diff --git a/.gitignore b/.gitignore
index c15ad42a..2584b70d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,7 +1,8 @@
 build
-cmake-build-*
+cmake-build*
 .DS_Store
-
+.vscode
+.vs
 *.swp
 *.kdev4
 
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 33e09423..a85d880d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -4,39 +4,6 @@ project(Sunshine)
 
 set(CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR})
 
-# On MSYS2, building a stand-alone binary that links with ffmpeg is not possible,
-# Therefore, ffmpeg, libx264 and libx265 must be build from source
-if(WIN32)
-	file(
-		DOWNLOAD "https://github.com/TheElixZammuto/sunshine-prebuilt/releases/download/1.0.0/pre-compiled.zip" "${CMAKE_CURRENT_BINARY_DIR}/pre-compiled.zip"
-		TIMEOUT 60
-		EXPECTED_HASH SHA256=5d59986bd7f619eaaf82b2dd56b5127b747c9cbe8db61e3b898ff6b485298ed6)
-
-	file(ARCHIVE_EXTRACT
-		INPUT "${CMAKE_CURRENT_BINARY_DIR}/pre-compiled.zip"
-		DESTINATION ${CMAKE_CURRENT_BINARY_DIR}/pre-compiled)
-	set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -static")
-	
-	if(NOT DEFINED SUNSHINE_PREPARED_BINARIES)
-		set(SUNSHINE_PREPARED_BINARIES "${CMAKE_CURRENT_BINARY_DIR}/pre-compiled/windows")
-	endif()
-
-	set(FFMPEG_INCLUDE_DIRS
-		${SUNSHINE_PREPARED_BINARIES}/include)
-	set(FFMPEG_LIBRARIES
-		${SUNSHINE_PREPARED_BINARIES}/lib/libavcodec.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libavdevice.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libavfilter.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libavformat.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libavutil.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libpostproc.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libswresample.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libswscale.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libx264.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libx265.a
-		${SUNSHINE_PREPARED_BINARIES}/lib/libhdr10plus.a
-		z lzma bcrypt libiconv.a)
-endif()
 add_subdirectory(Simple-Web-Server)
 add_subdirectory(moonlight-common-c/enet)
 
@@ -108,7 +75,7 @@ if(WIN32)
 		wsock32
 		ws2_32
 		iphlpapi
-		d3d11 dxgi
+		d3d11 dxgi D3DCompiler
 		setupapi
 		)
 
diff --git a/assets/shaders/ConvertUVPS.hlsl b/assets/shaders/ConvertUVPS.hlsl
new file mode 100644
index 00000000..2b72cddf
--- /dev/null
+++ b/assets/shaders/ConvertUVPS.hlsl
@@ -0,0 +1,33 @@
+Texture2D image : register(t0);
+
+SamplerState def_sampler : register(s0);
+
+struct FragTexWide {
+	float3 uuv : TEXCOORD0;
+};
+
+cbuffer ColorMatrix : register(b0) {
+	float4 color_vec_y;
+	float4 color_vec_u;
+	float4 color_vec_v;
+	float2 range_y;
+	float2 range_uv;
+};
+
+//--------------------------------------------------------------------------------------
+// Pixel Shader
+//--------------------------------------------------------------------------------------
+float2 main_ps(FragTexWide input) : SV_Target
+{
+	float3 rgb_left = image.Sample(def_sampler, input.uuv.xz).rgb;
+	float3 rgb_right = image.Sample(def_sampler, input.uuv.yz).rgb;
+	float3 rgb = (rgb_left + rgb_right) * 0.5;
+	
+	float u = dot(color_vec_u.xyz, rgb) + color_vec_u.w;
+	float v = dot(color_vec_v.xyz, rgb) + color_vec_v.w;
+
+	u = u * range_uv.x + range_uv.y;
+	v = v * range_uv.x + range_uv.y;
+
+	return float2(u, v * 224.0f/256.0f + 0.0625);
+}
\ No newline at end of file
diff --git a/assets/shaders/ConvertUVVS.hlsl b/assets/shaders/ConvertUVVS.hlsl
new file mode 100644
index 00000000..66d97d43
--- /dev/null
+++ b/assets/shaders/ConvertUVVS.hlsl
@@ -0,0 +1,29 @@
+struct VertTexPosWide {
+	float3 uuv : TEXCOORD;
+	float4 pos : SV_POSITION;
+};
+
+cbuffer info : register(b0) {
+	float width_i;
+};
+
+//--------------------------------------------------------------------------------------
+// Vertex Shader
+//--------------------------------------------------------------------------------------
+VertTexPosWide main_vs(uint vI : SV_VERTEXID)
+{
+	float idHigh = float(vI >> 1);
+	float idLow = float(vI & uint(1));
+
+	float x = idHigh * 4.0 - 1.0;
+	float y = idLow * 4.0 - 1.0;
+
+	float u_right = idHigh * 2.0;
+	float u_left = u_right - width_i;
+	float v = 1.0 - idLow * 2.0;
+
+	VertTexPosWide vert_out;
+	vert_out.uuv = float3(u_left, u_right, v);
+	vert_out.pos = float4(x, y, 0.0, 1.0);
+	return vert_out;
+}
\ No newline at end of file
diff --git a/assets/shaders/ConvertYPS.hlsl b/assets/shaders/ConvertYPS.hlsl
new file mode 100644
index 00000000..386133c8
--- /dev/null
+++ b/assets/shaders/ConvertYPS.hlsl
@@ -0,0 +1,25 @@
+Texture2D image : register(t0);
+
+SamplerState def_sampler : register(s0);
+
+cbuffer ColorMatrix : register(b0) {
+	float4 color_vec_y;
+	float4 color_vec_u;
+	float4 color_vec_v;
+	float2 range_y;
+	float2 range_uv;
+};
+
+struct PS_INPUT
+{
+	float4 pos : SV_POSITION;
+	float2 tex : TEXCOORD;
+};
+
+float main_ps(PS_INPUT frag_in) : SV_Target
+{
+	float3 rgb = image.Sample(def_sampler, frag_in.tex, 0).rgb;
+	float y = dot(color_vec_y.xyz, rgb);
+
+	return y * range_y.x + range_y.y;
+}
\ No newline at end of file
diff --git a/assets/shaders/ScenePS.hlsl b/assets/shaders/ScenePS.hlsl
new file mode 100644
index 00000000..aa601231
--- /dev/null
+++ b/assets/shaders/ScenePS.hlsl
@@ -0,0 +1,14 @@
+Texture2D image : register(t0);
+
+SamplerState def_sampler : register(s0);
+
+struct PS_INPUT
+{
+	float4 pos : SV_POSITION;
+	float2 tex : TEXCOORD;
+};
+
+float4 main_ps(PS_INPUT frag_in) : SV_Target
+{
+	return image.Sample(def_sampler, frag_in.tex, 0);
+}
\ No newline at end of file
diff --git a/assets/shaders/SceneVS.hlsl b/assets/shaders/SceneVS.hlsl
new file mode 100644
index 00000000..51319ddb
--- /dev/null
+++ b/assets/shaders/SceneVS.hlsl
@@ -0,0 +1,22 @@
+struct PS_INPUT
+{
+	float4 pos : SV_POSITION;
+	float2 tex : TEXCOORD;
+};
+
+PS_INPUT main_vs(uint vI : SV_VERTEXID)
+{
+	float idHigh = float(vI >> 1);
+	float idLow = float(vI & uint(1));
+
+	float x = idHigh * 4.0 - 1.0;
+	float y = idLow * 4.0 - 1.0;
+
+	float u = idHigh * 2.0;
+	float v = 1.0 - idLow * 2.0;
+
+	PS_INPUT vert_out;
+	vert_out.pos = float4(x, y, 0.0, 1.0);
+	vert_out.tex = float2(u, v);
+	return vert_out;
+}
\ No newline at end of file
diff --git a/assets/sunshine.conf b/assets/sunshine.conf
index 356acd79..12377eda 100644
--- a/assets/sunshine.conf
+++ b/assets/sunshine.conf
@@ -183,13 +183,11 @@
 # amd_preset = balanced
 #
 ####### rate control #####
-# auto      -- let ffmpeg decide rate control
-# constqp   -- constant QP mode
-# vbr       -- variable bitrate
-# cbr       -- constant bitrate
-# cbr_hq    -- cbr high quality
-# cbr_ld_hq -- cbr low delay high quality
-# vbr_hq    -- vbr high quality
+# auto        -- let ffmpeg decide rate control
+# constqp     -- constant QP mode
+# vbr_latency -- Latency Constrained Variable Bitrate
+# vbr_peak    -- Peak Contrained Variable Bitrate
+# cbr         -- constant bitrate
 ##########################
 # amd_rc = auto
 
diff --git a/sunshine/config.cpp b/sunshine/config.cpp
index 88cc4502..9408578c 100644
--- a/sunshine/config.cpp
+++ b/sunshine/config.cpp
@@ -94,12 +94,10 @@ enum quality_e : int {
 };
 
 enum rc_e : int {
-  constqp   = 0x0,       /**< Constant QP mode */
-  vbr       = 0x1,       /**< Variable bitrate mode */
-  cbr       = 0x2,       /**< Constant bitrate mode */
-  cbr_ld_hq = 0x8,       /**< low-delay CBR, high quality */
-  cbr_hq    = 0x10,      /**< CBR, high quality (slower) */
-  vbr_hq    = 0x20       /**< VBR, high quality (slower) */
+  constqp,     /**< Constant QP mode */
+  vbr_latency, /**< Latency Constrained Variable Bitrate */
+  vbr_peak,    /**< Peak Contrained Variable Bitrate */
+  cbr,         /**< Constant bitrate mode */
 };
 
 enum coder_e : int {
@@ -121,11 +119,9 @@ std::optional<quality_e> quality_from_view(const std::string_view &quality) {
 std::optional<rc_e> rc_from_view(const std::string_view &rc) {
 #define _CONVERT_(x) if(rc == #x##sv) return x
   _CONVERT_(constqp);
-  _CONVERT_(vbr);
+  _CONVERT_(vbr_latency);
+  _CONVERT_(vbr_peak);
   _CONVERT_(cbr);
-  _CONVERT_(cbr_hq);
-  _CONVERT_(vbr_hq);
-  _CONVERT_(cbr_ld_hq);
 #undef _CONVERT_
   return std::nullopt;
 }
diff --git a/sunshine/crypto.cpp b/sunshine/crypto.cpp
index 70a98289..398c8e09 100644
--- a/sunshine/crypto.cpp
+++ b/sunshine/crypto.cpp
@@ -187,10 +187,10 @@ x509_t x509(const std::string_view &x) {
 
   BIO_write(io.get(), x.data(), x.size());
 
-  X509 *p = nullptr;
+  x509_t p;
   PEM_read_bio_X509(io.get(), &p, nullptr, nullptr);
 
