Optimze copy tensor with padding #32461
Open
+85
−10
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
susbhere
force-pushed
the
optimize_padded_copy
branch
2 times, most recently
from
820f1d4 to
31ed86d
Compare
|
build_jenkins |
susbhere
force-pushed
the
optimize_padded_copy
branch
6 times, most recently
from
98db555 to
f3da61f
Compare
yeonbok
reviewed
Oct 17, 2025
yeonbok
reviewed
Oct 17, 2025
susbhere
force-pushed
the
optimize_padded_copy
branch
6 times, most recently
from
58fd670 to
8ec08c0
Compare
Tensor layout related properties are calculated once and used those cached values during per element offset calculation. This brings ~200x improvement in wait time between two queries for PhiSlica model. That means a user has to wait only for 0.36 sec (instead of 74 sec !!!) between two queries. These numbers are from LNL. JIRA: https://jira.devtools.intel.com/browse/CVS-174810
susbhere
force-pushed
the
optimize_padded_copy
branch
from
8ec08c0 to
f056da3
Compare
|
build_jenkins |
yeonbok
reviewed
Oct 21, 2025
| fmt == bfvuwzyx); | ||
| } | ||
|
|
||
| static void get_axes_map(const format& fmt, int64_t* axes_map, size_t& map_size) { |
There was a problem hiding this comment.
static std::vector<int64_t> get_internal_dims(const format& fmt) const{
to use more aligned naming and use safe container usage.
There was a problem hiding this comment.
I want to use aray to make it the fastest. Even a small perf difference between vector and array adds up to some noticable difference. If I use vector here, I have to copy it to a local array in the caller.
yeonbok
reviewed
Oct 21, 2025
| for (int64_t y = 0; y < size.spatial[1]; y++) { | ||
| for (int64_t x = 0; x < size.spatial[0]; x++) { | ||
| *dst++ = static_cast<dst_t>(src[layout.get_linear_offset(cldnn::tensor(b, f, x, y, z, w))]); | ||
| void convert_and_copy_padded_source(const src_t* src, dst_t* dst, layout& layout) { |
There was a problem hiding this comment.
This only works for plain format. So please apply it for plain format only. For other (e.g., blocked) formats, please use original method.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Tensor layout related properties are calculated once and used those
cached values during per element offset calculation. This brings ~200x improvement in wait time between two queries for PhiSlica model. That means a user has to wait only for 0.36 sec (instead of 74 sec !!!) between two queries. These numbers are from LNL.
JIRA: https://jira.devtools.intel.com/browse/CVS-174810