Add support for gfx900

- Route AQL doorbell directly to HW doorbell
- Reuse precompiled Gfx8 shaders on Gfx9 (ISA is compatible)
- Add a warning for unimplemented code cache invalidation

Change-Id: I92096584a1404e35779c96ae6bdc3e0e7fd04721

runtime/hsa-runtime/core/runtime/amd_aql_queue.cpp

@@ -370,6 +370,12 @@ uint64_t AqlQueue::AddWriteIndexRelease(uint64_t value) {
 }
 
 void AqlQueue::StoreRelaxed(hsa_signal_value_t value) {
+  if (doorbell_type_ == 2) {
+    // Hardware doorbell supports AQL semantics.
+    atomic::Store(signal_.hardware_doorbell_ptr, uint64_t(value), std::memory_order_release);
+    return;
+  }
+
   // Acquire spinlock protecting the legacy doorbell.
   while (atomic::Cas(&amd_queue_.legacy_doorbell_lock, 1U, 0U,
                      std::memory_order_acquire) != 0) {

runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

@@ -201,6 +201,7 @@ void GpuAgent::AssembleShader(const char* src_sp3, const char* func_name,
       asic_shader = &compiled_shader_it->second.compute_7;
       break;
     case 8:
+    case 9: // ISA-compatible with 8
       asic_shader = &compiled_shader_it->second.compute_8;
       break;
     default:
@@ -1176,6 +1177,12 @@ void GpuAgent::InvalidateCodeCaches() {
       // Microcode is handling code cache invalidation.
       return;
     }
+  } else if (isa_->GetMajorVersion() == 9) {
+    static std::once_flag once;
+    std::call_once(once, []() {
+      fprintf(stderr, "warning: code cache invalidation not implemented\n");
+    });
+    return;
   } else {
     assert(false && "Code cache invalidation not implemented for this agent");
   }