-  return x509_t { p };
+  return p;
 }
 
 pkey_t pkey(const std::string_view &k) {
@@ -198,10 +198,10 @@ pkey_t pkey(const std::string_view &k) {
 
   BIO_write(io.get(), k.data(), k.size());
 
-  EVP_PKEY *p = nullptr;
+  pkey_t p = nullptr;
   PEM_read_bio_PrivateKey(io.get(), &p, nullptr, nullptr);
 
-  return pkey_t { p };
+  return p;
 }
 
 std::string pem(x509_t &x509) {
diff --git a/sunshine/input.cpp b/sunshine/input.cpp
index 402d4ce3..9a41dd03 100644
--- a/sunshine/input.cpp
+++ b/sunshine/input.cpp
@@ -402,11 +402,18 @@ void passthrough(std::shared_ptr<input_t> &input, std::vector<std::uint8_t> &&in
   task_pool.push(passthrough_helper, input, util::cmove(input_data));
 }
 
-void reset(){
-    for(auto& kp : key_press){
+void reset() {
+  if(task_id) {
+    task_pool.cancel(task_id);
+  }
+
+  // Ensure input is synchronous
+  task_pool.push([]() {
+    for(auto& kp : key_press) {
       platf::keyboard(platf_input, kp.first & 0x00FF, true);
       key_press[kp.first] = false;
     }
+  });
 }
 
 void init() {
diff --git a/sunshine/main.cpp b/sunshine/main.cpp
index 2938e69f..15248deb 100644
--- a/sunshine/main.cpp
+++ b/sunshine/main.cpp
@@ -128,6 +128,10 @@ int main(int argc, char *argv[]) {
   proc::refresh(config::stream.file_apps);
   
   auto deinit_guard = platf::init();
+  if(!deinit_guard) {
+    return 4;
+  }
+
   input::init();
   reed_solomon_init();
   if(video::init()) {
@@ -141,6 +145,8 @@ int main(int argc, char *argv[]) {
   stream::rtpThread(shutdown_event);
 
   httpThread.join();
+  task_pool.stop();
+  task_pool.join();
 
   return 0;
 }
diff --git a/sunshine/nvhttp.cpp b/sunshine/nvhttp.cpp
index da402706..98f5a79d 100644
--- a/sunshine/nvhttp.cpp
+++ b/sunshine/nvhttp.cpp
@@ -846,8 +846,22 @@ void start(std::shared_ptr<safe::signal_t> shutdown_event) {
     return;
   }
 
-  std::thread ssl { &https_server_t::accept_and_run, &https_server };
-  std::thread tcp { &http_server_t::accept_and_run, &http_server };
+  auto accept_and_run = [&](auto *http_server) {
+    try {
+      http_server->accept_and_run();
+    } catch(boost::system::system_error &err) {
+      // It's possible the exception gets thrown after calling http_server->stop() from a different thread
+      if(shutdown_event->peek()) {
+        return;
+      }
+
+      BOOST_LOG(fatal) << "Couldn't start http server to ports ["sv << PORT_HTTPS << ", "sv << PORT_HTTP << "]: "sv << err.what();
+      shutdown_event->raise(true);
+      return;
+    }
+  };
+  std::thread ssl { accept_and_run, &https_server };
+  std::thread tcp { accept_and_run, &http_server };
 
   // Wait for any event
   shutdown_event->view();
diff --git a/sunshine/platform/common.h b/sunshine/platform/common.h
index 140b54b1..fa6b3be7 100644
--- a/sunshine/platform/common.h
+++ b/sunshine/platform/common.h
@@ -43,6 +43,21 @@ enum class pix_fmt_e {
   unknown
 };
 
+inline std::string_view from_pix_fmt(pix_fmt_e pix_fmt) {
+using namespace std::literals;
+#define _CONVERT(x) case pix_fmt_e:: x : return  #x ## sv
+  switch(pix_fmt) {
+    _CONVERT(yuv420p);
+    _CONVERT(yuv420p10);
+    _CONVERT(nv12);
+    _CONVERT(p010);
+    _CONVERT(unknown);
+  }
+#undef _CONVERT
+
+  return "unknown"sv;
+}
+
 struct gamepad_state_t {
   std::uint16_t buttonFlags;
   std::uint8_t lt;
diff --git a/sunshine/platform/linux/input.cpp b/sunshine/platform/linux/input.cpp
index f5696522..6bb5d495 100644
--- a/sunshine/platform/linux/input.cpp
+++ b/sunshine/platform/linux/input.cpp
@@ -502,5 +502,5 @@ void freeInput(void *p) {
   delete input;
 }
 
-std::unique_ptr<deinit_t> init() { return nullptr; }
+std::unique_ptr<deinit_t> init() { return std::make_unique<deinit_t>(); }
 }
diff --git a/sunshine/platform/windows/audio.cpp b/sunshine/platform/windows/audio.cpp
index a71e4a77..63369ccb 100644
--- a/sunshine/platform/windows/audio.cpp
+++ b/sunshine/platform/windows/audio.cpp
@@ -81,50 +81,43 @@ public:
 
     HRESULT status;
 
-    device_enum_t::pointer device_enum_p{};
     status = CoCreateInstance(
       CLSID_MMDeviceEnumerator,
       nullptr,
       CLSCTX_ALL,
       IID_IMMDeviceEnumerator,
-      (void **) &device_enum_p);
-    device_enum.reset(device_enum_p);
+      (void **) &device_enum);
 
-    if (FAILED(status)) {
+    if(FAILED(status)) {
       BOOST_LOG(error) << "Couldn't create Device Enumerator [0x"sv << util::hex(status).to_string_view() << ']';
 
       return -1;
     }
 
-    device_t::pointer device_p{};
-
     if(config::audio.sink.empty()) {
       status = device_enum->GetDefaultAudioEndpoint(
         eRender,
         eConsole,
-        &device_p);
+        &device);
     }
     else {
       std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>, wchar_t> converter;
       auto wstring_device_id = converter.from_bytes(config::audio.sink);
 
-      status = device_enum->GetDevice(wstring_device_id.c_str(), &device_p);
+      status = device_enum->GetDevice(wstring_device_id.c_str(), &device);
     }
-    device.reset(device_p);
 
-    if (FAILED(status)) {
+    if(FAILED(status)) {
       BOOST_LOG(error) << "Couldn't create audio Device [0x"sv << util::hex(status).to_string_view() << ']';
 
       return -1;
     }
 
-    audio_client_t::pointer audio_client_p{};
     status = device->Activate(
       IID_IAudioClient,
       CLSCTX_ALL,
       nullptr,
-      (void **) &audio_client_p);
-    audio_client.reset(audio_client_p);
+      (void **) &audio_client);
 
     if (FAILED(status)) {
       BOOST_LOG(error) << "Couldn't activate audio Device [0x"sv << util::hex(status).to_string_view() << ']';
@@ -132,11 +125,8 @@ public:
       return -1;
     }
 
-    wave_format_t::pointer wave_format_p{};
-    status = audio_client->GetMixFormat(&wave_format_p);
-    wave_format.reset(wave_format_p);
-
-    if (FAILED(status)) {
+    status = audio_client->GetMixFormat(&wave_format);
+    if(FAILED(status)) {
       BOOST_LOG(error) << "Couldn't acquire Wave Format [0x"sv << util::hex(status).to_string_view() << ']';
 
       return -1;
@@ -198,9 +188,7 @@ public:
     sample_buf = util::buffer_t<std::int16_t> { frames };
     sample_buf_pos = std::begin(sample_buf);
 
-    audio_capture_t::pointer audio_capture_p {};
-    status = audio_client->GetService(IID_IAudioCaptureClient, (void**)&audio_capture_p);
-    audio_capture.reset(audio_capture_p);
+    status = audio_client->GetService(IID_IAudioCaptureClient, (void**)&audio_capture);
     if (FAILED(status)) {
       BOOST_LOG(error) << "Couldn't initialize audio capture client [0x"sv << util::hex(status).to_string_view() << ']';
 
@@ -319,6 +307,12 @@ public:
 }
 
 namespace platf {
+
+// It's not big enough to justify it's own source file :/
+namespace dxgi {
+int init();
+}
+
 std::unique_ptr<mic_t> microphone(std::uint32_t sample_rate) {
   auto mic = std::make_unique<audio::mic_wasapi_t>();
 
@@ -330,6 +324,9 @@ std::unique_ptr<mic_t> microphone(std::uint32_t sample_rate) {
 }
 
 std::unique_ptr<deinit_t> init() {
+  if(dxgi::init()) {
+    return nullptr;
+  }
   return std::make_unique<platf::audio::co_init_t>();
 }
 }
diff --git a/sunshine/platform/windows/display.h b/sunshine/platform/windows/display.h
index 039e6a75..8bc7ae9e 100644
--- a/sunshine/platform/windows/display.h
+++ b/sunshine/platform/windows/display.h
@@ -32,6 +32,7 @@ using output_t      = util::safe_ptr<IDXGIOutput, Release<IDXGIOutput>>;
 using output1_t     = util::safe_ptr<IDXGIOutput1, Release<IDXGIOutput1>>;
 using dup_t         = util::safe_ptr<IDXGIOutputDuplication, Release<IDXGIOutputDuplication>>;
 using texture2d_t   = util::safe_ptr<ID3D11Texture2D, Release<ID3D11Texture2D>>;
+using texture1d_t   = util::safe_ptr<ID3D11Texture1D, Release<ID3D11Texture1D>>;
 using resource_t    = util::safe_ptr<IDXGIResource, Release<IDXGIResource>>;
 using multithread_t = util::safe_ptr<ID3D11Multithread, Release<ID3D11Multithread>>;
 
diff --git a/sunshine/platform/windows/display_base.cpp b/sunshine/platform/windows/display_base.cpp
index 29b53523..8644ed88 100644
--- a/sunshine/platform/windows/display_base.cpp
+++ b/sunshine/platform/windows/display_base.cpp
@@ -90,17 +90,10 @@ int display_base_t::init() {
     FreeLibrary(user32);
   });
 */
-  
-  dxgi::factory1_t::pointer   factory_p {};
-  dxgi::adapter_t::pointer    adapter_p {};
-  dxgi::output_t::pointer     output_p {};
-  dxgi::device_t::pointer     device_p {};
-  dxgi::device_ctx_t::pointer device_ctx_p {};
 
   HRESULT status;
 
-  status = CreateDXGIFactory1(IID_IDXGIFactory1, (void**)&factory_p);
-  factory.reset(factory_p);
+  status = CreateDXGIFactory1(IID_IDXGIFactory1, (void**)&factory);
   if(FAILED(status)) {
     BOOST_LOG(error) << "Failed to create DXGIFactory1 [0x"sv << util::hex(status).to_string_view() << ']';
     return -1;
@@ -111,7 +104,8 @@ int display_base_t::init() {
   auto adapter_name = converter.from_bytes(config::video.adapter_name);
   auto output_name = converter.from_bytes(config::video.output_name);
 
