Support MKL2017 DNN API #5
Assignees
Labels
Comments
|
Can Intel finally clarify relation roadmap against this and #363? At "marketing" level is quite confusional. |
|
But is mkl-dnn https://github.com/01org/mkl-dnn? |
|
Here mkl-dnn actually refers to the DNN primitives in Intel MKL 2017 (https://software.intel.com/en-us/articles/introducing-dnn-primitives-in-intelr-mkl) |
|
There will be a disambiguation between the two? |
|
Cause mkl dnn api and mkl-dnn api have a very similar marketing name. |
|
Ok probably it is still not clear the roadmap also if one is opensource and the other one is closed:
|
|
We plan to support the DNN primitives in Intel MKL 2017 at this moment; no plan for the https://github.com/01org/mkl-dnn support yet. |
wzhongyuan
pushed a commit
to wzhongyuan/BigDL
that referenced
this issue
Aug 28, 2017
refine inference process
i8run
added a commit
that referenced
this issue
Jun 28, 2018
This feature enables mkl-dnn support, which can speed up deep learning model. We wrapper the native c api in the java, which are in BigDL-core projects. And in BigDL, we integrated the convolution, batchnorm, maxpooling, avgpooling, relu, lrn, softmax, caddtable and concattable. Currently, it supports create the model which only contains dnn layer or container. Because the data layout is optimized in mkl-dnn. The mkl-dnn model will use `DnnTensor` which contains the native buffer as a default tensor. So there're some notations, 1. User should copy the data from jvm heap at the first layer and copy back to jvm heap at the last layer. 2. User should compile the model, which contains the phase (training/inference) and input tensor size. It will infer and allocate the other information. * fix: linear performance issue and serialization of java object in MklDnnTensor * memory leak refactor * memory leak and bn performance issues 1. Memory Leak The internal buffer with MklDnnTensor should not be re-assigned without releasing. So we should check it first. At first iteration or after the changing of input size, we create a new MklDnnTensor as a buffer. 2. Bn perf The JIT BatchNormalization only supports avx2 or avx512, which has much batter performance than ref version. The input and gradOutput format should be the same to get the best performance. * test: add some test cases for BatchNorm. The computation of float value is not the same as C/C++/Native with JVM. And batch norm will make it much greater such as 10^-8 -> 10^-4 -> 10^-1 * fix: rebase with upstream master: 1. Concat and ConcatTable should inherit from DynamicContainer. 2. updateParameters has been depricated. 3. zeroGradParameters should be final. But from now on, the Linear should use it. 4. Some other syntax or semantic errors. * perf: single node and single model performance * perf: single model * feat: add fusion for mkl-dnn * test: add test utils to compare dnn output * test: add some tests compared with caffe * add unit tests for dnn tensor * add unit test for reorder memory * test: fix the test regression errors * checkin reorder manager * add backward for sequential * fix some bugs * update core ref * add unit tests * refactor: move the static class DataType, AlgKind and so on to standalone class (#4) * refactor: delete MklDnn.MemoryFormat * refactor: move the static class DataType, AlgKind and so on to standalone class * fix: core refactor errors * refactor: spec errors (#5) * Mkl dnn dev (#6) * checkin reorder manager * add container and refine reorder manager * fix merge issue * add join table forward * refine inteface (#7) * add LRN and ReLU * add pooling * refactor: conv + linear + bn * add JoinTable backward * refactor: conv + linear + bn * add cAddTable concattable * fix: reorder failed on some of convs * refactor: softmax * refactor: fusion support * refactor: resnet_50 * refactor: move tests to this branch * refactor: delete unusefull files and enable the special old tests. refactor: delete unsed methods in MklDnnOps fix: scalastyle check * fix: rebase with upstream * fix: ignore the prototxt tests * fix: do not change the core commit ref * fix: move set num of threads for mkldnn to ResNet50Perf * fix: serialization disabled for mkldnn module
wzhongyuan
pushed a commit
to wzhongyuan/BigDL
that referenced
this issue
Jan 16, 2019
* feat: mkl-dnn initialize * fix: structure of building * fix: public final static * fix: delete the dependencies of environments * fix: skip tests * add update dnn wrappers * fix: dynamic load iomp5 * feat linear supports and some fix * add more wrapper * add lrn api * fix: add bn and softmax * fix: some fixes * fix: mkl-dnn build * feat: add get format api * fix: add getSize * feat: aligned memory * add conv fuse relu api * fix: add aligned storage * add concat api * fix: mkl envs for lib mkldnn * fix: add mkl add method with 2 ptrs * fix: update to Release * fix: batch norm infer mode * fix: update 0.5.0 -> 0.6.0 * add free (intel-analytics#5) * feat: affinity for java thread * fix: update core branch * fix: delete the memset constant value for debug, and add affinity * feat: add mkl-dnn fusion * fix: memory format enum consistent with dnn * feat: add auto format * refactor: delete the MemoryFormat in MklDnn * Memory should load MKLDnn (intel-analytics#6) * refactor: move enums to seprate classes (intel-analytics#7) * feat: add GetShape and GetFormat api * fix: delete printf * fix a bug * add sum * refactor: change name * refactor: change submodule infos * fix: set block time by default. A property to control to disable it
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Intel MKL release 2017 version and it contains a DNN API, which provide DNN operation optimized for IA architecture. We will add new layers which leverage these new APIs to get a better performance on CPU.
The text was updated successfully, but these errors were encountered: