add pad3d and pad3d_grad op for xpu, test=kunlun (#48306)

cmake/external/xpu.cmake

@@ -10,7 +10,7 @@ set(XPU_RT_LIB_NAME "libxpurt.so")
 if(NOT DEFINED XPU_BASE_URL)
   set(XPU_BASE_URL_WITHOUT_DATE
       "https://baidu-kunlun-product.su.bcebos.com/KL-SDK/klsdk-dev")
-  set(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20221120")
+  set(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20221124")
 else()
   set(XPU_BASE_URL "${XPU_BASE_URL}")
 endif()

paddle/fluid/platform/device/xpu/xpu2_op_list.h

@@ -433,6 +433,8 @@ XPUOpMap& get_kl2_ops() {
                      pOpKernelType(vartype::INT64, XPUPlace())})},
       {"p_norm", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"p_norm_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
+      {"pad3d_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
+      {"pad3d", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"pool2d_grad",
        XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),
                      pOpKernelType(vartype::FP16, XPUPlace())})},

paddle/phi/kernels/xpu/pad3d_grad_kernel.cc

+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/kernels/pad3d_grad_kernel.h"
+
+#include "paddle/phi/backends/xpu/enforce_xpu.h"
+#include "paddle/phi/backends/xpu/xpu_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void Pad3dGradKernel(const Context& dev_ctx,
+                     const DenseTensor& x,
+                     const DenseTensor& out_grad,
+                     const IntArray& paddings,
+                     const std::string& mode,
+                     float pad_value,
+                     const std::string& data_format,
+                     DenseTensor* x_grad) {
+  T value = static_cast<T>(pad_value);
+  std::vector<int64_t> pads = paddings.GetData();
+
+  auto* d_out = &out_grad;
+  auto* d_in = x_grad;
+  auto d_in_dims = d_in->dims();
+  const T* d_out_data = d_out->data<T>();
+  T* d_in_data = dev_ctx.template Alloc<T>(d_in);
+
+  bool is_ncdhw = true;
+  if (data_format == "NDHWC") {
+    is_ncdhw = false;
+  }
+
+  const int num = d_in_dims[0];  // n
+  int channels = d_in_dims[1];   // c
+  int in_depth = d_in_dims[2];   // xd
+  int in_height = d_in_dims[3];  // xh
+  int in_width = d_in_dims[4];   // xw
+  if (data_format == "NDHWC") {
+    channels = d_in_dims[4];
+    in_depth = d_in_dims[1];
+    in_height = d_in_dims[2];
+    in_width = d_in_dims[3];
+  }
+
+  std::vector<int> pads_xpu(6);
+  pads_xpu[0] = pads[4];  // pf
+  pads_xpu[1] = pads[5];  // pb
+  pads_xpu[2] = pads[2];  // pt
+  pads_xpu[3] = pads[3];  // pd
+  pads_xpu[4] = pads[0];  // pl
+  pads_xpu[5] = pads[1];  // pr
+
+  if (mode == "reflect") {
+    int r = xpu::reflection_pad3d_grad(dev_ctx.x_context(),
+                                       d_out_data,
+                                       d_in_data,
+                                       num,
+                                       channels,
+                                       in_depth,
+                                       in_height,
+                                       in_width,
+                                       pads_xpu,
+                                       is_ncdhw);
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "reflection_pad3d_grad");
+  } else if (mode == "replicate") {
+    int r = xpu::replication_pad3d_grad(dev_ctx.x_context(),
+                                        d_out_data,
+                                        d_in_data,
+                                        num,
+                                        channels,
+                                        in_depth,
+                                        in_height,
+                                        in_width,
+                                        pads_xpu,
+                                        is_ncdhw);
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "replication_pad3d_grad");
+  } else if (mode == "constant") {
+    int r = xpu::constant_pad3d_grad(dev_ctx.x_context(),
+                                     d_out_data,
+                                     d_in_data,
+                                     num,
+                                     channels,
+                                     in_depth,
+                                     in_height,
+                                     in_width,
+                                     pads_xpu,
+                                     value,
+                                     is_ncdhw);
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "constant_pad3d_grad");
+  }
