Add hsa_amd_queue_set_priority extension function

Controls dispatch and wavefront scheduling arbitration across quees.

Change-Id: I498f4898b544f79b8fb8514bf7e789ca9da29462

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

@@ -93,7 +93,9 @@ AqlQueue::AqlQueue(GpuAgent* agent, size_t req_size_pkts, HSAuint32 node_id, Scr
       is_kv_queue_(is_kv),
       pm4_ib_buf_(nullptr),
       pm4_ib_size_b_(0x1000),
-      dynamicScratchState(0) {
+      dynamicScratchState(0),
+      suspended_(false),
+      priority_(HSA_QUEUE_PRIORITY_NORMAL) {
   // When queue_full_workaround_ is set to 1, the ring buffer is internally
   // doubled in size. Virtual addresses in the upper half of the ring allocation
   // are mapped to the same set of pages backing the lower half.
@@ -150,8 +152,7 @@ AqlQueue::AqlQueue(GpuAgent* agent, size_t req_size_pkts, HSAuint32 node_id, Scr
   }
 
   HSAKMT_STATUS kmt_status;
-  kmt_status = hsaKmtCreateQueue(node_id, HSA_QUEUE_COMPUTE_AQL, 100,
-                                 HSA_QUEUE_PRIORITY_NORMAL, ring_buf_,
+  kmt_status = hsaKmtCreateQueue(node_id, HSA_QUEUE_COMPUTE_AQL, 100, priority_, ring_buf_,
                                  ring_buf_alloc_bytes_, NULL, &queue_rsrc);
   if (kmt_status != HSAKMT_STATUS_SUCCESS)
     throw AMD::hsa_exception(HSA_STATUS_ERROR_OUT_OF_RESOURCES,
@@ -686,7 +687,8 @@ int AqlQueue::CreateRingBufferFD(const char* ring_buf_shm_path,
 }
 
 void AqlQueue::Suspend() {
-  auto err = hsaKmtUpdateQueue(queue_id_, 0, HSA_QUEUE_PRIORITY_NORMAL, NULL, 0, NULL);
+  suspended_ = true;
+  auto err = hsaKmtUpdateQueue(queue_id_, 0, priority_, NULL, 0, NULL);
   assert(err == HSAKMT_STATUS_SUCCESS && "hsaKmtUpdateQueue failed.");
 }
 
@@ -700,6 +702,16 @@ hsa_status_t AqlQueue::Inactivate() {
   return HSA_STATUS_SUCCESS;
 }
 
+hsa_status_t AqlQueue::SetPriority(HSA_QUEUE_PRIORITY priority) {
+  if (suspended_) {
+    return HSA_STATUS_ERROR_INVALID_QUEUE;
+  }
+
+  priority_ = priority;
+  auto err = hsaKmtUpdateQueue(queue_id_, 100, priority_, ring_buf_, ring_buf_alloc_bytes_, NULL);
+  return (err == HSAKMT_STATUS_SUCCESS ? HSA_STATUS_SUCCESS : HSA_STATUS_ERROR_OUT_OF_RESOURCES);
+}
+
 bool AqlQueue::DynamicScratchHandler(hsa_signal_value_t error_code, void* arg) {
   AqlQueue* queue = (AqlQueue*)arg;
   hsa_status_t errorCode = HSA_STATUS_SUCCESS;

runtime/hsa-runtime/core/runtime/hsa_api_trace.cpp

@@ -380,6 +380,7 @@ void HsaApiTable::UpdateAmdExts() {
   amd_ext_api.hsa_amd_register_system_event_handler_fn = AMD::hsa_amd_register_system_event_handler;
   amd_ext_api.hsa_amd_queue_intercept_create_fn = AMD::hsa_amd_queue_intercept_create;
   amd_ext_api.hsa_amd_queue_intercept_register_fn = AMD::hsa_amd_queue_intercept_register;
+  amd_ext_api.hsa_amd_queue_set_priority_fn = AMD::hsa_amd_queue_set_priority;
 }
 
 class Init {

runtime/hsa-runtime/core/runtime/hsa_ext_amd.cpp

@@ -46,6 +46,7 @@
 #include <set>
 #include <utility>
 #include <memory>
+#include <map>
 
 #include "core/inc/runtime.h"
 #include "core/inc/agent.h"
@@ -852,4 +853,28 @@ hsa_status_t hsa_amd_register_system_event_handler(hsa_amd_system_event_callback
   CATCH;
 }
 
+hsa_status_t HSA_API hsa_amd_queue_set_priority(hsa_queue_t* queue,
+                                                hsa_amd_queue_priority_t priority) {
+  TRY;
+  IS_OPEN();
+  IS_BAD_PTR(queue);
+  core::Queue* cmd_queue = core::Queue::Convert(queue);
+  IS_VALID(cmd_queue);
+
+  static std::map<hsa_amd_queue_priority_t, HSA_QUEUE_PRIORITY> ext_kmt_priomap = {
+      {HSA_AMD_QUEUE_PRIORITY_LOW, HSA_QUEUE_PRIORITY_MINIMUM},
+      {HSA_AMD_QUEUE_PRIORITY_NORMAL, HSA_QUEUE_PRIORITY_NORMAL},
+      {HSA_AMD_QUEUE_PRIORITY_HIGH, HSA_QUEUE_PRIORITY_MAXIMUM},
+  };
+
+  auto priority_it = ext_kmt_priomap.find(priority);
+
+  if (priority_it == ext_kmt_priomap.end()) {
+    return HSA_STATUS_ERROR_INVALID_ARGUMENT;
+  }
+
+  return cmd_queue->SetPriority(priority_it->second);
+  CATCH;
+}
+
 } // end of AMD namespace