【SCU】【Paddle TensorRT No.48】Add pd_op.pad converter (#69673)

PolaKuma · web-flow · commit cd97c6d166b5 · 2025-02-24T16:28:59.000+08:00
* add_pad_op

* fix codestyle

* fix codestyle

* update

* Update trt_op_marker_pass.cc

* update

* update

* Update trt_op_marker_pass.cc

* fix codestyle

* fix codestyle

* fix opt

* Update manipulation.py

* Update trt_op_marker_pass.cc

* Update trt_op_marker_pass.cc

* update

* Update test_converter_manipulation.py

* Update test_converter_manipulation.py

* fix

* Update trt_op_marker_pass.cc

* fix codestyle

* fix codestyle

* Update trt_op_marker_pass.cc
diff --git a/paddle/fluid/pir/transforms/tensorrt/trt_op_marker_pass.cc b/paddle/fluid/pir/transforms/tensorrt/trt_op_marker_pass.cc
@@ -1065,6 +1065,7 @@ class FlattenOpPattern
     return true;
   }
 };
+
 class CastOpPattern : public pir::OpRewritePattern<paddle::dialect::CastOp> {
  public:
   using pir::OpRewritePattern<paddle::dialect::CastOp>::OpRewritePattern;
@@ -1092,6 +1093,63 @@ class CastOpPattern : public pir::OpRewritePattern<paddle::dialect::CastOp> {
   }
 };
 
+class PadOpPattern : public pir::OpRewritePattern<paddle::dialect::PadOp> {
+ public:
+  using pir::OpRewritePattern<paddle::dialect::PadOp>::OpRewritePattern;
+
+  bool MatchAndRewrite(paddle::dialect::PadOp op,
+                       pir::PatternRewriter &rewriter) const override {
+    if (op->HasAttribute(kCanRunTrtAttr) &&
+        op->attribute<pir::BoolAttribute>(kCanRunTrtAttr).data()) {
+      return false;
+    }
+    pir::Value pad_value_tensor = op.operand_source(1);
+    if (!op->HasAttribute("paddings") || !pad_value_tensor) {
+      VLOG(3) << "PadOp must has 'paddings' and 'pad_value'.";
+      return false;
+    }
+    if (pir::GetDefiningOpForInput(op, 1)->isa<paddle::dialect::FullOp>()) {
+      paddle::dialect::FullOp full_op =
+          pir::GetDefiningOpForInput(op, 1)
+              ->dyn_cast<paddle::dialect::FullOp>();
+      auto pad_value =
+          full_op->attribute<paddle::dialect::ScalarAttribute>("value")
+              .data()
+              .to<float>();
+      if (pad_value != 0.0f) {
+        VLOG(3) << "The pad layer of TRT only support zero.";
+        return false;
+      }
+    }
+    auto paddings_attr = op->attribute<pir::ArrayAttribute>("paddings");
+    std::vector<int> paddings;
+    for (const auto &attr : paddings_attr.AsVector()) {
+      paddings.push_back(attr.dyn_cast<pir::Int32Attribute>().data());
+    }
+    int pad_size = paddings.size();
+    pir::Value x = op.operand_source(0);
+    auto x_type = x.type().dyn_cast<paddle::dialect::DenseTensorType>();
+    auto input_shape = x_type.dims();
+    int nbDims = input_shape.size();
+    if (nbDims < 2) {
+      VLOG(3) << "Input must have at least 2 dimensions.";
+      return false;
+    }
+    if (nbDims * 2 != pad_size) {
+      VLOG(3) << "Padding size must be twice the number of input dimensions.";
+      return false;
+    }
+    for (int i = 0; i < pad_size - 4; i++) {
+      if (paddings[i] != 0) {
+        VLOG(3) << "Only the last two dimensions can have non-zero paddings.";
+        return false;
+      }
+    }
+    op->set_attribute(kCanRunTrtAttr, rewriter.bool_attr(true));
+    return true;
+  }
+};
+
 class SplitOpPattern : public pir::OpRewritePattern<paddle::dialect::SplitOp> {
  public:
   using pir::OpRewritePattern<paddle::dialect::SplitOp>::OpRewritePattern;
@@ -2914,6 +2972,7 @@ class TrtOpMarkerPass : public pir::PatternRewritePass {
     ps.Add(std::make_unique<LogicalXorOpPattern>(context));
     ps.Add(std::make_unique<CeluOpPattern>(context));
     ps.Add(std::make_unique<OneHotOpPattern>(context));
+    ps.Add(std::make_unique<PadOpPattern>(context));
     ps.Add(std::make_unique<TemporalShiftOpPattern>(context));
     ps.Add(std::make_unique<IndexPutOpPattern>(context));
     ps.Add(std::make_unique<InstanceNormOpPattern>(context));
diff --git a/python/paddle/tensorrt/impls/manipulation.py b/python/paddle/tensorrt/impls/manipulation.py
@@ -961,6 +961,17 @@ def roll_converter(network, paddle_op, inputs):
     return layer.get_output(0)
 
