Add full_like composite rule (#50794)

* implement composite full_like and simple unit test

* implement op tests for composite full_like op

* some modification as reviewers suggested
add cinn op test to CMakeLists.txt
fix code style

* fix code style

* modify input args of prim fill_any_like op

* resolve conflicts

* resolve conflicts

* modify python api and unit tests as suggested

* resolve conflicts

* resolve conflicts

* use framework.dtype to convert dtype in Op test

python/paddle/fluid/layers/tensor.py

@@ -147,6 +147,7 @@ def fill_constant(shape, dtype, value, force_cpu=False, out=None, name=None):
                 'int64',
                 'complex64',
                 'complex128',
+                'uint16',
             ],
             'fill_constant',
         )

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -1202,8 +1202,14 @@ if($ENV{USE_STANDALONE_EXECUTOR})
                        PROPERTIES ENVIRONMENT FLAGS_USE_STANDALONE_EXECUTOR=0)
 endif()
 
-set(TEST_CINN_OPS test_softmax_op test_expand_v2_op test_reduce_op
-                  test_slice_op test_activation_op)
+set(TEST_CINN_OPS
+    test_softmax_op
+    test_expand_v2_op
+    test_reduce_op
+    test_slice_op
+    test_activation_op
+    test_full_like_op
+    test_fill_any_like_op)
 
 foreach(TEST_CINN_OPS ${TEST_CINN_OPS})
   if(WITH_CINN)

python/paddle/fluid/tests/unittests/test_fill_any_like_op.py

@@ -19,24 +19,40 @@ from eager_op_test import OpTest, convert_float_to_uint16
 
 import paddle
 import paddle.fluid.core as core
+import paddle.framework.dtype as dtypes
+
+
+def fill_any_like_wrapper(x, value, out_dtype=None, name=None):
+    if isinstance(out_dtype, int):
+        tmp_dtype = dtypes.dtype(out_dtype)
+    elif out_dtype == np.complex64:
+        raise ValueError("Not supported dtype %s" % out_dtype)
+    else:
+        tmp_dtype = out_dtype
+    return paddle.full_like(x, value, tmp_dtype, name)
 
 
 class TestFillAnyLikeOp(OpTest):
     def setUp(self):
         self.op_type = "fill_any_like"
-        self.python_api = paddle.full_like
+        self.prim_op_type = "comp"
+        self.python_api = fill_any_like_wrapper
         self.dtype = np.int32
         self.value = 0.0
         self.init()
         self.inputs = {'X': np.random.random((219, 232)).astype(self.dtype)}
         self.attrs = {'value': self.value}
         self.outputs = {'Out': self.value * np.ones_like(self.inputs["X"])}
+        self.skip_cinn()
 
     def init(self):
         pass
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(check_prim=True)
+
+    def skip_cinn(self):
+        pass
 
 
 class TestFillAnyLikeOpFloat32(TestFillAnyLikeOp):
@@ -44,6 +60,9 @@ class TestFillAnyLikeOpFloat32(TestFillAnyLikeOp):
         self.dtype = np.float32
         self.value = 0.0
 
+    def skip_cinn(self):
+        self.enable_cinn = False
+
 
 @unittest.skipIf(
     not core.is_compiled_with_cuda(), "core is not compiled with CUDA"
@@ -51,7 +70,8 @@ class TestFillAnyLikeOpFloat32(TestFillAnyLikeOp):
 class TestFillAnyLikeOpBfloat16(OpTest):
     def setUp(self):
         self.op_type = "fill_any_like"
-        self.python_api = paddle.full_like
+        self.prim_op_type = "comp"
+        self.python_api = fill_any_like_wrapper
         self.dtype = np.uint16
         self.value = 0.0
         self.inputs = {'X': np.random.random((219, 232)).astype(np.float32)}
@@ -61,31 +81,45 @@ class TestFillAnyLikeOpBfloat16(OpTest):
                 self.value * np.ones_like(self.inputs["X"])
             )
         }
+        self.skip_cinn()
 
     def test_check_output(self):
         place = core.CUDAPlace(0)
-        self.check_output_with_place(place)
+        self.check_output_with_place(place, check_prim=True)
+
+    def skip_cinn(self):
+        self.enable_cinn = False
 
 
 class TestFillAnyLikeOpValue1(TestFillAnyLikeOp):
     def init(self):
         self.value = 1.0
 
+    def skip_cinn(self):
+        self.enable_cinn = False
+
 
 class TestFillAnyLikeOpValue2(TestFillAnyLikeOp):
     def init(self):
         self.value = 1e-10
 
+    def skip_cinn(self):
+        self.enable_cinn = False
+
 
 class TestFillAnyLikeOpValue3(TestFillAnyLikeOp):
     def init(self):
         self.value = 1e-100
 
