autotune

.github/workflows/format.yaml

@@ -43,12 +43,28 @@ jobs:
       - name: Install clang-format
         run: |
           sudo apt-get update
-          sudo apt-get install -y clang-format-17
+          sudo apt-get install -y clang-format-15
 
       - name: Check changed files
         run: |
+          set -euo pipefail
+
+          if [ "${{ github.event_name }}" = "pull_request" ]; then
+            diff_range="${{ github.event.pull_request.base.sha }}...HEAD"
+          elif [ "${{ github.event_name }}" = "push" ] && \
+               [ -n "${{ github.event.before }}" ] && \
+               [ "${{ github.event.before }}" != "0000000000000000000000000000000000000000" ]; then
+            diff_range="${{ github.event.before }}..${{ github.sha }}"
+          else
+            default_branch="${{ github.event.repository.default_branch }}"
+            git fetch --no-tags --prune origin \
+              "+refs/heads/${default_branch}:refs/remotes/origin/${default_branch}"
+            diff_range="origin/${default_branch}...HEAD"
+          fi
+
+          echo "Checking files changed in ${diff_range}"
           git diff --diff-filter=d --name-only \
-            origin/${{ github.base_ref }}...HEAD \
+            "${diff_range}" -- \
             > all_changed_files.txt
 
           awk '/\.(cc|cpp|cxx|h|hpp)$/ && $0 !~ /^third_party\// { print }' \
@@ -60,7 +76,7 @@ jobs:
             exit 0
           fi
 
-          xargs -a changed_files.txt clang-format-17 --dry-run -Werror
+          xargs -r -a changed_files.txt clang-format-15 --dry-run -Werror
 
   python-black:
     needs: markdownlint

backend/ascend_autotune_hooks.py

@@ -48,7 +48,7 @@ def _ascend_autotune_fn():
     """Lazy singleton accessor for ascend ``autotune``."""
     global _ASCEND_AUTOTUNE
     if _ASCEND_AUTOTUNE is None:
-        from .ascend_autotune_runtime.autotuner import autotune as _fn
+        from .ascend_autotune_runtime.ascend_kernel_autotuner import autotune as _fn
 
         _ASCEND_AUTOTUNE = _fn
     return _ASCEND_AUTOTUNE
@@ -58,7 +58,7 @@ def _ascend_max_autotune_fn():
     """Lazy singleton accessor for ascend ``max_autotune``."""
     global _ASCEND_MAX_AUTOTUNE
     if _ASCEND_MAX_AUTOTUNE is None:
-        from .ascend_autotune_runtime.autotuner import max_autotune as _fn
+        from .ascend_autotune_runtime.ascend_kernel_autotuner import max_autotune as _fn
 
         _ASCEND_MAX_AUTOTUNE = _fn
     return _ASCEND_MAX_AUTOTUNE
@@ -68,7 +68,7 @@ def _ascend_max_autotune_fn():
 # Proxies (installed once on import of this module)
 # ------------------------------------------------------------------
 def _autotune_proxy(configs, key, **kwargs):
-    if _is_ascend_backend():
+    if _is_ascend_backend() or kwargs.get("hints") is not None:
         return _ascend_autotune_fn()(configs=configs, key=key, **kwargs)
     from triton.runtime.autotuner import autotune as _stock
 

backend/ascend_autotune_runtime/__init__.py

@@ -4,7 +4,7 @@
 Adapted for DLCompiler (triton.backends.dicp_triton).
 """
 
-from .autoparser import (
+from .kernel_ast_analyzer import (
     AutoParser,
     AxesKeyParser,
     SplitAxesParser,
@@ -13,13 +13,22 @@
     LowDimsAxesParser,
     PtrNumsParser,
 )
-from .tile_generator import AxisInfo, BlockInfo, KernelMeta, TileGenerator
-from .compile_options import (
+from .tile_candidate_generator import AxisInfo, BlockInfo, KernelMeta, TileGenerator
+from .schedule_profiles import (
     CompileOptionsSpec,
+    CompileFailureRegionSet,
+    classify_compile_failure,
+    compile_profile_to_config,
+    effective_compile_profile_key,
     expand_compile_option_configs,
+    generate_linked_compile_neighbors,
+    get_stage1_probe_configs,
+    get_stage1_probe_profiles,
+    make_stage2_seed_profiles,
     parse_compile_options_hint,
+    validate_compile_profile,
 )
-from .autotuner import (
+from .ascend_kernel_autotuner import (
     AutoTilingTuner,
     autotune,
     max_autotune,
@@ -46,8 +55,17 @@
     "KernelMeta",
     "TileGenerator",
     "CompileOptionsSpec",
+    "CompileFailureRegionSet",
+    "classify_compile_failure",
+    "compile_profile_to_config",
+    "effective_compile_profile_key",
     "expand_compile_option_configs",
+    "generate_linked_compile_neighbors",
+    "get_stage1_probe_configs",
+    "get_stage1_probe_profiles",
+    "make_stage2_seed_profiles",
     "parse_compile_options_hint",
+    "validate_compile_profile",
     "AutoTilingTuner",
     "autotune",
     "max_autotune",