def tan_compute(input_x):
"""tan compute implemention"""
dtype = input_x.dtype
# cast to type float32 when type is float16 in cloud and mini, or int32 in cloud
if dtype == FLOAT_16 or dtype == FLOAT_32 or (dtype == INT_32
and not product_is_mini()):
input_x = topi.cast(input_x, FLOAT_32)
# adjust x to [-pi/2,pi/2] using x = x-round(x/pi)*pi
round_pi_div = akg.lang.ascend.round(
topi.multiply(input_x, tvm.const(1.0 / PI, FLOAT_32)))
round_pi_div = akg.lang.ascend.cast_to(round_pi_div, FLOAT_32)
input_x = topi.subtract(
input_x, topi.multiply(round_pi_div, tvm.const(PI, FLOAT_32)))
# cast to type float16 when type is int32 in mini
elif dtype == INT_32 and product_is_mini():
input_x = topi.cast(input_x, FLOAT_16)
# adjust x to [-pi/2,pi/2] using x = x-round(x/pi)*pi
round_pi_div = akg.lang.ascend.round(
topi.multiply(input_x, tvm.const(1.0 / PI, FLOAT_16)))
round_pi_div = akg.lang.ascend.cast_to(round_pi_div, FLOAT_16)
input_x = topi.subtract(
input_x, topi.multiply(round_pi_div, tvm.const(PI, FLOAT_16)))
res = _tan_2x_multi(input_x, TAN_2X_TIMES)
# cast the dtype to original dtype
res = topi.cast(res, dtype)
return res
def fused_bn_update(input1, input2, input3, input4, dtype, c1, c2, c3, c4):
"""
fused operator.
Args:
input1 ~ input4: tvm.tensor.Tensor.
dtype: dtype of Tensor.
c1 ~ c4: const.
Returns:
Three output (list of tvm.tensor.Tensor).
"""
const1 = tvm.const(c1, dtype)
mul0 = topi.multiply(input2, const1)
mul1 = topi.multiply(input1, const1)
mul2 = topi.multiply(mul1, mul1)
sigma2 = topi.subtract(mul0, mul2)
const2 = tvm.const(c2, dtype)
rsqrt_val = topi.rsqrt(topi.add(sigma2, const2))
const3 = tvm.const(c3, dtype)
mul3 = topi.multiply(sigma2, const3)
sub1 = topi.subtract(input3, mul3)
const4 = tvm.const(c4, dtype)
data1 = topi.multiply(const4, sub1)
sub2 = topi.subtract(input4, mul1)
data2 = topi.multiply(const4, sub2)
return (rsqrt_val, data1, data2)
def fused_relu_grad_bn_double_update_grad(data_1, data_2, data_3, data_4, data_5, data_6, data_7, layout='NHWC'):
transform_list = [data_2, data_4, data_5, data_6, data_7]
for i in transform_list:
if layout == "NCHW":
i = topi.transpose(i, axes=(0, 2, 3, 1))
elif layout != "NHWC":
raise NotImplementedError( 'Layout not supported {} '.format(layout))
data_tmp1 = topi.full_like(data_7, 0.0)
data_tmp2 = topi.greater(data_7, data_tmp1)
data_tmp3 = topi.add(data_5, data_6)
data_tmp4 = topi.where(data_tmp2, data_tmp3, data_tmp1)
data_tmp5 = topi.cast(data_tmp4, 'float32')
data_tmp7 = topi.sum(data_tmp5, axis=(0, 1, 2))
n, h, w, c = data_7.shape
data_tmp8 = topi.cast(data_2, 'float32')
data_tmp9 = topi.full_like(data_tmp7, 1.0/(n*h*w))
data_tmp10 = topi.multiply(data_1, data_tmp9)
data_tmp11 = topi.broadcast_to(data_tmp10, data_tmp8.shape)
