Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 b4b78060 编写于 作者: W wenbin 提交者： GitHub
浏览文件
操作

elementwise_floordiv (#47944) 

* elementwise_op

* add teller

* modify ut

* comments

* modify ut

* return

* modify
上级 a2a97cbb
显示空白变更内容
内联 并排
Showing with 121 addition and 36 deletion
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @ b4b78060
...@@ -2227,6 +2227,7 @@ USE_TRT_CONVERTER(elementwise_div_weight); ...@@ -2227,6 +2227,7 @@ USE_TRT_CONVERTER(elementwise_div_weight);
USE_TRT_CONVERTER(elementwise_min_weight); USE_TRT_CONVERTER(elementwise_min_weight);
USE_TRT_CONVERTER(elementwise_max_weight); USE_TRT_CONVERTER(elementwise_max_weight);
USE_TRT_CONVERTER(elementwise_pow_weight); USE_TRT_CONVERTER(elementwise_pow_weight);
USE_TRT_CONVERTER(elementwise_floordiv_weight);
USE_TRT_CONVERTER(elementwise_add_tensor); USE_TRT_CONVERTER(elementwise_add_tensor);
USE_TRT_CONVERTER(elementwise_sub_tensor); USE_TRT_CONVERTER(elementwise_sub_tensor);
USE_TRT_CONVERTER(elementwise_div_tensor); USE_TRT_CONVERTER(elementwise_div_tensor);
...@@ -2234,6 +2235,7 @@ USE_TRT_CONVERTER(elementwise_mul_tensor); ...@@ -2234,6 +2235,7 @@ USE_TRT_CONVERTER(elementwise_mul_tensor);
USE_TRT_CONVERTER(elementwise_max_tensor); USE_TRT_CONVERTER(elementwise_max_tensor);
USE_TRT_CONVERTER(elementwise_min_tensor); USE_TRT_CONVERTER(elementwise_min_tensor);
USE_TRT_CONVERTER(elementwise_pow_tensor); USE_TRT_CONVERTER(elementwise_pow_tensor);
USE_TRT_CONVERTER(elementwise_floordiv_tensor);
USE_TRT_CONVERTER(transpose); USE_TRT_CONVERTER(transpose);
USE_TRT_CONVERTER(transpose2); USE_TRT_CONVERTER(transpose2);
USE_TRT_CONVERTER(flatten); USE_TRT_CONVERTER(flatten);
......
paddle/fluid/inference/tensorrt/convert/elementwise_op.cc
浏览文件 @ b4b78060
...@@ -167,6 +167,7 @@ const std::unordered_map<std::string, nvinfer1::ElementWiseOperation> ...@@ -167,6 +167,7 @@ const std::unordered_map<std::string, nvinfer1::ElementWiseOperation>
{"min", nvinfer1::ElementWiseOperation::kMIN}, {"min", nvinfer1::ElementWiseOperation::kMIN},
{"pow", nvinfer1::ElementWiseOperation::kPOW}, {"pow", nvinfer1::ElementWiseOperation::kPOW},
{"max", nvinfer1::ElementWiseOperation::kMAX}, {"max", nvinfer1::ElementWiseOperation::kMAX},
{"floordiv", nvinfer1::ElementWiseOperation::kFLOOR_DIV},
}; };
class ElementwiseTensorAddOpConverter : public ElementwiseTensorOpConverter { class ElementwiseTensorAddOpConverter : public ElementwiseTensorOpConverter {
...@@ -204,6 +205,12 @@ class ElementwiseTensorPowOpConverter : public ElementwiseTensorOpConverter { ...@@ -204,6 +205,12 @@ class ElementwiseTensorPowOpConverter : public ElementwiseTensorOpConverter {
ElementwiseTensorPowOpConverter() { op_type_ = "pow"; } ElementwiseTensorPowOpConverter() { op_type_ = "pow"; }
}; };
class ElementwiseTensorFloorDivOpConverter
: public ElementwiseTensorOpConverter {
public:
ElementwiseTensorFloorDivOpConverter() { op_type_ = "floordiv"; }
};
