def quantize(self, op, **kwargs):
features, buffers = kwargs['features'], kwargs['buffers']
cfg_dict = kwargs['cfg_dict']
name, op_name = op.attr('name'), op.attr('op_name')
attr, childs = op.list_attr(), sym_iter(op.get_children())
cns = [c.attr('name') for c in childs]
xquant_type = cfg_dict[cns[0]]['quant_type']
wquant_type = cfg_dict[cns[1]]['quant_type']
xquant, wquant = \
get_quantizer(xquant_type), get_quantizer(wquant_type)
oprec = kwargs['op_input_precs'][op_name]
if xquant_type == wquant_type == USQuantizer.name:
X, xs = childs[0], buffers[cns[0]]
if xs != 1:
X, _, _ = xquant.quantize(X,
32,
oscale=1,
oname=name,
**kwargs)
W, _, ws = wquant.quantize(childs[1], oprec, oname=name, **kwargs)
features[name] = features[cns[1]]
buffers[name] = SBuffer(ws)
kwargs['precs'][name][OUT_KEY] = get_bit_exp(features[name].get() *
ws)
op = get_mxnet_op(op_name)(X, W, **attr, name=name)
else:
raise NotImplementedError(
"Quantization type not implementated," + \
" op: %20s, Xquant: %20s, Wquant: %20s",
op_name, [xquant_type, wquant_type])
return op
def quantize(self, op, **kwargs):
infer_shapes = kwargs['infer_shapes']
buffers = kwargs['buffers']
cfg_dict = kwargs['cfg_dict']
name, op_name = op.attr('name'), op.attr('op_name')
childs, attr = sym_iter(op.get_children()), op.list_attr()
cns = [c.attr('name') for c in childs] if childs else []
oshp = infer_shapes[name][get_entry_id(op)]
quant_type = cfg_dict[cns[0]]['quant_type']
assert quant_type == USQuantizer.name, (quant_type, name, op_name)
quant = get_quantizer(quant_type)
oprec = kwargs['op_input_precs'][op_name]
X, xprec, xs = quant.quantize(childs[0], oprec, oname=name, **kwargs)
buffers[name] = SBuffer(xs)
op = get_mxnet_op(op_name)(X, **attr, name=name)
ishp = infer_shapes[cns[0]][get_entry_id(childs[0])]
k = int(nd.prod(nd_array(ishp)).asscalar() / \
nd.prod(nd_array(oshp)).asscalar())
kprec = get_bit_cnt_exp(k)
infer_prec = kprec + xprec
kwargs['precs'][name][OUT_KEY] = infer_prec
logger = logging.getLogger('log.mrt.realize')
logger.debug("operator %-20s name=%-40s oscale=%s, iscale=%s",
op_name, name, buffers[name].serialize(), cns)
return op
def mergefunc(node, params, graph):
name, op_name = node.attr('name'), node.attr('op_name')
childs, attr = sutils.sym_iter(
node.get_children()), node.list_attr()
if op_name in attribute_deps:
attr_deps = attribute_deps[op_name]
for attr_name, v in attr_deps.items():
val = sutils.get_attr(attr, attr_name, 0)
attr[attr_name] = int(val * mrt_oscales[name_idx[v]])
node = sutils.get_mxnet_op(op_name)(*childs, **attr, name=name)
return node
def _quantize_scale(op, **kwargs):
features, precs = kwargs['features'], kwargs['precs']
buffers, cfg_dict = kwargs['buffers'], kwargs['cfg_dict']
name, op_name = op.attr('name'), op.attr('op_name')
attr, childs = op.list_attr(), sym_iter(op.get_children())
cns = [c.attr('name') for c in childs] if childs else []
assert all([features[cn].name == FT_TYPE_EXP for cn in cns])
absmax = max([features[cn].get() for cn in cns])
oprec = kwargs['op_input_precs'][op_name]
oscale = scale_exp(absmax, oprec)
buffers[name] = SBuffer(oscale)
nodes, cprecs = [], []
assert all([cfg_dict[cn]['quant_type'] == \
USQuantizer.name for cn in cns])
quant = get_quantizer(USQuantizer.name)
for c in childs:
c, cprec, _ = quant.quantize(c,
oprec,
oscale=oscale,
oname=name,