data_tmp12 = topi.subtract(data_tmp8, data_tmp11)
data_tmp13 = topi.multiply(data_tmp5, data_tmp12)
data_tmp15 = topi.sum(data_tmp13, axis=(0, 1, 2))
data_tmp16 = topi.cast(data_4, 'float32')
data_tmp17 = topi.multiply(data_3, data_tmp9)
data_tmp18 = topi.broadcast_to(data_tmp17, data_tmp16.shape)
data_tmp19 = topi.subtract(data_tmp16, data_tmp18)
data_tmp20 = topi.multiply(data_tmp5, data_tmp19)
data_tmp22 = topi.sum(data_tmp20, axis=(0, 1, 2))
return [data_tmp7, data_tmp15, data_tmp22]
def fused_relu_grad_bn_double_reduce_grad(data0, data1, data2, data3, data4, data5, data6, data7, data8,
data9, data10, data11, data12, data13, data14, data15, layout="NHWC",
out_dtype="float16", target=utils.CUDA):
if layout == 'NCHW':
data5 = topi.transpose(data5, (0, 2, 3, 1))
data9 = topi.transpose(data9, (0, 2, 3, 1))
data13 = topi.transpose(data13, (0, 2, 3, 1))
data14 = topi.transpose(data14, (0, 2, 3, 1))
data15 = topi.transpose(data15, (0, 2, 3, 1))
elif layout != 'NHWC':
raise NotImplementedError(
'Layout not supported {} '.format(layout))
inter_dtype = "float32"
n, h, w, c = data5.shape
scale = n * h * w
mul = topi.multiply(data2, data3)
mul1221 = topi.divide(mul, scale)
# ReluGrad
zero = tvm.const(0, data15.dtype)
add = topi.add(data13, data14)
addgrad = tvm.compute(add.shape, lambda *i: tvm.if_then_else(data15(*i) >= zero, add(*i), zero), tag=tag.INJECTIVE)
addgrad = topi.cast(addgrad, inter_dtype)
mul3283 = topi.multiply(scale, addgrad)
sub1159 = topi.subtract(mul3283, data6)
data5_cast = topi.cast(data5, inter_dtype)
mul2372 = topi.divide(data4, scale)
sub631 = topi.subtract(data5_cast, mul2372)
mul1220 = topi.multiply(sub631, data1)
div = topi.divide(mul1220, data0)
sub271 = topi.subtract(sub1159, div)
mul1218 = topi.multiply(mul1221, sub271)
mul1218_cast = topi.cast(mul1218, out_dtype)
mul1231 = topi.multiply(data11, data12)
mul1230 = topi.divide(mul1231, scale)
data9_cast = topi.cast(data9, inter_dtype)
mul2364 = topi.divide(data8, scale)
sub625 = topi.subtract(data9_cast, mul2364)
mul1229 = topi.multiply(data10, sub625)
div272 = topi.divide(mul1229, data7)
sub272 = topi.subtract(sub1159, div272)
mul1228 = topi.multiply(mul1230, sub272)
mul1228_cast = topi.cast(mul1228, out_dtype)
if layout == "NCHW":
mul1218_cast = topi.transpose(mul1218_cast, (0, 3, 1, 2))
mul1228_cast = topi.transpose(mul1228_cast, (0, 3, 1, 2))
return [mul1218_cast, mul1228_cast]
def fused_relu_grad_bn_reduce_grad(data_1,
data_2,
data_3,
data_4,
data_5,
data_6,
data_7,
data_8,
data_9,
layout='NHWC',
target=utils.CUDA):
"""
fused_relu_grad_bn_reduce_grad.
Args:
data_1~data_9: tvm.tensor.Tensor.
layout: input layout, only 'NCHW', 'NHWC' supported
Returns:
tvm.tensor.Tensor.
