version 0.40. Enable extra optimizations.

5 files changed, 866 insertions, 296 deletions

diff --git a/.SRCINFO b/.SRCINFO
index ab6d15ff9730..eb04e3cce0ec 100644
--- a/.SRCINFO
+++ b/.SRCINFO
@@ -1,6 +1,6 @@
 pkgbase = d9vk-mingw
 	pkgdesc = A d3d9 to vk layer based off DXVK's codebase, mingw version
-	pkgver = 0.40.rc.p
+	pkgver = 0.40
 	pkgrel = 1
 	url = https://github.com/Joshua-Ashton/d9vk
 	arch = x86_64
@@ -17,12 +17,16 @@ pkgbase = d9vk-mingw
 	depends = bash
 	provides = d9vk
 	conflicts = d9vk
-	source = git+https://github.com/Joshua-Ashton/d9vk.git#tag=0.40-rc-p
+	source = git+https://github.com/Joshua-Ashton/d9vk.git#tag=0.40
 	source = setup_d9vk
 	source = dxvk-async.patch
+	source = dxvk-mangohud.patch
+	source = extraopts.patch
 	sha256sums = SKIP
 	sha256sums = 7147644664ef33d04f7b18683c47be95b5664c57cf6d63fdc019d915deebd37a
-	sha256sums = 757fd5ca3faf476730891541de408593450fe4b490cc87bd90b342826d66c564
+	sha256sums = 6ff286091c20327e67252e1e6812830a42c990d1ee56541023eb217712209f3c
+	sha256sums = 2e335237623aaf006f8814fc9712f3a4be0d678cd0714879a3a4545f3bbf41ce
+	sha256sums = acde3a23166f79fa87ab090200be2aabaf16e5876ce19b8270ad1179bb0bc7a5
 
 pkgname = d9vk-mingw
 
diff --git a/PKGBUILD b/PKGBUILD
index 8885ce573661..41de98daede5 100644
--- a/PKGBUILD
+++ b/PKGBUILD
@@ -1,8 +1,8 @@
 # Maintainer: loathingkernel <loathingkernel @at gmail .dot com>
 
 pkgname=d9vk-mingw
-_tag=0.40-rc-p
-pkgver=${_tag//-/.}
+_tag=0.40
+pkgver=${_tag//-/_}
 pkgrel=1
 pkgdesc="A d3d9 to vk layer based off DXVK's codebase, mingw version"
 arch=('x86_64')
@@ -16,20 +16,56 @@ source=(
     "git+https://github.com/Joshua-Ashton/d9vk.git#tag=$_tag"
     "setup_d9vk"
     "dxvk-async.patch"
+    "dxvk-mangohud.patch"
+    "extraopts.patch"
 )
 sha256sums=(
     "SKIP"
     "7147644664ef33d04f7b18683c47be95b5664c57cf6d63fdc019d915deebd37a"
-    "757fd5ca3faf476730891541de408593450fe4b490cc87bd90b342826d66c564"
+    "6ff286091c20327e67252e1e6812830a42c990d1ee56541023eb217712209f3c"
+    "2e335237623aaf006f8814fc9712f3a4be0d678cd0714879a3a4545f3bbf41ce"
+    "acde3a23166f79fa87ab090200be2aabaf16e5876ce19b8270ad1179bb0bc7a5"
 )
 
 prepare() {
     cd d9vk
+    # If using -march=native and the CPU supports AVX, launching a d3d9
+    # game can cause an Unhandled exception. The cause seems to be the
+    # combination of AVX instructions and tree vectorization (implied by O3),
+    # all tested archictures from sandybridge to haswell are affected.
+    # Disabling either AVX (and AVX2 as a side-effect) or tree
+    # vectorization fixes the issue. I am not sure which one is better
+    # to disable so below you can choose. Append either of these flags.
+    # Relevant Wine issues
+    # https://bugs.winehq.org/show_bug.cgi?id=45289
+    # https://bugs.winehq.org/show_bug.cgi?id=43516
+    CFLAGS+=" -mno-avx"
+    # CFLAGS+=" -fno-tree-vectorize"
+    # Patch crossfiles with extra optimizations from makepkg.conf
+    # If building fails, comment the line below to disable them.
+    patch -p1 -i ../extraopts.patch
+    # Adjust optimization level in meson arguments. This is ignored
+    # anyways because meson sets its own optimization level.
+    CFLAGS="${CFLAGS/ -O*([0-3])/}"
+    # Doesn't compile with these flags in MingW so remove them.
+    # They are also filtered in Wine PKGBUILDs so remove them
+    # for winelib versions too.
+    CFLAGS="${CFLAGS/ -fno-plt/}"
+    LDFLAGS="${LDFLAGS/,-z,relro,-z,now/}"
+    sed -i build-win64.txt \
+        -e "s|@CARGS@|\'${CFLAGS// /\',\'}\'|g" \
+        -e "s|@LDARGS@|\'${LDFLAGS// /\',\'}\'|g"
+    sed -i build-win32.txt \
+        -e "s|@CARGS@|\'${CFLAGS// /\',\'}\'|g" \
+        -e "s|@LDARGS@|\'${LDFLAGS// /\',\'}\'|g"
     # Uncomment to enable dxvk async patch.
     # Enable at your own risk. If you don't know what it is,
     # and its implications, leave it as is. You have been warned.
     # I am not liable if anything happens to you by using it.
+    # Patch enables async by default. YOU HAVE BEEN WARNED.
     #patch -p1 -i ../dxvk-async.patch
+    # Uncomment to enable Mango HUD for dxvk
+    #patch -p1 -i ../dxvk-mangohud.patch
 }
 
 build() {
@@ -38,6 +74,7 @@ build() {
         --prefix "/usr/share/d9vk/x64" \
         --bindir "" --libdir "" \
         --buildtype "release" \
+        --optimization=3 \
         --strip \
         -Denable_d3d10=false \
         -Denable_d3d11=false \
@@ -50,6 +87,7 @@ build() {
         --prefix "/usr/share/d9vk/x32" \
         --bindir "" --libdir "" \
         --buildtype "release" \
+        --optimization=3 \
         --strip \
         -Denable_d3d10=false \
         -Denable_d3d11=false \
diff --git a/dxvk-async.patch b/dxvk-async.patch
index 21bf8325ba7c..19c506c4964f 100644
--- a/dxvk-async.patch
+++ b/dxvk-async.patch
@@ -1,14 +1,15 @@
-From 653b6b9d25d41df41acb9d5d50e06bf8f45cc56c Mon Sep 17 00:00:00 2001
+From dfac76f528d0fc88ecbfe27ccb87483840828b9b Mon Sep 17 00:00:00 2001
 From: Tk-Glitch <ti3nou@gmail.com>
-Date: Mon, 15 Jul 2019 05:46:41 +0200
-Subject: async patchset rebase against a93dd74f
+Date: Sat, 28 Sep 2019 18:59:04 +0200
+Subject: Async pipecompiler rebase against DXVK
+ d128d776ad906d6b8d3941eda7b7ee679346dbaf
 
 
 diff --git a/meson.build b/meson.build
-index f7dcd491..31665f93 100644
+index 9a519e46..dea82ee1 100644
 --- a/meson.build
 +++ b/meson.build
-@@ -90,7 +90,7 @@ else
+@@ -101,7 +101,7 @@ else
  endif
  
  dxvk_version = vcs_tag(
@@ -18,10 +19,10 @@ index f7dcd491..31665f93 100644
    output: 'version.h')
  
 diff --git a/src/dxvk/dxvk_context.cpp b/src/dxvk/dxvk_context.cpp
-index 6df6e3c3..bb802c96 100644
+index b4f679fa..13c86d8f 100644
 --- a/src/dxvk/dxvk_context.cpp
 +++ b/src/dxvk/dxvk_context.cpp
-@@ -648,7 +648,7 @@ namespace dxvk {
+@@ -606,7 +606,7 @@ namespace dxvk {
      const Rc<DxvkImageView>&    imageView,
            VkImageAspectFlags    clearAspects,
            VkClearValue          clearValue) {
@@ -30,7 +31,7 @@ index 6df6e3c3..bb802c96 100644
  
