Added sigmoid BF16 FWD/BWD kernels and gelu BF16 BWD kernel (#34216)

* added sigmoid BF16 FWD/BWD and gelu BF16 BWD

* added newline at EOF

* switched from lambdas to local functions

* changed function names

paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc

@@ -251,7 +251,9 @@ namespace ops = paddle::operators;
       ops::MKLDNNActivationKernel<ops::functor<paddle::platform::bfloat16>>); \
   REGISTER_OP_KERNEL(                                                         \
       act_type##_grad, MKLDNN, ::paddle::platform::CPUPlace,                  \
-      ops::MKLDNNActivationGradKernel<ops::grad_functor<float>>);
+      ops::MKLDNNActivationGradKernel<ops::grad_functor<float>>,              \
+      ops::MKLDNNActivationGradKernel<                                        \
+          ops::grad_functor<paddle::platform::bfloat16>>);
 
 #define FOR_EACH_MKLDNN_KERNEL_FUNCTOR(__macro)                           \
   __macro(relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor);                \
@@ -259,7 +261,6 @@ namespace ops = paddle::operators;
   __macro(leaky_relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor);          \
   __macro(swish, SwishMKLDNNFunctor, SwishMKLDNNGradFunctor);             \
   __macro(hardswish, HardSwishMKLDNNFunctor, HardSwishMKLDNNGradFunctor); \
-  __macro(sigmoid, SigmoidMKLDNNFunctor, SigmoidMKLDNNGradFunctor);       \
   __macro(tanh, TanhMKLDNNFunctor, TanhMKLDNNGradFunctor);                \
   __macro(sqrt, SqrtMKLDNNFunctor, SqrtMKLDNNGradFunctor);                \
   __macro(abs, AbsMKLDNNFunctor, AbsMKLDNNGradFunctor);
@@ -267,3 +268,5 @@ namespace ops = paddle::operators;
 FOR_EACH_MKLDNN_KERNEL_FUNCTOR(REGISTER_ACTIVATION_MKLDNN_KERNEL);
 REGISTER_ACTIVATION_MKLDNN_BF16_KERNEL(gelu, GeluMKLDNNFunctor,
                                        GeluMKLDNNGradFunctor);
+REGISTER_ACTIVATION_MKLDNN_BF16_KERNEL(sigmoid, SigmoidMKLDNNFunctor,
+                                       SigmoidMKLDNNGradFunctor);

python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py

@@ -16,9 +16,9 @@ from __future__ import print_function
 
 import unittest
 import numpy as np
-from scipy.special import expit
+from scipy.special import expit, erf
 import paddle.fluid.core as core
-from paddle.fluid.tests.unittests.op_test import OpTest, convert_float_to_uint16
+from paddle.fluid.tests.unittests.op_test import OpTest, OpTestTool, convert_float_to_uint16
 from paddle.fluid.tests.unittests.test_activation_op import TestActivation, TestRelu, TestTanh, TestSqrt, TestAbs, TestLeakyRelu, TestSwish, TestHardSwish, TestRelu6, TestSigmoid
 from paddle.fluid.tests.unittests.test_gelu_op import gelu
 from mkldnn_op_test import check_if_mkldnn_primitives_exist_in_bwd
@@ -79,46 +79,88 @@ class TestMKLDNNGeluDim2Approx(TestActivation):
         self.attrs = {"use_mkldnn": True, "approximate": True}
 
 
-@unittest.skipIf(not core.supports_bfloat16(),
-                 "place does not support BF16 evaluation")
-class TestMKLDNNGeluBf16Dim2(TestActivation):
+#Use it as a base class for BF16 activation tests, just override necessary functions
+class TestMKLDNNSigmoidBF16Op(TestActivation):
+    @OpTestTool.skip_if_not_cpu_bf16()
+    def config(self):
+        self.op_type = "sigmoid"
+
+    def op_forward(self, x):
+        return 1 / (1 + np.exp(-x))
+
+    def op_grad(self, dout, x):
+        return dout * self.op_forward(x) * (1 - self.op_forward(x))
+
+    def set_attrs(self):
+        self.attrs = {"use_mkldnn": True}
+
+    def init_data(self):
+        self.x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype(np.float32)
+
     def setUp(self):
-        self.op_type = "gelu"
         self.dtype = np.uint16
+        self.init_data()
+        self.config()
+        self.out = self.op_forward(self.x)
 