"""
transform_list = [data_7, data_8, data_9]
for i in transform_list:
if layout == "NCHW":
i = topi.transpose(i, axes=(0, 2, 3, 1))
elif layout != "NHWC":
raise NotImplementedError(
'Layout not supported {} '.format(layout))
data_tmp1 = topi.multiply(data_4, data_5)
n, h, w, c = data_9.shape
data_tmp2 = topi.full_like(data_tmp1, 1.0 / (n * h * w))
data_tmp3 = topi.multiply(data_tmp1, data_tmp2)
data_tmp5 = topi.full_like(data_9, 0.0)
data_tmp6 = topi.greater(data_9, data_tmp5)
data_tmp7 = topi.where(data_tmp6, data_8, data_tmp5)
data_tmp8 = topi.cast(data_tmp7, 'float32')
data_tmp9 = topi.full_like(data_tmp8, n * h * w)
data_tmp10 = topi.multiply(data_tmp8, data_tmp9)
data_tmp12 = topi.subtract(data_tmp10, data_3)
data_tmp14 = topi.cast(data_7, 'float32')
data_tmp15 = topi.multiply(data_6, data_tmp2)
data_tmp17 = topi.subtract(data_tmp14, data_tmp15)
data_tmp18 = topi.multiply(data_2, data_tmp17)
data_tmp20 = topi.divide(data_tmp18, data_1)
data_tmp21 = topi.subtract(data_tmp12, data_tmp20)
data_tmp22 = topi.multiply(data_tmp3, data_tmp21)
data_out = topi.cast(data_tmp22, 'float16')
return data_out
def sgd_compute(parameters, gradient, learning_rate, accum, momentum, stat,
dampening=0.0, weight_decay=0.0, nesterov=False):
"""sgd compute implementation"""
dtype = parameters.dtype
if dtype == "float16":
parameters = topi.cast(parameters, "float32")
accum = topi.cast(accum, "float32")
learning_rate = topi.cast(learning_rate, "float32")
gradient = topi.cast(gradient, "float32")
momentum = topi.cast(momentum, "float32")
stat = topi.cast(stat, "float32")
# if weight_decay != 0.0, need compute grad_delta to update gradient
if weight_decay != 0.0:
parameters = topi.multiply(parameters, tvm.const(1.0, 'float32'))
grad_delta = topi.multiply(parameters, weight_decay)
gradient = topi.add(gradient, grad_delta)
stat_mid = topi.multiply(stat, tvm.const(-1, "float32"))
stat_act = topi.add(stat_mid, tvm.const(1, "float32"))
dampening_t = topi.multiply(stat_act, dampening)
# update accum
accum_delta = tvm.compute(accum.shape, lambda *indice: accum(*indice) * momentum[0])
gradient_damp = topi.multiply(gradient, dampening_t)
accum_t = topi.add(accum_delta, gradient)
if dampening != 0.0:
accum_t = topi.subtract(accum_t, gradient_damp)
# update parameters
if nesterov:
parameters_delta = tvm.compute(gradient.shape, lambda *indice: gradient(*indice) * learning_rate[0])
parameters_delta_2 = tvm.compute(accum_t.shape, lambda *indice: accum_t(*indice) * momentum[0])
parameters_delta_2 = tvm.compute(parameters_delta_2.shape,
lambda *indice: parameters_delta_2(*indice) * learning_rate[0])
parameters_delta = topi.add(parameters_delta, parameters_delta_2)
parameters_t = topi.subtract(parameters, parameters_delta)
else:
parameters_delta = tvm.compute(accum_t.shape, lambda *indice: accum_t(*indice) * learning_rate[0])
parameters_t = topi.subtract(parameters, parameters_delta)
# update stat
stat_t = topi.multiply(stat_act, tvm.const(NUM_ZERO, 'float32'))
if dtype == "float16":
parameters_t = topi.cast(parameters_t, "float16")
accum_t = topi.cast(accum_t, "float16")
stat_t = topi.cast(stat_t, "float16")
return parameters_t, accum_t, stat_t
def _apply_ada_max_compute(var, m, v, grad, lr, beta1, beta1_power, beta2,
epsilon):
"""Compute ada_max."""
