def test_dnnlowp_elementwise_sum_int(self, N, M, is_empty, gc, dc):
if is_empty:
N = 0
# All inputs have scale 1, so exactly represented after quantization
inputs = M * [None]
X_names = M * [None]
X_q_names = M * [None]
for i in range(M):
X = np.random.randint(-128, 127, N, np.int8).astype(np.float32)
if N != 0:
X[0] = -128
X[-1] = 127
inputs[i] = X
X_names[i] = chr(ord("A") + i)
X_q_names[i] = X_names[i] + "_q"
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [("Sum", ""), ("Sum", "DNNLOWP"), ("Int8Sum", "DNNLOWP")]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
if engine == "DNNLOWP":
for i in range(M):
quantize = core.CreateOperator(
"Quantize",
X_names[i],
X_q_names[i],
engine=engine,
device_option=gc,
)
net.Proto().op.extend([quantize])
sum_ = core.CreateOperator(
op_type,
X_q_names if engine == "DNNLOWP" else X_names,
["Y_q" if engine == "DNNLOWP" else "Y"],
engine=engine,
device_option=gc,
)
net.Proto().op.extend([sum_])
if engine == "DNNLOWP":
dequantize = core.CreateOperator(
"Dequantize", ["Y_q"], ["Y"], engine=engine, device_option=gc
)
net.Proto().op.extend([dequantize])
for i in range(M):
self.ws.create_blob(X_names[i]).feed(X, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(), op_type=op_type, engine=engine)
)
check_quantized_results_close(outputs)
def test_dnnlowp_gather(self, dim1, dim2, is_empty, in_quantized,
out_quantized, gc, dc):
if is_empty:
dim2 = 0
# FIXME : DNNLOWP Gather doesn't support quantized input and
# dequantized output
if in_quantized:
out_quantized = True
data = (np.random.rand(dim1) * 2 - 1).astype(np.float32)
index = np.floor(np.random.rand(dim2) * dim1).astype(np.int32)
Output = collections.namedtuple("Output", ["out", "op_type", "engine"])
outputs = []
op_engine_list = [
("Gather", ""),
("Gather", "DNNLOWP"),
("Int8Gather", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
if do_quantize:
quantize_data = core.CreateOperator("Quantize", ["data"],
["data_q"],
engine=engine,
device_option=gc)
net.Proto().op.extend([quantize_data])
gather = core.CreateOperator(
op_type,
["data_q" if do_quantize else "data", "index"],
["out_q" if do_dequantize else "out"],
dequantize_output=not do_dequantize,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([gather])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["out_q"],
["out"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
self.ws.create_blob("data").feed(data, device_option=gc)
self.ws.create_blob("index").feed(index, device_option=gc)
self.ws.run(net)
outputs.append(
Output(out=self.ws.blobs["out"].fetch(),
op_type=op_type,
engine=engine))
check_quantized_results_close(outputs, ref=data)
def test_dnnlowp_relu(self, size, is_empty, gc, dc):
if is_empty:
size = 0
min_ = -10.0
max_ = 10.0
scale = (max_ - min_) / 255
zero_point = int(np.round(-min_ / scale))
X = (np.random.rand(size) * (max_ - min_) + min_).astype(np.float32)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [("Relu", ""), ("Relu", "DNNLOWP"),
("Int8Relu", "DNNLOWP")]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
if engine == "DNNLOWP":
quantize = core.CreateOperator(
"Quantize",
["X"],
["X_q"],
engine=engine,
device_option=gc,
Y_scale=scale,
Y_zero_point=zero_point,
)
net.Proto().op.extend([quantize])
relu = core.CreateOperator(
op_type,
["X_q" if engine == "DNNLOWP" else "X"],
["Y_q" if engine == "DNNLOWP" else "Y"],
engine=engine,
device_option=gc,
)
net.Proto().op.extend([relu])
if engine == "DNNLOWP":
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(),
op_type=op_type,
engine=engine))
# Y = max(0, X) so the only error is quantization of inputs
check_quantized_results_close(outputs, ref=X)
def test_dnnlowp_elementwise_add_broadcast_axis(self, gc, dc):
for bdim, axis in [
((3, 4), 1), # broadcasting intermediate dimensions
((2, ), 0), # broadcasting the first dimension
((1, 4, 1), 1),
]:
# broadcasting with single elem dimensions at both ends
min_ = -100
max_ = min_ + 255
A = np.round(np.random.rand(2, 3, 4, 5) * (max_ - min_) + min_)
A = A.astype(np.float32)
B = np.round(np.random.rand(*bdim) * 255 / 2 - 64).astype(
np.float32)
A.flat[0] = min_
A.flat[1] = max_
B.flat[0] = -64
B.flat[1] = 127.0 / 2
Output = collections.namedtuple("Output",
["Y", "op_type", "engine"])
outputs = []
op_engine_list = [("Add", ""), ("Add", "DNNLOWP"),
("Int8Add", "DNNLOWP")]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
add = core.CreateOperator(
op_type,
["A", "B"],
["Y"],
engine=engine,
device_option=gc,
broadcast=1,
axis=axis,
dequantize_output=1,
)
net.Proto().op.extend([add])
self.ws.create_blob("A").feed(A, device_option=gc)
self.ws.create_blob("B").feed(B, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(),
op_type=op_type,
engine=engine))
check_quantized_results_close(outputs)
def test_dnnlowp_elementwise_add_broadcast(self, gc, dc):
# Set broadcast and no axis, i.e. broadcasting last dimensions.
min_ = -100
max_ = min_ + 255
A = np.round(np.random.rand(2, 3, 4, 5) * (max_ - min_) + min_)
A = A.astype(np.float32)
A[0, 0, 0, 0] = min_
A[0, 0, 0, 1] = max_
B = np.round(np.random.rand(4, 5) * 255 / 2 - 64).astype(np.float32)
B[0, 0] = -64
B[0, 1] = 127.0 / 2
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [("Add", ""), ("Add", "DNNLOWP"),
("Int8Add", "DNNLOWP")]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
add = core.CreateOperator(
op_type,
["A", "B"],
["Y"],
engine=engine,
device_option=gc,
broadcast=1,
dequantize_output=1,
)
net.Proto().op.extend([add])
self.ws.create_blob("A").feed(A, device_option=gc)
self.ws.create_blob("B").feed(B, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(),
op_type=op_type,
engine=engine))
check_quantized_results_close(outputs)
def test_dnnlowp_dequantize(self, size, is_empty, gc, dc):
if is_empty:
size = 0
min_ = -10.0
max_ = 20.0
X = (np.random.rand(size) * (max_ - min_) + min_).astype(np.float32)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_type_list = ["Dequantize", "Int8Dequantize"]
engine = "DNNLOWP"
outputs.append(Output(X, op_type="", engine=""))
for op_type in op_type_list:
net = core.Net("test_net")
quantize = core.CreateOperator("Quantize", ["X"], ["X_q"],
engine=engine,
device_option=gc)
net.Proto().op.extend([quantize])
dequantize = core.CreateOperator(op_type, ["X_q"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(),
op_type=op_type,
engine=engine))
check_quantized_results_close(outputs)
def test_dnnlowp_concat_int(
self, dim1, dim2, axis, in_quantized, out_quantized, gc, dc
):
# X has scale 1, so exactly represented after quantization
min_ = -100
max_ = min_ + 255
X = np.round(np.random.rand(dim1, dim2) * (max_ - min_) + min_)
X = X.astype(np.float32)
if dim1 >= 1 and dim2 >= 2:
X[0, 0] = min_
X[0, 1] = max_
elif dim2 == 1:
return
# Y has scale 1/2, so exactly represented after quantization
Y = np.round(np.random.rand(dim1, dim2) * 255 / 2 - 64)
Y = Y.astype(np.float32)
if dim1 >= 1 and dim2 >= 2:
Y[0, 0] = -64
Y[0, 1] = 127.0 / 2
Output = collections.namedtuple("Output", ["Z", "op_type", "engine"])
outputs = []
op_engine_list = [
("Concat", ""),
("Concat", "DNNLOWP"),
("Int8Concat", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
if do_quantize:
quantize_x = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine=engine, device_option=gc
)
quantize_y = core.CreateOperator(
"Quantize", ["Y"], ["Y_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize_x, quantize_y])
concat = core.CreateOperator(
