Merge branch 'out-of-range' into 'master'

Use small input for out of range test for mali gpu.

See merge request !395

mace/core/runtime/opencl/opencl_runtime.cc

@@ -178,7 +178,47 @@ void OpenCLPrintfCallback(const char *buffer,
                           void *user_data) {
   fwrite(buffer, 1, length, stdout);
 }
+
+void GetAdrenoContextProperties(std::vector<cl_context_properties> *properties,
+                                GPUPerfHint gpu_perf_hint,
+                                GPUPriorityHint gpu_priority_hint) {
+  MACE_CHECK_NOTNULL(properties);
+  switch (gpu_perf_hint) {
+    case GPUPerfHint::PERF_LOW:
+      properties->push_back(CL_CONTEXT_PERF_HINT_QCOM);
+      properties->push_back(CL_PERF_HINT_LOW_QCOM);
+      break;
+    case GPUPerfHint::PERF_NORMAL:
+      properties->push_back(CL_CONTEXT_PERF_HINT_QCOM);
+      properties->push_back(CL_PERF_HINT_NORMAL_QCOM);
+      break;
+    case GPUPerfHint::PERF_HIGH:
+      properties->push_back(CL_CONTEXT_PERF_HINT_QCOM);
+      properties->push_back(CL_PERF_HINT_HIGH_QCOM);
+      break;
+    default:
+      break;
+  }
+  switch (gpu_priority_hint) {
+    case GPUPriorityHint::PRIORITY_LOW:
+      properties->push_back(CL_CONTEXT_PRIORITY_HINT_QCOM);
+      properties->push_back(CL_PRIORITY_HINT_LOW_QCOM);
+      break;
+    case GPUPriorityHint::PRIORITY_NORMAL:
+      properties->push_back(CL_CONTEXT_PRIORITY_HINT_QCOM);
+      properties->push_back(CL_PRIORITY_HINT_NORMAL_QCOM);
+      break;
+    case GPUPriorityHint::PRIORITY_HIGH:
+      properties->push_back(CL_CONTEXT_PRIORITY_HINT_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
 
 void OpenCLProfilingTimer::StartTiming() {}
 
@@ -210,7 +250,7 @@ GPUPriorityHint OpenCLRuntime::kGPUPriorityHint =
     GPUPriorityHint::PRIORITY_DEFAULT;
 
 OpenCLRuntime *OpenCLRuntime::Global() {
-  static OpenCLRuntime runtime(kGPUPerfHint, kGPUPriorityHint);
+  static OpenCLRuntime runtime;
   return &runtime;
 }
 
@@ -220,48 +260,8 @@ void OpenCLRuntime::Configure(GPUPerfHint gpu_perf_hint,
   OpenCLRuntime::kGPUPriorityHint = gpu_priority_hint;
 }
 
-void GetAdrenoContextProperties(std::vector<cl_context_properties> *properties,
-                                GPUPerfHint gpu_perf_hint,
-                                GPUPriorityHint gpu_priority_hint) {
-  MACE_CHECK_NOTNULL(properties);
-  switch (gpu_perf_hint) {
-    case GPUPerfHint::PERF_LOW:
-      properties->push_back(CL_CONTEXT_PERF_HINT_QCOM);
-      properties->push_back(CL_PERF_HINT_LOW_QCOM);
-      break;
-    case GPUPerfHint::PERF_NORMAL:
-      properties->push_back(CL_CONTEXT_PERF_HINT_QCOM);
-      properties->push_back(CL_PERF_HINT_NORMAL_QCOM);
-      break;
-    case GPUPerfHint::PERF_HIGH:
-      properties->push_back(CL_CONTEXT_PERF_HINT_QCOM);
-      properties->push_back(CL_PERF_HINT_HIGH_QCOM);
-      break;
-    default:
-      break;
-  }
-  switch (gpu_priority_hint) {
-    case GPUPriorityHint::PRIORITY_LOW:
-      properties->push_back(CL_CONTEXT_PRIORITY_HINT_QCOM);
-      properties->push_back(CL_PRIORITY_HINT_LOW_QCOM);
-      break;
-    case GPUPriorityHint::PRIORITY_NORMAL:
-      properties->push_back(CL_CONTEXT_PRIORITY_HINT_QCOM);
-      properties->push_back(CL_PRIORITY_HINT_NORMAL_QCOM);
-      break;
-    case GPUPriorityHint::PRIORITY_HIGH:
-      properties->push_back(CL_CONTEXT_PRIORITY_HINT_QCOM);
-      properties->push_back(CL_PRIORITY_HINT_HIGH_QCOM);
-      break;
-    default:
-      break;
-  }
-  // The properties list should be terminated with 0
-  properties->push_back(0);
-}
 