# cast to float32 for improved accuracy
inp_dtype = var.dtype
if inp_dtype == 'float16':
var = topi.cast(var, 'float32')
m = topi.cast(m, 'float32')
v = topi.cast(v, 'float32')
lr = topi.cast(lr, 'float32')
beta1_power = topi.cast(beta1_power, 'float32')
beta1 = topi.cast(beta1, 'float32')
beta2 = topi.cast(beta2, 'float32')
grad = topi.cast(grad, 'float32')
epsilon = tvm.const(epsilon, 'float32')
# m += (grad - m) * (1 - beta1)
rhs = tvm.compute(beta1.shape,
lambda *i: beta1(*i) * neg_one_const("float32"))
rhs = tvm.compute(rhs.shape, lambda *i: rhs(*i) + one_const("float32"))
lhs = topi.subtract(grad, m)
rhs = tvm.compute(lhs.shape, lambda *i: lhs(*i) * rhs[0])
m = topi.add(m, rhs)
# v = max(beta2*v, abs(grad))
lhs = tvm.compute(v.shape, lambda *i: v(*i) * beta2[0])
rhs = topi.abs(grad)
v = topi.maximum(lhs, rhs)
# var -= lr / (1 - beta1_power) * (m / (v + epsilon))
# lr * m / (1 - beta1_power) * (v + epsilon)
# v + epsilon
rhs = tvm.compute(v.shape, lambda *i: v(*i) + epsilon)
# 1 - beta1_power
lhs = tvm.compute(beta1_power.shape,
lambda *i: beta1_power(*i) * neg_one_const("float32"))
lhs = tvm.compute(lhs.shape, lambda *i: lhs(*i) + one_const("float32"))
# (1 - beta1_power) * (v + epsilon)
rhs = tvm.compute(rhs.shape, lambda *i: rhs(*i) * lhs[0])
# lr * m
lhs = tvm.compute(m.shape, lambda *i: m(*i) * lr[0])
# lr * m / (1 - beta1_power) * (v + epsilon)
rhs = reciprocal(rhs)
rhs = topi.multiply(lhs, rhs)
var = topi.subtract(var, rhs)
if inp_dtype == 'float16':
var = topi.cast(var, inp_dtype)
m = topi.cast(m, inp_dtype)
v = topi.cast(v, inp_dtype)
return var, m, v
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 fused_bn_reduce_grad(data0,
data1,
data2,
data3,
data4,
data5,
data6,
data7,
layout='NHWC',
out_dtype='float16',
target=utils.CUDA):
if layout == 'NCHW':
data3 = topi.transpose(data3, (0, 2, 3, 1))
data7 = topi.transpose(data7, (0, 2, 3, 1))
elif layout != 'NHWC':
raise NotImplementedError('Layout not supported {} '.format(layout))
n, h, w, c = data3.shape
const = n * h * w
inter_dtype = 'float32'
out1 = topi.multiply(data4, data5)
out1 = topi.divide(out1, const)
out1 = topi.expand_dims(out1, axis=0, num_newaxis=3)
out1 = topi.broadcast_to(out1, (n, h, w, c))
data3 = topi.cast(data3, inter_dtype)
data2 = topi.expand_dims(data2, axis=0, num_newaxis=3)
data2 = topi.broadcast_to(data2, (n, h, w, c))
out2 = topi.multiply(data3, const)
out2 = topi.subtract(out2, data2)
data1 = topi.expand_dims(data1, axis=0, num_newaxis=3)
data1 = topi.broadcast_to(data1, (n, h, w, c))
data7 = topi.cast(data7, inter_dtype)
out3 = topi.divide(data6, const)
out3 = topi.subtract(data7, out3)
out3 = topi.multiply(data1, out3)
out3 = topi.divide(out3, data0)
output = topi.subtract(out2, out3)
output = topi.multiply(output, out1)
output = topi.cast(output, out_dtype)
if layout == "NCHW":
output = topi.transpose(output, (0, 3, 1, 2))
return output
def _apply_adadelta_compute(var, accum, accum_update, grad, lr, rho, epsilon):
"""Compute apply_adadelta"""
dtype = var.dtype
if dtype == "float16":
var = topi.cast(var, "float32")
accum = topi.cast(accum, "float32")
accum_update = topi.cast(accum_update, "float32")
lr = topi.cast(lr, "float32")
rho = topi.cast(rho, "float32")
grad = topi.cast(grad, "float32")
epsilon = tvm.const(epsilon, "float32")
tensor_one = akg.lang.ascend.broadcast(tvm.const(1, "float32"), var.shape)
tensor_rho = topi.broadcast_to(rho, var.shape)
tensor_rho_gs = topi.subtract(tensor_one, tensor_rho)
