【PaddlePaddle Hackathon 4】No.56 : add fp16 test and bf16 for bernoulli and trunc

paddle/phi/kernels/gpu/bernoulli_kernel.cu

@@ -85,5 +85,11 @@ void BernoulliKernel(const Context& ctx,
 
 }  // namespace phi
 
-PD_REGISTER_KERNEL(
-    bernoulli, GPU, ALL_LAYOUT, phi::BernoulliKernel, float, double) {}
+PD_REGISTER_KERNEL(bernoulli,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::BernoulliKernel,
+                   float,
+                   double,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

paddle/phi/kernels/gpu/trunc_grad_kernel.cu

@@ -52,4 +52,6 @@ PD_REGISTER_KERNEL(trunc_grad,
                    float,
                    double,
                    int,
-                   int64_t) {}
+                   int64_t,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

paddle/phi/kernels/gpu/trunc_kernel.cu

@@ -78,5 +78,13 @@ void TruncKernel(const Context& dev_ctx,
 
 }  // namespace phi
 
-PD_REGISTER_KERNEL(
-    trunc, GPU, ALL_LAYOUT, phi::TruncKernel, float, double, int, int64_t) {}
+PD_REGISTER_KERNEL(trunc,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::TruncKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

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

@@ -31,10 +31,15 @@ def output_hist(out):
 class TestBernoulliOp(OpTest):
     def setUp(self):
         self.op_type = "bernoulli"
-        self.inputs = {"X": np.random.uniform(size=(1000, 784))}
+        self.inputs = {
+            "X": np.random.uniform(size=(1000, 784)).astype(self.dtype)
+        }
         self.attrs = {}
         self.outputs = {"Out": np.zeros((1000, 784)).astype("float32")}
 
+    def init_dtype(self):
+        self.dtype = np.float32
+
     def test_check_output(self):
         self.check_output_customized(self.verify_output)
 
@@ -98,5 +103,10 @@ def test_fixed_random_number(self):
         paddle.enable_static()
 
 
+class TestBernoulliFP16OP(TestBernoulliOp):
+    def init_dtype(self):
+        self.dtype = np.float16
+
+
 if __name__ == "__main__":
     unittest.main()

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

@@ -15,9 +15,10 @@
 import unittest
 
 import numpy as np
-from eager_op_test import OpTest
+from eager_op_test import OpTest, convert_float_to_uint16
 
 import paddle
+import paddle.fluid.core as core
 
 paddle.enable_static()
 
@@ -90,5 +91,38 @@ def test_errors(self):
             self.assertRaises(TypeError, paddle.trunc, x)
 
 
+class TestTruncFP16OP(TestTruncOp):
+    def init_dtype_type(self):
+        self.dtype = np.float16
+
+
+@unittest.skipIf(
+    not core.is_compiled_with_cuda()
+    or not core.is_bfloat16_supported(core.CUDAPlace(0)),
+    "core is not complied with CUDA and not support the bfloat16",
+)
+class TestTruncBF16OP(OpTest):
+    def setUp(self):
+        self.op_type = "trunc"
+        self.python_api = paddle.trunc
+        self.init_dtype_type()
+        np.random.seed(2021)
+        x = np.random.random((20, 20)).astype(np.float64)
+        out = np.trunc(x)
+        self.inputs = {'X': convert_float_to_uint16(x)}
+        self.outputs = {'Out': convert_float_to_uint16(out)}
+
+    def init_dtype_type(self):
+        self.dtype = np.uint16
+
+    def test_check_output(self):
+        place = core.CUDAPlace(0)
+        self.check_output_with_place(place)
+
+    def test_check_grad(self):
+        place = core.CUDAPlace(0)
+        self.check_grad_with_place(place, ['X'], 'Out')
+
+
 if __name__ == "__main__":
     unittest.main()