diff --git a/tests/test-conv2d.cpp b/tests/test-conv2d.cpp
index 41807a6b80..4bb6c692ba 100644
--- a/tests/test-conv2d.cpp
+++ b/tests/test-conv2d.cpp
@@ -297,34 +297,351 @@ std::vector<float> compute_graph(const test_model & model, ggml_gallocr_t allocr
 int main(void)
 {
     ggml_time_init();
+
+    double time_iter0 = 0.0, time_iter1 = 0.0;
     std::vector<std::tuple<int, int, int, int, int, int>> configs = {
-        std::make_tuple(64,64,48,64,3,3),
-        std::make_tuple(320,320,104,152,3,3),
-        std::make_tuple(640,640,52,76,3,3),
-        std::make_tuple(640,640,104,152,3,3),
-        std::make_tuple(960,320,104,152,3,3),
-        std::make_tuple(1280,1280,26,38,3,3),
-        std::make_tuple(4,320,96,128,3,3),
-        std::make_tuple(320,4,96,128,3,3),
-        std::make_tuple(4,320,64,96,3,3),
-        std::make_tuple(320,4,64,96,3,3),
-        std::make_tuple(640,640,96,128,3,3),
-        std::make_tuple(1280,1280,26,38,1,1),
-        std::make_tuple(256,128,768,1024,3,3),
-        std::make_tuple(128,3,768,1024,3,3),
-        std::make_tuple(256,128,768,1024,1,1),
-        std::make_tuple(512,256,384,512,1,1),
-        std::make_tuple(1280,640,52,76,3,3),
-        std::make_tuple(1920,1280,26,38,3,3),
-        std::make_tuple(2560,1280,26,38,3,3),
-        std::make_tuple(320,1280,26,38,3,3),
-        std::make_tuple(512,512,104,152,3,3),
-        std::make_tuple(512,512,208,304,3,3),
-        std::make_tuple(512,256,416,608,3,3),
-        std::make_tuple(256,128,832,1216,3,3),
-        std::make_tuple(256,256,832,1216,3,3),
+        // std::make_tuple(64,64,48,64,3,3),
+        // std::make_tuple(320,320,104,152,3,3),
+        // std::make_tuple(640,640,52,76,3,3),
+        // std::make_tuple(640,640,104,152,3,3),
+        // std::make_tuple(960,320,104,152,3,3),
+        // std::make_tuple(1280,1280,26,38,3,3),
+        // std::make_tuple(4,320,96,128,3,3),
+        // std::make_tuple(320,4,96,128,3,3),
+        // std::make_tuple(4,320,64,96,3,3),
+        // std::make_tuple(320,4,64,96,3,3),
+        // std::make_tuple(640,640,96,128,3,3),
+        // std::make_tuple(1280,1280,26,38,1,1),
+        // std::make_tuple(256,128,768,1024,3,3),
+        // std::make_tuple(128,3,768,1024,3,3),
+        // std::make_tuple(256,128,768,1024,1,1),
+        // std::make_tuple(512,256,384,512,1,1),
+        // std::make_tuple(1280,640,52,76,3,3),
+        // std::make_tuple(1920,1280,26,38,3,3),
+        // std::make_tuple(2560,1280,26,38,3,3),
+        // std::make_tuple(320,1280,26,38,3,3),
+        // std::make_tuple(512,512,104,152,3,3),
+        // std::make_tuple(512,512,208,304,3,3),
+        // std::make_tuple(512,256,416,608,3,3),
+        // std::make_tuple(256,128,832,1216,3,3),
+        // std::make_tuple(256,256,832,1216,3,3),
         // std::make_tuple(320,256,1024,1920)
         // std::make_tuple(32,64,58,58,3,3)
+
+        //512x512
+        std::make_tuple(4,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(320,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(640,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(640,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(2560,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(2560,1280,16,16,3,3),
+        std::make_tuple(2560,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(2560,1280,16,16,3,3),
+        std::make_tuple(1920,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1920,1280,16,16,3,3),
+        std::make_tuple(1280,1280,32,32,3,3),
+        std::make_tuple(1920,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(1920,640,32,32,3,3),
+        std::make_tuple(1280,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(1280,640,32,32,3,3),
+        std::make_tuple(960,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(960,640,32,32,3,3),
+        std::make_tuple(640,640,64,64,3,3),
+        std::make_tuple(960,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(960,320,64,64,3,3),
+        std::make_tuple(640,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(640,320,64,64,3,3),
+        std::make_tuple(640,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(640,320,64,64,3,3),
+        std::make_tuple(320,4,64,64,3,3),
+        std::make_tuple(4,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(320,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(320,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(640,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(640,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(2560,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(2560,1280,16,16,3,3),
+        std::make_tuple(2560,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(2560,1280,16,16,3,3),
+        std::make_tuple(1920,1280,16,16,3,3),
+        std::make_tuple(1280,1280,16,16,3,3),
+        std::make_tuple(1920,1280,16,16,3,3),
+        std::make_tuple(1280,1280,32,32,3,3),
+        std::make_tuple(1920,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(1920,640,32,32,3,3),
+        std::make_tuple(1280,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(1280,640,32,32,3,3),
+        std::make_tuple(960,640,32,32,3,3),
+        std::make_tuple(640,640,32,32,3,3),
+        std::make_tuple(960,640,32,32,3,3),
+        std::make_tuple(640,640,64,64,3,3),
+        std::make_tuple(960,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(960,320,64,64,3,3),
+        std::make_tuple(640,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(640,320,64,64,3,3),
+        std::make_tuple(640,320,64,64,3,3),
+        std::make_tuple(320,320,64,64,3,3),
+        std::make_tuple(640,320,64,64,3,3),
+        std::make_tuple(320,4,64,64,3,3),
+
+        //768x768
+        // std::make_tuple(4,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(320,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(640,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(640,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(2560,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(2560,1280,24,24,3,3),