**kwargs)
cprecs.append(cprec)
nodes.append(c)
if op_name in [
Concat.op_name, BroadcastAdd.op_name, ElemwiseAdd.op_name,
ElemwiseSub.op_name, SliceLike.op_name
]:
op = get_mxnet_op(op_name)(*nodes, **attr, name=name)
infer_prec = max(cprecs) if op_name == Concat.op_name \
else max(cprecs)+1
elif op_name == AddN.op_name:
while len(nodes) > 1:
tname = N.n('elemwise_add') if len(nodes) > 2 else name
a, b = nodes.pop(0), nodes.pop(0)
tmp = mx.sym.elemwise_add(a, b, name=tname)
nodes.append(tmp)
kprec = get_bit_cnt_exp(len(nodes))
infer_prec = max(cprecs) + kprec
op = nodes[0]
else:
raise NotImplementedError(
"symbol merge function of op_name: %s has not been " + \
"implemented, name: %s", op_name, name)
precs[name][OUT_KEY] = infer_prec
logger = logging.getLogger('log.mrt.realize')
logger.debug("operator %-20s name=%-40s oscale=%s, iscale=%s", op_name,
name, buffers[name].serialize(), cns)
return op
def _separate_pad(op, **kwargs):
name, op_name = op.attr('name'), op.attr('op_name')
attr, childs = op.list_attr(), sutils.sym_iter(op.get_children())
if op_name not in [Convolution.op_name]:
return op
if 'layout' in attr:
assert attr['layout'] == 'NCHW'
PH, PW = sutils.get_attr(attr, 'pad', (0, 0))
if 'pad' in attr:
del attr['pad']
if PH == 0 and PW == 0:
return sutils.get_mxnet_op(op_name)(*childs, **attr, name=name)
childs[0] = mx.sym.pad(childs[0],
pad_width=(0, 0, 0, 0, PH, PH, PW, PW),
mode='constant',
constant_value=0,
name=N.n('pad'))
op = sutils.get_mxnet_op(op_name)(*childs, **attr, name=name)
return op
def quantize(self, op, **kwargs):
precs, buffers = kwargs['precs'], kwargs['buffers']
name, op_name = op.attr('name'), op.attr('op_name')
cfg_dict = kwargs['cfg_dict']
childs, attr = sym_iter(op.get_children()), op.list_attr()
cns = [c.attr('name') for c in childs] if childs else []
oprec = kwargs['op_input_precs'][op_name]
xquant_type, bquant_type = \
cfg_dict[cns[0]]['quant_type'], cfg_dict[cns[1]]['quant_type']
xquant, bquant = \
get_quantizer(xquant_type), get_quantizer(bquant_type)
if xquant_type == bquant_type == USQuantizer.name:
X, xprec, xs = xquant.quantize(childs[0],
oprec,
oname=name,
**kwargs)
B, bprec, bs = bquant.quantize(childs[1],
oprec,
oname=name,
**kwargs)
op = get_mxnet_op(op_name)(X, B, **attr, name=name)
if bprec == 1 and bs == 1:
# special case: childs[1] is 0
buffers[name] = SBuffer(1)
precs[name][OUT_KEY] = 1
else:
buffers[name] = SBuffer(xs * bs)
infer_prec = xprec + bprec
precs[name][OUT_KEY] = infer_prec
else:
raise NotImplementedError(
"Quantization type not implementated," + \
" op: %20s, Xquant: %20s, Wquant: %20s",
op_name, [xquant_type, bquant_type])
logger = logging.getLogger('log.mrt.realize')
logger.debug("operator %-20s name=%-40s oscale=%s, iscale=%s",
op_name, name, buffers[name].serialize(), cns)
return op
def restore(op, **kwargs):
features, precs, buffers = \
kwargs['features'], kwargs['precs'], kwargs['buffers']
name, op_name = op.attr('name'), op.attr('op_name')
childs, attr = sutils.sym_iter(op.get_children()), op.list_attr()
childs = [] if childs is None else childs
new_childs = []
for c in childs:
cname = c.attr('name')
sc = buffers[c.attr('name')].get() \
if cname in buffers else 1
new_childs.append(c if sc == 1 else c / sc)
out = sutils.get_mxnet_op(op_name)(*new_childs, **attr, name=name)
ft = features[name]
assert ft.name == FT_TYPE_EXP
