[BugFix] Align default (C) tirx.round lowering to ties-to-even - #20131
Merged
Conversation
np.random.rand never produces exact midpoints, so the existing input cannot distinguish ties-to-even (np.round, constant folding, and every other backend) from ties-away-from-zero (C round/roundf). Reuse the midpoint vector from test_round_ties_to_even verbatim. This test fails against current main.
tirx.round is ties-to-even, and constant folding implements it with std::nearbyint, but the default.FLowerIntrinsic rule lowered it through FloatSuffix to the C library round()/roundf(), which is ties away from zero. Rename to nearbyint before the float suffix is applied, mirroring the existing CUDA rule. Fixes the C target's disagreement with the constant folder and with every other backend after apache#19368.
Contributor
Author
tlopex
approved these changes
Sep 1, 2026
tlopex
left a comment
tlopex
left a comment
Member
There was a problem hiding this comment.
LGTM Thanks so much for the fix!
LngelKyo
added a commit
to LngelKyo/tvm
that referenced
this pull request
Sep 1, 2026
test_round_ties_to_even was pinned to target="llvm". The shape of the bug apache#19368 left in the C host (fixed by apache#20131) was that a backend could implement tirx.round without ties-to-even and no test would notice: the ties behaviour was only exercised on llvm. The C host passed everything for four months because of exactly this. Parametrize the test over every backend that registers tirx.round (llvm, c, cuda, nvptx, rocm, metal, opencl, vulkan, webgpu, hexagon). The target-list convention (gpu-marked pytest.params for device targets) follows test_clz in the same file; "c" is always runnable, so it is the one target exempt from the device_enabled skip, as in test_codegen_loop_step. The midpoint vector is the one apache#19368 added.
LngelKyo
added a commit
to LngelKyo/tvm
that referenced
this pull request
Sep 1, 2026
apache#20131 left a 3-line wrapped call in test_round that ruff-format (line-length = 100) rejoins into one line, so pre-commit run --all-files fails on every open PR. Reformat the line; no other change. Reproduce with: uvx ruff@0.12.3 format --check tests/python/codegen/test_target_codegen_c_host.py
LngelKyo
added a commit
to LngelKyo/tvm
that referenced
this pull request
Sep 1, 2026
test_round_ties_to_even was pinned to target="llvm". The shape of the bug apache#19368 left in the C host (fixed by apache#20131) was that a backend could implement tirx.round without ties-to-even and no test would notice: the ties behaviour was only exercised on llvm. The C host passed everything for four months because of exactly this. Parametrize the test over the two host targets the regression can recur on ("c" and "llvm"), following the test_codegen_loop_step convention; device backends have their own lowering rules and belong in their own codegen tests. The midpoint vector is the one apache#19368 added.
LngelKyo
added a commit
to LngelKyo/tvm
that referenced
this pull request
Sep 1, 2026
apache#20131 left a 3-line wrapped call in test_round that ruff-format (line-length = 100) rejoins into one line, so pre-commit run --all-files fails on every open PR. Reformat the line; no other change. Reproduce with: uvx ruff@0.12.3 format --check tests/python/codegen/test_target_codegen_c_host.py
LngelKyo
added a commit
to LngelKyo/tvm
that referenced
this pull request
Sep 1, 2026
test_round_ties_to_even was pinned to target="llvm". The shape of the bug apache#19368 left in the C host (fixed by apache#20131) was that a backend could implement tirx.round without ties-to-even and no test would notice: the ties behaviour was only exercised on llvm. The C host passed everything for four months because of exactly this. Parametrize the test over the two host targets the regression can recur on ("c" and "llvm"), following the test_codegen_loop_step convention; device backends have their own lowering rules and belong in their own codegen tests. The midpoint vector is the one apache#19368 added.
