAN_TAYLOR_COEF[i], dtype))
res = topi.multiply(res, ddata)
res = topi.add(res, tvm.const(CONST_PI_BY_EIGHT, dtype))
square_data = topi.multiply(data, data)
res2 = tvm.const(ATAN_TAYLOR_COEF[CONST_ITERTOR2], dtype)
for i in reversed(range(CONST_ITERTOR2)):
res2 = topi.multiply(res2, square_data)
res2 = topi.add(res2, tvm.const(ATAN_TAYLOR_COEF[i], dtype))
return topi.minimum(res, topi.multiply(res2, data))