op_type,
["X_q", "Y_q"] if do_quantize else ["X", "Y"],
["Z_q" if do_dequantize else "Z", "split"],
dequantize_output=not do_dequantize,
engine=engine,
device_option=gc,
axis=axis,
)
net.Proto().op.extend([concat])
if do_dequantize:
dequantize = core.CreateOperator(
"Dequantize", ["Z_q"], ["Z"], engine=engine, device_option=gc
)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.create_blob("Y").feed(Y, device_option=gc)
self.ws.create_blob("split")
self.ws.run(net)
outputs.append(
Output(Z=self.ws.blobs["Z"].fetch(), op_type=op_type, engine=engine)
)
check_quantized_results_close(outputs)
def test_dnnlowp_conv_acc16_outlier(
self,
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
weight_quantized,
prepack_weight,
nbits_in_non_outlier,
share_col_buffer,
preserve_activation_sparsity,
preserve_weight_sparsity,
gc,
dc,
):
assume(group == 1 or dilation == 1)
assume(size >= dilation * (kernel - 1) + 1)
input_channels = input_channels_per_group * group
output_channels = output_channels_per_group * group
X_min = 0 if preserve_activation_sparsity else -77
X_max = X_min + 255
X = np.random.rand(batch_size, size, size, input_channels) * 4 + X_min
X = np.round(X).astype(np.float32)
X[..., 0] = X_min
if batch_size != 0:
X[0, 0, 0, 1] = X_max
if preserve_weight_sparsity:
W_min = -128
W_max = 100
else:
W_min = -100
W_max = W_min + 255
W = (np.random.rand(output_channels, kernel, kernel,
input_channels_per_group) * 4 - 2 + W_min + 128)
W = np.round(W).astype(np.float32)
W[0, 0, 0, 0] = W_min
W[1, 0, 0, 0] = W_max
W[..., 1] = W_min + 128 # "zeros"
if order == "NCHW":
X = utils.NHWC2NCHW(X)
W = utils.NHWC2NCHW(W)
b = np.round(np.random.randn(output_channels)).astype(np.float32)
Output = collections.namedtuple("Output",
["Y", "op_type", "engine", "order"])
outputs = []
op_engine_list = [
("Conv", ""),
("Conv", "DNNLOWP_ACC16"),
("Int8Conv", "DNNLOWP_ACC16"),
]
for op_type, engine in op_engine_list:
init_net = core.Net("test_init_net")
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine
do_dequantize = "DNNLOWP" in engine
do_quantize_weight = "DNNLOWP" in engine and weight_quantized
do_prepack_weight = "DNNLOWP" in engine and prepack_weight
if do_quantize:
quantize = core.CreateOperator(
"Quantize",
["X"],
["X_q"],
preserve_activation_sparsity=preserve_activation_sparsity,
engine="DNNLOWP",
device_option=gc,
)
net.Proto().op.extend([quantize])
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
x_q_param = dnnlowp_utils.choose_quantization_params(
X_min, X_max, preserve_activation_sparsity)
if do_quantize_weight:
int8_given_tensor_fill, w_q_param = dnnlowp_utils.create_int8_given_tensor_fill(
W, "W_q", preserve_weight_sparsity)
init_net.Proto().op.extend([int8_given_tensor_fill])
# Bias
int8_bias_tensor_fill = dnnlowp_utils.create_int8_bias_tensor_fill(
b, "b_q", x_q_param, w_q_param)
init_net.Proto().op.extend([int8_bias_tensor_fill])
if do_prepack_weight:
inputs = ["W_q" if do_quantize_weight else "W"]
if do_dequantize:
inputs += ["b_q" if do_quantize_weight else "b"]
pack = core.CreateOperator(
"Int8ConvPackWeight",
inputs,
["W_packed"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
nbits_in_non_outlier=nbits_in_non_outlier,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
in_scale=x_q_param.scale,
)
init_net.Proto().op.extend([pack])
conv = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"W_packed" if do_prepack_weight else
("W_q" if do_quantize_weight else "W"),
"b_q" if do_quantize_weight else "b",
],
["Y_q" if do_dequantize else "Y"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
nbits_in_non_outlier=nbits_in_non_outlier,
shared_buffer=(1 if share_col_buffer else 0),
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
device_option=gc,
)
if do_dequantize or do_quantize_weight or do_prepack_weight:
# When quantized weight is provided, we can't rescale the
# output dynamically by looking at the range of output of each
# batch, so here we provide the range of output observed from
# fp32 reference implementation
dnnlowp_utils.add_quantization_param_args(
conv, outputs[0][0], preserve_activation_sparsity)
net.Proto().op.extend([conv])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine="DNNLOWP",
device_option=gc)
net.Proto().op.extend([dequantize])
run_conv_or_fc(self, init_net, net, X, W, b, op_type, engine,
order, gc, outputs)
check_quantized_results_close(outputs,
symmetric=preserve_activation_sparsity)
def test_dnnlowp_depthwise_3x3_conv(
self,
stride,
size,
group,
batch_size,
prepack_weight,
share_col_buffer,
preserve_activation_sparsity,
preserve_weight_sparsity,
quantize_groupwise,
relu,
gc,
dc,
):
pad = 1
kernel = 3
dilation = 1
input_channels_per_group = 1
output_channels_per_group = 1
order = "NHWC"
X, W, b = generate_conv_inputs(
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
groupwise_quantization=quantize_groupwise,
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
)
Output = collections.namedtuple("Output",
["Y", "op_type", "engine", "order"])
outputs = []
if relu:
op_engine_list = [
("Conv", ""),
("ConvRelu", "DNNLOWP"),
("Int8ConvRelu", "DNNLOWP"),
]
else:
op_engine_list = [
("Conv", ""),
("Conv", "DNNLOWP"),
("Int8Conv", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
init_net = core.Net("test_init_net")
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine
do_dequantize = "DNNLOWP" in engine
do_prepack_weight = engine == "DNNLOWP" and prepack_weight
if do_quantize:
quantize = core.CreateOperator(
"Quantize",
["X"],
["X_q"],
preserve_activation_sparsity=preserve_activation_sparsity,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([quantize])
if do_prepack_weight:
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
x_q_param = dnnlowp_utils.choose_quantization_params(
X_min, X_max, preserve_activation_sparsity)
inputs = ["W"]
if do_dequantize:
inputs += ["b"]
pack = core.CreateOperator(
"Int8ConvPackWeight",
inputs,
["W_packed"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
quantize_groupwise=quantize_groupwise,
in_scale=x_q_param.scale,
)
init_net.Proto().op.extend([pack])
conv = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X", "W", "b"],
["Y_q" if do_dequantize else "Y"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
shared_buffer=(1 if share_col_buffer else 0),
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
quantize_groupwise=quantize_groupwise,
device_option=gc,
)
if do_dequantize or do_prepack_weight:
dnnlowp_utils.add_quantization_param_args(
conv, outputs[0][0], preserve_activation_sparsity)
net.Proto().op.extend([conv])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
elif relu:
relu_op = core.CreateOperator("Relu", ["Y"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([relu_op])
run_conv_or_fc(self, init_net, net, X, W, b, op_type, engine,
order, gc, outputs)
check_quantized_results_close(outputs,
symmetric=preserve_activation_sparsity)
def test_rowwise_dnnlowp_fully_connected_int(
self,
input_channels,
output_channels,
batch_size,
in_quantized,
out_quantized,
prepack_weight,
gc,
dc,
):
# X has scale 1, so exactly represented after quantization
X_min = -77
X_max = X_min + 255
X = np.round(
np.random.rand(batch_size, input_channels) * (X_max - X_min) +
X_min)
X = X.astype(np.float32)
# input channels 0 and 1 are all X_min to avoid overflow from vpmaddubsw
# when multiplied with W_min and W_max
X[:, 0:2] = X_min
if batch_size != 0:
