def add_node_to_graph_recursive(self, current: Any, graph: Graph, visited: Set[Any], added: Dict[str, Operator],
data_format: str, nodes_to_remove) \
-> Operator:
if current in visited:
return added[current.name]
# return current
added_op_dic: Dict[str, Operator] = {}
current_format, input_formats = self._get_format(current, data_format)
inputs = self.find_inputs(current)
for in_put, in_format in zip(inputs, input_formats):
in_op = self.add_node_to_graph_recursive(in_put, graph, visited,
added, in_format,
nodes_to_remove)
added_op_dic[in_op.name] = in_op
op = self.create_new_op(current, added_op_dic, current_format,
input_formats, nodes_to_remove)
graph.add_op(op)
visited.add(current)
added[op.name] = op
return op
def pass_constant_folding(graph: Graph) -> None:
"""Given a node N, if the value of each input of N is known at compilation time then N will be executed.
The node N and its inputs will be replaced with a Constant node which holds the computed output of N.
Args:
graph (Graph): The input graph. It will be modified in-place.
processed_nodes (list): The list of the processed nodes so far.
"""
done = False
processed_nodes = []
while not done:
exec_list = sort_graph(graph)
processed_before_precompute = len(processed_nodes)
to_be_removed = []
for m in exec_list:
if m in processed_nodes:
continue
# We want operators with inputs
if not m.input_nodes:
continue
precomputable = True
for input_node in m.input_nodes:
if input_node.op_type != 'Constant':
precomputable = False
if not precomputable:
continue
processed_nodes += m.input_nodes
processed_nodes.append(m)
data = m.run_forward()
new_constant = Constant(m.name + '_new',
m.dtype,
data,
dimension_format=m.dimension)
graph.add_op(new_constant)
# get nodes to be removed after being disconnected
get_nodes_in_branch(m, None, to_be_removed)
new_constant.add_outputs({'output': m.output_ops.values()})
for output_name, consumer_list in m.output_ops.items():
for consumer_node in consumer_list:
for input_name, input_node in consumer_node.input_ops.items(
):
if input_node == m:
consumer_node.add_input(input_name, new_constant)
break
for op in to_be_removed:
graph.remove_op(op)
done = len(processed_nodes) == processed_before_precompute
def test_graph_conv(self) -> None:
"""Test code for making a simple graph with Conv."""
graph = Graph()
# two inputs
x = Input(
'input',
[1, 5, 5, 3],
Float32(),
)
w = Constant('weight', Float32(), np.zeros([1, 2, 2, 3]))
# Conv
conv = Conv(
'conv',
[1, 4, 4, 3],
Float32(),
{
'X': x,
'W': w
}, # you can get these keys by 'Conv.input_names'
kernel_shape=[2, 2])
# One output
y = Output(
'output',
[1, 4, 4, 3],
Float32(),
{'input': conv} # you can get this key by 'Output.input_names'
)
# add ops to the graph
graph.add_op(x)
graph.add_op(w)
graph.add_op(conv)
graph.add_op(y)
self.assertTrue(graph.check_nodes(),
"All inputs of operators must match their outputs.")
print("Graph test passed!")
def pass_lookup(graph: Graph) -> None:
"""Lookup.
Parameters
----------
graph : Graph
The input graph. It will be modified in-place.
"""
quantization_types = [
'QTZ_binary_mean_scaling', 'QTZ_linear_mid_tread_half',
'QTZ_binary_channel_wise_mean_scaling'
]
to_be_removed = []
exec_list = [
n for n in sort_graph(graph) if n.op_type in quantization_types
]
placeholder = [n for n in sort_graph(graph) if n.op_type in 'Input']
for m in exec_list:
quantizer = m
p1 = quantizer.input_nodes[0]
if p1.op_type != 'Reshape':
continue
p2 = p1.input_nodes[0]
if p2.op_type != 'Reshape':
continue
p3 = p2.input_nodes[0]
if p3.op_type != 'Gather':
continue
p4 = p3.input_nodes[0]
if p4.op_type != 'Gather':
continue
gather_params = p4.input_nodes[0]
if gather_params.rank != 2 or gather_params.shape[0] != 256:
continue
params = gather_params.data
data = {'data': params}
qtz_data = quantizer.run(**data)['data']
word_size = 32
lu_bitwidth = quantizer.nbit
packer = Packer(lu_bitwidth, word_size)
lsb = np.zeros((256, ), np.uint32)
msb = np.zeros((256, ), np.uint32)
idx = 0
for p in qtz_data:
data = packer.run(p.astype(np.float32), p.shape).flatten()
lsb[idx] = data[0]
msb[idx] = data[1]
idx += 1
pe_lsb = Constant('pe_lsb_new',
QUANTIZED_PACKED_KERNEL(),
lsb,
dimension_format='TC',
packed=True,
actual_shape=[256, word_size])
pe_msb = Constant('pe_msb_new',
QUANTIZED_PACKED_KERNEL(),
msb,
dimension_format='TC',
packed=True,
actual_shape=[256, word_size])
n, h, w, c = quantizer.shape
shape = [1, h, w, 2, word_size]
pe = Lookup('Lookup',
shape,
QUANTIZED_PACKED(), {
'input': placeholder[0],
'lsb': pe_lsb,
'msb': pe_msb
},
dimension_format='ChHWBCl')
get_nodes_in_branch(quantizer, placeholder[0], to_be_removed)
placeholder[0].remove_output('output')
placeholder[0].add_output('output', pe)
pe.add_outputs(quantizer.output_ops)
output_op = quantizer.output_op_list[0]
target_input_name = 'X'
for input_name in output_op._input_names:
if quantizer.equals(output_op._input_ops[input_name]):
target_input_name = input_name
break
output_op.add_input(target_input_name, pe)
graph.add_op(pe_lsb)
graph.add_op(pe_msb)
graph.add_op(pe)
for op in to_be_removed:
graph.remove_op(op)
def pass_pack_weights(graph: Graph) -> None:
"""Given a Quantized convolution node C, it will pack the weights of C into 32 bit words.
