372 lines
12 KiB
Python
372 lines
12 KiB
Python
# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
import logging
|
|
|
|
import numpy as np
|
|
import tensorrt as trt
|
|
|
|
from paddle.base.log_helper import get_logger
|
|
from paddle.tensorrt.converter_utils import (
|
|
WithFp16,
|
|
add_1D_constant_layer,
|
|
get_axes_for_reduce_op,
|
|
get_dynamic_dims,
|
|
get_trt_plugin,
|
|
has_dynamic_shape,
|
|
set_layer_name,
|
|
trt_expand,
|
|
trt_prod,
|
|
trt_reshape,
|
|
trt_sum,
|
|
)
|
|
from paddle.tensorrt.register import converter_registry
|
|
from paddle.tensorrt.util import (
|
|
RefitManager,
|
|
RefitRole,
|
|
TensorRTConstantManager,
|
|
support_fp32_mix_precision,
|
|
)
|
|
|
|
_logger = get_logger(
|
|
__name__, logging.INFO, fmt='%(asctime)s-%(levelname)s: %(message)s'
|
|
)
|
|
|
|
|
|
@converter_registry.register(
|
|
"pd_op.layer_norm", trt_version="trt_version_ge=8.6"
|
|
)
|
|
def layernorm_converter(network, paddle_op, inputs):
|
|
input_a, scale, bias = inputs
|
|
|
|
begin_norm_axis = paddle_op.attrs().get("begin_norm_axis", 0)
|
|
epsilon = paddle_op.attrs().get("epsilon", 1e-5)
|
|
assert len(paddle_op.operands()) == 3
|
|
scale_shape = paddle_op.operands()[1].source().shape
|
|
|
|
if isinstance(scale, trt.Weights):
|
|
scale_tensor = network.add_constant(scale_shape, scale)
|
|
set_layer_name(scale_tensor, paddle_op)
|
|
scale_tensor = scale_tensor.get_output(0)
|
|
bias_shape = paddle_op.operands()[2].source().shape
|
|
bias_tensor = network.add_constant(bias_shape, bias)
|
|
set_layer_name(bias_tensor, paddle_op)
|
|
bias_tensor = bias_tensor.get_output(0)
|
|
else:
|
|
scale_tensor = scale
|
|
bias_tensor = bias
|
|
|
|
dims = list(range(len(input_a.shape)))[begin_norm_axis:]
|
|
axes = get_axes_for_reduce_op(dims)
|
|
|
|
broadcast_shape = [1] * begin_norm_axis
|
|
normalized_shape = list(input_a.shape)[begin_norm_axis:]
|
|
broadcast_shape.extend(normalized_shape)
|
|
|
|
scale_reshape = network.add_shuffle(scale_tensor)
|
|
scale_reshape.reshape_dims = tuple(broadcast_shape)
|
|
set_layer_name(scale_reshape, [paddle_op.name(), "scale_reshape"])
|
|
scale_tensor = scale_reshape.get_output(0)
|
|
|
|
bias_reshape = network.add_shuffle(bias_tensor)
|
|
bias_reshape.reshape_dims = tuple(broadcast_shape)
|
|
set_layer_name(bias_reshape, [paddle_op.name(), "bias_reshape"])
|
|
bias_tensor = bias_reshape.get_output(0)
|
|
|
|
layer_norm = network.add_normalization(
|
|
input_a, scale_tensor, bias_tensor, axes
|
|
)
|
|
layer_norm.epsilon = epsilon
|
|
set_layer_name(layer_norm, paddle_op)
|
|
support_fp32_mix_precision(paddle_op.name(), layer_norm)
|
|