} // namespace tensorrt } // namespace tensorrt
} // namespace inference } // namespace inference
} // namespace paddle } // namespace paddle
...@@ -222,6 +229,8 @@ REGISTER_TRT_OP_CONVERTER(elementwise_min_weight, ...@@ -222,6 +229,8 @@ REGISTER_TRT_OP_CONVERTER(elementwise_min_weight,
ElementwiseTensorMinOpConverter); ElementwiseTensorMinOpConverter);
REGISTER_TRT_OP_CONVERTER(elementwise_pow_weight, REGISTER_TRT_OP_CONVERTER(elementwise_pow_weight,
ElementwiseTensorPowOpConverter); ElementwiseTensorPowOpConverter);
REGISTER_TRT_OP_CONVERTER(elementwise_floordiv_weight,
ElementwiseTensorFloorDivOpConverter);
REGISTER_TRT_OP_CONVERTER(elementwise_add_tensor, REGISTER_TRT_OP_CONVERTER(elementwise_add_tensor,
ElementwiseTensorAddOpConverter); ElementwiseTensorAddOpConverter);
...@@ -237,3 +246,5 @@ REGISTER_TRT_OP_CONVERTER(elementwise_min_tensor, ...@@ -237,3 +246,5 @@ REGISTER_TRT_OP_CONVERTER(elementwise_min_tensor,
ElementwiseTensorMinOpConverter); ElementwiseTensorMinOpConverter);
REGISTER_TRT_OP_CONVERTER(elementwise_pow_tensor, REGISTER_TRT_OP_CONVERTER(elementwise_pow_tensor,
ElementwiseTensorPowOpConverter); ElementwiseTensorPowOpConverter);
REGISTER_TRT_OP_CONVERTER(elementwise_floordiv_tensor,
ElementwiseTensorFloorDivOpConverter);
paddle/fluid/inference/tensorrt/op_teller.cc
浏览文件 @ b4b78060
...@@ -1244,7 +1244,7 @@ struct SimpleOpTypeSetTeller : public Teller { ...@@ -1244,7 +1244,7 @@ struct SimpleOpTypeSetTeller : public Teller {
if (op_type == "elementwise_add" || op_type == "elementwise_mul" || if (op_type == "elementwise_add" || op_type == "elementwise_mul" ||
op_type == "elementwise_sub" || op_type == "elementwise_div" || op_type == "elementwise_sub" || op_type == "elementwise_div" ||
op_type == "elementwise_pow" || op_type == "elementwise_min" || op_type == "elementwise_pow" || op_type == "elementwise_min" ||
op_type == "elementwise_max") { op_type == "elementwise_max" || op_type == "elementwise_floordiv") {
if (desc.Input("X").size() != 1) { if (desc.Input("X").size() != 1) {
VLOG(3) << "The input op's Input(\"X\").size() " VLOG(3) << "The input op's Input(\"X\").size() "
"should equal to 1, but received Input(\"X\").size() = " "should equal to 1, but received Input(\"X\").size() = "
...@@ -2288,6 +2288,7 @@ struct SimpleOpTypeSetTeller : public Teller { ...@@ -2288,6 +2288,7 @@ struct SimpleOpTypeSetTeller : public Teller {
"elementwise_pow", "elementwise_pow",
"elementwise_min", "elementwise_min",
"elementwise_max", "elementwise_max",
"elementwise_floordiv",
"equal", "equal",
"dropout", "dropout",
"prelu", "prelu",
...@@ -2413,6 +2414,7 @@ struct SimpleOpTypeSetTeller : public Teller { ...@@ -2413,6 +2414,7 @@ struct SimpleOpTypeSetTeller : public Teller {
"elementwise_pow", "elementwise_pow",
"elementwise_min", "elementwise_min",
"elementwise_max", "elementwise_max",
"elementwise_floordiv",
"equal", "equal",
"dropout", "dropout",
"prelu", "prelu",
......