 
+@converter_registry.register("pd_op.pad", trt_version="8.x")
+def pad_converter(network, paddle_op, inputs):
+    input_tensor = inputs[0]
+    paddings = paddle_op.attrs()["paddings"]
+    pad_size = len(paddings)
+    pre_pad = [paddings[pad_size - 4], paddings[pad_size - 2]]
+    post_pad = [paddings[pad_size - 3], paddings[pad_size - 1]]
+    layer = network.add_padding_nd(input_tensor, pre_pad, post_pad)
+    return layer.get_output(0)
+
+
 @converter_registry.register("pd_op.pad3d", trt_version="8.x")
 def pad3d_converter(network, paddle_op, inputs):
     input_tensor, paddings = inputs
diff --git a/test/tensorrt/test_converter_manipulation.py b/test/tensorrt/test_converter_manipulation.py
@@ -743,6 +743,90 @@ def test_trt_result(self):
         self.check_trt_result()
 
 
+def wrapper_pad_error(x, padding, mode, pad_value):
+    return paddle.nn.functional.pad(
+        x=paddle.to_tensor(np.random.randn(1, 1, 1, 2, 3).astype("float32")),
+        pad=[0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
+        mode='constant',
+        value=0,
+    )
+
+
+class TestPadCaseTRTPattern(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = paddle.nn.functional.pad
+        self.api_args = {
+            "x": np.random.randn(1, 1, 1, 2, 3).astype("float32"),
+            "paddings": [0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
+            "mode": "constant",
+            "pad_value": np.array([0], dtype="float32"),
+        }
+        self.program_config = {"feed_list": ["x", "pad_value"]}
+        self.min_shape = {"x": [1, 1, 1, 2, 3]}
+        self.opt_shape = {"x": [1, 1, 1, 2, 3]}
+        self.max_shape = {"x": [5, 1, 1, 2, 3]}
+
+    def test_trt_result_fp16(self):
+        self.check_trt_result(precision_mode="fp16")
+
+    def test_trt_result_fp32(self):
+        self.check_trt_result()
+
+
+class TestPadError1TRTPattern(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = paddle.nn.functional.pad
+        self.api_args = {
+            "x": np.random.randn(1, 1, 1, 2, 3).astype("float32"),
+            "paddings": [0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
+            "mode": "constant",
+            "pad_value": np.array([1], dtype="float32"),
+        }
+        self.program_config = {"feed_list": ["x", "pad_value"]}
+        self.min_shape = {"x": [1, 1, 1, 2, 3]}
+        self.opt_shape = {"x": [1, 1, 1, 2, 3]}
+        self.max_shape = {"x": [5, 1, 1, 2, 3]}
+
+    def test_trt_result(self):
+        self.check_marker(expected_result=False)
+
+
+class TestPadError2TRTPattern(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = wrapper_pad_error
+        self.api_args = {
+            "x": np.random.randn(1, 1, 1, 2, 3).astype("float32"),
+            "paddings": [1, 1, 1, 0, 0, 0, 1, 1, 0, 0],
+            "mode": "constant",
+            "pad_value": np.array([1], dtype="float32"),
+        }
+        self.program_config = {"feed_list": ["x", "pad_value"]}
+        self.min_shape = {"x": [1, 1, 1, 2, 3]}
+        self.opt_shape = {"x": [1, 1, 1, 2, 3]}
+        self.max_shape = {"x": [5, 1, 1, 2, 3]}
+
+    def test_trt_result(self):
+        self.check_marker(expected_result=False)
+
+
+class TestPadError3TRTPattern(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = wrapper_pad_error
+        self.api_args = {
+            "x": np.random.randn(1, 1).astype("float32"),
+            "paddings": [0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
+            "mode": "constant",
+            "pad_value": np.array([0], dtype="float32"),
+        }
+        self.program_config = {"feed_list": ["x", "pad_value"]}
+        self.min_shape = {"x": [1, 1, 1, 2, 3]}
+        self.opt_shape = {"x": [1, 1, 1, 2, 3]}
+        self.max_shape = {"x": [5, 1, 1, 2, 3]}
+
+    def test_trt_result(self):
+        self.check_marker(expected_result=False)
+
+
 def wrapper_pad3d(x, paddings, mode, value, data_format):
     pad3d = paddle.nn.Pad3D(
         padding=[1, 0, 1, 2, 0, 0],