layer_norm.compute_precision = trt.float32
|
|
return layer_norm.get_output(0)
|
|
|
|
|
|
@converter_registry.register("pd_op.batch_norm")
|
|
@converter_registry.register("pd_op.batch_norm_")
|
|
def batch_norm_converter(network, paddle_op, inputs):
|
|
constant_manager = TensorRTConstantManager()
|
|
refit_manager = RefitManager()
|
|
|
|
input_tensor, mean, variance, scale, bias = inputs
|
|
|
|
scale_shape = paddle_op.operands()[3].source().shape
|
|
eps = paddle_op.attrs().get("epsilon", 1e-8)
|
|
|
|
scale_name = None
|
|
bias_name = None
|
|
if isinstance(mean, trt.ITensor):
|
|
mean_name = (
|
|
paddle_op.operands()[1]
|
|
.source()
|
|
.get_defining_op()
|
|
.attrs()['parameter_name']
|
|
)
|
|
variance_name = (
|
|
paddle_op.operands()[2]
|
|
.source()
|
|
.get_defining_op()
|
|
.attrs()['parameter_name']
|
|
)
|
|
scale_name = (
|
|
paddle_op.operands()[3]
|
|
.source()
|
|
.get_defining_op()
|
|
.attrs()['parameter_name']
|
|
)
|
|
bias_name = (
|
|
paddle_op.operands()[4]
|
|
.source()
|
|
.get_defining_op()
|
|
.attrs()['parameter_name']
|
|
)
|
|
mean_np = constant_manager.get_constant_value(mean_name)
|
|
variance_np = constant_manager.get_constant_value(variance_name)
|
|
scale_np = constant_manager.get_constant_value(scale_name)
|
|
bias_np = constant_manager.get_constant_value(bias_name)
|
|
else:
|
|
mean_np = mean.numpy()
|
|
variance_np = variance.numpy()
|
|
scale_np = scale.numpy()
|
|
bias_np = bias.numpy()
|
|
|
|
actual_scale_np = scale_np / np.sqrt(variance_np + eps)
|
|
actual_bias_np = bias_np - mean_np * actual_scale_np
|
|
|
|
bias = trt.Weights(actual_bias_np)
|
|
scale = trt.Weights(actual_scale_np)
|
|
power = trt.Weights(np.ones(scale_shape, dtype='float32'))
|
|
|
|
input_tensor_shape = paddle_op.operands()[0].source().shape
|
|
if has_dynamic_shape(input_tensor_shape):
|
|
assert input_tensor.shape[1] != -1, (
|
|
"Channel dim can't be dynamic for batch norm."
|
|
)
|
|
|
|
output_shape = input_tensor_shape
|
|
|
|
if not network.has_implicit_batch_dimension and len(input_tensor_shape) < 4:
|
|
assert len(get_dynamic_dims(input_tensor.shape)) <= 1, (
|
|
"BatchNorm1D with more than one dynamic dims is not currently supported."
|
|
)
|
|
reshape_layer = network.add_shuffle(input_tensor)
|
|
if len(input_tensor_shape) == 2:
|
|
reshape_layer.reshape_dims = (
|
|
input_tensor_shape[0],
|
|
input_tensor_shape[1],
|
|
1,
|
|
1,
|
|
)
|
|
else: # len(input_tensor_shape) ==3
|
|
reshape_layer.reshape_dims = (
|
|
input_tensor_shape[0],
|
|
input_tensor_shape[1],
|
|
input_tensor_shape[2],
|
|
1,
|
|
)
|
|
set_layer_name(reshape_layer, paddle_op)
|
|
input_tensor = reshape_layer.get_output(0)
