Build ROCr core and image libraries as one shared object

Change-Id: I3a16c1227e7db2e386ab33886965596fa0fb0c87

runtime/hsa-runtime/core/util/flag.h

@@ -113,6 +113,9 @@ class Flag {
 
     var = os::GetEnvVar("HSA_NO_SCRATCH_THREAD_LIMITER");
     no_scratch_thread_limit_ = (var == "1") ? true : false;
+
+    var = os::GetEnvVar("HSA_DISABLE_IMAGE");
+    disable_image_ = (var == "1") ? true : false;
   }
 
   bool check_flat_scratch() const { return check_flat_scratch_; }
@@ -153,6 +156,8 @@ class Flag {
 
   std::string tools_lib_names() const { return tools_lib_names_; }
 
+  bool disable_image() const { return disable_image_; }
+
  private:
   bool check_flat_scratch_;
   bool enable_vm_fault_message_;
@@ -167,6 +172,7 @@ class Flag {
   bool fine_grain_pcie_;
   bool no_scratch_reclaim_;
   bool no_scratch_thread_limit_;
+  bool disable_image_;
 
   std::string enable_sdma_;
 

runtime/hsa-runtime/core/util/utils.h

@@ -326,6 +326,50 @@ static __forceinline std::string& rtrim(std::string& s) {
 
 static __forceinline std::string& trim(std::string& s) { return ltrim(rtrim(s)); }
 
+template <uint32_t lowBit, uint32_t highBit, typename T>
+static __forceinline uint32_t BitSelect(T p) {
+  static_assert(sizeof(T) <= sizeof(uintptr_t), "Type out of range.");
+  static_assert(highBit < sizeof(uintptr_t) * 8, "Bit index out of range.");
+
+  uintptr_t ptr = p;
+  if (highBit != (sizeof(uintptr_t) * 8 - 1))
+    return (uint32_t)((ptr & ((1ull << (highBit + 1)) - 1)) >> lowBit);
+  else
+    return (uint32_t)(ptr >> lowBit);
+}
+
+inline uint32_t PtrLow16Shift8(const void* p) {
+  uintptr_t ptr = reinterpret_cast<uintptr_t>(p);
+  return (uint32_t)((ptr & 0xFFFFULL) >> 8);
+}
+
+inline uint32_t PtrHigh64Shift16(const void* p) {
+  uintptr_t ptr = reinterpret_cast<uintptr_t>(p);
+  return (uint32_t)((ptr & 0xFFFFFFFFFFFF0000ULL) >> 16);
+}
+
+inline uint32_t PtrLow40Shift8(const void* p) {
+  uintptr_t ptr = reinterpret_cast<uintptr_t>(p);
+  return (uint32_t)((ptr & 0xFFFFFFFFFFULL) >> 8);
+}
+
+inline uint32_t PtrHigh64Shift40(const void* p) {
+  uintptr_t ptr = reinterpret_cast<uintptr_t>(p);
+  return (uint32_t)((ptr & 0xFFFFFF0000000000ULL) >> 40);
+}
+
+inline uint32_t PtrLow32(const void* p) {
+  return static_cast<uint32_t>(reinterpret_cast<uintptr_t>(p));
+}
+
+inline uint32_t PtrHigh32(const void* p) {
+  uint32_t ptr = 0;
+#ifdef HSA_LARGE_MODEL
+  ptr = static_cast<uint32_t>(reinterpret_cast<uintptr_t>(p) >> 32);
+#endif
+  return ptr;
+}
+
 #include "atomic_helpers.h"
 
-#endif  // HSA_RUNTIME_CORE_UTIL_UTIIS_H_
+#endif  // HSA_RUNTIME_CORE_UTIL_UTILS_H_