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(pad3d_grad, XPU, ALL_LAYOUT, phi::Pad3dGradKernel, float) {}

paddle/phi/kernels/xpu/pad3d_kernel.cc

+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/kernels/pad3d_kernel.h"
+
+#include "paddle/phi/backends/xpu/enforce_xpu.h"
+#include "paddle/phi/backends/xpu/xpu_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void Pad3dKernel(const Context& dev_ctx,
+                 const DenseTensor& x,
+                 const IntArray& paddings,
+                 const std::string& mode,
+                 float pad_value,
+                 const std::string& data_format,
+                 DenseTensor* out) {
+  T value = static_cast<T>(pad_value);
+  std::vector<int64_t> pads = paddings.GetData();
+
+  auto in_dims = x.dims();
+  const T* in_data = x.data<T>();
+
+  bool is_ncdhw = true;
+  if (data_format == "NCDHW") {
+    out->Resize({in_dims[0],
+                 in_dims[1],
+                 in_dims[2] + pads[4] + pads[5],
+                 in_dims[3] + pads[2] + pads[3],
+                 in_dims[4] + pads[0] + pads[1]});
+  } else {
+    is_ncdhw = false;
+    out->Resize({in_dims[0],
+                 in_dims[1] + pads[4] + pads[5],
+                 in_dims[2] + pads[2] + pads[3],
+                 in_dims[3] + pads[0] + pads[1],
+                 in_dims[4]});
+  }
+
+  T* out_data = dev_ctx.template Alloc<T>(out);
+
+  const int num = in_dims[0];  // n
+  int channels = in_dims[1];   // c
+  int in_depth = in_dims[2];   // xd
+  int in_height = in_dims[3];  // xh
+  int in_width = in_dims[4];   // xw
+  if (data_format == "NDHWC") {
+    channels = in_dims[4];
+    in_depth = in_dims[1];
+    in_height = in_dims[2];
+    in_width = in_dims[3];
+  }
+
+  if (mode == "circular") {
+    PADDLE_THROW(phi::errors::External(
+        "XPU is not support circular padding mode in pad3d"));
+  }
+
+  if (mode == "reflect") {
+    PADDLE_ENFORCE_GT(
+        in_depth,
+        pads[4],
+        errors::InvalidArgument("The depth of Input(X)'s dimension should be "
+                                "greater than pad_front"
+                                " in reflect mode"
+                                ", but received depth(%d) and pad_front(%d).",
+                                in_depth,
+                                pads[4]));
+    PADDLE_ENFORCE_GT(
+        in_depth,
+        pads[5],
+        errors::InvalidArgument("The depth of Input(X)'s dimension should be "
+                                "greater than pad_back"
+                                " in reflect mode"
+                                ", but received depth(%d) and pad_back(%d).",
+                                in_depth,
+                                pads[5]));
+
+    PADDLE_ENFORCE_GT(
+        in_height,
+        pads[2],
+        errors::InvalidArgument("The height of Input(X)'s dimension should be "
+                                "greater than pad_top"
+                                " in reflect mode"
+                                ", but received depth(%d) and pad_top(%d).",
+                                in_height,
+                                pads[2]));
+    PADDLE_ENFORCE_GT(
+        in_height,
+        pads[3],
+        errors::InvalidArgument("The height of Input(X)'s dimension should be "
+                                "greater than pad_bottom"
+                                " in reflect mode"
+                                ", but received depth(%d) and pad_bottom(%d).",
+                                in_height,
+                                pads[3]));
+
+    PADDLE_ENFORCE_GT(
+        in_width,
+        pads[0],
+        errors::InvalidArgument("The width of Input(X)'s dimension should be "
+                                "greater than pad_left"
+                                " in reflect mode"
+                                ", but received depth(%d) and pad_left(%d).",
+                                in_width,
+                                pads[0]));
+    PADDLE_ENFORCE_GT(
+        in_width,
+        pads[1],
+        errors::InvalidArgument("The width of Input(X)'s dimension should be "
+                                "greater than pad_right"
+                                " in reflect mode"
+                                ", but received depth(%d) and pad_right(%d).",
+                                in_width,
+                                pads[1]));
+  } else if (mode == "replicate") {
+    PADDLE_ENFORCE_NE(in_depth * in_height * in_width,
+                      0,
+                      errors::InvalidArgument(
+                          "The input tensor size can not be 0 for circular "
