[ARM CPU] ACL TBB scheduler (openvinotoolkit#17445)

src/plugins/intel_cpu/src/config.cpp

@@ -256,12 +256,6 @@ void Config::readProperties(const std::map<std::string, std::string> &prop) {
         streamExecutorConfig._streams_changed = true;
     }
 
-#if defined(OPENVINO_ARCH_ARM) || defined(OPENVINO_ARCH_ARM64)
-    // TODO: multi-stream execution has functional issues on ARM target
-    streamExecutorConfig._streams = 1;
-    streamExecutorConfig._streams_changed = true;
-#endif
-
     CPU_DEBUG_CAP_ENABLE(applyDebugCapsProperties());
     updateProperties();
 }

src/plugins/intel_cpu/src/nodes/executors/acl/acl_ie_scheduler.cpp

@@ -0,0 +1,43 @@
+// Copyright (C) 2020-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "acl_ie_scheduler.hpp"
+
+#include "arm_compute/core/CPP/ICPPKernel.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include <ie_parallel.hpp>
+
+namespace ov {
+namespace intel_cpu {
+
+using namespace arm_compute;
+
+ACLScheduler::ACLScheduler() = default;
+
+unsigned int ACLScheduler::num_threads() const { return parallel_get_num_threads(); }
+
+void ACLScheduler::set_num_threads(unsigned int num_threads) {}
+
+void ACLScheduler::schedule(ICPPKernel *kernel, const Hints &hints) {
+    ITensorPack tensors;
+    schedule_common(kernel, hints, kernel->window(), tensors);
+}
+
+void ACLScheduler::schedule_op(ICPPKernel *kernel, const Hints &hints, const Window &window, ITensorPack &tensors) {
+    schedule_common(kernel, hints, window, tensors);
+}
+
+void ACLScheduler::run_workloads(std::vector<arm_compute::IScheduler::Workload> &workloads) {
+    InferenceEngine::parallel_for(workloads.size(), [&] (int wid) {
+        ThreadInfo info;
+        info.cpu_info    = &cpu_info();
+        info.num_threads = parallel_get_num_threads();
+        info.thread_id   = wid;
+        workloads[wid](info);
+    });
+}
+
+} // namespace intel_cpu
+} // namespace ov

src/plugins/intel_cpu/src/nodes/executors/acl/acl_ie_scheduler.hpp

@@ -0,0 +1,28 @@
+// Copyright (C) 2020-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <arm_compute/runtime/Scheduler.h>
+#include <arm_compute/core/CPP/ICPPKernel.h>
+#include <arm_compute/core/ITensorPack.h>
+
+namespace ov {
+namespace intel_cpu {
+
+using namespace arm_compute;
+
+class ACLScheduler final : public IScheduler {
+public:
+    ACLScheduler();
+    ~ACLScheduler() override = default;
+    std::uint32_t num_threads() const override;
+    void set_num_threads(unsigned int num_threads) override;
+    void schedule(ICPPKernel *kernel, const Hints &hints) override;
+    void schedule_op(ICPPKernel *kernel, const Hints &hints, const Window &window, ITensorPack &tensors) override;
+protected:
+    void run_workloads(std::vector<Workload> &workloads) override;
+};
+}  //  namespace intel_cpu
+}  //  namespace ov

src/plugins/intel_cpu/src/plugin.cpp

@@ -38,6 +38,11 @@
 #include <cpu/x64/cpu_isa_traits.hpp>
 #include <itt.h>
 
+#if defined(OV_CPU_WITH_ACL)
+#include "nodes/executors/acl/acl_ie_scheduler.hpp"
+#include "arm_compute/runtime/CPP/CPPScheduler.h"
+#endif
+
 using namespace InferenceEngine;
 
 #define IE_CPU_PLUGIN_THROW(...) IE_THROW(__VA_ARGS__) << "CPU plugin: "
@@ -142,6 +147,10 @@ Engine::Engine() :
     specialSetup(new CPUSpecialSetup) {
     _pluginName = "CPU";
     extensionManager->AddExtension(std::make_shared<Extension>());
+#if defined(OV_CPU_WITH_ACL)
+    acl_scheduler = std::make_unique<ACLScheduler>();
+    arm_compute::Scheduler::set(acl_scheduler);
+#endif
 }
 
 Engine::~Engine() {

src/plugins/intel_cpu/src/plugin.h

@@ -61,6 +61,10 @@ class Engine : public InferenceEngine::IInferencePlugin {
     const std::string deviceFullName;
 
     std::shared_ptr<void> specialSetup;
+
+#if defined(OV_CPU_WITH_ACL)
+    std::shared_ptr<arm_compute::IScheduler> acl_scheduler;
+#endif
 };
 
 }   // namespace intel_cpu

src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp

@@ -196,16 +196,6 @@ std::vector<std::string> disabledTestPatterns() {
     retVector.emplace_back(R"(.*HeteroSyntheticTest.*)");
     retVector.emplace_back(R"(.*IEClassBasicTestP.*)");
 #elif defined(OPENVINO_ARCH_ARM64) || defined(OPENVINO_ARCH_ARM)
-    {
-        // TODO: enable once streams / tput mode is supported
-        retVector.emplace_back(R"(OVClassConfigTestCPU.smoke_CpuExecNetworkCheck(Model|Core)StreamsHasHigherPriorityThanLatencyHint.*)");
-        retVector.emplace_back(R"(smoke_BehaviorTests/CorrectConfigCheck.canSetConfigAndCheckGetConfig.*CPU_THROUGHPUT_STREAMS=8.*)");
-        retVector.emplace_back(R"(smoke_BehaviorTests/CorrectConfigCheck.canSetConfigTwiceAndCheckGetConfig.*CPU_THROUGHPUT_STREAMS=8.*)");
-        retVector.emplace_back(R"(smoke_CPU_OVClassLoadNetworkAndCheckWithSecondaryPropertiesTest/OVClassLoadNetworkAndCheckSecondaryPropertiesTest.LoadNetworkAndCheckSecondaryPropertiesTest.*)");
-        retVector.emplace_back(R"(smoke_CPU_OVClassLoadNetworkAndCheckWithSecondaryPropertiesDoubleTest/OVClassLoadNetworkAndCheckSecondaryPropertiesTest.LoadNetworkAndCheckSecondaryPropertiesTest.*)");
-        retVector.emplace_back(R"(smoke_CPU_OVClassCompileModelAndCheckSecondaryPropertiesTest.*)");
-        retVector.emplace_back(R"(smoke_CPU_OVClassCompileModelAndCheckWithSecondaryPropertiesDoubleTest.*)");
-    }
     // invalid test: checks u8 precision for runtime graph, while it should be f32
     retVector.emplace_back(R"(smoke_NegativeQuantizedMatMulMultiplyFusion.*)");
     // int8 specific