Add XNNPACK backend option for workspace sharing (#11748)

Summary:
Add a backend option for XNNPACK to enable runtime control of workspace sharing. I've added 3 mode options - Disabled, PerModel, and Global. PerModel shares the workspace between all CALL_DELEGATE instances in a model, keyed by memory allocator address (see below). Global uses a single workspace instance.

I've written the code to allow for the active workspace mode to be safely changed at any time. The workspace instance is resolved at delegate instance init time (model load) and is stored in the XNNExecutor instance. This design will also allow us to set per-model sharing options in the future. I've introduced a wrapper class (XNNWorkspace) to help with synchronization.

With regard to the PerModel behavior, I am using the address of the runtime allocator to disambiguate the model. This is not ideal in the long-run, but there is some larger discussion around generating IDs in a coherent manner in multithreaded environments without synchronization in the core runtime. This might require PAL changes (exposing a thread ID, for example), so I intend to come back to this.

It should be possible to transparently update this logic in the future. The program ID can collide or change without affecting correctness, but may increase memory (for collisions) or enforce extra synchronization (if unstable between delegate instances in a method).

I'd like to add a PerMethod mode as a follow-up. This should be keyed to the specific method instance (not name), such that multiple method instances for the same method can be loaded for execution on different threads without forcing synchronization, but still allow sharing between call delegate instances in each method instance. This will require a unique method identifier.


Test Plan:
CI. I've also added a set of dedicated tests for getting/setting the option, running PTEs in each mode, switching modes at runtime, and I've also updated the multithreaded stress test to run in each mode.

Rollback Plan:

Differential Revision: D76789804

Pulled By: GregoryComer

Author: GregoryComer

backends/test/multi_method_delegate_test.cpp

@@ -5,13 +5,22 @@
 #include <thread>
 #include <vector>
 
+#include <executorch/backends/xnnpack/runtime/XNNPACKBackend.h>
+
+#include <executorch/runtime/backend/interface.h>
+#include <executorch/runtime/backend/options.h>
 #include <executorch/runtime/executor/program.h>
 #include <executorch/runtime/platform/runtime.h>
 
 #include <executorch/extension/data_loader/file_data_loader.h>
 #include <executorch/extension/memory_allocator/malloc_memory_allocator.h>
 #include <executorch/extension/runner_util/inputs.h>
 
+using executorch::backends::xnnpack::workspace_sharing_mode_option_key;
+using executorch::backends::xnnpack::WorkspaceSharingMode;
+using executorch::backends::xnnpack::xnnpack_backend_key;
+
+using executorch::runtime::BackendOptions;
 using executorch::runtime::Error;
 using executorch::runtime::EValue;
 using executorch::runtime::HierarchicalAllocator;
@@ -126,34 +135,61 @@ class XNNPACKMultiDelegateTest : public ETPTEMethodRunBaseTest {
     num_threads = 40;
     kMethodName = "forward";
   }
-};
 
