Variable thread count for multi-threaded GEMMs #1316
Merged
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
|
Looks good to me, but could you please provide some performance figures just to illustrate how significant "significant" is ? (And if I read your patch correctly, it should not cause a slowdown on less well endowed systems ?) |
m, n, and k can be set to arbitrary constants. A and B matrices can be transposed independently.
|
Here are some scaling studies on a 36-core system (2 x Intel 18-core Xeon E5-2695 v4, gcc 4.9.3, TARGET=HASWELL USE_OPENMP=1 NUM_THREADS=36 INTERFACE64=0):
|
|
Here are some experiments on an 8-core system (Intel Core i7, default build for OSX):
|
|
There is another threading threshold in interface/gemm.c |
|
I forgot to mention that you are on right path, fixed threading threshold as currently is great impairment to say OpenCV/SGEMM |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in 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Outdated suggestions cannot be applied.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.
When performing smallish GEMMs (M~1000, N~100, K~100) on a 36-core system, I've found that I can achieve a significant speedup by reducing the number of OpenMP threads. It turns out that multi-threading is disabled if
(m < nthreads * SWITCH_RATIO) || (n < nthreads * SWITCH_RATIO), so we maintain parallel execution with no algorithmic changes ifnthreadsis changed tomin(m / SWITCH_RATIO, n / SWITCH_RATIO, nthreads). This pull request makes this change automatically and also clears some unused code that dates back to GotoBLAS.