def _test_dnnlowp_nd_int(
self,
stride,
pad,
kernels,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
gc,
dc,
):
assume(group == 1 or dilation == 1)
ndim = len(kernels)
X, W, b = generate_convnd_inputs(
(stride, ) * ndim,
(pad, ) * ndim,
kernels,
(dilation, ) * ndim,
(size, ) * ndim,
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", ""), ("Conv", "DNNLOWP_16"),
("Int8Conv", "DNNLOWP")]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
fall_back_to_NCHW = "DNNLOWP" not in engine and order == "NHWC"
if fall_back_to_NCHW:
X_nchw = nhwc2nchw(X)
W_nchw = nhwc2nchw(W)
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 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, w_q_param = dnnlowp_utils.create_int8_given_tensor_fill(
W, "W_q")
net.Proto().op.extend([int8_given_tensor_fill])
# Bias
x_q_param = hardcode_scale_zp.choose_quantization_params(
X.min(), X.max())
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"],
strides=[stride] * ndim,
kernels=kernels,
dilations=[dilation] * ndim,
pads=[pad] * (ndim * 2),
order="NCHW" if fall_back_to_NCHW else order,
dequantize_output=not do_dequantize,
engine=engine,
group=group,
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(conv, outputs[0][0])
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])
self.ws.create_blob("X").feed(X_nchw if fall_back_to_NCHW else X,
device_option=gc)
self.ws.create_blob("W").feed(W_nchw if fall_back_to_NCHW else W,
device_option=gc)
self.ws.create_blob("b").feed(b, device_option=gc)
self.ws.run(net)
Y = self.ws.blobs["Y"].fetch()
if fall_back_to_NCHW:
Y = nchw2nhwc(Y)
outputs.append(
Output(Y=Y, op_type=op_type, engine=engine, order=order))
check_quantized_results_close(outputs)
def test_dnnlowp_depthwise_3x3x3_conv(
self,
stride_0,
stride_1,
stride_2,
size,
group,
batch_size,
prepack_weight,
fuse_relu,
share_col_buffer,
preserve_activation_sparsity,
preserve_weight_sparsity,
quantize_groupwise,
gc,
dc,
):
pad = 1
kernel = 3
dilation = 1
input_channels_per_group = 1
output_channels_per_group = 1
order = "NHWC"
X, W, b = generate_convnd_inputs(
(stride_0, stride_1, stride_2),
(pad,) * 3,
(kernel,) * 3,
(dilation,) * 3,
(size,) * 3,
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 = []
op = "ConvRelu" if fuse_relu else "Conv"
op_engine_list = [(op, ""), (op, "DNNLOWP"), ("Int8" + op, "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"],
group=group,
quantize_groupwise=quantize_groupwise,
preserve_weight_sparsity=preserve_weight_sparsity,
in_scale=x_q_param.scale,
engine=engine,
)
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"],
strides=[stride_0, stride_1, stride_2],
kernels=[kernel] * 3,
dilations=[dilation] * 3,
pads=[pad] * (3 * 2),
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])
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_3x3x3_conv(
self,
stride,
size,
group,
batch_size,
prepack_weight,
fuse_relu,
share_col_buffer,
preserve_activation_sparsity,
preserve_weight_sparsity,
gc,
dc,
):
pad = 1
kernel = 3
dilation = 1
input_channels_per_group = 1
output_channels_per_group = 1
order = "NHWC"
X, W, b = generate_convnd_inputs(
(stride, ) * 3,
(pad, ) * 3,
(kernel, ) * 3,
(dilation, ) * 3,
(size, ) * 3,
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 = "ConvRelu" if fuse_relu else "Conv"
op_engine_list = [(op, ""), (op, "DNNLOWP"), ("Int8" + op, "DNNLOWP")]
for op_type, engine in op_engine_list:
init_net = core.Net("test_init_net")
net = core.Net("test_net")
# TODO: no fall back to NCHW
fall_back_to_NCHW = "DNNLOWP" not in engine
if fall_back_to_NCHW:
X_nchw = nhwc2nchw(X)
W_nchw = nhwc2nchw(W)
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_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"],
group=group,
preserve_weight_sparsity=preserve_weight_sparsity,
in_scale=x_q_param.scale,
engine=engine,
)
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"],
strides=[stride] * 3,
kernels=[kernel] * 3,
dilations=[dilation] * 3,
pads=[pad] * (3 * 2),
order="NCHW" if fall_back_to_NCHW else 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_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])
self.ws.create_blob("X").feed(X_nchw if fall_back_to_NCHW else X,
device_option=gc)
self.ws.create_blob("W").feed(W_nchw if fall_back_to_NCHW else W,
device_option=gc)
self.ws.create_blob("b").feed(b, device_option=gc)
self.ws.run(init_net)
self.ws.run(net)
Y = self.ws.blobs["Y"].fetch()
if fall_back_to_NCHW:
Y = nchw2nhwc(Y)
outputs.append(
Output(Y=Y, op_type=op_type, engine=engine, order=order))
check_quantized_results_close(outputs,
symmetric=preserve_activation_sparsity)
def _test_dnnlowp_nd_int(
self,
stride,
pad,
kernels,
dilation,
size,
group,
input_channels_per_group,
output_channels_per_group,
batch_size,
order,
prepack_weight,
gc,
dc,
):
assume(group == 1 or dilation == 1)
assume((not prepack_weight) or order == "NHWC")
ndim = len(kernels)
X, W, b = generate_convnd_inputs(
(stride,) * ndim,
(pad,) * ndim,
kernels,
(dilation,) * ndim,
(size,) * ndim,
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", ""), ("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 len(outputs) > 0
do_prepack_weight = engine == "DNNLOWP" and prepack_weight
if do_quantize:
quantize = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize])
x_q_param = dnnlowp_utils.choose_quantization_params(X.min(), X.max())
if do_quantize_weight:
int8_given_tensor_fill, w_q_param = dnnlowp_utils.create_int8_given_tensor_fill(
W, "W_q"
)
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"],
group=group,
in_scale=x_q_param.scale,
engine=engine,
)
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"],
strides=[stride] * ndim,
kernels=kernels,
dilations=[dilation] * ndim,
pads=[pad] * (ndim * 2),
order=order,
dequantize_output=not do_dequantize,
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])
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)
