[Bert] xlarge model tile size

Author: YunseonShin

PyTorchSimFrontend/mlir/mlir_template.py

@@ -214,9 +214,9 @@ def gemm_combination_mapping(self, M, N, K, n_extra_node=0, n_prologue_node=0, p
                     input_size_per_lane = self.get_spad_size_per_lane(tile_M * (1 + n_prologue_node), tile_K)
                     output_size_per_lane = self.get_spad_size_per_lane(tile_M * (1 + n_extra_node), tile_N)
                     used_spad_size_per_lane = (weight_size_per_lane + input_size_per_lane + output_size_per_lane) * self.precision
-                    n_tile = math.ceil(N / tile_N)
+                    n_tile = math.ceil(M / tile_M) * math.ceil(N / tile_N)
                     check_spad_size = (used_spad_size < max_spad_size and used_spad_size_per_lane < max_spad_per_lane)
-                    if check_spad_size and max_used_spad_size < used_spad_size and maximize_i_j <= tile_M * tile_N and n_tile >= minimum_n_tile and tile_N // tile_M < 10 and tile_N < 2048:
+                    if check_spad_size and max_used_spad_size < used_spad_size and maximize_i_j <= tile_M * tile_N and n_tile >= minimum_n_tile and tile_N // tile_M < 10:
                         max_used_spad_size = used_spad_size
                         maximize_i_j = tile_M * tile_N
                         mapping = (tile_M, tile_N, tile_K)

validation/gemm_candidates/gemm_512_64_512.txt

@@ -0,0 +1,17 @@
+{
+    "512_2048_8192" : {
+        "TILE_M" : 512,
+        "TILE_K" : 512,
+        "TILE_N" : 1024
+    },
+    "512_2048_2048" : {
+        "TILE_M" : 512,
+        "TILE_K" : 512,
+        "TILE_N" : 1024
+    },
+    "2048_2048_512" : {
+        "TILE_M" : 1024,
+        "TILE_K" : 512,
+        "TILE_N" : 512
+    }
+}