Fixed checking CC for enabling Tensor Cores

build/darknet/yolo_cpp_dll.vcxproj

@@ -93,8 +93,8 @@
       <Optimization>Disabled</Optimization>
       <SDLCheck>true</SDLCheck>
       <AdditionalIncludeDirectories>..\..\include;..\..\3rdparty\include;%(AdditionalIncludeDirectories);$(CudaToolkitIncludeDir);$(cudnn)\include</AdditionalIncludeDirectories>
-      <PreprocessorDefinitions>_MBCS;LIB_EXPORTS;_TIMESPEC_DEFINED;_SCL_SECURE_NO_WARNINGS;_CRT_SECURE_NO_WARNINGS;GPU;WIN32;NDEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
-      <UndefinePreprocessorDefinitions>CUDNN</UndefinePreprocessorDefinitions>
+      <PreprocessorDefinitions>_MBCS;LIB_EXPORTS;_TIMESPEC_DEFINED;_SCL_SECURE_NO_WARNINGS;_CRT_SECURE_NO_WARNINGS;GPU;WIN32;DEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <UndefinePreprocessorDefinitions>CUDNN_HALF;CUDNN</UndefinePreprocessorDefinitions>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <ForcedIncludeFiles>stdlib.h;crtdbg.h;%(ForcedIncludeFiles)</ForcedIncludeFiles>
     </ClCompile>
@@ -106,7 +106,7 @@
       <AssemblyDebug>true</AssemblyDebug>
     </Link>
     <CudaCompile>
-      <CodeGeneration>compute_30,sm_30;compute_52,sm_52</CodeGeneration>
+      <CodeGeneration>compute_30,sm_30;compute_75,sm_75</CodeGeneration>
       <TargetMachinePlatform>64</TargetMachinePlatform>
     </CudaCompile>
   </ItemDefinitionGroup>
@@ -137,7 +137,7 @@
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <SDLCheck>true</SDLCheck>
       <AdditionalIncludeDirectories>C:\opencv_3.0\opencv\build\include;..\..\include;..\..\3rdparty\include;%(AdditionalIncludeDirectories);$(CudaToolkitIncludeDir);$(cudnn)\include</AdditionalIncludeDirectories>
-      <PreprocessorDefinitions>CUDNN;GPU;LIB_EXPORTS;_TIMESPEC_DEFINED;_SCL_SECURE_NO_WARNINGS;_CRT_SECURE_NO_WARNINGS;WIN32;NDEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>CUDNN_HALF;CUDNN;GPU;LIB_EXPORTS;_TIMESPEC_DEFINED;_SCL_SECURE_NO_WARNINGS;_CRT_SECURE_NO_WARNINGS;WIN32;NDEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <CLanguageStandard>c11</CLanguageStandard>
       <CppLanguageStandard>c++1y</CppLanguageStandard>
       <PrecompiledHeaderCompileAs>CompileAsCpp</PrecompiledHeaderCompileAs>
@@ -156,7 +156,7 @@
     </Link>
     <CudaCompile>
       <TargetMachinePlatform>64</TargetMachinePlatform>
-      <CodeGeneration>compute_30,sm_30;compute_52,sm_52</CodeGeneration>
+      <CodeGeneration>compute_30,sm_30;compute_75,sm_75</CodeGeneration>
     </CudaCompile>
   </ItemDefinitionGroup>
   <ItemGroup>

src/cuda.c

@@ -309,7 +309,7 @@ int get_gpu_compute_capability(int i)
     cudaDeviceProp prop;
     cudaError_t status = cudaGetDeviceProperties(&prop, i);
     CHECK_CUDA(status);
-    int cc = prop.major * 10 + prop.minor;
+    int cc = prop.major * 100 + prop.minor * 10;    // __CUDA_ARCH__ format
     return cc;
 }
 

src/parser.c

@@ -671,8 +671,10 @@ void parse_net_options(list *options, network *net)
     net->policy = get_policy(policy_s);
     net->burn_in = option_find_int_quiet(options, "burn_in", 0);
 #ifdef CUDNN_HALF
-    if(get_gpu_compute_capability(net->gpu_index) >= 7000) net->cudnn_half = 1;
+    int compute_capability = get_gpu_compute_capability(net->gpu_index);
+    if(get_gpu_compute_capability(net->gpu_index) >= 700) net->cudnn_half = 1;
     else net->cudnn_half = 0;
+    fprintf(stderr, " compute_capability = %d, cudnn_half = %d \n", compute_capability, net->cudnn_half);
 #endif
     if(net->policy == STEP){
         net->step = option_find_int(options, "step", 1);