      // Prepare attachment ops
      DxvkColorAttachmentOps colorOp;
-@@ -2406,7 +2406,7 @@ namespace dxvk {
+@@ -2433,7 +2433,7 @@ namespace dxvk {
            VkExtent3D            extent,
            VkImageAspectFlags    aspect,
            VkClearValue          value) {
@@ -39,16 +40,16 @@ index 6df6e3c3..bb802c96 100644
  
      // Find out if the render target view is currently bound,
      // so that we can avoid spilling the render pass if it is.
-@@ -3527,7 +3527,7 @@ namespace dxvk {
-       // Retrieve and bind actual Vulkan pipeline handle
-       m_gpActivePipeline = m_state.gp.pipeline != nullptr && m_state.om.framebuffer != nullptr
-         ? m_state.gp.pipeline->getPipelineHandle(m_state.gp.state,
--            m_state.om.framebuffer->getRenderPass())
-+            m_state.om.framebuffer->getRenderPass(), this->checkAsyncCompilationCompat())
-         : VK_NULL_HANDLE;
-       
-       if (m_gpActivePipeline != VK_NULL_HANDLE) {
-@@ -3794,7 +3794,7 @@ namespace dxvk {
+@@ -3652,7 +3652,7 @@ namespace dxvk {
+       : DxvkContextFlag::GpDirtyStencilRef);
+     
+     // Retrieve and bind actual Vulkan pipeline handle
+-    m_gpActivePipeline = m_state.gp.pipeline->getPipelineHandle(m_state.gp.state, m_state.om.framebuffer->getRenderPass());
++    m_gpActivePipeline = m_state.gp.pipeline->getPipelineHandle(m_state.gp.state, m_state.om.framebuffer->getRenderPass(), this->checkAsyncCompilationCompat());
+ 
+     if (unlikely(!m_gpActivePipeline))
+       return false;
+@@ -3969,7 +3969,7 @@ namespace dxvk {
    }
    
    
@@ -56,8 +57,8 @@ index 6df6e3c3..bb802c96 100644
 +  void DxvkContext::updateFramebuffer(bool isDraw) {
      if (m_flags.test(DxvkContextFlag::GpDirtyFramebuffer)) {
        m_flags.clr(DxvkContextFlag::GpDirtyFramebuffer);
-       
-@@ -3813,6 +3813,11 @@ namespace dxvk {
+ 
+@@ -3988,6 +3988,11 @@ namespace dxvk {
            : VkComponentMapping();
        }
  
@@ -69,7 +70,7 @@ index 6df6e3c3..bb802c96 100644
        m_flags.set(DxvkContextFlag::GpDirtyPipelineState);
      }
    }
-@@ -4040,7 +4045,7 @@ namespace dxvk {
+@@ -4205,7 +4210,7 @@ namespace dxvk {
    template<bool Indexed>
    void DxvkContext::commitGraphicsState() {
      if (m_flags.test(DxvkContextFlag::GpDirtyFramebuffer))
@@ -78,7 +79,7 @@ index 6df6e3c3..bb802c96 100644
  
      if (!m_flags.test(DxvkContextFlag::GpRenderPassBound))
        this->startRenderPass();
-@@ -4288,4 +4293,13 @@ namespace dxvk {
+@@ -4465,4 +4470,13 @@ namespace dxvk {
      }
    }
    
@@ -95,10 +96,10 @@ index 6df6e3c3..bb802c96 100644
 +  }
 +}
 diff --git a/src/dxvk/dxvk_context.h b/src/dxvk/dxvk_context.h
-index 79a818e9..935cc5a4 100644
+index f54f5321..0984d085 100644
 --- a/src/dxvk/dxvk_context.h
 +++ b/src/dxvk/dxvk_context.h
-@@ -1138,7 +1138,7 @@ namespace dxvk {
+@@ -1141,7 +1141,7 @@ namespace dxvk {
              VkDescriptorSet         set,
        const DxvkPipelineLayout*     layout);
  
@@ -107,50 +108,21 @@ index 79a818e9..935cc5a4 100644
      
      void updateIndexBufferBinding();
      void updateVertexBufferBindings();
-@@ -1163,6 +1163,8 @@ namespace dxvk {
-     
-     void commitGraphicsPostBarriers();
+@@ -1180,6 +1180,7 @@ namespace dxvk {
+ 
+     void trackDrawBuffer();
  
 +    bool checkAsyncCompilationCompat();
-+
-     void emitMemoryBarrier(
-             VkPipelineStageFlags      srcStages,
-             VkAccessFlags             srcAccess,
-diff --git a/src/dxvk/dxvk_framebuffer.h b/src/dxvk/dxvk_framebuffer.h
-index 49133845..ff80297b 100644
---- a/src/dxvk/dxvk_framebuffer.h
-+++ b/src/dxvk/dxvk_framebuffer.h
-@@ -120,8 +120,8 @@ namespace dxvk {
-      * \brief Retrieves render pass
-      * \returns Render pass reference
-      */
--    const DxvkRenderPass& getRenderPass() const {
--      return *m_renderPass;
-+    const Rc<DxvkRenderPass>& getRenderPass() const {
-+      return m_renderPass;
-     }
-     
-     /**
-@@ -221,4 +221,4 @@ namespace dxvk {
-     
    };
    
 -}
 \ No newline at end of file
 +}
 diff --git a/src/dxvk/dxvk_graphics.cpp b/src/dxvk/dxvk_graphics.cpp
-index 92403199..367b955c 100644
+index 5e05b425..8586c794 100644
 --- a/src/dxvk/dxvk_graphics.cpp
 +++ b/src/dxvk/dxvk_graphics.cpp
-@@ -4,6 +4,7 @@
- #include "dxvk_device.h"
- #include "dxvk_graphics.h"
- #include "dxvk_pipemanager.h"
-+#include "dxvk_pipecompiler.h"
- #include "dxvk_spec_const.h"
- #include "dxvk_state_cache.h"
- 
-@@ -79,8 +80,6 @@ namespace dxvk {
+@@ -75,8 +75,6 @@ namespace dxvk {
    
    
    DxvkGraphicsPipeline::~DxvkGraphicsPipeline() {
@@ -159,151 +131,38 @@ index 92403199..367b955c 100644
    }
    
    
-@@ -100,8 +99,9 @@ namespace dxvk {
+@@ -96,7 +94,8 @@ namespace dxvk {
  
    VkPipeline DxvkGraphicsPipeline::getPipelineHandle(
      const DxvkGraphicsPipelineStateInfo& state,
--    const DxvkRenderPass&                renderPass) {
--    VkRenderPass renderPassHandle = renderPass.getDefaultHandle();
-+    const Rc<DxvkRenderPass>&            renderPass,
+-    const DxvkRenderPass*                renderPass) {
++    const DxvkRenderPass*                renderPass,
 +          bool                           async) {
-+    VkRenderPass renderPassHandle = renderPass->getDefaultHandle();
-     
-     VkPipeline newPipelineHandle = VK_NULL_HANDLE;
+     DxvkGraphicsPipelineInstance* instance = nullptr;
  
-@@ -119,26 +119,47 @@ namespace dxvk {
+     { std::lock_guard<sync::Spinlock> lock(m_mutex);
+@@ -105,10 +104,13 @@ namespace dxvk {
        