+                          "or replicate padding mode."));
+  }
+
+  std::vector<int> pads_xpu(6);
+  pads_xpu[0] = pads[4];  // pf
+  pads_xpu[1] = pads[5];  // pb
+  pads_xpu[2] = pads[2];  // pt
+  pads_xpu[3] = pads[3];  // pd
+  pads_xpu[4] = pads[0];  // pl
+  pads_xpu[5] = pads[1];  // pr
+
+  if (mode == "reflect") {
+    int r = xpu::reflection_pad3d(dev_ctx.x_context(),
+                                  in_data,
+                                  out_data,
+                                  num,
+                                  channels,
+                                  in_depth,
+                                  in_height,
+                                  in_width,
+                                  pads_xpu,
+                                  is_ncdhw);
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "reflection_pad3d");
+  } else if (mode == "replicate") {
+    int r = xpu::replication_pad3d(dev_ctx.x_context(),
+                                   in_data,
+                                   out_data,
+                                   num,
+                                   channels,
+                                   in_depth,
+                                   in_height,
+                                   in_width,
+                                   pads_xpu,
+                                   is_ncdhw);
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "replication_pad3d");
+  } else if (mode == "constant") {
+    int r = xpu::constant_pad3d(dev_ctx.x_context(),
+                                in_data,
+                                out_data,
+                                num,
+                                channels,
+                                in_depth,
+                                in_height,
+                                in_width,
+                                pads_xpu,
+                                value,
+                                is_ncdhw);
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "constant_pad3d");
+  }
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(pad3d, XPU, ALL_LAYOUT, phi::Pad3dKernel, float) {}

python/paddle/fluid/tests/unittests/xpu/test_pad3d_op_xpu.py

+#   Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import numpy as np
+import sys
+
+sys.path.append("..")
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+from op_test_xpu import XPUOpTest
+from xpu.get_test_cover_info import (
+    create_test_class,
+    get_xpu_op_support_types,
+    XPUOpTestWrapper,
+)
+
+from paddle.fluid import Program, program_guard, Executor, default_main_program
+
+paddle.enable_static()
+
+
+class XPUTestPad3dOp(XPUOpTestWrapper):
+    def __init__(self):
+        self.op_name = 'pad3d'
+
+    class TestPad3dOp(XPUOpTest):
+        def setUp(self):
+            paddle.enable_static()
+            self.op_type = "pad3d"
+            self.dtype = self.in_type
+            self.place = paddle.XPUPlace(0)
+
+            self.value = 0.0
+            self.initTestCase()
+            self.python_api = paddle.nn.functional.pad
+            self.inputs = {'X': np.random.random(self.shape).astype(self.dtype)}
+            self.attrs = {}
+            if self.variable_paddings:
+                self.attrs['paddings'] = []
+                self.inputs['Paddings'] = (
+                    np.array(self.paddings).flatten().astype("int32")
+                )
+            else:
+                self.attrs['paddings'] = (
+                    np.array(self.paddings).flatten().astype("int32")
+                )
+            self.attrs['value'] = self.value
+            self.attrs['mode'] = self.mode
+            self.attrs['data_format'] = self.data_format
+            if self.data_format == "NCDHW":
+                paddings = [
+                    (0, 0),
+                    (0, 0),
+                    (self.paddings[4], self.paddings[5]),
+                    (self.paddings[2], self.paddings[3]),
+                    (self.paddings[0], self.paddings[1]),
+                ]
+            else:
+                paddings = [
+                    (0, 0),
+                    (self.paddings[4], self.paddings[5]),
+                    (self.paddings[2], self.paddings[3]),
+                    (self.paddings[0], self.paddings[1]),
+                    (0, 0),
+                ]
+            if self.mode == "constant":
+                out = np.pad(
+                    self.inputs['X'],
+                    paddings,
+                    mode=self.mode,
+                    constant_values=self.value,
+                )
+            elif self.mode == "reflect":
