From e1dd02625a45a7549b4c39a7e479152e79340ec9 Mon Sep 17 00:00:00 2001
From: Andrew Adams <andrew.b.adams@gmail.com>
Date: Tue, 24 Dec 2024 13:46:16 -0800
Subject: [PATCH 1/2] Hopefully fix flaky test

---
 test/autoschedulers/mullapudi2016/reorder.cpp | 24 +++++++++++--------
 1 file changed, 14 insertions(+), 10 deletions(-)

diff --git a/test/autoschedulers/mullapudi2016/reorder.cpp b/test/autoschedulers/mullapudi2016/reorder.cpp
index 37476a0ff0dc..b383312e12a4 100644
--- a/test/autoschedulers/mullapudi2016/reorder.cpp
+++ b/test/autoschedulers/mullapudi2016/reorder.cpp
@@ -7,10 +7,17 @@ using namespace Halide::Tools;
 double run_test_1(bool auto_schedule) {
     Var x("x"), y("y"), dx("dx"), dy("dy"), c("c");
 
+    int W = 1024;
+    int H = 1920;
+    int search_area = 7;
+
+    Buffer<uint32_t> im(2048);
+    im.fill(17);
+
+
     Func f("f");
-    f(x, y, dx, dy) = x + y + dx + dy;
+    f(x, y, dx, dy) = im(x) + im(y + 1) + im(dx + search_area / 2) + im(dy + search_area / 2);
 
-    int search_area = 7;
     RDom dom(-search_area / 2, search_area, -search_area / 2, search_area, "dom");
 
     // If 'f' is inlined into 'r', the only storage layout that the auto scheduler
@@ -23,23 +30,20 @@ double run_test_1(bool auto_schedule) {
 
     if (auto_schedule) {
         // Provide estimates on the pipeline output
-        r.set_estimates({{0, 1024}, {0, 1024}, {0, 3}});
+        r.set_estimates({{0, W}, {0, H}, {0, 3}});
         // Auto-schedule the pipeline
         p.apply_autoscheduler(target, {"Mullapudi2016"});
     } else {
-        /*
+        Var par;
         r.update(0).fuse(c, y, par).parallel(par).reorder(x, dom.x, dom.y).vectorize(x, 4);
-        r.fuse(c, y, par).parallel(par).vectorize(x, 4); */
-
-        // The sequential schedule in this case seems to perform best which is
-        // odd have to investigate this further.
+        r.fuse(c, y, par).parallel(par).vectorize(x, 4);
     }
 
     // Inspect the schedule (only for debugging))
     // r.print_loop_nest();
 
     // Run the schedule
-    Buffer<int> out(1024, 1024, 3);
+    Buffer<int> out(W, H, 3);
     double t = benchmark(3, 10, [&]() {
         p.realize(out);
     });
@@ -154,7 +158,7 @@ int main(int argc, char **argv) {
         double manual_time = run_test_1(false);
         double auto_time = run_test_1(true);
 
-        const double slowdown_factor = 15.0;  // TODO: whoa
+        const double slowdown_factor = 2.0;
         if (!get_jit_target_from_environment().has_gpu_feature() && auto_time > manual_time * slowdown_factor) {
             std::cerr << "Autoscheduler time (1) is slower than expected:\n"
                       << "======================\n"

From eac3a293466385d069f678917e6ac3f9b34aa783 Mon Sep 17 00:00:00 2001
From: Andrew Adams <andrew.b.adams@gmail.com>
Date: Tue, 24 Dec 2024 14:01:34 -0800
Subject: [PATCH 2/2] clang-format

---
 test/autoschedulers/mullapudi2016/reorder.cpp | 1 -
 1 file changed, 1 deletion(-)

diff --git a/test/autoschedulers/mullapudi2016/reorder.cpp b/test/autoschedulers/mullapudi2016/reorder.cpp
index b383312e12a4..942da121b8b1 100644
--- a/test/autoschedulers/mullapudi2016/reorder.cpp
+++ b/test/autoschedulers/mullapudi2016/reorder.cpp
@@ -14,7 +14,6 @@ double run_test_1(bool auto_schedule) {
     Buffer<uint32_t> im(2048);
     im.fill(17);
 
-
     Func f("f");
     f(x, y, dx, dy) = im(x) + im(y + 1) + im(dx + search_area / 2) + im(dy + search_area / 2);