X[0, 2] = X_max
# Each row of W has scale 1 but with different offset, so row-wise
# quantization shouldn't have any input quantization error.
W = np.zeros((output_channels, input_channels))
W = W.astype(np.float32)
for i in range(output_channels):
W_min = -100 + i
W_max = W_min + 255
W[i, :] = np.round(
np.random.rand(input_channels) * (W_max - W_min) + W_min)
W[i, 0] = W_min
W[i, 1] = W_max
# Make sure we won't have overflows from vpmaddubsw instruction used in
# fbgemm
avoid_vpmaddubsw_overflow_fc(
batch_size,
input_channels,
1,
X,
X_min,
X_max,
W[i:i + 1, ],
W_min,
W_max,
)
if i % 2 == 0:
W[i, :] = (W[i, :] - W_min) * 2 + W_min
b = np.random.randn(output_channels).astype(np.float32)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("FC", ""),
("FC", "DNNLOWP_ROWWISE"),
("FC", "DNNLOWP_ROWWISE_16"),
("Int8FC", "DNNLOWP_ROWWISE"),
]
for op_type, engine in op_engine_list:
init_net = core.Net("test_init_net")
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
do_prepack_weight = engine == "DNNLOWP_ROWWISE" and prepack_weight
if do_quantize:
quantize = core.CreateOperator("Quantize", ["X"], ["X_q"],
engine=engine,
device_option=gc)
net.Proto().op.extend([quantize])
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
x_q_param = dnnlowp_utils.choose_quantization_params(X_min, X_max)
if do_prepack_weight:
inputs = ["W"]
if do_dequantize:
inputs += ["b"]
pack = core.CreateOperator(
"Int8FCPackWeight",
inputs,
["W_packed"],
in_scale=x_q_param.scale,
engine=engine,
)
init_net.Proto().op.extend([pack])
fc = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"W_packed" if do_prepack_weight else "W",
"b",
],
["Y_q" if do_dequantize else "Y"],
dequantize_output=not do_dequantize,
engine=engine,
device_option=gc,
)
if do_prepack_weight:
# When pre-packed quantized weight is provided, we can't rescale
# the output dynamically by looking at the range of output of
# each batch, so here we provide the range of output observed
# from fp32 reference implementation
dnnlowp_utils.add_quantization_param_args(fc, outputs[0][0])
net.Proto().op.extend([fc])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
run_conv_or_fc(self, init_net, net, X, W, b, op_type, engine, None,
gc, outputs)
check_quantized_results_close(outputs)
def test_groupwise_dnnlowp_conv_int(
self,
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
prepack_weight,
preserve_activation_sparsity,
preserve_weight_sparsity,
gc,
dc,
):
assume(group == 1 or dilation == 1)
assume((not prepack_weight) or order == "NHWC")
X, W, b = generate_conv_inputs(
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
groupwise_quantization=True,
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
)
Output = collections.namedtuple("Output",
["Y", "op_type", "engine", "order"])
outputs = []
op_engine_list = [
("Conv", ""),
("Conv", "DNNLOWP"),
("Conv", "DNNLOWP_16"),
("Int8Conv", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
init_net = core.Net("test_init_net")
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine
do_dequantize = "DNNLOWP" in engine
do_prepack_weight = engine == "DNNLOWP" and prepack_weight
if do_quantize:
quantize = core.CreateOperator(
"Quantize",
["X"],
["X_q"],
preserve_activation_sparsity=preserve_activation_sparsity,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([quantize])
if do_prepack_weight:
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
x_q_param = hardcode_scale_zp.choose_quantization_params(
X_min, X_max)
inputs = ["W"]
if do_dequantize:
inputs += ["b"]
pack = core.CreateOperator(
"Int8ConvPackWeight",
inputs,
["W_packed"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
quantize_groupwise=1,
in_scale=x_q_param.scale,
)
init_net.Proto().op.extend([pack])
conv = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"W_packed" if do_prepack_weight else "W",
"b",
],
["Y_q" if do_dequantize else "Y"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
quantize_groupwise=1,
device_option=gc,
)
if do_dequantize or do_prepack_weight:
# groupwise quantization only works with static quantization
# so we need to set quantization parameters
dnnlowp_utils.add_quantization_param_args(
conv, outputs[0][0], preserve_activation_sparsity)
net.Proto().op.extend([conv])
if do_dequantize:
dequantize = core.CreateOperator(
"Dequantize",
["Y_q"],
["Y"],
preserve_activation_sparsity=preserve_activation_sparsity,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([dequantize])
run_conv_or_fc(self, init_net, net, X, W, b, op_type, engine,
order, gc, outputs)
check_quantized_results_close(outputs,
symmetric=preserve_activation_sparsity)
def test_dnnlowp_spatial_bn_int(
self,
size,
input_channels,
output_channels,
batch_size,
order,
in_quantized,
out_quantized,
fuse_relu,
gc,
dc,
):
X_min = -77
X_max = X_min + 255
X = np.round(np.random.rand(batch_size, size, size, input_channels)).astype(
np.float32
)
if batch_size != 0:
X[0, 0, 0, 0] = X_min
X[0, 0, 0, 1] = X_max
epsilon = np.abs(np.random.rand())
scale = np.random.rand(input_channels).astype(np.float32)
bias = np.random.rand(input_channels).astype(np.float32)
mean = np.random.rand(input_channels).astype(np.float32)
var = np.random.rand(input_channels).astype(np.float32)
if order == "NCHW":
X = utils.NHWC2NCHW(X)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("SpatialBN", ""),
]
if fuse_relu:
op_engine_list += [
("Int8SpatialBNRelu", "DNNLOWP"),
]
else:
op_engine_list += [
("SpatialBN", "DNNLOWP"),
("Int8SpatialBN", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
if do_quantize:
quantize = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine=engine
)
net.Proto().op.extend([quantize])
bn = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X", "scale", "bias", "mean", "var"],
["Y_q" if do_dequantize else "Y"],
is_test=True,
epsilon=epsilon,
order=order,
engine=engine,
dequantize_output=not do_dequantize,
)
net.Proto().op.extend([bn])
if "DNNLOWP" in engine:
dnnlowp_utils.add_quantization_param_args(bn, outputs[0][0])
if fuse_relu and "DNNLOWP" not in engine:
net.Relu(["Y"], "Y")
if do_dequantize:
dequantize = core.CreateOperator(
"Dequantize", ["Y_q"], ["Y"], engine=engine
)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.create_blob("scale").feed(scale, device_option=gc)
self.ws.create_blob("bias").feed(bias, device_option=gc)
self.ws.create_blob("mean").feed(mean, device_option=gc)
self.ws.create_blob("var").feed(var, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(), op_type=op_type, engine=engine)
)
check_quantized_results_close(outputs)
def test_dnnlowp_conv_relu_int(
self,
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
share_col_buffer,
gc,
dc,
):
assume(group == 1 or dilation == 1)
X, W, b = generate_conv_inputs(
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
)
Output = collections.namedtuple("Output",
["Y", "op_type", "engine", "order"])
outputs = []
op_engine_list = [
("Conv", ""),
("ConvRelu", "DNNLOWP"),
("ConvRelu", "DNNLOWP_16"),
("Int8ConvRelu", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
if "DNNLOWP" in engine:
quantize = core.CreateOperator("Quantize", ["X"], ["X_q"],
engine=engine,
device_option=gc)
net.Proto().op.extend([quantize])
conv = core.CreateOperator(
op_type,
["X_q", "W", "b"],
["Y_q"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
engine=engine,
shared_buffer=(1 if share_col_buffer else 0),
group=group,
device_option=gc,
)
net.Proto().op.extend([conv])
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
else:
conv = core.CreateOperator(
op_type,
["X", "W", "b"],
["Y"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
shared_buffer=(1 if share_col_buffer else 0),
engine=engine,
group=group,
device_option=gc,
)
net.Proto().op.extend([conv])
relu = core.CreateOperator("Relu", ["Y"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([relu])
run_conv_or_fc(self, None, net, X, W, b, op_type, engine, order,
gc, outputs)
check_quantized_results_close(outputs)
def test_groupwise_dnnlowp_conv_acc16_int(
self,
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
share_col_buffer,
preserve_activation_sparsity,
preserve_weight_sparsity,
gc,
dc,
):
assume(group == 1 or dilation == 1)
assume(size >= dilation * (kernel - 1) + 1)