+                out = np.pad(self.inputs['X'], paddings, mode=self.mode)
+            elif self.mode == "replicate":
+                out = np.pad(self.inputs['X'], paddings, mode="edge")
+            elif self.mode == "circular":
+                out = np.pad(self.inputs['X'], paddings, mode="wrap")
+            self.outputs = {'Out': out}
+
+        def test_check_output(self):
+            self.check_output(check_eager=True)
+
+        def test_check_grad_normal(self):
+            self.check_grad(['X'], 'Out', check_eager=True)
+
+        def initTestCase(self):
+            self.shape = (2, 3, 4, 5, 6)
+            self.paddings = [0, 0, 0, 0, 0, 0]
+            self.mode = "constant"
+            self.data_format = "NCDHW"
+            self.pad_value = 0.0
+            self.variable_paddings = False
+
+    class TestCase1(TestPad3dOp):
+        def initTestCase(self):
+            self.shape = (2, 3, 4, 5, 6)
+            self.paddings = [0, 1, 2, 3, 4, 5]
+            self.mode = "constant"
+            self.data_format = "NCDHW"
+            self.value = 1.0
+            self.variable_paddings = False
+
+    class TestCase2(TestPad3dOp):
+        def initTestCase(self):
+            self.shape = (2, 3, 4, 5, 6)
+            self.paddings = [1, 1, 1, 1, 1, 1]
+            self.mode = "constant"
+            self.data_format = "NDHWC"
+            self.value = 1.0
+            self.variable_paddings = False
+
+    class TestCase3(TestPad3dOp):
+        def initTestCase(self):
+            self.shape = (2, 3, 4, 5, 6)
+            self.paddings = [0, 1, 1, 0, 2, 3]
+            self.mode = "reflect"
+            self.data_format = "NCDHW"
+            self.variable_paddings = False
+
+    class TestCase4(TestPad3dOp):
+        def initTestCase(self):
+            self.shape = (4, 4, 4, 4, 4)
+            self.paddings = [0, 1, 2, 1, 2, 3]
+            self.mode = "reflect"
+            self.data_format = "NDHWC"
+            self.variable_paddings = False
+
+    class TestCase5(TestPad3dOp):
+        def initTestCase(self):
+            self.shape = (2, 3, 4, 5, 6)
+            self.paddings = [0, 1, 2, 3, 2, 1]
+            self.mode = "replicate"
+            self.data_format = "NCDHW"
+            self.variable_paddings = False
+
+    class TestCase6(TestPad3dOp):
+        def initTestCase(self):
+            self.shape = (4, 4, 4, 4, 4)
+            self.paddings = [5, 4, 2, 1, 2, 3]
+            self.mode = "replicate"
+            self.data_format = "NDHWC"
+            self.variable_paddings = False
+
+    class TestCase7(TestPad3dOp):
+        def initTestCase(self):
+            self.shape = (2, 3, 4, 5, 6)
+            self.paddings = [0, 1, 2, 3, 4, 5]
+            self.mode = "constant"
+            self.data_format = "NCDHW"
+            self.value = 1.0
+            self.variable_paddings = True
+
+    class TestCase8(TestPad3dOp):
+        def initTestCase(self):
+            self.shape = (2, 3, 4, 5, 6)
+            self.paddings = [0, 1, 2, 3, 4, 5]
+            self.mode = "constant"
+            self.data_format = "NDHWC"
+            self.value = 1.0
+            self.variable_paddings = True
+
+    class TestPadAPI(unittest.TestCase):
+        def setUp(self):
+            self.places = [paddle.XPUPlace(0)]
+            self.dtype = self.in_type
+
+        def check_static_result_1(self, place):
+            paddle.enable_static()
+            with program_guard(Program(), Program()):
+                input_shape = (1, 2, 3, 4, 5)
+                pad = [1, 2, 1, 1, 3, 4]
+                mode = "constant"
+                value = 100
+                input_data = np.random.rand(*input_shape).astype(self.dtype)
+                x = paddle.fluid.data(name="x", shape=input_shape)
+                result = F.pad(
+                    x=x, pad=pad, value=value, mode=mode, data_format="NCDHW"
+                )
+                exe = Executor(place)
+                fetches = exe.run(
+                    default_main_program(),
+                    feed={"x": input_data},
+                    fetch_list=[result],
+                )
+
+                np_out = self._get_numpy_out(input_data, pad, mode, value)
+                np.testing.assert_allclose(fetches[0], np_out, rtol=1e-05)
+
+        def check_static_result_2(self, place):
+            paddle.enable_static()
+            with program_guard(Program(), Program()):
+                input_shape = (2, 3, 4, 5, 6)
+                pad = [1, 2, 1, 1, 1, 2]
+                mode = "reflect"
+                input_data = np.random.rand(*input_shape).astype(self.dtype)