-       // If no pipeline instance exists with the given state
-       // vector, create a new one and add it to the list.
--      newPipelineHandle = this->compilePipeline(state, renderPass);
-+      Rc<DxvkGraphicsPipelineInstance> newPipeline =
-+        new DxvkGraphicsPipelineInstance(m_pipeMgr->m_device->vkd(),
-+          state, renderPassHandle, VK_NULL_HANDLE);
+       if (instance)
+         return instance->pipeline();
+-      
+-      instance = this->createInstance(state, renderPass);
 +
 +      if (async && m_pipeMgr->m_compiler != nullptr)
-+        m_pipeMgr->m_compiler->queueCompilation(this, newPipeline, renderPass);
++        m_pipeMgr->m_compiler->queueCompilation(this, state, renderPass);
 +      else
-+        newPipelineHandle = this->compileInstance(newPipeline, renderPass);
- 
-       // Add new pipeline to the set
--      m_pipelines.emplace_back(state, renderPassHandle, newPipelineHandle);
-+      m_pipelines.push_back(newPipeline);
-       m_pipeMgr->m_numGraphicsPipelines += 1;
++        instance = this->createInstance(state, renderPass);
      }
-     
--    if (newPipelineHandle != VK_NULL_HANDLE)
--      this->writePipelineStateToCache(state, renderPass.format());
-+    return newPipelineHandle;
-+  }
-+  
-+  
-+  VkPipeline DxvkGraphicsPipeline::compileInstance(
-+    const Rc<DxvkGraphicsPipelineInstance>& instance,
-+    const Rc<DxvkRenderPass>&               renderPass) {
-+    VkPipeline newPipelineHandle = this->compilePipeline(
-+      instance->m_stateVector, *renderPass);
-+
-+    if (!instance->setPipeline(newPipelineHandle)) {
-+      m_vkd->vkDestroyPipeline(m_vkd->device(), newPipelineHandle, nullptr);
-+    } else {
-+      if (newPipelineHandle != VK_NULL_HANDLE)
-+        this->writePipelineStateToCache(instance->m_stateVector, renderPass->format());
-+    }
-     
-     return newPipelineHandle;
-   }
-   
-   
--  const DxvkGraphicsPipelineInstance* DxvkGraphicsPipeline::findInstance(
-+  DxvkGraphicsPipelineInstance* DxvkGraphicsPipeline::findInstance(
-     const DxvkGraphicsPipelineStateInfo& state,
-           VkRenderPass                   renderPass) const {
-     for (const auto& instance : m_pipelines) {
--      if (instance.isCompatible(state, renderPass))
--        return &instance;
-+      if (instance->isCompatible(state, renderPass))
-+        return instance.ptr();
-     }
-     
-     return nullptr;
+-    
++
+     if (!instance)
+       return VK_NULL_HANDLE;
+ 
 diff --git a/src/dxvk/dxvk_graphics.h b/src/dxvk/dxvk_graphics.h
-index 976cf590..2d0fbd2b 100644
+index 168e9714..07f0883d 100644
 --- a/src/dxvk/dxvk_graphics.h
 +++ b/src/dxvk/dxvk_graphics.h
-@@ -132,8 +132,8 @@ namespace dxvk {
-    * Stores a state vector and the
-    * corresponding pipeline handle.
-    */
--  class DxvkGraphicsPipelineInstance {
--
-+  class DxvkGraphicsPipelineInstance: public RcObject {
-+    friend class DxvkGraphicsPipeline;
-   public:
- 
-     DxvkGraphicsPipelineInstance()
-@@ -141,13 +141,19 @@ namespace dxvk {
-       m_renderPass  (VK_NULL_HANDLE),
-       m_pipeline    (VK_NULL_HANDLE) { }
-     DxvkGraphicsPipelineInstance(
-+      const Rc<vk::DeviceFn>&               vkd,
-       const DxvkGraphicsPipelineStateInfo&  state,
-             VkRenderPass                    rp,
-             VkPipeline                      pipe)
--    : m_stateVector (state),
-+    : m_vkd         (vkd),
-+      m_stateVector (state),
-       m_renderPass  (rp),
-       m_pipeline    (pipe) { }
- 
-+    ~DxvkGraphicsPipelineInstance() {
-+      m_vkd->vkDestroyPipeline(m_vkd->device(), m_pipeline, nullptr);
-+    }
-+
-     /**
-      * \brief Checks for matching pipeline state
-      * 
-@@ -162,23 +168,38 @@ namespace dxvk {
-           && m_renderPass  == rp;
-     }
- 
-+    /**
-+     * \brief Sets the pipeline handle
-+     *
-+     * If a pipeline handle has already been
-+     * set up, this method will fail and the new pipeline
-+     * handle should be destroyed.
-+     * \param [in] pipeline The pipeline
-+     */
-+    bool setPipeline(VkPipeline pipeline) {
-+      VkPipeline expected = VK_NULL_HANDLE;
-+      return m_pipeline.compare_exchange_strong(expected, pipeline);
-+    }
-+
-     /**
-      * \brief Retrieves pipeline
-      * \returns The pipeline handle
-      */
-     VkPipeline pipeline() const {
--      return m_pipeline;
-+      return m_pipeline.load();
-     }
- 
-   private:
- 
-+    const Rc<vk::DeviceFn>        m_vkd;
-     DxvkGraphicsPipelineStateInfo m_stateVector;
-     VkRenderPass                  m_renderPass;
--    VkPipeline                    m_pipeline;
-+    std::atomic<VkPipeline>       m_pipeline;
- 
-   };
- 
-   
-+
-   /**
-    * \brief Graphics pipeline
-    * 
-@@ -237,19 +258,19 @@ namespace dxvk {
+@@ -247,11 +247,13 @@ namespace dxvk {
       * state. If necessary, a new pipeline will be created.
       * \param [in] state Pipeline state vector
       * \param [in] renderPass The render pass
@@ -312,50 +171,17 @@ index 976cf590..2d0fbd2b 100644
       */
      VkPipeline getPipelineHandle(
        const DxvkGraphicsPipelineStateInfo&    state,
--      const DxvkRenderPass&                   renderPass);
-+      const Rc<DxvkRenderPass>&               renderPass,
+-      const DxvkRenderPass*                   renderPass);
++      const DxvkRenderPass*                   renderPass,
 +            bool                              async);
      
--  private:
-+    VkPipeline compileInstance(
-+      const Rc<DxvkGraphicsPipelineInstance>& instance,
-+      const Rc<DxvkRenderPass>&               renderPass);
-     
--    struct PipelineStruct {
--      DxvkGraphicsPipelineStateInfo stateVector;
--      VkRenderPass                  renderPass;
--      VkPipeline                    pipeline;
--    };
-+  private:
-     
-     Rc<vk::DeviceFn>          m_vkd;
-     DxvkPipelineManager*      m_pipeMgr;
-@@ -270,10 +291,10 @@ namespace dxvk {
-     DxvkGraphicsCommonPipelineStateInfo m_common;
-     
-     // List of pipeline instances, shared between threads
--    alignas(CACHE_LINE_SIZE) sync::Spinlock   m_mutex;
--    std::vector<DxvkGraphicsPipelineInstance> m_pipelines;
-+    alignas(CACHE_LINE_SIZE) sync::Spinlock       m_mutex;
-+    std::vector<Rc<DxvkGraphicsPipelineInstance>> m_pipelines;
-     
--    const DxvkGraphicsPipelineInstance* findInstance(
-+    DxvkGraphicsPipelineInstance* findInstance(
-       const DxvkGraphicsPipelineStateInfo& state,
-             VkRenderPass                   renderPass) const;
-     
-@@ -301,4 +322,4 @@ namespace dxvk {
-     
-   };
-   
--}
-\ No newline at end of file
-+}
+     /**
+      * \brief Compiles a pipeline
 diff --git a/src/dxvk/dxvk_image.h b/src/dxvk/dxvk_image.h
-index 1035a23d..209f4f06 100644
+index 8f82d65a..156f6054 100644
 --- a/src/dxvk/dxvk_image.h
 +++ b/src/dxvk/dxvk_image.h
-@@ -439,6 +439,37 @@ namespace dxvk {
+@@ -442,6 +442,37 @@ namespace dxvk {
        return result;
      }
  