tensor_epsilon = akg.lang.ascend.broadcast(epsilon, var.shape)
# accum = accum * rho + grad ** 2 * (1 - rho)
rhs = topi.multiply(accum, tensor_rho)
lhs = topi.multiply(grad, grad)
lhs = topi.multiply(lhs, tensor_rho_gs)
accum_res = akg.lang.ascend.vadd(lhs, rhs)
# update = (accum_update + epsilon).sqrt * (accum + epsilon).rsqrt * grad
rhs = topi.add(accum_update, tensor_epsilon)
rhs = sqrt(rhs, target=utils.CCE)
lhs = topi.add(accum_res, tensor_epsilon)
lhs = rsqrt(lhs, target=utils.CCE)
lhs = topi.multiply(grad, lhs)
update = topi.multiply(lhs, rhs)
# var -= update * lr
var_res = topi.broadcast_to(lr, var.shape)
var_res = topi.multiply(update, var_res)
var_res = topi.subtract(var, var_res)
# accum_update = rho * accum_update + (1 - rho) * update.square
rhs = topi.multiply(accum_update, tensor_rho)
lhs = topi.multiply(update, update)
lhs = topi.multiply(lhs, tensor_rho_gs)
accum_update_res = akg.lang.ascend.vadd(lhs, rhs)
if dtype == "float16":
var_res = topi.cast(var_res, "float16")
accum_res = topi.cast(accum_res, "float16")
accum_update_res = topi.cast(accum_update_res, "float16")
return var_res, accum_res, accum_update_res
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 fused_bn_follow(data0, data1, data2, data3, data4, target=utils.CUDA):
"""
input:
data: length is 5
data0: param0 beta
data1: param1 gamma
data2: param2 BNupdate: xi_variance
data3: param6 BNreduce: xi_mean
data4: param7 xi_conv2d
layout: (N, C, H, W)
output:
beta + gamma * xi_variance * ( xi - xi_mean/(N*H*W) )
"""
n, h, w, c = data4.shape
const = n * h * w
inter_dtype = 'float32'
data4 = topi.cast(data4, inter_dtype)
multiply0 = topi.divide(data3, const)
multiply0 = topi.expand_dims(multiply0, axis=0, num_newaxis=3)
multiply0 = topi.broadcast_to(multiply0, (n, h, w, c))
subtract0 = topi.subtract(data4, multiply0)
multiply1 = topi.multiply(subtract0, data2)
multiply2 = topi.multiply(multiply1, data1)
add0 = topi.add(multiply2, data0)
return add0
def _apply_gradient_descent_compute(var, alpha, delta):
"""Compute gradient_descent"""
# step 1: calculate delta * alpha
var_change = tvm.compute(delta.shape,
lambda *indices: delta(*indices) * alpha[0])
# step 2: calculate var - delta * alpha
reuse_var = topi.subtract(var, var_change)
return reuse_var
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 _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 _log_taylor(data):
"""log algrithm is log(1+x) = ((((0.2x - 0.25)x + 0.33333)x - 0.5)x + 1)x"""
data = topi.subtract(data, 1)
taylor_params = [0.2, -0.25, 1 / 3, -0.5, 1]
taylor_five = topi.multiply(data, taylor_params[0])
taylor_four_1 = topi.add(taylor_five, taylor_params[1])
taylor_four_2 = topi.multiply(taylor_four_1, data)
taylor_three_1 = topi.add(taylor_four_2, taylor_params[2])
taylor_three_2 = topi.multiply(taylor_three_1, data)
taylor_two_1 = topi.add(taylor_three_2, taylor_params[3])
taylor_two_2 = topi.multiply(taylor_two_1, data)
taylor_one = topi.add(taylor_two_2, taylor_params[4])
taylor = topi.multiply(taylor_one, data)
return taylor
def bn_gamma_grad(head, in_data, data_sum, layout="NHWC"):
if layout == "NCHW":
head = topi.tranpose(head, (0, 2, 3, 1))
n, h, w, c = head.shape
n = n.value
h = h.value
w = w.value
c = c.value
scale = tvm.const(n * h * w, head.dtype)
mean = topi.divide(data_sum, scale)
x_hat = topi.subtract(in_data, mean)
x_hat_mul = topi.multiply(x_hat, head)
bn_gamma_grad = topi.sum(x_hat_mul, axis=(0, 1, 2))
return bn_gamma_grad
def sinh_compute(x):
"""Compute sinh."""
dtype = x.dtype
# in order to get the precise calcuate result
if dtype == "float16":
x = topi.cast(x, "float32")
data_exp = Exp(x, utils.CCE)
negative_data = topi.multiply(x, -1)
negative_data_exp = Exp(negative_data, utils.CCE)
data_exp_sub = topi.subtract(data_exp, negative_data_exp)
res = topi.multiply(data_exp_sub, tvm.const(0.5, "float32"))
if dtype == "float16":
res = topi.cast(res, "float16")
return res
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 _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))
def fused_minimum_or_maximum_grad(dz, x, y, grad_x, grad_y, op_type):
"""
Gradient for minimum or maximum operation between two input tensors `x` and `y`.
Args:
dz (tvm.tensor.Tensor): Type float16, float32, int32.
x (tvm.tensor.Tensor): Type float16, float32, int32.
y (tvm.tensor.Tensor): Type float16, float32, int32.
grad_x (bool): Whether calculate dx.
grad_y (bool): Whether calculate dy.
op_type (str): The type of the op, "GE" for MaximumGrad or "LE" for MinimumGrad.
Note:
At least one of grad_x and grad_y is True.
Returns:
dx, tvm.tensor.Tensor of the same type as inputs, it will be returned if grad_x is True.
dy, tvm.tensor.Tensor of the same type as inputs, it will be returned if grad_y is True.
"""
vc_util.check_shape(x)
vc_util.check_shape(y)
vc_util.check_shape(dz)
vc_util.ops_dtype_check([x.dtype, y.dtype, dz.dtype],
[vc_util.DtypeForDavinci.ALL_FLOAT, vc_util.DtypeForDavinci.INT32])
vc_util.broadcast_check(x, dz)
vc_util.broadcast_check(y, dz)
# check op types
check_list = ["GE", "LE"]
if op_type not in check_list:
raise ValueError("FusedMinimumOrMaximumGrad only support %s while op type is %s" %
(",".join(check_list), op_type))
if not grad_x and not grad_y:
raise ValueError("At least one of grad_x and grad_y is True.")
x_shape = get_shape(x)
y_shape = get_shape(y)
dz_shape = get_shape(dz)
ori_dtype = dz.dtype
# get greater compute
x = akg.lang.cce.broadcast(x, dz_shape)
y = akg.lang.cce.broadcast(y, dz_shape)
if utils.product_is_mini() and ori_dtype != "float16":
x = cast(x, "float16")
y = cast(y, "float16")
dz = cast(dz, "float16")
elif ori_dtype == "int32":
x = cast(x, "float32")
y = cast(y, "float32")
dz = cast(dz, "float32")
zero = zero_const(dz.dtype)
if op_type == "LE":
dx = tvm.compute(dz_shape, lambda *i: tvm.expr.Select((x(*i) <= y(*i)), dz(*i), zero), name='dx')
dy = topi.subtract(dz, dx)
elif op_type == "GE":
dx = tvm.compute(dz_shape, lambda *i: tvm.expr.Select((x(*i) >= y(*i)), dz(*i), zero), name='dx')
dy = topi.subtract(dz, dx)
if dx.dtype == "float16":
# cast to fp32 for higher precision of reduce_sum.
if get_shape(dx) != x_shape:
dx = cast(dx, "float32")
if get_shape(dy) != y_shape:
dy = cast(dy, "float32")
dx = sum.sum_by_shape(dx, x_shape)
dy = sum.sum_by_shape(dy, y_shape)
if ori_dtype != dx.dtype:
dx = cast(dx, ori_dtype)
if ori_dtype != dy.dtype:
dy = cast(dy, ori_dtype)
attrs = get_default_attrs()
if grad_x and grad_y:
return dx, dy, attrs
if grad_x:
return dx, attrs
return dy, attrs
def _bool_negate(input_bool):
"""Negate every value"""
return topi.subtract(dc.one_const(input_bool.dtype), input_bool)
def nudged_min_max_compute(min_broadcast, max_broadcast, num_bits,
narrow_range):
"""
Calculate the maximum and minimum values of the quantization.
Notes:
Each channel scale[i] euqal to (max_broadcast[i] - min_broadcast[i]) / (quant_max - quant_min).