LngelKyo
added a commit
to LngelKyo/tvm
that referenced
this pull request
Sep 1, 2026
apache#20131 left a 3-line wrapped call in test_round that ruff-format (line-length = 100) rejoins into one line, so pre-commit run --all-files fails on every open PR. Reformat the line; no other change. Reproduce with: uvx ruff@0.12.3 format --check tests/python/codegen/test_target_codegen_c_host.py
LngelKyo
added a commit
to LngelKyo/tvm
that referenced
this pull request
Sep 1, 2026
test_round_ties_to_even was pinned to target="llvm". The shape of the bug apache#19368 left in the C host (fixed by apache#20131) was that a backend could implement tirx.round without ties-to-even and no test would notice: the ties behaviour was only exercised on llvm. The C host passed everything for four months because of exactly this. Parametrize the test over the two host targets the regression can recur on ("c" and "llvm"), following the test_codegen_loop_step convention; device backends have their own lowering rules and belong in their own codegen tests. The midpoint vector is the one apache#19368 added.
tlopex
pushed a commit
that referenced
this pull request
Sep 1, 2026
#20252) #19368 aligned `tir.round` to ties-to-even across LLVM, CUDA, NVPTX, ROCm, Hexagon, Metal, OpenCL, SPIR-V and WebGPU, and #20131 fixed the C host — but the shape of that bug can recur: `test_round_ties_to_even` (tests/python/tirx-base/test_tir_intrin.py) is pinned to `target="llvm"`, so the C host's ties behaviour is not exercised anywhere. The C host passed every test for four months for exactly this reason. **What this PR does** Parametrize `test_round_ties_to_even` over the two host targets the regression can recur on — `c` and `llvm` — following the `test_codegen_loop_step` convention. Device backends have their own lowering rules and belong in their own codegen tests. The midpoint vector is the one #19368 added, reused verbatim. **Verification.** On a local build: reverting the C rule to its pre-#20131 ties-away-from-zero form makes `test_round_ties_to_even[c]` fail with the expected midpoint mismatches (`0.5 → 1.0 vs 0.0`, `2.5 → 3.0 vs 2.0`, `−0.5 → −1.0 vs −0.0`), while `[llvm]` still passes — i.e. the test now catches the class of bug that #20131 fixed. With the rule restored, both targets pass. Formatting checked with ruff format/ruff check — both the CI-pinned ruff 0.12.3 and 0.15.14 pass clean on the touched file. **Second commit** is docs-only: it pins the floating-point-environment assumption from the #20131 discussion into `src/target/intrin_rule.cc`. nearbyint is ties-to-even under the default `FE_TONEAREST`, and the same holds for the LLVM rule, which lowers to `llvm.nearbyint` — rounding-mode sensitive, like nearbyint itself (`llvm.roundeven` is the mode-independent one). Every other backend that registers `tirx.round` — cuda, nvptx, rocm, hexagon, metal, opencl, vulkan, webgpu — emits code for a separate device whose rounding mode is fixed at RNE, so `fesetround()` in the host process cannot reach it. **Third commit** is a one-line CI restore: #20131 left a 3-line wrapped call in `test_target_codegen_c_host.py` that ruff-format rejoins under `line-length = 100`, so `pre-commit run --all-files` fails on every open PR. This PR carries the rejoin so that merging it leaves main lint-green. Reproduce: `uvx ruff@0.12.3 format --check tests/python/codegen/test_target_codegen_c_host.py`. --------- Co-authored-by: LngelKyo <lngel.kyo@a2o-labs.com>
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
3 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.
#19368 aligned tir.round to ties-to-even "across all backends" and moved LLVM, CUDA, NVPTX, ROCm, Hexagon, Metal, OpenCL, SPIR-V and WebGPU. It did not touch src/target/intrin_rule.cc, so the default.FLowerIntrinsic rule — which the c target uses — still lowers through FloatSuffix and emits round/roundf, i.e. ties away from zero.
This disagrees with:
tir.roundto ties-to-even across all backends #19368.Why CI did not catch it. test_target_codegen_c_host.py::test_round compiles with target="c" and asserts against np.round, which is ties-to-even — so the intended semantics was already encoded. It passes only because its input is np.random.rand, which does not produce exact midpoints (0 occurrences in 2,048,000 sampled values). The ties-to-even test added by #19368 (test_tir_intrin.py::test_round_ties_to_even) only runs target="llvm", so no test covered the C path's tie behaviour.
The first commit adds the midpoint vector — reused verbatim from test_round_ties_to_even — to the existing C-host test and it fails:
The second commit renames round→nearbyint before the float suffix, mirroring the existing CUDA rule, and it passes.
One note for reviewers: nearbyint honours the current floating-point environment and is ties-to-even under the default FE_TONEAREST. That is the same guarantee LLVM, CUDA, Metal and OpenCL already rely on after #19368; this PR does not introduce a new assumption.