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_elementwise.py
浏览文件 @ b4b78060
...@@ -28,10 +28,21 @@ class TrtConvertElementwiseTest_one_input_special_case0(TrtLayerAutoScanTest): ...@@ -28,10 +28,21 @@ class TrtConvertElementwiseTest_one_input_special_case0(TrtLayerAutoScanTest):
return True return True
def sample_program_configs(self): def sample_program_configs(self):
def generate_input(shape): def generate_input(shape, op_type):
# elementwise_floordiv is integer only
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=shape, dtype=np.int32
)
else:
return np.random.random(shape).astype(np.float32) return np.random.random(shape).astype(np.float32)
def generate_weight(): def generate_weight(op_type):
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=[1, 32, 1, 1], dtype=np.int32
)
else:
return np.random.randn(1, 32, 1, 1).astype(np.float32) return np.random.randn(1, 32, 1, 1).astype(np.float32)
for batch in [1, 4]: for batch in [1, 4]:
...@@ -44,6 +55,7 @@ class TrtConvertElementwiseTest_one_input_special_case0(TrtLayerAutoScanTest): ...@@ -44,6 +55,7 @@ class TrtConvertElementwiseTest_one_input_special_case0(TrtLayerAutoScanTest):
"elementwise_pow", "elementwise_pow",
"elementwise_min", "elementwise_min",
"elementwise_max", "elementwise_max",
"elementwise_floordiv",
]: ]:
for axis in [-1]: for axis in [-1]:
self.dims = len(shape) self.dims = len(shape)
...@@ -65,12 +77,14 @@ class TrtConvertElementwiseTest_one_input_special_case0(TrtLayerAutoScanTest): ...@@ -65,12 +77,14 @@ class TrtConvertElementwiseTest_one_input_special_case0(TrtLayerAutoScanTest):
ops=ops, ops=ops,
weights={ weights={
"weight": TensorConfig( "weight": TensorConfig(
data_gen=partial(generate_weight) data_gen=partial(generate_weight, op_type)
) )
}, },
inputs={ inputs={
"input_data": TensorConfig( "input_data": TensorConfig(
data_gen=partial(generate_input, shape) data_gen=partial(
generate_input, shape, op_type
)
), ),
}, },
outputs=["output_data"], outputs=["output_data"],
...@@ -142,10 +156,22 @@ class TrtConvertElementwiseTest_one_input_special_case1(TrtLayerAutoScanTest): ...@@ -142,10 +156,22 @@ class TrtConvertElementwiseTest_one_input_special_case1(TrtLayerAutoScanTest):
return True return True
def sample_program_configs(self): def sample_program_configs(self):
def generate_input(shape): def generate_input(shape, op_type):
# elementwise_floordiv is integer only
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=shape, dtype=np.int32
)
else:
return np.random.random(shape).astype(np.float32) return np.random.random(shape).astype(np.float32)
def generate_weight(): def generate_weight(op_type):
# elementwise_floordiv is integer only
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=[1], dtype=np.int32
)
else:
return np.random.randn(1).astype(np.float32) return np.random.randn(1).astype(np.float32)
for shape in [[32]]: for shape in [[32]]:
...@@ -157,6 +183,7 @@ class TrtConvertElementwiseTest_one_input_special_case1(TrtLayerAutoScanTest): ...@@ -157,6 +183,7 @@ class TrtConvertElementwiseTest_one_input_special_case1(TrtLayerAutoScanTest):
"elementwise_pow", "elementwise_pow",
"elementwise_min", "elementwise_min",
"elementwise_max", "elementwise_max",
"elementwise_floordiv",
]: ]:
for axis in [-1]: for axis in [-1]:
self.dims = len(shape) self.dims = len(shape)
...@@ -175,12 +202,12 @@ class TrtConvertElementwiseTest_one_input_special_case1(TrtLayerAutoScanTest): ...@@ -175,12 +202,12 @@ class TrtConvertElementwiseTest_one_input_special_case1(TrtLayerAutoScanTest):
ops=ops, ops=ops,
weights={ weights={
"weight": TensorConfig( "weight": TensorConfig(
data_gen=partial(generate_weight) data_gen=partial(generate_weight, op_type)
) )
}, },
inputs={ inputs={
"input_data": TensorConfig( "input_data": TensorConfig(
