Add mkldnn OP for slice #13730
Add mkldnn OP for slice #13730
Conversation
|
@huangzhiyuan please verify the performance in #12303 and other tests from your local. |
| offsets[i] = s; | ||
| } | ||
| auto in_mem = in.GetMKLDNNData(); | ||
| auto in_mem_pd = in_mem->get_primitive_desc(); |
There was a problem hiding this comment.
Please merge in_mem and in_mem_pd into one line as out_mem_pd in case the in_mem doesn't use in here.
There was a problem hiding this comment.
This is just for the overall code style, I can also modify it if you insist.
There was a problem hiding this comment.
Have merge im_mem and im_mem_pd into one line in the new commit.
|
@pengzhao-intel No specific test case for this change should be added, the current test case is enough to cover. Thanks for your review :) |
| offsets[i] = s; | ||
| } | ||
| auto in_mem = in.GetMKLDNNData(); | ||
| auto in_mem_pd = in_mem->get_primitive_desc(); |
| static MX_THREAD_LOCAL std::unordered_map<MKLDNNSliceSignature, MKLDNNSliceFwd, OpHash> fwds; | ||
| #endif | ||
| MKLDNNSliceSignature key(param); | ||
| key.AddSign(in_data); |
There was a problem hiding this comment.
Do we need to have is_train and out_data into the key?
There was a problem hiding this comment.
In_data is enough for the key to cache. And I have put is_train and out_data into the key in new commit.
src/operator/tensor/matrix_op-inl.h
Outdated
|
|
||
| if (in_stype == kDefaultStorage) { | ||
| #if MXNET_USE_MKLDNN == 1 | ||
| if (dev_mask == Context::kCPU && MKLDNNEnvSet() && SupportMKLDNNSlice(param)) { |
|
Closes #12303 . |
* add mkldnn slice * fix lint * fix lint * mv SliceEx to matrix_op.cc * fix lint * optimize dispatch_mode * retrigger ci * fix indent
* add mkldnn slice * fix lint * fix lint * mv SliceEx to matrix_op.cc * fix lint * optimize dispatch_mode * retrigger ci * fix indent
Description
Add mkldnn implement for slice OP
Checklist
Essentials
Please feel free to remove inapplicable items for your PR.
Changes
test_slice has passed
@pengzhao-intel , @ZhennanQin , @TaoLv