[Zero-Dim] Support 0-D tensor for some oneDNN unary kernels #51687

YangQun1 · 2023-03-15T05:26:12Z

PR types

New features

PR changes

OPs

Describe

Partially solve #51364, support 0-d tensor for the following onednn elementwise kernels:

kernel	0D Usage need to be supported	status
abs	element-wise unary, 0D->0D	DONE
elu	element-wise unary, 0D->0D	DONE
exp	element-wise unary, 0D->0D	DONE
gelu	element-wise unary, 0D->0D	DONE
hardswish	element-wise unary, 0D->0D	DONE
leaky_relu	element-wise unary, 0D->0D	DONE
mish	element-wise unary, 0D->0D	DONE
relu	element-wise unary, 0D->0D	DONE
relu6	element-wise unary, 0D->0D	DONE
sigmoid	element-wise unary, 0D->0D	DONE
sqrt	element-wise unary, 0D->0D	DONE
swish	element-wise unary, 0D->0D	DONE
tanh	element-wise unary, 0D->0D	DONE
round	element-wise unary, 0D->0D	DONE
softplus	element-wise unary, 0D->0D	DONE
cast	element-wise unary, 0D->0D	DONE
clip	element-wise unary, 0D->0D	DONE
scale	element-wise unary, 0D->0D	DONE
softmax	when input 0D, outputs is 0D and value is 1.0, grad is 0D and value is 0.0	DONE
log_softmax	when input 0D, outputs is 0D and value is 0.0, grad is 0D and value is 0.0	DONE

paddle-bot · 2023-03-15T05:26:16Z

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

CLAassistant · 2023-03-15T05:26:19Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

CLAassistant · 2023-03-15T05:26:20Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

tsocha · 2023-03-17T10:54:25Z

paddle/phi/kernels/funcs/data_layout_transform.cc

@@ -65,7 +65,7 @@ void TransDataLayoutFromOneDNN(DataLayout in_layout,
  auto& cpu_engine = dev_ctx->GetEngine();

  auto in_tz = vectorize<int64_t>(in.dims());
-  auto out_tz = in_tz;
+  auto out_tz = in_tz.size() != 0 ? in_tz : std::vector<int64_t>{1};


This fragment of code is so common that I think it's worthy to make it a small helper function. Later if oneDNN will add support 0D-dim tensors we will have only one place to update.

added, pls help to review

tsocha · 2023-03-17T10:56:14Z

paddle/phi/kernels/onednn/log_softmax_kernel.cc

+    int rank = x.dims().size() != 0 ? x.dims().size() : 1;
+    const int canonical_axis = funcs::CanonicalAxis(axis, rank);


as above.
BTW rank can also be const in this case.

tsocha · 2023-03-17T11:01:33Z

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

+        out = (
+            np.apply_along_axis(ref_log_softmax, self.axis, x)
+            if len(self.shape) > 0
+            else np.array(0.0).astype(self.dtype)
+        )


Suggested change

out = (

np.apply_along_axis(ref_log_softmax, self.axis, x)

if len(self.shape) > 0

else np.array(0.0).astype(self.dtype)

)

out = (

np.apply_along_axis(ref_log_softmax, self.axis, x)

if len(self.shape) < 0

np.array(0.0).astype(self.dtype)

)

why we can't use python's ternary operator?

Ok, now I see what you did here.
Because of its multi-lining I got confused :)

tsocha · 2023-03-17T11:04:01Z

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

@@ -59,6 +68,7 @@ def test_check_grad(self):
        self.check_grad(['X'], 'Out')


+# FIXME(xx) no use_mkldnn attr, does this case run into oneDNN?


Please remove this comment. It looks like: "I will never fix it".

tsocha · 2023-03-17T11:04:05Z

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

@@ -77,6 +87,7 @@ def test_check_grad(self):
        self.check_grad(['X'], 'Out')


+# FIXME(xx) no use_mkldnn attr, does this case run into oneDNN?


Please remove this comment. It looks like: "I will never fix it".

tsocha · 2023-03-17T11:05:09Z

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

@@ -122,6 +122,7 @@ def setUp(self):
        self.use_mkldnn = False
        # explicilty use float32 for ROCm, as MIOpen does not yet support float64
        self.dtype = np.float32 if core.is_compiled_with_rocm() else np.float64
+        self.init_kernel_type()


Do you know why this line is needed now?

I overwrite this method in the derived class, to reset the self.use_mkldnn and self.dtype member.

https://github.com/PaddlePaddle/Paddle/pull/51687/files/1870101dcd9d61dead302a84d1d1c3b5a2d6e4a7#diff-33756190a3e238a856bc0bb361cd700e50832b9fd63686b5ce0b3e122b2ae51eR127

xinyu-intel · 2023-03-20T01:50:15Z

@zhouwei25 Can you please take a review? Thanks:)

zhwesky2010

I have no much problem, LGTM

zhwesky2010 · 2023-03-20T08:09:47Z

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

+        self.python_api = F.gelu
+        self.dtype = np.float32
+
+        x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype)


这个case还是0D不

Thank you, has fixed it now.

zhwesky2010 · 2023-03-20T08:10:32Z

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

+        self.python_api = F.mish
+        self.dtype = np.float32
+
+        x = np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype(self.dtype)


这个case还是0D不

zhwesky2010 · 2023-03-20T08:11:03Z

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

+        self.python_api = F.elu
+        self.set_alpha()
+
+        x = np.random.random((5, 5, 4)).astype("float32")


这个case还是0D不（is this case still 0D?）

zhwesky2010 · 2023-03-20T08:11:15Z

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

+    def setUp(self):
+        self.op_type = "exp"
+        self.python_api = paddle.exp
+        x = np.random.random((5, 5, 4)).astype("float32")


is this case still 0D?

