[cherry-pick] Add xpu transpose2 op.test=kunlun (#28096)

paddle/fluid/operators/transpose_op_xpu.cc

+/* Copyright (c) 2020 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. */
+
+#ifdef PADDLE_WITH_XPU
+#include "paddle/fluid/operators/transpose_op.h"
+#include <memory>
+#include <string>
+#include <vector>
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+bool XPUSupported(int ndims, const std::vector<int>& axis) {
+  /*
+   * XPU currently support:
+   * permute = {0, 2, 1}, permute = {1, 0},
+   * permute = {0, 2, 1, 3}, permute = {1, 0, 2},
+   * permute = {0, 2, 3, 1}
+   */
+  bool is_supported = false;
+  std::vector<int> permute_10(2, 0);
+  std::vector<int> permute_102(3, 0);
+  std::vector<int> permute_021(3, 0);
+  std::vector<int> permute_210(3, 0);
+  std::vector<int> permute_0213(4, 0);
+  std::vector<int> permute_0231(4, 0);
+  std::vector<int> permute_0312(4, 0);
+  std::vector<int> permute_3201(4, 0);
+  permute_10[0] = 1;
+  permute_102[0] = 1;
+  permute_102[2] = 2;
+  permute_021[1] = 2;
+  permute_021[2] = 1;
+  permute_210[0] = 2;
+  permute_210[1] = 1;
+  permute_0213[1] = 2;
+  permute_0213[2] = 1;
+  permute_0213[3] = 3;
+  permute_0231[1] = 2;
+  permute_0231[2] = 3;
+  permute_0231[3] = 1;
+  permute_0312[1] = 3;
+  permute_0312[2] = 1;
+  permute_0312[3] = 2;
+  permute_3201[0] = 3;
+  permute_3201[1] = 2;
+  permute_3201[3] = 1;
+  switch (ndims) {
+    case 2:
+      if (axis == permute_10) {
+        is_supported = true;
+      }
+      break;
+    case 3:
+      if ((axis == permute_021) || (axis == permute_102) ||
+          (axis == permute_210)) {
+        is_supported = true;
+      }
+      break;
+    case 4:
+      if ((axis == permute_0213) || (axis == permute_0231) ||
+          (axis == permute_0312) || (axis == permute_3201)) {
+        is_supported = true;
+      }
+      break;
+    default:
+      PADDLE_THROW(platform::errors::Unimplemented(
+          "Tensors with rank only 2, 3 and 4 are supported on XPU"));
+  }
+  return is_supported;
+}
+
+template <typename DeviceContext, typename T>
+class TransposeXPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto x = context.Input<framework::Tensor>("X");
+    auto out = context.Output<framework::Tensor>("Out");
+    // axis is permute
+    auto axis = context.Attr<std::vector<int>>("axis");
+    int ndims = axis.size();
+    const auto x_dims = x->dims();
+
+    const T* x_data = x->data<T>();
+    T* y_data = out->mutable_data<T>(context.GetPlace());
+    if (!XPUSupported(ndims, axis)) {
+      VLOG(0) << "XPU does not support the permute, try to do on cpu";
+      framework::Tensor x_cpu;
+      framework::Tensor out_cpu;
+      auto x_cpu_data = x_cpu.mutable_data<T>(x->dims(), platform::CPUPlace());
+      auto out_cpu_data =
+          out_cpu.mutable_data<T>(out->dims(), platform::CPUPlace());
+      memory::Copy(platform::CPUPlace(), reinterpret_cast<void*>(x_cpu_data),
+                   BOOST_GET_CONST(platform::XPUPlace, context.GetPlace()),
+                   (const void*)x_data, x->numel() * sizeof(T));
+
+      const platform::CPUDeviceContext* cpu_dev_ctx =
+          static_cast<const platform::CPUDeviceContext*>(
+              platform::DeviceContextPool::Instance().Get(
+                  platform::CPUPlace()));
+      TransCompute<platform::CPUDeviceContext, T>(ndims, *cpu_dev_ctx, x_cpu,
+                                                  &out_cpu, axis);
+      memory::Copy(BOOST_GET_CONST(platform::XPUPlace, context.GetPlace()),
+                   reinterpret_cast<void*>(y_data), platform::CPUPlace(),
+                   (const void*)out_cpu_data, out->numel() * sizeof(T));
+      return;
+    }
+
+    std::vector<int> x_shape_host(ndims, 0);
+    for (int i = 0; i < ndims; ++i) {
+      x_shape_host[i] = x_dims[i];
+    }
+    int* permute_host = axis.data();
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    int r = xpu::transpose(dev_ctx.x_context(), x_data, y_data,
+                           x_shape_host.data(), permute_host, ndims);
+    PADDLE_ENFORCE_EQ(
+        r, xpu::Error_t::SUCCESS,
+        platform::errors::External("XPU kernel error! error code=%d", r));
+  }
+};
+
+template <typename DeviceContext, typename T>
+class TransposeGradXPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* out_grad =
+        context.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* x_grad =
