Add num_instructions program function

Summary:
Add num_instructions program function to executorch program methods, because FDLA finds this information useful. 

FDLA can use this information to gauge how big the method allocator needs to be. While FDLA doesn't have an exact formula yet, most models are fine with a small allocation of 16Kb, whereas some require much more (LLM CRIA requires .3MB, Surface Typing requires .5MB). 
Checking the number of instructions in the model is a fairly good indicator of this need, as most FDLA models have under 100 instructions, and the latter models have ~1000 instructions. This order of magnitude difference means that, while not an exact science, FDLA can read the number of instructions and surmise that models with above 100 instructions deserve a large size method allocator block, and regular models can take the default size method allocator block.

FDLA's method allocator is not dynamically sized, FDLA grants all models with the same size method allocator right now, but it would be an egregious waste of memory to grant all models with a .5 MB allocator (the new top line with Surface Typing model) when most models only need 16 KB. FDLA's method allocator is not dynamic because the core who manages the memory (FDLAMCU) is different from the core who executes the inference (FDLADSP).

This thread discusses the same https://fb.workplace.com/groups/pytorch.edge.users/permalink/1523803391823282/

Differential Revision: D66504180

Author: Regan Stehle

runtime/executor/method_meta.cpp

@@ -210,5 +210,25 @@ Result<int64_t> MethodMeta::memory_planned_buffer_size(size_t index) const {
   return s_plan_->non_const_buffer_sizes()->Get(index + 1);
 }
 
+size_t MethodMeta::num_instructions() const {
+  const auto chains = s_plan_->chains();
+  if (chains == nullptr) {
+    return 0;
+  }
+  const auto num_chains = chains->size();
+  auto num_instructions = 0;
+  for (size_t i = 0; i < num_chains; ++i) {
+    auto s_chain = chains->Get(i);
+    if (s_chain == nullptr) {
+      continue;
+    }
+    auto s_instructions = s_chain->instructions();
+    if (s_instructions != nullptr) {
+      num_instructions += s_instructions->size();
+    }
+  }
+  return num_instructions;
+}
+
 } // namespace runtime
 } // namespace executorch

runtime/executor/method_meta.h

@@ -184,6 +184,13 @@ class MethodMeta final {
    * @returns The size in bytes on success, or an error on failure.
    */
   Result<int64_t> memory_planned_buffer_size(size_t index) const;
+  
+  /**
+   * Get the number of instructions in this method.
+   *
+   * @returns The number of instructions.
+   */
+  ET_EXPERIMENTAL size_t num_instructions() const;
 
   /**
    * DEPRECATED: Use num_memory_planned_buffers() instead.

runtime/executor/test/method_meta_test.cpp

@@ -92,6 +92,9 @@ TEST_F(MethodMetaTest, MethodMetaApi) {
       method_meta->non_const_buffer_size(1).error(),
       Error::InvalidArgument); // Deprecated API
 
+  // Number instructions in method is nonzero 
+  EXPECT_NE(method_meta->num_instructions(), 0);
+
   // Missing method fails
   EXPECT_EQ(
       program_->method_meta("not_a_method").error(), Error::InvalidArgument);