apeforest opened a new pull request #14570: [WIP] use a compile flag to use int64 tensor size URL: https://github.com/apache/incubator-mxnet/pull/14570 ## Description ## This PR is to fix the performance degradation reported in https://github.com/apache/incubator-mxnet/issues/14496. The performance degradation is introduced by PR https://github.com/apache/incubator-mxnet/pull/11742. I have verified the performance degradation in one of the operators transpose using a script below: ``` import mxnet as mx import time import numpy as np sizes = [10, 50, 100,200,500] iters = [10000,1000,500,200,20] times = [] for size in range(len(sizes)): data = [] s = sizes[size] print(s) for i in range(iters[size]): x = mx.nd.ones((s,s,s)) mx.nd.waitall() start = time.time() y = mx.nd.transpose(x,(2,0,1)) mx.nd.waitall() data.append((time.time() - start)*1000) #print(data[-1]) times.append(data) print('mxnet version: %s' % mx.__version__) for s in range(len(sizes)): print('--------------------') print('size: %s' % str(sizes[s])) print('p50: %4.2f ms' % np.percentile(times[s],50)) print('p90: %4.2f ms' % np.percentile(times[s],90)) print('p99: %4.2f ms' % np.percentile(times[s],99)) ``` By changing the `index_t` type from `int64_t` to `int32_t` can consistently change the 50-percentile runtime of tranpose of size 500 from 5000ms to 2400ms. By changing the data type in the operator (https://github.com/dmlc/mshadow/blob/master/mshadow/extension/transpose.h#L70) alone, can also reduce the 50-percentile runtime of size 500 to 2600ms. I thereforce come to the conclusion that the performance degradation is caused by the runtime of integer arithmetic operations between 32-bit integers and 64-bit integers. To further experiment with the arithmetic operations alone, I tested using a small program [here](https://github.com/apeforest/doraemon/blob/master/perf32vs64.cc). The runtime results are shown below: ``` result = 49995000 Add 32-bit time in ms 1359 result = 49995000 Add 64-bit time in ms 1971 result = 349965000 Add Mul 32-bit time in ms 1196 result = 349965000 Add Mul 64-bit time in ms 3477 result = 7137858 Add Div 32-bit time in ms 2878 result = 7137858 Add Div 64-bit time in ms 8499 ``` There are a few solutions to this problem: (1) Add a compilation flag to choose data types for tensor size (This PR) (2) Add an environment variable to choose data type for tensor size at runtime (3) Choose data type for tensor size at runtime based on the size of the tensor Given the expression template used in mshadow for operators, either (2) or (3) requries a significant change in the mshadow library. (1) can be used as a quick solution to fix performance degradation reported in several issues: https://github.com/apache/incubator-mxnet/issues/14496, https://github.com/apache/incubator-mxnet/issues/13928, https://github.com/apache/incubator-mxnet/issues/14563, https://github.com/apache/incubator-mxnet/issues/14569 Any other suggestion is appreciated. ## Checklist ## ### Essentials ### Please feel free to remove inapplicable items for your PR. - [ ] The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant [JIRA issue](https://issues.apache.org/jira/projects/MXNET/issues) created (except PRs with tiny changes) - [ ] Changes are complete (i.e. I finished coding on this PR) - [ ] All changes have test coverage: - Unit tests are added for small changes to verify correctness (e.g. adding a new operator) - Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore) - Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL) - [ ] Code is well-documented: - For user-facing API changes, API doc string has been updated. - For new C++ functions in header files, their functionalities and arguments are documented. - For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable - Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html - [ ] To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change ### Changes ### - [ ] Feature1, tests, (and when applicable, API doc) - [ ] Feature2, tests, (and when applicable, API doc) ## Comments ## - If this change is a backward incompatible change, why must this change be made. - Interesting edge cases to note here
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services