@@ -393,7 +219,7 @@ index 1035a23d..209f4f06 100644
    private:
      
      Rc<vk::DeviceFn>  m_vkd;
-@@ -447,6 +478,9 @@ namespace dxvk {
+@@ -450,6 +481,9 @@ namespace dxvk {
      DxvkImageViewCreateInfo m_info;
      VkImageView             m_views[ViewCount];
  
@@ -404,19 +230,19 @@ index 1035a23d..209f4f06 100644
      
    };
 diff --git a/src/dxvk/dxvk_options.cpp b/src/dxvk/dxvk_options.cpp
-index f18fd70f..e051c06d 100644
+index 394feabb..df05ecd8 100644
 --- a/src/dxvk/dxvk_options.cpp
 +++ b/src/dxvk/dxvk_options.cpp
 @@ -5,6 +5,7 @@ namespace dxvk {
    DxvkOptions::DxvkOptions(const Config& config) {
      enableStateCache      = config.getOption<bool>    ("dxvk.enableStateCache",       true);
-     enableTransferQueue   = config.getOption<bool>    ("dxvk.enableTransferQueue",    true);
-+    useAsync              = config.getOption<bool>    ("dxvk.useAsync",               false);
+     enableOpenVR          = config.getOption<bool>    ("dxvk.enableOpenVR",           true);
++    useAsync              = config.getOption<bool>    ("dxvk.useAsync",               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);
 diff --git a/src/dxvk/dxvk_options.h b/src/dxvk/dxvk_options.h
-index aec003e4..0fbb6dd0 100644
+index 447294b5..571d6c15 100644
 --- a/src/dxvk/dxvk_options.h
 +++ b/src/dxvk/dxvk_options.h
 @@ -10,6 +10,7 @@ namespace dxvk {
@@ -429,7 +255,7 @@ index aec003e4..0fbb6dd0 100644
      bool enableTransferQueue;
 diff --git a/src/dxvk/dxvk_pipecompiler.cpp b/src/dxvk/dxvk_pipecompiler.cpp
 new file mode 100644
-index 00000000..a70c9908
+index 00000000..cfef228a
 --- /dev/null
 +++ b/src/dxvk/dxvk_pipecompiler.cpp
 @@ -0,0 +1,69 @@
@@ -468,11 +294,11 @@ index 00000000..a70c9908
 +
 +
 +  void DxvkPipelineCompiler::queueCompilation(
-+    const Rc<DxvkGraphicsPipeline>&         pipeline,
-+    const Rc<DxvkGraphicsPipelineInstance>& instance,
-+    const Rc<DxvkRenderPass>&               renderPass) {
++    DxvkGraphicsPipeline*                   pipeline,
++    const DxvkGraphicsPipelineStateInfo&    state,
++    const DxvkRenderPass*                   renderPass) {
 +    std::unique_lock<std::mutex> lock(m_compilerLock);
-+    m_compilerQueue.push({ pipeline, instance, renderPass });
++    m_compilerQueue.push({ pipeline, state, renderPass });
 +    m_compilerCond.notify_one();
 +  }
 +
@@ -496,15 +322,15 @@ index 00000000..a70c9908
 +        }
 +      }
 +
-+      if (entry.pipeline != nullptr && entry.instance != nullptr)
-+        entry.pipeline->compileInstance(entry.instance, entry.renderPass);
++      if (entry.pipeline != nullptr && entry.renderPass != nullptr)
++        entry.pipeline->compilePipeline(entry.state, entry.renderPass);
 +    }
 +  }
 +
 +}
 diff --git a/src/dxvk/dxvk_pipecompiler.h b/src/dxvk/dxvk_pipecompiler.h
 new file mode 100644
-index 00000000..26e983cc
+index 00000000..a82fff60
 --- /dev/null
 +++ b/src/dxvk/dxvk_pipecompiler.h
 @@ -0,0 +1,60 @@
@@ -521,7 +347,7 @@ index 00000000..26e983cc
 +namespace dxvk {
 +
 +  class DxvkGraphicsPipeline;
-+  class DxvkGraphicsPipelineInstance;
++  class DxvkGraphicsPipelineStateInfo;
 +
 +  /**
 +   * \brief Pipeline compiler
@@ -541,20 +367,20 @@ index 00000000..26e983cc
 +     * This should be used to compile graphics
 +     * pipeline instances asynchronously.
 +     * \param [in] pipeline The pipeline object
-+     * \param [in] instance The pipeline instance
++     * \param [in] state The pipeline state info object
 +     * \param [in] renderPass
 +     */
 +    void queueCompilation(
-+      const Rc<DxvkGraphicsPipeline>&         pipeline,
-+      const Rc<DxvkGraphicsPipelineInstance>& instance,
-+      const Rc<DxvkRenderPass>&               renderPass);
++      DxvkGraphicsPipeline*                   pipeline,
++      const DxvkGraphicsPipelineStateInfo&    state,
++      const DxvkRenderPass*                   renderPass);
 +
 +  private:
 +
 +    struct PipelineEntry {
-+      Rc<DxvkGraphicsPipeline>         pipeline;
-+      Rc<DxvkGraphicsPipelineInstance> instance;
-+      Rc<DxvkRenderPass>               renderPass;
++      DxvkGraphicsPipeline*                   pipeline = nullptr;
++      DxvkGraphicsPipelineStateInfo           state;
++      const DxvkRenderPass*                   renderPass = nullptr;
 +    };
 +
 +    std::atomic<bool>           m_compilerStop = { false };
@@ -569,7 +395,7 @@ index 00000000..26e983cc
 +
 +}
 diff --git a/src/dxvk/dxvk_pipemanager.cpp b/src/dxvk/dxvk_pipemanager.cpp
-index 682db83d..cb7e993b 100644
+index 378bb3d1..3323f926 100644
 --- a/src/dxvk/dxvk_pipemanager.cpp
 +++ b/src/dxvk/dxvk_pipemanager.cpp
 @@ -45,9 +45,13 @@ namespace dxvk {
@@ -587,10 +413,10 @@ index 682db83d..cb7e993b 100644
    
    
 diff --git a/src/dxvk/dxvk_pipemanager.h b/src/dxvk/dxvk_pipemanager.h
-index fb62cb9a..c3b846cf 100644
+index f0087d15..29e758c7 100644
 --- a/src/dxvk/dxvk_pipemanager.h
 +++ b/src/dxvk/dxvk_pipemanager.h
-@@ -5,6 +5,7 @@
+@@ -6,6 +6,7 @@
  