Then compute nudged_zero_point:
nudged_zero_point = floor(between_min_max_float + 0.5) + less_quant_min_float + more_quant_max_float,
between_min_max_float is first calculated by:
zero_point_from_min = (quant_min_float - min_broadcast) / scale,
then between_min_max_float = zero_point_from_min, which min_broadcast <= zero_point_from_min <= max_broadcast.
Besides, the value of less_quant_min_float is equal to quant_min or zero, zero_point_from_min < quant_min_float,
the value is quant_min, else is 0. The same as more_quant_max_float.
Finally according to scale and nudged_zero_point to compute nudged_min and nudged_max:
nudged_min = (quant_min - nudged_zero_point) * scale
nudged_max = (quant_max - nudged_zero_point) * scale
Args:
min_broadcast (tvm.tensor.Tensor): minimum value to be quantified for each channel.
max_broadcast (tvm.tensor.Tensor): maximum value to be quantified for each channel.
num_bits (int): num_bits is the bitwidth of the quantization, range [2,16].
narrow_range (bool): if True, for each channel, quantized into the quantization range [0, 2^num_bits - 1] else
quantized into the quantization range [1, 2^num_bits - 1].
Returns:
nudged_min (tvm.tensor.Tensor): The same type and shape as min_broadcast.
nudged_max (tvm.tensor.Tensor): The same type and shape as max_broadcast.
scale (tvm.tensor.Tensor): The same type and shape as max_broadcast.
"""
dtype = min_broadcast.dtype
quant_min = 1 if narrow_range else 0
quant_max = (2**num_bits) - 1
# because of need compute each channel, so quant_min and quant_max need to broadcast.
quant_min_float = topi.full(min_broadcast.shape, dtype,
tvm.const(quant_min, dtype))
quant_max_float = topi.full(min_broadcast.shape, dtype,
tvm.const(quant_max, dtype))
# caculate each channel max and min difference.
max_sub_min = topi.subtract(max_broadcast, min_broadcast)
quant_max_sub_quant_min = topi.subtract(quant_max_float, quant_min_float)
# compute scale = (max_broadcast - min_broadcast) / (quant_max - quant_min)
# and min_div_scale = min_broadcast / scale
if product_is_mini():
scale = mul(max_sub_min,
reciprocal(quant_max_sub_quant_min),
target=utils.CCE)
min_div_scale = Mul(min_broadcast, reciprocal(scale), target=utils.CCE)
else:
scale = Divide(max_sub_min, quant_max_sub_quant_min, target=utils.CCE)
min_div_scale = Divide(min_broadcast, scale, target=utils.CCE)
# zero_point_from_min = quant_min_float - min_broadcast / scale
zero_point_from_min = topi.subtract(quant_min_float, min_div_scale)
# if zero_point_from_min < quant_min_float, bool_less_quant_min_float = 1 else 0
bool_less_quant_min_float = less_compare_float32(zero_point_from_min,
quant_min_float)
# if quant_max_float < zero_point_from_min, bool_more_quant_max_float = 1 else 0
bool_more_quant_max_float = less_compare_float32(quant_max_float,
zero_point_from_min)
# according to above bool param to select effective value
less_quant_min_float = topi.multiply(quant_min_float,
bool_less_quant_min_float)
more_quant_max_float = topi.multiply(quant_max_float,
bool_more_quant_max_float)
# compute which num is not less than quant_min_float and not large than quant_max_float
tensor_one = topi.full(min_broadcast.shape, dtype, dc.one_const(dtype))
bool_not_less_quant_min_float = topi.subtract(tensor_one,
bool_less_quant_min_float)
bool_not_more_quant_max_float = topi.subtract(tensor_one,
bool_more_quant_max_float)
bool_between_min_max = topi.multiply(bool_not_less_quant_min_float,
bool_not_more_quant_max_float)
between_min_max_float = topi.multiply(zero_point_from_min,
bool_between_min_max)
