Conversation
|
great to see that oneDNN is making its way here! And thanks for the port @jacobkahn :-) Do you have any particular workloads where you see oneDNN with MKL faster than oneDNN without MKL? |
|
@mgouicem — happy to help 👍 - we're about to use the port in flashlight for our new CPU/OpenCL backends I have zero current benchmarks or supporting evidence that linking with MKL is faster besides that it used to be a bit faster with very old versions of MKL-DNN that I last benchmarked on. If it's just as good without, I'll happily gut the dependency (it makes life easier!). Thanks for the pointer there. |
Summary: Fix the Docker image build - Use multiple `RUN` commands with `docker prune` to allow checkingpointing the base image builds to make it easier to debug stuff in the future - Make `MKLROOT` a global env var - Remove the MKL BLAS vendor for oneDNN (apparently faster without MKL) - see microsoft/vcpkg#15068 (comment) - Enable building tests in the CPU CI build since it works now and `gtest_discover_tests` is removed so we don't have runtime `libmlk_rt` issues. We also shouldn't have those issues anymore in general because we are no longer linking MKL to the MKL-DNN/oneDNN build Differential Revision: D25514612 fbshipit-source-id: d0f52bced60c5265f322f80363bd888e3d50ab5a
Summary: Pull Request resolved: flashlight#330 Fix the Docker image build - Use multiple `RUN` commands with `docker prune` to allow checkingpointing the base image builds to make it easier to debug stuff in the future - Make `MKLROOT` a global env var - Remove the MKL BLAS vendor for oneDNN (apparently faster without MKL) - see microsoft/vcpkg#15068 (comment) - Enable building tests in the CPU CI build since it works now and `gtest_discover_tests` is removed so we don't have runtime `libmlk_rt` issues. We also shouldn't have those issues anymore in general because we are no longer linking MKL to the MKL-DNN/oneDNN build Reviewed By: tlikhomanenko Differential Revision: D25514612 fbshipit-source-id: a14ba9113b7d3f36ecf9a7478613ec1354518458
Summary: Pull Request resolved: #330 Fix the Docker image build - Use multiple `RUN` commands with `docker prune` to allow checkingpointing the base image builds to make it easier to debug stuff in the future - Make `MKLROOT` a global env var - Remove the MKL BLAS vendor for oneDNN (apparently faster without MKL) - see microsoft/vcpkg#15068 (comment) - Enable building tests in the CPU CI build since it works now and `gtest_discover_tests` is removed so we don't have runtime `libmlk_rt` issues. We also shouldn't have those issues anymore in general because we are no longer linking MKL to the MKL-DNN/oneDNN build Reviewed By: tlikhomanenko Differential Revision: D25514612 fbshipit-source-id: 9f371a24e8c4de0c6779092539ffc544eaf21812
Summary: Pull Request resolved: flashlight/flashlight#330 Fix the Docker image build - Use multiple `RUN` commands with `docker prune` to allow checkingpointing the base image builds to make it easier to debug stuff in the future - Make `MKLROOT` a global env var - Remove the MKL BLAS vendor for oneDNN (apparently faster without MKL) - see microsoft/vcpkg#15068 (comment) - Enable building tests in the CPU CI build since it works now and `gtest_discover_tests` is removed so we don't have runtime `libmlk_rt` issues. We also shouldn't have those issues anymore in general because we are no longer linking MKL to the MKL-DNN/oneDNN build Reviewed By: tlikhomanenko Differential Revision: D25514612 fbshipit-source-id: 9f371a24e8c4de0c6779092539ffc544eaf21812
Summary: Pull Request resolved: flashlight/flashlight#330 Fix the Docker image build - Use multiple `RUN` commands with `docker prune` to allow checkingpointing the base image builds to make it easier to debug stuff in the future - Make `MKLROOT` a global env var - Remove the MKL BLAS vendor for oneDNN (apparently faster without MKL) - see microsoft/vcpkg#15068 (comment) - Enable building tests in the CPU CI build since it works now and `gtest_discover_tests` is removed so we don't have runtime `libmlk_rt` issues. We also shouldn't have those issues anymore in general because we are no longer linking MKL to the MKL-DNN/oneDNN build Reviewed By: tlikhomanenko Differential Revision: D25514612 fbshipit-source-id: 9f371a24e8c4de0c6779092539ffc544eaf21812
Summary: Pull Request resolved: flashlight/flashlight#330 Fix the Docker image build - Use multiple `RUN` commands with `docker prune` to allow checkingpointing the base image builds to make it easier to debug stuff in the future - Make `MKLROOT` a global env var - Remove the MKL BLAS vendor for oneDNN (apparently faster without MKL) - see microsoft/vcpkg#15068 (comment) - Enable building tests in the CPU CI build since it works now and `gtest_discover_tests` is removed so we don't have runtime `libmlk_rt` issues. We also shouldn't have those issues anymore in general because we are no longer linking MKL to the MKL-DNN/oneDNN build Reviewed By: tlikhomanenko Differential Revision: D25514612 fbshipit-source-id: 9f371a24e8c4de0c6779092539ffc544eaf21812
Summary: Pull Request resolved: flashlight/flashlight#330 Fix the Docker image build - Use multiple `RUN` commands with `docker prune` to allow checkingpointing the base image builds to make it easier to debug stuff in the future - Make `MKLROOT` a global env var - Remove the MKL BLAS vendor for oneDNN (apparently faster without MKL) - see microsoft/vcpkg#15068 (comment) - Enable building tests in the CPU CI build since it works now and `gtest_discover_tests` is removed so we don't have runtime `libmlk_rt` issues. We also shouldn't have those issues anymore in general because we are no longer linking MKL to the MKL-DNN/oneDNN build Reviewed By: tlikhomanenko Differential Revision: D25514612 fbshipit-source-id: 9f371a24e8c4de0c6779092539ffc544eaf21812
Adds a new port for the oneDNN deep learning library.
cc @vpirogov, @dzarukin, @echeresh, @mgouicem, @akharito, @densamoilov
cc @syhw, @padentomasello, @tlikhomanenko, @vineelpratap, @xuqiantong, @avidov
x64 & !uwp. Tested myself onx64-linux.Yes