+    def skip_cinn(self):
+        self.enable_cinn = False
+
 
 class TestFillAnyLikeOpType(TestFillAnyLikeOp):
     def setUp(self):
         self.op_type = "fill_any_like"
-        self.python_api = paddle.full_like
+        self.prim_op_type = "comp"
+        self.python_api = fill_any_like_wrapper
         self.dtype = np.int32
         self.value = 0.0
         self.init()
@@ -99,11 +133,19 @@ class TestFillAnyLikeOpType(TestFillAnyLikeOp):
             * np.ones_like(self.inputs["X"]).astype(np.float32)
         }
 
+        self.skip_cinn()
+
+    def skip_cinn(self):
+        self.enable_cinn = False
+
 
 class TestFillAnyLikeOpFloat16(TestFillAnyLikeOp):
     def init(self):
         self.dtype = np.float16
 
+    def skip_cinn(self):
+        self.enable_cinn = False
+
 
 if __name__ == "__main__":
     paddle.enable_static()

python/paddle/fluid/tests/unittests/test_full_like_op.py

@@ -19,10 +19,21 @@ from op_test import OpTest
 
 import paddle
 import paddle.fluid.core as core
+import paddle.framework.dtype as dtypes
 from paddle.fluid.framework import convert_np_dtype_to_dtype_
 from paddle.static import Program, program_guard
 
 
+def fill_any_like_wrapper(x, value, out_dtype=None, name=None):
+    if isinstance(out_dtype, int):
+        tmp_dtype = dtypes.dtype(out_dtype)
+    elif out_dtype == np.complex64:
+        raise ValueError("Not supported dtype %s" % out_dtype)
+    else:
+        tmp_dtype = out_dtype
+    return paddle.full_like(x, value, tmp_dtype, name)
+
+
 class TestFullOp(unittest.TestCase):
     """Test fill_any_like op(whose API is full_like) for attr out."""
 
@@ -100,8 +111,10 @@ class TestFullLikeOp1(OpTest):
     # test basic
     def setUp(self):
         self.op_type = "fill_any_like"
-        self.python_api = paddle.full_like
+        self.prim_op_type = "comp"
+        self.python_api = fill_any_like_wrapper
         self.init_data()
+        self.skip_cinn()
 
         x = np.zeros(self.shape)
         out = np.full_like(x, self.fill_value, self.dtype)
@@ -119,7 +132,10 @@ class TestFullLikeOp1(OpTest):
         self.dtype = np.float32
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=True, check_prim=True)
+
+    def skip_cinn(self):
+        pass
 
 
 class TestFullLikeOp2(TestFullLikeOp1):
@@ -128,6 +144,9 @@ class TestFullLikeOp2(TestFullLikeOp1):
         self.shape = [1024, 1024]
         self.dtype = np.float64
 
+    def skip_cinn(self):
+        self.enable_cinn = False
+
 
 class TestFullLikeOp3(TestFullLikeOp1):
     def init_data(self):
@@ -135,6 +154,9 @@ class TestFullLikeOp3(TestFullLikeOp1):
         self.shape = [5000, 5000]
         self.dtype = np.int64
 
+    def skip_cinn(self):
+        self.enable_cinn = False
+
 
 @unittest.skipIf(
     not core.is_compiled_with_cuda(), "core is not compiled with CUDA"

python/paddle/incubate/autograd/composite_rules.py

@@ -20,6 +20,7 @@
 import functools
 import operator
 
+import paddle.framework.dtype as dtypes
 from paddle.fluid import core
 
 from .primitives import *  # noqa: F403
@@ -269,3 +270,13 @@ def silu_composite(x):
     sum_temp = 1 + exp(-x)
     res = x / sum_temp
     return res
+
+
+@REGISTER_COMPOSITE('fill_any_like')
+def fill_any_like(x, fill_value, dtype, place=None):
+    """define composite rule of op full_like."""
+    """op name: full_like  op type name: fill_any_like."""
+    """arg place is not used, add it here to keep same as python api."""
+    dtype = dtypes.dtype(dtype)
+    val = full(x.shape, fill_value, dtype)
+    return val

python/paddle/tensor/creation.py

@@ -827,6 +827,7 @@ def full_like(x, fill_value, dtype=None, name=None):
                 'int16',
                 'int32',
                 'int64',
+                'uint16',
             ],
             'full_like',
         )
@@ -841,6 +842,7 @@ def full_like(x, fill_value, dtype=None, name=None):
                 'int16',
                 'int32',
                 'int64',
+                'uint16',
             ],
             'full_like/zeros_like/ones_like',
         )