-  for(int x = 0; factory_p->EnumAdapters1(x, &adapter_p) != DXGI_ERROR_NOT_FOUND; ++x) {
+  adapter_t::pointer adapter_p;
+  for(int x = 0; factory->EnumAdapters1(x, &adapter_p) != DXGI_ERROR_NOT_FOUND; ++x) {
     dxgi::adapter_t adapter_tmp { adapter_p };
 
     DXGI_ADAPTER_DESC1 adapter_desc;
@@ -121,8 +115,9 @@ int display_base_t::init() {
       continue;
     }
 
+    dxgi::output_t::pointer output_p;
     for(int y = 0; adapter_tmp->EnumOutputs(y, &output_p) != DXGI_ERROR_NOT_FOUND; ++y) {
-      dxgi::output_t output_tmp {output_p };
+      dxgi::output_t output_tmp { output_p };
 
       DXGI_OUTPUT_DESC desc;
       output_tmp->GetDesc(&desc);
@@ -173,14 +168,12 @@ int display_base_t::init() {
     D3D11_CREATE_DEVICE_VIDEO_SUPPORT,
     featureLevels, sizeof(featureLevels) / sizeof(D3D_FEATURE_LEVEL),
     D3D11_SDK_VERSION,
-    &device_p,
+    &device,
     &feature_level,
-    &device_ctx_p);
+    &device_ctx);
 
   adapter_p->Release();
 
-  device.reset(device_p);
-  device_ctx.reset(device_ctx_p);
   if(FAILED(status)) {
     BOOST_LOG(error) << "Failed to create D3D11 device [0x"sv << util::hex(status).to_string_view() << ']';
 
@@ -216,7 +209,7 @@ int display_base_t::init() {
       tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
 
       if (!AdjustTokenPrivileges(token, false, &tp, sizeof(tp), NULL, NULL)) {
-        BOOST_LOG(error) << "Could not set privilege to increase GPU priority";
+        BOOST_LOG(warning) << "Could not set privilege to increase GPU priority";
       }
     }
 
@@ -229,17 +222,15 @@ int display_base_t::init() {
       if (fn) {
         status = fn(GetCurrentProcess(), D3DKMT_SCHEDULINGPRIORITYCLASS_REALTIME);
         if (FAILED(status)) {
-          BOOST_LOG(error) << "Failed to set realtime GPU priority. Please run application as administrator for optimal performance.";
+          BOOST_LOG(warning) << "Failed to set realtime GPU priority. Please run application as administrator for optimal performance.";
         }
       }
     }
 
-    dxgi::dxgi_t::pointer dxgi_p {};
-    status = device->QueryInterface(IID_IDXGIDevice, (void**)&dxgi_p);
-    dxgi::dxgi_t dxgi { dxgi_p };
-
+    dxgi::dxgi_t dxgi;
+    status = device->QueryInterface(IID_IDXGIDevice, (void**)&dxgi);
     if(FAILED(status)) {
-      BOOST_LOG(error) << "Failed to query DXGI interface from device [0x"sv << util::hex(status).to_string_view() << ']';
+      BOOST_LOG(warning) << "Failed to query DXGI interface from device [0x"sv << util::hex(status).to_string_view() << ']';
       return -1;
     }
 
@@ -248,25 +239,24 @@ int display_base_t::init() {
 
   // Try to reduce latency
   {
-    dxgi::dxgi1_t::pointer dxgi_p {};
-    status = device->QueryInterface(IID_IDXGIDevice, (void**)&dxgi_p);
-    dxgi::dxgi1_t dxgi { dxgi_p };
-
+    dxgi::dxgi1_t dxgi {};
+    status = device->QueryInterface(IID_IDXGIDevice, (void**)&dxgi);
     if(FAILED(status)) {
       BOOST_LOG(error) << "Failed to query DXGI interface from device [0x"sv << util::hex(status).to_string_view() << ']';
       return -1;
     }
 
-    dxgi->SetMaximumFrameLatency(1);
+    status = dxgi->SetMaximumFrameLatency(1);
+    if(FAILED(status)) {
+      BOOST_LOG(warning) << "Failed to set maximum frame latency [0x"sv << util::hex(status).to_string_view() << ']';
+    }
   }
 
   //FIXME: Duplicate output on RX580 in combination with DOOM (2016) --> BSOD
   //TODO: Use IDXGIOutput5 for improved performance
   {
-    dxgi::output1_t::pointer output1_p {};
-    status = output->QueryInterface(IID_IDXGIOutput1, (void**)&output1_p);
-    dxgi::output1_t output1 {output1_p };
-
+    dxgi::output1_t output1 {};
+    status = output->QueryInterface(IID_IDXGIOutput1, (void**)&output1);
     if(FAILED(status)) {
       BOOST_LOG(error) << "Failed to query IDXGIOutput1 from the output"sv;
       return -1;
@@ -274,10 +264,8 @@ int display_base_t::init() {
 
     // We try this twice, in case we still get an error on reinitialization
     for(int x = 0; x < 2; ++x) {
-      dxgi::dup_t::pointer dup_p {};
-      status = output1->DuplicateOutput((IUnknown*)device.get(), &dup_p);
+      status = output1->DuplicateOutput((IUnknown*)device.get(), &dup.dup);
       if(SUCCEEDED(status)) {
-        dup.reset(dup_p);
         break;
       }
       std::this_thread::sleep_for(200ms);
diff --git a/sunshine/platform/windows/display_ram.cpp b/sunshine/platform/windows/display_ram.cpp
index 15d608fe..ffcbbd25 100644
--- a/sunshine/platform/windows/display_ram.cpp
+++ b/sunshine/platform/windows/display_ram.cpp
@@ -203,9 +203,8 @@ capture_e display_ram_t::snapshot(::platf::img_t *img_base, std::chrono::millise
   // If frame has been updated
   if (frame_info.LastPresentTime.QuadPart != 0) {
     {
-      texture2d_t::pointer src_p {};
-      status = res->QueryInterface(IID_ID3D11Texture2D, (void **)&src_p);
-      texture2d_t src{src_p};
+      texture2d_t src {};
+      status = res->QueryInterface(IID_ID3D11Texture2D, (void **)&src);
 
       if (FAILED(status)) {
         BOOST_LOG(error) << "Couldn't query interface [0x"sv << util::hex(status).to_string_view() << ']';
@@ -279,10 +278,7 @@ int display_ram_t::init() {
   t.Format = format;
   t.CPUAccessFlags = D3D11_CPU_ACCESS_READ;
 
-  dxgi::texture2d_t::pointer tex_p {};
-  auto status = device->CreateTexture2D(&t, nullptr, &tex_p);
-
-  texture.reset(tex_p);
+  auto status = device->CreateTexture2D(&t, nullptr, &texture);
 
   if(FAILED(status)) {
     BOOST_LOG(error) << "Failed to create texture [0x"sv << util::hex(status).to_string_view() << ']';
diff --git a/sunshine/platform/windows/display_vram.cpp b/sunshine/platform/windows/display_vram.cpp
index 0d87b40c..00e0c69d 100644
--- a/sunshine/platform/windows/display_vram.cpp
+++ b/sunshine/platform/windows/display_vram.cpp
@@ -1,14 +1,131 @@
+#include <codecvt>
+
+#include <d3dcompiler.h>
+#include <directxmath.h>
+
 #include "sunshine/main.h"
 #include "display.h"
 
+#define SUNSHINE_SHADERS_DIR SUNSHINE_ASSETS_DIR "/shaders"
 namespace platf {
 using namespace std::literals;
 }
 
 namespace platf::dxgi {
+constexpr float aquamarine[] { 0.498039246f, 1.000000000f, 0.831372619f, 1.000000000f };
+
+using input_layout_t        = util::safe_ptr<ID3D11InputLayout, Release<ID3D11InputLayout>>;
+using render_target_t       = util::safe_ptr<ID3D11RenderTargetView, Release<ID3D11RenderTargetView>>;
+using shader_res_t          = util::safe_ptr<ID3D11ShaderResourceView, Release<ID3D11ShaderResourceView>>;
+using buf_t                 = util::safe_ptr<ID3D11Buffer, Release<ID3D11Buffer>>;
+using blend_t               = util::safe_ptr<ID3D11BlendState, Release<ID3D11BlendState>>;
+using raster_state_t        = util::safe_ptr<ID3D11RasterizerState, Release<ID3D11RasterizerState>>;
+using sampler_state_t       = util::safe_ptr<ID3D11SamplerState, Release<ID3D11SamplerState>>;
+using vs_t                  = util::safe_ptr<ID3D11VertexShader, Release<ID3D11VertexShader>>;
+using ps_t                  = util::safe_ptr<ID3D11PixelShader, Release<ID3D11PixelShader>>;
+using blob_t                = util::safe_ptr<ID3DBlob, Release<ID3DBlob>>;
+using depth_stencil_state_t = util::safe_ptr<ID3D11DepthStencilState, Release<ID3D11DepthStencilState>>;
+using depth_stencil_view_t  = util::safe_ptr<ID3D11DepthStencilView, Release<ID3D11DepthStencilView>>;
+
+using float4 = DirectX::XMFLOAT4;
+using float3 = DirectX::XMFLOAT3;
+using float2 = DirectX::XMFLOAT2;
+struct __attribute__ ((__aligned__ (16))) color_t {
+  float4 color_vec_y;
+  float4 color_vec_u;
+  float4 color_vec_v;
+  float2 range_y;
+  float2 range_uv;
+};
+
+color_t make_color_matrix(float Cr, float Cb, float U_max, float V_max, float add_Y, float add_UV, float2 range_Y, float2 range_UV) {
+  float Cg = 1.0f - Cr - Cb;
+
+  float Cr_i = 1.0f - Cr;
+  float Cb_i = 1.0f - Cb;
+
+  float shift_y = range_Y.x / 256.0f;
+  float shift_uv = range_UV.x / 256.0f;
+
+  float scale_y = (range_Y.y - range_Y.x) / 256.0f;
+  float scale_uv = (range_UV.y - range_UV.x) / 256.0f;
+  return {
+    { Cr, Cg, Cb, add_Y },
+    { -(Cr * U_max / Cb_i), -(Cg * U_max / Cb_i), U_max, add_UV },
+    { V_max, -(Cg * V_max / Cr_i), -(Cb * V_max / Cr_i), add_UV },
+    { scale_y, shift_y },
+    { scale_uv, shift_uv },
+  };
+}
+
+color_t colors[] {
+  make_color_matrix(0.299f, 0.114f, 0.436f, 0.615f, 0.0625, 0.5f, { 16.0f, 235.0f }, { 16.0f, 240.0f }),   // BT601 MPEG
+  make_color_matrix(0.299f, 0.114f, 0.5f, 0.5f, 0.0f, 0.5f, { 0.0f, 255.0f }, { 0.0f, 255.0f }),           // BT601 JPEG
+  make_color_matrix(0.2126f, 0.0722f, 0.436f, 0.615f, 0.0625, 0.5f, { 16.0f, 235.0f }, { 16.0f, 240.0f }), //BT701 MPEG
+  make_color_matrix(0.2126f, 0.0722f, 0.5f, 0.5f, 0.0f, 0.5f, { 0.0f, 255.0f }, { 0.0f, 255.0f }),         //BT701 JPEG
+};
+
+template<class T>
+buf_t make_buffer(device_t::pointer device, const T& t) {
+  static_assert(sizeof(T) % 16 == 0, "Buffer needs to be aligned on a 16-byte alignment");
+
+  D3D11_BUFFER_DESC buffer_desc {
+    sizeof(T),
+    D3D11_USAGE_IMMUTABLE,
+    D3D11_BIND_CONSTANT_BUFFER
+  };
+
+  D3D11_SUBRESOURCE_DATA init_data {
+    &t
+  };
+
+  buf_t::pointer buf_p;
+  auto status = device->CreateBuffer(&buffer_desc, &init_data, &buf_p);
+  if(status) {
+    BOOST_LOG(error) << "Failed to create buffer: [0x"sv << util::hex(status).to_string_view() << ']';
+    return nullptr;
+  }
+
+  return buf_t { buf_p };
+}
+
+blend_t make_blend(device_t::pointer device, bool enable) {
+  D3D11_BLEND_DESC bdesc {};
+  auto &rt = bdesc.RenderTarget[0];
+  rt.BlendEnable = enable;
+  rt.RenderTargetWriteMask = D3D11_COLOR_WRITE_ENABLE_ALL;
+
+  if(enable) {
+    rt.BlendOp = D3D11_BLEND_OP_ADD;
+    rt.BlendOpAlpha = D3D11_BLEND_OP_ADD;
+
+    rt.SrcBlend = D3D11_BLEND_SRC_ALPHA;
+    rt.DestBlend = D3D11_BLEND_INV_SRC_ALPHA;
+
+    rt.SrcBlendAlpha = D3D11_BLEND_ZERO;
+    rt.DestBlendAlpha = D3D11_BLEND_ZERO;
+  }
+
+  blend_t blend;
+  auto status = device->CreateBlendState(&bdesc, &blend);
+  if(status) {
+    BOOST_LOG(error) << "Failed to create blend state: [0x"sv << util::hex(status).to_string_view() << ']';
+    return nullptr;
+  }
+
+  return blend;
+}
+
+blob_t convert_UV_vs_hlsl;
+blob_t convert_UV_ps_hlsl;
+blob_t scene_vs_hlsl;
+blob_t convert_Y_ps_hlsl;
+blob_t scene_ps_hlsl;
+
 struct img_d3d_t : public platf::img_t {
-  std::shared_ptr<platf::display_t> display;
+  shader_res_t input_res;
   texture2d_t texture;
+  std::shared_ptr<platf::display_t> display;
 
