Adjusted gpu curls parameters

include/libgmshdg/kernels_gpu.h

@@ -21,7 +21,7 @@ struct kernel_gpu_sizes<1>
     static const size_t num_bf              = num_dofs_3D(1); //4
     static const size_t num_fluxes          = num_dofs_2D(1); //3
 
-    static const size_t curl_threads        = 256;
+    static const size_t curl_threads        = 512;
     static const size_t deriv_threads       = 128;
     static const size_t lifting_threads     = 128;
 
@@ -37,7 +37,7 @@ struct kernel_gpu_sizes<2>
     static const size_t num_bf              = num_dofs_3D(2); //10
     static const size_t num_fluxes          = num_dofs_2D(2); //6
 
-    static const size_t curl_threads        = 256;
+    static const size_t curl_threads        = 128;
     static const size_t deriv_threads       = 128;
     static const size_t lifting_threads     = 128;
 
@@ -53,7 +53,7 @@ struct kernel_gpu_sizes<3>
     static const size_t num_bf              = num_dofs_3D(3); //20
     static const size_t num_fluxes          = num_dofs_2D(3); //10
 
-    static const size_t curl_threads        = 256;
+    static const size_t curl_threads        = 128;
     static const size_t deriv_threads       = 128;
     static const size_t lifting_threads     = 128;
 
@@ -69,7 +69,7 @@ struct kernel_gpu_sizes<4>
     static const size_t num_bf              = num_dofs_3D(4); //35
     static const size_t num_fluxes          = num_dofs_2D(4);
 
-    static const size_t curl_threads        = 256;
+    static const size_t curl_threads        = 128;
     static const size_t deriv_threads       = 512;
     static const size_t lifting_threads     = 128;
 
@@ -85,7 +85,7 @@ struct kernel_gpu_sizes<5>
     static const size_t num_bf              = num_dofs_3D(5);
     static const size_t num_fluxes          = num_dofs_2D(5);
 
-    static const size_t curl_threads        = 128;
+    static const size_t curl_threads        = 64;
     static const size_t deriv_threads       = 512;
     static const size_t lifting_threads     = 128;
 
@@ -101,7 +101,7 @@ struct kernel_gpu_sizes<6>
     static const size_t num_bf              = num_dofs_3D(6);
     static const size_t num_fluxes          = num_dofs_2D(6);
 
-    static const size_t curl_threads        = 128;
+    static const size_t curl_threads        = 64;
     static const size_t deriv_threads       = 512;
     static const size_t lifting_threads     = 128;