|
|
|
|
batch_norm_layer = network.add_scale(
|
|
input_tensor, trt.ScaleMode.CHANNEL, bias, scale, power
|
|
)
|
|
support_fp32_mix_precision(paddle_op.name(), batch_norm_layer)
|
|
set_layer_name(batch_norm_layer, paddle_op)
|
|
if isinstance(mean, trt.ITensor):
|
|
refit_manager.set_mapping(
|
|
bias_name, batch_norm_layer.name, RefitRole.SHIFT
|
|
)
|
|
refit_manager.set_mapping(
|
|
scale_name, batch_norm_layer.name, RefitRole.SCALE
|
|
)
|
|
|
|
if not network.has_implicit_batch_dimension and len(output_shape) < 4:
|
|
reshape_output_layer = network.add_shuffle(
|
|
batch_norm_layer.get_output(0)
|
|
)
|
|
reshape_output_layer.reshape_dims = tuple(output_shape)
|
|
batch_norm_layer = reshape_output_layer
|
|
set_layer_name(batch_norm_layer, paddle_op)
|
|
|
|
return batch_norm_layer.get_output(0)
|
|
|
|
|
|
@converter_registry.register("pd_op.instance_norm")
|
|
def instance_norm_converter(network, paddle_op, inputs):
|
|
eps = paddle_op.attrs().get("epsilon", 1e-8)
|
|
instance_norm_inputs = [inputs[0], inputs[1], inputs[2]]
|
|
plugin_fields = [
|
|
trt.PluginField(
|
|
"epsilon",
|
|
np.array(eps, dtype=np.float32),
|
|
trt.PluginFieldType.FLOAT32,
|
|
),
|
|
]
|
|
plugin_field_collection = trt.PluginFieldCollection(plugin_fields)
|
|
plugin_name = "pir_instance_norm"
|
|
plugin_version = "1"
|
|
plugin = get_trt_plugin(
|
|
plugin_name, plugin_field_collection, plugin_version
|
|
)
|
|
instance_norm_layer = network.add_plugin_v2(instance_norm_inputs, plugin)
|
|
set_layer_name(instance_norm_layer, paddle_op)
|
|
return instance_norm_layer.get_output(0)
|
|
|
|
|
|
@converter_registry.register(
|
|
"pd_op.fused_bias_dropout_residual_layer_norm",
|
|
trt_version="trt_version_ge=8.0",
|
|
)
|
|
def fused_bias_dropout_residual_layer_norm_converter(
|
|
network, paddle_op, inputs
|
|
):
|
|
input1, input2, ele_bias, scale, bias = inputs
|
|
if isinstance(ele_bias, trt.ITensor):
|
|
refit_manager = RefitManager
|
|
ele_bias = refit_manager.get_trt_weight_tensor(ele_bias.name)
|
|
scale = refit_manager.get_trt_weight_tensor(scale.name)
|
|
bias = refit_manager.get_trt_weight_tensor(bias.name)
|
|
else:
|
|
ele_bias = ele_bias
|
|
scale = scale
|
|
bias = bias
|
|
has_bias = ele_bias is not None
|
|
bias_size = bias.size
|
|
scale_size = scale.size
|
|
ele_bias_size = ele_bias.size if has_bias else 0
|
|
epsilon = paddle_op.attrs().get("ln_epsilon", 1e-5)
|
|
with_fp16 = int(WithFp16())
|
|
# TODO: FusedBiasDropoutResidualLayerNorm will support FP16 UT in the future.
|
|
if with_fp16 == 1:
|
|
raise NotImplementedError(
|
|
"FusedBiasDropoutResidualLayerNorm will support FP16 UT in the future."
