feat(rocke) : Ask the compiler which intrinsics each GPU can actually run, and commit the answer

dnn-providers/hip-kernel-provider/rocke/TESTING.md

@@ -324,6 +324,42 @@ Key semantics:
   ([`fuzz_diff.py`](platform/tests/instances/differential/fuzz_diff.py)) feeds the
   differential oracle with generated `(spec, arch)` inputs rather than a fixed list.
 
+### 4.4 Is the emitted IR legal? (the toolchain gate)
+
+§4.2 and §4.3 both compare rocKE's output against *another copy of rocKE's
+output* — a golden sha, or the other engine. Neither asks the AMDGPU toolchain
+whether the IR is valid at all, so an illegal construct that is emitted stably
+and by both engines is green in both. The **emitted-IR validity gate**,
+[`check_ir_validity.py`](platform/tools/check_ir_validity.py), asks the external
+question: it lowers every case in the representative corpus
+([`rocke_ir_parity_harness.cases()`](platform/tests/instances/rocke_ir_parity_harness.py),
+the same corpus §4.2 hashes) and pushes each module through `clang` to a linked
+hsaco.
+
+Three design points, each of them load-bearing:
+
+- **The oracle is a LINK, not a verify and not a codegen.** A `declare` for an
+  intrinsic that does not exist passes `opt -passes=verify`, *and* passes
+  `clang -S` — the backend silently treats the unknown `llvm.*` name as an
+  ordinary external function and emits a GOT-relative call. The undefined symbol
+  appears only when the relocatable is linked. Stopping anywhere earlier makes
+  the gate blind to the whole *fictional-intrinsic* bug class.
+- **Every compile is a subprocess.** A backend failure is a
+  `report_fatal_error`, not an exception: it takes the process down. Compiling
+  in-process (via comgr, which is otherwise faster) means one bad module kills
+  the run and emits no report. The isolation is a correctness requirement, and
+  it makes the per-module diagnostic free.
+- **Only the host's own LLVM flavor can be validated.** Modules lower at any
+  flavor, but there is no local compiler for the others; those report
+  `UNVALIDATED`, never green. On a host with no LLVM tools at all the gate
+  self-skips loudly — pass `--strict` to make that a failure instead.
+
+Failures that are known and owned live in a `KNOWN_BAD` allowlist, same
+convention and same rule as `KNOWN_VIOLATIONS` in
+[`test_library_layering.py`](library/tests/test_library_layering.py): **it only
+shrinks.** An entry that starts compiling is itself reported as a failure, so a
+fix cannot leave dead weight behind.
+
 ---
 
 ## 5. Execution tiers & gating
@@ -332,13 +368,13 @@ Four distinct things run here; **do not conflate them**:
 
 | Tier | What | Gated? |
 |---|---|---|
-| **1. Gate** | relative-path guard → byte-identity gate → pytest (`platform/tests`) → ctest | ✅ blocking |
+| **1. Gate** | relative-path guard → byte-identity gate → emitted-IR validity gate → pytest (`platform/tests`) → ctest | ✅ blocking |
 | **2. Diagnostics** | IR-canonical diff, fuzz diff, per-config golden check | ❌ opt-in |
 | **3. GPU / numeric** | reference-oracle kernel-correctness lanes | ❌ skipped off-device |
 | **4. Manual demos/tools** | hand-compiled CLIs / demos | ❌ |
 
 **Two entrypoints with different environments and selection.** [`run_all.py`](platform/tests/run_all.py) is
-the **developer** runner (guard → gate → pytest → ctest). **CI does not run
+the **developer** runner (guard → gate → IR validity → pytest → ctest). **CI does not run
 `run_all.py`** — it runs
 **ctest** against the installed artifact, after the component script installs
 the packaged wheels. Its registered pytest entries include the platform suite

dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt

@@ -391,7 +391,10 @@ if(ROCKE_INSTALL_TESTS)
     # Install the full pytest suite for installed/artifact test execution.
     # Excludes: C/C++ parity emitters (compiled against rocke_core at build time),
     # the differential harness (needs compiler + source), the WaveScope sidecar
-    # and LDS analyzer tests (load tools from the adjacent dsl_docs source tree).
+    # and LDS analyzer tests (load tools from the adjacent dsl_docs source tree),
+    # and the arch-domain artifact gate (imports tools/gen_arch_domain.py, which
+    # is not installed -- and an unresolvable import is a collection error that
+    # aborts the whole installed pytest session, not one skipped test).
     # Include run_all.py: installed runner unit tests import it as a dependency.
     # cmake-lint: disable=E1122
     block()
@@ -403,7 +406,8 @@ if(ROCKE_INSTALL_TESTS)
               PATTERN "*.cpp" EXCLUDE
               PATTERN "differential" EXCLUDE
               PATTERN "test_wavescope_sidecar.py" EXCLUDE
-              PATTERN "test_lds_analyzer.py" EXCLUDE)
+              PATTERN "test_lds_analyzer.py" EXCLUDE
+              PATTERN "test_arch_domain_artifact.py" EXCLUDE)
 