   ~img_d3d_t() override = default;
 };
@@ -97,9 +214,43 @@ util::buffer_t<std::uint8_t> make_cursor_image(util::buffer_t<std::uint8_t> &&im
   return cursor_img;
 }
 
+blob_t compile_shader(LPCSTR file, LPCSTR entrypoint, LPCSTR shader_model) {
+  blob_t::pointer msg_p = nullptr;
+  blob_t::pointer compiled_p;
+
+  DWORD flags = D3DCOMPILE_ENABLE_STRICTNESS;
+
+#ifndef NDEBUG
+  flags |= D3DCOMPILE_DEBUG | D3DCOMPILE_SKIP_OPTIMIZATION;
+#endif
+  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>, wchar_t> converter;
+
+  auto wFile = converter.from_bytes(file);
+  auto status = D3DCompileFromFile(wFile.c_str(), nullptr, nullptr, entrypoint, shader_model, flags, 0, &compiled_p, &msg_p);
+
+  if(msg_p) {
+    BOOST_LOG(warning) << std::string_view { (const char *)msg_p->GetBufferPointer(), msg_p->GetBufferSize() - 1 };
+    msg_p->Release();
+  }
+
+  if(status) {
+    BOOST_LOG(error) << "Couldn't compile ["sv << file << "] [0x"sv << util::hex(status).to_string_view() << ']';
+    return nullptr;
+  }
+
+  return blob_t { compiled_p };
+}
+
+blob_t compile_pixel_shader(LPCSTR file) {
+  return compile_shader(file, "main_ps", "ps_5_0");
+}
+
+blob_t compile_vertex_shader(LPCSTR file) {
+  return compile_shader(file, "main_vs", "vs_5_0");
+}
+
 class hwdevice_t : public platf::hwdevice_t {
 public:
-
   hwdevice_t(std::vector<hwdevice_t*> *hwdevices_p) : hwdevices_p { hwdevices_p } {}
   hwdevice_t() = delete;
 
@@ -110,81 +261,118 @@ public:
       return;
     }
 
-    LONG x = ((double)rel_x) * out_width / (double)in_width;
-    LONG y = ((double)rel_y) * out_height / (double)in_height;
+    auto x = ((float)rel_x);
+    auto y = ((float)rel_y);
 
-    // Ensure it's within bounds
-    auto left_out   = std::min<LONG>(out_width, std::max<LONG>(0, x));
-    auto top_out    = std::min<LONG>(out_height, std::max<LONG>(0, y));
-    auto right_out  = std::max<LONG>(0, std::min<LONG>(out_width, x + cursor_scaled_width));
-    auto bottom_out = std::max<LONG>(0, std::min<LONG>(out_height, y + cursor_scaled_height));
-
-    auto left_in   = std::max<LONG>(0, -rel_x);
-    auto top_in    = std::max<LONG>(0, -rel_y);
-    auto right_in  = std::min<LONG>(in_width - rel_x, cursor_width);
-    auto bottom_in = std::min<LONG>(in_height - rel_y, cursor_height);
-
-    RECT rect_in { left_in, top_in, right_in, bottom_in };
-    RECT rect_out { left_out, top_out, right_out, bottom_out };
-
-    ctx->VideoProcessorSetStreamSourceRect(processor.get(), 1, TRUE, &rect_in);
-    ctx->VideoProcessorSetStreamDestRect(processor.get(), 1, TRUE, &rect_out);
+    cursor_view.TopLeftX = x;
+    cursor_view.TopLeftY = y;
   }
 
   int set_cursor_texture(texture2d_t::pointer texture, LONG width, LONG height) {
-    D3D11_VIDEO_PROCESSOR_INPUT_VIEW_DESC input_desc = { 0, (D3D11_VPIV_DIMENSION)D3D11_VPIV_DIMENSION_TEXTURE2D, { 0, 0 } };
+    auto device = (device_t::pointer)data;
 
-    video::processor_in_t::pointer processor_in_p;
-    auto status = device->CreateVideoProcessorInputView(texture, processor_e.get(), &input_desc, &processor_in_p);
+    cursor_view.Width = width;
+    cursor_view.Height = height;
+
+    D3D11_SHADER_RESOURCE_VIEW_DESC desc {
+      DXGI_FORMAT_B8G8R8A8_UNORM,
+      D3D11_SRV_DIMENSION_TEXTURE2D
+    };
+    desc.Texture2D.MipLevels = 1;
+
+    auto status = device->CreateShaderResourceView(texture, &desc, &img.input_res);
     if(FAILED(status)) {
-      BOOST_LOG(error) << "Failed to create cursor VideoProcessorInputView [0x"sv << util::hex(status).to_string_view() << ']';
+      BOOST_LOG(error) << "Failed to create cursor shader resource view [0x"sv << util::hex(status).to_string_view() << ']';
       return -1;
     }
 
-    cursor_in.reset(processor_in_p);
-
-    cursor_width  = width;
-    cursor_height = height;
-    cursor_scaled_width = ((double)width) / in_width * out_width;
-    cursor_scaled_height = ((double)height) / in_height * out_height;
-
     return 0;
   }
 
   int convert(platf::img_t &img_base) override {
     auto &img = (img_d3d_t&)img_base;
 
-    auto it = texture_to_processor_in.find(img.texture.get());
-    if(it == std::end(texture_to_processor_in)) {
-      D3D11_VIDEO_PROCESSOR_INPUT_VIEW_DESC input_desc = { 0, (D3D11_VPIV_DIMENSION)D3D11_VPIV_DIMENSION_TEXTURE2D, { 0, 0 } };
+    if(!img.input_res) {
+      auto device = (device_t::pointer)data;
 
-      video::processor_in_t::pointer processor_in_p;
-      auto status = device->CreateVideoProcessorInputView(img.texture.get(), processor_e.get(), &input_desc, &processor_in_p);
+      D3D11_SHADER_RESOURCE_VIEW_DESC desc {
+        DXGI_FORMAT_B8G8R8A8_UNORM,
+        D3D11_SRV_DIMENSION_TEXTURE2D
+      };
+      desc.Texture2D.MipLevels = 1;
+
+      auto status = device->CreateShaderResourceView(img.texture.get(), &desc, &img.input_res);
       if(FAILED(status)) {
-        BOOST_LOG(error) << "Failed to create VideoProcessorInputView [0x"sv << util::hex(status).to_string_view() << ']';
+        BOOST_LOG(error) << "Failed to create input shader resource view [0x"sv << util::hex(status).to_string_view() << ']';
         return -1;
       }
-      it = texture_to_processor_in.emplace(img.texture.get(), processor_in_p).first;
     }
-    auto &processor_in = it->second;
 
-    D3D11_VIDEO_PROCESSOR_STREAM stream[] {
-      { TRUE, 0, 0, 0, 0, nullptr, processor_in.get() },
-      { TRUE, 0, 0, 0, 0, nullptr, cursor_in.get() }
-    };
+    auto input_res_p = img.input_res.get();
 
-    auto status = ctx->VideoProcessorBlt(processor.get(), processor_out.get(), 0, cursor_visible ? 2 : 1, stream);
-    if(FAILED(status)) {
-      BOOST_LOG(error) << "Failed size and color conversion [0x"sv << util::hex(status).to_string_view() << ']';
-      return -1;
+    if(cursor_visible) {
+      _init_view_port(img.width, img.height);
+
+      device_ctx_p->OMSetRenderTargets(1, &scene_rt, nullptr);
+      device_ctx_p->VSSetShader(scene_vs.get(), nullptr, 0);
+      device_ctx_p->PSSetShader(scene_ps.get(), nullptr, 0);
+      device_ctx_p->PSSetShaderResources(0, 1, &input_res_p);
+
+      device_ctx_p->Draw(3, 0);
+
+      device_ctx_p->OMSetBlendState(blend_enable.get(), nullptr, 0xFFFFFFFFu);
+      device_ctx_p->RSSetViewports(1, &cursor_view);
+      device_ctx_p->PSSetShaderResources(0, 1, &this->img.input_res);
+      device_ctx_p->Draw(3, 0);
+      device_ctx_p->OMSetBlendState(blend_disable.get(), nullptr, 0xFFFFFFFFu);
+
+      input_res_p = scene_sr.get();
     }
 