input_channels = input_channels_per_group * group
output_channels = output_channels_per_group * group
# X and W have scale 1, so exactly represented after quantization
# This was made sure by having at least one 0 and one 255 for unsigned
# 8-bit tensors, and at least one -128 and one 127 for signed 8-bit
# tensors.
# Since fbgemm_acc16 accumulates to 16-bit, To avoid overflow, we use
# small numbers except for those 0, 255, -128, and 127, for this test
# We also make sure 255, -128, or 127 are not multiplied together by
# putting them in different input channels and the corresponding input
# channel in other matrix is 0.
# For example, we put 255 in input channel 1 in X, so we make the
# corresponding input channel in W all zeros.
X_min = 0 if preserve_activation_sparsity else -77
X_max = X_min + 255
X = np.random.rand(batch_size, size, size, input_channels) * 4 + X_min
X = np.round(X).astype(np.float32)
X[..., 0] = X_min
if batch_size != 0:
X[0, 0, 0, 1] = X_max
if preserve_weight_sparsity:
W_min = -128
W_max = 100
else:
W_min = -100
W_max = W_min + 255
W = (np.random.rand(output_channels, kernel, kernel,
input_channels_per_group) * 4 - 2 + W_min + 128)
W = np.round(W).astype(np.float32)
W[..., 1] = W_min + 128 # "zeros"
for g in range(group):
W[g * output_channels_per_group, 0, 0, 0] = W_min
W[g * output_channels_per_group + 1, 0, 0, 0] = W_max
if not preserve_weight_sparsity:
W[g * output_channels_per_group:(g + 1) *
output_channels_per_group, ] += g
if order == "NCHW":
X = utils.NHWC2NCHW(X)
W = utils.NHWC2NCHW(W)
# No input quantization error in bias
b = np.round(np.random.randn(output_channels)).astype(np.float32)
Output = collections.namedtuple("Output",
["Y", "op_type", "engine", "order"])
outputs = []
op_engine_list = [
("Conv", ""),
("Conv", "DNNLOWP_ACC16"),
("Int8Conv", "DNNLOWP_ACC16"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine
do_dequantize = "DNNLOWP" in engine
if do_quantize:
quantize = core.CreateOperator(
"Quantize",
["X"],
["X_q"],
preserve_activation_sparsity=preserve_activation_sparsity,
engine="DNNLOWP",
device_option=gc,
)
net.Proto().op.extend([quantize])
conv = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X", "W", "b"],
["Y_q" if do_dequantize else "Y"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
shared_buffer=(1 if share_col_buffer else 0),
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
quantize_groupwise=1,
device_option=gc,
)
if do_dequantize:
# groupwise quantization only works with static quantization
# so we need to set quantization parameters
dnnlowp_utils.add_quantization_param_args(
conv, outputs[0][0], preserve_activation_sparsity)
net.Proto().op.extend([conv])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine="DNNLOWP",
device_option=gc)
net.Proto().op.extend([dequantize])
run_conv_or_fc(self, None, net, X, W, b, op_type, engine, order,
gc, outputs)
check_quantized_results_close(outputs,
symmetric=preserve_activation_sparsity)
def test_dnnlowp_elementwise_mul_int(
self, N, is_empty, in_quantized, out_quantized, in_place, gc, dc
):
if is_empty:
N = 0
# FIXME: DNNLOWP Mul doesn't support inplace operation and
# dequantize_output=1 at the same time
if in_place[0] or in_place[1]:
in_quantized = True
out_quantized = True
# All inputs have scale 1, so exactly represented after quantization
min_ = -100
max_ = min_ + 255
A = np.round(np.random.rand(N) * (max_ - min_) + min_)
A = A.astype(np.float32)
if N != 0:
A[0] = min_
A[1] = max_
B = np.round(np.random.rand(N) * 255 - 128).astype(np.float32)
if N != 0:
B[0] = -128
B[1] = 127
Output = collections.namedtuple("Output", ["Y", "engine"])
outputs = []
engine_list = ["", "DNNLOWP"]
for engine in engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
if do_quantize:
quantize_A = core.CreateOperator(
"Quantize", ["A"], ["A_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize_A])
quantize_B = core.CreateOperator(
"Quantize", ["B"], ["B_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize_B])
out = "Y"
if in_place[0]:
out = "A"
elif in_place[1]:
out = "B"
mul = core.CreateOperator(
"Mul",
["A_q", "B_q"] if do_quantize else ["A", "B"],
[(out + "_q") if do_dequantize else out],
dequantize_output=not do_dequantize,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([mul])
if do_dequantize:
dequantize = core.CreateOperator(
"Dequantize", [out + "_q"], [out], engine=engine, device_option=gc
)
net.Proto().op.extend([dequantize])
self.ws.create_blob("A").feed(A, device_option=gc)
self.ws.create_blob("B").feed(B, device_option=gc)
self.ws.run(net)
outputs.append(Output(Y=self.ws.blobs[out].fetch(), engine=engine))
check_quantized_results_close(outputs)
def test_dnnlowp_fully_connected_acc16_int(
self,
input_channels,
output_channels,
batch_size,
in_quantized,
out_quantized,
gc,
dc,
):
# X and W have scale 1, so exactly represented after quantization
# This was made sure by having at least one 0 and one 255 for unsigned
# 8-bit tensors, and at least one -128 and one 127 for signed 8-bit
# tensors.
# Since fbgemm_acc16 accumulates to 16-bit, To avoid overflow, we use
# small numbers except for those 0, 255, -128, and 127, for this test
# We also make sure 255, -128, or 127 are not multiplied together by
# putting them in different input channels and the corresponding input
# channel in other matrix is 0.
# For example, we put 255 in input channel 1 in X, so we make the
# corresponding input channel in W all zeros.
X_min = -77
X_max = X_min + 255
X = np.round(np.random.rand(batch_size, input_channels) * 4 + X_min)
X = X.astype(np.float32)
X[:, 0] = X_min
if batch_size != 0:
X[0, 1] = X_max
W_min = -100
W_max = W_min + 255
W = np.round(
np.random.rand(output_channels, input_channels) * 4 - 2 + W_min +
128)
W = W.astype(np.float32)
W[0, 0] = W_min
W[1, 0] = W_max
W[:, 1] = W_min + 128
# No input quantization error in bias
b = np.round(np.random.randn(output_channels)).astype(np.float32)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("FC", ""),
("FC", "DNNLOWP_ACC16"),
("Int8FC", "DNNLOWP_ACC16"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
if do_quantize:
quantize = core.CreateOperator("Quantize", ["X"], ["X_q"],
engine="DNNLOWP",
device_option=gc)
net.Proto().op.extend([quantize])
fc = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X", "W", "b"],
["Y_q" if do_dequantize else "Y"],
dequantize_output=(0 if do_dequantize else 1),
engine=engine,
device_option=gc,
)
net.Proto().op.extend([fc])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine="DNNLOWP",
device_option=gc)
net.Proto().op.extend([dequantize])
run_conv_or_fc(self, None, net, X, W, b, op_type, engine, None, gc,
outputs)
check_quantized_results_close(outputs)
def test_dnnlowp_conv_int(
self,
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
weight_quantized,
prepack_weight,
share_col_buffer,
preserve_activation_sparsity,
preserve_weight_sparsity,
gc,
dc,
):
assume(group == 1 or dilation == 1)
assume((not prepack_weight) or order == "NHWC")
X, W, b = generate_conv_inputs(
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
)
Output = collections.namedtuple("Output",
["Y", "op_type", "engine", "order"])
outputs = []
op_engine_list = [
("Conv", ""),
("Conv", "DNNLOWP"),
("Conv", "DNNLOWP_16"),
("Int8Conv", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
init_net = core.Net("test_init_net")
