added DeviceCacheSettings

examples_ptx/deferred/kernels.cu

@@ -133,7 +133,7 @@ struct Uniform
     data[chunkIdx] = shdata[programIndex];
   }
 };
-#elif 1
+#elif 0
 template<typename T, int N>
 struct Uniform
 {
@@ -174,7 +174,7 @@ __shared__ int shdata_full[4*MAX_LIGHTS];
 template<typename T, int N>
 struct Uniform
 {
-  volatile T *shdata;
+  /* volatile */ T *shdata;
 
   __device__ Uniform()
   {

examples_ptx/ptxcc/PTXParser.h

@@ -141,6 +141,16 @@ namespace parser
         s << printArgumentList();
         s << "\n )\n";
         s << "{\n   ";
+//        s << " cudaFuncSetCacheConfig (" << calleeName << ", ";
+        s << " cudaDeviceSetCacheConfig (";
+#if 1
+        s << " cudaFuncCachePreferEqual ";
+#elif 1
+        s << " cudaFuncCachePreferL1 ";
+#else
+        s << " cudaFuncCachePreferShared ";
+#endif
+        s << ");\n";
         s << calleeName;
         s << "<<<1,32>>>(\n";
         s << printArgumentList(false);