[Relay][Hexagon] Add per-channel FixedPointMultiply operation #13080
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
Collaborator
|
Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment. Generated by tvm-bot |
ibsidorenko
force-pushed
the
fpm-per-channel
branch
from
7f61048 to
c0c66c8
Compare
ibsidorenko
changed the title
Member
|
@tvm-bot rerun |
masahi
reviewed
Oct 14, 2022
| right shift. This is because we are rounding twice instead than only once. I.e.: | ||
|
|
||
| * original q_multiply_shift: round(x*y*2^-s) | ||
| * hexagon q_multiply_shift: round(round(x*y)*2^-s) |
Member
There was a problem hiding this comment.
cc @kparzysz-quic @jverma-quic on this HVX implementation.
Contributor
Author
There was a problem hiding this comment.
I will add test to demonstrate issue with accuracy.
Contributor
Author
There was a problem hiding this comment.
I have fixed issue with accuracy drop.
For the case when we need both right and left shifts at the same time I use "old" approach and lower this oper to the sequence left_shift/multipy/add/right_shift (64bit arithmetic). Right now I have no idea how to implement this case through vector HVX instructions without accuracy drop.
areusch
added
needs-triage
ibsidorenko
force-pushed
the
fpm-per-channel
branch
from
c0c66c8 to
123ad38
Compare
Contributor
Author
|
@tvm-bot rerun |
Main goal of this commit is to improve performance for Hexagon target and preserve performance/accuracy for x86, GPU and etc. targets. "qnn.requantize" operation is lowered into the sequence of multiply, add, shift during QNN canonicalization pass if scale quantization parameter is the vector of scalars. This commit adds new Relay per-channel/per-axis FixedPointMultiply operation and is used in "qnn.requantize" operation lowering. per-channel/per-axis FixedPointMultiply is implemented through tir.q_multiply_shift_per_axis intrinsic. For Hexagon target it overrides default implementation and generates HVX vmpye/vmpyo instruction (see _q_multiply_shift_per_axis_hexagon). For all other targets it uses default implementation (64 bits arithmetic). Performance/accuracy measurement: CPU(x86) target: accuracy and performance are the same. For other targets should be the same (otherwise it is bug). Hexagon target: speedup of qnn.requantize 7x-9x times (Snapdragon 888, 3.08 ms -> 0.39 ms)
ibsidorenko
force-pushed
the
fpm-per-channel
branch
from
123ad38 to
483a1be
Compare
masahi
reviewed
Oct 26, 2022
src/target/intrin_rule.cc
Outdated
| PrimExpr right_shift = call->args[3]; | ||
| PrimExpr q = call->args[4]; | ||
| PrimExpr is_lshift_required = call->args[5]; | ||
| // Note, 7th argument is "is_rshift_required" flag, but we do need that here. |
Contributor
Author
There was a problem hiding this comment.
Oh... yes, exactly. My bad, this is typo in comment.
masahi
approved these changes
Oct 26, 2022
Member
There was a problem hiding this comment.
@kparzysz-quic This PR improves performance on int8 resnet50 from the PR #12911 while preserving accuracy.
Manual schedules (no tuning): 146 msec (before) -> 92 msec.
Tuned schedules (vrmpy auto tensorization): 105 msec -> 58 msec.
Very cool!
xinetzone
pushed a commit
to daobook/tvm
that referenced
this pull request
Nov 10, 2022
…#13080) * [Relay][Hexagon] Add per-channel FixedPointMultiply operation Main goal of this commit is to improve performance for Hexagon target and preserve performance/accuracy for x86, GPU and etc. targets. "qnn.requantize" operation is lowered into the sequence of multiply, add, shift during QNN canonicalization pass if scale quantization parameter is the vector of scalars. This commit adds new Relay per-channel/per-axis FixedPointMultiply operation and is used in "qnn.requantize" operation lowering. per-channel/per-axis FixedPointMultiply is implemented through tir.q_multiply_shift_per_axis intrinsic. For Hexagon target it overrides default implementation and generates HVX vmpye/vmpyo instruction (see _q_multiply_shift_per_axis_hexagon). For all other targets it uses default implementation (64 bits arithmetic). Performance/accuracy measurement: CPU(x86) target: accuracy and performance are the same. For other targets should be the same (otherwise it is bug). Hexagon target: speedup of qnn.requantize 7x-9x times (Snapdragon 888, 3.08 ms -> 0.39 ms) * Address code review comments
xinetzone
pushed a commit
to daobook/tvm
that referenced
this pull request
Nov 25, 2022
…#13080) * [Relay][Hexagon] Add per-channel FixedPointMultiply operation Main goal of this commit is to improve performance for Hexagon target and preserve performance/accuracy for x86, GPU and etc. targets. "qnn.requantize" operation is lowered into the sequence of multiply, add, shift during QNN canonicalization pass if scale quantization parameter is the vector of scalars. This commit adds new Relay per-channel/per-axis FixedPointMultiply operation and is used in "qnn.requantize" operation lowering. per-channel/per-axis FixedPointMultiply is implemented through tir.q_multiply_shift_per_axis intrinsic. For Hexagon target it overrides default implementation and generates HVX vmpye/vmpyo instruction (see _q_multiply_shift_per_axis_hexagon). For all other targets it uses default implementation (64 bits arithmetic). Performance/accuracy measurement: CPU(x86) target: accuracy and performance are the same. For other targets should be the same (otherwise it is bug). Hexagon target: speedup of qnn.requantize 7x-9x times (Snapdragon 888, 3.08 ms -> 0.39 ms) * Address code review comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
Main goal of this commit is to improve performance for Hexagon target and preserve performance/accuracy for x86, GPU and etc. targets.
"qnn.requantize" operation is lowered into the sequence of multiply, add, shift during QNN canonicalization pass if scale quantization parameter is vector of scalars. This commit adds new Relay per-channel/per-axis FixedPointMultiply operation and is used in "qnn.requantize" operation lowering.
per-channel/per-axis FixedPointMultiply is implemented through tir.q_multiply_shift_per_axis intrinsic. For Hexagon target it overrides default implementation and generate HVX vmpye/vmpyo instruction (see _q_multiply_shift_per_axis_hexagon). For all other targets it uses default implementation (64 bits arithmetic).
Performance/accuracy measurement:
CPU(x86) target: accuracy and performance are the same. For other targets should be the same (otherwise it is bug).
Hexagon target: speedup of qnn.requantize 7x-9x times (Snapdragon 888, 4.4 ms -> 0.5 ms)