Register emulated kernel implementations for RandomStandardNormal and TruncatedNormal #120
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
We need to track that these aren't truly implemented. Can you add some metadata to the ops table to track emulated kernel implementations? |
|
@jstoecker I'll add something to the op report for these kernels. |
jstoecker
added a commit
that referenced
this pull request
Dec 15, 2020
Merges some of the recent changes from the directml branch: * Use compute queue for AMD devices (#102) * Register List Kernels for DML (#95) * Update DirectMLX to latest (#104) * Remove extra rows from test email (#106) * Fix DML's Select kernel for int64 (#113) * Fix list kernels and tensor array ops registration (#114) * Simplify CI scripts (#112) * Fix StridedSlice's input size coalescing (#115) * Disable int64 image test (#116) * Fix network share copy path (#117) * Pipeline should continue if a test job fails (#118) * Switch network share path to use build number instead of build ID * Add missing HostMemory int32 registrations for _Arg and _RetVal (#122) * Implement all the arithmetic Scatter and ResourceScatter operators (#121) * Register emulated kernel implementations for RandomStandardNormal and TruncatedNormal (#120)
# for free
to join this conversation on GitHub.
Already have an account?
# to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
For some reason, on some models, TensorFlow has a habit of forcibly colocating kernels like ApplyAdam with RandomUniform, RandomStandardNormal, and TruncatedNormal. This may be because the initial weights, which is what Adam optimizes, are initialized at training start by one of the Random operators.
This change registers a set of kernels to "emulate" support for RandomStandardNormal and TruncatedNormal. These kernels re-use the CPU implementations and merely upload the values to a GPU tensor. This means that computation is still done on the CPU (which is okay, since it's usually only done once during initialization), but the DML registration means they can now be colocated with other operators, like ApplyAdam.
This change should improve our AI-Benchmark scores by about 5-10%.