+                x = paddle.fluid.data(name="x", shape=input_shape)
+                result1 = F.pad(x=x, pad=pad, mode=mode, data_format="NCDHW")
+                result2 = F.pad(x=x, pad=pad, mode=mode, data_format="NDHWC")
+                exe = Executor(place)
+                fetches = exe.run(
+                    default_main_program(),
+                    feed={"x": input_data},
+                    fetch_list=[result1, result2],
+                )
+
+                np_out1 = self._get_numpy_out(
+                    input_data, pad, mode, data_format="NCDHW"
+                )
+                np_out2 = self._get_numpy_out(
+                    input_data, pad, mode, data_format="NDHWC"
+                )
+                np.testing.assert_allclose(fetches[0], np_out1, rtol=1e-05)
+                np.testing.assert_allclose(fetches[1], np_out2, rtol=1e-05)
+
+        def check_static_result_3(self, place):
+            paddle.enable_static()
+            with program_guard(Program(), Program()):
+                input_shape = (2, 3, 4, 5, 6)
+                pad = [1, 2, 1, 1, 3, 4]
+                mode = "replicate"
+                input_data = np.random.rand(*input_shape).astype(self.dtype)
+                x = paddle.fluid.data(name="x", shape=input_shape)
+                result1 = F.pad(x=x, pad=pad, mode=mode, data_format="NCDHW")
+                result2 = F.pad(x=x, pad=pad, mode=mode, data_format="NDHWC")
+                exe = Executor(place)
+                fetches = exe.run(
+                    default_main_program(),
+                    feed={"x": input_data},
+                    fetch_list=[result1, result2],
+                )
+
+                np_out1 = self._get_numpy_out(
+                    input_data, pad, mode, data_format="NCDHW"
+                )
+                np_out2 = self._get_numpy_out(
+                    input_data, pad, mode, data_format="NDHWC"
+                )
+                np.testing.assert_allclose(fetches[0], np_out1, rtol=1e-05)
+                np.testing.assert_allclose(fetches[1], np_out2, rtol=1e-05)
+
+        def _get_numpy_out(
+            self, input_data, pad, mode, value=0, data_format="NCDHW"
+        ):
+            if mode == "constant" and len(pad) == len(input_data.shape) * 2:
+                pad = np.reshape(pad, (-1, 2)).tolist()
+            elif data_format == "NCDHW":
+                pad = [
+                    (0, 0),
+                    (0, 0),
+                    (pad[4], pad[5]),
+                    (pad[2], pad[3]),
+                    (pad[0], pad[1]),
+                ]
+            elif data_format == "NDHWC":
+                pad = [
+                    (0, 0),
+                    (pad[4], pad[5]),
+                    (pad[2], pad[3]),
+                    (pad[0], pad[1]),
+                    (0, 0),
+                ]
+            elif data_format == "NCHW":
+                pad = [
+                    (0, 0),
+                    (0, 0),
+                    (pad[2], pad[3]),
+                    (pad[0], pad[1]),
+                ]
+            elif data_format == "NHWC":
+                pad = [
+                    (0, 0),
+                    (pad[2], pad[3]),
+                    (pad[0], pad[1]),
+                    (0, 0),
+                ]
+            elif data_format == "NCL":
+                pad = [
+                    (0, 0),
+                    (0, 0),
+                    (pad[0], pad[1]),
+                ]
+            elif data_format == "NLC":
+                pad = [
+                    (0, 0),
+                    (pad[0], pad[1]),
+                    (0, 0),
+                ]
+
+            if mode == "constant":
+                out = np.pad(input_data, pad, mode=mode, constant_values=value)
+            elif mode == "reflect":
+                out = np.pad(input_data, pad, mode=mode)
+            elif mode == "replicate":
+                out = np.pad(input_data, pad, mode="edge")
+            elif mode == "circular":
+                out = np.pad(input_data, pad, mode="wrap")
+
+            return out
+
+        def test_static(self):
+            for place in self.places:
+                self.check_static_result_1(place=place)
+                self.check_static_result_2(place=place)
+                self.check_static_result_3(place=place)
+
+        def test_dygraph_1(self):
+            paddle.disable_static()
+            input_shape = (1, 2, 3, 4, 5)
+            pad = [1, 2, 1, 1, 3, 4]
+            pad_3 = [1, 2, 1, 1, 3, 4, 5, 6, 7, 8]
+            mode = "constant"
