[Cherry-pick Fleety_12] Slice fix (#75596、#75794、#76004)

paddle/fluid/pybind/eager_method.cc

@@ -1434,7 +1434,8 @@ static PyObject* tensor_method_set_underline_tensor(TensorObject* self,
           static_cast<phi::DenseTensor*>(self->tensor.impl().get());
       if (self->tensor.has_allocation() && self->tensor.initialized() &&
           (!dst_tensor->meta().is_contiguous() ||
-           !src_tensor->meta().is_contiguous())) {
+           !src_tensor->meta().is_contiguous()) &&
+          dst_tensor->place().GetType() == src_tensor->place().GetType()) {
         VLOG(8) << "set_tensor() method , src or dst tensor is not contiguous ";
         if (!FLAGS_use_stride_kernel) {
           PADDLE_THROW(common::errors::Fatal(
@@ -1451,6 +1452,17 @@ static PyObject* tensor_method_set_underline_tensor(TensorObject* self,
                   dst_tensor);
             }));
       } else {
+        if (!dst_tensor->meta().is_contiguous()) {
+          PADDLE_THROW(common::errors::Fatal(
+              "dst_tensor is not contiguous and src_tesnor has different place "
+              "with dst_tensor, so Strided kernel "
+              "can't be called, please change src_tensor'place as same as "
+              "dst_tensor'place or change dst_tensor to be contiguous"));
+        } else if (!src_tensor->meta().is_contiguous()) {
+          VLOG(6) << "src_tensor is not contiguous, so dst_tensor will be not "
+                     "contiguous after set_value ";
+        }
+
         if (dst_tensor->place().GetType() != phi::AllocationType::UNDEFINED) {
           framework::TensorCopy(*src_tensor, dst_tensor->place(), dst_tensor);
         } else if (src_tensor->place().GetType() !=

paddle/fluid/pybind/slice_utils.h

@@ -802,7 +802,7 @@ static paddle::Tensor getValueForBoolTensor(const paddle::Tensor& tensor,
   }
 
   auto bool_2_idx = nonzero_ad_func(bool_index);
-  if (FLAGS_use_stride_kernel) {
+  if (FLAGS_use_stride_kernel && self_tensor.is_contiguous()) {
     std::vector<paddle::Tensor> indices =
         PrepareIndices(tensor, bool_2_idx, bool_index);
     for (int i = 0; i < pos_of_new_dim; ++i) {
@@ -1302,7 +1302,8 @@ static void ApplyGetitem(const int index_size,
       }
     }
 
-    if (FLAGS_use_stride_kernel && !has_empty_index) {
+    if (FLAGS_use_stride_kernel && !has_empty_index &&
+        self_tensor->is_contiguous()) {
       const phi::distributed::ProcessMesh* mesh = nullptr;
       if (InputsContainDistTensor(
               &mesh, *self_tensor, *transed_tensor, *transed_index)) {

paddle/phi/kernels/funcs/stride_utils.h

@@ -330,7 +330,7 @@ static inline void CopyStride(
 
   coalesce_dimensions<N>(ndim, strides_array, &stride_size, desired_shape);
 
-  int num = 1;
+  int64_t num = 1;
   for (size_t i = 0; i < desired_shape->size(); i++) {
     num *= (*desired_shape)[i];
   }
@@ -385,7 +385,7 @@ static inline void IndexPutStride(
 
   coalesce_dimensions<N>(ndim, strides_array, &stride_size, desired_shape);
 
-  int num = 1;
+  int64_t num = 1;
   for (size_t i = 0; i < desired_shape->size(); i++) {
     num *= (*desired_shape)[i];
   }
@@ -444,7 +444,7 @@ static inline void IndexGetStride(
 
   coalesce_dimensions<N>(ndim, strides_array, &stride_size, desired_shape);
 
-  int num = 1;
+  int64_t num = 1;
   for (size_t i = 0; i < desired_shape->size(); i++) {
     num *= (*desired_shape)[i];
   }
@@ -539,8 +539,8 @@ static inline void ScatterAddStride(
 
   coalesce_dimensions<N>(ndim, strides_array, &stride_size, desired_shape);
 
-  int num = 1;
-  for (int i = 0; i < desired_shape->size(); i++) {
+  int64_t num = 1;
+  for (size_t i = 0; i < desired_shape->size(); i++) {
     num *= (*desired_shape)[i];
   }
   *numel = num;

test/indexing/test_getitem.py

@@ -409,6 +409,24 @@ def test_indexing_is_boolean_false(self):
 
         np.testing.assert_allclose(y.numpy(), np_res)
 
+    def test_input_strided_tensor(self):
+        base = paddle.to_tensor(
+            [5.0, 5.0, 6.0, 5.0, 5.0, 6.0], dtype=paddle.float64
+        )
+        foo_strided = paddle.as_strided(base, shape=(2, 1), stride=(2, 1))
+
+        base2 = paddle.to_tensor(
+            [0, 0, 1, 0, 1, 0, 0, 5, 5, 5, 5], dtype=paddle.int64
+        )
+        atype = paddle.as_strided(base2, shape=(2, 3), stride=(4, 1))
+
+        result = foo_strided[atype]
+        expected_result = paddle.to_tensor(
+            [[[5.0], [5.0], [6.0]], [[6.0], [5.0], [5.0]]], dtype=paddle.float64
+        )
+
+        np.testing.assert_allclose(result.numpy(), expected_result.numpy())
+
 
 class TestMultipleIndexing(TestGetitemInDygraph):
     def test_indexing_with_all_possible_start_end_step_dygraph(self):

test/legacy_test/test_set_value_op.py

@@ -17,7 +17,11 @@
 import unittest
 
 import numpy as np
-from op_test import OpTest, convert_float_to_uint16, get_devices
+from op_test import (
+    OpTest,
+    convert_float_to_uint16,
+    get_devices,
+)
 
 import paddle
 from paddle.base import core