[Paddle TensorRT No.9-10] Add pd_op.(argmin,argsort) converter (#69261)

ooooo-create · web-flow · commit 5c4e4b48035f · 2024-11-15T10:40:54.000+08:00
* [Paddle TensorRT] No.8,9 Add pd_op.(argmin,argsort) converter

* fix

* Fix typos

* Apply review

* Fix typos

* Fix Typos

* Add more tests to pass coverage ci

* Fix tests

* Fix tests
diff --git a/paddle/fluid/pir/transforms/tensorrt/trt_op_marker_pass.cc b/paddle/fluid/pir/transforms/tensorrt/trt_op_marker_pass.cc
@@ -1216,12 +1216,13 @@ class ArgmaxOpPattern
                  "data in arg_max.";
       return false;
     }
-    auto x = op.x();
-    auto x_tensor_type = x.type().dyn_cast<paddle::dialect::DenseTensorType>();
-    auto data_type = paddle::dialect::TransToPhiDataType(x_tensor_type.dtype());
-    if (!(data_type == phi::DataType::FLOAT32 ||
-          data_type == phi::DataType::FLOAT16 ||
-          data_type == phi::DataType::FLOAT64)) {
+    pir::Value x = op.x();
+    auto data_type = pir::GetDataTypeFromValue(x);
+    if (!(data_type.isa<pir::Float32Type>() ||
+          data_type.isa<pir::Float16Type>() ||
+          data_type.isa<pir::Float64Type>())) {
+      VLOG(3) << "At present, pd_op.argmax only support float32 or float16 or "
+                 "float64 into trt.";
       return false;
     }
     int axis = static_cast<int>(op.axis()
@@ -1233,13 +1234,93 @@ class ArgmaxOpPattern
     phi::DataType dtype =
         op.attribute<paddle::dialect::DataTypeAttribute>("dtype").data();
     if (axis == 0 || flatten ||
-        (dtype != phi::DataType::INT32 && dtype != phi::DataType::INT64))
+        (dtype != phi::DataType::INT32 && dtype != phi::DataType::INT64)) {
+      VLOG(3) << "Skipping TRT conversion in pd_op.argmax: axis is zero, "
+                 "flatten is True, or dtype isn't int32/int64";
       return false;
+    }
     op->set_attribute(kCanRunTrtAttr, rewriter.bool_attr(true));
     return true;
   }
 };
 
+class ArgminOpPattern
+    : public pir::OpRewritePattern<paddle::dialect::ArgminOp> {
+ public:
+  using pir::OpRewritePattern<paddle::dialect::ArgminOp>::OpRewritePattern;
+  bool MatchAndRewrite(paddle::dialect::ArgminOp op,
+                       pir::PatternRewriter &rewriter) const override {
+    if (op->HasAttribute(kCanRunTrtAttr) &&
+        op.attribute<pir::BoolAttribute>(kCanRunTrtAttr).data()) {
+      return false;
+    }
+    if (!op.axis().defining_op()->isa<paddle::dialect::FullOp>()) {
+      VLOG(3) << "Skip to convert into TRT while found axis is not a constant "
+                 "data in arg_mix.";
+      return false;
+    }
+    pir::Value x = op.x();
+    auto data_type = pir::GetDataTypeFromValue(x);
+    if (!(data_type.isa<pir::Float32Type>() ||
+          data_type.isa<pir::Float16Type>() ||
+          data_type.isa<pir::Float64Type>())) {
+      VLOG(3) << "At present, pd_op.argmin only support float32 or float16 or "
+                 "float64 into trt.";
+      return false;
+    }
+    int axis = static_cast<int>(op.axis()
+                                    .defining_op()
+                                    ->attribute<pir::DoubleAttribute>("value")
+                                    .data());
+
+    bool flatten = op.attribute<pir::BoolAttribute>("flatten").data();
+    phi::DataType dtype =
+        op.attribute<paddle::dialect::DataTypeAttribute>("dtype").data();
+    if (axis == 0 || flatten ||
+        (dtype != phi::DataType::INT32 && dtype != phi::DataType::INT64)) {
+      VLOG(3) << "Skipping TRT conversion in pd_op.argmin: axis is zero, "
+                 "flatten is True, or dtype isn't int32/int64";
+      return false;
+    }
+
+    op->set_attribute(kCanRunTrtAttr, rewriter.bool_attr(true));
+    return true;
+  }
+};
+
+class ArgsortOpPattern
+    : public pir::OpRewritePattern<paddle::dialect::ArgsortOp> {
+ public:
+  using pir::OpRewritePattern<paddle::dialect::ArgsortOp>::OpRewritePattern;
+  bool MatchAndRewrite(paddle::dialect::ArgsortOp op,
+                       pir::PatternRewriter &rewriter) const override {
+    if (op->HasAttribute(kCanRunTrtAttr) &&
+        op.attribute<pir::BoolAttribute>(kCanRunTrtAttr).data()) {
+      return false;
+    }
+    const std::vector<std::string> required_attrs = {"axis", "descending"};
+    for (const auto &attr : required_attrs) {
+      if (!op->HasAttribute(attr)) {
+        VLOG(3) << "pd_op.argsort " << attr << " attribute does not exist";
+        return false;
+      }
+    }
+    pir::Value x = op.x();
+    auto x_type = x.type().dyn_cast<paddle::dialect::DenseTensorType>();
+    auto x_shape = x_type.dims();
+    int axis = op->attribute<pir::Int32Attribute>("axis").data();
+    if (axis < 0) {
+      axis += x_shape.size();
+    }
+    if (x_shape[axis] > 3840) {
+      VLOG(3)
+          << "In pd_op.argsort,the axis dim of input should be less than 3840";
+      return false;
+    }
+    op->set_attribute(kCanRunTrtAttr, rewriter.bool_attr(true));
+    return true;
+  }
+};
 class BilinearInterpV2Pattern
     : public pir::OpRewritePattern<paddle::dialect::BilinearInterpOp> {
  public:
@@ -1809,6 +1890,8 @@ class TrtOpMarkerPass : public pir::PatternRewritePass {
     ps.Add(std::make_unique<RemainderOpPattern>(context));
     ps.Add(std::make_unique<MulticlassNms3OpPattern>(context));
     ps.Add(std::make_unique<ArgmaxOpPattern>(context));
+    ps.Add(std::make_unique<ArgminOpPattern>(context));
+    ps.Add(std::make_unique<ArgsortOpPattern>(context));
     ps.Add(std::make_unique<MaxOpPattern>(context));
     ps.Add(std::make_unique<MinOpPattern>(context));
     ps.Add(std::make_unique<AllOpPattern>(context));
diff --git a/python/paddle/tensorrt/impls/search.py b/python/paddle/tensorrt/impls/search.py
@@ -16,8 +16,11 @@
 import tensorrt as trt
 