+        context.Output<framework::Tensor>(framework::GradVarName("X"));
+    if (!x_grad) return;
+
+    x_grad->mutable_data<T>(context.GetPlace());
+    std::vector<int> axis = context.Attr<std::vector<int>>("axis");
+    std::vector<int> reversed_axis(axis);
+    for (size_t i = 0; i < axis.size(); i++) {
+      reversed_axis[axis[i]] = i;
+    }
+
+    int ndims = axis.size();
+    if (!XPUSupported(ndims, reversed_axis)) {
+      PADDLE_THROW(
+          platform::errors::Unimplemented("XPU does not support the permute"));
+    }
+
+    std::vector<int> out_shape_host(ndims, 0);
+    for (int i = 0; i < ndims; ++i) {
+      out_shape_host[i] = out_grad->dims()[i];
+    }
+    int* permute_host = reversed_axis.data();
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    int r = xpu::transpose(dev_ctx.x_context(), out_grad->data<T>(),
+                           x_grad->data<T>(), out_shape_host.data(),
+                           permute_host, ndims);
+    PADDLE_ENFORCE_EQ(
+        r, xpu::Error_t::SUCCESS,
+        platform::errors::External("XPU kernel error! error code=%d", r));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_XPU_KERNEL(
+    transpose,
+    ops::TransposeXPUKernel<paddle::platform::XPUDeviceContext, float>);
+REGISTER_OP_XPU_KERNEL(
+    transpose_grad,
+    ops::TransposeGradXPUKernel<paddle::platform::XPUDeviceContext, float>);
+REGISTER_OP_XPU_KERNEL(
+    transpose2,
+    ops::TransposeXPUKernel<paddle::platform::XPUDeviceContext, float>);
+REGISTER_OP_XPU_KERNEL(
+    transpose2_grad,
+    ops::TransposeGradXPUKernel<paddle::platform::XPUDeviceContext, float>);
+
+#endif  // PADDLE_WITH_XPU

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

+#   Copyright (c) 2018 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.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import sys
+
+sys.path.append("..")
+from op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid import compiler, Program, program_guard
+
+
+class TestXPUTransposeOp(OpTest):
+    def setUp(self):
+        self.init_op_type()
+        self.initTestCase()
+        self.inputs = {'X': np.random.random(self.shape).astype("float64")}
+        self.attrs = {
+            'axis': list(self.axis),
+            'use_mkldnn': False,
+            'use_xpu': True
+        }
+        self.outputs = {
+            'XShape': np.random.random(self.shape).astype("float64"),
+            'Out': self.inputs['X'].transpose(self.axis)
+        }
+
+    def init_op_type(self):
+        self.op_type = "transpose2"
+        self.use_mkldnn = False
+
+    def test_check_output(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_output_with_place(place=place, no_check_set=['XShape'])
+
+    def test_check_grad(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_grad_with_place(place, ['X'], 'Out')
+
+    def initTestCase(self):
+        self.shape = (3, 40)
+        self.axis = (1, 0)
+
+
+class TestCase0(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (100, )
+        self.axis = (0, )
+
+
+class TestCase1(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (3, 4, 10)
+        self.axis = (0, 2, 1)
+
+
+class TestCase2(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (2, 3, 4, 5)
+        self.axis = (0, 2, 3, 1)
+
+
+class TestCase3(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (2, 3, 4, 5, 6)
+        self.axis = (4, 2, 3, 1, 0)
+
+
+class TestCase4(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (2, 3, 4, 5, 6, 1)
+        self.axis = (4, 2, 3, 1, 0, 5)
+
+
+class TestCase5(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (2, 16, 96)
+        self.axis = (0, 2, 1)
+
+
+class TestCase6(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (2, 10, 12, 16)
+        self.axis = (3, 1, 2, 0)
+
+
+class TestCase7(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (2, 10, 2, 16)
+        self.axis = (0, 1, 3, 2)
+
+
+class TestCase8(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (2, 3, 2, 3, 2, 4, 3, 3)
+        self.axis = (0, 1, 3, 2, 4, 5, 6, 7)
+
+
+class TestCase9(TestXPUTransposeOp):
+    def initTestCase(self):
+        self.shape = (2, 3, 2, 3, 2, 4, 3, 3)
+        self.axis = (6, 1, 3, 5, 0, 2, 4, 7)
+
+
+class TestTransposeOpError(unittest.TestCase):
+    def test_errors(self):
+        with program_guard(Program(), Program()):
+            x = fluid.layers.data(name='x', shape=[10, 5, 3], dtype='float64')
