Merge branch 'power' into 'master'

Support Hexagon power control

See merge request !1218

include/mace/public/mace.h

@@ -90,6 +90,18 @@ enum CPUAffinityPolicy {
   AFFINITY_POWER_SAVE = 4,
 };
 
+// Voltage corners for clock frequencies, please refer to
+// docs/Hap_power_set_dcvs_2.html in Hexagon SDK for more detailed information.
+enum HexagonNNCornerType {
+  HEXAGON_NN_CORNER_RELEASE,
+  HEXAGON_NN_CORNER_TURBO,
+  HEXAGON_NN_CORNER_NOMPLUS,
+  HEXAGON_NN_CORNER_NOMINAL,
+  HEXAGON_NN_CORNER_SVSPLUS,
+  HEXAGON_NN_CORNER_SVS,
+  HEXAGON_NN_CORNER_SVS2,
+};
+
 struct CallStats {
   int64_t start_micros;
   int64_t end_micros;
@@ -281,7 +293,7 @@ class MACE_API MaceEngineConfig {
   ///
   /// Just use one GPUContext for multiple models run on GPU.
   /// \param context created use GPUContextBuilder
-  /// \return MaceStatus::MACE_SUCCESS for success, other for failed.
+  /// \return MaceStatus::MACE_SUCCESS for success, other for failure.
   MaceStatus SetGPUContext(std::shared_ptr<GPUContext> context);
 
   /// \brief Set GPU hints, currently only supports Adreno GPU.
@@ -291,7 +303,7 @@ class MACE_API MaceEngineConfig {
   ///
   /// \param perf_hint  performance hint
   /// \param priority_hint  priority hint
-  /// \return MaceStatus::MACE_SUCCESS for success, other for failed.
+  /// \return MaceStatus::MACE_SUCCESS for success, other for failure.
   MaceStatus SetGPUHints(GPUPerfHint perf_hint,
                          GPUPriorityHint priority_hint);
 
@@ -312,10 +324,27 @@ class MACE_API MaceEngineConfig {
   /// \param status MACE_SUCCESS for successful, or it can't reliabley
   /// detect big-LITTLE cores (see GetBigLittleCoreIDs). In such cases, it's
   /// suggested to use AFFINITY_NONE to use all cores.
-  /// \return MaceStatus::MACE_SUCCESS for success, other for failed.
+  /// \return MaceStatus::MACE_SUCCESS for success, other for failure.
   MaceStatus SetCPUThreadPolicy(int num_threads_hint,
                                 CPUAffinityPolicy policy);
 
+  /// \brief Set Hexagon DSP power parameters
+  ///
+  /// Caution: this function may hurt performance if improper
+  /// parameters provided. For most performance critical applications, set
+  /// HexagonNNCornerType to HEXAGON_NN_CORNER_TURBO, enable dynamic clock
+  /// voltage scaling(DCVS) and set sleep latency to 100us works just fine.
+  /// If a more balanced scheme between performance and power consumption
+  /// is needed, these three parameters may be tweaked to achieve that.
+  /// \param corner DCVS voltage target corner, can be set even when DCVS
+  /// is disabled.
+  /// \param dcvs_enable enable or disable DCVS.
+  /// \param latency sleep latency, in micro seconds.
+  /// \return MaceStatus::MACE_SUCCESS for success, other for failure.
+  MaceStatus SetHexagonPower(HexagonNNCornerType corner,
+                             bool dcvs_enable,
+                             int latency);
+
  private:
   class Impl;
   std::unique_ptr<Impl> impl_;

mace/core/CMakeLists.txt

@@ -37,7 +37,7 @@ endif(MACE_ENABLE_HEXAGON_DSP)
 
 if(MACE_ENABLE_HEXAGON_HTA)
   set(CORE_SRCS ${CORE_SRCS} runtime/hexagon/hexagon_hta_wrapper.cc)
-  set(EXTRA_LINK_LIBS ${EXTRA_LINK_LIBS} hta_controller hta_hexagon_runtime npu)
+  set(EXTRA_LINK_LIBS ${EXTRA_LINK_LIBS} hta_hexagon_runtime)
 endif(MACE_ENABLE_HEXAGON_HTA)
 
