Add ROCm release fuzzing wrapper

Author: jlebar

amdgpu/README.md

@@ -54,6 +54,17 @@ Scaling from 64 to 256 workers did not reach 1k exec/s; CPU remained mostly
 idle, so the current bottleneck appears to be HIP/module execution rather than
 host compilation.
 
+For ROCm 7.2.3 release fuzzing, use the release wrapper:
+
+```bash
+scripts/run_rocm_7_2_3_release_fuzzer.sh -max_total_time=900 -max_len=1024 -rss_limit_mb=8192 -use_value_profile=1
+```
+
+That wrapper keeps the release-reproducing bugs suppressed (`m001`, `m013`, and
+`m017`), but explicitly re-enables the idioms that only failed on LLVM HEAD /
+ROCm HEAD in the checked matrix (`m002` through `m012`, plus `m014` through
+`m016`).
+
 Candidate compiler crashes, runner failures, or output mismatches are saved
 under `findings/`. Generated corpora and findings are local artifacts and are
 ignored by git.
@@ -92,6 +103,7 @@ rediscovering the same issue.
 | `scripts/build_directed_fuzzer.sh` | Builds the C++ GPU differential libFuzzer target. |
 | `scripts/run_directed_fuzzer.sh` | Runs the C++ directed fuzzer on one GPU. |
 | `scripts/run_directed_multigpu_fuzzer.sh` | Runs one or more C++ directed fuzzer processes per selected GPU. |
+| `scripts/run_rocm_7_2_3_release_fuzzer.sh` | Runs the C++ directed fuzzer with the ROCm 7.2.3 release suppression policy. |
 | `fuzzer/` | LLVM API plus HIP differential libFuzzer target. |
 | `runner/hip_module_runner.cpp` | HIP module loader used to execute generated HSACO files. |
 | `known-miscompiles/` | Reduced or standalone reproducers for confirmed findings. |

amdgpu/scripts/run_rocm_7_2_3_release_fuzzer.sh

@@ -0,0 +1,40 @@
+#!/usr/bin/env bash
+# Run the GPU differential fuzzer against ROCm 7.2.3 release semantics.
+#
+# This mode re-enables known idioms that reproduce only on LLVM HEAD / ROCm HEAD,
+# while keeping ROCm 7.2.3 release failures suppressed.
+
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+ROOT="$(pwd)"
+
+FUZZER_BIN="${FUZZER_BIN:-$ROOT/build/fuzzer-rocm-7.2.3-cov-release/llvm_amdgpu_diff_fuzzer}"
+CORPUS_ROOT="${CORPUS_ROOT:-$ROOT/corpus/rocm-7.2.3-cov-release-fuzz}"
+ARTIFACT_ROOT="${ARTIFACT_ROOT:-$ROOT/findings/rocm-7.2.3-cov-release-fuzz-artifacts}"
+FUZZX_FINDINGS_DIR="${FUZZX_FINDINGS_DIR:-$ROOT/findings/rocm-7.2.3-cov-release-fuzz}"
+LOG_DIR="${LOG_DIR:-$ROOT/build/rocm-7.2.3-cov-release-fuzz-logs/$(date +%Y%m%d-%H%M%S)}"
+GPUS="${GPUS:-0 1 2 3 4 5 6 7}"
+WORKERS_PER_GPU="${WORKERS_PER_GPU:-32}"
+
+# These known cases are HEAD-only for the checked toolchain matrix and should
+# stay enabled while fuzzing the ROCm release.
+export FUZZX_ALLOW_M002_I8_CLEAR_XOR=1
+export FUZZX_ALLOW_M003_SHL3_ADD_CHAIN=1
+export FUZZX_ALLOW_M004_VECTOR_IDENTITY_XOR=1
+export FUZZX_ALLOW_M005_SHL_ADD_CHAIN=1
+export FUZZX_ALLOW_M006_I8_CLEAR_XOR=1
+export FUZZX_ALLOW_M007_VECTOR_IDENTITY_XOR=1
+export FUZZX_ALLOW_M008_I8_CLEAR_XOR=1
+export FUZZX_ALLOW_M009_I16_CLEAR_XOR=1
+export FUZZX_ALLOW_M010_I16_SEXT_CLEAR_XOR=1
+export FUZZX_ALLOW_M011_I8_SEXT_CLEAR_XOR=1
+export FUZZX_ALLOW_M012_ADD_SHL_LADDER=1
+export FUZZX_ALLOW_M014_SHL_ADD_CTPOP=1
+export FUZZX_ALLOW_M015_SCALAR_FSHL_ZERO=1
+export FUZZX_ALLOW_M016_SCALAR_FSHL=1
+
+export FUZZER_BIN CORPUS_ROOT ARTIFACT_ROOT FUZZX_FINDINGS_DIR LOG_DIR GPUS
+export WORKERS_PER_GPU
+
+exec "$ROOT/scripts/run_directed_multigpu_fuzzer.sh" "$@"