 from paddle.tensorrt.converter_utils import (
+    get_shape_tensor_element,
     squeeze_trt,
     trt_cast,
+    trt_reshape,
+    trt_shape,
     unsqueeze_trt,
 )
 from paddle.tensorrt.register import converter_registry
@@ -66,6 +69,77 @@ def argmax_converter(network, paddle_op, inputs):
         return squeeze_layer.get_output(0)
 
 
+@converter_registry.register("pd_op.argmin", trt_version="8.x")
+def argmin_converter(network, paddle_op, inputs):
+    x = inputs[0]
+    input_dims = x.shape
+    rank = len(input_dims)
+    axis = int(
+        paddle_op.operands()[1]
+        .source()
+        .get_defining_op()
+        .attrs()
+        .get("value", -1)
+    )
+    keepdims = paddle_op.attrs()["keepdims"]
+
+    if axis < 0:
+        axis += rank
+
+    topk_layer = network.add_topk(
+        input=x, op=trt.TopKOperation.MIN, k=1, axes=(1 << axis)
+    )
+
+    if keepdims:
+        return topk_layer.get_output(1)
+    else:
+        squeeze_layer = network.add_shuffle(topk_layer.get_output(1))
+        output_dims = []
+        for i in range(len(input_dims)):
+            if i == axis:
+                continue
+            output_dims.append(input_dims[i])
+        squeeze_layer.reshape_dims = tuple(output_dims)
+        return squeeze_layer.get_output(0)
+
+
+@converter_registry.register("pd_op.argsort", trt_version="8.x")
+def argsort_converter(network, paddle_op, inputs):
+    input_tensor = inputs[0]
+    input_shape = input_tensor.shape
+    in_type = input_tensor.dtype
+    in_rank = len(input_shape)
+    axis = paddle_op.attrs()["axis"]
+    descending = paddle_op.attrs()["descending"]
+    if axis < 0:
+        axis += len(input_shape)
+    topk_op = trt.TopKOperation.MAX if descending else trt.TopKOperation.MIN
+    need_cast = True if in_type != trt.DataType.FLOAT else False
+    if in_rank == 1:
+        unsqueeze_shape = trt.Dims([1, -1])
+        input_tensor = trt_reshape(
+            network, input_tensor, unsqueeze_shape, is_shape_tensor=False
+        )
+        axis = 1
+    if need_cast:
+        input_tensor = trt_cast(network, input_tensor, trt.DataType.FLOAT)
+    topk_layer = network.add_topk(input_tensor, topk_op, 1, 1 << axis)
+    shape = trt_shape(network, input_tensor)
+    k_tensor = get_shape_tensor_element(network, shape, axis, True)
+    topk_layer.set_input(1, k_tensor)
+    out = topk_layer.get_output(0)
+    indices = topk_layer.get_output(1)
+    if in_rank == 1:
+        squeeze_shape = trt.Dims([-1])
+        out = trt_reshape(network, out, squeeze_shape, is_shape_tensor=False)
+        indices = trt_reshape(
+            network, indices, squeeze_shape, is_shape_tensor=False
+        )
+    out_tensor = trt_cast(network, out, in_type)
+    indices_tensor = trt_cast(network, indices, indices.dtype)
+    return out_tensor, indices_tensor
+
+
 @converter_registry.register("pd_op.where", trt_version="8.x")
 def where_converter(network, paddle_op, inputs):
     condition = inputs[0]
diff --git a/test/tensorrt/test_converter_search.py b/test/tensorrt/test_converter_search.py