-        x = np.random.uniform(-1, 1, [11, 17]).astype(np.float32)
-        out = convert_float_to_uint16(gelu(x, False))
+        self.inputs = {'X': convert_float_to_uint16(self.x)}
+        self.outputs = {'Out': self.out}
+        self.set_attrs()
 
-        self.inputs = {'X': convert_float_to_uint16(x)}
-        self.outputs = {'Out': out}
-        self.attrs = {"use_mkldnn": True}
+    def calculate_grads(self):
+        self.dx = self.op_grad(self.out, self.x)
 
     def test_check_output(self):
         self.check_output_with_place(core.CPUPlace())
 
     def test_check_grad(self):
-        pass
+        self.calculate_grads()
+        self.check_grad_with_place(
+            core.CPUPlace(), ["X"],
+            "Out",
+            user_defined_grads=[self.dx],
+            user_defined_grad_outputs=[convert_float_to_uint16(self.out)])
 
 
-@unittest.skipIf(not core.supports_bfloat16(),
-                 "place does not support BF16 evaluation")
-class TestMKLDNNGeluBf16Dim2Approx(TestActivation):
-    def setUp(self):
+class TestMKLDNNGeluErfBF16Op(TestMKLDNNSigmoidBF16Op):
+    def config(self):
         self.op_type = "gelu"
-        self.dtype = np.uint16
 
-        x = np.random.uniform(-1, 1, [11, 17]).astype(np.float32)
-        out = convert_float_to_uint16(gelu(x, True))
+    def op_forward(self, x):
+        return gelu(x, False)
 
-        self.inputs = {'X': convert_float_to_uint16(x)}
-        self.outputs = {'Out': out}
+    def op_grad(self, dout, x):
+        return (dout *
+                (0.5 + 0.5 * erf(x / np.sqrt(2)) +
+                 (x / np.sqrt(2 * np.pi) * np.exp(-0.5 * np.power(x, 2)))))
+
+
+class TestMKLDNNGeluErfDim2BF16Op(TestMKLDNNGeluErfBF16Op):
+    def init_data(self):
+        self.x = np.random.uniform(-1, 1, [11, 17]).astype(np.float32)
+
+
+class TestMKLDNNGeluTanhBF16Op(TestMKLDNNSigmoidBF16Op):
+    def config(self):
+        self.op_type = "gelu"
+
+    def op_forward(self, x):
+        return gelu(x, True)
+
+    def op_grad(self, dout, x):
+        grad_part = np.tanh(
+            np.sqrt(2 / np.pi) * (x + 0.044715 * np.power(x, 3)))
+        return dout * 0.5 * (1 + grad_part) * (1 + np.sqrt(2 / np.pi) *
+                                               (x + 0.134145 * np.power(x, 3)) *
+                                               (1 - grad_part))
+
+    def set_attrs(self):
         self.attrs = {"use_mkldnn": True, "approximate": True}
 
-    def test_check_output(self):
-        self.check_output_with_place(core.CPUPlace())
 
-    def test_check_grad(self):
-        pass
+class TestMKLDNNGeluTanhDim2BF16Op(TestMKLDNNGeluTanhBF16Op):
+    def init_data(self):
+        self.x = np.random.uniform(-1, 1, [11, 17]).astype(np.float32)
 
 
 class TestMKLDNNTanhDim2(TestTanh):

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

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 from scipy.special import expit, erf
 
-from op_test import OpTest, convert_float_to_uint16
+from paddle.fluid.tests.unittests.op_test import OpTest, convert_float_to_uint16, skip_check_grad_ci
 import paddle
 import paddle.nn as nn
 import paddle.nn.functional as F
@@ -1619,7 +1619,6 @@ class TestHardSwish(TestActivation):
         self.op_type = 'hard_swish'
         self.init_dtype()
 
-        from op_test import skip_check_grad_ci
         skip_check_grad_ci(reason="not implemented yet")
 
         np.random.seed(1024)