Repository navigation
Restore vectorized PRelu for per-channel slopes on CPU - #33996
Open
Chinmay Shrivastava (JonSnow1807) wants to merge 2 commits into
Open
Chinmay Shrivastava (JonSnow1807) wants to merge 2 commits into
Chinmay Shrivastava (JonSnow1807) wants to merge 2 commits into
Conversation
microsoft#28750 replaced the branch-free PRelu formula with a per-element scalar ternary loop to stop +/-inf inputs producing NaN. The loop is not vectorized, which made per-channel PRelu (for example X [1,64,56,56] with slope [64,1,1]) roughly 20x slower on the CPU EP. Express the same per-lane selection with Eigen select() in the two span functors. select() evaluates both arms and keeps one, never adding them, so a non-finite value in the discarded arm cannot leak into the result and the inf/NaN behavior from microsoft#28750 is unchanged. Add a per-channel test with spans longer than a SIMD packet and inf, NaN and signed-zero inputs. Fixes microsoft#33831
|
Azure Pipelines: There may be pipelines that require an authorized user to comment /azp run to run. |
Contributor
There was a problem hiding this comment.
🟡 Changes recommended
The signed-zero tests can miss incorrect output signs because they use EXPECT_NEAR and only positive slopes, requiring explicit signbit checks and a negative-slope case.
1 open issue
What changed in this PR
Restores vectorized CPU float PRelu processing to address #33831’s performance regression while preserving the infinity and NaN handling introduced by #28750.
Changes:
- Replaces scalar loops with Eigen
select()for scalar-slope and elementwise-slope spans. - Adds a multi-channel 4-D regression test with finite values, infinities, NaNs, and signed zeros.
| File | Description |
|---|---|
onnxruntime/test/providers/cpu/activation/activation_op_test.cc |
Adds per-channel special-value coverage across vector-length spans. |
onnxruntime/core/providers/cpu/math/element_wise_ops.cc |
Uses Eigen selection to restore vectorization in two PRelu broadcast paths. |
Review effort: Balanced
Comment on lines
+686
to
+701
| const std::vector<float> slopes{0.5f, 2.0f}; | ||
|
|
||
| std::vector<float> inputs; | ||
| std::vector<float> outputs; | ||
| const float specials[] = {pos_inf, neg_inf, nan_val, 0.0f, -0.0f, 3.0f, -3.0f}; | ||
| for (int i = 0; i < 50; ++i) { | ||
| const float x = (i % 3 == 0) ? specials[(i / 3) % 7] : static_cast<float>(i) - 25.0f; | ||
| inputs.push_back(x); | ||
| // NaN compares false, so it takes the multiply branch and stays NaN. | ||
| outputs.push_back(x >= 0.0f ? x : x * slopes[i / 25]); | ||
| } | ||
|
|
||
| OpTester test("PRelu"); | ||
| test.AddInput<float>("X", x_dims, inputs); | ||
| test.AddInput<float>("slope", slope_dims, slopes); | ||
| test.AddOutput<float>("Y", x_dims, outputs); |
There was a problem hiding this comment.
Agreed, the float checker treats +0 and -0 as equal and both slopes were positive. The second slope is now negative, and a custom verifier checks every value (NaN, exact infinity, the rest with EXPECT_FLOAT_EQ) plus, on the CPU EP, the sign of each zero output. The sign check is CPU-only because CUDA's PRelu (a > 0 ? a : a * b) gives -0 for a +0 input with a negative slope. I checked that the test fails when the kernel gets the sign of zero wrong and passes with this PR.
Author
|
@microsoft-github-policy-service agree |
The default float checker compares with EXPECT_NEAR, so +0 and -0 are equal and a kernel that multiplies a zero input by the slope passed the test. Make the second slope negative so such a multiply flips the sign, and verify the output in a custom verifier: shape, NaN, exact infinity, the other values with EXPECT_FLOAT_EQ, and on CPU the sign of each zero.
This branch has not been deployed
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
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.
Description
Fixes #33831. Since #28750, the float PRelu span functors use a per-element ternary to keep
x * slopefrom turning ±inf into NaN. The ternary doesn’t vectorize, so PRelu with a per-channel slope became ~21x slower. This replaces both functors with Eigenselect(), which keeps #28750’s inf/NaN behaviour and vectorizes again.Motivation and Context
Measured with the issue’s repro (4-D input, per-channel slope), one thread,
taskseton one core, median of 9 fresh processes, GCC 14, x86-64: 1345 µs → 65 µs. The same-shape slope case goes 1349 µs → 63 µs. Output is bit-exact againstnumpy.where(x >= 0, x, slope * x)including inf, NaN and -0. New testActivationOpTest.PRelu_MultiChannel4D_SpecialValues; the PRelu, activation and broadcast tests pass, 548 tests, 0 failures. MSVC was not measured.AI usage: code drafted by both human and AI, and reviewed by human.