+
+            def test_x_Variable_check():
+                # the Input(x)'s type must be Variable
+                fluid.layers.transpose("not_variable", perm=[1, 0, 2])
+
+            self.assertRaises(TypeError, test_x_Variable_check)
+
+            def test_x_dtype_check():
+                # the Input(x)'s dtype must be one of [float16, float32, float64, int32, int64]
+                x1 = fluid.layers.data(
+                    name='x1', shape=[10, 5, 3], dtype='bool')
+                fluid.layers.transpose(x1, perm=[1, 0, 2])
+
+            self.assertRaises(TypeError, test_x_dtype_check)
+
+            def test_perm_list_check():
+                # Input(perm)'s type must be list
+                fluid.layers.transpose(x, perm="[1, 0, 2]")
+
+            self.assertRaises(TypeError, test_perm_list_check)
+
+            def test_perm_length_and_x_dim_check():
+                # Input(perm) is the permutation of dimensions of Input(input)
+                # its length should be equal to dimensions of Input(input)
+                fluid.layers.transpose(x, perm=[1, 0, 2, 3, 4])
+
+            self.assertRaises(ValueError, test_perm_length_and_x_dim_check)
+
+            def test_each_elem_value_check():
+                # Each element in Input(perm) should be less than Input(x)'s dimension
+                fluid.layers.transpose(x, perm=[3, 5, 7])
+
+            self.assertRaises(ValueError, test_each_elem_value_check)
+
+
+class TestTAPI(unittest.TestCase):
+    def test_out(self):
+        with fluid.program_guard(fluid.Program()):
+            data = fluid.data(shape=[10], dtype="float64", name="data")
+            data_t = paddle.t(data)
+            place = fluid.CPUPlace()
+            exe = fluid.Executor(place)
+            data_np = np.random.random([10]).astype("float64")
+            result, = exe.run(feed={"data": data_np}, fetch_list=[data_t])
+            expected_result = np.transpose(data_np)
+        self.assertEqual((result == expected_result).all(), True)
+
+        with fluid.program_guard(fluid.Program()):
+            data = fluid.data(shape=[10, 5], dtype="float64", name="data")
+            data_t = paddle.t(data)
+            place = fluid.CPUPlace()
+            exe = fluid.Executor(place)
+            data_np = np.random.random([10, 5]).astype("float64")
+            result, = exe.run(feed={"data": data_np}, fetch_list=[data_t])
+            expected_result = np.transpose(data_np)
+        self.assertEqual((result == expected_result).all(), True)
+
+        with fluid.program_guard(fluid.Program()):
+            data = fluid.data(shape=[1, 5], dtype="float64", name="data")
+            data_t = paddle.t(data)
+            place = fluid.CPUPlace()
+            exe = fluid.Executor(place)
+            data_np = np.random.random([1, 5]).astype("float64")
+            result, = exe.run(feed={"data": data_np}, fetch_list=[data_t])
+            expected_result = np.transpose(data_np)
+        self.assertEqual((result == expected_result).all(), True)
+
+        with fluid.dygraph.guard():
+            np_x = np.random.random([10]).astype("float64")
+            data = fluid.dygraph.to_variable(np_x)
+            z = paddle.t(data)
+            np_z = z.numpy()
+            z_expected = np.array(np.transpose(np_x))
+        self.assertEqual((np_z == z_expected).all(), True)
+
+        with fluid.dygraph.guard():
+            np_x = np.random.random([10, 5]).astype("float64")
+            data = fluid.dygraph.to_variable(np_x)
+            z = paddle.t(data)
+            np_z = z.numpy()
+            z_expected = np.array(np.transpose(np_x))
+        self.assertEqual((np_z == z_expected).all(), True)
+
+        with fluid.dygraph.guard():
+            np_x = np.random.random([1, 5]).astype("float64")
+            data = fluid.dygraph.to_variable(np_x)
+            z = paddle.t(data)
+            np_z = z.numpy()
+            z_expected = np.array(np.transpose(np_x))
+        self.assertEqual((np_z == z_expected).all(), True)
+
+    def test_errors(self):
+        with fluid.program_guard(fluid.Program()):
+            x = fluid.data(name='x', shape=[10, 5, 3], dtype='float64')
+
+            def test_x_dimension_check():
+                paddle.t(x)
+
+            self.assertRaises(ValueError, test_x_dimension_check)
+
+
+if __name__ == "__main__":
+    unittest.main()