|
|
)
|
|
ele_bias_data = (
|
|
ele_bias.numpy().astype('float16') if with_fp16 else ele_bias.numpy()
|
|
)
|
|
plugin_fields = [
|
|
trt.PluginField("bias", bias.numpy(), trt.PluginFieldType.FLOAT32),
|
|
trt.PluginField("scale", scale.numpy(), trt.PluginFieldType.FLOAT32),
|
|
trt.PluginField(
|
|
"ele_bias",
|
|
ele_bias_data,
|
|
(
|
|
trt.PluginFieldType.FLOAT16
|
|
if with_fp16
|
|
else trt.PluginFieldType.FLOAT32
|
|
),
|
|
),
|
|
trt.PluginField(
|
|
"bias_size",
|
|
np.array([bias_size], dtype=np.int32),
|
|
trt.PluginFieldType.INT32,
|
|
),
|
|
trt.PluginField(
|
|
"scale_size",
|
|
np.array([scale_size], dtype=np.int32),
|
|
trt.PluginFieldType.INT32,
|
|
),
|
|
trt.PluginField(
|
|
"ele_bias_size",
|
|
np.array([ele_bias_size], dtype=np.int32),
|
|
trt.PluginFieldType.INT32,
|
|
),
|
|
trt.PluginField(
|
|
"epsilon",
|
|
np.array([epsilon], dtype=np.float32),
|
|
trt.PluginFieldType.FLOAT32,
|
|
),
|
|
trt.PluginField(
|
|
"with_fp16",
|
|
np.array([with_fp16], dtype=np.bool_),
|
|
trt.PluginFieldType.INT32,
|
|
),
|
|
]
|
|
plugin_field_collection = trt.PluginFieldCollection(plugin_fields)
|
|
plugin_name = "pir_preln_residual_bias_plugin_dynamic"
|
|
plugin_version = "1"
|
|
plugin = get_trt_plugin(
|
|
plugin_name, plugin_field_collection, plugin_version
|
|
)
|
|
plugin_inputs = [input1, input2]
|
|
layer = network.add_plugin_v2(plugin_inputs, plugin)
|
|
set_layer_name(layer, paddle_op)
|
|
return layer.get_output(0)
|
|
|
|
|
|
@converter_registry.register(
|
|
"pd_op.group_norm", trt_version="trt_version_ge=8.6"
|
|
)
|
|
def group_norm_converter(network, paddle_op, inputs):
|
|
x, scale, bias = inputs
|
|
groups = paddle_op.attrs().get("groups", 1)
|
|
eps = paddle_op.attrs().get("epsilon", 1e-05)
|
|
|
|
axes_mask = 0
|
|
x_shape = paddle_op.operands()[0].source().shape
|
|
rank_x = len(x_shape)
|
|
|
|
fake_shape = [1, groups] + [1] * (rank_x - 2)
|
|
broadcast_shape = [1, x_shape[1]] + [1] * (rank_x - 2)
|
|
for d in range(2, rank_x):
|
|
axes_mask |= 1 << d
|
|
|
|
weight_one = add_1D_constant_layer(
|
|
network, 1.0, np.float32, name=[paddle_op.name(), 'weight_one']
|
|
)
|
|
bias_zero = add_1D_constant_layer(
|
|
network, 0.0, np.float32, name=[paddle_op.name(), 'bias_zero']
|
|
)
|
|
fake_shape = add_1D_constant_layer(
|
|
network, fake_shape, np.int32, name=[paddle_op.name(), 'fake_shape']
|
|
)
|
|
weight_one = trt_expand(
|
|
network,
|
|
weight_one,
|
|
1,
|
|
fake_shape,
|
|
rank_x,
|
|
name=[paddle_op.name(), 'weight_one'],
|
|
)
|
|
bias_zero = trt_expand(
|
|
network,
|
|
bias_zero,
|
|
1,
|
|
fake_shape,
|
|
rank_x,
|
|
name=[paddle_op.name(), 'bias_zero'],
|
|
)
|
|
layer = network.add_normalization(x, weight_one, bias_zero, axes_mask)
|
|
layer.num_groups = groups
|
|
layer.epsilon = eps
|
|
set_layer_name(layer, paddle_op)
|
|
output = layer.get_output(0)
|
|
if scale is not None:
|
|
scale = trt_reshape(
|
|
network, scale, broadcast_shape, name=[paddle_op.name(), 'scale']
|
|
)
|
|
output = trt_prod(
|
|
network, output, scale, name=[paddle_op.name(), 'output']
|
|
)
|
|
if bias is not None:
|
|
bias = trt_reshape(
|
|
network, bias, broadcast_shape, name=[paddle_op.name(), 'bias']
|
|
)
|
|
output = trt_sum(
|
|
network, output, bias, name=[paddle_op.name(), 'output']
|
|
)
|
|
|
|
return output
|