+    _init_view_port(out_width, out_height);
+    device_ctx_p->OMSetRenderTargets(1, &nv12_Y_rt, nullptr);
+    device_ctx_p->VSSetShader(scene_vs.get(), nullptr, 0);
+    device_ctx_p->PSSetShader(convert_Y_ps.get(), nullptr, 0);
+    device_ctx_p->PSSetShaderResources(0, 1, &input_res_p);
+    device_ctx_p->Draw(3, 0);
+
+    _init_view_port(out_width / 2, out_height / 2);
+    device_ctx_p->OMSetRenderTargets(1, &nv12_UV_rt, nullptr);
+    device_ctx_p->VSSetShader(convert_UV_vs.get(), nullptr, 0);
+    device_ctx_p->PSSetShader(convert_UV_ps.get(), nullptr, 0);
+    device_ctx_p->PSSetShaderResources(0, 1, &input_res_p);
+    device_ctx_p->Draw(3, 0);
+
     return 0;
   }
 
   void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) override {
-    colorspace |= (color_range >> 4);
-    ctx->VideoProcessorSetOutputColorSpace(processor.get(), (D3D11_VIDEO_PROCESSOR_COLOR_SPACE*)&colorspace);
+    switch (colorspace) {
+      case 5: // SWS_CS_SMPTE170M
+        color_p = &colors[0];
+        break;
+      case 1: // SWS_CS_ITU709
+        color_p = &colors[2];
+        break;
+      case 9: // SWS_CS_BT2020
+      default:
+        BOOST_LOG(warning) << "Colorspace: ["sv << colorspace << "] not yet supported: switching to default"sv;
+        color_p = &colors[0];
+    };
+
+    if(color_range > 1) {
+      // Full range
+      ++color_p;
+    }
+
+    auto color_matrix = make_buffer((device_t::pointer)data, *color_p);
+    if(!color_matrix) {
+      BOOST_LOG(warning) << "Failed to create color matrix"sv;
+      return;
+    }
+
+    device_ctx_p->PSSetConstantBuffers(0, 1, &color_matrix);
+    this->color_matrix = std::move(color_matrix);
   }
 
   int init(
@@ -194,62 +382,82 @@ public:
   ) {
     HRESULT status;
 
+    device_p->AddRef();
+    data = device_p;
+
+    this->device_ctx_p = device_ctx_p;
+
     cursor_visible = false;
+    cursor_view.MinDepth = 0.0f;
+    cursor_view.MaxDepth = 1.0f;
 
     platf::hwdevice_t::img = &img;
 
     this->out_width  = out_width;
     this->out_height = out_height;
-    this->in_width   = in_width;
-    this->in_height  = in_height;
 
-    video::device_t::pointer vdevice_p;
-    status = device_p->QueryInterface(IID_ID3D11VideoDevice, (void**)&vdevice_p);
-    if(FAILED(status)) {
-      BOOST_LOG(error) << "Failed to query ID3D11VideoDevice interface [0x"sv << util::hex(status).to_string_view() << ']';
+    status = device_p->CreateVertexShader(scene_vs_hlsl->GetBufferPointer(), scene_vs_hlsl->GetBufferSize(), nullptr, &scene_vs);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create scene vertex shader [0x"sv << util::hex(status).to_string_view() << ']';
       return -1;
     }
-    device.reset(vdevice_p);
 
-    video::ctx_t::pointer ctx_p;
-    status = device_ctx_p->QueryInterface(IID_ID3D11VideoContext, (void**)&ctx_p);
-    if(FAILED(status)) {
-      BOOST_LOG(error) << "Failed to query ID3D11VideoContext interface [0x"sv << util::hex(status).to_string_view() << ']';
+    status = device_p->CreatePixelShader(convert_Y_ps_hlsl->GetBufferPointer(), convert_Y_ps_hlsl->GetBufferSize(), nullptr, &convert_Y_ps);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create convertY pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
       return -1;
     }
-    ctx.reset(ctx_p);
 
-    D3D11_VIDEO_PROCESSOR_CONTENT_DESC contentDesc {
-      D3D11_VIDEO_FRAME_FORMAT_PROGRESSIVE,
-      { 1, 1 }, (UINT)in_width, (UINT)in_height,
-      { 1, 1 }, (UINT)out_width, (UINT)out_height,
-      D3D11_VIDEO_USAGE_OPTIMAL_QUALITY
+    status = device_p->CreatePixelShader(convert_UV_ps_hlsl->GetBufferPointer(), convert_UV_ps_hlsl->GetBufferSize(), nullptr, &convert_UV_ps);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create convertUV pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    status = device_p->CreateVertexShader(convert_UV_vs_hlsl->GetBufferPointer(), convert_UV_vs_hlsl->GetBufferSize(), nullptr, &convert_UV_vs);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create convertUV vertex shader [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    status = device_p->CreatePixelShader(scene_ps_hlsl->GetBufferPointer(), scene_ps_hlsl->GetBufferSize(), nullptr, &scene_ps);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create scene pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    blend_disable = make_blend(device_p, false);
+    blend_enable = make_blend(device_p, true);
+
+    if(!blend_disable || !blend_enable) {
+      return -1;
+    }
+
+    if(_init_rt(scene_sr, scene_rt, in_width, in_height, DXGI_FORMAT_B8G8R8A8_UNORM)) {
+      return -1;
+    }
+
+    color_matrix = make_buffer(device_p, colors[0]);
+    if(!color_matrix) {
+      BOOST_LOG(error) << "Failed to create color matrix buffer"sv;
+      return -1;
+    }
+
+    float info_in[16 / sizeof(float)] { 1.0f / (float)out_width }; //aligned to 16-byte
+    info_scene = make_buffer(device_p, info_in);
+    if(!info_in) {
+      BOOST_LOG(error) << "Failed to create info scene buffer"sv;
+      return -1;
+    }
+
+    D3D11_INPUT_ELEMENT_DESC layout_desc {
+      "SV_Position", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0
     };
 
-    video::processor_enum_t::pointer vp_e_p;
-    status = device->CreateVideoProcessorEnumerator(&contentDesc, &vp_e_p);
-    if(FAILED(status)) {
-      BOOST_LOG(error) << "Failed to create video processor enumerator [0x"sv << util::hex(status).to_string_view() << ']';
-      return -1;
-    }
-    processor_e.reset(vp_e_p);
-
-    D3D11_VIDEO_PROCESSOR_CAPS proc_caps;
-    processor_e->GetVideoProcessorCaps(&proc_caps);
-    if(!(proc_caps.FeatureCaps & D3D11_VIDEO_PROCESSOR_FEATURE_CAPS_ALPHA_STREAM)) {
-      BOOST_LOG(warning) << "VideoProcessorSetStreamAlpha() not supported, hardware accelerated mouse cannot be added to the video stream"sv;
-    }
-
-    video::processor_t::pointer processor_p;
-    status = device->CreateVideoProcessor(processor_e.get(), 0, &processor_p);
-    if(FAILED(status)) {
-      BOOST_LOG(error) << "Failed to create video processor [0x"sv << util::hex(status).to_string_view() << ']';
-      return -1;
-    }
-    processor.reset(processor_p);
-
-    // Tell video processor alpha values need to be enabled
-    ctx->VideoProcessorSetStreamAlpha(processor.get(), 1, TRUE, 1.0f);
+    status = device_p->CreateInputLayout(
+      &layout_desc, 1,
+      convert_UV_vs_hlsl->GetBufferPointer(), convert_UV_vs_hlsl->GetBufferSize(),
+      &input_layout);
 
     D3D11_TEXTURE2D_DESC t {};
     t.Width  = out_width;
@@ -259,34 +467,61 @@ public:
     t.SampleDesc.Count = 1;
     t.Usage = D3D11_USAGE_DEFAULT;
     t.Format = pix_fmt == pix_fmt_e::nv12 ? DXGI_FORMAT_NV12 : DXGI_FORMAT_P010;
-    t.BindFlags = D3D11_BIND_RENDER_TARGET | D3D11_BIND_VIDEO_ENCODER;
+    t.BindFlags = D3D11_BIND_RENDER_TARGET;
 
-    dxgi::texture2d_t::pointer tex_p {};
-    status = device_p->CreateTexture2D(&t, nullptr, &tex_p);
+    status = device_p->CreateTexture2D(&t, nullptr, &img.texture);
     if(FAILED(status)) {
-      BOOST_LOG(error) << "Failed to create video output texture [0x"sv << util::hex(status).to_string_view() << ']';
+      BOOST_LOG(error) << "Failed to create render target texture [0x"sv << util::hex(status).to_string_view() << ']';
       return -1;
     }
 
-    img.texture.reset(tex_p);
     img.display = std::move(display);
     img.width = out_width;
     img.height = out_height;
-    img.data = (std::uint8_t*)tex_p;
+    img.data = (std::uint8_t*)img.texture.get();
     img.row_pitch = out_width;
     img.pixel_pitch = 1;
 
-    D3D11_VIDEO_PROCESSOR_OUTPUT_VIEW_DESC output_desc { D3D11_VPOV_DIMENSION_TEXTURE2D, 0 };
-    video::processor_out_t::pointer processor_out_p;
-    status = device->CreateVideoProcessorOutputView(img.texture.get(), processor_e.get(), &output_desc, &processor_out_p);
+    D3D11_RENDER_TARGET_VIEW_DESC nv12_rt_desc {
+      DXGI_FORMAT_R8_UNORM,
+      D3D11_RTV_DIMENSION_TEXTURE2D
+    };
+
+    status = device_p->CreateRenderTargetView(img.texture.get(), &nv12_rt_desc, &nv12_Y_rt);
     if(FAILED(status)) {
-      BOOST_LOG(error) << "Failed to create VideoProcessorOutputView [0x"sv << util::hex(status).to_string_view() << ']';
+      BOOST_LOG(error) << "Failed to create render target view [0x"sv << util::hex(status).to_string_view() << ']';
       return -1;
     }
-    processor_out.reset(processor_out_p);
 
-    device_p->AddRef();
-    data = device_p;
+    nv12_rt_desc.Format = DXGI_FORMAT_R8G8_UNORM;
+    status = device_p->CreateRenderTargetView(img.texture.get(), &nv12_rt_desc, &nv12_UV_rt);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to create render target view [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    D3D11_SAMPLER_DESC sampler_desc {};
+    sampler_desc.Filter = D3D11_FILTER_MIN_MAG_MIP_LINEAR;
+    sampler_desc.AddressU = D3D11_TEXTURE_ADDRESS_CLAMP;
+    sampler_desc.AddressV = D3D11_TEXTURE_ADDRESS_CLAMP;
+    sampler_desc.AddressW = D3D11_TEXTURE_ADDRESS_WRAP;
+    sampler_desc.ComparisonFunc = D3D11_COMPARISON_NEVER;
+    sampler_desc.MinLOD = 0;
+    sampler_desc.MaxLOD = D3D11_FLOAT32_MAX;
+
+    status = device_p->CreateSamplerState(&sampler_desc, &sampler_linear);
+    if(FAILED(status)) {
+      BOOST_LOG(error) << "Failed to create point sampler state [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    device_ctx_p->OMSetBlendState(blend_disable.get(), nullptr, 0xFFFFFFFFu);
+    device_ctx_p->PSSetSamplers(0, 1, &sampler_linear);
+    device_ctx_p->PSSetConstantBuffers(0, 1, &color_matrix);
+    device_ctx_p->VSSetConstantBuffers(0, 1, &info_scene);
+    device_ctx_p->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+    device_ctx_p->IASetInputLayout(input_layout.get());
+
     return 0;
   }
 
