diff --git a/tile_engine/CMakeLists.txt b/tile_engine/CMakeLists.txt
index cd1a192a74..6e92b2125a 100644
--- a/tile_engine/CMakeLists.txt
+++ b/tile_engine/CMakeLists.txt
@@ -3,3 +3,5 @@ include_directories(BEFORE
     )
 
 add_subdirectory(ops)
+
+add_subdirectory(01_gemm)
diff --git a/tile_engine/ops/gemm/codegen_utils.py b/tile_engine/ops/gemm/codegen_utils.py
index 98595933b8..0487d3d798 100644
--- a/tile_engine/ops/gemm/codegen_utils.py
+++ b/tile_engine/ops/gemm/codegen_utils.py
@@ -71,7 +71,9 @@ CSHUFFLE_EPILOGUE = """
 """
 
 PIPELINE_MAP = {
-    "mem": ["ck_tile::BaseGemmPipelineAgBgCrMem", "ck_tile::GemmPipelineAgBgCrMem"],
+    "mem": ["ck_tile::BaseGemmPipelineAgBgCrMem", 
+            "ck_tile::GemmPipelineAgBgCrMem"
+    ],
     "compv3": [
         "ck_tile::BaseGemmPipelineAgBgCrCompV3",
         "ck_tile::GemmPipelineAgBgCrCompV3",
diff --git a/tile_engine/ops/gemm/single_instance/generate_kernel_instance.py b/tile_engine/ops/gemm/single_instance/generate_kernel_instance.py
index 073acc3f7b..12d1319e54 100644
--- a/tile_engine/ops/gemm/single_instance/generate_kernel_instance.py
+++ b/tile_engine/ops/gemm/single_instance/generate_kernel_instance.py
@@ -1,8 +1,28 @@
-
 class GemmKernelBuilder:
 
     def _generate_kernel_instance(self, is_header=True):
 
+        # Map pipeline names to base pipeline for hot loop detection
+        base_pipeline_map = {
+            "mem": "ck_tile::BaseGemmPipelineAgBgCrMem",
+            "compv3": "ck_tile::BaseGemmPipelineAgBgCrCompV3",
+            "compv4": "ck_tile::BaseGemmPipelineAgBgCrCompV4",
+        }      
+
+        # Map pipeline names to the correct pipeline implementation
+        pipeline_impl_map = {
+            "mem": "ck_tile::GemmPipelineAgBgCrMem",
+            "compv3": "ck_tile::GemmPipelineAgBgCrCompV3",
+            "compv4": "ck_tile::GemmPipelineAgBgCrCompV4",
+        }          
+
+        # Map scheduler names to the correct enum values
+        scheduler_type_map = {
+            "intrawave": "ck_tile::GemmPipelineScheduler::Intrawave",
+            "interwave": "ck_tile::GemmPipelineScheduler::Interwave",
+            "default": "ck_tile::GemmPipelineScheduler::Default",
+        }
+
         a_type = "ck_tile::fp16_t"
         b_type = "ck_tile::fp16_t"
         c_type = "ck_tile::fp16_t"
@@ -14,19 +34,36 @@ class GemmKernelBuilder:
         b_layout = "ck_tile::tensor_layout::gemm::ColumnMajor"
         c_layout = "ck_tile::tensor_layout::gemm::RowMajor"
 
-        # BlkTile: 128x256x64
+        # BlkTile: 128x256x64, WaveMap: 4x4, WaveTile: 16x16,         
         tile_config = {
             "tile_m": 128,
             "tile_n": 256,
             "tile_k": 64,
-            "warp_m": warp_m,
-            "warp_n": warp_n,
-            "warp_k": warp_k,
-            "warp_tile_m": warp_tile_m,
-            "warp_tile_n": warp_tile_n,
-            "warp_tile_k": warp_tile_k,
+            "warp_m": 4,
+            "warp_n": 4,
+            "warp_k": 1,
+            "warp_tile_m": 16,
+            "warp_tile_n": 16,
+            "warp_tile_k": 16,
         }
 
+        pad_m = "false"
+        pad_n = "false"
+        pad_k = "false" 
+
+        persistent = "false"
+
+        # BlkGemmPipelineVersion: v1, -> ?????
+        pipeline="mem"
+
+        # BlkGemmPipelineScheduler: Intrawave
+        scheduler = "intrawave"
+
+        # DeviceGemm_Wmma_CShuffleV3<Default, RCR> 
+        epilogue = "cshuffle"
+
+        # VmemReadVec: 8x8  ->  ?????
+
         kernel_name = f"gemm"
 
         # Generate kernel instance code using the correct API
diff --git a/tile_engine/ops/gemm/validation_utils.py b/tile_engine/ops/gemm/validation_utils.py
index c0e109bf11..65f69ab7de 100644
--- a/tile_engine/ops/gemm/validation_utils.py
+++ b/tile_engine/ops/gemm/validation_utils.py
@@ -128,10 +128,11 @@ WARP_SUPPORTED_COMBINATIONS = {
         [4, 1, 1],
     ],
     "gfx1201": [
-        [2, 4, 1], 
-        [1, 8, 1], 
-        [8, 1, 1], 
-        [4, 2, 1],
+        [4, 4, 1],
+        # [2, 4, 1], 
+        # [1, 8, 1], 
+        # [8, 1, 1], 
+        # [4, 2, 1],
     ],    
 }