[LITE][OPENCL] add gpu perf mode, priority level for qcom… (#3507)

* [LITE][OPENCL] add gpu perf mode, priority level for qcom adreno. test=develop

README.md

@@ -61,7 +61,8 @@ For demands of Apple's GPU Metal and web front end inference, please see `./meta
 Paddle Lite has referenced the following open-source projects:
 
 - [ARM compute library](http://agroup.baidu.com/paddle-infer/md/article/%28https://github.com/ARM-software/ComputeLibrary%29)
-- [Anakin](https://github.com/PaddlePaddle/Anakin). The optimizations under Anakin has been incorporated into Paddle Lite, and so there will not be any future updates of Anakin. As another high-performance inference project under PaddlePaddle, Anakin has been forward-looking and helpful to the making of Paddle Lite. 
+- [Anakin](https://github.com/PaddlePaddle/Anakin). The optimizations under Anakin has been incorporated into Paddle Lite, and so there will not be any future updates of Anakin. As another high-performance inference project under PaddlePaddle, Anakin has been forward-looking and helpful to the making of Paddle Lite.  
+
 
 
 ## Feedback and Community Support

lite/backends/opencl/cl_runtime.cc

@@ -129,6 +129,26 @@ bool CLRuntime::InitializePlatform() {
   return true;
 }
 
+GpuType CLRuntime::ParseGpuTypeFromDeviceName(std::string device_name) {
+  const std::string kMALI_PATTERN_STR = "Mali";
+  const std::string kADRENO_PATTERN_STR = "QUALCOMM Adreno(TM)";
+  const std::string kPOWERVR_PATTERN_STR = "PowerVR";
+
+  if (device_name == kADRENO_PATTERN_STR) {
+    LOG(INFO) << "adreno gpu";
+    return GpuType::QUALCOMM_ADRENO;
+  } else if (device_name.find(kMALI_PATTERN_STR) != std::string::npos) {
+    LOG(INFO) << "mali gpu";
+    return GpuType::ARM_MALI;
+  } else if (device_name.find(kPOWERVR_PATTERN_STR) != std::string::npos) {
+    LOG(INFO) << "powerVR gpu";
+    return GpuType::IMAGINATION_POWERVR;
+  } else {
+    LOG(INFO) << "others gpu";
+    return GpuType::UNKNOWN;
+  }
+}
+
 bool CLRuntime::InitializeDevice() {
   // ===================== BASIC =====================
   // CL_DEVICE_TYPE_GPU
@@ -148,6 +168,7 @@ bool CLRuntime::InitializeDevice() {
 
   auto device_name = device_->getInfo<CL_DEVICE_NAME>();
   LOG(INFO) << "Using device: " << device_name;
+  gpu_type_ = ParseGpuTypeFromDeviceName(device_name);
 
   cl_device_type device_type = device_->getInfo<CL_DEVICE_TYPE>();
   auto device_type_to_str = [](cl_device_type t) -> std::string {
@@ -296,5 +317,53 @@ std::map<std::string, size_t>& CLRuntime::GetDeviceInfo() {
   return device_info_;
 }
 
+void CLRuntime::GetAdrenoContextProperties(
+    std::vector<cl_context_properties>* properties,
+    GPUPerfMode gpu_perf_mode,
+    GPUPriorityLevel gpu_priority_level) {
+  CHECK(properties) << "cl_context_properties is nullptr";
+  properties->reserve(5);
+  switch (gpu_perf_mode) {
+    case GPUPerfMode::PERF_LOW:
+      LOG(INFO) << "GPUPerfMode::PERF_LOW";
+      properties->push_back(CL_CONTEXT_PERF_MODE_QCOM);
+      properties->push_back(CL_PERF_MODE_LOW_QCOM);
+      break;
+    case GPUPerfMode::PERF_NORMAL:
+      LOG(INFO) << "GPUPerfMode::PERF_NORMAL";
+      properties->push_back(CL_CONTEXT_PERF_MODE_QCOM);
+      properties->push_back(CL_PERF_MODE_NORMAL_QCOM);
+      break;
+    case GPUPerfMode::PERF_HIGH:
+      LOG(INFO) << "GPUPerfMode::PERF_HIGH";
+      properties->push_back(CL_CONTEXT_PERF_MODE_QCOM);
+      properties->push_back(CL_PERF_MODE_HIGH_QCOM);
+      break;
+    default:
+      break;
+  }
+  switch (gpu_priority_level) {
+    case GPUPriorityLevel::PRIORITY_LOW:
+      LOG(INFO) << "GPUPriorityLevel::PRIORITY_LOW";
+      properties->push_back(CL_CONTEXT_PRIORITY_LEVEL_QCOM);
+      properties->push_back(CL_PRIORITY_HINT_LOW_QCOM);
+      break;
+    case GPUPriorityLevel::PRIORITY_NORMAL:
+      LOG(INFO) << "GPUPriorityLevel::PRIORITY_NORMAL";
+      properties->push_back(CL_CONTEXT_PRIORITY_LEVEL_QCOM);
+      properties->push_back(CL_PRIORITY_HINT_NORMAL_QCOM);
+      break;
+    case GPUPriorityLevel::PRIORITY_HIGH:
+      LOG(INFO) << "GPUPriorityLevel::PRIORITY_HIGH";
+      properties->push_back(CL_CONTEXT_PRIORITY_LEVEL_QCOM);
+      properties->push_back(CL_PRIORITY_HINT_HIGH_QCOM);
+      break;
+    default:
+      break;
+  }
+  // The properties list should be terminated with 0
+  properties->push_back(0);
+}
+
 }  // namespace lite
 }  // namespace paddle