@@ -300,25 +535,104 @@ public:
       hwdevices_p->erase(it);
     }
   }
+private:
+  void _init_view_port(float x, float y, float width, float height) {
+    D3D11_VIEWPORT view {
+      x, y,
+      width, height,
+      0.0f, 1.0f
+    };
+
+    device_ctx_p->RSSetViewports(1, &view);
+  }
+
+  void _init_view_port(float width, float height) {
+    _init_view_port(0.0f, 0.0f, width, height);
+  }
+
+  int _init_rt(shader_res_t &shader_res, render_target_t &render_target, int width, int height, DXGI_FORMAT format) {
+    D3D11_TEXTURE2D_DESC desc {};
+
+    desc.Width            = width;
+    desc.Height           = height;
+    desc.Format           = format;
+    desc.Usage            = D3D11_USAGE_DEFAULT;
+    desc.BindFlags        = D3D11_BIND_RENDER_TARGET | D3D11_BIND_SHADER_RESOURCE;
+    desc.MipLevels        = 1;
+    desc.ArraySize        = 1;
+    desc.SampleDesc.Count = 1;
+
+    auto device = (device_t::pointer)data;
+
+    texture2d_t tex;
+    auto status = device->CreateTexture2D(&desc, nullptr, &tex);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create render target texture for luma [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+
+    D3D11_SHADER_RESOURCE_VIEW_DESC shader_resource_desc {
+      format,
+      D3D11_SRV_DIMENSION_TEXTURE2D
+    };
+    shader_resource_desc.Texture2D.MipLevels = 1;
+
+    device->CreateShaderResourceView(tex.get(), &shader_resource_desc, &shader_res);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create render target texture for luma [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    D3D11_RENDER_TARGET_VIEW_DESC render_target_desc {
+      format,
+      D3D11_RTV_DIMENSION_TEXTURE2D
+    };
+
+    device->CreateRenderTargetView(tex.get(), &render_target_desc, &render_target);
+    if(status) {
+      BOOST_LOG(error) << "Failed to create render target view [0x"sv << util::hex(status).to_string_view() << ']';
+      return -1;
+    }
+
+    return 0;
+  }
+
+public:
+  color_t *color_p;
+
+  blend_t blend_enable;
+  blend_t blend_disable;
+
+  buf_t info_scene;
+  buf_t color_matrix;
+
+  sampler_state_t sampler_linear;
+
+  input_layout_t input_layout;
+
+  render_target_t nv12_Y_rt;
+  render_target_t nv12_UV_rt;
+
+  render_target_t scene_rt;
+  shader_res_t scene_sr;
 
   img_d3d_t img;
-  video::device_t device;
-  video::ctx_t ctx;
-  video::processor_enum_t processor_e;
-  video::processor_t processor;
-  video::processor_out_t processor_out;
-  std::unordered_map<texture2d_t::pointer, video::processor_in_t> texture_to_processor_in;
 
-  video::processor_in_t cursor_in;
+  vs_t convert_UV_vs;
+  ps_t convert_UV_ps;
+  ps_t convert_Y_ps;
+  ps_t scene_ps;
+  vs_t scene_vs;
 
+  D3D11_VIEWPORT cursor_view;
   bool cursor_visible;
 
-  LONG cursor_width, cursor_height;
-  LONG cursor_scaled_width, cursor_scaled_height;
+  float out_width, out_height;
 
-  LONG in_width, in_height;
-  double out_width, out_height;
+  device_ctx_t::pointer device_ctx_p;
 
+  // The destructor will remove itself from the list of hardware devices, this is done synchronously
   std::vector<hwdevice_t*> *hwdevices_p;
 };
 
@@ -375,18 +689,17 @@ capture_e display_vram_t::snapshot(platf::img_t *img_base, std::chrono::millisec
     t.SampleDesc.Count = 1;
     t.Usage = D3D11_USAGE_DEFAULT;
     t.Format = DXGI_FORMAT_B8G8R8A8_UNORM;
-    t.BindFlags = D3D11_BIND_RENDER_TARGET;
+    t.BindFlags = D3D11_BIND_SHADER_RESOURCE;
 
-    dxgi::texture2d_t::pointer tex_p {};
-    auto status = device->CreateTexture2D(&t, &data, &tex_p);
+    texture2d_t texture;
+    auto status = device->CreateTexture2D(&t, &data, &texture);
     if(FAILED(status)) {
       BOOST_LOG(error) << "Failed to create mouse texture [0x"sv << util::hex(status).to_string_view() << ']';
       return capture_e::error;
     }
-    texture2d_t texture { tex_p };
 
     for(auto *hwdevice : hwdevices) {
-      if(hwdevice->set_cursor_texture(tex_p, t.Width, t.Height)) {
+      if(hwdevice->set_cursor_texture(texture.get(), t.Width, t.Height)) {
         return capture_e::error;
       }
     }
@@ -403,15 +716,14 @@ capture_e display_vram_t::snapshot(platf::img_t *img_base, std::chrono::millisec
   }
 
   if(frame_update_flag) {
-    texture2d_t::pointer src_p {};
-    status = res->QueryInterface(IID_ID3D11Texture2D, (void **)&src_p);
+    texture2d_t src;
+    status = res->QueryInterface(IID_ID3D11Texture2D, (void **)&src);
 
     if(FAILED(status)) {
       BOOST_LOG(error) << "Couldn't query interface [0x"sv << util::hex(status).to_string_view() << ']';
       return capture_e::error;
     }
 
-    texture2d_t src { src_p };
     device_ctx->CopyResource(img->texture.get(), src.get());
   }
 
@@ -429,21 +741,19 @@ std::shared_ptr<platf::img_t> display_vram_t::alloc_img() {
   t.SampleDesc.Count = 1;
   t.Usage = D3D11_USAGE_DEFAULT;
   t.Format = format;
-  t.BindFlags = D3D11_BIND_RENDER_TARGET;
+  t.BindFlags = D3D11_BIND_SHADER_RESOURCE;
 
-  dxgi::texture2d_t::pointer tex_p {};
-  auto status = device->CreateTexture2D(&t, nullptr, &tex_p);
+  auto status = device->CreateTexture2D(&t, nullptr, &img->texture);
   if(FAILED(status)) {
     BOOST_LOG(error) << "Failed to create img buf texture [0x"sv << util::hex(status).to_string_view() << ']';
     return nullptr;
   }
 
-  img->data        = (std::uint8_t*)tex_p;
+  img->data        = (std::uint8_t*)img->texture.get();
   img->row_pitch   = 0;
   img->pixel_pitch = 4;
   img->width       = 0;
   img->height      = 0;
-  img->texture.reset(tex_p);
   img->display     = shared_from_this();
 
   return img;
@@ -456,8 +766,7 @@ int display_vram_t::dummy_img(platf::img_t *img_base) {
   auto dummy_data = std::make_unique<int[]>(width * height);
   D3D11_SUBRESOURCE_DATA data {
     dummy_data.get(),
-    (UINT)img->row_pitch,
-    0
+    (UINT)img->row_pitch
   };
 
   D3D11_TEXTURE2D_DESC t {};
@@ -468,17 +777,17 @@ int display_vram_t::dummy_img(platf::img_t *img_base) {
   t.SampleDesc.Count = 1;
   t.Usage = D3D11_USAGE_DEFAULT;
   t.Format = format;
-  t.BindFlags = D3D11_BIND_RENDER_TARGET;
+  t.BindFlags = D3D11_BIND_SHADER_RESOURCE;
 
-  dxgi::texture2d_t::pointer tex_p {};
-  auto status = device->CreateTexture2D(&t, &data, &tex_p);
+  dxgi::texture2d_t tex;
+  auto status = device->CreateTexture2D(&t, &data, &tex);
   if(FAILED(status)) {
     BOOST_LOG(error) << "Failed to create dummy texture [0x"sv << util::hex(status).to_string_view() << ']';
     return -1;
   }
 
-  img->data        = (std::uint8_t*)tex_p;
-  img->texture.reset(tex_p);
+  img->texture     = std::move(tex);
+  img->data        = (std::uint8_t*)img->texture.get();
   img->height      = height;
   img->width       = width;
   img->pixel_pitch = 4;
@@ -487,8 +796,8 @@ int display_vram_t::dummy_img(platf::img_t *img_base) {
 }
 
 std::shared_ptr<platf::hwdevice_t> display_vram_t::make_hwdevice(int width, int height, pix_fmt_e pix_fmt) {
-  if(pix_fmt != platf::pix_fmt_e::nv12 && pix_fmt != platf::pix_fmt_e::p010) {
-    BOOST_LOG(error) << "display_vram_t doesn't support pixel format ["sv << (int)pix_fmt << ']';
+  if(pix_fmt != platf::pix_fmt_e::nv12) {
+    BOOST_LOG(error) << "display_vram_t doesn't support pixel format ["sv << from_pix_fmt(pix_fmt) << ']';
 
     return nullptr;
   }
@@ -515,4 +824,44 @@ std::shared_ptr<platf::hwdevice_t> display_vram_t::make_hwdevice(int width, int
 