If the node Q that apply quantization to the weights of C quantizes, for example, into 1 bit values
then one 32 bit word will contain 32 weights.
Parameters
----------
graph : Graph
The input graph. It will be modified in-place.
"""
exec_list = [n for n in sort_graph(graph) if n.op_type == 'Conv']
quantization_types = [
'QTZ_binary_mean_scaling', 'QTZ_linear_mid_tread_half',
'QTZ_binary_channel_wise_mean_scaling'
]
word_size = 32
weight_bitwidth = 1
packer = Packer(weight_bitwidth, word_size)
to_be_removed = []
b = 32
for m in exec_list:
conv_node = m
# check if this is a quantized convolution
if not conv_node.quantizer or not conv_node.a_quantizer:
continue
# Check if we support this kind of quantizer
weight_quantizer = conv_node.quantizer
if weight_quantizer.op_type not in quantization_types:
continue
# Quantize the weights
weight_quantizer.run_forward()
def pad_to_multiple_of_b(tensor, axis, b):
shape = list(tensor.shape)
pad = (((shape[axis] + b - 1) // b) * b) - shape[axis]
shape[axis] = pad
return np.zeros(shape) if pad else None
padded_data = np.copy(weight_quantizer.data)
for axis in [0, 3]:
pad_tensor = pad_to_multiple_of_b(padded_data, axis, b)
if pad_tensor is not None:
padded_data = np.append(padded_data, pad_tensor, axis=axis)
tca_output = np.copy(padded_data)
oc, kh, kw, kd = padded_data.shape[:]
padded_data = padded_data.flatten()
tca_output = tca_output.flatten()
out_index = 0
for g in range(oc // b):
for p in range(kd // b):
for h in range(kh):
for w in range(kw):
for o in range(b):
for d in range(b):
idx = g * (kw * kh * kd * b) + p * b + h * (
kw * kd) + w * kd + o * (kw * kh * kd) + d
tca_output[out_index] = padded_data[idx]
out_index += 1
kn2row_output = np.zeros(oc * kh * kw * kd)
out_index = 0
for h in range(kh):
for w in range(kw):
for o in range(oc):
for i in range(kd):
idx = o * kh * kw * kd + h * kw * kd + w * kd + i
kn2row_output[out_index] = padded_data[idx]
out_index += 1
op_data = weight_quantizer.binarizer(padded_data)
data = packer.run(op_data.astype(np.float32),
weight_quantizer.dimension)
tca_binarized_data = weight_quantizer.binarizer(tca_output)
tca_packed_data = packer.run(tca_binarized_data.astype(np.float32),
weight_quantizer.dimension)
kn2row_binarized_data = weight_quantizer.binarizer(kn2row_output)
kn2row_data = packer.run(kn2row_binarized_data.astype(np.float32),
weight_quantizer.dimension)
shape = [oc, kh, kw, kd]
tca_shape = [oc // b, kd // b, kh, kw, b, b]
kn2row_shape = [kh, kw, oc, kd]
# Create the new constant with the quantized weights
quantized_constant = Constant(
weight_quantizer.name + '_new',
PackedUint32(),
data=np.vectorize(lambda k: (~k) & ((0x1 << 32) - 1))(data),
dimension_format="NHWC",
transposed_dimension_format="OhIhHWOlIl",
packed=True,
actual_shape=shape,
transposed_shape=tca_shape,
transposed_data=[(~k) & ((0x1 << 32) - 1)
for k in tca_packed_data.flatten()],
kn2row_data=[k for k in kn2row_data.flatten()],
kn2row_shape=kn2row_shape,
kn2row_dimension_format="HWNC")
# get nodes to be removed after being disconnected
get_nodes_in_branch(weight_quantizer, None, to_be_removed)
# Add the constant to the graph and connect the new constant
graph.add_op(quantized_constant)
quantized_constant.add_outputs(weight_quantizer.output_ops)
for output_name, consumer_list in weight_quantizer.output_ops.items():
for consumer_node in consumer_list:
for input_name, input_node in consumer_node.input_ops.items():
if input_node == weight_quantizer:
consumer_node.add_input(input_name, quantized_constant)
break
for op in to_be_removed:
graph.remove_op(op)
def pass_pack_weights(graph: Graph) -> None:
"""Given a Quantized convolution node C, it will pack the weights of C into 32 bit words.