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine
do_dequantize = "DNNLOWP" in engine
# If output scale/zp aren't set, it gets computed from ref fp32 op
# in DNNLOWP, which isn't possible when we quantize input weights.
# Make sure atleast one output is collected to compute output
# scale/zp.
do_quantize_weight = (engine == "DNNLOWP" and weight_quantized
and len(outputs) > 0)
do_prepack_weight = engine == "DNNLOWP" and prepack_weight
if do_quantize:
quantize = core.CreateOperator(
"Quantize",
["X"],
["X_q"],
preserve_activation_sparsity=preserve_activation_sparsity,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([quantize])
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
x_q_param = dnnlowp_utils.choose_quantization_params(
X_min, X_max, preserve_activation_sparsity)
if do_quantize_weight:
int8_given_tensor_fill, w_q_param = dnnlowp_utils.create_int8_given_tensor_fill(
W, "W_q", preserve_weight_sparsity)
init_net.Proto().op.extend([int8_given_tensor_fill])
# Bias
int8_bias_tensor_fill = dnnlowp_utils.create_int8_bias_tensor_fill(
b, "b_q", x_q_param, w_q_param)
init_net.Proto().op.extend([int8_bias_tensor_fill])
if do_prepack_weight:
inputs = ["W_q" if do_quantize_weight else "W"]
if do_dequantize:
inputs += ["b_q" if do_quantize_weight else "b"]
pack = core.CreateOperator(
"Int8ConvPackWeight",
inputs,
["W_packed"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
in_scale=x_q_param.scale,
)
init_net.Proto().op.extend([pack])
conv = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"W_packed" if do_prepack_weight else
("W_q" if do_quantize_weight else "W"),
"b_q" if do_quantize_weight else "b",
],
["Y_q" if do_dequantize else "Y"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
shared_buffer=(1 if share_col_buffer else 0),
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
device_option=gc,
)
if do_quantize_weight or do_prepack_weight:
# When quantized weight is provided, we can't rescale the
# output dynamically by looking at the range of output of each
# batch, so here we provide the range of output observed from
# fp32 reference implementation
dnnlowp_utils.add_quantization_param_args(
conv, outputs[0][0], preserve_activation_sparsity)
net.Proto().op.extend([conv])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
run_conv_or_fc(self, init_net, net, X, W, b, op_type, engine,
order, gc, outputs)
check_quantized_results_close(outputs,
symmetric=preserve_activation_sparsity)
def test_dnnlowp_batch_matmul_int(self, m, n, k, batch_size, gc, dc):
# A and B have scale 1, so exactly represented after quantization
A_min = -77
A_max = A_min + 255
A = np.round(np.random.rand(batch_size, m, k) * 255 + A_min)
A = A.astype(np.float32)
# input channels 0 and 1 are all A_min to avoid overflow from vpmaddubsw
# when multiplied with B_min and B_max
if batch_size > 0 and m > 0:
A[0, :, 0] = A_min
A[0, 0, 1] = A_max
B_min = -100
B_max = B_min + 255
B = np.round(np.random.rand(batch_size, n, k) * 255 + B_min)
B = B.astype(np.float32)
if batch_size > 0:
B[0, 0, 0] = B_min
B[0, 1, 0] = B_max
for i in range(batch_size):
avoid_vpmaddubsw_overflow_fc(m, k, n, A[i, ], A_min, A_max, B[i, ],
B_min, B_max)
for trans_a, trans_b in product([0, 1], [0, 1]):
Output = collections.namedtuple("Output",
["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("BatchMatMul", ""),
("BatchMatMul", "DNNLOWP"),
("BatchMatMul", "DNNLOWP_16"),
("Int8BatchMatMul", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
if "DNNLOWP" in engine:
quantize_A = core.CreateOperator("Quantize", ["A"],
["A_q"],
engine=engine,
device_option=gc)
net.Proto().op.extend([quantize_A])
quantize_B = core.CreateOperator("Quantize", ["B"],
["B_q"],
engine=engine,
device_option=gc)
net.Proto().op.extend([quantize_B])
batch_matmul = core.CreateOperator(
op_type,
[
"A_q" if "DNNLOWP" in engine else "A",
"B_q" if "DNNLOWP" in engine else "B",
],
["Y_q" if "DNNLOWP" in engine else "Y"],
trans_a=trans_a,
trans_b=trans_b,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([batch_matmul])
if "DNNLOWP" in engine:
dequantize = core.CreateOperator("Dequantize", ["Y_q"],
["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
self.ws.create_blob("A").feed(
np.transpose(A, (0, 2, 1)) if trans_a else A,
device_option=gc)
self.ws.create_blob("B").feed(
B if trans_b else np.transpose(B, (0, 2, 1)),
device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(),
op_type=op_type,
engine=engine))
check_quantized_results_close(outputs)
def test_dnnlowp_batch_matmul_int_constant_B(self, m, n, k, C_1, C_2,
A_quantized, B_quantized,
out_quantized, gc, dc):
batch_dims = tuple(np.random.randint(3, size=max(C_1, C_2)))
batch_dims_A = batch_dims[-C_1:]
batch_dims_B = batch_dims[-C_2:]
A = np.zeros(batch_dims_A + (m, k)).astype(np.float32)
B = np.zeros(batch_dims_B + (n, k)).astype(np.float32)
if np.prod(batch_dims) > 0:
for index in np.ndindex(batch_dims_A):
# When both input and output are float, each input of the batch has
# scale 1 but with different offset, so input-wise quantization
# shouldn't have any input quantization error
# A_min = -77 if (A_quantized or out_quantized) else -77 + i
A_min = -77
A_max = A_min + 255
A[index] = np.round(np.random.rand(m, k) * 255 + A_min)
# input channels 0 and 1 are all A_min to avoid overflow from vpmaddubsw
# when multiplied with B_min and B_max
A[index][:, 0] = A_min
if m != 0:
A[index][0, 1] = A_max
i = 0
for index in np.ndindex(batch_dims_B):