+            value = 100
+            input_data = np.random.rand(*input_shape).astype(self.dtype)
+            np_out1 = self._get_numpy_out(
+                input_data, pad, mode, value, data_format="NCDHW"
+            )
+            np_out2 = self._get_numpy_out(
+                input_data, pad, mode, value, data_format="NDHWC"
+            )
+            np_out3 = self._get_numpy_out(
+                input_data, pad_3, mode, value, data_format="NCDHW"
+            )
+            tensor_data = paddle.to_tensor(input_data)
+
+            y1 = F.pad(
+                tensor_data,
+                pad=pad,
+                mode=mode,
+                value=value,
+                data_format="NCDHW",
+            )
+            y2 = F.pad(
+                tensor_data,
+                pad=pad,
+                mode=mode,
+                value=value,
+                data_format="NDHWC",
+            )
+            y3 = F.pad(
+                tensor_data,
+                pad=pad_3,
+                mode=mode,
+                value=value,
+                data_format="NCDHW",
+            )
+
+            np.testing.assert_allclose(y1.numpy(), np_out1, rtol=1e-05)
+            np.testing.assert_allclose(y2.numpy(), np_out2, rtol=1e-05)
+            np.testing.assert_allclose(y3.numpy(), np_out3, rtol=1e-05)
+
+        def test_dygraph_2(self):
+            paddle.disable_static()
+            input_shape = (2, 3, 4, 5)
+            pad = [1, 1, 3, 4]
+            pad_3 = [1, 2, 1, 1, 3, 4, 5, 6]
+            mode = "constant"
+            value = 100
+            input_data = np.random.rand(*input_shape).astype(self.dtype)
+            np_out1 = self._get_numpy_out(
+                input_data, pad, mode, value, data_format="NCHW"
+            )
+            np_out2 = self._get_numpy_out(
+                input_data, pad, mode, value, data_format="NHWC"
+            )
+            np_out3 = self._get_numpy_out(
+                input_data, pad_3, mode, value, data_format="NCHW"
+            )
+
+            tensor_data = paddle.to_tensor(input_data)
+            tensor_pad = paddle.to_tensor(pad, dtype="int32")
+
+            y1 = F.pad(
+                tensor_data,
+                pad=tensor_pad,
+                mode=mode,
+                value=value,
+                data_format="NCHW",
+            )
+            y2 = F.pad(
+                tensor_data,
+                pad=tensor_pad,
+                mode=mode,
+                value=value,
+                data_format="NHWC",
+            )
+            y3 = F.pad(
+                tensor_data,
+                pad=pad_3,
+                mode=mode,
+                value=value,
+                data_format="NCHW",
+            )
+
+            np.testing.assert_allclose(y1.numpy(), np_out1, rtol=1e-05)
+            np.testing.assert_allclose(y2.numpy(), np_out2, rtol=1e-05)
+            np.testing.assert_allclose(y3.numpy(), np_out3, rtol=1e-05)
+
+        def test_dygraph_3(self):
+            paddle.disable_static()
+            input_shape = (3, 4, 5)
+            pad = [3, 4]
+            pad_3 = [3, 4, 5, 6, 7, 8]
+            mode = "constant"
+            value = 100
+            input_data = np.random.rand(*input_shape).astype(self.dtype)
+            np_out1 = self._get_numpy_out(
+                input_data, pad, mode, value, data_format="NCL"
+            )
+            np_out2 = self._get_numpy_out(
+                input_data, pad, mode, value, data_format="NLC"
+            )
+            np_out3 = self._get_numpy_out(
+                input_data, pad_3, mode, value, data_format="NCL"
+            )
+            tensor_data = paddle.to_tensor(input_data)
+            tensor_pad = paddle.to_tensor(pad, dtype="int32")
+
+            y1 = F.pad(
+                tensor_data,
+                pad=tensor_pad,
+                mode=mode,
+                value=value,
+                data_format="NCL",
+            )
+            y2 = F.pad(
+                tensor_data,
+                pad=tensor_pad,
+                mode=mode,
+                value=value,
+                data_format="NLC",
+            )
+            y3 = F.pad(
+                tensor_data,
+                pad=pad_3,
+                mode=mode,
+                value=value,
+                data_format="NCL",
+            )
+
+            np.testing.assert_allclose(y1.numpy(), np_out1, rtol=1e-05)
+            np.testing.assert_allclose(y2.numpy(), np_out2, rtol=1e-05)
+            np.testing.assert_allclose(y3.numpy(), np_out3, rtol=1e-05)
+
+    class TestPad3dAPI(unittest.TestCase):
