[inference][trt] add reduce_all and reduce_any (#53088)

paddle/fluid/inference/api/analysis_predictor.cc

@@ -2615,6 +2615,8 @@ USE_TRT_CONVERTER(reduce_max);
 USE_TRT_CONVERTER(reduce_min);
 USE_TRT_CONVERTER(reduce_sum);
 USE_TRT_CONVERTER(reduce_prod);
+USE_TRT_CONVERTER(reduce_any);
+USE_TRT_CONVERTER(reduce_all);
 USE_TRT_CONVERTER(tile);
 USE_TRT_CONVERTER(conv3d);
 USE_TRT_CONVERTER(conv3d_transpose);

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

@@ -95,6 +95,8 @@ const std::unordered_map<std::string, std::vector<nvinfer1::ReduceOperation>>
         {"reduce_max", {nvinfer1::ReduceOperation::kMAX}},
         {"reduce_min", {nvinfer1::ReduceOperation::kMIN}},
         {"reduce_prod", {nvinfer1::ReduceOperation::kPROD}},
+        {"reduce_any", {nvinfer1::ReduceOperation::kMAX}},
+        {"reduce_all", {nvinfer1::ReduceOperation::kMIN}},
 };
 
 class ReduceSumOpConverter : public ReduceOpConverter {
@@ -122,6 +124,80 @@ class ReduceProdOpConverter : public ReduceOpConverter {
   ReduceProdOpConverter() { op_type = "reduce_prod"; }
 };
 
+class ReduceAnyOpConverter : public ReduceOpConverter {
+ public:
+  ReduceAnyOpConverter() { op_type = "reduce_any"; }
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope,
+                  bool test_mode) override {
+    VLOG(4) << "convert a paddle " << op_type << " op to tensorrt reduce layer";
+    framework::OpDesc op_desc(op, nullptr);
+    auto reduce_type = ops_.find(op_type);
+    auto* x = engine_->GetITensor(op_desc.Input("X").front());
+    // Cast the DataType to float
+    nvinfer1::IReduceLayer* reduce_layer = nullptr;
+    auto* cast_layer = TRT_ENGINE_ADD_LAYER(engine_, Identity, *x);
+    cast_layer->setOutputType(0, nvinfer1::DataType::kINT32);
+    cast_layer->getOutput(0)->setType(nvinfer1::DataType::kINT32);
+
+    nvinfer1::Dims input_shape = x->getDimensions();
+    int input_dims = input_shape.nbDims;
+    // Discriminate DataType between int and bool.
+    bool keep_dim = PADDLE_GET_CONST(bool, op_desc.GetAttr("keep_dim"));
+    std::vector<int32_t> dim =
+        PADDLE_GET_CONST(std::vector<int32_t>, op_desc.GetAttr("dim"));
+    bool reduce_all = PADDLE_GET_CONST(bool, op_desc.GetAttr("reduce_all"));
+
+    if (reduce_all) {
+      uint32_t reduce_dim = 0;
+      for (int i = 0; i < input_dims; ++i) {
+        reduce_dim |= 1 << i;
+      }
+      reduce_layer = TRT_ENGINE_ADD_LAYER(engine_,
+                                          Reduce,
+                                          *cast_layer->getOutput(0),
+                                          reduce_type->second.front(),
+                                          reduce_dim,
+                                          keep_dim);
+    } else {
+      auto CvtToBitMask = [&](const std::vector<int32_t>& dims) -> uint32_t {
+        uint32_t res = 0;
+        for (auto x : dims) {
+          if (x < 0) {
+            res |= 1 << (x + input_dims);
+          } else {
+            if (!engine_->with_dynamic_shape()) x = x - 1;
+            res |= 1 << x;
+          }
+        }
+        return res;
+      };
+      reduce_layer = TRT_ENGINE_ADD_LAYER(engine_,
+                                          Reduce,
+                                          *cast_layer->getOutput(0),
+                                          reduce_type->second.front(),
+                                          CvtToBitMask(dim),
+                                          keep_dim);
+    }
+
+    auto output_name = op_desc.Output("Out")[0];
+
+    auto* layer =
+        TRT_ENGINE_ADD_LAYER(engine_, Identity, *reduce_layer->getOutput(0));
+    layer->setOutputType(0, nvinfer1::DataType::kBOOL);
+    layer->getOutput(0)->setType(nvinfer1::DataType::kBOOL);
+    // Ensure that the output type and input type are consistent.
+    layer->getOutput(0)->setType(cast_layer->getInput(0)->getType());
+
+    RreplenishLayerAndOutput(layer, op_type, {output_name}, test_mode);
+  };
+};
+
+class ReduceAllOpConverter : public ReduceAnyOpConverter {
+ public:
+  ReduceAllOpConverter() { op_type = "reduce_all"; }
+};
+
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle
@@ -131,3 +207,5 @@ REGISTER_TRT_OP_CONVERTER(reduce_mean, ReduceMeanOpConverter);
 REGISTER_TRT_OP_CONVERTER(reduce_max, ReduceMaxOpConverter);
 REGISTER_TRT_OP_CONVERTER(reduce_min, ReduceMinOpConverter);
 REGISTER_TRT_OP_CONVERTER(reduce_prod, ReduceProdOpConverter);
+REGISTER_TRT_OP_CONVERTER(reduce_any, ReduceAnyOpConverter);
+REGISTER_TRT_OP_CONVERTER(reduce_all, ReduceAllOpConverter);

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -2193,7 +2193,8 @@ struct SimpleOpTypeSetTeller : public Teller {
 
     if (op_type == "reduce_sum" || op_type == "reduce_mean" ||
         op_type == "reduce_max" || op_type == "reduce_min" ||
-        op_type == "reduce_prod") {
+        op_type == "reduce_prod" || op_type == "reduce_any" ||
+        op_type == "reduce_all") {
       if (!desc.HasAttr("dim", /*with_attr_var=*/false)) {
         VLOG(3) << "Skip to convert into TRT while found Attribute('dim') is "
                    "Variable type in "
@@ -2234,8 +2235,21 @@ struct SimpleOpTypeSetTeller : public Teller {
           return false;
       }
 
-#if IS_TRT_VERSION_LT(7000)
       auto dtype = x_var_desc->GetDataType();
+      if (op_type == "reduce_all" || op_type == "reduce_any") {
+        if (dtype != framework::proto::VarType::BOOL) {
+          VLOG(3)
+              << "reduce_all and reduce_any op input data type must be bool";
+          return false;
+        }
+      } else {
+#if IS_TRT_VERSION_GE(7000)
+        if (dtype != framework::proto::VarType::INT32 &&
+            dtype != framework::proto::VarType::FP32) {
+          VLOG(3) << "reduce op input data type must be int32 or float32";
+          return false;
+        }
+#else
         if (dtype != framework::proto::VarType::FP32) {
           VLOG(3) << "reduce op input data type must be float32 using TensorRT "
                      "< 7.0";
@@ -2243,6 +2257,7 @@ struct SimpleOpTypeSetTeller : public Teller {
         }
 #endif
       }
+    }
 #if IS_TRT_VERSION_GE(7000)
     if (op_type == "tile") {
       // Paddle-TRT does not support the input tensors.
@@ -2804,8 +2819,12 @@ struct SimpleOpTypeSetTeller : public Teller {
       "nearest_interp",
       "anchor_generator",
       "reduce_max",
+      "reduce_min",
       "reduce_mean",
       "reduce_sum",
+      "reduce_prod",
+      "reduce_any",
+      "reduce_all",
       "conv3d",
       "conv3d_transpose",
       "mish",
@@ -2961,8 +2980,12 @@ struct SimpleOpTypeSetTeller : public Teller {
       "nearest_interp",
       "anchor_generator",
       "reduce_max",
+      "reduce_min",
       "reduce_mean",
       "reduce_sum",
+      "reduce_prod",
+      "reduce_any",
+      "reduce_all",
       "conv3d",
       "conv3d_transpose",
       "mish",

test/ir/inference/test_trt_convert_reduce.py

@@ -51,6 +51,8 @@ class TrtConvertReduceTest(TrtLayerAutoScanTest):
                 return np.random.random([1, 3, 64, 64]).astype(np.float32)
             elif dtype == 2:
                 return np.random.random([1, 3, 64, 64]).astype(np.int32)
+            elif dtype == 0:
+                return np.random.random([1, 3, 64, 64]).astype(np.bool_)
 
         for keep_dim in [True, False]:
             for dim in [
@@ -65,25 +67,24 @@ class TrtConvertReduceTest(TrtLayerAutoScanTest):
                 [3, 4, 5],
             ]:
                 for reduce_all in [True, False]:
-                    for out_dtype in [-1, 2, 5]:
-                        for op_type in [
+                    for out_dtype in [-1, 0, 2, 5]:
+                        if out_dtype != 0:
+                            reduce_type_list = [
                                 "reduce_max",
                                 "reduce_min",
                                 "reduce_mean",
                                 "reduce_sum",
                                 "reduce_prod",
-                        ]:
-                            dics1 = [
-                                {
-                                    "keep_dim": keep_dim,
-                                    "dim": dim,
-                                    "reduce_all": reduce_all,
-                                    "out_dtype": out_dtype,
-                                    "in_dtype": out_dtype,
-                                },
-                                {},
                             ]
-                            dics2 = [
+                        else:
+                            reduce_type_list = [
+                                "reduce_all",
+                                "reduce_any",
+                            ]
+
+                        for op_type in reduce_type_list:
+
+                            dics = [
                                 {
                                     "keep_dim": keep_dim,
                                     "dim": dim,
@@ -93,7 +94,7 @@ class TrtConvertReduceTest(TrtLayerAutoScanTest):
                                 },
                                 {},
                             ]
-                            for dics in [dics1, dics2]:
+
                             ops_config = [
                                 {
                                     "op_type": op_type,
@@ -104,6 +105,10 @@ class TrtConvertReduceTest(TrtLayerAutoScanTest):
                                     "op_attrs": dics[0],
                                 }
                             ]
+                            if op_type in ["reduce_any", "reduce_all"]:
+                                ops_config[0]["outputs_dtype"] = {
+                                    "reduce_output_data": np.bool_
+                                }
                             ops = self.generate_op_config(ops_config)
 
                             program_config = ProgramConfig(