【Paddle Toolkit Development Competition No.2】[Custom OP] Add contiguous api support to paddle::Tensor -part (#68748)

BeingGod · web-flow · commit f495a6cb474a · 2024-11-07T14:40:50.000+08:00
* [Custom OP] Add contiguous api support to paddle::Tensor

* fix

* fix

* CI retrigger

* add null ptr check

* fix null ptr

* fix coverage
diff --git a/paddle/phi/api/include/tensor.h b/paddle/phi/api/include/tensor.h
@@ -637,6 +637,24 @@ class PADDLE_API Tensor final {
    */
   Tensor to_dense() const;
 
+  /* Part 12: Contiguous methods */
+
+  /**
+   * @brief Determine whether tensor is contiguous
+   *
+   * @return bool
+   */
+  bool is_contiguous() const;
+
+  /**
+   * @brief Returns a contiguous in memory tensor containing the same data as
+   * current Tensor. If self tensor is already contiguous, this function returns
+   * the current Tensor.
+   *
+   * @return Tensor
+   */
+  Tensor contiguous();
+
  private:
   /**
    * [ Why use abstract TensorImpl interface here? ]
diff --git a/paddle/phi/api/lib/tensor.cc b/paddle/phi/api/lib/tensor.cc
@@ -22,6 +22,7 @@ limitations under the License. */
 
 #include "paddle/common/ddim.h"
 #include "paddle/phi/api/include/context_pool.h"
+#include "paddle/phi/api/lib/data_transform.h"
 #include "paddle/phi/api/lib/utils/allocator.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/backends/gpu/gpu_info.h"
@@ -508,4 +509,51 @@ void Tensor::reset_inplace_version(bool set_to_zero) {
   }
 }
 
+/* Part 7: Contiguous methods */
+
+bool Tensor::is_contiguous() const {
+  if (is_dense_tensor() || is_dist_tensor()) {
+    phi::DenseTensor *dense_tensor = nullptr;
+    if (is_dist_tensor()) {
+      dense_tensor = static_cast<phi::distributed::DistTensor *>(impl_.get())
+                         ->unsafe_mutable_value();
+    } else {
+      dense_tensor = static_cast<phi::DenseTensor *>(impl_.get());
+    }
+    return dense_tensor->meta().is_contiguous();
+  } else {
+    PADDLE_THROW(
+        common::errors::Unimplemented("Only support is_contiguous operation on "
+                                      "DenseTensor or DistTensor now."));
+  }
+}
+
+Tensor Tensor::contiguous() {
+  if (is_dense_tensor() || is_dist_tensor()) {
+    phi::DenseTensor *dense_tensor = nullptr;
+    if (is_dist_tensor()) {
+      dense_tensor = static_cast<phi::distributed::DistTensor *>(impl_.get())
+                         ->unsafe_mutable_value();
+    } else {
+      dense_tensor = static_cast<phi::DenseTensor *>(impl_.get());
+    }
+    PADDLE_ENFORCE_NOT_NULL(dense_tensor,
+                            common::errors::InvalidArgument(
+                                "TensorImpl with nullptr is not supported"));
+    if (!dense_tensor->meta().is_contiguous()) {
+      auto new_dense_tensor = std::make_shared<phi::DenseTensor>();
+      *new_dense_tensor = paddle::experimental::Trans2Contiguous(*dense_tensor);
+
+      return Tensor(std::shared_ptr<phi::TensorBase>(new_dense_tensor),
+                    autograd_meta_,
+                    name_);
+    } else {
+      return *this;
+    }
+  } else {
+    PADDLE_THROW(common::errors::Unimplemented(
+        "Only support contiguous operation on DenseTensor or DistTensor now."));
+  }
+}
+
 }  // namespace paddle
diff --git a/test/custom_op/CMakeLists.txt b/test/custom_op/CMakeLists.txt
@@ -71,6 +71,8 @@ if(WITH_TESTING)
   py_test(test_custom_tanh_double_grad SRCS test_custom_tanh_double_grad.py)
   py_test(test_custom_inplace SRCS test_custom_inplace.py)
   py_test(test_custom_optional SRCS test_custom_optional.py)
+  py_test(test_custom_contiguous SRCS test_custom_contiguous.py)
+  set_tests_properties(test_custom_contiguous PROPERTIES TIMEOUT 180)
 