+        // std::make_tuple(2560,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(2560,1280,24,24,3,3),
+        // std::make_tuple(1920,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1920,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,48,48,3,3),
+        // std::make_tuple(1920,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(1920,640,48,48,3,3),
+        // std::make_tuple(1280,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(1280,640,48,48,3,3),
+        // std::make_tuple(960,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(960,640,48,48,3,3),
+        // std::make_tuple(640,640,96,96,3,3),
+        // std::make_tuple(960,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(960,320,96,96,3,3),
+        // std::make_tuple(640,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(640,320,96,96,3,3),
+        // std::make_tuple(640,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(640,320,96,96,3,3),
+        // std::make_tuple(320,4,96,96,3,3),
+        // std::make_tuple(4,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(320,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(320,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(640,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(640,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(2560,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(2560,1280,24,24,3,3),
+        // std::make_tuple(2560,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(2560,1280,24,24,3,3),
+        // std::make_tuple(1920,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,24,24,3,3),
+        // std::make_tuple(1920,1280,24,24,3,3),
+        // std::make_tuple(1280,1280,48,48,3,3),
+        // std::make_tuple(1920,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(1920,640,48,48,3,3),
+        // std::make_tuple(1280,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(1280,640,48,48,3,3),
+        // std::make_tuple(960,640,48,48,3,3),
+        // std::make_tuple(640,640,48,48,3,3),
+        // std::make_tuple(960,640,48,48,3,3),
+        // std::make_tuple(640,640,96,96,3,3),
+        // std::make_tuple(960,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(960,320,96,96,3,3),
+        // std::make_tuple(640,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(640,320,96,96,3,3),
+        // std::make_tuple(640,320,96,96,3,3),
+        // std::make_tuple(320,320,96,96,3,3),
+        // std::make_tuple(640,320,96,96,3,3),
+        // std::make_tuple(320,4,96,96,3,3),
+
+
+        //1024x1024
+        // std::make_tuple(4,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(320,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(640,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(640,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(2560,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(2560,1280,32,32,3,3),
+        // std::make_tuple(2560,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(2560,1280,32,32,3,3),
+        // std::make_tuple(1920,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1920,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,64,64,3,3),
+        // std::make_tuple(1920,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(1920,640,64,64,3,3),
+        // std::make_tuple(1280,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(1280,640,64,64,3,3),
+        // std::make_tuple(960,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(960,640,64,64,3,3),
+        // std::make_tuple(640,640,128,128,3,3),
+        // std::make_tuple(960,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(960,320,128,128,3,3),
+        // std::make_tuple(640,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(640,320,128,128,3,3),
+        // std::make_tuple(640,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(640,320,128,128,3,3),
+        // std::make_tuple(320,4,128,128,3,3),
+        // std::make_tuple(4,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(320,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(320,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(640,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(640,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(2560,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(2560,1280,32,32,3,3),
+        // std::make_tuple(2560,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(2560,1280,32,32,3,3),
+        // std::make_tuple(1920,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,32,32,3,3),
+        // std::make_tuple(1920,1280,32,32,3,3),
+        // std::make_tuple(1280,1280,64,64,3,3),
+        // std::make_tuple(1920,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(1920,640,64,64,3,3),
+        // std::make_tuple(1280,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(1280,640,64,64,3,3),
+        // std::make_tuple(960,640,64,64,3,3),
+        // std::make_tuple(640,640,64,64,3,3),
+        // std::make_tuple(960,640,64,64,3,3),
+        // std::make_tuple(640,640,128,128,3,3),
+        // std::make_tuple(960,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(960,320,128,128,3,3),
+        // std::make_tuple(640,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(640,320,128,128,3,3),
+        // std::make_tuple(640,320,128,128,3,3),
+        // std::make_tuple(320,320,128,128,3,3),
+        // std::make_tuple(640,320,128,128,3,3),
+        // std::make_tuple(320,4,128,128,3,3),
+
+
     };
 
     int k = 0;
@@ -377,6 +694,10 @@ int main(void)
             fprintf(stderr, "| --- | --- | --- | --- | --- \n");
         }
 
+        time_iter0 += run_time0;
+        time_iter1 += run_time1;
+
+
         fprintf(stderr, " | (%d, %d, %d, %d, %d, %d) | %.2f ms | %.2f MB | %.2f ms | %.2f MB\n",
                 std::get<0>(c), std::get<1>(c), std::get<2>(c), std::get<3>(c), std::get<4>(c), std::get<5>(c),
                 run_time0, mem_size0/1024.0f/1024.0f,
@@ -401,6 +722,7 @@ int main(void)
         ggml_gallocr_free(allocr);
 
     }
+    printf("| 1 unet iter takes| %.2f ms |   | %.2f ms | \n", time_iter0, time_iter1);
 
     // printf("\nPerforming test:\n");
     return 0;