[Inference][Trt]fix bilinear_v2 (#56043)

* fix trt bilinear_interp_v2_op

* add trt 8.0 condition

* add trt 8.0 condition

test bilinear

add trt 8.0 condition

* code style

paddle/fluid/inference/tensorrt/convert/bilinear_interp_v2_op.cc

@@ -48,10 +48,21 @@ class BilinearInterpolateV2OpConverter : public OpConverter {
     auto out_w = PADDLE_GET_CONST(int, op_desc.GetAttr("out_w"));
 
     auto layer = TRT_ENGINE_ADD_LAYER(engine_, Resize, *input);
-    if (align_mode == 0 && !align_corners) {
+    if (align_mode == 0) {
       layer->setResizeMode(nvinfer1::ResizeMode::kLINEAR);
     }
-
+#if IS_TRT_VERSION_GE(8000)
+    if (align_corners == true) {
+      layer->setCoordinateTransformation(
+          nvinfer1::ResizeCoordinateTransformation::kALIGN_CORNERS);
+    } else {
+      layer->setCoordinateTransformation(
+          nvinfer1::ResizeCoordinateTransformation::kHALF_PIXEL);
+    }
+#endif
+#if !IS_TRT_VERSION_GE(8000)
+    layer->setAlignCorners(align_corners);
+#endif
     auto in_dim = input->getDimensions();
     float scale_h = -1.f;
     float scale_w = -1.f;
@@ -95,7 +106,7 @@ class BilinearInterpolateV2OpConverter : public OpConverter {
       }
     }
 
-    if (out_h > 0 && out_w > 0) {
+    if (out_h > 0 && out_w > 0 && !(scale_w > 0. && scale_h > 0.)) {
       scale_h =
           static_cast<float>(out_h) / static_cast<float>(in_dim.d[h_axis]);
       scale_w =

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -919,7 +919,6 @@ struct SimpleOpTypeSetTeller : public Teller {
           return false;
         }
       }
-
       if (resize_inputs.find("OutSize") != resize_inputs.end()) {
         if (!with_dynamic_shape) {
           VLOG(3) << "Static shape don't support the OutSize for op_type "
@@ -944,18 +943,11 @@ struct SimpleOpTypeSetTeller : public Teller {
         return false;
       }
 
-      auto align_corners =
-          PADDLE_GET_CONST(bool, desc.GetAttr("align_corners"));
-      if (align_corners != false) {
-        VLOG(3)
-            << "The bilinear_interp_v2 only supports align_corners with false.";
-        return false;
-      }
-
       bool has_scale_input_size =
           (resize_inputs.find("Scale") != resize_inputs.end());
 
-      if (has_scale_input_size && desc.Input("Scale").size() != 1) {
+      if (!has_scale_input_size ||
+          (has_scale_input_size && desc.Input("Scale").size() != 1)) {
         const std::vector<float> scale =
             PADDLE_GET_CONST(std::vector<float>, desc.GetAttr("scale"));
         if (scale.size() <= 1) {

test/ir/inference/test_trt_convert_bilinear_interp_v2.py

@@ -38,7 +38,9 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest):
 
     def sample_program_configs(self):
         def generate_input1(attrs: List[Dict[str, Any]]):
-            return np.ones([1, 3, 64, 64]).astype(np.float32)
+            return np.random.uniform(
+                low=0.0, high=1.0, size=[1, 3, 64, 64]
+            ).astype(np.float32)
 
         def generate_input2(attrs: List[Dict[str, Any]]):
             return np.random.uniform(low=0.5, high=6.0, size=(2)).astype(
@@ -46,20 +48,19 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest):
             )
 
         for data_layout in ["NCHW", "NHWC"]:
+            for align_corners in [False, True]:
                 for scale_y in [2.0, 1.0]:
                     for scale_x in [2.0]:
                         scale = [scale_y, scale_x]
-                    for out_h in [32, 128]:
-                        for out_w in [64]:
                         dics = [
                             {
                                 "data_layout": data_layout,
                                 "interp_method": "bilinear",
-                                    "align_corners": False,
+                                "align_corners": align_corners,
                                 "align_mode": 0,
                                 "scale": scale,
-                                    "out_h": out_h,
-                                    "out_w": out_w,
+                                "out_h": -1,
+                                "out_w": -1,
                             }
                         ]
 