data_gen=partial(generate_input, shape) data_gen=partial(generate_input, shape, op_type)
), ),
}, },
outputs=["output_data"], outputs=["output_data"],
...@@ -245,10 +272,22 @@ class TrtConvertElementwiseTest_one_input(TrtLayerAutoScanTest): ...@@ -245,10 +272,22 @@ class TrtConvertElementwiseTest_one_input(TrtLayerAutoScanTest):
return True return True
def sample_program_configs(self): def sample_program_configs(self):
def generate_input(shape): def generate_input(shape, op_type):
# elementwise_floordiv is integer only
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=shape, dtype=np.int32
)
else:
return np.random.random(shape).astype(np.float32) return np.random.random(shape).astype(np.float32)
def generate_weight(): def generate_weight(op_type):
# elementwise_floordiv is integer only
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=[32], dtype=np.int32
)
else:
return np.random.randn(32).astype(np.float32) return np.random.randn(32).astype(np.float32)
for batch in [1, 4]: for batch in [1, 4]:
...@@ -266,6 +305,7 @@ class TrtConvertElementwiseTest_one_input(TrtLayerAutoScanTest): ...@@ -266,6 +305,7 @@ class TrtConvertElementwiseTest_one_input(TrtLayerAutoScanTest):
"elementwise_pow", "elementwise_pow",
"elementwise_min", "elementwise_min",
"elementwise_max", "elementwise_max",
"elementwise_floordiv",
]: ]:
for axis in [-1 if len(shape) == 1 else 1]: for axis in [-1 if len(shape) == 1 else 1]:
self.dims = len(shape) self.dims = len(shape)
...@@ -287,12 +327,14 @@ class TrtConvertElementwiseTest_one_input(TrtLayerAutoScanTest): ...@@ -287,12 +327,14 @@ class TrtConvertElementwiseTest_one_input(TrtLayerAutoScanTest):
ops=ops, ops=ops,
weights={ weights={
"weight": TensorConfig( "weight": TensorConfig(
data_gen=partial(generate_weight) data_gen=partial(generate_weight, op_type)
) )
}, },
inputs={ inputs={
"input_data": TensorConfig( "input_data": TensorConfig(
data_gen=partial(generate_input, shape) data_gen=partial(
generate_input, shape, op_type
)
), ),
}, },
outputs=["output_data"], outputs=["output_data"],
...@@ -379,7 +421,13 @@ class TrtConvertElementwiseTest_two_input_without_broadcast( ...@@ -379,7 +421,13 @@ class TrtConvertElementwiseTest_two_input_without_broadcast(
return True return True
def sample_program_configs(self): def sample_program_configs(self):
def generate_input(shape): def generate_input(shape, op_type):
# elementwise_floordiv is integer only
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=shape, dtype=np.int32
)
else:
return np.random.random(shape).astype(np.float32) return np.random.random(shape).astype(np.float32)
for shape in [[4], [4, 32], [2, 32, 16], [1, 8, 16, 32]]: for shape in [[4], [4, 32], [2, 32, 16], [1, 8, 16, 32]]:
...@@ -391,6 +439,7 @@ class TrtConvertElementwiseTest_two_input_without_broadcast( ...@@ -391,6 +439,7 @@ class TrtConvertElementwiseTest_two_input_without_broadcast(
"elementwise_pow", "elementwise_pow",
"elementwise_min", "elementwise_min",
"elementwise_max", "elementwise_max",
"elementwise_floordiv",
]: ]:
for axis in [0, -1]: for axis in [0, -1]:
self.dims = len(shape) self.dims = len(shape)
...@@ -413,10 +462,10 @@ class TrtConvertElementwiseTest_two_input_without_broadcast( ...@@ -413,10 +462,10 @@ class TrtConvertElementwiseTest_two_input_without_broadcast(
weights={}, weights={},
inputs={ inputs={
"input_data1": TensorConfig( "input_data1": TensorConfig(
data_gen=partial(generate_input, shape) data_gen=partial(generate_input, shape, op_type)
), ),
"input_data2": TensorConfig( "input_data2": TensorConfig(
data_gen=partial(generate_input, shape) data_gen=partial(generate_input, shape, op_type)
), ),
}, },
outputs=["output_data"], outputs=["output_data"],
