【SCU】【Paddle TensorRT No.57】Add pd_op.temporal_shift converter (#69848)

PolaKuma · web-flow · commit 15259f9b8dde · 2025-01-14T14:42:44.000+08:00
* add

* fix codestyle

* update

* Update trt_op_marker_pass.cc

* add_fp16

* Update trt_op_marker_pass.cc

* update

* Update trt_op_marker_pass.cc

* fix

* fix

* fix

* fix codestyle

* add_test

* add_optshape

* update

* Update test_converter_others.py

* delete size
diff --git a/paddle/fluid/pir/transforms/tensorrt/trt_op_marker_pass.cc b/paddle/fluid/pir/transforms/tensorrt/trt_op_marker_pass.cc
@@ -2181,6 +2181,27 @@ class OneHotOpPattern
   }
 };
 
+class TemporalShiftOpPattern
+    : public pir::OpRewritePattern<paddle::dialect::TemporalShiftOp> {
+ public:
+  using pir::OpRewritePattern<
+      paddle::dialect::TemporalShiftOp>::OpRewritePattern;
+
+  bool MatchAndRewrite(paddle::dialect::TemporalShiftOp op,
+                       pir::PatternRewriter &rewriter) const override {
+    if (op->HasAttribute(kCanRunTrtAttr) &&
+        op.attribute<pir::BoolAttribute>(kCanRunTrtAttr).data()) {
+      return false;
+    }
+    if (!op->HasAttribute("shift_ratio") || !op->HasAttribute("seg_num")) {
+      VLOG(3) << "temporal shift need attributes : shift_ratio and seg_num";
+      return false;
+    }
+    op->set_attribute(kCanRunTrtAttr, rewriter.bool_attr(true));
+    return true;
+  }
+};
+
 class InstanceNormOpPattern
     : public pir::OpRewritePattern<paddle::dialect::InstanceNormOp> {
  public:
@@ -2388,6 +2409,7 @@ class TrtOpMarkerPass : public pir::PatternRewritePass {
     ps.Add(std::make_unique<TanhOpPattern>(context));
     ps.Add(std::make_unique<CeluOpPattern>(context));
     ps.Add(std::make_unique<OneHotOpPattern>(context));
+    ps.Add(std::make_unique<TemporalShiftOpPattern>(context));
     ps.Add(std::make_unique<InstanceNormOpPattern>(context));
     ps.Add(std::make_unique<AffineChannelOpPattern>(context));
     return ps;
diff --git a/python/paddle/tensorrt/impls/others.py b/python/paddle/tensorrt/impls/others.py
@@ -27,6 +27,7 @@
     trt_concat,
     trt_prod,
     trt_shape,
+    trt_sub,
     trt_sum,
 )
 from paddle.tensorrt.register import converter_registry
@@ -274,6 +275,112 @@ def share_data_converter(network, paddle_op, inputs):
     return identity_layer.get_output(0)
 
