update tensorrt impls (#78833)

Author: co63oc

python/paddle/tensorrt/impls/activation.py

@@ -38,9 +38,9 @@
 }
 
 
-@converter_registry.register("pd_op.relu", trt_version="trt_version_ge=8.0")
-@converter_registry.register("pd_op.tanh", trt_version="trt_version_ge=8.0")
-@converter_registry.register("pd_op.sigmoid", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.relu")
+@converter_registry.register("pd_op.tanh")
+@converter_registry.register("pd_op.sigmoid")
 def activation_converter(network, paddle_op, inputs):
     layer = network.add_activation(
         inputs[0], activation_type_map[paddle_op.name()]
@@ -49,9 +49,7 @@ def activation_converter(network, paddle_op, inputs):
     return layer.get_output(0)
 
 
-@converter_registry.register(
-    "pd_op.logsigmoid", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.logsigmoid")
 def logsigmoid_converter(network, paddle_op, inputs):
     sigmoid_layer = network.add_activation(
         inputs[0], trt.ActivationType.SIGMOID
@@ -64,7 +62,7 @@ def logsigmoid_converter(network, paddle_op, inputs):
     return layer.get_output(0)
 
 
-@converter_registry.register("pd_op.relu6", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.relu6")
 def relu6_converter(network, paddle_op, inputs):
     layer = network.add_activation(inputs[0], trt.ActivationType.CLIP)
     layer.alpha = 0.0
@@ -73,7 +71,7 @@ def relu6_converter(network, paddle_op, inputs):
     return layer.get_output(0)
 
 
-@converter_registry.register("pd_op.softmax", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.softmax")
 def softmax_converter(network, paddle_op, inputs):
     from paddle.tensorrt.util import support_fp32_mix_precision
 
@@ -126,7 +124,7 @@ def softmax_converter(network, paddle_op, inputs):
     return layer.get_output(0)
 
 
-@converter_registry.register("pd_op.gelu", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.gelu")
 def gelu_converter(network, paddle_op, inputs):
     input_val = inputs[0]
     approximate = paddle_op.attrs()["approximate"]
@@ -261,9 +259,7 @@ def gelu_converter(network, paddle_op, inputs):
         return y
 
 
-@converter_registry.register(
-    "pd_op.hardsigmoid", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.hardsigmoid")
 def hardsigmoid_converter(network, paddle_op, inputs):
     x = inputs[0]
     slope = paddle_op.attrs()["slope"]
@@ -277,9 +273,7 @@ def hardsigmoid_converter(network, paddle_op, inputs):
     return hardsigmoid_layer.get_output(0)
 
 
-@converter_registry.register(
-    "pd_op.hardswish", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.hardswish")
 def hardswish_converter(network, paddle_op, inputs):
     x = inputs[0]
     threshold = 6.0

python/paddle/tensorrt/impls/attribute.py

@@ -16,12 +16,12 @@
 from paddle.tensorrt.register import converter_registry
 
 
-@converter_registry.register("pd_op.shape", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.shape")
 def shape_converter(network, paddle_op, inputs):
     return trt_shape(network, inputs[0], name=[paddle_op.name(), 'trt_shape'])
 
 
-@converter_registry.register("pd_op.shape64", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.shape64")
 def shape64_converter(network, paddle_op, inputs):
     input_tensor = inputs[0]
     shape_layer = network.add_shape(input_tensor)

python/paddle/tensorrt/impls/common.py

@@ -58,9 +58,7 @@ def dropout_converter(network, paddle_op, inputs):
     return scale_layer.get_output(0)
 
 
-@converter_registry.register(
-    "pd_op.bilinear_interp", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.bilinear_interp")
 def bilinear_interp_converter(network, paddle_op, inputs):
     input_tensor = inputs[0]
     input_shape_tensor = network.add_shape(input_tensor)
@@ -275,9 +273,7 @@ def bilinear_interp_converter(network, paddle_op, inputs):
     return resize_layer.get_output(0)
 
 
-@converter_registry.register(
-    "pd_op.embedding", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.embedding")
 def embedding_converter(network, paddle_op, inputs):
     x = inputs[0]
     weight = inputs[1]
@@ -286,7 +282,7 @@ def embedding_converter(network, paddle_op, inputs):
     return gather_layer.get_output(0)
 
 
-@converter_registry.register("pd_op.unbind", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.unbind")
 def unbind_converter(network, paddle_op, inputs):
     x = inputs[0]
     input_shape = x.shape
@@ -361,9 +357,7 @@ def unbind_converter(network, paddle_op, inputs):
     return outputs
 
 
-@converter_registry.register(
-    "pd_op.nearest_interp", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.nearest_interp")
 def nearest_interp_converter(network, paddle_op, inputs):
     input_tensor = inputs[0]
     input_shape_tensor = network.add_shape(input_tensor)
@@ -468,9 +462,7 @@ def nearest_interp_converter(network, paddle_op, inputs):
     return resize_layer.get_output(0)
 