 if(MACE_ENABLE_MTK_APU)

mace/core/runtime/hexagon/hexagon_dsp_wrapper.cc

@@ -87,17 +87,53 @@ std::string FloatToString(const FloatType v, const int32_t precision) {
   stream << std::fixed << std::setprecision(precision) << v;
   return stream.str();
 }
+
+hexagon_nn_corner_type TransformCornerType(HexagonNNCornerType corner) {
+  switch (corner) {
+    case HEXAGON_NN_CORNER_RELEASE: return NN_CORNER_RELEASE;
+    case HEXAGON_NN_CORNER_TURBO: return NN_CORNER_TURBO;
+    case HEXAGON_NN_CORNER_NOMPLUS: return NN_CORNER_NOMPLUS;
+    case HEXAGON_NN_CORNER_NOMINAL: return NN_CORNER_NOMINAL;
+    case HEXAGON_NN_CORNER_SVSPLUS: return NN_CORNER_SVSPLUS;
+    case HEXAGON_NN_CORNER_SVS: return NN_CORNER_SVS;
+    case HEXAGON_NN_CORNER_SVS2: return NN_CORNER_SVS2;
+    default:
+      LOG(FATAL) << "Wrong Hexagon NN corner type: " << corner;
+      return NN_CORNER_TURBO;
+  }
+}
+
 }  // namespace
 
+HexagonDSPWrapper::HexagonDSPWrapper() {
+  std::string env_log_execute_time_str;
+  GetEnv("MACE_DSP_LOG_EXECUTE_TIME", &env_log_execute_time_str);
+  if (env_log_execute_time_str.empty()) {
+    log_execute_time_ = false;
+  } else {
+    log_execute_time_ = static_cast<bool>(std::stoi(env_log_execute_time_str));
+  }
+}
+
+
 int HexagonDSPWrapper::GetVersion() {
   int version;
   MACE_CHECK(hexagon_nn_version(&version) == 0, "get version error");
   return version;
 }
 
+bool HexagonDSPWrapper::SetPower(HexagonNNCornerType corner,
+                                 bool dcvs_enable,
+                                 int latency) {
+  int ret = hexagon_nn_set_clocks(TransformCornerType(corner),
+                                  dcvs_enable ? NN_DCVS_ENABLE
+                                              : NN_DCVS_DISABLE,
+                                  static_cast<uint32_t>(std::max(0, latency)));
+  return ret == 0;
+}
+
 bool HexagonDSPWrapper::Config() {
   LOG(INFO) << "Hexagon config";
-  MACE_CHECK(hexagon_nn_set_powersave_level(0) == 0, "hexagon power error");
   MACE_CHECK(hexagon_nn_config() == 0, "hexagon config error");
   return true;
 }
@@ -111,7 +147,7 @@ bool HexagonDSPWrapper::Init() {
 
 bool HexagonDSPWrapper::Finalize() {
   LOG(INFO) << "Hexagon finalize";
-  return hexagon_nn_set_powersave_level(1) == 0;
+  return hexagon_nn_remove_clocks() == 0;
 }
 
 bool HexagonDSPWrapper::SetupGraph(const NetDef &net_def,
@@ -432,6 +468,11 @@ bool HexagonDSPWrapper::ExecuteGraph(const Tensor &input_tensor,
   }
   MACE_CHECK(output_bytes == output_tensor->raw_size(),
              "wrong output bytes inferred.");
+
+  if (log_execute_time_) {
+    LOG(INFO) << "dsp cycles: " << GetLastExecuteCycles();
+  }
+
   return true;
 }
 
