def act_infos(gen,
pnode,
fnode,
act_params,
act_q,
extra1=0,
extra2=0,
extra3=0,
extra4=0):
if isinstance(pnode, FilterParameters):
comment = str.format("BiasQ: {}", extra1)
elif isinstance(pnode, MatrixAddParameters):
comment = str.format(
"In1Scale: {} In1ScaleN: {} OutScale: {} OutScaleN: {}", extra1,
extra2, extra3, extra4)
else:
comment = ""
if act_params is None:
contents = np.array([0, 0, 0, 0, 0, extra1, extra2, extra3, extra4],
dtype=np.int8)
elif isinstance(act_params, ReluActivationParameters):
actscale = act_q.scale_mul_biases_q.qbiases[0]
actscalen = act_q.scale_mul_biases_q.qnorms[0]
if act_params.upper_bound is None: # or fnode is not None:
contents = np.array(
[actscale, actscalen, 0, 0, 0, extra1, extra2, extra3, extra4],
dtype=np.int8)
if len(comment) == 0:
comment = "all 0"
else:
fac_1 = act_q.in_qs[0].quantize(act_params.upper_bound)
contents = np.array([
actscale, actscalen, fac_1, 0, 0, extra1, extra2, extra3,
extra4
],
dtype=np.int8)
comment += str.format("in: {:05f} out: {:05f} A0: {} B0: 0 C0: 0",
act_q.in_qs[0].scale[0],
act_q.out_qs[0].scale[0], fac_1[0])
elif isinstance(act_params, HSigmoidActivationParameters):
# currently combines all scaling factors into one scale and shift
fac_1, upper_bound, _ = hsigmoid_mult_gen_factors(act_params, act_q)
contents = np.array([
act_q.scale_mul_biases_q.qbiases[0],
act_q.scale_mul_biases_q.qnorms[0], upper_bound, fac_1, 1, extra1,
extra2, extra3, extra4
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: {} B0: {} C0: 1",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.scale_mul_biases_q.qbiases[0],
act_q.scale_mul_biases_q.qnorms[0], upper_bound[0], fac_1[0])
elif isinstance(act_params, HSwishActivationParameters):
# currently combines all scaling factors into one scale and shift
fac_1, upper_bound, _ = hswish_mult_gen_factors(act_q)
contents = np.array([
act_q.scale_mul_biases_q.qbiases[0],
act_q.scale_mul_biases_q.qnorms[0], upper_bound, fac_1, 1, extra1,
extra2, extra3, extra4
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: {} B0: {} C0: 1",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.scale_mul_biases_q.qbiases[0],
act_q.scale_mul_biases_q.qnorms[0], upper_bound[0], fac_1[0])
elif isinstance(act_params, SoftMaxParameters):
norm = 15 + np.ceil(np.log2(act_q.in_qs[0].scale))
contents = np.array([norm, 0, 0, 0, 0, extra1, extra2, extra3, extra4],
dtype=np.int8)
comment += str.format("in: {:05f} out: {:05f} NORM: {}",
act_q.in_qs[0].scale[0],
act_q.out_qs[0].scale[0], int(norm[0]))
elif isinstance(act_params, LeakyActivationParameters):
act_q.set_scale()
leak_factor_quant = leak_mult_gen_factor_q7(act_params)
contents = np.array([
act_q.scale_mul_biases_q.qbiases[0],
act_q.scale_mul_biases_q.qnorms[0], leak_factor_quant, 0, 0,
extra1, extra2, extra3, extra4
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: {} B0: x C0: x",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.scale_mul_biases_q.qbiases[0],
act_q.scale_mul_biases_q.qnorms[0], leak_factor_quant)
else:
raise NotImplementedError("activation tye not implemented")
if isinstance(pnode, (GlobalPoolParameters, PoolingParameters)):
contents = np.array([
act_q.scale_mul_biases_q.qbiases[0],
act_q.scale_mul_biases_q.qnorms[0], 0, 0, 0, extra1, extra2,
extra3, extra4
],
dtype=np.int8)
comment += str.format("in: {:05f} out: {:05f}",
act_q.in_qs[0].scale[0],
act_q.out_qs[0].scale[0])
cname, file_name = gen_constant(gen, pnode, fnode, INFOS)
const_info = ConstantInfo(file_name,
QType(bits=8, q=0, signed=True),
contents=contents)
gen.globals.append(
GlobalArgInfo("int8",
cname,
gen.opts['default_global_home_location'],
gen.opts['default_global_exec_location'],
const_info=const_info,
comment=comment))
def gen_act_infos(act_params, act_q):
comment = ""
if isinstance(act_params, ReluActivationParameters):
compute_in_out_scale(act_q)
actscale = act_q.cache['scale_mul_biases_q'].qbiases[0]
actscalen = act_q.cache['scale_mul_biases_q'].qnorms[0]
if act_params.upper_bound is None: # or fnode is not None:
if act_q.in_qs[0].zero_point == 0:
contents = np.array([actscale, actscalen, 0, 0, 0],
dtype=np.int8)
if len(comment) == 0:
comment = "all 0"
else:
fac_1 = act_q.in_qs[0].zero_point
contents = np.array([actscale, actscalen, fac_1, 0, 0],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} A0: {} B0: 0 C0: 0",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