absmax = features[name].get()
precs[name][OUT_KEY] = get_bit_exp(absmax)
buffers[name] = get_buffer_exp(1)
return out
def _separate_bias(op, **kwargs):
name, op_name = op.attr('name'), op.attr('op_name')
attr, childs = op.list_attr(), sutils.sym_iter(op.get_children())
if childs and len(childs) < 3 or op_name not in \
[Convolution.op_name, FullyConnected.op_name]:
return op
attr['no_bias'] = True
op = sutils.get_mxnet_op(op_name)(childs[0],
childs[1],
**attr,
name=N.n(name))
bn = childs[2].attr('name')
if op_name == Convolution.op_name:
if 'layout' in attr:
assert attr['layout'] == 'NCHW'
B = mx.sym.expand_dims(childs[2], axis=0, name=N.n('expand_dims'))
B = mx.sym.expand_dims(B, axis=-1, name=N.n('expand_dims'))
B = mx.sym.expand_dims(B, axis=-1, name=N.n(bn))
else:
B = mx.sym.expand_dims(childs[2], axis=0, name=N.n(bn))
op = mx.sym.broadcast_add(op, B, name=name)
return op
def _quantize_xw(op, **kwargs):
features, buffers = kwargs['features'], kwargs['buffers']
precs = kwargs['precs']
cfg_dict, params = kwargs['cfg_dict'], kwargs['params']
name, op_name = op.attr('name'), op.attr('op_name')
childs, attr = sym_iter(op.get_children()), op.list_attr()
cns = [c.attr('name') for c in childs] if childs else []
X, W = childs
Xquant = get_quantizer(cfg_dict[cns[0]]['quant_type'])
Wquant = get_quantizer(cfg_dict[cns[1]]['quant_type'])
oprec = kwargs['op_input_precs'][op_name]
Xq, xprec, xscale = Xquant.quantize(X, oprec, oname=name, **kwargs)
Wq, wprec, wscale = Wquant.quantize(W, oprec, oname=name, **kwargs)
buffers[name] = get_buffer_exp(xscale * wscale)
op = get_mxnet_op(op_name)(Xq, Wq, **attr, name=name)
shp = params[cns[1]].shape
k = int(nd.prod(nd_array(shp[1:])).asscalar())
kprec = get_bit_cnt_exp(k)
infer_prec = kprec + xprec + wprec
precs[name][OUT_KEY] = infer_prec
return op
def quantize(self, op, **kwargs):
features, buffers = kwargs['features'], kwargs['buffers']
precs, graph = kwargs['precs'], kwargs['graph']
cfg_dict, params = kwargs['cfg_dict'], kwargs['params']
name, op_name = op.attr('name'), op.attr('op_name')
childs, attr = sym_iter(op.get_children()), op.list_attr()
cns = [c.attr('name') for c in childs] if childs else []
# assert len(childs) == 2 and 'pad' not in attr
assert len(childs) == 2
xquant_type = cfg_dict[cns[0]]['quant_type']
wquant_type = cfg_dict[cns[1]]['quant_type']
X, W = childs
xquant, wquant = \
get_quantizer(xquant_type), get_quantizer(wquant_type)
oprec = kwargs['op_input_precs'][op_name]
if xquant_type == wquant_type == USQuantizer.name:
op = _quantize_xw(op, **kwargs)
elif xquant_type == USQuantizer.name and \
wquant_type == UAQuantizer.name:
Xq, xprec, xscale = xquant.quantize(X, oprec, oname=name, **kwargs)
Wq, wprec, wscale, wzpoint = wquant.quantize(W,
oprec,
oname=name,
**kwargs)
buffers[name] = get_buffer_exp(xscale * wscale)
Ye1 = mx.sym.Convolution(Xq, Wq, **attr, name=N.n('Convolution'))
wshp = params[cns[1]].shape
pd = int(np.product(wshp[1:]))
infer_prec1 = get_bit_cnt_exp(pd) + xprec + wprec
W1 = nd_full_const(1, wshp, graph, params)
Ye2 = mx.sym.Convolution(Xq, W1, **attr, name=N.n('Convolution'))
wzint = round(wzpoint * wscale)
Wz = nd_const(wzint, graph, params)
Ye2 = mx.sym.broadcast_mul(Wz, Ye2, name=N.n('broadcast_mul'))
infer_prec2 = get_bit_cnt_exp(pd) + xprec + get_bit_exp(wzint)
op = mx.sym.elemwise_add(Ye1, Ye2, name=name)