If the node Q that apply quantization to the weights of C quantizes, for example, into 1 bit values
then one 32 bit word will contain 32 weights.
Parameters
----------
graph : Graph
The input graph. It will be modified in-place.
"""
exec_list = [n for n in sort_graph(graph) if n.op_type == 'Conv']
quantization_types = [
'QTZ_binary_mean_scaling', 'QTZ_linear_mid_tread_half',
'QTZ_binary_channel_wise_mean_scaling'
]
word_size = 32
weight_bitwidth = 1
packer = Packer(weight_bitwidth, word_size)
to_be_removed = []
for m in exec_list:
conv_node = m
# check if this is a quantized convolution
if not conv_node.quantizer or not conv_node.a_quantizer:
continue
# Check if we support this kind of quantizer
weight_quantizer = conv_node.quantizer
if weight_quantizer.op_type not in quantization_types:
continue
# Quantize the weights
weight_quantizer.run_forward()
op_data = weight_quantizer.binarizer(weight_quantizer.data)
data = packer.run(op_data.astype(np.float32),
weight_quantizer.dimension)
# Create the new constant with the quantized weights
oh = conv_node.height
ow = conv_node.width
od = conv_node.channel
kh = conv_node.kernel_height
kw = conv_node.kernel_width
kd = conv_node.input_ops['X'].channel
quantized_constant = Constant(weight_quantizer.name + '_new',
Uint32(),
data,
packed=True,
actual_shape=weight_quantizer.shape,
transposed_data=_transpose_kernels(
data, oh, ow, od, kh, kw, kd))
# get nodes to be removed after being disconnected
get_nodes_in_branch(weight_quantizer, None, to_be_removed)
# Add the constant to the graph and connect the new constant
graph.add_op(quantized_constant)
quantized_constant.add_outputs(weight_quantizer.output_ops)
for output_name, consumer_list in weight_quantizer.output_ops.items():
for consumer_node in consumer_list:
for input_name, input_node in consumer_node.input_ops.items():
if input_node == weight_quantizer:
consumer_node.add_input(input_name, quantized_constant)
break
for op in to_be_removed:
graph.remove_op(op)
def pass_simplify_batchnorm(graph: Graph) -> None:
"""Simplify BarchNorm operator.
"""
exec_list = [
x for x in sort_graph(graph) if x.op_type == 'BatchNormalization'
]
to_be_removed = []
for node in exec_list:
scale = node.input_ops['scale']
if scale.op_type != 'Constant':
raise RuntimeError('scale for BatchNormalization must be Constant')
B = node.input_ops['B']
if B.op_type != 'Constant':
raise RuntimeError('B for BatchNormalization must be Constant')
mean = node.input_ops['mean']
if mean.op_type != 'Constant':
raise RuntimeError('mean for BatchNormalization must be Constant')
var = node.input_ops['var']
if var.op_type != 'Constant':
raise RuntimeError('var for BatchNormalization must be Constant')
new_name = node.name + '_optimized'
new_scale_data = scale.data / np.sqrt(var.data + node.epsilon)
new_scale = Constant(new_name + '_scale',
scale.dtype,
new_scale_data,
dimension_format=scale.dimension)
new_bias_data = B.data - new_scale_data * mean.data
new_bias = Constant(new_name + '_bias',
B.dtype,
new_bias_data,
dimension_format=B.dimension)
new_op = BatchNormalizationOptimized(new_name,
node.shape,
node.dtype, {
'X': node.input_ops['X'],
'scale': new_scale,
'bias': new_bias
},
dimension_format=node.dimension)
new_scale.add_output('output', new_op)
new_bias.add_output('output', new_op)
input_op = node.input_ops['X']
update_key = None
new_outputs = [new_op]
for key, inout_ops in input_op.output_ops.items():
if node in inout_ops:
update_key = key
for op in inout_ops:
if op != node:
new_outputs.append(op)
if update_key is not None:
input_op.remove_output(update_key)
input_op.add_outputs({update_key: new_outputs})
out_ops = node.output_op_list
for op in out_ops:
update_key = None
for key, outin_op in op.input_ops.items():
if outin_op == node:
update_key = key
if update_key is not None:
op.add_input(update_key, new_op)
new_op.add_output('Y', op)
graph.add_op(new_scale)
graph.add_op(new_bias)
graph.add_op(new_op)
to_be_removed += [node, scale, B, mean, var]
for node in to_be_removed:
graph.remove_op(node)