# When weight is quantized in a lazy manner, each input of the batch has
# scale 1 but with different offset, so input-wise quantization
# shouldn't have any input quantization error when weight is quantized
# in a lazy manner.
B_min = -100 if B_quantized else -100 + i
# B_min = -100
B_max = B_min + 255
B[index] = np.round(np.random.rand(n, k) * 255 + B_min)
B[index][0, 0] = B_min
B[index][1, 0] = B_max
if C_1 > C_2:
# A has more dims
for outer_index in np.ndindex(batch_dims_A[:C_1 - C_2]):
avoid_vpmaddubsw_overflow_fc(
m,
k,
n,
A[outer_index] if C_2 == 0 else A[outer_index +
index],
A_min,
A_max,
B[index],
B_min,
B_max,
)
else:
avoid_vpmaddubsw_overflow_fc(m, k, n, A[index[-C_1:]],
A_min, A_max, B[index], B_min,
B_max)
i += 1
for trans_a, trans_b in product([0, 1], [0, 1]):
Output = collections.namedtuple("Output",
["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("BatchMatMul", ""),
("BatchMatMul", "DNNLOWP"),
("Int8BatchMatMul", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize_A = "DNNLOWP" in engine and A_quantized
do_quantize_B = "DNNLOWP" in engine and B_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
if do_quantize_A:
quantize_A = core.CreateOperator("Quantize", ["A"],
["A_q"],
engine=engine,
device_option=gc)
net.Proto().op.extend([quantize_A])
if do_quantize_B:
int8_given_tensor_fill, B_q_param = dnnlowp_utils.create_int8_given_tensor_fill(
B if trans_b else B.swapaxes(-1, -2), "B_q")
net.Proto().op.extend([int8_given_tensor_fill])
batch_matmul = core.CreateOperator(
op_type,
[
"A_q" if do_quantize_A else "A",
"B_q" if do_quantize_B else "B"
],
["Y_q" if do_dequantize else "Y"],
trans_a=trans_a,
trans_b=trans_b,
broadcast=True,
constant_B=True,
dequantize_output=not do_dequantize,
engine=engine,
device_option=gc,
)
if do_quantize_B:
# When quantized weight is provided, we can't rescale the
# output dynamically by looking at the range of output of each
# batch, so here we provide the range of output observed from
# fp32 reference implementation
dnnlowp_utils.add_quantization_param_args(
batch_matmul, outputs[0][0])
net.Proto().op.extend([batch_matmul])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"],
["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
self.ws.create_blob("A").feed(
A.swapaxes(-1, -2) if trans_a else A, device_option=gc)
self.ws.create_blob("B").feed(
B if trans_b else B.swapaxes(-1, -2), device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(),
op_type=op_type,
engine=engine))
if np.prod(batch_dims) > 0:
check_quantized_results_close(outputs)
def test_dnnlowp_conv_acc16_int(
self,
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
weight_quantized,
share_col_buffer,
preserve_activation_sparsity,
preserve_weight_sparsity,
gc,
dc,
):
assume(group == 1 or dilation == 1)
assume(size >= dilation * (kernel - 1) + 1)
input_channels = input_channels_per_group * group
output_channels = output_channels_per_group * group
# X and W have scale 1, so exactly represented after quantization
# This was made sure by having at least one 0 and one 255 for unsigned
# 8-bit tensors, and at least one -128 and one 127 for signed 8-bit
# tensors.
# Since fbgemm_acc16 accumulates to 16-bit, To avoid overflow, we use
# small numbers except for those 0, 255, -128, and 127, for this test
# We also make sure 255, -128, or 127 are not multiplied together by
# putting them in different input channels and the corresponding input
# channel in other matrix is 0.
# For example, we put 255 in input channel 1 in X, so we make the
# corresponding input channel in W all zeros.
X_min = 0 if preserve_activation_sparsity else -77
X_max = X_min + 255
X = np.random.rand(batch_size, size, size, input_channels) * 4 + X_min
X = np.round(X).astype(np.float32)
X[..., 0] = X_min
if batch_size != 0:
X[0, 0, 0, 1] = X_max
if preserve_weight_sparsity:
W_min = -128
W_max = 100
else:
W_min = -100
W_max = W_min + 255
W = (np.random.rand(output_channels, kernel, kernel,
input_channels_per_group) * 4 - 2 + W_min + 128)
W = np.round(W).astype(np.float32)
W[0, 0, 0, 0] = W_min
W[1, 0, 0, 0] = W_max
W[..., 1] = W_min + 128 # "zeros"
if order == "NCHW":
X = utils.NHWC2NCHW(X)
W = utils.NHWC2NCHW(W)
# No input quantization error in bias
b = np.round(np.random.randn(output_channels)).astype(np.float32)
Output = collections.namedtuple("Output",
["Y", "op_type", "engine", "order"])
outputs = []
op_engine_list = [
("Conv", ""),
("Conv", "DNNLOWP_ACC16"),
("Int8Conv", "DNNLOWP_ACC16"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine
do_dequantize = "DNNLOWP" in engine
do_quantize_weight = ("DNNLOWP" in engine and weight_quantized
and len(outputs) > 0)
if do_quantize:
quantize = core.CreateOperator(
"Quantize",
["X"],
["X_q"],
preserve_activation_sparsity=preserve_activation_sparsity,
engine="DNNLOWP",
device_option=gc,
)
net.Proto().op.extend([quantize])
if do_quantize_weight:
int8_given_tensor_fill, w_q_param = dnnlowp_utils.create_int8_given_tensor_fill(
W, "W_q", preserve_weight_sparsity)
net.Proto().op.extend([int8_given_tensor_fill])
# Bias
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
x_q_param = dnnlowp_utils.choose_quantization_params(
X_min, X_max, preserve_activation_sparsity)
int8_bias_tensor_fill = dnnlowp_utils.create_int8_bias_tensor_fill(
b, "b_q", x_q_param, w_q_param)
net.Proto().op.extend([int8_bias_tensor_fill])
conv = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"W_q" if do_quantize_weight else "W",
"b_q" if do_quantize_weight else "b",
],
["Y_q" if do_dequantize else "Y"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
shared_buffer=(1 if share_col_buffer else 0),
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
group=group,
device_option=gc,
)
if do_dequantize or do_quantize_weight:
# When quantized weight is provided, we can't rescale the
# output dynamically by looking at the range of output of each
# batch, so here we provide the range of output observed from
# fp32 reference implementation
dnnlowp_utils.add_quantization_param_args(
conv, outputs[0][0], preserve_activation_sparsity)
net.Proto().op.extend([conv])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine="DNNLOWP",
device_option=gc)
net.Proto().op.extend([dequantize])
run_conv_or_fc(self, None, net, X, W, b, op_type, engine, order,
gc, outputs)
check_quantized_results_close(outputs,
symmetric=preserve_activation_sparsity)
def test_groupwise_dnnlowp_conv_relu_int(
self,
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
gc,
dc,
):
assume(group == 1 or dilation == 1)
X, W, b = generate_conv_inputs(
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
True, # group-wise
)
Output = collections.namedtuple("Output",
["Y", "op_type", "engine", "order"])
outputs = []
op_engine_list = [
("Conv", ""),
("ConvRelu", "DNNLOWP"),
("ConvRelu", "DNNLOWP_16"),
("Int8ConvRelu", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
if "DNNLOWP" in engine:
quantize = core.CreateOperator("Quantize", ["X"], ["X_q"],
engine=engine,
device_option=gc)
net.Proto().op.extend([quantize])
conv = core.CreateOperator(
op_type,
["X_q", "W", "b"],
["Y_q"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
engine=engine,
group=group,
quantize_groupwise=1,
device_option=gc,
)
# groupwise quantization only works with static quantization
# so we need to set quantization parameters
dnnlowp_utils.add_quantization_param_args(conv, outputs[0][0])
net.Proto().op.extend([conv])
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
else:
conv = core.CreateOperator(
op_type,
["X", "W", "b"],
["Y"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
engine=engine,
group=group,
device_option=gc,
)