-// This test is to validate the assumption that the delegate is thread safe.
-// That includes the following:
-// 1. The delegate can be initilized by multiple threads in parallel.
-// 2. The delegate can be executed by multiple threads in parallel.
-// 3. The delegate can be destroyed by multiple threads in parallel.
-// Regardless of the underlying implementation of the delegate.
-// This is particularly important when we have shared resources across
-// delegate instances through a singleton backend instance.
-TEST_F(XNNPACKMultiDelegateTest, MultipleThreads) {
-  ASSERT_NE(kTestPTE1Path.size(), 0);
-  ASSERT_NE(kTestPTE2Path.size(), 0);
-  ASSERT_NE(num_threads, 0);
-  ASSERT_NE(kMethodName.size(), 0);
-
-  std::vector<std::thread> threads(num_threads);
-  std::atomic<size_t> count{0};
-
-  for (int i = 0; i < num_threads; i++) {
-    threads[i] = std::thread([&, i]() {
-      run(i, i % 7 ? kTestPTE1Path : kTestPTE2Path, kMethodName, count);
-    });
+  // This test is to validate the assumption that the delegate is thread safe.
+  // That includes the following:
+  // 1. The delegate can be initilized by multiple threads in parallel.
+  // 2. The delegate can be executed by multiple threads in parallel.
+  // 3. The delegate can be destroyed by multiple threads in parallel.
+  // Regardless of the underlying implementation of the delegate.
+  // This is particularly important when we have shared resources across
+  // delegate instances through a singleton backend instance.
+  void runStressTest() {
+    ASSERT_NE(kTestPTE1Path.size(), 0);
+    ASSERT_NE(kTestPTE2Path.size(), 0);
+    ASSERT_NE(num_threads, 0);
+    ASSERT_NE(kMethodName.size(), 0);
+
+    std::vector<std::thread> threads(num_threads);
+    std::atomic<size_t> count{0};
+
+    for (int i = 0; i < num_threads; i++) {
+      threads[i] = std::thread([&, i]() {
+        run(i, i % 7 ? kTestPTE1Path : kTestPTE2Path, kMethodName, count);
+      });
+    }
+    for (int i = 0; i < num_threads; i++) {
+      threads[i].join();
+    }
+    ASSERT_EQ(count, num_threads);
   }
-  for (int i = 0; i < num_threads; i++) {
-    threads[i].join();
+
+  void setWorkspaceSharingMode(WorkspaceSharingMode mode) {
+    executorch::runtime::runtime_init();
+
+    BackendOptions<1> backend_options;
+    backend_options.set_option(
+        workspace_sharing_mode_option_key, static_cast<int>(mode));
+
+    auto status = executorch::runtime::set_option(
+        xnnpack_backend_key, backend_options.view());
+    ASSERT_EQ(status, Error::Ok);
   }
-  ASSERT_EQ(count, num_threads);
+};
+
+TEST_F(XNNPACKMultiDelegateTest, MultipleThreadsSharingDisabled) {
+  setWorkspaceSharingMode(WorkspaceSharingMode::Disabled);
+  runStressTest();
+}
+
+TEST_F(XNNPACKMultiDelegateTest, MultipleThreadsPerModelSharing) {
+  setWorkspaceSharingMode(WorkspaceSharingMode::PerModel);
+  runStressTest();
+}
+
+TEST_F(XNNPACKMultiDelegateTest, MultipleThreadsGlobalSharing) {
+  setWorkspaceSharingMode(WorkspaceSharingMode::Global);
+  runStressTest();
 }
 
 // TODO(T208989291): Add more tests here. For example,

backends/xnnpack/runtime/XNNCompiler.cpp

@@ -1891,24 +1891,16 @@ ET_NODISCARD Error XNNCompiler::compileModel(
   xnn_weights_cache_t weights_cache_ptr = nullptr;
 #endif
 
-#ifdef ENABLE_XNNPACK_SHARED_WORKSPACE
-  ET_CHECK_OR_RETURN_ERROR(
-      workspace != nullptr, Internal, "Failed to initialize XNNPACK workspace");
+  // NOLINTBEGIN(facebook-hte-NullableDereference) - weights cache is allowed to
+  // be null
   status = xnn_create_runtime_v4(
       subgraph.get(),
       weights_cache_ptr,
       workspace,
       ::executorch::extension::threadpool::get_pthreadpool(),
       runtime_flags,
       &runtime_ptr);
-#else
-  status = xnn_create_runtime_v3(
-      subgraph.get(),
-      weights_cache_ptr,
-      ::executorch::extension::threadpool::get_pthreadpool(),
-      runtime_flags,
-      &runtime_ptr);
-#endif
+  // NOLINTEND(facebook-hte-NullableDereference)
 
   ET_CHECK_OR_RETURN_ERROR(
       xnn_status_success == status,

backends/xnnpack/runtime/XNNExecutor.h

@@ -9,13 +9,13 @@
 #pragma once
 
 #include <executorch/backends/xnnpack/runtime/XNNStatus.h>
+#include <executorch/backends/xnnpack/runtime/XNNWorkspace.h>
 #include <executorch/backends/xnnpack/runtime/profiling/XNNProfiler.h>
 #include <executorch/runtime/backend/interface.h>
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/exec_aten/util/tensor_util.h>
 
 #include <xnnpack.h>
-#include <map>
 #include <memory>
 #include <vector>
 
@@ -35,9 +35,11 @@ class XNNExecutor {
   std::vector<uint32_t> output_ids_;
   std::vector<xnn_external_value> externals_;
   std::vector<std::string> packed_data_names_;
+  std::shared_ptr<XNNWorkspace> workspace_;
 
  public:
-  XNNExecutor() = default;
+  XNNExecutor(std::shared_ptr<XNNWorkspace> workspace)
+      : workspace_(workspace) {}
 
   inline size_t getNumInputs() {
     return input_ids_.size();
@@ -51,6 +53,10 @@ class XNNExecutor {
     return packed_data_names_;
   }
 
+  inline XNNWorkspace& get_workspace() {
+    return *workspace_;
+  }
+
   /**
    * Initialize the XNNExecutor with a given runtime and input/output ids.
    * The input/output ids are expected to be sorted in order of their