...@@ -530,7 +579,13 @@ class TrtConvertElementwiseTest_two_input_with_broadcast(TrtLayerAutoScanTest): ...@@ -530,7 +579,13 @@ class TrtConvertElementwiseTest_two_input_with_broadcast(TrtLayerAutoScanTest):
return True return True
def sample_program_configs(self): def sample_program_configs(self):
def generate_input(shape): def generate_input(shape, op_type):
# elementwise_floordiv is integer only
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=shape, dtype=np.int32
)
else:
return np.random.random(shape).astype(np.float32) return np.random.random(shape).astype(np.float32)
input1_shape_list = [[4, 32], [2, 4, 32], [4, 2, 4, 32]] input1_shape_list = [[4, 32], [2, 4, 32], [4, 2, 4, 32]]
...@@ -575,6 +630,7 @@ class TrtConvertElementwiseTest_two_input_with_broadcast(TrtLayerAutoScanTest): ...@@ -575,6 +630,7 @@ class TrtConvertElementwiseTest_two_input_with_broadcast(TrtLayerAutoScanTest):
"elementwise_pow", "elementwise_pow",
"elementwise_min", "elementwise_min",
"elementwise_max", "elementwise_max",
"elementwise_floordiv",
]: ]:
for axis in axis_list[j][i]: for axis in axis_list[j][i]:
self.shape1 = input1_shape self.shape1 = input1_shape
...@@ -599,12 +655,12 @@ class TrtConvertElementwiseTest_two_input_with_broadcast(TrtLayerAutoScanTest): ...@@ -599,12 +655,12 @@ class TrtConvertElementwiseTest_two_input_with_broadcast(TrtLayerAutoScanTest):
inputs={ inputs={
"input_data1": TensorConfig( "input_data1": TensorConfig(
data_gen=partial( data_gen=partial(
generate_input, input1_shape generate_input, input1_shape, op_type
) )
), ),
"input_data2": TensorConfig( "input_data2": TensorConfig(
data_gen=partial( data_gen=partial(
generate_input, input2_shape generate_input, input2_shape, op_type
) )
), ),
}, },
...@@ -676,11 +732,22 @@ class TrtConvertElementwiseTest_one_input_corner_case(TrtLayerAutoScanTest): ...@@ -676,11 +732,22 @@ class TrtConvertElementwiseTest_one_input_corner_case(TrtLayerAutoScanTest):
return True return True
def sample_program_configs(self): def sample_program_configs(self):
def generate_input(shape): def generate_input(shape, op_type):
# elementwise_floordiv is integer only
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=shape, dtype=np.int32
)
else:
return np.random.random(shape).astype(np.float32) return np.random.random(shape).astype(np.float32)
# use rand not randn to avoiding pow producing `NAN` # use rand not randn to avoiding pow producing `NAN`
def generate_weight(): def generate_weight(op_type):
if op_type == "elementwise_floordiv":
return np.random.randint(
low=1, high=10000, size=[32], dtype=np.int32
)
else:
return np.random.rand(32).astype(np.float32) return np.random.rand(32).astype(np.float32)
for batch in [1, 2, 4]: for batch in [1, 2, 4]:
...@@ -698,6 +765,7 @@ class TrtConvertElementwiseTest_one_input_corner_case(TrtLayerAutoScanTest): ...@@ -698,6 +765,7 @@ class TrtConvertElementwiseTest_one_input_corner_case(TrtLayerAutoScanTest):
"elementwise_pow", "elementwise_pow",
"elementwise_min", "elementwise_min",
"elementwise_max", "elementwise_max",
"elementwise_floordiv",
]: ]:
self.op_type = op_type self.op_type = op_type
for axis in [-1 if len(shape) == 1 else 1]: for axis in [-1 if len(shape) == 1 else 1]:
...@@ -720,12 +788,14 @@ class TrtConvertElementwiseTest_one_input_corner_case(TrtLayerAutoScanTest): ...@@ -720,12 +788,14 @@ class TrtConvertElementwiseTest_one_input_corner_case(TrtLayerAutoScanTest):
ops=ops, ops=ops,
weights={ weights={
"weight": TensorConfig( "weight": TensorConfig(
data_gen=partial(generate_weight) data_gen=partial(generate_weight, op_type)
) )
}, },
inputs={ inputs={
"input_data": TensorConfig( "input_data": TensorConfig(
data_gen=partial(generate_input, shape) data_gen=partial(
generate_input, shape, op_type
)
), ),
}, },
outputs=["output_data"], outputs=["output_data"],
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或