# add 0.5 to num which min <= num <= max and then floor them.
between_min_max_add_half_one = topi.add(between_min_max_float,
dc.half_const(dtype))
between_min_max_round = akg.lang.ascend.floor(between_min_max_add_half_one)
if product_is_mini():
between_min_max_round = topi.cast(between_min_max_round, "float16")
between_min_max_round = topi.cast(between_min_max_round, "float32")
# calculate the maximum and minimum values of the quantization
nudged_zero_point_tmp = topi.add(less_quant_min_float,
more_quant_max_float)
nudged_zero_point = topi.add(nudged_zero_point_tmp, between_min_max_round)
nudged_min_tmp = topi.subtract(quant_min_float, nudged_zero_point)
nudged_max_tmp = topi.subtract(quant_max_float, nudged_zero_point)
nudged_min = topi.multiply(nudged_min_tmp, scale)
nudged_max = topi.multiply(nudged_max_tmp, scale)
res = [nudged_min, nudged_max, scale]
return res
def _apply_adagrad_da_compute(var, gradient_accum, gradient_squared_accum,
grad, lr, l1, l2, global_step):
"""Compute adagrad_da."""
dtype = var.dtype
# cast to float32 for higher precision
if dtype == "float16":
gradient_accum = topi.cast(gradient_accum, "float32")
gradient_squared_accum = topi.cast(gradient_squared_accum, "float32")
grad = topi.cast(grad, "float32")
lr = topi.cast(lr, "float32")
l1 = topi.cast(l1, "float32")
l2 = topi.cast(l2, "float32")
if product_is_mini():
global_step = topi.cast(global_step, "float16")
global_step = topi.cast(global_step, "float32")
else:
global_step = topi.cast(global_step, "float32")
# 1.grad_accum += grad
gradient_accum = topi.add(gradient_accum, grad)
# 2.grad_squared_accum += grad * grad
gs = topi.multiply(grad, grad)
gradient_squared_accum = topi.add(gradient_squared_accum, gs)
# 3.if l1 > 0: tmp_val = Sign(grad_accum) * max(|grad_accum|-l1*global_step, 0)
# else: tmp_val = grad_accum
sign_val = Sign(gradient_accum)
abs_val = topi.abs(gradient_accum)
mul_val = topi.multiply(global_step, l1)
sub_val = topi.subtract(abs_val, mul_val)
max_val = topi.maximum(sub_val, tvm.const(0, sub_val.dtype))
tmp_val = topi.multiply(sign_val, max_val)
def select(l1, tmp_val, gradient_accum):
"""Returns tmp_val if l1 > 0 else gradient_accum."""
if product_is_mini():
l1 = topi.cast(l1, "float16")
tmp_val = topi.cast(tmp_val, "float16")
gradient_accum = topi.cast(gradient_accum, "float16")
tmp_val = akg.tvm.compute(
tmp_val.shape, lambda *i: tvm.expr.Select(l1[0] > 0, tmp_val(*i),
gradient_accum(*i)))
return topi.cast(tmp_val, "float32") if product_is_mini() else tmp_val
tmp_val = select(l1, tmp_val, gradient_accum)
# 4.x_value = -1 * lr * tmp_val
x_value = topi.multiply(lr, tvm.const(-1, "float32"))
x_value = topi.multiply(x_value, tmp_val)
# 5.y_value = l2 * global_step * lr + sqrt(grad_squared_accum)
pro_val = topi.multiply(l2, global_step)
pro_val = topi.multiply(pro_val, lr)
sqrt_val = sqrt(gradient_squared_accum, target=utils.CCE)
y_value = topi.add(pro_val, sqrt_val)
# 6.var = x_value / y_value
if product_is_mini():
y_rec = reciprocal(y_value, target=utils.CCE)
var_out = topi.multiply(x_value, y_rec)
else:
var_out = topi.divide(x_value, y_value)
if dtype == "float16":
var_out = akg.lang.ascend.cast_to(var_out, "float16")
gradient_accum = akg.lang.ascend.cast_to(gradient_accum, "float16")
gradient_squared_accum = akg.lang.ascend.cast_to(
gradient_squared_accum, "float16")
return var_out, gradient_accum, gradient_squared_accum
def _compute_var(var, lr, update):
"""Update var."""
lt_tmp = tvm.compute(update.shape, lambda *indice: update(*indice) * lr[0])
var_t = topi.subtract(var, lt_tmp)
return var_t
