Add promote_storage_type pass to compute storage-only types in float

[pfultz2]

Motivation

On the GPU, bf16 does not support any VALU computations. It only supports matrix ops. So the compiler does an implicit conversion to float and then back to bf16 every time making it much slower. This adds a pass to do this directly in the graph, so we dont generate all these intermediate conversions.

Constant propagation also needs to cooperate: folding a widening convert
would materialize the literal in the wider type, enlarging it and losing the
narrow storage type, so those converts must stay as runtime conversions.

Technical Details

• Add a promote_storage_type pass that takes a list of storage-only types.
  For every pointwise or reduction instruction of such a type (detected via
  the pointwise/reduce attributes, excluding convert, bit_cast,
  layout, and identity, which carry the pointwise attribute but perform no
  computation), it converts the storage-typed inputs to float and converts the
  result back to the storage type. It then runs eliminate_convert, which
  cancels the storage/float convert pairs between adjacent promoted
  instructions, leaving the intermediate values in float.
• Generalize quantize_module from truncate_float.cpp into a reusable
  replace_data_type(module, src_types, target_type, predicate) template in
  include/migraphx/replace_data_type.hpp. Both truncate_float_pass and
  promote_storage_type are now thin wrappers over it, so the tuple-output
  and module-input handling is shared instead of duplicated.
• Update propagate_constant to skip folding converts to a wider type
  (mirroring the existing dequantizelinear handling), keeping literals in
  the narrow storage type with a runtime conversion. Narrowing converts still
  fold.
• Tests: unit tests for the new pass (pointwise, reduce, adjacent
  pointwise+reduce with no intermediate converts, where keeping its bool
  condition, unlisted types unchanged, idempotency) and for the
  propagate_constant change (widening convert of a broadcast literal is
  kept, narrowing convert still folds). The literal_add quantization test is
  updated for the new semantics: quantize_fp16 with {"add"} const-folds
  the add in fp32 before truncation, so nothing remains to quantize and the
  folded literal stays float, while the hand-built half program now keeps its
  half literal with a runtime widening convert.

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]

Pull request overview

This PR improves GPU performance for storage-only data types (notably bf16) by introducing a compiler pass that promotes eligible computations to run in float, reducing repeated implicit bf16↔float conversions and preserving narrow storage where appropriate during constant folding.

Changes:

• Added promote_storage_type pass to insert storage→float→storage conversions around pointwise/reduction computations and then cancel adjacent convert pairs via eliminate_convert.
• Refactored truncation/quantization type-rewrite logic into a reusable replace_data_type(...) helper template and rewired truncate_float_pass to use it.
• Updated constant propagation to avoid folding widening convert ops (to preserve narrow storage literals), and added/updated unit tests to match the new semantics.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
test/quantization.cpp	Updates literal_add expectations to reflect fp32 const-folding behavior before truncation/quantization.
test/propagate_constant_test.cpp	Adds regression coverage for skipping widening convert folding and still folding narrowing converts.
test/promote_storage_type_test.cpp	Adds unit tests for the new promote_storage_type pass (pointwise, reduce, adjacency, idempotency, where-condition handling).
src/truncate_float.cpp	Replaces the local quantization rewrite helper with the new shared replace_data_type(...) utility.
src/propagate_constant.cpp	Prevents constant folding of widening convert to keep narrow storage literals and avoid literal size/type expansion.
src/promote_storage_type.cpp	Implements the promote_storage_type module pass and runs eliminate_convert to remove redundant convert pairs.
src/include/migraphx/replace_data_type.hpp	Introduces the shared type-rewrite helper template used by both truncation and promotion.
src/include/migraphx/promote_storage_type.hpp	Declares the new promote_storage_type pass API.
src/CMakeLists.txt	Adds promote_storage_type.cpp to the core library build.

---

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Additional details and impacted files

@@           Coverage Diff            @@
##           develop    #5138   +/-   ##
========================================
  Coverage    93.13%   93.13%           
========================================
  Files          625      628    +3     
  Lines        33252    33270   +18     
========================================
+ Hits         30967    30985   +18     
  Misses        2285     2285           

Files with missing lines	Coverage Δ
src/include/migraphx/promote_storage_type.hpp	100.00% <100.00%> (ø)
src/include/migraphx/replace_data_type.hpp	100.00% <100.00%> (ø)
src/promote_storage_type.cpp	100.00% <100.00%> (ø)
src/propagate_constant.cpp	84.48% <100.00%> (+0.55%)	⬆️
src/truncate_float.cpp	100.00% <100.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[bdevorem]

lgtm

[TedThemistokleous]

Add a changelog entry for this