 
+@converter_registry.register("pd_op.temporal_shift", trt_version="8.x")
+def temporal_shift_converter(network, paddle_op, inputs):
+    input_tensor = inputs[0]
+    shift_ratio = paddle_op.attrs()["shift_ratio"]
+    T = paddle_op.attrs()["seg_num"]
+    data_format = paddle_op.attrs().get("data_format", "NCHW")
+
+    if data_format == "NHWC":
+        # Transpose input to [N, C, H, W]
+        transpose_layer = network.add_shuffle(input_tensor)
+        transpose_layer.first_transpose = trt.Permutation([0, 3, 1, 2])
+        input_tensor = transpose_layer.get_output(0)
+
+    input_dims = input_tensor.shape
+    C, H, W = input_dims[1], input_dims[2], input_dims[3]
+
+    # Reshape input to [N, T, C, H, W]
+    reshape_layer = network.add_shuffle(input_tensor)
+    reshape_layer.reshape_dims = trt.Dims([-1, T, C, H, W])
+    input_tensor = reshape_layer.get_output(0)
+
+    # Pad input to [N, T + 2, C, H, W]
+    pre_pad = add_1D_constant_layer(network, [0, 1, 0, 0, 0])
+    post_pad = add_1D_constant_layer(network, [0, 1, 0, 0, 0])
+    dims = 5
+    zeros = add_1D_constant_layer(network, [0] * dims)
+    start = trt_sub(network, zeros, pre_pad)
+    total_padding = trt_sum(network, pre_pad, post_pad)
+    input_shape = trt_shape(network, input_tensor)
+    size = trt_sum(network, input_shape, total_padding)
+    stride = [1] * dims
+    dummy = stride
+
+    slice_layer = network.add_slice(input_tensor, dummy, dummy, stride)
+    slice_layer.set_input(1, start)
+    slice_layer.set_input(2, size)
+
+    trt_version = trt.__version__.split('.')
+    if int(trt_version[0]) > 8 or (
+        int(trt_version[0]) == 8 and int(trt_version[1]) >= 5
+    ):
+        slice_layer.mode = trt.SampleMode.FILL
+    else:
+        slice_layer.mode = trt.SliceMode.FILL
+
+    slice_c = int(C * shift_ratio)
+    slice_c2 = int(C * shift_ratio * 2)
+
+    slice_start1 = zeros
+    slice_start2 = add_1D_constant_layer(network, [0, 2, slice_c, 0, 0])
+    slice_start3 = add_1D_constant_layer(network, [0, 1, slice_c2, 0, 0])
+
+    slice_size_base = trt_shape(network, input_tensor)
+    sub_size1 = add_1D_constant_layer(network, [0, 0, C - slice_c, 0, 0])
+    sub_size2 = add_1D_constant_layer(
+        network, [0, 0, C + slice_c - slice_c2, 0, 0]
+    )
+    sub_size3 = add_1D_constant_layer(network, [0, 0, slice_c2, 0, 0])
+
+    slice_size1 = trt_sub(network, slice_size_base, sub_size1)
+    slice_size2 = trt_sub(network, slice_size_base, sub_size2)
+    slice_size3 = trt_sub(network, slice_size_base, sub_size3)
+
+    slice1_layer = network.add_slice(
+        slice_layer.get_output(0), start=dummy, shape=dummy, stride=stride
+    )
+    slice1_layer.set_input(1, slice_start1)
+    slice1_layer.set_input(2, slice_size1)
+    slice2_layer = network.add_slice(
+        slice_layer.get_output(0), start=dummy, shape=dummy, stride=stride
+    )
+    slice2_layer.set_input(1, slice_start2)
+    slice2_layer.set_input(2, slice_size2)
+    slice3_layer = network.add_slice(
+        slice_layer.get_output(0), start=dummy, shape=dummy, stride=stride
+    )
+    slice3_layer.set_input(1, slice_start3)
+    slice3_layer.set_input(2, slice_size3)
+
+    concat_inputs = [slice2_layer.get_output(0), slice3_layer.get_output(0)]
+    if slice_c == 0:
+        concat_layer = network.add_concatenation(concat_inputs)
+        concat_layer.axis = 2
+    else:
+        concat_inputs = [
+            slice1_layer.get_output(0),
+            slice2_layer.get_output(0),
+            slice3_layer.get_output(0),
+        ]
+        concat_layer = network.add_concatenation(concat_inputs)
+        concat_layer.axis = 2
+
+    # Reshape output to [N*T,C,H,W]
+    reshape_layer3 = network.add_shuffle(concat_layer.get_output(0))
+    reshape_layer3.reshape_dims = trt.Dims([-1, C, H, W])
+
+    if data_format == "NHWC":
+        transpose_layer2 = network.add_shuffle(reshape_layer3.get_output(0))
+        transpose_layer2.first_transpose = trt.Permutation([0, 2, 3, 1])
+        output_tensor = transpose_layer2.get_output(0)
+    else:
+        output_tensor = reshape_layer3.get_output(0)
+
+    return output_tensor
+
+
 @converter_registry.register("pd_op.anchor_generator", trt_version="8.x")
 def anchor_generator_converter(network, paddle_op, inputs):
     inputs = inputs[0]
diff --git a/test/tensorrt/test_converter_others.py b/test/tensorrt/test_converter_others.py
@@ -406,6 +406,152 @@ def test_trt_result(self):
         self.check_trt_result()
 