 
-@converter_registry.register(
-    "pd_op.linear_interp", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.linear_interp")
 def linear_interp_converter(network, paddle_op, inputs):
     input_tensor = inputs[0]
     data_layout = paddle_op.attrs().get("data_format")

python/paddle/tensorrt/impls/conv.py

@@ -20,26 +20,16 @@
 from paddle.tensorrt.register import converter_registry
 
 
-@converter_registry.register(
-    "pd_op.depthwise_conv2d", trt_version="trt_version_ge=8.0"
-)
-@converter_registry.register("pd_op.conv2d", trt_version="trt_version_ge=8.0")
-@converter_registry.register(
-    "pd_op.fused_conv2d_add_act", trt_version="trt_version_ge=8.0"
-)
-@converter_registry.register(
-    "pd_op.conv2d_transpose", trt_version="trt_version_ge=8.0"
-)
-@converter_registry.register(
-    "pd_op.depthwise_conv2d_transpose", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.depthwise_conv2d")
+@converter_registry.register("pd_op.conv2d")
+@converter_registry.register("pd_op.fused_conv2d_add_act")
+@converter_registry.register("pd_op.conv2d_transpose")
+@converter_registry.register("pd_op.depthwise_conv2d_transpose")
 def conv2d_converter(network, paddle_op, inputs):
     return convert_conv2d(network, paddle_op, inputs)
 
 
-@converter_registry.register(
-    "pd_op.conv3d_transpose", trt_version="trt_version_ge=8.0"
-)
-@converter_registry.register("pd_op.conv3d", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.conv3d_transpose")
+@converter_registry.register("pd_op.conv3d")
 def conv3d_converter(network, paddle_op, inputs):
     return convert_conv3d(network, paddle_op, inputs)

python/paddle/tensorrt/impls/creation.py

@@ -35,9 +35,7 @@
 from paddle.tensorrt.register import converter_registry
 
 
-@converter_registry.register(
-    "pd_op.full_int_array", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.full_int_array")
 def full_int_array_converter(network, paddle_op, inputs):
     value = paddle_op.attrs()["value"]
     if len(value) == 0:
@@ -48,7 +46,7 @@ def full_int_array_converter(network, paddle_op, inputs):
     return full_int_array_layer.get_output(0)
 
 
-@converter_registry.register("pd_op.full", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.full")
 def full_converter(network, paddle_op, inputs):
     shape = paddle_op.attrs()["shape"]
     value = paddle_op.attrs().get("value", 1.0)

python/paddle/tensorrt/impls/linalg.py

@@ -27,7 +27,7 @@
 from paddle.tensorrt.util import support_fp32_mix_precision
 
 
-@converter_registry.register("pd_op.matmul", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.matmul")
 def matmul_converter(network, paddle_op, inputs):
     weight_shape = paddle_op.operands()[1].source().shape
     transpose_x = paddle_op.attrs()["transpose_x"]
@@ -71,9 +71,7 @@ def matmul_converter(network, paddle_op, inputs):
     return out.get_output(0)
 
 
-@converter_registry.register(
-    "pd_op.transpose", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.transpose")
 def transpose_converter(network, paddle_op, inputs):
     perm = paddle_op.attrs()["perm"]
     transposed_tensor = network.add_shuffle(inputs[0])

python/paddle/tensorrt/impls/manipulation.py

@@ -47,7 +47,7 @@
 from ..util import get_trt_version_list
 
 
-@converter_registry.register("pd_op.reshape", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.reshape")
 def reshape_converter(network, paddle_op, inputs):
     x = inputs[0]
     is_constant_shape = False
@@ -112,7 +112,7 @@ def gather_nd_converter(network, paddle_op, inputs):
     return non_zero_layer.get_output(0)
 
 
-@converter_registry.register("pd_op.flatten", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.flatten")
 def flatten_converter(network, paddle_op, inputs):
     input_val = inputs[0]
     input_val_shape = paddle_op.operands()[0].source().shape
@@ -199,7 +199,7 @@ def flatten_converter(network, paddle_op, inputs):
 
 
 # In the converter, pd_op.concat has three inputs, because builtin.combine has two inputs.
-@converter_registry.register("pd_op.concat", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.concat")
 def concat_converter(network, paddle_op, inputs):
     input_tensors = inputs[0]
     axis_tensor = inputs[1]
@@ -215,12 +215,8 @@ def concat_converter(network, paddle_op, inputs):
     return concat_layer.get_output(0)
 