@@ -71,9 +72,7 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest):
                                     "Scale": ["input_scale"],
                                 },
                                 "op_outputs": {
-                                        "Out": [
-                                            "bilinear_interp_v2_output_data"
-                                        ]
+                                    "Out": ["bilinear_interp_v2_output_data"]
                                 },
                                 "op_attrs": dics[0],
                             }
@@ -136,9 +135,125 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest):
         program_config.set_input_type(np.float32)
         yield self.create_inference_config(), generate_trt_nodes_num(
             attrs, True
+        ), (1e-5, 1e-5)
+        self.trt_param.precision = paddle_infer.PrecisionType.Half
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, True
+        ), 1e-2
+
+    def test(self):
+        self.run_test()
+
+
+class TrtConvertBilinearInterpV2Test1(TrtLayerAutoScanTest):
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        inputs = program_config.inputs
+        weights = program_config.weights
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+        ver = paddle_infer.get_trt_compile_version()
+        # here is consistent with op_teller.cc
+        if ver[0] * 1000 + ver[1] * 100 + ver[2] * 10 < 7100:
+            return False
+        return True
+
+    def sample_program_configs(self):
+        self.workspace_size = 1 << 32
+
+        def generate_input1(attrs: List[Dict[str, Any]]):
+            return np.random.uniform(
+                low=0.0, high=1.0, size=[1, 18, 144, 144]
+            ).astype(np.float32)
+
+        for data_layout in ["NCHW", "NHWC"]:
+            for align_corners in [False, True]:
+                for out_h in [128, 288]:
+                    for out_w in [288]:
+                        dics = [
+                            {
+                                "data_layout": data_layout,
+                                "interp_method": "bilinear",
+                                "align_corners": align_corners,
+                                "align_mode": 0,
+                                "scale": [],
+                                "out_h": out_h,
+                                "out_w": out_w,
+                            }
+                        ]
+
+                        ops_config = [
+                            {
+                                "op_type": "bilinear_interp_v2",
+                                "op_inputs": {
+                                    "X": ["input_data"],
+                                },
+                                "op_outputs": {
+                                    "Out": ["bilinear_interp_v2_output_data"]
+                                },
+                                "op_attrs": dics[0],
+                            }
+                        ]
+                        ops = self.generate_op_config(ops_config)
+
+                        program_config = ProgramConfig(
+                            ops=ops,
+                            weights={},
+                            inputs={
+                                "input_data": TensorConfig(
+                                    data_gen=partial(generate_input1, dics)
+                                )
+                            },
+                            outputs=["bilinear_interp_v2_output_data"],
+                        )
+
+                        yield program_config
+
+    def sample_predictor_configs(
+        self, program_config
+    ) -> (paddle_infer.Config, List[int], float):
+        def generate_dynamic_shape(attrs):
+            self.dynamic_shape.min_input_shape = {
+                "input_data": [1, 18, 144, 144]
+            }
+            self.dynamic_shape.max_input_shape = {
+                "input_data": [8, 18, 144, 144]
+            }
+            self.dynamic_shape.opt_input_shape = {
+                "input_data": [4, 18, 144, 144]
+            }
+
+        def clear_dynamic_shape():
+            self.dynamic_shape.min_input_shape = {}
+            self.dynamic_shape.max_input_shape = {}
+            self.dynamic_shape.opt_input_shape = {}
+
+        def generate_trt_nodes_num(attrs, dynamic_shape):
+            return 1, 2
+
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+
+        # for static_shape
+        clear_dynamic_shape()
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, False
         ), 1e-5
         self.trt_param.precision = paddle_infer.PrecisionType.Half
         program_config.set_input_type(np.float16)
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, False
+        ), 1e-2
+
+        # for dynamic_shape
+        generate_dynamic_shape(attrs)
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, True
+        ), (1e-5, 1e-5)
+        self.trt_param.precision = paddle_infer.PrecisionType.Half
         yield self.create_inference_config(), generate_trt_nodes_num(
             attrs, True
         ), 1e-2