-OpenCLRuntime::OpenCLRuntime(GPUPerfHint gpu_perf_hint,
-                             GPUPriorityHint gpu_priority_hint):
+OpenCLRuntime::OpenCLRuntime():
     storage_(nullptr), is_profiling_enabled_(false) {
   LoadOpenCLLibrary();
 
@@ -319,8 +319,9 @@ OpenCLRuntime::OpenCLRuntime(GPUPerfHint gpu_perf_hint,
   if (gpu_type_ == GPUType::QUALCOMM_ADRENO) {
     std::vector<cl_context_properties> context_properties;
     context_properties.reserve(5);
-    GetAdrenoContextProperties(&context_properties, gpu_perf_hint,
-                               gpu_priority_hint);
+    GetAdrenoContextProperties(&context_properties,
+                               OpenCLRuntime::kGPUPerfHint,
+                               OpenCLRuntime::kGPUPriorityHint);
     context_ = std::shared_ptr<cl::Context>(
         new cl::Context({*device_}, context_properties.data(),
                         nullptr, nullptr, &err));

mace/core/runtime/opencl/opencl_runtime.h

@@ -90,7 +90,7 @@ class OpenCLRuntime {
   const bool is_profiling_enabled() const;
 
  private:
-  OpenCLRuntime(GPUPerfHint, GPUPriorityHint);
+  OpenCLRuntime();
   ~OpenCLRuntime();
   OpenCLRuntime(const OpenCLRuntime &) = delete;
   OpenCLRuntime &operator=(const OpenCLRuntime &) = delete;

mace/kernels/opencl/out_of_range_check_test.cc

@@ -128,34 +128,31 @@ class OutOfRangeCheckTest : public ::testing::Test {
 };
 
 TEST(OutOfRangeCheckTest, RandomTest) {
-  static unsigned int seed = time(NULL);
-  for (int round = 0; round < 10; ++round) {
-    index_t batch = 11 + rand_r(&seed) % 10;
-    index_t height = 12 + rand_r(&seed) % 100;
-    index_t width = 13 + rand_r(&seed) % 100;
-    index_t channels = 14 + rand_r(&seed) % 50;
-
-    std::vector<index_t> buffer_shape = {batch, height, width, channels};
-    Workspace ws;
-    Tensor *buffer = ws.CreateTensor("Buffer",
-                                     GetDeviceAllocator(DeviceType::OPENCL),
-                                     DataTypeToEnum<float>::v());
-    buffer->Resize(buffer_shape);
-
-    std::vector<size_t> image_shape;
-    Tensor *image = ws.CreateTensor("Image",
-                                     GetDeviceAllocator(DeviceType::OPENCL),
-                                     DataTypeToEnum<float>::v());
-    CalImage2DShape(buffer->shape(), IN_OUT_CHANNEL, &image_shape);
-    image->ResizeImage(buffer->shape(), image_shape);
-    ASSERT_FALSE(BufferToImageOpImpl(buffer, image, image_shape));
-
-    std::vector<size_t> overflow_image_shape = image_shape;
-    for (int i = 0; i < overflow_image_shape.size(); ++i) {
-      overflow_image_shape[i] += 1;
-    }
-    ASSERT_TRUE(BufferToImageOpImpl(buffer, image, overflow_image_shape));
+  index_t batch = 3;
+  index_t height = 5;
+  index_t width = 7;
+  index_t channels = 11;
+
+  std::vector<index_t> buffer_shape = {batch, height, width, channels};
+  Workspace ws;
+  Tensor *buffer = ws.CreateTensor("Buffer",
+                                   GetDeviceAllocator(DeviceType::OPENCL),
+                                   DataTypeToEnum<float>::v());
+  buffer->Resize(buffer_shape);
+
+  std::vector<size_t> image_shape;
+  Tensor *image = ws.CreateTensor("Image",
+                                  GetDeviceAllocator(DeviceType::OPENCL),
+                                  DataTypeToEnum<float>::v());
+  CalImage2DShape(buffer->shape(), IN_OUT_CHANNEL, &image_shape);
+  image->ResizeImage(buffer->shape(), image_shape);
+  ASSERT_FALSE(BufferToImageOpImpl(buffer, image, image_shape));
+
+  std::vector<size_t> overflow_image_shape = image_shape;
+  for (int i = 0; i < overflow_image_shape.size(); ++i) {
+    overflow_image_shape[i] += 1;
   }
+  ASSERT_TRUE(BufferToImageOpImpl(buffer, image, overflow_image_shape));
 }
 
 }  // namespace kernels