+        def _get_numpy_out(
+            self, input_data, pad, mode, value=0.0, data_format="NCDHW"
+        ):
+            if data_format == "NCDHW":
+                pad = [
+                    (0, 0),
+                    (0, 0),
+                    (pad[4], pad[5]),
+                    (pad[2], pad[3]),
+                    (pad[0], pad[1]),
+                ]
+            else:
+                pad = [
+                    (0, 0),
+                    (pad[4], pad[5]),
+                    (pad[2], pad[3]),
+                    (pad[0], pad[1]),
+                    (0, 0),
+                ]
+
+            if mode == "constant":
+                out = np.pad(input_data, pad, mode=mode, constant_values=value)
+            elif mode == "reflect":
+                out = np.pad(input_data, pad, mode=mode)
+            elif mode == "replicate":
+                out = np.pad(input_data, pad, mode="edge")
+            elif mode == "circular":
+                out = np.pad(input_data, pad, mode="wrap")
+
+            return out
+
+        def setUp(self):
+            self.places = [paddle.XPUPlace(0)]
+            self.dtype = self.in_type
+
+        def test_class(self):
+            paddle.disable_static()
+            for place in self.places:
+                input_shape = (3, 4, 5, 6, 7)
+                pad = [1, 2, 2, 1, 1, 0]
+                pad_int = 1
+                value = 100
+                input_data = np.random.rand(*input_shape).astype(self.dtype)
+
+                pad_reflection = nn.Pad3D(padding=pad, mode="reflect")
+                pad_replication = nn.Pad3D(padding=pad, mode="replicate")
+                pad_constant = nn.Pad3D(
+                    padding=pad, mode="constant", value=value
+                )
+                pad_constant_int = nn.Pad3D(
+                    padding=pad_int, mode="constant", value=value
+                )
+                pad_circular = nn.Pad3D(padding=pad, mode="circular")
+
+                data = paddle.to_tensor(input_data)
+
+                output = pad_reflection(data)
+                np_out = self._get_numpy_out(
+                    input_data, pad, "reflect", data_format="NCDHW"
+                )
+                np.testing.assert_allclose(output.numpy(), np_out, rtol=1e-05)
+
+                output = pad_replication(data)
+                np_out = self._get_numpy_out(
+                    input_data, pad, "replicate", data_format="NCDHW"
+                )
+                np.testing.assert_allclose(output.numpy(), np_out, rtol=1e-05)
+
+                output = pad_constant(data)
+                np_out = self._get_numpy_out(
+                    input_data,
+                    pad,
+                    "constant",
+                    value=value,
+                    data_format="NCDHW",
+                )
+                np.testing.assert_allclose(output.numpy(), np_out, rtol=1e-05)
+
+                output = pad_constant_int(data)
+                np_out = self._get_numpy_out(
+                    input_data,
+                    [pad_int] * 6,
+                    "constant",
+                    value=value,
+                    data_format="NCDHW",
+                )
+                np.testing.assert_allclose(output.numpy(), np_out, rtol=1e-05)
+
+        def test_pad_tensor(self):
+            paddle.disable_static()
+            for place in self.places:
+                input_shape = (3, 4, 5, 6, 7)
+                pad = [1, 2, 2, 1, 1, 0]
+                pad_tensor = paddle.to_tensor(pad)
+                input_data = np.random.rand(*input_shape).astype(np.float32)
+
+                pad_reflection_ncdhw = nn.Pad3D(
+                    padding=pad_tensor, mode="reflect", data_format="NCDHW"
+                )
+                pad_reflection_ndhwc = nn.Pad3D(
+                    padding=pad_tensor, mode="reflect", data_format="NDHWC"
+                )
+                data = paddle.to_tensor(input_data)
+
+                output = pad_reflection_ncdhw(data)
+                np_out = self._get_numpy_out(
+                    input_data, pad, "reflect", data_format="NCDHW"
+                )
+                np.testing.assert_allclose(output.numpy(), np_out, rtol=1e-05)
+
+                output = pad_reflection_ndhwc(data)
+                np_out = self._get_numpy_out(
+                    input_data, pad, "reflect", data_format="NDHWC"
+                )
+                np.testing.assert_allclose(output.numpy(), np_out, rtol=1e-05)
+
+    class TestPad3dOpError(unittest.TestCase):
+        def setUp(self):