     endblock()
 

dnn-providers/hip-kernel-provider/rocke/platform/cpp/core/lower_llvm/data.cpp

@@ -277,9 +277,6 @@ const rocke_ll_decl_t ROCKE_LL_INTRINSIC_DECLS[] = {
     {"update.dpp.i32",
      "declare i32 @llvm.amdgcn.update.dpp.i32(i32, i32, i32 immarg, i32 immarg, i32 immarg, i1 "
      "immarg)"},
-    {"global.atomic.fadd.v2bf16",
-     "declare <2 x bfloat> @llvm.amdgcn.global.atomic.fadd.v2bf16.p1("
-     "ptr addrspace(1), <2 x bfloat>)"},
     {"global.atomic.fadd.v2f16",
      "declare <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1("
      "ptr addrspace(1), <2 x half>)"},
@@ -344,10 +341,6 @@ const rocke_ll_decl_t ROCKE_LL_INTRINSIC_DECLS[] = {
      "declare <2 x float> @llvm.amdgcn.cvt.scalef32.pk.f32.fp8(i32, float, i1)"},
     {"amdgcn.cvt.scalef32.pk.f32.bf8",
      "declare <2 x float> @llvm.amdgcn.cvt.scalef32.pk.f32.bf8(i32, float, i1)"},
-    {"amdgcn.cvt.scalef32.pk.fp8.f32",
-     "declare i32 @llvm.amdgcn.cvt.scalef32.pk.fp8.f32(i32, <2 x float>, float, i1)"},
-    {"amdgcn.cvt.scalef32.pk.bf8.f32",
-     "declare i32 @llvm.amdgcn.cvt.scalef32.pk.bf8.f32(i32, <2 x float>, float, i1)"},
     {"amdgcn.ds.swizzle", "declare i32 @llvm.amdgcn.ds.swizzle(i32, i32 immarg)"},
     /* Not overloaded, so no name suffix, but the flags are immarg like every
      * other permlane* flag pair. */

dnn-providers/hip-kernel-provider/rocke/platform/cpp/instances/gfx942/attention_tiled_2d_public_entry_glue.cpp

@@ -580,6 +580,9 @@ static rocke_status_t
             hkv_buf,
             s->dtype ? s->dtype : "",
             kv_buf,
+            /* Python emits "fnuz" under the *same* condition as the kv part,
+             * not under a format test, so it keys off kv_buf being non-empty. */
+            kv_buf[0] ? "fnuz" : "",
             s->use_sinks ? "sinks" : "",
             sw_buf,
             s->has_softcap ? "softcap" : "",

dnn-providers/hip-kernel-provider/rocke/platform/cpp/instances/gfx942/attention_tiled_3d_gfx942_attn_tiled_3d_public_glue.cpp

@@ -231,7 +231,8 @@ int rocke_unified_attention_3d_tiled_spec_kernel_name(
     const rocke_unified_attention_3d_tiled_spec_t* s, char* buf, size_t cap)
 {
     /* kernel_name_join("rocke_uattn3d_tiled_gfx942", d.., b.., h..kv.., seg..,
-     *   dtype, kv<...> if kv_storage_dtype else "", "sinks" if use_sinks else "",
+     *   dtype, kv<...> if kv_storage_dtype else "", "fnuz" if kv_storage_dtype
+     *   else "", "sinks" if use_sinks else "",
      *   "sw<sw>" if sw>0 else "", "softcap" if has_softcap else "",
      *   "alibi" if use_alibi else "", "qqb" if use_qq_bias else "",
      *   "hoist" if use_invariant_hoist else "", "wkv" if use_wide_kv_load else "")
@@ -270,6 +271,9 @@ int rocke_unified_attention_3d_tiled_spec_kernel_name(
     {
         snprintf(kv_part, sizeof(kv_part), "kv%s", s->kv_storage_dtype);
         parts[np++] = kv_part;
+        /* Python emits "fnuz" under the *same* condition as the kv part, not
+         * under a format test, so it is tagged here rather than separately. */
+        parts[np++] = "fnuz";
     }
     if(s->use_sinks)
     {