From a558f82b5f56cdf55f897a3092bb1da8f9085d2c Mon Sep 17 00:00:00 2001
From: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Date: Sun, 4 Aug 2019 16:28:32 +0200
Subject: [PATCH] [dxvk] Implement asynchronous presentation option in the
 backend

---
 dxvk.conf                 | 13 +++++++++++++
 src/dxvk/dxvk_options.cpp |  1 +
 src/dxvk/dxvk_options.h   |  3 +++
 src/dxvk/dxvk_queue.cpp   | 26 ++++++++++++++++++++------
 src/dxvk/dxvk_queue.h     |  1 +
 5 files changed, 38 insertions(+), 6 deletions(-)

diff --git a/dxvk.conf b/dxvk.conf
index b33036cc..ed5839b6 100644
--- a/dxvk.conf
+++ b/dxvk.conf
@@ -183,6 +183,19 @@
 # dxvk.numCompilerThreads = 0
 
 
+# Toggles asynchronous present.
+#
+# Off-loads presentation to the queue submission thread in
+# order to reduce stalling on the main rendering thread and
+# improve performance.
+#
+# Supported values:
+# - Auto: Enable on certain drivers
+# - True / False: Always enable / disable
+
+# dxvk.asyncPresent = Auto
+
+
 # Toggles raw SSBO usage.
 # 
 # Uses storage buffers to implement raw and structured buffer
diff --git a/src/dxvk/dxvk_options.cpp b/src/dxvk/dxvk_options.cpp
index 61fe1018..394feabb 100644
--- a/src/dxvk/dxvk_options.cpp
+++ b/src/dxvk/dxvk_options.cpp
@@ -6,6 +6,7 @@ namespace dxvk {
     enableStateCache      = config.getOption<bool>    ("dxvk.enableStateCache",       true);
     enableTransferQueue   = config.getOption<bool>    ("dxvk.enableTransferQueue",    true);
     numCompilerThreads    = config.getOption<int32_t> ("dxvk.numCompilerThreads",     0);
+    asyncPresent          = config.getOption<Tristate>("dxvk.asyncPresent",           Tristate::Auto);
     useRawSsbo            = config.getOption<Tristate>("dxvk.useRawSsbo",             Tristate::Auto);
     useEarlyDiscard       = config.getOption<Tristate>("dxvk.useEarlyDiscard",        Tristate::Auto);
     hud                   = config.getOption<std::string>("dxvk.hud", "");
diff --git a/src/dxvk/dxvk_options.h b/src/dxvk/dxvk_options.h
index 5b58eead..447294b5 100644
--- a/src/dxvk/dxvk_options.h
+++ b/src/dxvk/dxvk_options.h
@@ -18,6 +18,9 @@ namespace dxvk {
     /// when using the state cache
     int32_t numCompilerThreads;
 
+    /// Asynchronous presentation
+    Tristate asyncPresent;
+
     /// Shader-related options
     Tristate useRawSsbo;
     Tristate useEarlyDiscard;
diff --git a/src/dxvk/dxvk_queue.cpp b/src/dxvk/dxvk_queue.cpp
index e2b1a5f6..0c88f87b 100644
--- a/src/dxvk/dxvk_queue.cpp
+++ b/src/dxvk/dxvk_queue.cpp
@@ -7,7 +7,12 @@ namespace dxvk {
   : m_device(device),
     m_submitThread([this] () { submitCmdLists(); }),
     m_finishThread([this] () { finishCmdLists(); }) {
-    
+    // Asynchronous presentation seems to increase the
+    // likelyhood of hangs on Nvidia for some reason.
+    m_asyncPresent = !m_device->adapter()->matchesDriver(
+      DxvkGpuVendor::Nvidia, VK_DRIVER_ID_NVIDIA_PROPRIETARY_KHR, 0, 0);
+
+    applyTristate(m_asyncPresent, m_device->config().asyncPresent);
   }
   
   
@@ -43,12 +48,21 @@ namespace dxvk {
   void DxvkSubmissionQueue::present(DxvkPresentInfo presentInfo, DxvkSubmitStatus* status) {
     std::unique_lock<std::mutex> lock(m_mutex);
 
-    DxvkSubmitEntry entry = { };
-    entry.status  = status;
-    entry.present = std::move(presentInfo);
+    if (m_asyncPresent) {
+      DxvkSubmitEntry entry = { };
+      entry.status  = status;
+      entry.present = std::move(presentInfo);
 
-    m_submitQueue.push(std::move(entry));
-    m_appendCond.notify_all();
+      m_submitQueue.push(std::move(entry));
+      m_appendCond.notify_all();
+    } else {
+      m_submitCond.wait(lock, [this] {
+        return m_submitQueue.empty();
+      });
+
+      VkResult result = presentInfo.presenter->presentImage(presentInfo.waitSync);
+      status->result.store(result);
+    }
   }
 
 
diff --git a/src/dxvk/dxvk_queue.h b/src/dxvk/dxvk_queue.h
index 50ea26b5..bef9494e 100644
--- a/src/dxvk/dxvk_queue.h
+++ b/src/dxvk/dxvk_queue.h
@@ -156,6 +156,7 @@ namespace dxvk {
   private:
 
     DxvkDevice*             m_device;
+    bool                    m_asyncPresent;
     
     std::atomic<bool>       m_stopped = { false };
     std::atomic<uint32_t>   m_pending = { 0u };