net.Proto().op.extend([conv])
relu = core.CreateOperator("Relu", ["Y"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([relu])
run_conv_or_fc(self, None, net, X, W, b, op_type, engine, order,
gc, outputs)
check_quantized_results_close(outputs)
def test_dnnlowp_average_pool(
self,
ndim,
stride,
pad,
kernel,
size,
input_channels,
batch_size,
order,
in_quantized,
gc,
dc,
):
kernel = 2 # Only kernel size 2 is supported
assume(kernel <= size)
assume(pad < kernel)
C = input_channels
N = batch_size
strides = (stride,) * ndim
pads = (pad,) * (ndim * 2)
kernels = (kernel,) * ndim
sizes = (size,) * ndim
# X has scale 1, so no input quantization error
min_ = -100
max_ = min_ + 255
if order == "NCHW":
X = np.round(np.random.rand(*((N, C) + sizes)) * (max_ - min_) + min_)
X = X.astype(np.float32)
X[(0,) * (ndim + 2)] = min_
X[(0,) * (ndim + 1) + (1,)] = max_
elif order == "NHWC":
X = np.round(np.random.rand(*((N,) + sizes + (C,))) * (max_ - min_) + min_)
X = X.astype(np.float32)
X[(0,) * (ndim + 2)] = min_
X[(0, 1) + (0,) * ndim] = max_
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("AveragePool", ""),
("AveragePool", "DNNLOWP"),
("Int8AveragePool", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
if do_quantize:
quantize = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize])
max_pool = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X"],
["Y_q" if engine == "DNNLOWP" else "Y"],
strides=strides,
kernels=kernels,
pads=pads,
order=order,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([max_pool])
if engine == "DNNLOWP":
dequantize = core.CreateOperator(
"Dequantize", ["Y_q"], ["Y"], engine=engine, device_option=gc
)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(), op_type=op_type, engine=engine)
)
check_quantized_results_close(outputs)
def test_groupwise_dnnlowp_conv_acc16_outlier(
self,
stride,
pad,
kernel,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
prepack_weight,
nbits_in_non_outlier,
share_col_buffer,
gc,
dc,
):
assume(group == 1 or dilation == 1)
assume(size >= dilation * (kernel - 1) + 1)
input_channels = input_channels_per_group * group
output_channels = output_channels_per_group * group
X_min = -77
X_max = X_min + 255
X = np.random.rand(batch_size, size, size, input_channels) * 4 + X_min
X = np.round(X).astype(np.float32)
X[..., 0] = X_min
if batch_size != 0:
X[0, 0, 0, 1] = X_max
W_min = -100
W_max = W_min + 255
W = (np.random.rand(output_channels, kernel, kernel,
input_channels_per_group) * 4 - 2 + W_min + 128)
W = np.round(W).astype(np.float32)
W[..., 1] = W_min + 128 # "zeros"
for g in range(group):
W[g * output_channels_per_group, 0, 0, 0] = W_min
W[g * output_channels_per_group + 1, 0, 0, 0] = W_max
W[g * output_channels_per_group:(g + 1) *
output_channels_per_group, ] += g
if order == "NCHW":
X = utils.NHWC2NCHW(X)
W = utils.NHWC2NCHW(W)
b = np.round(np.random.randn(output_channels)).astype(np.float32)
Output = collections.namedtuple("Output",
["Y", "op_type", "engine", "order"])
outputs = []
op_engine_list = [
("Conv", ""),
("Conv", "DNNLOWP_ACC16"),
("Int8Conv", "DNNLOWP_ACC16"),
]
for op_type, engine in op_engine_list:
init_net = core.Net("test_init_net")
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine
do_dequantize = "DNNLOWP" in engine
do_prepack_weight = "DNNLOWP" in engine and prepack_weight
if do_quantize:
quantize = core.CreateOperator("Quantize", ["X"], ["X_q"],
engine="DNNLOWP",
device_option=gc)
net.Proto().op.extend([quantize])
if do_prepack_weight:
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
x_q_param = dnnlowp_utils.choose_quantization_params(
X_min, X_max)
inputs = ["W"]
if do_dequantize:
inputs += ["b"]
pack = core.CreateOperator(
"Int8ConvPackWeight",
inputs,
["W_packed"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
nbits_in_non_outlier=nbits_in_non_outlier,
engine=engine,
group=group,
quantize_groupwise=1,
in_scale=x_q_param.scale,
)
init_net.Proto().op.extend([pack])
conv = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"W_packed" if do_prepack_weight else "W",
"b",
],
["Y_q" if do_dequantize else "Y"],
stride=stride,
kernel=kernel,
dilation=dilation,
pad=pad,
order=order,
nbits_in_non_outlier=nbits_in_non_outlier,
shared_buffer=(1 if share_col_buffer else 0),
engine=engine,
group=group,
quantize_groupwise=1,
device_option=gc,
)
if do_dequantize or do_prepack_weight:
# groupwise quantization only works with static quantization
# so we need to set quantization parameters
dnnlowp_utils.add_quantization_param_args(conv, outputs[0][0])
net.Proto().op.extend([conv])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine="DNNLOWP",
device_option=gc)
net.Proto().op.extend([dequantize])
run_conv_or_fc(self, init_net, net, X, W, b, op_type, engine,
order, gc, outputs)
check_quantized_results_close(outputs)
def test_dnnlowp_fully_connected_int(
self,
input_channels,
output_channels,
batch_size,
in_quantized,
out_quantized,
weight_quantized,
prepack_weight,
preserve_activation_sparsity,
preserve_weight_sparsity,
fuse_relu,
output_packed_bias,
use_input_qparam,
gc,
dc,
):
# X and W have scale 1, so exactly represented after quantization
X_min = 0 if preserve_activation_sparsity else -77
X_max = X_min + 255
X = np.round(
np.random.rand(batch_size, input_channels) * (X_max - X_min) +
X_min)
X = X.astype(np.float32)
# input channels 0 and 1 are all X_min to avoid overflow from vpmaddubsw
# when multiplied with W_min and W_max
X[:, 0] = X_min
if batch_size != 0:
X[0, 1] = X_max
if preserve_weight_sparsity:
W_min = -128
W_max = 100
else:
W_min = -100
W_max = W_min + 255
W = np.round(
np.random.rand(output_channels, input_channels) * (W_max - W_min) +
W_min)
W = W.astype(np.float32)
W[0, 0] = W_min
W[1, 0] = W_max
# Make sure we won't have overflows from vpmaddubsw instruction used in
# fbgemm
avoid_vpmaddubsw_overflow_fc(
batch_size,
input_channels,
output_channels,
X,
X_min,
X_max,
W,
W_min,
W_max,
)
b = np.random.randn(output_channels).astype(np.float32)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [("FC", "")]
if fuse_relu:
op_engine_list += [("Int8FCRelu", "DNNLOWP")]
else:
op_engine_list += [
("FC", "DNNLOWP"),
("FC", "DNNLOWP_16"),
("Int8FC", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
init_net = core.Net("test_init_net")
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
do_quantize_weight = (engine == "DNNLOWP" and weight_quantized
and len(outputs) > 0)
do_prepack_weight = engine == "DNNLOWP" and prepack_weight
if do_quantize:
quantize = core.CreateOperator(
"Quantize",
["X"],
["X_q"],
preserve_activation_sparsity=preserve_activation_sparsity,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([quantize])
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
x_q_param = dnnlowp_utils.choose_quantization_params(
X_min, X_max, preserve_activation_sparsity)
w_q_param = None
if do_quantize_weight:
(
int8_given_tensor_fill,
w_q_param,
) = dnnlowp_utils.create_int8_given_tensor_fill(
W, "W_q", preserve_weight_sparsity)
init_net.Proto().op.extend([int8_given_tensor_fill])
# Bias
int8_bias_tensor_fill = dnnlowp_utils.create_int8_bias_tensor_fill(
b, "b_q", x_q_param, w_q_param)
init_net.Proto().op.extend([int8_bias_tensor_fill])
if do_prepack_weight:
inputs = ["W_q" if do_quantize_weight else "W"]
if do_dequantize:
inputs += ["b_q" if do_quantize_weight else "b"]
pack = core.CreateOperator(
"Int8FCPackWeight",
inputs,
["W_packed", "B_q32"]
if do_dequantize and output_packed_bias else ["W_packed"],
preserve_weight_sparsity=preserve_weight_sparsity,
in_scale=x_q_param.scale,
engine=engine,
)
init_net.Proto().op.extend([pack])
if use_input_qparam and do_dequantize and op_type != "FC":