  #include "dxvk_compute.h"
  #include "dxvk_graphics.h"
@@ -598,7 +424,7 @@ index fb62cb9a..c3b846cf 100644
  
  namespace dxvk {
  
-@@ -139,6 +140,7 @@ namespace dxvk {
+@@ -107,6 +108,7 @@ namespace dxvk {
      const DxvkDevice*         m_device;
      Rc<DxvkPipelineCache>     m_cache;
      Rc<DxvkStateCache>        m_stateCache;
@@ -606,24 +432,11 @@ index fb62cb9a..c3b846cf 100644
  
      std::atomic<uint32_t>     m_numComputePipelines  = { 0 };
      std::atomic<uint32_t>     m_numGraphicsPipelines = { 0 };
-diff --git a/src/dxvk/dxvk_state_cache.cpp b/src/dxvk/dxvk_state_cache.cpp
-index a9249a3a..4d296c19 100644
---- a/src/dxvk/dxvk_state_cache.cpp
-+++ b/src/dxvk/dxvk_state_cache.cpp
-@@ -263,7 +263,7 @@ namespace dxvk {
-         const auto& entry = m_entries[e->second];
- 
-         auto rp = m_passManager->getRenderPass(entry.format);
--        pipeline->getPipelineHandle(entry.gpState, *rp);
-+        pipeline->getPipelineHandle(entry.gpState, rp, false);
-       }
-     } else {
-       auto pipeline = m_pipeManager->createComputePipeline(item.cs);
 diff --git a/src/dxvk/meson.build b/src/dxvk/meson.build
-index 604434c2..bf90af2d 100644
+index 867f3002..f4c4336c 100644
 --- a/src/dxvk/meson.build
 +++ b/src/dxvk/meson.build
-@@ -77,6 +77,7 @@ dxvk_src = files([
+@@ -82,6 +82,7 @@ dxvk_src = files([
    'dxvk_openvr.cpp',
    'dxvk_options.cpp',
    'dxvk_pipecache.cpp',
@@ -631,23 +444,3 @@ index 604434c2..bf90af2d 100644
    'dxvk_pipelayout.cpp',
    'dxvk_pipemanager.cpp',
    'dxvk_queue.cpp',
-diff --git a/dxvk.conf b/dxvk.conf
-index 25d27cb9..4e409075 100644
---- a/dxvk.conf
-+++ b/dxvk.conf
-@@ -159,6 +159,15 @@
- # dxvk.numCompilerThreads = 0
- 
- 
-+# Compile pipelines asynchronously if possible. This may reduce stuttering
-+# in some games, but may also introduce rendering issues that might become
-+# apparent over time. Do not report bugs with this option enabled.
-+#
-+# Supported values: True, False
-+
-+# dxvk.useAsync = False
-+
-+
- # Toggles raw SSBO usage.
- # 
- # Uses storage buffers to implement raw and structured buffer
diff --git a/dxvk-mangohud.patch b/dxvk-mangohud.patch
new file mode 100644
index 000000000000..13fbdee8db9a
--- /dev/null
+++ b/dxvk-mangohud.patch
@@ -0,0 +1,665 @@
+diff --git a/README.md b/README.md
+index 25d03230..e8a53432 100644
+--- a/README.md
++++ b/README.md
+@@ -1,3 +1,29 @@
++# MangoHud Specific
++
++- `DXVK_HUD_OFFSET_X` Set X offset of the DVXK Hud.
++- `DXVK_HUD_OFFSET_Y` Set Y offset of the DVXK Hud.
++- `DXVK_LOG_TO_FILE` Turn on logging and select path/filename (Fps,Cpu load,Gpu load)
++- Logging Gpu load requires either mangogpuload or gpuload hud options
++
++# Hud options
++- `mangogpuload` : Shows current gpu load.
++- `mangocpuload` : Shows current cpu load.
++
++# Keybinds
++- `F2`  : Toggle Logging on/off
++- `F12` : Toggle Hud on/off
++
++# MangoLog file
++
++ When you press F2, a file is created with your chosen name + date/time stamp.
++ this file can be uploaded to https://flightlessmango.com/logs/new to create graphs automatically.
++ you can share the created page with others, just link it.
++ 
++ #### Multiple log files
++ 
++ It's possible to upload multiple files, you can rename them to your preferred names and upload them in a batch.
++ The graphs will then use those names in the data.
++
+ # DXVK
+ 
+ A Vulkan-based translation layer for Direct3D 10/11 which allows running 3D applications on Linux using Wine.
+diff --git a/src/dxvk/dxvk_cpu.h b/src/dxvk/dxvk_cpu.h
+new file mode 100644
+index 00000000..b2c8736a
+--- /dev/null
++++ b/src/dxvk/dxvk_cpu.h
+@@ -0,0 +1,172 @@
++#include <cmath>
++#include <iomanip>
++#include <array>
++#include <vector>
++#include <algorithm>
++#include <iterator>
++#include <thread>
++#include <sstream>
++#include <fstream>
++using namespace std;
++
++const int NUM_CPU_STATES = 10;
++
++struct Cpus{
++  size_t num;
++  string name;
++  float value;
++  string output;
++};
++
++size_t numCpuCores = dxvk::thread::hardware_concurrency();
++size_t arraySize = numCpuCores + 1;
++std::vector<Cpus> cpuArray;
++
++void coreCounting(){
++  cpuArray.push_back({0, "CPU:"});
++  for (size_t i = 0; i < arraySize; i++) {
++    size_t offset = i;
++    stringstream ss;
++    ss << "CPU" << offset << ":";
++    string cpuNameString = ss.str();
++    cpuArray.push_back({i+1 , cpuNameString});
++  }
++}
++
++std::string m_cpuUtilizationString;
++
++enum CPUStates
++{
++	S_USER = 0,
++	S_NICE,
++	S_SYSTEM,
++	S_IDLE,
++	S_IOWAIT,
++	S_IRQ,
++	S_SOFTIRQ,
++	S_STEAL,
++	S_GUEST,
++	S_GUEST_NICE
++};
++
++typedef struct CPUData
++{
++	std::string cpu;
++	size_t times[NUM_CPU_STATES];
++} CPUData;
++
++void ReadStatsCPU(std::vector<CPUData> & entries)
++{
++	std::ifstream fileStat("/proc/stat");
++
++	std::string line;
++
++	const std::string STR_CPU("cpu");
++	const std::size_t LEN_STR_CPU = STR_CPU.size();
++	const std::string STR_TOT("tot");
++
++	while(std::getline(fileStat, line))
++	{
++		// cpu stats line found
++		if(!line.compare(0, LEN_STR_CPU, STR_CPU))
++		{
++			std::istringstream ss(line);
++
++			// store entry
++			entries.emplace_back(CPUData());
++			CPUData & entry = entries.back();
++
++			// read cpu label
++			ss >> entry.cpu;
++
++			if(entry.cpu.size() > LEN_STR_CPU)
++				entry.cpu.erase(0, LEN_STR_CPU);
++			else
++				entry.cpu = STR_TOT;
++
++			// read times
++			for(int i = 0; i < NUM_CPU_STATES; ++i)
++				ss >> entry.times[i];
++		}
++	}
++}
++
++size_t GetIdleTime(const CPUData & e)
++{
++	return	e.times[S_IDLE] +
++			e.times[S_IOWAIT];
++}
++
++size_t GetActiveTime(const CPUData & e)
++{
++	return	e.times[S_USER] +
++			e.times[S_NICE] +
++			e.times[S_SYSTEM] +
++			e.times[S_IRQ] +
++			e.times[S_SOFTIRQ] +
++			e.times[S_STEAL] +
++			e.times[S_GUEST] +
++			e.times[S_GUEST_NICE];
++}
++
++void PrintStats(const std::vector<CPUData> & entries1, const std::vector<CPUData> & entries2)
++{
++	const size_t NUM_ENTRIES = entries1.size();
++
++	for(size_t i = 0; i < NUM_ENTRIES; ++i)
++	{
++		const CPUData & e1 = entries1[i];
++		const CPUData & e2 = entries2[i];
++
++		const float ACTIVE_TIME	= static_cast<float>(GetActiveTime(e2) - GetActiveTime(e1));
++		const float IDLE_TIME	= static_cast<float>(GetIdleTime(e2) - GetIdleTime(e1));