fac_1[0])
else:
if act_q.in_qs[0].zero_point == 0:
fac_1 = act_q.in_qs[0].quantize(act_params.upper_bound)
contents = np.array([actscale, actscalen, fac_1, 0, 0],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} A0: {} B0: 0 C0: 0",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
fac_1[0])
else:
fac_1 = act_q.in_qs[0].zero_point
fac_2 = act_q.in_qs[0].quantize(act_params.upper_bound)
contents = np.array([actscale, actscalen, fac_1, fac_2, 0],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} A0: {} B0: {} C0: 0",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
fac_1[0], fac_2[0])
elif isinstance(act_params, HSigmoidActivationParameters):
# currently combines all scaling factors into one scale and shift
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
fac_1, upper_bound, _ = hsigmoid_mult_gen_factors(act_params, act_q)
contents = np.array([
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], upper_bound, fac_1, 1
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: {} B0: {} C0: 1",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], upper_bound[0],
fac_1[0])
elif isinstance(act_params, HSwishActivationParameters):
# currently combines all scaling factors into one scale and shift
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
fac_1, upper_bound, _ = hswish_mult_gen_factors(act_q)
contents = np.array([
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], upper_bound, fac_1, 1
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: {} B0: {} C0: 1",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], upper_bound[0],
fac_1[0])
elif isinstance(act_params, SoftMaxParameters):
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
norm = 15 + np.ceil(np.log2(act_q.in_qs[0].scale))
contents = np.array([norm, 0, 0, 0, 0], dtype=np.int8)
comment += str.format("in: {:05f} out: {:05f} NORM: {}",
act_q.in_qs[0].scale[0],
act_q.out_qs[0].scale[0], int(norm[0]))
elif isinstance(act_params, LeakyActivationParameters):
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
compute_in_out_scale(act_q)
leak_factor_quant = leak_mult_gen_factor_q7(act_params)
contents = np.array([
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], leak_factor_quant, 0,
0
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: {} B0: x C0: x",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], leak_factor_quant)
elif isinstance(act_params,
(SigmoidActivationParameters, TanHActivationParameters)):
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
compute_in_out_scale(
act_q,
extra_scale=QType.Pow2(bits=32, q=7, signed=True).scale /
act_q.in_qs[0].scale)
contents = np.array([
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], 0, 0, 0
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: x B0: x C0: x",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0])
else:
raise NotImplementedError("activation tye not implemented")
return contents, comment
def act_infos(gen,
pnode,
fnode,
act_params,
act_q,
extra1=0,
extra2=0,
extra3=0,
extra4=0,
extra5=None,
extra6=None,
prenorm=0,
extra_name='',
for_ne16=False,
in_zero_point=0):
if isinstance(pnode, FilterParameters):
comment = str.format("BiasQ: {}", extra1)
elif isinstance(pnode, MatrixAddParameters):
comment = str.format(
"In1Scale: {} In1ScaleN: {} OutScale: {} OutScaleN: {}", extra1,
extra2, extra3, extra4)
else:
comment = ""
if act_params is None:
contents = np.array([0, 0, 0, 0, 0], dtype=np.int8)
elif isinstance(act_params, ReluActivationParameters):
compute_in_out_scale(act_q)
actscale = act_q.cache['scale_mul_biases_q'].qbiases[0]
actscalen = act_q.cache['scale_mul_biases_q'].qnorms[0]
if act_params.upper_bound is None: # or fnode is not None:
if act_q.in_qs[0].zero_point == 0:
contents = np.array([actscale, actscalen, 0, 0, 0],
dtype=np.int8)
if len(comment) == 0:
comment = "all 0"
else:
fac_1 = act_q.in_qs[0].zero_point
contents = np.array([actscale, actscalen, fac_1, 0, 0],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} A0: {} B0: 0 C0: 0",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
fac_1[0])
else:
if act_q.in_qs[0].zero_point == 0:
fac_1 = act_q.in_qs[0].quantize(act_params.upper_bound)
contents = np.array([actscale, actscalen, fac_1, 0, 0],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} A0: {} B0: 0 C0: 0",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