precs[name][OUT_KEY] = max(infer_prec1, infer_prec2) + 1
buffers[name] = get_buffer_exp(xscale * wscale)
elif xquant_type == UAQuantizer.name and \
wquant_type == USQuantizer.name:
Xq, xprec, xscale, Xzp = xquant.quantize(X,
oprec,
oname=name,
**kwargs)
Wq, wprec, wscale = wquant.quantize(W, oprec, oname=name, **kwargs)
buffers[name] = get_buffer_exp(xscale * wscale)
Y1 = mx.sym.Convolution(Xq, Wq, **attr, name=N.n('Convolution'))
wshp = params[cns[1]].shape
pd = np.product(wshp[1:])
infer_prec1 = get_bit_cnt_exp(pd) + xprec + wprec + 1
xshp = params[cns[0]].shape
X1 = nd_full(1, xshp, graph, params)
Y2 = mx.sym.Convolution(X1, Wq, **attr, name=N.n('Convolution'))
xzp = params[Xzp.attr('name')].asscalar()
infer_prec2 = get_bit_cnt_exp(abs(xzp) * pd) + wprec
op = mx.sym.elemwise_add(Y1, Y2, name=N.n('elemwise_add'))
infer_prec = max(infer_prec1, infer_prec2) + 1
precs[name][OUT_KEY] = infer_prec
elif xquant_type == wquant_type == UAQuantizer.name:
Xq, xprec, xscale, Xzp = xquant.quantize(X,
oprec,
oname=name,
**kwargs)
Wq, wprec, wscale, Wzp = wquant.quantize(W,
oprec,
oname=name,
**kwargs)
buffers[name] = get_buffer_exp(xscale * wscale)
nodes, infer_precs = [], []
Y1 = mx.sym.Convolution(Xq, Wq, **attr, name=N.n('Convolution'))
nodes.append(Y1)
wshp = params[cns[1]].shape
pd = np.product(wshp[1:])
infer_prec1 = get_bit_cnt_exp(pd) + xprec + wprec + 2
infer_precs.append(infer_prec1)
W1 = nd_full_const(1, wshp, graph, params)
Y2 = mx.sym.Convolution(Xq, W1, **attr, name=N.n('Convolution'))
Y2 = mx.sym.broadcast_mul(Wzp, Y2, name=N.n('broadcast_mul'))
nodes.append(Y2)
wzp = params[Wzp.attr('name')].asscalar()
infer_prec2 = get_bit_cnt_exp(abs(wzp) * pd) + xprec + 1
infer_precs.append(infer_prec2)
xshp = params[cns[0]].shape
X1 = nd_full_const(1, xshp, graph, params)
Y3 = mx.sym.Convolution(X1, Wq, graph, params)
Y3 = mx.sym.broadcast_mul(Xzp, Y3, name=N.n('broadcast_mul'))
nodes.append(Y3)
xzp = params[Xzp.attr('name')].asscalar()
infer_prec3 = get_bit_cnt_exp(abs(xzp) * pd) + wprec + 1
infer_precs.append(infer_prec3)
val = pd * abs(xzp) * abs(wzp)
Y4 = nd_const(val, graph, params)
nodes.append(Y4)
infer_prec4 = get_bit_cnt_exp(val)
infer_precs.append(infer_prec4)
while len(nodes) > 1:
a, b = nodes.pop(), nodes.pop()
node = mx.sym.broadcast_add(a, b, name=N.n('broadcast_add'))
nodes.append(node)
op = nodes[0]
infer_prec = max(infer_precs) + 2
precs[name][OUT_KEY] = infer_prec
elif xquant_type == GroupConvQuant.name and \
wquant_type == GroupConvQuant.name:
num_groups_x = cfg_dict[cns[0]]['gn_info']['num_groups']
num_groups_w = cfg_dict[cns[1]]['gn_info']['num_groups']
assert num_groups_x == num_groups_w, \
"num_groups of x and weight should be equal, " + \
"num_groups of x: {}, num_groups of weight: {}".format(
num_groups_x, num_groups_w)
Xq, xprec_list, xscale_list = xquant.quantize(
X, oprec, oname=name, num_groups=num_groups_x, **kwargs)
Wq, wprec_list, wscale_list = wquant.quantize(
W, oprec, oname=name, num_groups=num_groups_w, **kwargs)
op = get_mxnet_op(op_name)(Xq, Wq, **attr, name=name)
IPG = kwargs['infer_shapes'][cns[1]][get_entry_id(X)][1]
kprec = get_bit_cnt_exp(IPG)
infer_prec_list = [
kprec + wprec_list[i] + xprec_list[i] \
for i in range(len(wprec_list))
]
oscale_list = [
xscale_list[i] * wscale_list[i] \
for i in range(len(wscale_list))
]
assert False, "implementing..."