++		const float TOTAL_TIME	= ACTIVE_TIME + IDLE_TIME;
++
++    cpuArray[i].value = (truncf(100.f * ACTIVE_TIME / TOTAL_TIME) * 10 / 10);
++	}
++}
++
++int getCpuUsage()
++{
++	std::vector<CPUData> entries1;
++	std::vector<CPUData> entries2;
++
++	// snapshot 1
++	ReadStatsCPU(entries1);
++
++	// 100ms pause
++	std::this_thread::sleep_for(std::chrono::milliseconds(100));
++
++	// snapshot 2
++	ReadStatsCPU(entries2);
++
++	// print output
++	PrintStats(entries1, entries2);
++
++	return 0;
++}
++
++
++void updateCpuStrings(){
++  for (size_t i = 0; i < arraySize; i++) {
++    size_t spacing = 10;
++    string value = to_string(cpuArray[i].value);
++    value.erase( value.find_last_not_of('0') + 1, std::string::npos );
++    size_t correctionValue = (spacing - cpuArray[i].name.length()) - value.length();
++    string correction = "";
++    for (size_t i = 0; i < correctionValue; i++) {
++          correction.append(" ");
++        }
++        stringstream ss;
++        if (i < 11) {
++          if (i == 0) {
++            ss << cpuArray[i].name << correction << cpuArray[i].value << "%";
++          } else {
++            ss << cpuArray[i].name << correction << cpuArray[i].value << "%";
++          }
++        } else {
++          ss << cpuArray[i].name << correction << cpuArray[i].value << "%";
++        }
++        cpuArray[i].output = ss.str();
++      }
++    }
+\ No newline at end of file
+diff --git a/src/dxvk/hud/dxvk_hud.cpp b/src/dxvk/hud/dxvk_hud.cpp
+index 4fcd3bd2..b6e23c4c 100644
+--- a/src/dxvk/hud/dxvk_hud.cpp
++++ b/src/dxvk/hud/dxvk_hud.cpp
+@@ -1,9 +1,13 @@
+-#include <cstring>
+ #include <version.h>
+-
+ #include "dxvk_hud.h"
++#include <windows.h>
++#include <time.h>
+ 
+ namespace dxvk::hud {
++  float Hud::offset_x_float = 0;
++  float Hud::offset_y_float = 0;
++  bool show_hud = true;
++  int lastPress;
+   
+   Hud::Hud(
+     const Rc<DxvkDevice>& device,
+@@ -60,6 +64,14 @@ namespace dxvk::hud {
+   
+   Rc<Hud> Hud::createHud(const Rc<DxvkDevice>& device) {
+     std::string hudElements = env::getEnvVar("DXVK_HUD");
++    std::string offset_x = env::getEnvVar("DXVK_HUD_OFFSET_X");
++    std::string offset_y = env::getEnvVar("DXVK_HUD_OFFSET_Y");
++    
++    if (!offset_x.empty())
++      offset_x_float = stof(offset_x);
++    
++    if (!offset_y.empty())
++      offset_y_float = stof(offset_y);
+ 
+     if (hudElements.empty())
+       hudElements = device->config().hud;
+@@ -82,33 +94,45 @@ namespace dxvk::hud {
+     m_renderer.beginFrame(ctx, m_uniformData.surfaceSize);
+   }
+ 
+-
+   void Hud::renderHudElements(const Rc<DxvkContext>& ctx) {
+-    HudPos position = { 8.0f, 24.0f };
++    if(GetAsyncKeyState(VK_F12) & 0x8000)
++    {
++      if (GetTickCount() - lastPress > 500){
++        lastPress = GetTickCount();
++        if (show_hud){
++          show_hud = false;
++        } else {
++          show_hud = true;
++        }
++      }
++    } 
+     
+-    if (m_config.elements.test(HudElement::DxvkVersion)) {
+-      m_renderer.drawText(ctx, 16.0f,
+-        { position.x, position.y },
+-        { 1.0f, 1.0f, 1.0f, 1.0f },
+-        "DXVK " DXVK_VERSION);
+-      position.y += 24.0f;
+-    }
+-
+-    if (m_config.elements.test(HudElement::DxvkClientApi)) {
+-      m_renderer.drawText(ctx, 16.0f,
+-        { position.x, position.y },
+-        { 1.0f, 1.0f, 1.0f, 1.0f },
+-        m_device->clientApi());
+-      position.y += 24.0f;
+-    }
+-
+-    if (m_config.elements.test(HudElement::DeviceInfo)) {
+-      position = m_hudDeviceInfo.render(
+-        ctx, m_renderer, position);
+-    }
++    HudPos position = { offset_x_float + 8.0f, offset_y_float + 24.0f };
+     
+-    position = m_hudFramerate.render(ctx, m_renderer, position);
+-    position = m_hudStats    .render(ctx, m_renderer, position);
++    if (show_hud){
++      if (m_config.elements.test(HudElement::DxvkVersion)) {
++        m_renderer.drawText(ctx, 16.0f,
++          { position.x, position.y },
++          { 1.0f, 1.0f, 1.0f, 1.0f },
++          "DXVK " DXVK_VERSION);
++        position.y += 24.0f;
++      }
++
++      if (m_config.elements.test(HudElement::DxvkClientApi)) {
++        m_renderer.drawText(ctx, 16.0f,
++          { position.x, position.y },
++          { 1.0f, 1.0f, 1.0f, 1.0f },
++          m_device->clientApi());
++        position.y += 24.0f;
++      }
++
++      if (m_config.elements.test(HudElement::DeviceInfo)) {
++        position = m_hudDeviceInfo.render(
++          ctx, m_renderer, position);
++      }
++      position = m_hudFramerate.render(ctx, m_renderer, position);
++      position = m_hudStats    .render(ctx, m_renderer, position);
++    }
+   }
+   
+   
+diff --git a/src/dxvk/hud/dxvk_hud.h b/src/dxvk/hud/dxvk_hud.h
+index 3b702c3b..71779977 100644
+--- a/src/dxvk/hud/dxvk_hud.h
++++ b/src/dxvk/hud/dxvk_hud.h
+@@ -29,6 +29,8 @@ namespace dxvk::hud {
+   class Hud : public RcObject {
+     
+   public:
++    static float offset_x_float;
++    static float offset_y_float;
+     
+     Hud(
+       const Rc<DxvkDevice>& device,
+diff --git a/src/dxvk/hud/dxvk_hud_config.cpp b/src/dxvk/hud/dxvk_hud_config.cpp
+index fe1745bd..020a394b 100644
+--- a/src/dxvk/hud/dxvk_hud_config.cpp
++++ b/src/dxvk/hud/dxvk_hud_config.cpp
+@@ -16,6 +16,9 @@ namespace dxvk::hud {
+     { "version",      HudElement::DxvkVersion       },
+     { "api",          HudElement::DxvkClientApi     },
+     { "compiler",     HudElement::CompilerActivity  },
++    { "mangogpuload", HudElement::GpuLoad           },
++    { "mangocpuload", HudElement::CpuLoad           },
++    { "mangocpuload", HudElement::Logging           },
+   }};
+   
+   
+diff --git a/src/dxvk/hud/dxvk_hud_config.h b/src/dxvk/hud/dxvk_hud_config.h
+index 05a1e4b6..5d571663 100644
+--- a/src/dxvk/hud/dxvk_hud_config.h
++++ b/src/dxvk/hud/dxvk_hud_config.h
+@@ -22,6 +22,9 @@ namespace dxvk::hud {
+     DxvkVersion       = 8,
+     DxvkClientApi     = 9,
+     CompilerActivity  = 10,
++    GpuLoad           = 11,
++    CpuLoad           = 12,
++    Logging           = 13,
+   };
+   
+   using HudElements = Flags<HudElement>;
+diff --git a/src/dxvk/hud/dxvk_hud_fps.cpp b/src/dxvk/hud/dxvk_hud_fps.cpp
+index f8cb6e7d..84eb2db6 100644
+--- a/src/dxvk/hud/dxvk_hud_fps.cpp
