optimize workgroup size for convolution depthwise and innerproduct pack4

src/layer/convolutiondepthwise.cpp

@@ -659,7 +659,7 @@ int ConvolutionDepthWise::create_pipeline()
         if (num_output % 4 == 0)
         {
             pipeline_convolutiondepthwise_pack4 = new Pipeline(vkdev);
-            pipeline_convolutiondepthwise_pack4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 8));
+            pipeline_convolutiondepthwise_pack4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 4));
             pipeline_convolutiondepthwise_pack4->create("convolutiondepthwise_pack4", specializations, 4, 10);
         }
 

src/layer/innerproduct.cpp

@@ -411,7 +411,7 @@ int InnerProduct::create_pipeline()
     if (num_input % 4 == 0 && num_output % 4 != 0)
     {
         pipeline_innerproduct_pack4to1 = new Pipeline(vkdev);
-        pipeline_innerproduct_pack4to1->set_optimal_local_size_xyz(num_output / 4, 1, 1);
+        pipeline_innerproduct_pack4to1->set_optimal_local_size_xyz(num_output, 1, 1);
         pipeline_innerproduct_pack4to1->create("innerproduct_pack4to1", specializations, 4, 10);
     }