   return hwdevice;
 }
+
+int init() {
+  for(auto &color : colors) {
+    BOOST_LOG(debug) << "Color Matrix"sv;
+    BOOST_LOG(debug) << "Y ["sv << color.color_vec_y.x << ", "sv << color.color_vec_y.y << ", "sv << color.color_vec_y.z << ", "sv << color.color_vec_y.w << ']';
+    BOOST_LOG(debug) << "U ["sv << color.color_vec_u.x << ", "sv << color.color_vec_u.y << ", "sv << color.color_vec_u.z << ", "sv << color.color_vec_u.w << ']';
+    BOOST_LOG(debug) << "V ["sv << color.color_vec_v.x << ", "sv << color.color_vec_v.y << ", "sv << color.color_vec_v.z << ", "sv << color.color_vec_v.w << ']';
+    BOOST_LOG(debug) << "range Y  ["sv << color.range_y.x << ", "sv << color.range_y.y << ']';
+    BOOST_LOG(debug) << "range UV ["sv << color.range_uv.x << ", "sv << color.range_uv.y << ']';
+  }
+
+  BOOST_LOG(info) << "Compiling shaders..."sv;
+  scene_vs_hlsl = compile_vertex_shader(SUNSHINE_SHADERS_DIR "/SceneVS.hlsl");
+  if(!scene_vs_hlsl) {
+    return -1;
+  }
+
+  convert_Y_ps_hlsl = compile_pixel_shader(SUNSHINE_SHADERS_DIR "/ConvertYPS.hlsl");
+  if(!convert_Y_ps_hlsl) {
+    return -1;
+  }
+
+  convert_UV_ps_hlsl = compile_pixel_shader(SUNSHINE_SHADERS_DIR "/ConvertUVPS.hlsl");
+  if(!convert_UV_ps_hlsl) {
+    return -1;
+  }
+
+  convert_UV_vs_hlsl = compile_vertex_shader(SUNSHINE_SHADERS_DIR "/ConvertUVVS.hlsl");
+  if(!convert_UV_vs_hlsl) {
+    return -1;
+  }
+
+  scene_ps_hlsl = compile_pixel_shader(SUNSHINE_SHADERS_DIR "/ScenePS.hlsl");
+  if(!scene_ps_hlsl) {
+    return -1;
+  }
+  BOOST_LOG(info) << "Compiled shaders"sv;
+
+  return 0;
+}
 }
\ No newline at end of file
diff --git a/sunshine/utility.h b/sunshine/utility.h
index 0c03c51d..63be108c 100644
--- a/sunshine/utility.h
+++ b/sunshine/utility.h
@@ -76,37 +76,6 @@ struct __either<false, X, Y> {
 template<bool V, class X, class Y>
 using either_t = typename __either<V, X, Y>::type;
 
-template<class T, class V = void>
-struct __false_v;
-
-template<class T>
-struct __false_v<T, std::enable_if_t<instantiation_of_v<std::optional, T>>> {
-  static constexpr std::nullopt_t value = std::nullopt;
-};
-
-template<class T>
-struct __false_v<T, std::enable_if_t<
-  (std::is_pointer_v<T> || instantiation_of_v<std::unique_ptr, T> || instantiation_of_v<std::shared_ptr, T>)
-  >> {
-  static constexpr std::nullptr_t value = nullptr;
-};
-
-template<class T>
-struct __false_v<T, std::enable_if_t<std::is_same_v<T, bool>>> {
-  static constexpr bool value = false;
-};
-
-template<class T>
-static constexpr auto false_v = __false_v<T>::value;
-
-template<class T>
-using optional_t = either_t<
-  (std::is_same_v<T, bool> ||
-   instantiation_of_v<std::unique_ptr, T> ||
-   instantiation_of_v<std::shared_ptr, T> ||
-   std::is_pointer_v<T>),
-  T, std::optional<T>>;
-
 template<class... Ts> struct overloaded : Ts... { using Ts::operator()...; };
 template<class... Ts> overloaded(Ts...) -> overloaded<Ts...>;
 
@@ -362,35 +331,6 @@ auto enm(T& val) -> std::underlying_type_t<T>& {
   return *reinterpret_cast<std::underlying_type_t<T>*>(&val);
 }
 
-template<class ReturnType, class ...Args>
-struct Function {
-  typedef ReturnType (*type)(Args...);
-};
-
-template<class T, class ReturnType, typename Function<ReturnType, T>::type function>
-struct Destroy {
-  typedef T pointer;
-  
-  void operator()(pointer p) {
-    function(p);
-  }
-};
-
-template<class T, typename Function<void, T*>::type function>
-using safe_ptr = std::unique_ptr<T, Destroy<T*, void, function>>;
-
-// You cannot specialize an alias
-template<class T, class ReturnType, typename Function<ReturnType, T*>::type function>
-using safe_ptr_v2 = std::unique_ptr<T, Destroy<T*, ReturnType, function>>;
-
-template<class T>
-void c_free(T *p) {
-  free(p);
-}
-
-template<class T>
-using c_ptr = safe_ptr<T, c_free<T>>;
-
 inline std::int64_t from_chars(const char *begin, const char *end) {
   std::int64_t res {};
   std::int64_t mul = 1;
@@ -436,6 +376,163 @@ public:
   }
 };
 
+// Compared to std::unique_ptr, it adds the ability to get the address of the pointer itself
+template <typename T, typename D = std::default_delete<T>>
+class uniq_ptr {
+public:
+  using element_type = T;
+  using pointer = element_type*;
+  using deleter_type = D;
+
+  constexpr uniq_ptr() noexcept : _p { nullptr } {}
+  constexpr uniq_ptr(std::nullptr_t) noexcept : _p { nullptr } {}
+
+  uniq_ptr(const uniq_ptr &other) noexcept = delete;
+  uniq_ptr &operator=(const uniq_ptr &other) noexcept = delete;
+
+  template<class V>
+  uniq_ptr(V *p) noexcept : _p { p } {
+    static_assert(std::is_same_v<element_type, void> || std::is_same_v<element_type, V> || std::is_base_of_v<element_type, V>, "element_type must be base class of V");
+  }
+
+  template<class V>
+  uniq_ptr(std::unique_ptr<V, deleter_type> &&uniq) noexcept : _p { uniq.release() } {
+    static_assert(std::is_same_v<element_type, void> || std::is_same_v<T, V> || std::is_base_of_v<element_type, V>, "element_type must be base class of V");
+  }
+
+  template<class V>
+  uniq_ptr(uniq_ptr<V, deleter_type> &&other) noexcept : _p { other.release() } {
+    static_assert(std::is_same_v<element_type, void> || std::is_same_v<T, V> || std::is_base_of_v<element_type, V>, "element_type must be base class of V");
+  }
+
+  template<class V>
+  uniq_ptr &operator=(uniq_ptr<V, deleter_type> &&other) noexcept {
+    static_assert(std::is_same_v<element_type, void> || std::is_same_v<T, V> || std::is_base_of_v<element_type, V>, "element_type must be base class of V");
+    reset(other.release());
+
+    return *this;
+  }
+
+  template<class V>
+  uniq_ptr &operator=(std::unique_ptr<V, deleter_type> &&uniq) noexcept {
+    static_assert(std::is_same_v<element_type, void> || std::is_same_v<T, V> || std::is_base_of_v<element_type, V>, "element_type must be base class of V");
+
+    reset(uniq.release());
+
+    return *this;
+  }
+
+  ~uniq_ptr() {
+    reset();
+  }
+
+  void reset(pointer p = pointer()) {
+    if(_p) {
+      _deleter(_p);
+    }
+
+    _p = p;
+  }
+
+  pointer release() {
+    auto tmp = _p;
+    _p = nullptr;
+    return tmp;
+  }
+
+  pointer get() {
+    return _p;
+  }
+
+  const pointer get() const {
+    return _p;
+  }
+
+  const std::add_lvalue_reference_t<element_type> operator*() const {
+    return *_p;
+  }
+  std::add_lvalue_reference_t<element_type> operator*() {
+    return *_p;
+  }
+  const pointer operator->() const {
+    return _p;
+  }
+  pointer operator->() {
+    return _p;
+  }
+  pointer *operator&() const {
+    return &_p;
+  }
+
+  pointer *operator&() {
+    return &_p;
+  }
+
+  deleter_type& get_deleter() {
+    return _deleter;
+  }
+
+  const deleter_type& get_deleter() const {
+    return _deleter;
+  }
+
+  explicit operator bool() const {
+    return _p != nullptr;
+  }
+protected:
+  pointer _p;
+  deleter_type _deleter;
+};
+
+template<class T1, class D1, class T2, class D2>
+bool operator==(const uniq_ptr<T1, D1>& x, const uniq_ptr<T2, D2>& y) {
+  return x.get() == y.get();
+}
+
+template<class T1, class D1, class T2, class D2>
+bool operator!=(const uniq_ptr<T1, D1>& x, const uniq_ptr<T2, D2>& y) {
+  return x.get() != y.get();
+}
+
+template<class T1, class D1, class T2, class D2>
+bool operator==(const std::unique_ptr<T1, D1>& x, const uniq_ptr<T2, D2>& y) {
+  return x.get() == y.get();
+}
+
+template<class T1, class D1, class T2, class D2>
+bool operator!=(const std::unique_ptr<T1, D1>& x, const uniq_ptr<T2, D2>& y) {
+  return x.get() != y.get();
+}
+
+template<class T1, class D1, class T2, class D2>
+bool operator==(const uniq_ptr<T1, D1>& x, const std::unique_ptr<T1, D1>& y) {
+  return x.get() == y.get();
+}
+
+template<class T1, class D1, class T2, class D2>
+bool operator!=(const uniq_ptr<T1, D1>& x, const std::unique_ptr<T1, D1>& y) {
+  return x.get() != y.get();
+}
+
+template<class T, class D>
+bool operator==(const uniq_ptr<T, D>& x, std::nullptr_t) {
+  return !(bool)x;
+}
+
+template<class T, class D>
+bool operator!=(const uniq_ptr<T, D>& x, std::nullptr_t) {
+  return (bool)x;
+}
+
+template<class T, class D>
+bool operator==(std::nullptr_t, const uniq_ptr<T, D>& y) {
+  return !(bool)y;
+}
+
+template<class T, class D>
+bool operator!=(std::nullptr_t, const uniq_ptr<T, D>& y) {
+  return (bool)y;
+}
 
 template<class T>
 class wrap_ptr {
@@ -510,6 +607,43 @@ private:
   pointer _p;
 };
 
+template<class T, class V = void>
+struct __false_v;
+
+template<class T>
+struct __false_v<T, std::enable_if_t<instantiation_of_v<std::optional, T>>> {
+  static constexpr std::nullopt_t value = std::nullopt;
+};
+
+template<class T>
+struct __false_v<T, std::enable_if_t<
+  (
+    std::is_pointer_v<T> ||
+    instantiation_of_v<std::unique_ptr, T> ||
+    instantiation_of_v<std::shared_ptr, T> ||
+    instantiation_of_v<uniq_ptr, T>
+  )
+  >> {
+  static constexpr std::nullptr_t value = nullptr;
+};
+
+template<class T>
+struct __false_v<T, std::enable_if_t<std::is_same_v<T, bool>>> {
+  static constexpr bool value = false;
+};
+
+template<class T>
+static constexpr auto false_v = __false_v<T>::value;
+
+template<class T>
+using optional_t = either_t<
+  (std::is_same_v<T, bool> ||
+   instantiation_of_v<std::unique_ptr, T> ||
+   instantiation_of_v<std::shared_ptr, T> ||
+   instantiation_of_v<uniq_ptr, T> ||
+   std::is_pointer_v<T>),
+  T, std::optional<T>>;
+
 template<class T>
 class buffer_t {
 public:
@@ -569,6 +703,35 @@ T either(std::optional<T> &&l, T &&r) {
   return std::forward<T>(r);
 }
 
