From 2c6bd4ad48093371da81f43892f81281c1d9b04e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?S=C5=82awomir=20Siwek?= <slawomir.siwek@intel.com>
Date: Tue, 11 Oct 2022 15:03:51 +0200
Subject: [PATCH] hard_swish grad (#46857)

---
 .../kernels/onednn/activation_grad_kernel.cc  |  2 +-
 .../mkldnn/test_activation_mkldnn_op.py       | 20 ++++++-------------
 2 files changed, 7 insertions(+), 15 deletions(-)

diff --git a/paddle/phi/kernels/onednn/activation_grad_kernel.cc b/paddle/phi/kernels/onednn/activation_grad_kernel.cc
index 9e183abf028..cc7f71ff364 100644
--- a/paddle/phi/kernels/onednn/activation_grad_kernel.cc
+++ b/paddle/phi/kernels/onednn/activation_grad_kernel.cc
@@ -243,7 +243,7 @@ void HardSwishGradKernel(const Context& dev_ctx,
                          float offset,
                          DenseTensor* dx) {
   HardSwishOneDNNGradFunctor<T> functor;
-  functor(dev_ctx, x, dout, threshold, 0, dx);
+  functor(dev_ctx, x, dout, 0, 0, dx);
 }
 
 template <typename T, typename Context>
diff --git a/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py
index ac880b02c45..b2e58514dca 100644
--- a/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py
+++ b/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py
@@ -128,19 +128,11 @@ class TestMKLDNNSwishDim2(TestSwish):
         self.dtype = np.float32
 
 
-@skip_check_grad_ci(reason="not implemented yet")
 class TestMKLDNNHardSwishDim2(TestHardSwish):
 
     def setUp(self):
         super(TestMKLDNNHardSwishDim2, self).setUp()
-
-        self.attrs["use_mkldnn"] = True
-
-    def init_dtype(self):
-        self.dtype = np.float32
-
-    def test_check_grad(self):
-        pass
+        self.attrs = {"use_mkldnn": True}
 
 
 class TestMKLDNNSigmoidDim2(TestSigmoid):
@@ -317,11 +309,14 @@ class TestMKLDNNSwishDim4(TestSwish):
 
 
 def ref_hardswish(x, threshold=6.0, scale=6.0, offset=3.0):
+    x_dtype = x.dtype
+    if x_dtype == 'float16':
+        x_dtype = 'float16'
+        x = x.astype('float32')
     return (x * np.minimum(np.maximum(x + offset, 0.), threshold) /
-            scale).astype(x.dtype)
+            scale).astype(x_dtype)
 
 
-@skip_check_grad_ci(reason="not implemented yet")
 class TestMKLDNNHardSwishDim4(TestHardSwish):
 
     def setUp(self):
@@ -343,9 +338,6 @@ class TestMKLDNNHardSwishDim4(TestHardSwish):
     def init_dtype(self):
         self.dtype = np.float32
 
-    def test_check_grad(self):
-        pass
-
 
 class TestMKLDNNMish(TestActivation):
 
-- 
GitLab