Skip to content

fix: restore getCaptureStatus() — was hardcoded to return None#1210

Open
pvelesko wants to merge 3 commits intomainfrom
fix-hipgraph-stub
Open

fix: restore getCaptureStatus() — was hardcoded to return None#1210
pvelesko wants to merge 3 commits intomainfrom
fix-hipgraph-stub

Conversation

@pvelesko
Copy link
Collaborator

Summary

getCaptureStatus() in CHIPBackend.hh had its real implementation commented out with a TODO, hardcoded to always return hipStreamCaptureStatusNone. This caused hipStreamEndCapture to always fail with hipErrorIllegalState (401).

Impact

  • 384 rocSPARSE graph_test failures (all stream-capture-based tests)
  • rocRAND test_rocrand_hipgraphs
  • Any library test using hipStreamBeginCapture/hipStreamEndCapture

Test plan

  • Standalone stream capture test: begin + end returns hipSuccess
  • Stream capture with memset: graph instantiate + launch works
  • rocSPARSE axpby/scatter/gather graph tests: 680/680 passing

LLVM 22 SPIR-V triggers an IGC driver bug on i915 igpus with emulated
double precision (IGC_EnableDPEmulation). The testInt2Double kernel
causes ZE_RESULT_ERROR_DEVICE_LOST / GPU hang on UHD 730/770.

The SPIR-V for the kernel itself is identical to LLVM 21, but the
module-level SPIR-V differences (rounding-mode conversion kernels)
cause IGC to produce bad machine code for the igpu.

Test passes on dgpu (Arc A380/A770) and on igpu with LLVM 21.
Excluded via --regex-exclude on igpu only, until IGC is fixed.
@pvelesko pvelesko force-pushed the fix-hipgraph-stub branch from 3ba5085 to ba0f14d Compare March 19, 2026 07:37
getCaptureStatus() had its real implementation commented out with a
TODO, always returning hipStreamCaptureStatusNone. This caused
hipStreamEndCapture to always fail with hipErrorIllegalState (401).

Unblocks 384 rocSPARSE graph_test failures and rocRAND hipgraph test.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant