[ESIMD] Fix perf regression caused by assumed align in block_load(usm) #11850
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
The element-size address alignment is valid from correctness point of view, but using 1-byte and 2-byte alignment implicitly causes performance regression for block_load(const int8_t *, ...) and block_load(const int16_t *, ...) because GPU BE have to generate slower GATHER instead of more efficient BLOCK-LOAD. Without this fix block-load causes up to 44% performance slow-down on some apps that used block_load() with alignment assumptions used before block_load(usm, ..., compile_time_props) was implemented. The reasoning for the expected/assumed alignment from element-size to 4-bytes for byte- and word-vectors is such: The idea of block_load() call (opposing to gather() call) is to have efficient block-load, and thus the assumed alignment is such that allows to generate block-load. This is a bit more tricky for user but that is how block_load/store API always worked before: block-load had restrictions that needed to be honored. To be on safer side, user can always pass the guaranteed alignment. Signed-off-by: Klochkov, Vyacheslav N <[email protected]>
Signed-off-by: Klochkov, Vyacheslav N <[email protected]>
v-klochkov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
v-klochkov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
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Outdated suggestions cannot be applied.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.
The element-size address alignment is valid from correctness point of view, but using 1-byte and 2-byte alignment implicitly causes performance regression for block_load(const int8_t *, ...) and block_load(const int16_t *, ...) because GPU BE have to generate slower GATHER instead of more efficient BLOCK-LOAD. Without this fix block-load causes up to 44% performance slow-down on some apps that used block_load() with alignment assumptions used before block_load(usm, ..., compile_time_props) was implemented.
The reasoning for the expected/assumed alignment from element-size to 4-bytes for byte- and word-vectors is such:
The idea of block_load() call (opposing to gather() call) is to have
efficient block-load, and thus the assumed alignment is such that
allows to generate block-load. This is a bit more tricky for user
but that is how block_load/store API always worked before: block-load
had restrictions that needed to be honored.
To be on safer side, user can always pass the guaranteed alignment.