fc = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"W_packed" if do_prepack_weight else
("W_q" if do_quantize_weight else "W"),
"b_q" if do_quantize_weight else "b",
"quant_param",
],
["Y_q" if do_dequantize else "Y"],
dequantize_output=not do_dequantize,
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
device_option=gc,
)
else:
fc = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"W_packed" if do_prepack_weight else
("W_q" if do_quantize_weight else "W"),
"b_q" if do_quantize_weight else "b",
],
["Y_q" if do_dequantize else "Y"],
dequantize_output=not do_dequantize,
preserve_activation_sparsity=preserve_activation_sparsity,
preserve_weight_sparsity=preserve_weight_sparsity,
engine=engine,
device_option=gc,
)
if do_quantize_weight or do_prepack_weight:
# When quantized weight is provided, we can't rescale the
# output dynamically by looking at the range of output of each
# batch, so here we provide the range of output observed from
# fp32 reference implementation
dnnlowp_utils.add_quantization_param_args(
fc, outputs[0][0], preserve_activation_sparsity)
net.Proto().op.extend([fc])
if fuse_relu and "DNNLOWP" not in engine:
net.Relu(["Y"], "Y")
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
if use_input_qparam and do_dequantize and op_type != "FC":
ref_output = outputs[0][0]
ref_output_min = 0 if ref_output.size == 0 else ref_output.min(
)
ref_output_max = 0 if ref_output.size == 0 else ref_output.max(
)
q_param = dnnlowp_utils.choose_quantization_params(
ref_output_min, ref_output_max,
preserve_activation_sparsity)
run_conv_or_fc(
self,
init_net,
net,
X,
W,
b,
op_type,
engine,
None,
gc,
outputs,
q_param.scale,
q_param.zero_point,
)
else:
run_conv_or_fc(self, init_net, net, X, W, b, op_type, engine,
None, gc, outputs)
if output_packed_bias and do_prepack_weight and do_dequantize:
bias_int32 = self.ws.blobs["B_q32"].fetch()
if do_quantize_weight:
np.testing.assert_equal(
bias_int32[0],
np.round(b / (x_q_param.scale * w_q_param.scale)))
np.testing.assert_equal(bias_int32[0].dtype, np.int32)
shapes, types = workspace.InferShapesAndTypes(
[init_net, net],
blob_dimensions={
"X": [batch_size, input_channels],
"W": [output_channels, input_channels],
"b": [output_channels],
"quant_param": [1],
},
blob_types={
"X": core.DataType.FLOAT,
"W": core.DataType.FLOAT,
"b": core.DataType.FLOAT,
"quant_param": core.DataType.FLOAT,
},
)
assert ("Y" in shapes
and "Y" in types), "Failed to infer the shape or type of Y"
self.assertEqual(shapes["Y"], [batch_size, output_channels])
self.assertEqual(types["Y"], core.DataType.FLOAT)
check_quantized_results_close(outputs,
symmetric=preserve_activation_sparsity)
def test_dnnlowp_max_pool(
self,
stride,
pad,
kernel,
size,
input_channels,
batch_size,
order,
in_quantized,
gc,
dc,
):
assume(kernel <= size)
assume(pad < kernel)
C = input_channels
N = batch_size
H = W = size
min_ = -10
max_ = 20
if order == "NCHW":
X = np.round(np.random.rand(N, C, H, W) * (max_ - min_) + min_)
elif order == "NHWC":
X = np.round(np.random.rand(N, H, W, C) * (max_ - min_) + min_)
X = X.astype(np.float32)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("MaxPool", ""),
("MaxPool", "DNNLOWP"),
("Int8MaxPool", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
if do_quantize:
quantize = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize])
max_pool = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X"],
["Y_q" if engine == "DNNLOWP" else "Y"],
stride=stride,
kernel=kernel,
pad=pad,
order=order,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([max_pool])
if engine == "DNNLOWP":
dequantize = core.CreateOperator(
"Dequantize", ["Y_q"], ["Y"], engine=engine, device_option=gc
)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(), op_type=op_type, engine=engine)
)
# Y_i = max(X_j) so the only error is in quantization of inputs
check_quantized_results_close(outputs, ref=X)
def test_dnnlowp_group_norm(
self,
N,
G,
K,
H,
W,
order,
in_quantized,
out_quantized,
weight_quantized,
gc,
dc,
):
C = G * K
X = np.random.rand(N, C, H, W).astype(np.float32) * 5.0 - 1.0
if order == "NHWC":
X = utils.NCHW2NHWC(X)
gamma = np.random.rand(C).astype(np.float32) * 2.0 - 1.0
beta = np.random.randn(C).astype(np.float32) - 0.5
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("GroupNorm", ""),
("GroupNorm", "DNNLOWP"),
("Int8GroupNorm", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
do_quantize_weight = (engine == "DNNLOWP" and weight_quantized
and len(outputs) > 0)
if do_quantize:
quantize = core.CreateOperator("Quantize", ["X"], ["X_q"],
engine=engine,
device_option=gc)
net.Proto().op.extend([quantize])
if do_quantize_weight:
int8_given_tensor_fill, gamma_q_param = dnnlowp_utils.create_int8_given_tensor_fill(
gamma, "gamma_q")
net.Proto().op.extend([int8_given_tensor_fill])
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
X_q_param = dnnlowp_utils.choose_quantization_params(
X_min, X_max)
int8_bias_tensor_fill = dnnlowp_utils.create_int8_bias_tensor_fill(
beta, "beta_q", X_q_param, gamma_q_param)
net.Proto().op.extend([int8_bias_tensor_fill])
group_norm = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"gamma_q" if do_quantize_weight else "gamma",
"beta_q" if do_quantize_weight else "beta",
],
["Y_q" if do_dequantize else "Y"],
dequantize_output=0 if do_dequantize else 1,
group=G,
order=order,
is_test=True,
engine=engine,
device_option=gc,
)
if do_quantize_weight:
# When quantized weight is provided, we can't rescale the
# output dynamically by looking at the range of output of each
# batch, so here we provide the range of output observed from
# fp32 reference implementation
dnnlowp_utils.add_quantization_param_args(
group_norm, outputs[0][0])
net.Proto().op.extend([group_norm])
if do_dequantize:
dequantize = core.CreateOperator("Dequantize", ["Y_q"], ["Y"],
engine=engine,
device_option=gc)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.create_blob("gamma").feed(gamma, device_option=gc)
self.ws.create_blob("beta").feed(beta, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(),
op_type=op_type,
engine=engine))
check_quantized_results_close(outputs, atol_scale=2.0)
def test_dnnlowp_elementwise_linear_int(
self, N, D, empty_batch, in_quantized, out_quantized, gc, dc
):
if empty_batch:
N = 0
# All inputs have scale 1, so exactly represented after quantization
min_ = -100
max_ = min_ + 255
X = np.round(np.random.rand(N, D) * (max_ - min_) + min_)
X = X.astype(np.float32)
if N != 0:
X[0, 0] = min_
X[0, 1] = max_
a = np.round(np.random.rand(D) * 255 - 128).astype(np.float32)
a[0] = -128
a[1] = 127
b = np.round(np.random.rand(D) * 255 - 128).astype(np.float32)
b[0] = -128
b[1] = 127
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("ElementwiseLinear", ""),
("ElementwiseLinear", "DNNLOWP"),
("Int8ElementwiseLinear", "DNNLOWP"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
if do_quantize:
quantize = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize])
eltwise_linear = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X", "a", "b"],
["Y_q" if do_dequantize else "Y"],
dequantize_output=not do_dequantize,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([eltwise_linear])
if do_dequantize:
dequantize = core.CreateOperator(
"Dequantize", ["Y_q"], ["Y"], engine=engine, device_option=gc
)
net.Proto().op.extend([dequantize])
self.ws.create_blob("X").feed(X, device_option=gc)
self.ws.create_blob("a").feed(a, device_option=gc)
self.ws.create_blob("b").feed(b, device_option=gc)
self.ws.run(net)
outputs.append(
Output(Y=self.ws.blobs["Y"].fetch(), op_type=op_type, engine=engine)
)
check_quantized_results_close(outputs)