 
+class TestTemporalShiftTRTPatternBasic(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = paddle.nn.functional.temporal_shift
+        self.api_args = {
+            "x": np.random.random([4, 9, 7, 7]).astype(np.float32),
+            "seg_num": 2,
+            "shift_ratio": 0.2,
+            "data_format": "NCHW",
+        }
+        self.program_config = {"feed_list": ["x"]}
+        self.min_shape = {"x": [2, 9, 7, 7]}
+        self.opt_shape = {"x": [2, 9, 7, 7]}
+        self.max_shape = {"x": [8, 9, 7, 7]}
+
+    def test_trt_result_fp16(self):
+        self.check_trt_result(precision_mode="fp16")
+
+    def test_trt_result_fp32(self):
+        self.check_trt_result()
+
+
+class TestTemporalShiftTRTPatternZeroSlice(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = paddle.nn.functional.temporal_shift
+        self.api_args = {
+            "x": np.random.random([4, 2, 7, 7]).astype(np.float32),
+            "seg_num": 2,
+            "shift_ratio": 0.2,
+            "data_format": "NCHW",
+        }
+        self.program_config = {"feed_list": ["x"]}
+        self.min_shape = {"x": [2, 2, 7, 7]}
+        self.opt_shape = {"x": [2, 2, 7, 7]}
+        self.max_shape = {"x": [8, 2, 7, 7]}
+
+    def test_trt_result_fp16(self):
+        self.check_trt_result(precision_mode="fp16")
+
+    def test_trt_result_fp32(self):
+        self.check_trt_result()
+
+
+class TestTemporalShiftTRTPatternDifferentSegNum(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = paddle.nn.functional.temporal_shift
+        self.api_args = {
+            "x": np.random.random([4, 9, 7, 7]).astype(np.float32),
+            "seg_num": 4,
+            "shift_ratio": 0.2,
+            "data_format": "NCHW",
+        }
+        self.program_config = {"feed_list": ["x"]}
+        self.min_shape = {"x": [4, 9, 7, 7]}
+        self.opt_shape = {"x": [4, 9, 7, 7]}
+        self.max_shape = {"x": [8, 9, 7, 7]}
+
+    def test_trt_result_fp16(self):
+        self.check_trt_result(precision_mode="fp16")
+
+    def test_trt_result_fp32(self):
+        self.check_trt_result()
+
+
+class TestTemporalShiftTRTPatternDifferentShiftRatio(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = paddle.nn.functional.temporal_shift
+        self.api_args = {
+            "x": np.random.random([4, 9, 7, 7]).astype(np.float32),
+            "seg_num": 2,
+            "shift_ratio": 0.4,
+            "data_format": "NCHW",
+        }
+        self.program_config = {"feed_list": ["x"]}
+        self.min_shape = {"x": [2, 9, 7, 7]}
+        self.opt_shape = {"x": [2, 9, 7, 7]}
+        self.max_shape = {"x": [8, 9, 7, 7]}
+
+    def test_trt_result_fp16(self):
+        self.check_trt_result(precision_mode="fp16")
+
+    def test_trt_result_fp32(self):
+        self.check_trt_result()
+
+
+class TestTemporalShiftTRTPatternDifferentDataFormat(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = paddle.nn.functional.temporal_shift
+        self.api_args = {
+            "x": np.random.random([4, 9, 7, 7]).astype(np.float32),
+            "seg_num": 2,
+            "shift_ratio": 0.2,
+            "name": None,
+            "data_format": "NHWC",
+        }
+        self.program_config = {"feed_list": ["x"]}
+        self.min_shape = {"x": [2, 9, 7, 7]}
+        self.opt_shape = {"x": [2, 9, 7, 7]}
+        self.max_shape = {"x": [8, 9, 7, 7]}
+
+    def test_trt_result_fp16(self):
+        self.check_trt_result(precision_mode="fp16")
+
+    def test_trt_result_fp32(self):
+        self.check_trt_result()
+
+
+class TestTemporalShiftTRTPatternMinMaxShape(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = paddle.nn.functional.temporal_shift
+        self.api_args = {
+            "x": np.random.random([4, 9, 7, 7]).astype(np.float32),
+            "seg_num": 2,
+            "shift_ratio": 0.2,
+            "data_format": "NCHW",
+        }
+        self.program_config = {"feed_list": ["x"]}
+        self.min_shape = {"x": [2, 9, 7, 7]}
+        self.opt_shape = {"x": [2, 9, 7, 7]}
+        self.max_shape = {"x": [10, 9, 7, 7]}
+
+    def test_trt_result_fp16(self):
+        self.check_trt_result(precision_mode="fp16")
+
+    def test_trt_result_fp32(self):
+        self.check_trt_result()
+
+
+def wrapper_temporal_shift(x):
+    return paddle.nn.functional.temporal_shift(x=x, seg_num=2, shift_ratio=0.2)
+
+
+class TestTemporalShiftTRTPatternError1(TensorRTBaseTest):
+    def setUp(self):
+        self.python_api = wrapper_temporal_shift
+        self.api_args = {
+            "x": np.random.random([4, 9, 7, 7]).astype(np.float32),
+        }
+        self.program_config = {"feed_list": ["x"]}
+        self.min_shape = {"x": [2, 9, 7, 7]}
+        self.opt_shape = {"x": [2, 9, 7, 7]}
+        self.max_shape = {"x": [10, 9, 7, 7]}
+
+    def test_trt_result(self):
+        self.check_marker(expected_result=False)
+
+
 def affine_channel(x, scale_shape, bias_shape, layout):
     scale = paddle.static.create_parameter(
         shape=scale_shape, dtype='float32', name="scale"