lite/backends/opencl/cl_runtime.h

@@ -19,6 +19,45 @@ limitations under the License. */
 #include "lite/backends/opencl/cl_include.h"
 #include "lite/backends/opencl/cl_utility.h"
 
+typedef enum {
+  UNKNOWN = 0,
+  QUALCOMM_ADRENO = 1,
+  ARM_MALI = 2,
+  IMAGINATION_POWERVR = 3,
+  OTHERS = 4,
+} GpuType;
+
+typedef enum {
+  PERF_DEFAULT = 0,
+  PERF_LOW = 1,
+  PERF_NORMAL = 2,
+  PERF_HIGH = 3
+} GPUPerfMode;
+
+typedef enum {
+  PRIORITY_DEFAULT = 0,
+  PRIORITY_LOW = 1,
+  PRIORITY_NORMAL = 2,
+  PRIORITY_HIGH = 3
+} GPUPriorityLevel;
+
+// Adreno extensions
+// Adreno performance hints
+typedef cl_uint cl_perf_hint;
+#define CL_CONTEXT_PERF_MODE_QCOM 0x40C2
+#define CL_PERF_MODE_HIGH_QCOM 0x40C3
+#define CL_PERF_MODE_NORMAL_QCOM 0x40C4
+#define CL_PERF_MODE_LOW_QCOM 0x40C5
+
+// Adreno priority hints
+typedef cl_uint cl_priority_hint;
+
+#define CL_PRIORITY_HINT_NONE_QCOM 0
+#define CL_CONTEXT_PRIORITY_LEVEL_QCOM 0x40C9
+#define CL_PRIORITY_HINT_HIGH_QCOM 0x40CA
+#define CL_PRIORITY_HINT_NORMAL_QCOM 0x40CB
+#define CL_PRIORITY_HINT_LOW_QCOM 0x40CC
+
 namespace paddle {
 namespace lite {
 
@@ -63,9 +102,28 @@ class CLRuntime {
 
   bool InitializeDevice();
 
+  void GetAdrenoContextProperties(
+      std::vector<cl_context_properties>* properties,
+      GPUPerfMode gpu_perf_mode,
+      GPUPriorityLevel gpu_priority_level);
+
   std::shared_ptr<cl::Context> CreateContext() {
-    auto context = std::make_shared<cl::Context>(
-        std::vector<cl::Device>{device()}, nullptr, nullptr, nullptr, &status_);
+    // note(ysh329): gpu perf mode and priority level of adreno gpu referred
+    // from xiaomi/mace.
+    // However, no performance gain after `PERF_HIGH` and `PRIORITY_HIGH` set.
+    auto perf_mode = GPUPerfMode::PERF_HIGH;
+    auto priority_level = GPUPriorityLevel::PRIORITY_HIGH;
+    std::vector<cl_context_properties> context_properties;
+    if (gpu_type_ == GpuType::QUALCOMM_ADRENO) {
+      GetAdrenoContextProperties(
+          &context_properties, perf_mode, priority_level);
+    }
+    auto context =
+        std::make_shared<cl::Context>(std::vector<cl::Device>{device()},
+                                      context_properties.data(),
+                                      nullptr,
+                                      nullptr,
+                                      &status_);
     CL_CHECK_FATAL(status_);
     return context;
   }
@@ -83,8 +141,12 @@ class CLRuntime {
     return queue;
   }
 
+  GpuType ParseGpuTypeFromDeviceName(std::string device_name);
+
   std::map<std::string, size_t> device_info_;
 
+  GpuType gpu_type_{GpuType::UNKNOWN};
+
   std::string cl_path_;
 
   std::shared_ptr<cl::Platform> platform_{nullptr};