else:
raise NotImplementedError(
"Quantization type not implementated," + \
" op: {}, Xquant: {}, Wquant: {}".format(
op_name, xquant_type, wquant_type))
logger = logging.getLogger('log.mrt.realize')
logger.debug("operator %-20s name=%-40s oscale=%s, iscale=%s",
op_name, name, buffers[name].serialize(), cns)
return op
def slice_channel(self, op, **kwargs):
name, op_name = op.attr('name'), op.attr('op_name')
attr, childs = op.list_attr(), sym_iter(op.get_children())
cns = [c.attr('name') for c in childs]
cfg_dict = kwargs['cfg_dict']
infer_shapes = kwargs['infer_shapes']
gn_info = cfg_dict[name]['gn_info']
ichannel, step = gn_info['ichannel'], gn_info['step']
assert ichannel == 1
assert len(childs) == 2
X, W = childs
xshp = infer_shapes[cns[0]][get_entry_id(childs[0])]
wshp = infer_shapes[cns[1]][get_entry_id(childs[1])]
oshp = infer_shapes[name][get_entry_id(op)]
assert len(xshp) == len(wshp) == 4 and xshp[1] % step == 0
xi_cfg_info, wi_cfg_info = cfg_dict[cns[0]], cfg_dict[cns[1]]
xi_cfg_info['gn_info'] = {'gn_type': LAYER_WISE_TYPE}
wi_cfg_info['gn_info'] = {'gn_type': LAYER_WISE_TYPE}
yi_cfg_info = {
'gn_info': {
'gn_type': LAYER_WISE_TYPE
},
'quant_type': US_QUANT_TYPE,
'opt_info': cfg_dict[name]['opt_info'],
}
num_group = eval(attr['num_group'])
C, IC, OC = xshp[1], wshp[1], wshp[0]
assert num_group * IC == C and OC >= num_group and OC % num_group == 0
if num_group == 1:
xs = sym_slice(X, ichannel, step, **kwargs)
ws = sym_slice(W, ichannel, step, **kwargs)
nodes = []
j = 0
for i in range(0, C, step):
suffix = '_' + str(i) + '-' + str(i + step)
xni = xs[j].attr('name')
cfg_dict[xni] = xi_cfg_info
wni = ws[j].attr('name')
cfg_dict[wni] = wi_cfg_info
yni = N.n(name + suffix)
Yi = get_mxnet_op(op_name)(xs[j], ws[j], **attr, name=yni)
cfg_dict[yni] = yi_cfg_info
nodes.append(Yi)
j += 1
assert len(nodes) > 1
op = mx.sym.add_n(*nodes, name=name)
# # transpose and reshape weight
# Wt = mx.sym.transpose(W, axes=(1,0,2,3), name=N.n('transpose'))
# rshp = (OC*IC,1,) + wshp[2:]
# wrn = N.n('reshape')
# cfg_dict[wrn] = wi_cfg_info
# Wr = mx.sym.reshape(Wt, shape=rshp, name=wrn)
# # groupwise convolution
# nattr = attr.copy()
# nattr['num_group'] = IC
# nattr['num_filter'] = IC * OC
# conv_name = N.n('groupwise_convolution')
# cfg_dict[conv_name] = yi_cfg_info
# print(nattr, name)
# op = mx.sym.Convolution(X, Wr, **nattr, name=conv_name)
# # reshape output
# rname = N.n('reshape')
# cfg_dict[rname] = yi_cfg_info
# rshp = (-1, IC, OC,) + oshp[2:]
# op = mx.sym.reshape(op, shape=rshp, name=rname)
# # sum
# sum_name = N.n('sum')
# cfg_dict[sum_name] = yi_cfg_info
# op = mx.sym.sum(op, axis=1, keepdims=False, name=sum_name)
else:
assert step == 1
xs = sym_slice(X, ichannel, step, **kwargs)
ws = kernel_slice_2d(W, **kwargs)
OPG = OC // num_group
nattr = attr.copy()
nattr['num_group'] = '1'
nattr['num_filter'] = '1'