++++ b/src/dxvk/hud/dxvk_hud_fps.cpp
+@@ -1,7 +1,22 @@
+ #include "dxvk_hud_fps.h"
++#include "dxvk_hud_stats.h"
++#include "../dxvk_cpu.h"
++#include <time.h>
+ 
+ #include <cmath>
+ #include <iomanip>
++using namespace std;
++time_t now_log = time(0);
++tm *log_time = localtime(&now_log);
++fstream f;
++
++struct logData{
++  float fps;
++  float cpu;
++  uint64_t gpu;
++};
++
++std::vector<logData> logArray;
+ 
+ namespace dxvk::hud {
+   
+@@ -9,8 +24,8 @@ namespace dxvk::hud {
+   : m_elements  (elements),
+     m_fpsString ("FPS: "),
+     m_prevFpsUpdate(Clock::now()),
+-    m_prevFtgUpdate(Clock::now()) {
+-    
++    m_prevFtgUpdate(Clock::now()),
++    m_prevLogUpdate(Clock::now()) {
+   }
+   
+   
+@@ -25,11 +40,48 @@ namespace dxvk::hud {
+     TimePoint now = Clock::now();
+     TimeDiff elapsedFps = std::chrono::duration_cast<TimeDiff>(now - m_prevFpsUpdate);
+     TimeDiff elapsedFtg = std::chrono::duration_cast<TimeDiff>(now - m_prevFtgUpdate);
++    TimeDiff elapsedLog = std::chrono::duration_cast<TimeDiff>(now - m_prevLogUpdate);
+     m_prevFtgUpdate = now;
++
++    if(GetAsyncKeyState(VK_F2) & 0x8000)
++    {
++      elapsedF2 = std::chrono::duration_cast<TimeDiff>(now - m_prevF2Press);
++      if (elapsedF2.count() > UpdateInterval || elapsedF2.count() == 0) {
++          if (mango_logging){
++            m_prevF2Press = now;
++            mango_logging = false;
++            for (size_t i = 0; i < logArray.size(); i++) {
++              f << logArray[i].fps << "," << logArray[i].cpu << "," << logArray[i].gpu << endl;
++            }
++            f.close();
++            logArray.clear();
++          } else {
++            m_prevF2Press = now;
++            now_log = time(0);
++            log_time = localtime(&now_log);
++            mango_logging = true;
++            string date = to_string(log_time->tm_year + 1900) + "-" + to_string(1 + log_time->tm_mon) + "-" + to_string(log_time->tm_mday) + "_" + to_string(1 + log_time->tm_hour) + "-" + to_string(1 + log_time->tm_min) + "-" + to_string(1 + log_time->tm_sec);
++            f.open(logging + "_" + date, f.out | f.app);
++          }
++        } 
++      }
++    
++    if (elapsedLog.count() >= LogUpdateInterval) {
++      fps = (10'000'000ll * m_frameCount) / elapsedFps.count();
++      if (!logging.empty()){
++        if (mango_logging){
++          logArray.push_back({float(fps / 10 + (float(fps % 10) / 10)), cpuArray[0].value, gpuLoad});
++        }
++      }
++      m_prevLogUpdate = now;
++    }
+     
+-    // Update FPS string
+     if (elapsedFps.count() >= UpdateInterval) {
+-      const int64_t fps = (10'000'000ll * m_frameCount) / elapsedFps.count();
++    // Update FPS string
++      coreCounting();
++      dxvk::thread([this] () { getCpuUsage();});
++      updateCpuStrings();
++      m_cpuUtilizationString = str::format(cpuArray[0].output);
+       m_fpsString = str::format("FPS: ", fps / 10, ".", fps % 10);
+       
+       m_prevFpsUpdate = now;
+@@ -46,32 +98,81 @@ namespace dxvk::hud {
+     const Rc<DxvkContext>&  context,
+           HudRenderer&      renderer,
+           HudPos            position) {
++    if (m_elements.test(HudElement::GpuLoad)) {
++      position = this->renderGpuText(
++        context, renderer, position);
++      }
++    if (m_elements.test(HudElement::CpuLoad)) {
++      position = this->renderCpuText(
++        context, renderer, position);
++    }
+     if (m_elements.test(HudElement::Framerate)) {
+       position = this->renderFpsText(
+         context, renderer, position);
+-    }
++      }
+     
+     if (m_elements.test(HudElement::Frametimes)) {
+       position = this->renderFrametimeGraph(
+         context, renderer, position);
+     }
+     
++    if (mango_logging && !logging.empty()) {
++      this->renderLogging(context, renderer,
++        { float(renderer.surfaceSize().width) - 250.0f, float(renderer.surfaceSize().height) - 20.0f });
++    }
++    
+     return position;
+   }
+   
+   
+-  HudPos HudFps::renderFpsText(
+-    const Rc<DxvkContext>&  context,
+-          HudRenderer&      renderer,
+-          HudPos            position) {
++  
++  HudPos HudFps::renderGpuText(
++  const Rc<DxvkContext>&  context,
++        HudRenderer&      renderer,
++        HudPos            position) {
++  renderer.drawText(context, 16.0f,
++    { position.x, position.y },
++    { 1.0f, 1.0f, 1.0f, 1.0f },
++    m_gpuLoadString);
++
++  return HudPos { position.x, position.y + 24 };
++}
++
++HudPos HudFps::renderCpuText(
++const Rc<DxvkContext>&  context,
++      HudRenderer&      renderer,
++      HudPos            position) {
++renderer.drawText(context, 16.0f,
++  { position.x, position.y },
++  { 1.0f, 1.0f, 1.0f, 1.0f },
++  m_cpuUtilizationString);
++
++return HudPos { position.x, position.y + 24 };
++}  
++
++HudPos HudFps::renderFpsText(
++  const Rc<DxvkContext>&  context,
++  HudRenderer&      renderer,
++  HudPos            position) {
+     renderer.drawText(context, 16.0f,
+       { position.x, position.y },
+       { 1.0f, 1.0f, 1.0f, 1.0f },
+       m_fpsString);
++      
++      return HudPos { position.x, position.y + 24 };
++    }
+     
+-    return HudPos { position.x, position.y + 24 };
+-  }
+-  
++  HudPos HudFps::renderLogging(
++    const Rc<DxvkContext>&  context,
++    HudRenderer&      renderer,
++    HudPos            position) {
++      renderer.drawText(context, 16.0f,
++        { position.x, position.y },
++        { 1.0f, 1.0f, 1.0f, 1.0f },
++        "Currently Logging...");
++        
++        return HudPos { position.x, position.y};
++      }
+   
+   HudPos HudFps::renderFrametimeGraph(
+     const Rc<DxvkContext>&  context,
+diff --git a/src/dxvk/hud/dxvk_hud_fps.h b/src/dxvk/hud/dxvk_hud_fps.h
+index c8c4b984..b1476889 100644
+--- a/src/dxvk/hud/dxvk_hud_fps.h
++++ b/src/dxvk/hud/dxvk_hud_fps.h
+@@ -17,8 +17,9 @@ namespace dxvk::hud {
+     using TimeDiff  = std::chrono::microseconds;
+     using TimePoint = typename Clock::time_point;
+     
+-    constexpr static uint32_t NumDataPoints  = 300;
+-    constexpr static int64_t  UpdateInterval = 500'000;
++    constexpr static uint32_t NumDataPoints     = 300;
++    constexpr static int64_t  UpdateInterval    = 500'000;
++    constexpr static int64_t  LogUpdateInterval = 100'000;