@@ -439,8 +480,8 @@ bool HexagonDSPWrapper::ExecuteGraphNew(
     const std::map<std::string, Tensor*> &input_tensors,
     std::map<std::string, Tensor*> *output_tensors) {
   VLOG(2) << "Execute graph new: " << nn_id_;
-  uint32_t num_inputs = static_cast<uint32_t>(input_tensors.size());
-  uint32_t num_outputs = static_cast<uint32_t>(output_tensors->size());
+  auto num_inputs = static_cast<uint32_t>(input_tensors.size());
+  auto num_outputs = static_cast<uint32_t>(output_tensors->size());
   MACE_CHECK(num_inputs_ == static_cast<int>(num_inputs), "Wrong inputs num");
   MACE_CHECK(num_outputs_ == static_cast<int>(num_outputs),
              "Wrong outputs num");
@@ -519,7 +560,18 @@ bool HexagonDSPWrapper::ExecuteGraphNew(
                " wrong output bytes inferred.");
   }
 
+  if (log_execute_time_) {
+    LOG(INFO) << "dsp cycles: " << GetLastExecuteCycles();
+  }
+
   return true;
 }
 
+uint64_t HexagonDSPWrapper::GetLastExecuteCycles() {
+  uint32_t cycle_lo;
+  uint32_t cycle_hi;
+  hexagon_nn_last_execution_cycles(nn_id_, &cycle_lo, &cycle_hi);
+  return (static_cast<uint64_t>(cycle_hi) << 32) | cycle_lo;
+}
+
 }  // namespace mace

mace/core/runtime/hexagon/hexagon_dsp_wrapper.h

@@ -27,7 +27,7 @@ namespace mace {
 
 class HexagonDSPWrapper : public HexagonControlWrapper {
  public:
-  HexagonDSPWrapper() = default;
+  HexagonDSPWrapper();
 
   int GetVersion() override;
   bool Config() override;
@@ -46,6 +46,15 @@ class HexagonDSPWrapper : public HexagonControlWrapper {
   void ResetPerfInfo() override;
   void SetDebugLevel(int level) override;
 
+  static bool SetPower(HexagonNNCornerType corner,
+                       bool dcvs_enable,
+                       int latency);
+
+ private:
+  uint64_t GetLastExecuteCycles();
+
+  bool log_execute_time_;
+
   MACE_DISABLE_COPY_AND_ASSIGN(HexagonDSPWrapper);
 };
 }  // namespace mace

mace/libmace/mace.cc

@@ -34,7 +34,10 @@
 #include "mace/core/runtime/opencl/opencl_runtime.h"
 #endif  // MACE_ENABLE_OPENCL
 
-#if defined(MACE_ENABLE_HEXAGON) || defined(MACE_ENABLE_HTA)
+#if defined(MACE_ENABLE_HEXAGON)
+#include "mace/core/runtime/hexagon/hexagon_device.h"
+#include "mace/core/runtime/hexagon/hexagon_dsp_wrapper.h"
+#elif defined(MACE_ENABLE_HTA)
 #include "mace/core/runtime/hexagon/hexagon_device.h"
 #endif
 
@@ -189,6 +192,10 @@ class MaceEngineConfig::Impl {
   MaceStatus SetCPUThreadPolicy(int num_threads_hint,
                                 CPUAffinityPolicy policy);
 
+  MaceStatus SetHexagonPower(HexagonNNCornerType corner,
+                             bool dcvs_enable,
+                             int latency);
+
   inline DeviceType device_type() const {
     return device_type_;
   }
@@ -228,7 +235,13 @@ MaceEngineConfig::Impl::Impl(const DeviceType device_type)
       cpu_affinity_policy_(CPUAffinityPolicy::AFFINITY_NONE),
       gpu_context_(nullptr),
       gpu_priority_hint_(GPUPriorityHint::PRIORITY_LOW),
-      gpu_perf_hint_(GPUPerfHint::PERF_NORMAL) {}
+      gpu_perf_hint_(GPUPerfHint::PERF_NORMAL) {
+#ifdef MACE_ENABLE_HEXAGON
+  if (!HexagonDSPWrapper::SetPower(HEXAGON_NN_CORNER_TURBO, true, 100)) {
+    LOG(WARNING) << "Hexagon set default clocks failed!";
+  }
+#endif
+}
 
 MaceStatus MaceEngineConfig::Impl::SetGPUContext(
     std::shared_ptr<GPUContext> context) {
@@ -252,6 +265,20 @@ MaceStatus MaceEngineConfig::Impl::SetCPUThreadPolicy(
   return MaceStatus::MACE_SUCCESS;
 }
 