nodes = []
for o in range(OC):
nnodes = []
j = int(o / OPG) * IC
for i in range(IC):
suffix = '_' + str(o) + '-' + str(i)
k = i + j
xk, woi = xs[k], ws[o][i]
xnk, wnoi = xk.attr('name'), woi.attr('name')
cfg_dict[xnk] = xi_cfg_info
cfg_dict[wnoi] = wi_cfg_info
ynoi = N.n(name + suffix)
yoi = mx.sym.Convolution(xk, woi, **nattr, name=ynoi)
cfg_dict[ynoi] = yi_cfg_info
nnodes.append(yoi)
if len(nnodes) > 1:
zni = N.n(name + '_add_n_' + str(o))
zi = mx.sym.add_n(*nnodes, name=zni)
cfg_dict[zni] = yi_cfg_info
else:
zi = nnodes[0]
nodes.append(zi)
assert len(nodes) > 1
op = mx.sym.concat(*nodes, dim=1, name=name)
return op
def _quantize_scale_zp(op, **kwargs):
features, precs = kwargs['features'], kwargs['precs']
buffers, cfg_dict = kwargs['buffers'], kwargs['cfg_dict']
graph, params = kwargs['graph'], kwargs['params']
name, op_name = op.attr('name'), op.attr('op_name')
attr, childs = op.list_attr(), sym_iter(op.get_children())
cns = [c.attr('name') for c in childs] if childs else []
oprec = kwargs['op_input_precs'][op_name]
oscales = []
for c in childs:
cquant_type = cfg_dict[c.attr('name')]['quant_type']
cquant = get_quantizer(cquant_type)
ft = features[c.attr('name')]
oscale = cquant.get_scale(oprec, ft)
oscales.append(oscale)
oscale = min(oscales)
buffers[name] = SBuffer(oscale)
nodes, cprecs = [], []
for c in childs:
cquant_type = cfg_dict[c.attr('name')]['quant_type']
cquant = get_quantizer(cquant_type)
if cquant.name == USQuantizer.name:
c, cprec, _ = cquant.quantize(c,
oprec,
oscale=oscale,
oname=name,
**kwargs)
elif cquant.name == UAQuantizer.name:
c, cprec, cscale, czpoint = cquant.quantize(c,
oprec,
oscale=oscale,
oname=name,
**kwargs)
czint = round(czpoint * cscale)
Cz = nd_const(czint, graph, params)
nodes.append(Cz)
cprecs.append(get_bit_exp(czint))
cprecs.append(cprec)
nodes.append(c)
if op_name in [Concat.op_name]:
op = get_mxnet_op(op_name)(*nodes, **attr, name=name)
infer_prec = max(cprecs)
elif op_name in [BroadcastAdd.op_name]:
while len(nodes) > 1:
tname = N.n('broadcast_add') if len(nodes) > 2 else name
a, b = nodes.pop(0), nodes.pop(0)
tmp = mx.sym.broadcast_add(a, b, name=tname)
nodes.append(tmp)
kprec = get_bit_cnt_exp(len(nodes))
infer_prec = max(cprecs) + kprec
op = nodes[0]
elif op_name in [AddN.op_name]:
while len(nodes) > 1:
tname = N.n('elemwise_add') if len(nodes) > 2 else name
a, b = nodes.pop(0), nodes.pop(0)
tmp = mx.sym.elemwise_add(a, b, name=tname)
nodes.append(tmp)
kprec = get_bit_cnt_exp(len(nodes))
infer_prec = max(cprecs) + kprec
op = nodes[0]
else:
raise NotADirectoryError(
"symbol merge function of op_name: %s has not been " + \
"implemented, name: %s", op_name, name)
precs[name][OUT_KEY] = infer_prec
logger = logging.getLogger('log.mrt.realize')
logger.debug("operator %-20s name=%-40s oscale=%s, iscale=%s", op_name,
name, buffers[name].serialize(), cns)
return op