Update to new version with new 16x4 kleidi kernels (pytorch#6101)

Summary: * This xnnpack has kleidi v0.3.0 * Doesn't yet enable Kleidi yet by default `-DEXECUTORCH_XNNPACK_ENABLE_KLEIDI=OFF` * Minor fix for model load stat with warmup which was reporting 0.0 Pull Request resolved: pytorch#6101 Reviewed By: mergennachin Differential Revision: D64220150 Pulled By: digantdesai fbshipit-source-id: 98c8c4e079ee09dfd869970ec1abb9dccae29deb
kirklandsign · Oct 14, 2024 · 8957dc8 · 8957dc8
1 parent cd2d2b4
commit 8957dc8
Show file tree

Hide file tree

Showing 3 changed files with 12 additions and 5 deletions.
diff --git a/backends/xnnpack/third-party/XNNPACK b/backends/xnnpack/third-party/XNNPACK
diff --git a/backends/xnnpack/third-party/cpuinfo b/backends/xnnpack/third-party/cpuinfo
diff --git a/extension/llm/runner/stats.h b/extension/llm/runner/stats.h
@@ -59,9 +59,16 @@ struct ET_EXPERIMENTAL Stats {
     aggregate_sampling_timer_start_timestamp = 0;
   }
 
-  void reset() {
-    model_load_start_ms = 0;
-    model_load_end_ms = 0;
+  void reset(bool all_stats = false) {
+    // Not resetting model_load_start_ms and model_load_end_ms because reset is
+    // typically called after warmup and before running the actual run.
+    // However, we don't load the model again during the actual run after
+    // warmup. So, we don't want to reset these timestamps unless we are
+    // resetting everything.
+    if (all_stats) {
+      model_load_start_ms = 0;
+      model_load_end_ms = 0;
+    }
     inference_start_ms = 0;
     prompt_eval_end_ms = 0;
     first_token_ms = 0;
+1 −1		include/cpuinfo-mock.h
+54 −0		include/cpuinfo.h
+1 −1		src/arm/linux/aarch32-isa.c
+20 −2		src/arm/linux/aarch64-isa.c
+12 −6		src/arm/linux/api.h
+8 −8		src/arm/linux/hwcap.c
+4 −2		src/arm/linux/init.c