From 4d9c540f4fe57551dbfc3cb1b8bf85ee3c19f551 Mon Sep 17 00:00:00 2001
From: Christophe Paquot <christophe.paquot@amd.com>
Date: Thu, 9 Apr 2020 12:44:21 -0700
Subject: [PATCH] Remove a map lookup whenever we were getting the default
 stream

Change-Id: I64b6d1deea41d81e94a58a83de287e78923656b3
---
 vdi/hip_context.cpp  | 24 ++++--------------------
 vdi/hip_device.cpp   | 14 ++++++++++++++
 vdi/hip_internal.hpp |  5 +++--
 3 files changed, 21 insertions(+), 22 deletions(-)

diff --git a/vdi/hip_context.cpp b/vdi/hip_context.cpp
index 6151f68c6a..440c3f4b47 100644
--- a/vdi/hip_context.cpp
+++ b/vdi/hip_context.cpp
@@ -34,8 +34,6 @@ thread_local hipError_t g_lastError = hipSuccess;
 std::once_flag g_ihipInitialized;
 Device* host_device = nullptr;
 
-std::map<Device*, amd::HostQueue*> g_nullStreams;
-
 void init() {
   if (!amd::Runtime::initialized()) {
     amd::IS_HIP = true;
@@ -93,24 +91,10 @@ amd::HostQueue* getQueue(hipStream_t stream) {
   }
 }
 
-amd::HostQueue* getNullStream(Device& dev) {
-  auto stream = g_nullStreams.find(&dev);
-  if (stream == g_nullStreams.end()) {
-    amd::Device* device = dev.devices()[0];
-    cl_command_queue_properties properties = CL_QUEUE_PROFILING_ENABLE;
-    amd::HostQueue* queue = new amd::HostQueue(*dev.asContext(), *device, properties,
-                                               amd::CommandQueue::RealTimeDisabled,
-                                               amd::CommandQueue::Priority::Normal);
-    g_nullStreams[&dev] = queue;
-    return queue;
-  }
-  return stream->second;
-}
-
 amd::HostQueue* getNullStream(amd::Context& ctx) {
- for (auto& it : g_nullStreams) {
-   if (it.first->asContext() == &ctx) {
-     return it.second;
+ for (auto& it : g_devices) {
+   if (it->asContext() == &ctx) {
+     return it->defaultStream();
    }
  }
  return nullptr;
@@ -118,7 +102,7 @@ amd::HostQueue* getNullStream(amd::Context& ctx) {
 
 amd::HostQueue* getNullStream() {
   Device* device = getCurrentDevice();
-  return device ? getNullStream(*device) : nullptr;
+  return device ? device->defaultStream() : nullptr;
 }
 
 };
diff --git a/vdi/hip_device.cpp b/vdi/hip_device.cpp
index 30b2292271..cc27fef05e 100644
--- a/vdi/hip_device.cpp
+++ b/vdi/hip_device.cpp
@@ -22,6 +22,20 @@
 
 #include "hip_internal.hpp"
 
+namespace hip {
+
+amd::HostQueue* Device::defaultStream() {
+  if (defaultStream_ == nullptr) {
+    const cl_command_queue_properties properties = CL_QUEUE_PROFILING_ENABLE;
+    defaultStream_ = new amd::HostQueue(*asContext(), *devices()[0], properties,
+                                        amd::CommandQueue::RealTimeDisabled,
+                                        amd::CommandQueue::Priority::Normal);
+  }
+  return defaultStream_;
+}
+
+};
+
 hipError_t hipDeviceGet(hipDevice_t *device, int deviceId) {
   HIP_INIT_API(hipDeviceGet, device, deviceId);
 
diff --git a/vdi/hip_internal.hpp b/vdi/hip_internal.hpp
index a2851fd978..ed99d6911e 100755
--- a/vdi/hip_internal.hpp
+++ b/vdi/hip_internal.hpp
@@ -84,6 +84,8 @@ namespace hip {
     amd::Monitor lock_{"Device lock"};
     /// VDI context
     amd::Context* context_;
+    /// VDI host queue for default streams
+    amd::HostQueue* defaultStream_;
     /// Device's ID
     /// Store it here so we don't have to loop through the device list every time
     int deviceId_;
@@ -117,6 +119,7 @@ namespace hip {
         return hipErrorPeerAccessNotEnabled;
       }
     }
+    amd::HostQueue* defaultStream();
   };
 
   extern std::once_flag g_ihipInitialized;
@@ -135,8 +138,6 @@ namespace hip {
   /// Note: This follows the CUDA spec to sync with default streams
   ///       and Blocking streams
   extern amd::HostQueue* getQueue(hipStream_t s);
-  /// Get default stream of the device
-  extern amd::HostQueue* getNullStream(Device&);
   /// Get default stream associated with the VDI context
   extern amd::HostQueue* getNullStream(amd::Context&);
   /// Get default stream of the thread