Have convolution_backwards error on indivisible groups and reference operator accumulate in double

[CharlieL7]

Motivation

• convolution_backwards was hitting accuracy errors with allclose due to reference accumulator inaccuracy. It was accumulating in the low precision types on the reference side.
• Also fixed correctness for convolution_backwards for invalid groups settings.

Technical Details

• Change accumulator to use double type on reference convolution_backwards.

Changelog Category

Add a CHANGELOG.md entry for any option other than Not Applicable

• • Added: New functionality.
• • Changed: Changes to existing functionality.
• • Removed: Functionality or support that has been removed. (Compared to a previous release)
• • Optimized: Component performance that has been optimized or improved.
• • Resolved Issues: Known issues from a previous version that have been resolved.
• • Not Applicable: This PR is not to be included in the changelog.

Follow the LLVM AI Tool Use Policy for contributions using AI.

[Copilot]

🟢 Approval recommended

Remaining feedback is limited to minor comment clarifications.

Pull request overview

Fixes convolution_backwards validation and improves reference accumulation precision.

Changes:

• Validates invalid groups and zero strides.
• Uses higher-precision accumulators.
• Adds regression tests and changelog coverage.

File summaries

File	Summary	Review notes
test/ref/convolution_backwards.cpp	Numerical and grouped-convolution regression tests	No issues
test/op_shape_test.cpp	Invalid stride and group tests	Nit: clarify the out-of-range group mapping
src/include/migraphx/op/convolution_backwards.hpp	Accumulation and validation fixes	Nit: clarify the invalid-group failure rationale
CHANGELOG.md	Documents the resolved issue	No issues

Review details

Suppressed comments (2)

src/include/migraphx/op/convolution_backwards.hpp:102

• The rationale here describes the failure as leaving channels unread, but compute() uses group_id = w / (wei_n / group): with a non-divisible channel count, the trailing w values map to group_id >= group, so idx_out goes past the output channel dimension and can write/read out of bounds. Update the comment to document the actual invalid-group failure this check prevents.

        // compute() walks a group's input channels as one contiguous block of
        // weights_channels / group, so an inexact division leaves part of every group unread.

test/op_shape_test.cpp:1005

• This test comment has the same incorrect failure description: integer division does not simply leave part of each group unread; the trailing input channel is assigned to an invalid group and makes idx_out exceed the output channel bounds. Please describe the out-of-range group mapping instead.

// compute() splits the input channels into equal per-group blocks, so a group that does not
// divide them would leave part of every group unread.

• Files reviewed: 4/4 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[eddieliao]

Overall LGTM; minor nitpick with changelog entry.