+   public:
+     
+     HudFps(HudElements elements);
+@@ -36,23 +37,46 @@ namespace dxvk::hud {
+     const HudElements m_elements;
+     
+     std::string m_fpsString;
++    bool mango_logging = false;
++    time_t lastPress = time(0);
++    std::string logging = env::getEnvVar("DXVK_LOG_TO_FILE");
++    int64_t fps;
+     
+     TimePoint m_prevFpsUpdate;
+     TimePoint m_prevFtgUpdate;
++    TimePoint m_prevLogUpdate;
++    TimePoint m_prevF2Press;
++    TimeDiff  elapsedF2;
+     int64_t   m_frameCount = 0;
+     
+     std::array<float, NumDataPoints>  m_dataPoints  = {};
+     uint32_t                          m_dataPointId = 0;
++
++    HudPos renderGpuText(
++      const Rc<DxvkContext>&  context,
++      HudRenderer&      renderer,
++      HudPos            position);
++      
++    HudPos renderCpuText(
++      const Rc<DxvkContext>&  context,
++            HudRenderer&      renderer,
++            HudPos            position);
+     
+     HudPos renderFpsText(
+       const Rc<DxvkContext>&  context,
+             HudRenderer&      renderer,
+             HudPos            position);
+     
++    
+     HudPos renderFrametimeGraph(
+       const Rc<DxvkContext>&  context,
+             HudRenderer&      renderer,
+             HudPos            position);
++            
++    HudPos renderLogging(
++      const Rc<DxvkContext>&  context,
++            HudRenderer&      renderer,
++            HudPos            position);
+     
+   };
+   
+diff --git a/src/dxvk/hud/dxvk_hud_stats.cpp b/src/dxvk/hud/dxvk_hud_stats.cpp
+index 995f186b..be8a5eb7 100644
+--- a/src/dxvk/hud/dxvk_hud_stats.cpp
++++ b/src/dxvk/hud/dxvk_hud_stats.cpp
+@@ -1,5 +1,7 @@
+ #include "dxvk_hud_stats.h"
+ 
++std::string m_gpuLoadString = "GPU: ";
++uint64_t gpuLoad;
+ namespace dxvk::hud {
+   
+   HudStats::HudStats(HudElements elements)
+@@ -24,7 +26,7 @@ namespace dxvk::hud {
+ 
+     // GPU load is a bit more complex than that since
+     // we don't want to update this every frame
+-    if (m_elements.test(HudElement::StatGpuLoad))
++    if (m_elements.test(HudElement::GpuLoad) || m_elements.test(HudElement::StatGpuLoad))
+       this->updateGpuLoad();
+   }
+   
+@@ -70,8 +72,8 @@ namespace dxvk::hud {
+       uint64_t busyTicks = ticks > m_diffGpuIdleTicks
+         ? uint64_t(ticks - m_diffGpuIdleTicks)
+         : uint64_t(0);
+-
+-      m_gpuLoadString = str::format("GPU: ", (100 * busyTicks) / ticks, "%");
++      gpuLoad = 100 * busyTicks / ticks;
++      m_gpuLoadString = str::format("GPU:   ", (100 * busyTicks) / ticks, "%");
+     }
+   }
+ 
+@@ -224,7 +226,8 @@ namespace dxvk::hud {
+       HudElement::StatPipelines,
+       HudElement::StatMemory,
+       HudElement::StatGpuLoad,
+-      HudElement::CompilerActivity);
++      HudElement::CompilerActivity,
++      HudElement::GpuLoad);
+   }
+   
+ }
+diff --git a/src/dxvk/hud/dxvk_hud_stats.h b/src/dxvk/hud/dxvk_hud_stats.h
+index 227f600c..b7d740f1 100644
+--- a/src/dxvk/hud/dxvk_hud_stats.h
++++ b/src/dxvk/hud/dxvk_hud_stats.h
+@@ -7,6 +7,8 @@
+ #include "dxvk_hud_config.h"
+ #include "dxvk_hud_renderer.h"
+ 
++extern std::string m_gpuLoadString;
++extern uint64_t gpuLoad;
+ namespace dxvk::hud {
+   
+   /**
+@@ -44,8 +46,6 @@ namespace dxvk::hud {
+     uint64_t m_prevGpuIdleTicks = 0;
+     uint64_t m_diffGpuIdleTicks = 0;
+     
+-    std::string m_gpuLoadString = "GPU: ";
+-
+     void updateGpuLoad();
+     
+     HudPos printDrawCallStats(
diff --git a/extraopts.patch b/extraopts.patch
new file mode 100644
index 000000000000..43a8fd6d10ec
--- /dev/null
+++ b/extraopts.patch
@@ -0,0 +1,70 @@
+diff --git a/build-win32.txt b/build-win32.txt
+index 0865fc5e..b93a280f 100644
+--- a/build-win32.txt
++++ b/build-win32.txt
+@@ -5,10 +5,10 @@ ar = 'i686-w64-mingw32-ar'
+ strip = 'i686-w64-mingw32-strip'
+ 
+ [properties]
+-c_args=['-msse', '-msse2']
+-cpp_args=['-msse', '-msse2']
+-c_link_args = ['-static', '-static-libgcc']
+-cpp_link_args = ['-static', '-static-libgcc', '-static-libstdc++']
++c_args=['-msse', '-msse2', @CARGS@]
++cpp_args=['-msse', '-msse2', @CARGS@]
++c_link_args = ['-static', '-static-libgcc', @LDARGS@]
++cpp_link_args = ['-static', '-static-libgcc', '-static-libstdc++', @LDARGS@]
+ needs_exe_wrapper = true
+ 
+ [host_machine]
+diff --git a/build-win64.txt b/build-win64.txt
+index 2a7fbee3..9e3bfc2e 100644
+--- a/build-win64.txt
++++ b/build-win64.txt
+@@ -5,8 +5,10 @@ ar = 'x86_64-w64-mingw32-ar'
+ strip = 'x86_64-w64-mingw32-strip'
+ 
+ [properties]
+-c_link_args = ['-static', '-static-libgcc']
+-cpp_link_args = ['-static', '-static-libgcc', '-static-libstdc++']
++c_args=[@CARGS@]
++cpp_args=[@CARGS@]
++c_link_args = ['-static', '-static-libgcc', @LDARGS@]
++cpp_link_args = ['-static', '-static-libgcc', '-static-libstdc++', @LDARGS@]
+ needs_exe_wrapper = true
+ 
+ [host_machine]
+diff --git a/build-wine32.txt b/build-wine32.txt
+index ba8d34a8..ecec2eb4 100644
+--- a/build-wine32.txt
++++ b/build-wine32.txt
+@@ -7,9 +7,9 @@ strip = 'strip'
+ [properties]
+ needs_exe_wrapper = true
+ 
+-c_args=['-m32', '-msse', '-msse2', '-fvisibility=hidden']
+-cpp_args=['-m32', '-msse', '-msse2', '-fvisibility=hidden', '-fvisibility-inlines-hidden', '-D__WIDL_objidl_generated_name_0000000C=']
+-cpp_link_args=['-m32', '-mwindows']
++c_args=['-m32', '-msse', '-msse2', '-fvisibility=hidden', @CARGS@]
++cpp_args=['-m32', '-msse', '-msse2', '-fvisibility=hidden', '-fvisibility-inlines-hidden', '-D__WIDL_objidl_generated_name_0000000C=', @CARGS@]
++cpp_link_args=['-m32', '-mwindows', @LDARGS@]
+ 
+ [host_machine]
+ system = 'linux'
+diff --git a/build-wine64.txt b/build-wine64.txt
+index b3e028bb..fb9d98fd 100644
+--- a/build-wine64.txt
++++ b/build-wine64.txt
+@@ -7,9 +7,9 @@ strip = 'strip'
+ [properties]
+ needs_exe_wrapper = true
+ 
+-c_args=['-m64', '-fvisibility=hidden']
+-cpp_args=['-m64', '-fvisibility=hidden', '-fvisibility-inlines-hidden', '-D__WIDL_objidl_generated_name_0000000C=']
+-cpp_link_args=['-m64', '-mwindows']
++c_args=['-m64', '-fvisibility=hidden', @CARGS@]
++cpp_args=['-m64', '-fvisibility=hidden', '-fvisibility-inlines-hidden', '-D__WIDL_objidl_generated_name_0000000C=', @CARGS@]
++cpp_link_args=['-m64', '-mwindows', @LDARGS@]
+ 
+ [host_machine]
+ system = 'linux'