tsocha · 2023-03-20T08:47:19Z

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

+        out = (
+            np.apply_along_axis(ref_log_softmax, self.axis, x)
+            if len(self.shape) > 0
+            else np.array(0.0).astype(self.dtype)
+        )


Ok, now I see what you did here.
Because of its multi-lining I got confused :)

jczaja · 2023-03-20T09:22:01Z

@YangQun1 We have finished validation of this PR on our internal testing system and it passes, so this PR should be functionally correct according to our workload tests.

jczaja

LGTM

paddle-bot · 2023-03-22T12:34:39Z

你的PR已合入Paddle库，请关注后续测试结果。
Your PR has been merged into the repository. An official integration test will be conducted later. Stay tuned.

paddle-bot bot added contributor External developers status: proposed labels Mar 15, 2023

YangQun1 force-pushed the yangqun/eltwise-op-support-0d-tensor branch from d9c4044 to 5d7efbd Compare March 15, 2023 05:32

YangQun1 marked this pull request as draft March 15, 2023 07:48

YangQun1 added 4 commits March 15, 2023 20:21

support 0-d tensor for element wise unary ops

f630bee

fix python code style check

7274eda

fix approval check

13d6ee3

support 0-d tensor for onednn softmax and logsoftmax kernels

1870101

YangQun1 force-pushed the yangqun/eltwise-op-support-0d-tensor branch from d366376 to 1870101 Compare March 15, 2023 14:13

YangQun1 changed the title ~~[Zero-Dim] Support 0-D tensor for oneDNN elementwise kernels~~ [Zero-Dim] Support 0-D tensor for some oneDNN unary kernels Mar 15, 2023

jczaja added the Intel label Mar 17, 2023

YangQun1 marked this pull request as ready for review March 17, 2023 08:41

jczaja requested review from Silv3S, tsocha and jczaja March 17, 2023 10:46

tsocha suggested changes Mar 17, 2023

View reviewed changes

fix commnets

8c54371

zhwesky2010 reviewed Mar 20, 2023

View reviewed changes

fix some unittests

edbd918

tsocha approved these changes Mar 20, 2023

View reviewed changes

jczaja approved these changes Mar 20, 2023

View reviewed changes

onecatcn assigned weishengying Mar 21, 2023

zhwesky2010 approved these changes Mar 22, 2023

View reviewed changes

jczaja merged commit 2a3d75b into PaddlePaddle:develop Mar 22, 2023

paddle-bot bot removed the status: proposed label Mar 22, 2023

paddle-bot bot added the status: accepted label Mar 22, 2023

YangQun1 mentioned this pull request Mar 30, 2023

[Zero-Dim] support 0-D tensor for reduce/reshape/stack/prelu/expand_v2/gaussion onednn kernels #52185

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Zero-Dim] Support 0-D tensor for some oneDNN unary kernels #51687

[Zero-Dim] Support 0-D tensor for some oneDNN unary kernels #51687

YangQun1 commented Mar 15, 2023 •

edited

Loading

paddle-bot bot commented Mar 15, 2023

CLAassistant commented Mar 15, 2023 •

edited

Loading

CLAassistant commented Mar 15, 2023

tsocha Mar 17, 2023

YangQun1 Mar 19, 2023

tsocha Mar 17, 2023

YangQun1 Mar 19, 2023

tsocha Mar 17, 2023

YangQun1 Mar 17, 2023

tsocha Mar 20, 2023

tsocha Mar 17, 2023

YangQun1 Mar 19, 2023

tsocha Mar 17, 2023

YangQun1 Mar 19, 2023

tsocha Mar 17, 2023

YangQun1 Mar 17, 2023

xinyu-intel commented Mar 20, 2023

zhwesky2010 left a comment

zhwesky2010 Mar 20, 2023

YangQun1 Mar 20, 2023

zhwesky2010 Mar 20, 2023

YangQun1 Mar 20, 2023

zhwesky2010 Mar 20, 2023

YangQun1 Mar 20, 2023

zhwesky2010 Mar 20, 2023

YangQun1 Mar 20, 2023

tsocha Mar 20, 2023

jczaja commented Mar 20, 2023

jczaja left a comment

paddle-bot bot commented Mar 22, 2023

		int rank = x.dims().size() != 0 ? x.dims().size() : 1;
		const int canonical_axis = funcs::CanonicalAxis(axis, rank);

		@@ -59,6 +68,7 @@ def test_check_grad(self):
		self.check_grad(['X'], 'Out')


		# FIXME(xx) no use_mkldnn attr, does this case run into oneDNN?

		@@ -77,6 +87,7 @@ def test_check_grad(self):
		self.check_grad(['X'], 'Out')


		# FIXME(xx) no use_mkldnn attr, does this case run into oneDNN?

[Zero-Dim] Support 0-D tensor for some oneDNN unary kernels #51687

[Zero-Dim] Support 0-D tensor for some oneDNN unary kernels #51687

Conversation

YangQun1 commented Mar 15, 2023 • edited Loading

PR types

PR changes

Describe

paddle-bot bot commented Mar 15, 2023

CLAassistant commented Mar 15, 2023 • edited Loading

CLAassistant commented Mar 15, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

xinyu-intel commented Mar 20, 2023

zhwesky2010 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jczaja commented Mar 20, 2023

jczaja left a comment

Choose a reason for hiding this comment

paddle-bot bot commented Mar 22, 2023

YangQun1 commented Mar 15, 2023 •

edited

Loading

CLAassistant commented Mar 15, 2023 •

edited

Loading