test(miopen): add CTC loss numel>INT_MAX scale coverage (ALMIOPEN-2148)

[peymanr]

Summary

Add a new gtest file projects/miopen/test/gtest/ctc_large.cpp with a Scale-tier test that exercises MIOpen CTC loss at logit numel > INT_MAX.

Motivation

The existing gtest/ctc.cpp cases top out at approximately 64 M elements (batch=128, T=100, C=5001). MIOpen's CTC kernel accumulates the logit tensor dimensions through plain int locals in ctc.cpp:

int problog_offset = lb_prime_offset + batch_size * max_S_len;
int alpha_offset   = problog_offset + class_sz * batch_size * max_time_step;

When T * N * C exceeds INT_MAX (2 147 483 647), these multiplications wrap to negative values, producing out-of-bounds workspace reads/writes and silently wrong losses. Large-vocabulary or long-sequence ASR models (e.g. T=1000, vocab=50k, batch=128) are in this regime.

JIRA ID: ALMIOPEN-2148

What the test does

Parameter	Value	Notes
T (time steps)	1000	Matches ASR frame budget
N (batch)	2048	Large batch to cross the boundary
C (classes, excl. blank)	1048	class_sz = C+1 = 1049
numel = T × N × (C+1)	2 148 352 000	> INT_MAX = 2 147 483 647
Label length	2 per sample	Kept small so alpha workspace is bounded
Input length	10 per sample	Actual used time steps

• Queries hipMemGetInfo and calls GTEST_SKIP() when free GPU memory < 19 GiB; safe to run on smaller devices.
• Allocates logits/gradients/losses directly on GPU (no 8.59 GiB host buffer).
• Probs initialised with hipMemset(0) — uniform logits give a valid softmax input.
• Asserts every returned loss is finite. NaN or Inf signals integer overflow in the workspace offset computation.
• No CPU reference comparison (host-side 8.59 GiB allocation is prohibitive).

MIOpenDriver equivalent for manual verification on TW325 (gfx942):

MIOpenDriver ctc -T 1000 -N 2048 -C 1048 --forw 1 -V 0 -t 1

Test tier

The test uses a TEST_F fixture with a Scale_GPU_CTC_numel_gt_INT_MAX case name. It is picked up automatically by file(GLOB TESTS *.cpp) in the gtest CMakeLists.txt — no CMakeLists.txt modification required. The Scale_ prefix and the memory guard prevent accidental execution in Smoke/Standard CI.

Checklist

• New file only; no existing file modified
• GTEST_SKIP() guard for insufficient GPU memory
• static_assert confirms numel > INT_MAX at compile time
• JIRA ID: ALMIOPEN-2148 included in commit message and PR body

🤖 Assisted by PR Pundit

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	PR does not contain code files — Unit Test auto-passed
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

Pre-commit check failed

⛔ pre-commit failed

Please run locally:

• python -m pip install pre-commit
• pre-commit install
• pre-commit run --all-files --show-diff-on-failure

This repo uses .pre-commit-config.yaml.

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.