[ET-VK] Replacing use of adaptive_work_group_size function by create_local_wg_size function.

This diff replaces the use of the adaptive_work_group_size function with create_local_wg_size function, which is better tuned for improving shader performance.

Differential Revision: [D66308779](https://our.internmc.facebook.com/intern/diff/D66308779/)

[ghstack-poisoned]

Author: trivedivivek

backends/vulkan/runtime/graph/ops/impl/Copy.cpp

@@ -135,7 +135,7 @@ void add_copy_channel_offset_node(
         utils::safe_downcast<uint32_t>(dim_at<kWidth4D>(in_sizes)),
         utils::safe_downcast<uint32_t>(dim_at<kHeight4D>(in_sizes)),
         utils::safe_downcast<uint32_t>(dst_last_z - dst_first_z + 1)};
-    uvec3 local_size = adaptive_work_group_size(global_size);
+    uvec3 local_size = graph.create_local_wg_size(global_size);
 
     const struct Block final {
       ivec3 range;

backends/vulkan/runtime/graph/ops/impl/Linear.cpp

@@ -198,7 +198,7 @@ void add_addmm_optimized_node(
   } else {
     global_size = utils::divup_vec(global_size, {4, 4, 1});
   }
-  utils::uvec3 local_size = adaptive_work_group_size(global_size);
+  utils::uvec3 local_size = graph.create_local_wg_size(global_size);
 
   graph.execute_nodes().emplace_back(new DispatchNode(
       graph,

backends/vulkan/runtime/graph/ops/impl/MatMul.cpp

@@ -213,7 +213,7 @@ void add_matmul_optimized_node(
     global_size = utils::divup_vec(global_size, {4, 4, 1});
   }
 
-  utils::uvec3 local_size = adaptive_work_group_size(global_size);
+  utils::uvec3 local_size = graph.create_local_wg_size(global_size);
 
   graph.execute_nodes().emplace_back(new DispatchNode(
       graph,

backends/vulkan/runtime/graph/ops/impl/NativeLayerNorm.cpp

@@ -89,7 +89,7 @@ void add_native_layer_norm_node(
   std::vector<int64_t> in_sizes = t_input->sizes();
 
   utils::uvec3 global_size = t_mean->logical_limits();
-  utils::uvec3 local_size = adaptive_work_group_size(global_size);
+  utils::uvec3 local_size = graph.create_local_wg_size(global_size);
 
   std::string kernel_name("native_layer_norm");
   kernel_name.reserve(kShaderNameReserve);

backends/vulkan/runtime/graph/ops/impl/Pool.cpp

@@ -79,7 +79,7 @@ void add_max_pool2d_node(
   check_pool2d_args(*t_in, *t_out);
 
   utils::uvec3 global_size = t_out->logical_limits();
-  utils::uvec3 local_size = adaptive_work_group_size(global_size);
+  utils::uvec3 local_size = graph.create_local_wg_size(global_size);
 
   std::string kernel_name("max_pool2d");
   add_dtype_suffix(kernel_name, *t_out);
@@ -154,7 +154,7 @@ void add_avg_pool2d_node(
   check_pool2d_args(*t_in, *t_out);
 
   utils::uvec3 global_size = t_out->logical_limits();
-  utils::uvec3 local_size = adaptive_work_group_size(global_size);
+  utils::uvec3 local_size = graph.create_local_wg_size(global_size);
 
   std::string kernel_name("avg_pool2d");
   add_dtype_suffix(kernel_name, *t_out);

backends/vulkan/runtime/graph/ops/impl/Repeat.cpp

@@ -90,7 +90,7 @@ void add_repeat_channel_node(
   // Channel packed global work ids
   running_range[2] = out_whcn_sizes[3] * utils::div_up_4(out_whcn_sizes[2]);
   utils::uvec3 global_size = utils::make_uvec3(running_range);
-  utils::uvec3 local_size = adaptive_work_group_size(global_size);
+  utils::uvec3 local_size = graph.create_local_wg_size(global_size);
 
   const struct Block final {
     utils::ivec4 out_sizes;

backends/vulkan/runtime/graph/ops/impl/Slice.cpp

@@ -126,7 +126,7 @@ void add_slice_tensor_copy_node(
     add_dtype_suffix(kernel_name, *t_out);
 
     utils::uvec3 global_size = t_out->logical_limits();
-    utils::uvec3 local_size = adaptive_work_group_size(global_size);
+    utils::uvec3 local_size = graph.create_local_wg_size(global_size);
 
     const struct Block final {
       int dim;