You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by pa...@apache.org on 2020/09/23 12:44:37 UTC
[incubator-mxnet] branch v1.x updated: [v1.x] ElementWiseSum fix
for oneDNN (#18777) (#19199)
This is an automated email from the ASF dual-hosted git repository.
patriczhao pushed a commit to branch v1.x
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git
The following commit(s) were added to refs/heads/v1.x by this push:
new a6c357a [v1.x] ElementWiseSum fix for oneDNN (#18777) (#19199)
a6c357a is described below
commit a6c357a0da7ff01df2bfef09bad644ab3f1c0ae3
Author: bgawrych <ba...@intel.com>
AuthorDate: Wed Sep 23 14:42:52 2020 +0200
[v1.x] ElementWiseSum fix for oneDNN (#18777) (#19199)
* Fix ElementwiseSum for DNNL
* Fix sanity and replace push_back with emplace_back
* Change order of the data format conditions
* Add NOLINT to avoid readability error
* Add test for oneDNN ElemwiseSum
Co-authored-by: Bart Gawrych <ga...@intel.com>
Co-authored-by: Bart Gawrych <ga...@intel.com>
---
src/operator/tensor/elemwise_sum.cc | 15 ++++++++-------
tests/python/mkl/test_mkldnn.py | 24 +++++++++++++++++++++++-
2 files changed, 31 insertions(+), 8 deletions(-)
diff --git a/src/operator/tensor/elemwise_sum.cc b/src/operator/tensor/elemwise_sum.cc
index 5885d73..c513e65 100644
--- a/src/operator/tensor/elemwise_sum.cc
+++ b/src/operator/tensor/elemwise_sum.cc
@@ -113,7 +113,14 @@ void ElementWiseSumComputeExCPU(const nnvm::NodeAttrs& attrs,
CHECK_EQ(outputs.size(), 1U);
CHECK_EQ(req.size(), 1U);
if (req[0] == kNullOp) return;
- if (common::ContainsOnlyStorage(inputs, kRowSparseStorage) ||
+#if MXNET_USE_MKLDNN == 1
+ if (IsMKLDNNData(inputs)) {
+ MKLDNNRun(MKLDNNSumForward, attrs, ctx, inputs, req, outputs);
+ } else if (common::ContainsOnlyStorage(inputs, kDefaultStorage)) {
+ FallBackCompute(ElementWiseSumCompute<cpu>, attrs, ctx, inputs, req, outputs);
+ }
+#endif
+ else if (common::ContainsOnlyStorage(inputs, kRowSparseStorage) || // NOLINT(*)
(inputs.size() == 3U && inputs[0].storage_type() == kDefaultStorage &&
inputs[1].storage_type() == kCSRStorage && inputs[2].storage_type() == kDefaultStorage) ||
(inputs.size() > 4U && common::ContainsStorageType(inputs, kDefaultStorage) &&
@@ -123,12 +130,6 @@ void ElementWiseSumComputeExCPU(const nnvm::NodeAttrs& attrs,
ResourceRequest(ResourceRequest::kTempSpace));
NDArray out_nd = outputs[0];
mxnet::ndarray::ElementwiseSum<cpu>(s, rsc, inputs, &out_nd);
-#if MXNET_USE_MKLDNN == 1
- } else if (IsMKLDNNData(inputs)) {
- MKLDNNRun(MKLDNNSumForward, attrs, ctx, inputs, req, outputs);
- } else if (common::ContainsOnlyStorage(inputs, kDefaultStorage)) {
- FallBackCompute(ElementWiseSumCompute<cpu>, attrs, ctx, inputs, req, outputs);
-#endif
} else {
LogUnimplementedOp(attrs, ctx, inputs, req, outputs);
}
diff --git a/tests/python/mkl/test_mkldnn.py b/tests/python/mkl/test_mkldnn.py
index 731a761..213bcfb 100644
--- a/tests/python/mkl/test_mkldnn.py
+++ b/tests/python/mkl/test_mkldnn.py
@@ -254,6 +254,28 @@ def test_flatten_slice_after_conv():
print(p[0])
+def test_mkldnn_sum_with_mkldnn_layout():
+
+ x_shape = (32, 3, 224, 224)
+ x_npy = np.ones(x_shape)
+ w_shape = (32, 3, 3, 3)
+ w_npy = np.ones(w_shape)
+
+ x = mx.sym.Variable("x")
+ w = mx.sym.Variable("w")
+ z = mx.symbol.Convolution(data=x, weight=w, num_filter=32, kernel=(3, 3))
+ num_inputs = [2, 3, 4, 5]
+ for i in num_inputs:
+ inputs = []
+ for n in range(i):
+ inputs.append(z)
+ y = mx.sym.add_n(*inputs) # (only MKLDNN data input)
+ exe = y.simple_bind(ctx=mx.cpu(), x=x_shape, w=w_shape)
+ out = exe.forward(is_train=False, x=x_npy, w=np.ones(w_shape))[0]
+ #conv with kernel (3,3) on ones should give result=27
+ single_cov = 27.0
+ assert_almost_equal(out[0].asnumpy()[0, 0, 0], single_cov*i)
+
def test_mkldnn_sum_inplace_with_cpu_layout():
x_shape = (32, 3, 224, 224)
@@ -263,7 +285,7 @@ def test_mkldnn_sum_inplace_with_cpu_layout():
x = mx.sym.Variable("x")
y = mx.sym.Variable("y")
z = mx.symbol.Convolution(data=x, num_filter=32, kernel=(3, 3))
- z = mx.sym.add_n(z, y)
+ z = mx.sym.add_n(z, y) # (MKLDNN data, cpu data)
exe = z.simple_bind(ctx=mx.cpu(), x=x_shape, y=y_shape)
out = exe.forward(is_train=False, x=x_npy, y=y_npy)[0]
assert_almost_equal(out[0].asnumpy()[0, 0, 0], 1.0)