From c054608df3a03c5988603713038bef14967971be Mon Sep 17 00:00:00 2001
From: James Newling <james.newling@gmail.com>
Date: Fri, 28 Feb 2025 09:22:39 -0800
Subject: [PATCH] first commit

---
 build_tools/ci/cpu_comparison/run.py          | 22 ++-----------------
 .../AMD-AIE/iree-amd-aie/Target/XCLBinGen.cpp |  2 +-
 2 files changed, 3 insertions(+), 21 deletions(-)

diff --git a/build_tools/ci/cpu_comparison/run.py b/build_tools/ci/cpu_comparison/run.py
index f081a2b4f..c56a71ab6 100755
--- a/build_tools/ci/cpu_comparison/run.py
+++ b/build_tools/ci/cpu_comparison/run.py
@@ -2034,7 +2034,6 @@ def __init__(self):
                 "N": 512,
                 "K": 4096,
                 "use_ukernel": False,
-                "peano_opt_level": 2,
                 "outline": "none",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2045,7 +2044,6 @@ def __init__(self):
                 "N": 512,
                 "K": 4096,
                 "use_ukernel": False,
-                "peano_opt_level": 2,
                 "outline": "balanced",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2089,7 +2087,6 @@ def __init__(self):
                 "N": 4096,
                 "K": 512,
                 "use_ukernel": False,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2100,7 +2097,6 @@ def __init__(self):
                 "N": 4096,
                 "K": 512,
                 "use_ukernel": True,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2111,7 +2107,6 @@ def __init__(self):
                 "N": 4096,
                 "K": 512,
                 "use_ukernel": False,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "transpose_a": False,
                 "transpose_b": True,
@@ -2122,7 +2117,6 @@ def __init__(self):
                 "N": 512,
                 "K": 512,
                 "use_ukernel": False,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2133,7 +2127,6 @@ def __init__(self):
                 "N": 512,
                 "K": 512,
                 "use_ukernel": True,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2144,7 +2137,6 @@ def __init__(self):
                 "N": 512,
                 "K": 512,
                 "use_ukernel": False,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "transpose_a": True,
                 "transpose_b": False,
@@ -2158,7 +2150,6 @@ def __init__(self):
                 "N": 512,
                 "K": 512,
                 "use_ukernel": False,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "outline_to_empty_function": True,
                 "transpose_a": False,
@@ -2171,7 +2162,6 @@ def __init__(self):
                 "N": 4096,
                 "K": 512,
                 "use_ukernel": False,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2182,7 +2172,6 @@ def __init__(self):
                 "N": 4096,
                 "K": 512,
                 "use_ukernel": True,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2193,7 +2182,6 @@ def __init__(self):
                 "N": 4096,
                 "K": 512,
                 "use_ukernel": True,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2208,7 +2196,6 @@ def __init__(self):
                 "N": 4096,
                 "K": 512,
                 "use_ukernel": False,
-                "peano_opt_level": 3,
                 "outline": "balanced",
                 "outline_to_empty_function": True,
                 "transpose_a": False,
@@ -2226,7 +2213,6 @@ def __init__(self):
                 "in_dtype": "i8",
                 "out_dtype": "i32",
                 "use_ukernel": True,
-                "peano_opt_level": 3,
                 "outline": "all",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2240,7 +2226,6 @@ def __init__(self):
                 "in_dtype": "i8",
                 "out_dtype": "i32",
                 "use_ukernel": False,
-                "peano_opt_level": 3,
                 "outline": "all",
                 "outline_to_empty_function": True,
                 "transpose_a": False,
@@ -2256,7 +2241,6 @@ def __init__(self):
                 "in_dtype": "i8",
                 "out_dtype": "i32",
                 "use_ukernel": True,
-                "peano_opt_level": 3,
                 "outline": "all",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2270,7 +2254,6 @@ def __init__(self):
                 "in_dtype": "i8",
                 "out_dtype": "i32",
                 "use_ukernel": True,
-                "peano_opt_level": 3,
                 "outline": "all",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2285,7 +2268,6 @@ def __init__(self):
                 "in_dtype": "i8",
                 "out_dtype": "i32",
                 "use_ukernel": False,
-                "peano_opt_level": 3,
                 "outline": "all",
                 "outline_to_empty_function": True,
                 "transpose_a": False,
@@ -2301,7 +2283,6 @@ def __init__(self):
                 "in_dtype": "i8",
                 "out_dtype": "i32",
                 "use_ukernel": True,
-                "peano_opt_level": 3,
                 "outline": "all",
                 "transpose_a": False,
                 "transpose_b": False,
@@ -2317,7 +2298,8 @@ def __init__(self):
             N = test["N"]
             K = test["K"]
             use_ukernel = test["use_ukernel"]
-            peano_opt_level = test["peano_opt_level"]
+            # Default optimization level for benchmarking is 2:
+            peano_opt_level = test.get("peano_opt_level", 2)
             outline = test["outline"]
             transpose_a = test["transpose_a"]
             transpose_b = test["transpose_b"]
diff --git a/compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/XCLBinGen.cpp b/compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/XCLBinGen.cpp
index 49b58760c..89752396a 100644
--- a/compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/XCLBinGen.cpp
+++ b/compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/XCLBinGen.cpp
@@ -212,7 +212,7 @@ FailureOr<std::vector<std::string>> makePeanoOptArgs(
       // Extend the max limit of the search depth in BasicAA
       "-basic-aa-max-lookup-search-depth=10",
       //
-      "-O3",
+      "-O2",
       //
       "--inline-threshold=10",
       // missing from libc