   # other tests
   py_test(test_sysconfig SRCS test_sysconfig.py)
diff --git a/test/custom_op/custom_contiguous.cc b/test/custom_op/custom_contiguous.cc
@@ -0,0 +1,62 @@
+// Copyright (c) 2024 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
+//
+//     https://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 <iostream>
+#include <vector>
+
+#include "paddle/extension.h"
+
+static paddle::Tensor Transpose(const paddle::Tensor& t,
+                                int64_t dim0,
+                                int64_t dim1) {
+  int len = t.shape().size();
+  dim0 = dim0 >= 0 ? dim0 : len + dim0;
+  dim1 = dim1 >= 0 ? dim1 : len + dim1;
+  PD_CHECK(dim0 >= 0 && dim0 < len,
+           "dim0 not in range"
+           "dim0:%d ,range:%d",
+           dim0,
+           len);
+  PD_CHECK(dim1 >= 0 && dim1 < len,
+           "dim1 not in range"
+           "dim1:%d ,range:%d",
+           dim1,
+           len);
+  std::vector<int> transpose_perm(len);
+  std::iota(transpose_perm.begin(), transpose_perm.end(), 0);
+  transpose_perm[dim0] = dim1;
+  transpose_perm[dim1] = dim0;
+  // maybe there is another way to avoid experiment api
+  return paddle::experimental::transpose(t, transpose_perm);
+}
+
+std::vector<paddle::Tensor> ContiguousForward(paddle::Tensor& x) {  // NOLINT
+  PD_CHECK(x.shape().size() == 2, "x must be a 2-d tensor.");
+
+  x = x.contiguous();
+  PD_CHECK(x.is_contiguous(), "Check failed !");
+
+  auto non_contiguous_x = Transpose(x, 0, 1);
+  PD_CHECK(!non_contiguous_x.is_contiguous(), "Check failed !");
+
+  auto contiguous_x = non_contiguous_x.contiguous();
+  PD_CHECK(contiguous_x.is_contiguous(), "Check failed !");
+
+  return {contiguous_x};
+}
+
+PD_BUILD_OP(custom_contiguous)
+    .Inputs({"X"})
+    .Outputs({"Out"})
+    .SetKernelFn(PD_KERNEL(ContiguousForward));
diff --git a/test/custom_op/test_custom_contiguous.py b/test/custom_op/test_custom_contiguous.py
@@ -0,0 +1,73 @@
+# Copyright (c) 2023 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
+#
+#     https://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 os
+import unittest
+
+import numpy as np
+from utils import (
+    extra_cc_args,
+    extra_nvcc_args,
+    paddle_includes,
+    paddle_libraries,
+)
+
+import paddle
+from paddle.utils.cpp_extension import get_build_directory, load
+from paddle.utils.cpp_extension.extension_utils import run_cmd
+
+# Because Windows don't use docker, the shared lib already exists in the
+# cache dir, it will not be compiled again unless the shared lib is removed.
+file = f'{get_build_directory()}\\custom_contiguous\\custom_contiguous.pyd'
+if os.name == 'nt' and os.path.isfile(file):
+    cmd = f'del {file}'
+    run_cmd(cmd, True)
+
+custom_module = load(
+    name='custom_contiguous',
+    sources=['custom_contiguous.cc'],
+    extra_include_paths=paddle_includes,  # add for Coverage CI
+    extra_library_paths=paddle_libraries,
+    extra_cxx_cflags=extra_cc_args,  # test for cc flags
+    extra_cuda_cflags=extra_nvcc_args,  # test for nvcc flags
+    verbose=True,
+)
+
+
+def custom_contiguous_dynamic(device, np_x):
+    paddle.set_device(device)
+
+    x = paddle.to_tensor(np_x, dtype="float32")
+    x.stop_gradient = True
+
+    x = x.transpose((1, 0))
+
+    out = custom_module.custom_contiguous(x)
+
+    assert out.is_contiguous()
+
+
+class TestCustomCastOp(unittest.TestCase):
+    def setUp(self):
+        self.dtypes = ['float32', 'float64']
+        paddle.set_flags({"FLAGS_use_stride_kernel": 1})
+
+    def test_dynamic(self):
+        for dtype in self.dtypes:
+            x = np.random.uniform(-1, 1, [4, 8]).astype("float32")
+            custom_contiguous_dynamic('cpu', x)
+
+
+if __name__ == '__main__':
+    unittest.main()