 
-@converter_registry.register(
-    "pd_op.unsqueeze", trt_version="trt_version_ge=8.0"
-)
-@converter_registry.register(
-    "pd_op.unsqueeze_", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.unsqueeze")
+@converter_registry.register("pd_op.unsqueeze_")
 def unsqueeze_converter(network, paddle_op, inputs):
     x = inputs[0]
     input_dims = x.shape
@@ -277,8 +273,8 @@ def unsqueeze_converter(network, paddle_op, inputs):
     return layer.get_output(0)
 
 
-@converter_registry.register("pd_op.squeeze", trt_version="trt_version_ge=8.0")
-@converter_registry.register("pd_op.squeeze_", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.squeeze")
+@converter_registry.register("pd_op.squeeze_")
 def squeeze_converter(network, paddle_op, inputs):
     input_val = inputs[0]
     input_shape = input_val.shape
@@ -342,7 +338,7 @@ def squeeze_converter(network, paddle_op, inputs):
     return layer.get_output(0)
 
 
-@converter_registry.register("pd_op.expand", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.expand")
 def expand_converter(network, paddle_op, inputs):
     input = inputs[0]
     input_dims = input.shape
@@ -374,9 +370,7 @@ def expand_converter(network, paddle_op, inputs):
     )
 
 
-@converter_registry.register(
-    "pd_op.expand_as", trt_version="trt_version_ge=8.0"
-)
+@converter_registry.register("pd_op.expand_as")
 def expand_as_converter(network, paddle_op, inputs):
     input = inputs[0]
     input_dims = input.shape
@@ -434,7 +428,7 @@ def cast_converter(network, paddle_op, inputs):
     return cast_layer.get_output(0)
 
 
-@converter_registry.register("pd_op.slice", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.slice")
 def slice_converter(network, paddle_op, inputs):
     input_tensor = inputs[0]
     axes = paddle_op.attrs()["axes"]

python/paddle/tensorrt/impls/math.py

@@ -40,15 +40,15 @@
 from paddle.tensorrt.register import converter_registry
 
 
-@converter_registry.register("pd_op.add", trt_version="trt_version_ge=8.0")
-@converter_registry.register("pd_op.add_", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.add")
+@converter_registry.register("pd_op.add_")
 def add_converter(network, paddle_op, inputs):
     return add_elementwise_layer(
         network, paddle_op, inputs, trt.ElementWiseOperation.SUM
     )
 
 
-@converter_registry.register("pd_op.scale", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.scale")
 def scale_converter(network, paddle_op, inputs):
     x = inputs[0]
     bias = paddle_op.attrs().get("bias", 0.0)
@@ -152,7 +152,7 @@ def scale_converter(network, paddle_op, inputs):
     return layer.get_output(0)
 
 
-@converter_registry.register("pd_op.max", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.max")
 def max_converter(network, paddle_op, inputs):
     input_tensor = inputs[0]
     axis = get_input_constant_value(paddle_op, inputs, 1)
@@ -178,21 +178,21 @@ def max_converter(network, paddle_op, inputs):
     return layer.get_output(0)
 
 
-@converter_registry.register("pd_op.divide", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.divide")
 def divide_converter(network, paddle_op, inputs):
     return add_elementwise_layer(
         network, paddle_op, inputs, trt.ElementWiseOperation.DIV
     )
 
 
-@converter_registry.register("pd_op.subtract", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.subtract")
 def subtract_converter(network, paddle_op, inputs):
     return add_elementwise_layer(
         network, paddle_op, inputs, trt.ElementWiseOperation.SUB
     )
 
 
-@converter_registry.register("pd_op.multiply", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.multiply")
 def multiply_converter(network, paddle_op, inputs):
     return add_elementwise_layer(
         network, paddle_op, inputs, trt.ElementWiseOperation.PROD
@@ -257,7 +257,7 @@ def _get_constant_or_expand_tensor(
     return layer.get_output(0)
 
 
-@converter_registry.register("pd_op.pow", trt_version="trt_version_ge=8.0")
+@converter_registry.register("pd_op.pow")
 def pow_converter(network, paddle_op, inputs):
     from paddle.tensorrt.util import support_fp32_mix_precision
 

python/paddle/tensorrt/impls/norm.py

@@ -94,12 +94,8 @@ def layernorm_converter(network, paddle_op, inputs):
     return layer_norm.get_output(0)
 
 
-@converter_registry.register(
-    "pd_op.batch_norm", trt_version="trt_version_ge=8.0"
-)
-@converter_registry.register(
-    "pd_op.batch_norm_", trt_version="trt_version_ge=8.0"
-)
+@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()
@@ -207,9 +203,7 @@ def batch_norm_converter(network, paddle_op, inputs):
     return batch_norm_layer.get_output(0)
 
 
-@converter_registry.register(
-    "pd_op.instance_norm", trt_version="trt_version_ge=8.0"
-)
+@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]]