+template<class ReturnType, class ...Args>
+struct Function {
+  typedef ReturnType (*type)(Args...);
+};
+
+template<class T, class ReturnType, typename Function<ReturnType, T>::type function>
+struct Destroy {
+  typedef T pointer;
+  
+  void operator()(pointer p) {
+    function(p);
+  }
+};
+
+template<class T, typename Function<void, T*>::type function>
+using safe_ptr = uniq_ptr<T, Destroy<T*, void, function>>;
+
+// You cannot specialize an alias
+template<class T, class ReturnType, typename Function<ReturnType, T*>::type function>
+using safe_ptr_v2 = uniq_ptr<T, Destroy<T*, ReturnType, function>>;
+
+template<class T>
+void c_free(T *p) {
+  free(p);
+}
+
+template<class T>
+using c_ptr = safe_ptr<T, c_free<T>>;
+
 namespace endian {
 template<class T = void>
 struct endianness {
diff --git a/sunshine/video.cpp b/sunshine/video.cpp
index b2e23855..c6befe8d 100644
--- a/sunshine/video.cpp
+++ b/sunshine/video.cpp
@@ -58,20 +58,6 @@ enum class profile_hevc_e : int {
 };
 }
 
-namespace amd {
-
-enum class profile_h264_e : int {
-  main,
-  high,
-  constrained_baseline,
-  constrained_high,
-};
-
-enum class profile_hevc_e : int {
-  main,
-};
-}
-
 using ctx_t       = util::safe_ptr<AVCodecContext, free_ctx>;
 using frame_t     = util::safe_ptr<AVFrame, free_frame>;
 using buffer_t    = util::safe_ptr<AVBufferRef, free_buffer>;
@@ -82,10 +68,10 @@ platf::dev_type_e map_dev_type(AVHWDeviceType type);
 platf::pix_fmt_e map_pix_fmt(AVPixelFormat fmt);
 
 void sw_img_to_frame(const platf::img_t &img, frame_t &frame);
-void nv_d3d_img_to_frame(const platf::img_t &img, frame_t &frame);
-util::Either<buffer_t, int> nv_d3d_make_hwdevice_ctx(platf::hwdevice_t *hwdevice_ctx);
-void amd_d3d_img_to_frame(const platf::img_t &img, frame_t &frame);
-util::Either<buffer_t, int> amd_d3d_make_hwdevice_ctx(platf::hwdevice_t *hwdevice_ctx);
+void dxgi_img_to_frame(const platf::img_t &img, frame_t &frame);
+util::Either<buffer_t, int> dxgi_make_hwdevice_ctx(platf::hwdevice_t *hwdevice_ctx);
+void dxgi_img_to_frame(const platf::img_t &img, frame_t &frame);
+util::Either<buffer_t, int> dxgi_make_hwdevice_ctx(platf::hwdevice_t *hwdevice_ctx);
 
 util::Either<buffer_t, int> make_hwdevice_ctx(AVHWDeviceType type, void *hwdevice_ctx);
 int hwframe_ctx(ctx_t &ctx, buffer_t &hwdevice, AVPixelFormat format);
@@ -297,16 +283,16 @@ static encoder_t nvenc {
   false,
   true,
 
-  nv_d3d_img_to_frame,
-  nv_d3d_make_hwdevice_ctx
+  dxgi_img_to_frame,
+  dxgi_make_hwdevice_ctx
 };
 
 static encoder_t amdvce {
   "amdvce"sv,
-  { (int)amd::profile_h264_e::high, (int)amd::profile_hevc_e::main },
+  { FF_PROFILE_H264_HIGH, FF_PROFILE_HEVC_MAIN },
   AV_HWDEVICE_TYPE_D3D11VA,
   AV_PIX_FMT_D3D11,
-  AV_PIX_FMT_NV12, AV_PIX_FMT_YUV420P,
+  AV_PIX_FMT_NV12, AV_PIX_FMT_P010,
   {
     {
       { "header_insertion_mode"s, "idr"s },
@@ -331,8 +317,8 @@ static encoder_t amdvce {
   false,
   true,
 
-  amd_d3d_img_to_frame,
-  amd_d3d_make_hwdevice_ctx
+  dxgi_img_to_frame,
+  dxgi_make_hwdevice_ctx
 };
 #endif
 
@@ -373,11 +359,9 @@ static encoder_t software {
 static std::vector<encoder_t> encoders {
 #ifdef _WIN32
   nvenc,
-#endif
-  software,
-#ifdef _WIN32
   amdvce,
 #endif
+  software
 };
 
 void reset_display(std::shared_ptr<platf::display_t> &disp, AVHWDeviceType type) {
@@ -622,6 +606,7 @@ std::optional<session_t> make_session(const encoder_t &encoder, const config_t &
     case 0:
     default:
       // Rec. 601
+      BOOST_LOG(info) << "Color coding [Rec. 601]"sv;
       ctx->color_primaries = AVCOL_PRI_SMPTE170M;
       ctx->color_trc = AVCOL_TRC_SMPTE170M;
       ctx->colorspace = AVCOL_SPC_SMPTE170M;
@@ -630,6 +615,7 @@ std::optional<session_t> make_session(const encoder_t &encoder, const config_t &
 
     case 1:
       // Rec. 709
+      BOOST_LOG(info) << "Color coding [Rec. 709]"sv;
       ctx->color_primaries = AVCOL_PRI_BT709;
       ctx->color_trc = AVCOL_TRC_BT709;
       ctx->colorspace = AVCOL_SPC_BT709;
@@ -638,12 +624,14 @@ std::optional<session_t> make_session(const encoder_t &encoder, const config_t &
 
     case 2:
       // Rec. 2020
+      BOOST_LOG(info) << "Color coding [Rec. 2020]"sv;
       ctx->color_primaries = AVCOL_PRI_BT2020;
       ctx->color_trc = AVCOL_TRC_BT2020_10;
       ctx->colorspace = AVCOL_SPC_BT2020_NCL;
       sws_color_space = SWS_CS_BT2020;
       break;
   }
+  BOOST_LOG(info) << "Color range: ["sv << ((config.encoderCscMode & 0x1) ? "JPEG"sv : "MPEG"sv) << ']';
 
   AVPixelFormat sw_fmt;
   if(config.dynamicRange == 0) {
@@ -1196,7 +1184,7 @@ bool validate_encoder(encoder_t &encoder) {
   encoder.hevc[encoder_t::PASSED] = test_hevc;
 
   std::vector<std::pair<encoder_t::flag_e, config_t>> configs {
-    { encoder_t::DYNAMIC_RANGE, { 1920, 1080, 60, 1000, 1, 0, 1, 1, 1 } }
+    { encoder_t::DYNAMIC_RANGE, { 1920, 1080, 60, 1000, 1, 0, 3, 1, 1 } }
   };
   for(auto &[flag, config] : configs) {
     auto h264 = config;
@@ -1300,7 +1288,16 @@ int hwframe_ctx(ctx_t &ctx, buffer_t &hwdevice, AVPixelFormat format) {
 void sw_img_to_frame(const platf::img_t &img, frame_t &frame) {}
 
 #ifdef _WIN32
-void nv_d3d_img_to_frame(const platf::img_t &img, frame_t &frame) {
+}
+
+// Ugly, but need to declare for wio
+namespace platf::dxgi {
+void lock(void *hwdevice);
+void unlock(void *hwdevice);
+}
+void do_nothing(void*) {}
+namespace video {
+void dxgi_img_to_frame(const platf::img_t &img, frame_t &frame) {
   if(img.data == frame->data[0]) {
     return;
   }
@@ -1323,65 +1320,25 @@ void nv_d3d_img_to_frame(const platf::img_t &img, frame_t &frame) {
   frame->width = img.width;
 }
 
-void amd_d3d_img_to_frame(const platf::img_t &img, frame_t &frame) {
-  if(img.data == frame->data[0]) {
-    return;
-  }
-  
-  // Need to have something refcounted
-  if(!frame->buf[0]) {
-    frame->buf[0] = av_buffer_allocz(sizeof(AVD3D11FrameDescriptor));
-  }
-
-  auto desc = (AVD3D11FrameDescriptor*)frame->buf[0]->data;
-  desc->texture = (ID3D11Texture2D*)img.data;
-  desc->index = 0;
-
-  frame->data[0] = img.data;
-  frame->data[1] = 0;
-
-  frame->linesize[0] = img.row_pitch;
-
-  frame->height = img.height;
-  frame->width = img.width;
-}
-
-
-util::Either<buffer_t, int> nv_d3d_make_hwdevice_ctx(platf::hwdevice_t *hwdevice_ctx) {
+util::Either<buffer_t, int> dxgi_make_hwdevice_ctx(platf::hwdevice_t *hwdevice_ctx) {
   buffer_t ctx_buf { av_hwdevice_ctx_alloc(AV_HWDEVICE_TYPE_D3D11VA) };
   auto ctx = (AVD3D11VADeviceContext*)((AVHWDeviceContext*)ctx_buf->data)->hwctx;
   
   std::fill_n((std::uint8_t*)ctx, sizeof(AVD3D11VADeviceContext), 0);
 
   auto device = (ID3D11Device*)hwdevice_ctx->data;
+
   device->AddRef();
   ctx->device = device;
 
+  ctx->lock_ctx = (void*)1;
+  ctx->lock = do_nothing;
+  ctx->unlock = do_nothing;
+
   auto err = av_hwdevice_ctx_init(ctx_buf.get());
   if(err) {
     char err_str[AV_ERROR_MAX_STRING_SIZE] {0};
-    BOOST_LOG(error) << "Failed to create FFMpeg nvenc: "sv << av_make_error_string(err_str, AV_ERROR_MAX_STRING_SIZE, err);
-
-    return err;
-  }
-
-  return ctx_buf;
-}
-
-util::Either<buffer_t, int> amd_d3d_make_hwdevice_ctx(platf::hwdevice_t *hwdevice_ctx) {
-  buffer_t ctx_buf { av_hwdevice_ctx_alloc(AV_HWDEVICE_TYPE_D3D11VA) };
-  auto ctx = (AVD3D11VADeviceContext*)((AVHWDeviceContext*)ctx_buf->data)->hwctx;
-  
-  std::fill_n((std::uint8_t*)ctx, sizeof(AVD3D11VADeviceContext), 0);
-
-  auto device = (ID3D11Device*)hwdevice_ctx->data;
-  device->AddRef();
-  ctx->device = device;
-
-  auto err = av_hwdevice_ctx_init(ctx_buf.get());
-  if(err) {
-    char err_str[AV_ERROR_MAX_STRING_SIZE] {0};
-    BOOST_LOG(error) << "Failed to create FFMpeg amddech: "sv << av_make_error_string(err_str, AV_ERROR_MAX_STRING_SIZE, err);
+    BOOST_LOG(error) << "Failed to create FFMpeg hardware device context: "sv << av_make_error_string(err_str, AV_ERROR_MAX_STRING_SIZE, err);
 
     return err;
   }