Enable fast mode by default

src/targets/gpu/fuse_mlir.cpp

@@ -164,15 +164,21 @@ auto is_mlir_dot(mlir_mode mode)
             return false;
         if(mode != mlir_mode::fast)
             return true;
-        auto a = ins->inputs().front()->get_shape();
-        auto b = ins->inputs().back()->get_shape();
-        // auto m = a.lens()[a.lens().size() - 2];
-        // auto n = b.lens().back();
-        auto k = a.lens().back();
-        // Skipping GEMMs with a K dimension greater than 2048 is a course-grained strategy
-        // to avoid poor-performing GEMM kernels from MLIR
-        // To-do: Investigate a more precise strategy
-        return k <= 2048;
+        float a = ins->inputs().front()->get_shape();
+        float b = ins->inputs().back()->get_shape();
+        float m = a.lens()[a.lens().size() - 2];
+        float n = b.lens().back();
+        float k = a.lens().back();
+        if (k > 1024)
+            return false;
+        auto ratio = m*n/k;
+        if (ratio < 16384)
+            return false;
+        return true;
+        // // Skipping GEMMs with a K dimension greater than 2048 is a course-grained strategy
+        // // to avoid poor-performing GEMM kernels from MLIR
+        // // To-do: Investigate a more precise strategy
+        // return k <= 2048;
     });
 }
 
@@ -418,7 +424,7 @@ void fuse_mlir::apply(module_pass_manager& mpm) const
 
     match::find_matches(mpm,
                         find_mlir_fused_ops{.conv_mode = get_mode("fused", mlir_mode::fast),
-                                            .dot_mode  = get_mode("fused", mode)});
+                                            .dot_mode  = get_mode("fused", mlir_mode::fast)});
 
     match::find_matches(
         mpm,