fac_1[0])
else:
fac_1 = act_q.in_qs[0].zero_point
fac_2 = act_q.in_qs[0].quantize(act_params.upper_bound)
contents = np.array([actscale, actscalen, fac_1, fac_2, 0],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} A0: {} B0: {} C0: 0",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
fac_1[0], fac_2[0])
elif isinstance(act_params, HSigmoidActivationParameters):
# currently combines all scaling factors into one scale and shift
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
fac_1, upper_bound, _ = hsigmoid_mult_gen_factors(act_params, act_q)
contents = np.array([
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], upper_bound, fac_1, 1
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: {} B0: {} C0: 1",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], upper_bound[0],
fac_1[0])
elif isinstance(act_params, HSwishActivationParameters):
# currently combines all scaling factors into one scale and shift
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
fac_1, upper_bound, _ = hswish_mult_gen_factors(act_q)
contents = np.array([
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], upper_bound, fac_1, 1
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: {} B0: {} C0: 1",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], upper_bound[0],
fac_1[0])
elif isinstance(act_params, SoftMaxParameters):
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
norm = 15 + np.ceil(np.log2(act_q.in_qs[0].scale))
contents = np.array([norm, 0, 0, 0, 0], dtype=np.int8)
comment += str.format("in: {:05f} out: {:05f} NORM: {}",
act_q.in_qs[0].scale[0],
act_q.out_qs[0].scale[0], int(norm[0]))
elif isinstance(act_params, LeakyActivationParameters):
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
compute_in_out_scale(act_q)
leak_factor_quant = leak_mult_gen_factor_q7(act_params)
contents = np.array([
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], leak_factor_quant, 0,
0
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: {} B0: x C0: x",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], leak_factor_quant)
elif isinstance(act_params,
(SigmoidActivationParameters, TanHActivationParameters)):
assert act_q.in_qs[0].zero_point == 0 and act_q.out_qs[
0].zero_point == 0, "asymmetric not supported"
compute_in_out_scale(
act_q,
extra_scale=QType.Pow2(bits=32, q=7, signed=True).scale /
act_q.in_qs[0].scale)
contents = np.array([
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], 0, 0, 0
],
dtype=np.int8)
comment += str.format(
"in: {:05f} out: {:05f} qbias: {} qnorm: {} A0: x B0: x C0: x",
act_q.in_qs[0].scale[0], act_q.out_qs[0].scale[0],
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0])
else:
raise NotImplementedError("activation tye not implemented")
if isinstance(pnode, (GlobalPoolingParameters, PoolingParameters)):
contents = np.array([
act_q.cache['scale_mul_biases_q'].qbiases[0],
act_q.cache['scale_mul_biases_q'].qnorms[0], 0, 0, 0
],
dtype=np.int8)
contents = np.append(contents, [extra1, extra2, extra3, extra4])
if extra5 is not None:
contents = np.append(contents, [extra5])
if extra6 is not None:
contents = np.append(contents, [extra6])
if for_ne16:
# append weights_offset and pad_val for ne16
# TODO - default config maybe in future
if isinstance(pnode, (ConvFusionParameters, LinearFusionParameters)):
filt_q = gen.G.quantization[NodeId(pnode, fnode)]
else:
filt_q = gen.G.quantization[NodeId(pnode)]
pad_value = np.array(in_zero_point).astype(np.int16)
pad_value1 = np.bitwise_and(pad_value, 0xFF)
pad_value2 = np.bitwise_and(pad_value, 0xFF00) >> 8
w_offset = -np.array(filt_q.in_qs[1].zero_point).astype(np.int32)
w_offset1 = np.bitwise_and(w_offset, 0xFF)
w_offset2 = np.bitwise_and(w_offset, 0xFF00) >> 8
w_offset3 = np.bitwise_and(w_offset, 0xFF0000) >> 16
w_offset4 = np.bitwise_and(w_offset, 0xFF000000) >> 24
contents = np.append(
contents, [[prenorm] if prenorm else [0], pad_value1, pad_value2,
w_offset1, w_offset2, w_offset3, w_offset4])
cname, file_name = gen_constant(gen, pnode, fnode, INFOS, extra_name)
const_info = ConstantInfo(file_name,
QType.Pow2(bits=8, q=0, signed=True),
contents=contents)
gen.globals.append(
GlobalArgInfo("int8",
cname,
gen.opts['default_global_home_location'],
gen.opts['default_global_exec_location'],
const_info=const_info,
comment=comment))