+            self.places = [paddle.XPUPlace(0)]
+            self.dtype = self.in_type
+
+        def test_errors(self):
+            def test_variable():
+                input_shape = (1, 2, 3, 4, 5)
+                data = np.random.rand(*input_shape).astype(self.dtype)
+                y = F.pad(x=data, pad=[1, 1, 1, 1, 1, 1], data_format="NCDHW")
+
+            def test_reflect_1():
+                input_shape = (1, 2, 3, 4, 5)
+                data = np.random.rand(*input_shape).astype(self.dtype)
+                x = paddle.to_tensor(data)
+                y = F.pad(
+                    x,
+                    pad=[5, 6, 1, 1, 1, 1],
+                    value=1,
+                    mode='reflect',
+                    data_format="NCDHW",
+                )
+
+            def test_reflect_2():
+                input_shape = (1, 2, 3, 4, 5)
+                data = np.random.rand(*input_shape).astype(self.dtype)
+                x = paddle.to_tensor(data)
+                y = F.pad(
+                    x,
+                    pad=[1, 1, 4, 3, 1, 1],
+                    value=1,
+                    mode='reflect',
+                    data_format="NCDHW",
+                )
+
+            def test_reflect_3():
+                input_shape = (1, 2, 3, 4, 5)
+                data = np.random.rand(*input_shape).astype(self.dtype)
+                x = paddle.to_tensor(data)
+                y = F.pad(
+                    x,
+                    pad=[1, 1, 1, 1, 2, 3],
+                    value=1,
+                    mode='reflect',
+                    data_format="NCDHW",
+                )
+
+            def test_replicate_1():
+                input_shape = (1, 2, 0, 4, 5)
+                data = np.random.rand(*input_shape).astype(self.dtype)
+                x = paddle.to_tensor(data)
+                y = F.pad(
+                    x,
+                    pad=[1, 1, 1, 1, 2, 3],
+                    mode='replicate',
+                    data_format="NCDHW",
+                )
+
+            paddle.disable_static()
+            for place in self.places:
+                self.assertRaises(ValueError, test_variable)
+                self.assertRaises(Exception, test_reflect_1)
+                self.assertRaises(Exception, test_reflect_2)
+                self.assertRaises(Exception, test_reflect_3)
+                self.assertRaises(Exception, test_replicate_1)
+            paddle.enable_static()
+
+    class TestPadDataformatError(unittest.TestCase):
+        def test_errors(self):
+            def test_ncl():
+                input_shape = (1, 2, 3, 4)
+                pad = paddle.to_tensor(np.array([2, 1, 2, 1]).astype('int32'))
+                data = (
+                    np.arange(np.prod(input_shape), dtype=np.float64).reshape(
+                        input_shape
+                    )
+                    + 1
+                )
+                my_pad = nn.Pad1D(
+                    padding=pad, mode="replicate", data_format="NCL"
+                )
+                data = paddle.to_tensor(data)
+                result = my_pad(data)
+
+            def test_nchw():
+                input_shape = (1, 2, 4)
+                pad = paddle.to_tensor(np.array([2, 1, 2, 1]).astype('int32'))
+                data = (
+                    np.arange(np.prod(input_shape), dtype=np.float64).reshape(
+                        input_shape
+                    )
+                    + 1
+                )
+                my_pad = nn.Pad1D(
+                    padding=pad, mode="replicate", data_format="NCHW"
+                )
+                data = paddle.to_tensor(data)
+                result = my_pad(data)
+
+            def test_ncdhw():
+                input_shape = (1, 2, 3, 4)
+                pad = paddle.to_tensor(np.array([2, 1, 2, 1]).astype('int32'))
+                data = (
+                    np.arange(np.prod(input_shape), dtype=np.float64).reshape(
+                        input_shape
+                    )
+                    + 1
+                )
+                my_pad = nn.Pad1D(
+                    padding=pad, mode="replicate", data_format="NCDHW"
+                )
+                data = paddle.to_tensor(data)
+                result = my_pad(data)
+
+            self.assertRaises(AssertionError, test_ncl)
+            self.assertRaises(AssertionError, test_nchw)
+            self.assertRaises(AssertionError, test_ncdhw)
+
+
+support_types = get_xpu_op_support_types('pad3d')
+for stype in support_types:
+    create_test_class(globals(), XPUTestPad3dOp, stype)
+
+
+if __name__ == '__main__':
+    unittest.main()