[ET-VK] Enable dynamic operator registration

Pull Request resolved: #2305

This change follows
1. in the footsteps of #2222 for static initialization and
2. the popular `TorchLibraryImpl` for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using `REGISTER_OPERATORS` + `VK_REGISTER_OP()`, as shown in `Arithmetic.h/cpp`.

Typically in Linux/Android C++ environments, the symbols corresponding to `OperatorRegisterInit` static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to `link_whole = True` for the `vulkan_graph_runtime` library.

We update our Compute API tests to verify we can go through `OperatorRegistry` with proper static initialization.
ghstack-source-id: 217884083
@exported-using-ghexport

Differential Revision: [D54641117](https://our.internmc.facebook.com/intern/diff/D54641117/)

Author: jorgep31415

backends/vulkan/runtime/graph/ops/OperatorRegistry.cpp

@@ -15,28 +15,17 @@ namespace native {
 namespace vulkan {
 
 bool OperatorRegistry::has_op(const std::string& name) {
-  return OperatorRegistry::kTable.count(name) > 0;
+  return table_.count(name) > 0;
 }
 
 OperatorRegistry::OpFunction& OperatorRegistry::get_op_fn(
     const std::string& name) {
-  return OperatorRegistry::kTable.find(name)->second;
+  return table_.find(name)->second;
 }
 
-// @lint-ignore-every CLANGTIDY modernize-avoid-bind
-// clang-format off
-#define OPERATOR_ENTRY(name, function) \
-  { #name, std::bind(&function, std::placeholders::_1, std::placeholders::_2) }
-// clang-format on
-
-const OperatorRegistry::OpTable OperatorRegistry::kTable = {
-    OPERATOR_ENTRY(aten.add.Tensor, add),
-    OPERATOR_ENTRY(aten.sub.Tensor, sub),
-    OPERATOR_ENTRY(aten.mul.Tensor, mul),
-    OPERATOR_ENTRY(aten.div.Tensor, div),
-    OPERATOR_ENTRY(aten.div.Tensor_mode, floor_div),
-    OPERATOR_ENTRY(aten.pow.Tensor_Tensor, pow),
-};
+void OperatorRegistry::register_op(const std::string& name, OpFunction& fn) {
+  table_.insert(std::make_pair(name, fn));
+}
 
 OperatorRegistry& operator_registry() {
   static OperatorRegistry registry;

backends/vulkan/runtime/graph/ops/OperatorRegistry.h

@@ -20,6 +20,16 @@
 #define VK_GET_OP_FN(name) \
   ::at::native::vulkan::operator_registry().get_op_fn(name)
 
+#define VK_REGISTER_OP(name, function)                   \
+  ::at::native::vulkan::operator_registry().register_op( \
+      #name,                                             \
+      std::bind(&function, std::placeholders::_1, std::placeholders::_2))
+
+#define REGISTER_OPERATORS                              \
+  static void register_ops();                           \
+  static const OperatorRegisterInit reg(&register_ops); \
+  static void register_ops()
+
 namespace at {
 namespace native {
 namespace vulkan {
@@ -35,7 +45,7 @@ class OperatorRegistry final {
       const std::function<void(ComputeGraph&, const std::vector<ValueRef>&)>;
   using OpTable = std::unordered_map<std::string, OpFunction>;
 
-  static const OpTable kTable;
+  OpTable table_;
 
  public:
   /*
@@ -47,6 +57,20 @@ class OperatorRegistry final {
    * Given an operator name, return the Vulkan delegate function
    */
   OpFunction& get_op_fn(const std::string& name);
+
+  /*
+   * Register a function to a given operator name
+   */
+  void register_op(const std::string& name, OpFunction& fn);
+};
+
+class OperatorRegisterInit final {
+  using InitFn = void();
+
+ public:
+  explicit OperatorRegisterInit(InitFn* init_fn) {
+    init_fn();
+  }
 };
 
 // The Vulkan operator registry is global. It is retrieved using this function,

backends/vulkan/runtime/graph/ops/Utils.h

@@ -16,9 +16,6 @@ namespace at {
 namespace native {
 namespace vulkan {
 
-#define DECLARE_OP_FN(function) \
-  void function(ComputeGraph& graph, const std::vector<ValueRef>& args);
-
 api::utils::ivec4 get_size_as_ivec4(const vTensor& t);
 
 void bind_tensor_to_descriptor_set(

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

@@ -9,6 +9,7 @@
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/Arithmetic.h>
 
 #include <executorch/backends/vulkan/runtime/graph/ops/OpUtils.h>
+#include <executorch/backends/vulkan/runtime/graph/ops/OperatorRegistry.h>
 
