Thread count fix for noit sheba

srcCU/sfx-sheba.cu

@@ -286,8 +286,8 @@ void FluxSheba<T, memIn, memOut, MemType::GPU>::noit_compute_flux(const sfx_sheb
     const sfx_sheba_noit_numericsType_C numerics,
     const sfx_phys_constants phys)
 {
-    const int BlockCount = int(ceil(float(grid_size) / 1024.0));
-    dim3 cuBlock = dim3(1024, 1, 1);
+    const int BlockCount = int(ceil(float(grid_size) / 512.0));
+    dim3 cuBlock = dim3(512, 1, 1);
 	dim3 cuGrid = dim3(BlockCount, 1, 1);
 
     sfx_kernel::noit_compute_flux<T><<<cuGrid, cuBlock>>>(sfx, meteo, model,