[afagaj]

JIRA ID : AICK-3015

Summary: The C++ SDPA backward builders now refuse the same resource-gate specs Python refuses, using a C++ port of spec_digest and of both refusal tables.

Stacked on #14365; retarget to develop after it merges.

Motivation

Review comment on #14365: #14365 (comment)

Python keeps two tables of specs that compile but fail the resource gate (spills, AGPR copy traffic) or crash the code-object compiler: RESOURCE_REFUSED (2,283 main-kernel rows) and DQ_RESOURCE_REFUSED (144 dQ rows), both keyed by (arch, spec_digest(spec, arch)). In #14365 the C++ main builder skipped the main table (it mirrored _emit_main(..., main_kernel_refusals=False)) and the C++ dQ validator skipped the dQ table. Only comments said so, so a future C++ caller would build and launch those kernels without noticing.

Change

• platform/cpp/instances/common/_attention_bwd_refusals.cpp (new): sha256_hex, spec_digest / dq_spec_digest (the existing non_default_fields plus head_size, sorted, Python repr, SHA-256, first 16 hex digits), and binary-search lookups resource_refusal / dq_resource_refusal.
• _attention_bwd_refusals_table.inc (new, generated): both tables, written by library/tests/parity/attention_bwd_refusals_emit.py --print-c-table.
• Main validator and builder take main_kernel_refusals, with no default, so every caller has to choose:
  • rocke_attn_bwd_validate_spec(spec, arch, main_kernel_refusals, out, err) and rocke_build_attn_bwd_main_new(b, spec, arch, main_kernel_refusals). With true (Python's default) a refused spec is ROCKE_ERR_VALUE with Python's exact message (resource gate: this resolved spec fails on <arch> (<record>; RESOURCE_REFUSED)).
  • The dense and unified shells pass false, as their Python twins do. The IR parity emitters (attention_bwd_main/body/spec_emit.c) pass false, matching their Python side.
• The dQ validator always applies DQ_RESOURCE_REFUSED between DQ_SPILLING_COMBOS and DQ_SCHED_SPILLING, in Python's order. The dense dQ validator inherits this, as in Python.
• New C entry points: rocke_attn_bwd_spec_digest, rocke_attn_bwd_resource_refusal, rocke_attn_bwd_dq_spec_digest, rocke_attn_bwd_dq_resource_refusal.
• Headers, attention_bwd.md section 0.3 and the parity-emitter comments now describe the port instead of the gap.

Testing

• New parity pair attention_bwd_refusals_emit.{py,c}: 68 configs (11 main/dQ defaults, knob specs and early rejections, plus 50 main and 7 dQ recorded refusals covering every target, every failure kind and every spec field that appears in the recorded rows). Each config prints the spec digest, the resolved digest, the table record and the validator verdicts. C and Python agree on all 68. Enrolled in NativeGoldenTests.cmake, with goldens recorded at llvm20/22/23/24.
• New library/tests/test_attention_bwd_refusals_cpp_table.py: the committed .inc must equal the generator output (so all 2,283 + 144 rows match).
• check_byte_identity.py --only attention_bwd,attention_dense_bwd (llvm22): GREEN, 21 families, including attention_bwd_refusals.
• ctest: 50/50 passed (includes rocke_native_golden_ll).
• run_diff.py --python-only --check-golden at llvm20/22/23/24: GOLDEN OK, 0 unblessed changes.
• pytest test_attention_bwd_refusals_cpp_table.py, test_attention_bwd_refusals.py, test_attention_bwd_layout_spill_rows.py: all passed.
• pre-commit run --from-ref <stack head> --to-ref HEAD: clean.
• Negative checks:
  • Making the C++ main validator ignore the table made rocke_native_golden_ll fail on attention_bwd_refusals at all four flavors. Restoring it brought it back to pass.
  • Deleting one row from the .inc failed both table tests. Restoring it brought them back to pass.

🤖 Generated with Claude Code