Revert "add trt transpose and flatten converter (#31022) (#31139)" (#31343)

Pei Yang · web-flow · commit 325bfc374a0f · 2021-03-02T10:19:48.000+08:00
This reverts commit 20e68a2.
diff --git a/paddle/fluid/inference/analysis/ir_pass_manager.cc b/paddle/fluid/inference/analysis/ir_pass_manager.cc
@@ -141,10 +141,6 @@ void IRPassManager::CreatePasses(Argument *argument,
       pass->Set("optim_input_shape",
                 new std::map<std::string, std::vector<int>>(
                     argument->optim_input_shape()));
-      bool with_dynamic_shape = argument->max_input_shape().size() > 0 &&
-                                argument->min_input_shape().size() > 0 &&
-                                argument->optim_input_shape().size() > 0;
-      pass->Set("with_dynamic_shape", new bool(with_dynamic_shape));
       pass->Set("trt_disabled_ops", new std::vector<std::string>(
                                         argument->tensorrt_disabled_ops()));
       pass->Set("trt_use_dla", new bool(argument->tensorrt_use_dla()));
diff --git a/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc b/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
@@ -40,7 +40,6 @@ void analysis::TensorRtSubgraphPass::ApplyImpl(
   auto use_calib_mode = Get<bool>("use_calib_mode");
   bool no_calib_int8 = enable_int8 && !(use_calib_mode);
   auto trt_disabled_ops = Get<std::vector<std::string>>("trt_disabled_ops");
-  auto with_dynamic_shape = Get<bool>("with_dynamic_shape");
   auto teller = [&](const framework::ir::Node *node) {
     if (!node->IsOp() || !node->Op()) return false;
     if (find(trt_disabled_ops.begin(), trt_disabled_ops.end(),
@@ -49,8 +48,8 @@ void analysis::TensorRtSubgraphPass::ApplyImpl(
               << " is diabled by config in TensorRT";
       return false;
     }
-    return tensorrt::OpTeller::Global().Tell(node, no_calib_int8,
-                                             with_dynamic_shape);
+    return tensorrt::OpTeller::Global().Tell(node->Op()->Type(), *node->Op(),
+                                             no_calib_int8);
   };
 
   framework::ir::SubGraphFuser fuser(
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -1158,8 +1158,6 @@ USE_TRT_CONVERTER(elementwise_mul_tensor);
 USE_TRT_CONVERTER(elementwise_max_tensor);
 USE_TRT_CONVERTER(elementwise_min_tensor);
 USE_TRT_CONVERTER(elementwise_pow_tensor);
-USE_TRT_CONVERTER(transpose);
-USE_TRT_CONVERTER(flatten);
 USE_TRT_CONVERTER(matmul);
 USE_TRT_CONVERTER(conv2d);
 USE_TRT_CONVERTER(relu);
diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
@@ -3,9 +3,50 @@ nv_library(tensorrt_converter
            SRCS matmul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc
                 batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc group_norm_op.cc
                 pad_op.cc split_op.cc prelu_op.cc leaky_relu_op.cc gelu_op.cc layer_norm_op.cc multihead_matmul_op.cc
-                shuffle_channel_op.cc swish_op.cc instance_norm_op.cc stack_op.cc transpose_op.cc flatten_op.cc
+                shuffle_channel_op.cc swish_op.cc instance_norm_op.cc stack_op.cc
                 emb_eltwise_layernorm.cc skip_layernorm.cc scale_op.cc slice_op.cc hard_sigmoid_op.cc hard_swish_op.cc clip_op.cc
            DEPS tensorrt_engine tensorrt_plugin operator scope framework_proto op_registry)
 