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/Staging.h>
 
@@ -81,6 +82,15 @@ void add_arithmetic_node(
       std::move(params)));
 }
 
+REGISTER_OPERATORS {
+  VK_REGISTER_OP(aten.add.Tensor, add);
+  VK_REGISTER_OP(aten.sub.Tensor, sub);
+  VK_REGISTER_OP(aten.mul.Tensor, mul);
+  VK_REGISTER_OP(aten.div.Tensor, div);
+  VK_REGISTER_OP(aten.div.Tensor_mode, floor_div);
+  VK_REGISTER_OP(aten.pow.Tensor_Tensor, pow);
+}
+
 } // namespace vulkan
 } // namespace native
 } // namespace at

backends/vulkan/runtime/graph/ops/impl/Arithmetic.h

@@ -18,13 +18,6 @@ namespace at {
 namespace native {
 namespace vulkan {
 
-DECLARE_OP_FN(add);
-DECLARE_OP_FN(sub);
-DECLARE_OP_FN(mul);
-DECLARE_OP_FN(div);
-DECLARE_OP_FN(floor_div);
-DECLARE_OP_FN(pow);
-
 void add_arithmetic_node(
     ComputeGraph& graph,
     const ValueRef in1,

backends/vulkan/targets.bzl

@@ -56,6 +56,10 @@ def define_common_targets():
             "//caffe2:torch_vulkan_spv",
         ],
         define_static_target = False,
+        # Static initialization is used to register operators to the global operator registry,
+        # therefore link_whole must be True to make sure unused symbols are not discarded.
+        # @lint-ignore BUCKLINT: Avoid `link_whole=True`
+        link_whole = True,
     )
 
     runtime.cxx_library(
@@ -81,4 +85,6 @@ def define_common_targets():
         # VulkanBackend.cpp needs to compile with executor as whole
         # @lint-ignore BUCKLINT: Avoid `link_whole=True` (https://fburl.com/avoid-link-whole)
         link_whole = True,
+        # Define an soname that can be used for dynamic loading in Java, Python, etc.
+        soname = "libvulkan_graph_runtime.$(ext)",
     )

backends/vulkan/test/vulkan_compute_api_test.cpp

@@ -11,6 +11,8 @@
 #include <ATen/native/vulkan/api/api.h>
 
 #include <executorch/backends/vulkan/runtime/graph/ops/OpUtils.h>
+
+#include <executorch/backends/vulkan/runtime/graph/ops/OperatorRegistry.h>
 #include <executorch/backends/vulkan/runtime/graph/ops/StagingUtils.h>
 
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/Arithmetic.h>
@@ -585,8 +587,8 @@ TEST(VulkanComputeGraphTest, test_simple_graph) {
 
   out.value = graph.add_tensor(size_big, api::kFloat);
 
-  add_arithmetic_node(
-      graph, a.value, b.value, kDummyValueRef, out.value, VK_KERNEL(add));
+  auto addFn = VK_GET_OP_FN("aten.add.Tensor");
+  addFn(graph, {a.value, b.value, kDummyValueRef, out.value});
 
   out.staging = graph.set_output_tensor(out.value);
 
@@ -636,8 +638,11 @@ TEST(VulkanComputeGraphTest, test_simple_prepacked_graph) {
   ValueRef c = graph.add_tensor(size_big, api::kFloat);
   ValueRef e = graph.add_tensor(size_big, api::kFloat);
 
-  add_arithmetic_node(graph, a.value, w1, kDummyValueRef, c, VK_KERNEL(add));
-  add_arithmetic_node(graph, c, w2, kDummyValueRef, e, VK_KERNEL(mul));
+  auto addFn = VK_GET_OP_FN("aten.add.Tensor");
+  addFn(graph, {a.value, w1, kDummyValueRef, c});
+
+  auto mulFn = VK_GET_OP_FN("aten.mul.Tensor");
+  mulFn(graph, {c, w2, e});
 
   IOValueRef out = {};
   out.value = e;
@@ -697,8 +702,8 @@ TEST(VulkanComputeGraphTest, test_simple_shared_objects) {
       api::kFloat,
       /*shared_object_idx = */ 6);
 
-  add_arithmetic_node(
-      graph, a.value, b.value, kDummyValueRef, c, VK_KERNEL(add));
+  auto addFn = VK_GET_OP_FN("aten.add.Tensor");
+  addFn(graph, {a.value, b.value, kDummyValueRef, c});
 
   IOValueRef d = graph.add_input_tensor(
       size_small,
@@ -716,7 +721,8 @@ TEST(VulkanComputeGraphTest, test_simple_shared_objects) {
       api::kFloat,
       /*shared_object_idx = */ 4);
 
-  add_arithmetic_node(graph, c, d.value, kDummyValueRef, e, VK_KERNEL(mul));
+  auto mulFn = VK_GET_OP_FN("aten.mul.Tensor");
+  mulFn(graph, {c, d.value, e});
 
   IOValueRef out = {};
   out.value = e;