Add gather_nd tests (PaddlePaddle#1389)

* add gather_nd tests

* fix bug

* fix bug

python/tests/ops/test_gather_nd_op.py

@@ -14,76 +14,85 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import sys
 import unittest
 import numpy as np
 from op_test import OpTest, OpTestTool
 import paddle
+import cinn
 from cinn.frontend import *
 from cinn.common import *
+import logging
+import os
+from itertools import product
+
+logging.basicConfig(level=os.environ.get('LOG_LEVEL', 'INFO').upper())
+logger = logging.getLogger(name="gather_nd")
 
 
 @OpTestTool.skip_if(not is_compiled_with_cuda(),
                     "x86 test will be skipped due to timeout.")
 class TestGatherNdOp(OpTest):
     def setUp(self):
+        self.data = []
         self.init_case()
 
     def init_case(self):
-        self.inputs = {
-            'x': self.random([2, 3, 4], 'float32'),
-            'index': np.array([[1]], dtype='int32')
-        }
+        self.inputs = [{"x": [3, 4, 3], "index": [4, 1]}]
+        self.dtypes = ["float32"]
 
     def build_paddle_program(self, target):
-        x = paddle.to_tensor(self.inputs["x"], stop_gradient=False)
-        index = paddle.to_tensor(self.inputs["index"], stop_gradient=False)
-        out = paddle.gather_nd(x, index)
-        self.paddle_outputs = [out]
+        for inputs, dtype in product(self.inputs, self.dtypes):
+            x_shape = inputs["x"]
+            index_shape = inputs["index"]
+            x = np.random.randn(*x_shape).astype(dtype)
+            index = np.random.randint(0, min(x_shape),
+                                      index_shape).astype("int32")
+            self.data.append([x, index])
+            x = paddle.to_tensor(x, stop_gradient=False)
+            index = paddle.to_tensor(index, stop_gradient=False)
+            out = paddle.gather_nd(x, index)
+            logger.debug(" -- The output of Paddle:\n{}".format(out))
+            self.paddle_outputs.append(out)
 
     def build_cinn_program(self, target):
-        builder = NetBuilder("GatherNd")
-        x = builder.create_input(
-            self.nptype2cinntype(self.inputs["x"].dtype),
-            self.inputs["x"].shape, "x")
-        index = builder.create_input(
-            self.nptype2cinntype(self.inputs["index"].dtype),
-            self.inputs["index"].shape, "index")
-        out = builder.gather_nd(x, index)
-
-        prog = builder.build()
-        res = self.get_cinn_output(prog, target, [x, index],
-                                   [self.inputs["x"], self.inputs["index"]],
-                                   [out])
-
-        self.cinn_outputs = [res[0]]
+        for i, (inputs, dtype) in enumerate(product(self.inputs, self.dtypes)):
+            builder = NetBuilder("gather")
+            x = builder.create_input(
+                self.nptype2cinntype(dtype), inputs["x"], "x")
+            index = builder.create_input(Int(32), inputs["index"], "index")
+            out = builder.gather_nd(x, index)
+            prog = builder.build()
+            res = self.get_cinn_output(prog, target, [x, index], self.data[i],
+                                       [out])
+            logger.debug(" -- The output of CINN:\n{}".format(res))
+            self.cinn_outputs.extend(res)
 
     def test_check_results(self):
         self.check_outputs_and_grads(all_equal=True)
 
 
-class TestGatherNdCase1(TestGatherNdOp):
-    def init_case(self):
-        self.inputs = {
-            'x': self.random([2, 3, 4], 'float32'),
-            'index': np.array([[0, 2]], dtype='int32')
-        }
-
-
-class TestGatherNdCase2(TestGatherNdOp):
-    def init_case(self):
-        self.inputs = {
-            'x': self.random([2, 3, 4], 'float32'),
-            'index': np.array([[1, 2, 3]], dtype='int32')
-        }
-
-
-class TestGatherNdCase3(TestGatherNdOp):
+class TestGatherOpAll(TestGatherNdOp):
     def init_case(self):
-        self.inputs = {
-            'x': self.random([2, 3, 4], 'float64'),
-            'index': np.array([[1, 2, 3]], dtype='int64')
-        }
+        self.inputs = []
+        for x_shape in [
+            [16],
+            [8, 16],
+            [4, 8, 16],
+            [2, 4, 8, 16],
+            [2, 4, 8, 1],
+            [2, 4, 8, 1024],
+        ]:
+            for j in range(1, len(x_shape)):
+                self.inputs.append({"x": x_shape, "index": [8, j]})
+
+        self.dtypes = [
+            "float32",
+            "float64",
+            "int16",
+            "int32",
+            "int64",
+            # "uint8"  # note: some types is not supported in paddle now.
+        ]
 
 
 if __name__ == "__main__":