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_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_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_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_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)