+MaceStatus MaceEngineConfig::Impl::SetHexagonPower(
+    HexagonNNCornerType corner,
+    bool dcvs_enable,
+    int latency) {
+  MACE_UNUSED(corner);
+  MACE_UNUSED(dcvs_enable);
+  MACE_UNUSED(latency);
+  bool ret = false;
+#ifdef MACE_ENABLE_HEXAGON
+  ret = HexagonDSPWrapper::SetPower(corner, dcvs_enable, latency);
+#endif
+  return ret ? MaceStatus::MACE_SUCCESS : MaceStatus::MACE_RUNTIME_ERROR;
+}
+
 MaceEngineConfig::MaceEngineConfig(
     const DeviceType device_type)
     : impl_(new MaceEngineConfig::Impl(device_type)) {}
@@ -275,6 +302,13 @@ MaceStatus MaceEngineConfig::SetCPUThreadPolicy(
   return impl_->SetCPUThreadPolicy(num_threads_hint, policy);
 }
 
+MaceStatus MaceEngineConfig::SetHexagonPower(
+    HexagonNNCornerType corner,
+    bool dcvs_enable,
+    int latency) {
+  return impl_->SetHexagonPower(corner, dcvs_enable, latency);
+}
+
 // Mace Tensor
 class MaceTensor::Impl {
  public:

mace/tools/BUILD.bazel

@@ -19,7 +19,11 @@ cc_binary(
         "-Werror",
         "-Wextra",
         "-Wno-missing-field-initializers",
-    ] + if_opencl_enabled(["-DMACE_ENABLE_OPENCL"]),
+    ] + if_opencl_enabled([
+        "-DMACE_ENABLE_OPENCL",
+    ]) + if_hexagon_enabled([
+        "-DMACE_ENABLE_HEXAGON",
+    ]),
     linkstatic = 1,
     deps = [
         "//external:gflags_nothreads",
@@ -39,7 +43,11 @@ cc_binary(
         "-Werror",
         "-Wextra",
         "-Wno-missing-field-initializers",
-    ] + if_opencl_enabled(["-DMACE_ENABLE_OPENCL"]),
+    ] + if_opencl_enabled([
+        "-DMACE_ENABLE_OPENCL",
+    ]) + if_hexagon_enabled([
+        "-DMACE_ENABLE_HEXAGON",
+    ]),
     linkopts = [
         "-lm",
     ] + if_android([

mace/tools/mace_run.cc

@@ -196,7 +196,9 @@ bool RunModel(const std::string &model_name,
         static_cast<GPUPriorityHint>(FLAGS_gpu_priority_hint));
   }
 #endif  // MACE_ENABLE_OPENCL
-
+#ifdef MACE_ENABLE_HEXAGON
+  config.SetHexagonPower(HEXAGON_NN_CORNER_TURBO, true, 100);
+#endif
   std::unique_ptr<mace::port::ReadOnlyMemoryRegion> model_graph_data =
     make_unique<mace::port::ReadOnlyBufferMemoryRegion>();
   if (FLAGS_model_file != "") {

third_party/nnlib/hexagon_nn.h

@@ -165,6 +165,8 @@ __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_execute)(hexagon_nn_nn_id id,
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_teardown)(hexagon_nn_nn_id id) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_set_powersave_level)(unsigned int level) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_set_powersave_details)(hexagon_nn_corner_type corner, hexagon_nn_dcvs_type dcvs, unsigned int latency) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_set_clocks)(hexagon_nn_corner_type corner, hexagon_nn_dcvs_type dcvs, unsigned int latency) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_remove_clocks)() __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_get_perfinfo)(hexagon_nn_nn_id id, hexagon_nn_perfinfo* info_out, int info_outLen, unsigned int* n_items) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_reset_perfinfo)(hexagon_nn_nn_id id, unsigned int event) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(hexagon_nn_last_execution_cycles)(hexagon_nn_nn_id id, unsigned int* cycles_lo, unsigned int* cycles_hi) __QAIC_HEADER_ATTRIBUTE;