 nv_test(test_op_converter SRCS test_op_converter.cc DEPS
   paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine tensorrt_converter)
+
+# TODO(xingzhaolong): fix the the following ci ut error.
+
+#nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)
+#nv_test(test_trt_mul_op SRCS test_mul_op.cc mul_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine mul_op)
+#nv_test(test_trt_fc_op SRCS test_fc_op.cc fc_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine mul_op)
+#nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine activation_op)
+#nv_test(test_trt_conv_op SRCS test_conv2d_op.cc conv2d_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine conv_op conv_transpose_op)
+#nv_test(test_trt_pool2d_op SRCS test_pool2d_op.cc pool2d_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine pool_op tensorrt_plugin)
+#nv_test(test_trt_elementwise_op SRCS test_elementwise_op.cc elementwise_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine tensorrt_plugin
+#             elementwise_add_op elementwise_mul_op)
+#nv_test(test_trt_softmax_op SRCS test_softmax_op.cc softmax_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine softmax_op)
+#nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine batch_norm_op)
+#nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine concat_op)
+#nv_test(test_trt_dropout_op SRCS test_dropout_op.cc dropout_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine dropout_op)
+#nv_test(test_trt_pad_op SRCS test_pad_op.cc pad_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine pad_op)
+#nv_test(test_trt_split_op SRCS test_split_op.cc split_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine tensorrt_plugin
+#             split_op concat_op)
+#nv_test(test_trt_prelu_op SRCS test_prelu_op.cc prelu_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine tensorrt_plugin
+#        prelu_op)
+#nv_test(test_trt_leaky_relu_op SRCS test_leaky_relu_op.cc leaky_relu_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine activation_op)
+
+#nv_test(test_shuffle_channel_op SRCS test_shuffle_channel_op.cc shuffle_channel_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine shuffle_channel_op)
+
+#nv_test(test_swish_op SRCS test_swish_op.cc swish_op.cc
+#        DEPS paddle_framework ${GLOB_OPERATOR_DEPS} tensorrt_engine activation_op tensorrt_plugin)
diff --git a/paddle/fluid/inference/tensorrt/convert/flatten_op.cc b/paddle/fluid/inference/tensorrt/convert/flatten_op.cc
diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h
@@ -109,18 +109,7 @@ class OpConverter {
           it, platform::errors::Unimplemented("no OpConverter for optype [%s]",
                                               op_desc.Type()));
     }
-    if (op_desc.Type() == "transpose2") {
-      it = Registry<OpConverter>::Global().Lookup("transpose");
-      PADDLE_ENFORCE_NOT_NULL(
-          it, platform::errors::Unimplemented("no OpConverter for optype [%s]",
-                                              op_desc.Type()));
-    }
-    if (op_desc.Type() == "flatten2") {
-      it = Registry<OpConverter>::Global().Lookup("flatten");
-      PADDLE_ENFORCE_NOT_NULL(
-          it, platform::errors::Unimplemented("no OpConverter for optype [%s]",
-                                              op_desc.Type()));
-    }
+
     if (!it) {
       it = Registry<OpConverter>::Global().Lookup(op_desc.Type());
     }
diff --git a/paddle/fluid/inference/tensorrt/convert/transpose_op.cc b/paddle/fluid/inference/tensorrt/convert/transpose_op.cc
diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc
@@ -106,17 +106,11 @@ struct SimpleOpTypeSetTeller : public Teller {
       "layer_norm",
       "scale",
       "stack",
-      "transpose2",
-      "transpose",
-      "flatten2",
-      "flatten",
   };
 };
 
-bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
-                    bool with_dynamic_shape) {
-  const std::string op_type = node->Op()->Type();
-  const framework::OpDesc desc = *node->Op();
+bool OpTeller::Tell(const std::string& op_type, const framework::OpDesc& desc,
+                    bool use_no_calib_int8) {
   // do not support the op which is labeled the `skip_quant`
   if ((desc.HasAttr("namescope") &&
        BOOST_GET_CONST(std::string, desc.GetAttr("op_namescope")) ==
@@ -168,26 +162,6 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
         if (axis <= 0) return false;
       }
     }
-    if (op_type == "transpose2" || op_type == "transpose") {
-      if (!desc.HasAttr("axis")) {
-        return false;
-      } else {
-        std::vector<int> axis =
-            BOOST_GET_CONST(std::vector<int>, desc.GetAttr("axis"));
-        if (!with_dynamic_shape && axis[0] != 0) return false;
-        if (axis.size() >= nvinfer1::Dims::MAX_DIMS) return false;
-      }
-    }
-    if (op_type == "flatten2" || op_type == "flatten") {
-      // flatten doesn't support dynamic shape currently
-      if (!desc.HasAttr("axis")) {
-        return false;
-      } else {
-        if (with_dynamic_shape) return false;
-        int axis = BOOST_GET_CONST(int, desc.GetAttr("axis"));
-        if (axis != 1) return false;
-      }
-    }
     if ((*teller)(op_type, desc, use_no_calib_int8)) return true;
   }
   return false;
diff --git a/paddle/fluid/inference/tensorrt/op_teller.h b/paddle/fluid/inference/tensorrt/op_teller.h
@@ -17,7 +17,7 @@
 #include <string>
 #include <unordered_set>
 #include <vector>
-#include "paddle/fluid/framework/ir/node.h"
+
 #include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
 
@@ -65,8 +65,8 @@ class OpTeller {
     return *x;
   }
 
-  bool Tell(const framework::ir::Node* node, bool use_no_calib_int8 = false,
-            bool with_dynamic_shape = false);
+  bool Tell(const std::string& op_type, const framework::OpDesc& desc,
+            bool use_no_calib_int8 = false);
 
  private:
   OpTeller();
diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_subgraph_pass.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_subgraph_pass.py
@@ -287,59 +287,6 @@ def test_check_output(self):
                 PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
 
 
-class TensorRTSubgraphPassTransposeTest(InferencePassTest):
-    def setUp(self):
-        with fluid.program_guard(self.main_program, self.startup_program):
-            data = fluid.data(
-                name="data", shape=[-1, 6, 64, 64], dtype="float32")
-            transpose_out = self.append_transpose(data)
-            out = fluid.layers.batch_norm(transpose_out, is_test=True)
-        self.feeds = {
-            "data": np.random.random([1, 6, 64, 64]).astype("float32"),
-        }
-        self.enable_trt = True
-        self.trt_parameters = TensorRTSubgraphPassTransposeTest.TensorRTParam(
-            1 << 30, 32, 0, AnalysisConfig.Precision.Float32, False, False)
-        self.fetch_list = [out]
-
-    def append_transpose(self, data):
-        return fluid.layers.transpose(data, [0, 3, 1, 2])
-
-    def test_check_output(self):
-        if core.is_compiled_with_cuda():
-            use_gpu = True
-            self.check_output_with_option(use_gpu)
-            self.assertTrue(
-                PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
-
-
-class TensorRTSubgraphPassFlattenTest(InferencePassTest):
-    def setUp(self):
-        with fluid.program_guard(self.main_program, self.startup_program):
-            data = fluid.data(
-                name="data", shape=[-1, 6, 64, 64], dtype="float32")
-            flatten_out = self.append_flatten(data)
-            reshape_out = fluid.layers.reshape(flatten_out, [-1, 0, 1, 1])
-            out = fluid.layers.batch_norm(reshape_out, is_test=True)
-        self.feeds = {
-            "data": np.random.random([1, 6, 64, 64]).astype("float32"),
-        }
-        self.enable_trt = True
-        self.trt_parameters = TensorRTSubgraphPassFlattenTest.TensorRTParam(
-            1 << 30, 32, 0, AnalysisConfig.Precision.Float32, False, False)
-        self.fetch_list = [out]
-
-    def append_flatten(self, data):
-        return fluid.layers.flatten(data, axis=1)
-
-    def test_check_output(self):
-        if core.is_compiled_with_cuda():
-            use_gpu = True
-            self.check_output_with_option(use_gpu)
-            self.assertTrue(
-                PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
-
-
 class TensorRTSubgraphPassLayerNormTest(InferencePassTest):
     def setUp(self):
         self.set_params()
diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_transpose_flatten_concat_fuse_pass.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_transpose_flatten_concat_fuse_pass.py
