[VideoPlayer] WinRenderer: load 3dlut asynchronous

5 files changed, 73 insertions, 38 deletions

diff --git a/system/shaders/output_d3d.fx b/system/shaders/output_d3d.fx
index 6f3f1a3890..40d6252697 100644
--- a/system/shaders/output_d3d.fx
+++ b/system/shaders/output_d3d.fx
@@ -19,7 +19,7 @@
  */
 
 #if defined(KODI_3DLUT)
-float     m_CLUTsize;
+float2    m_CLUTParams; // x- scale, y- offset
 texture3D m_CLUT;
 
 SamplerState LutSampler : IMMUTABLE
@@ -45,9 +45,10 @@ SamplerState DitherSampler : IMMUTABLE
 float4 output4(float4 color, float2 uv)
 {
 #if defined(KODI_3DLUT)
-  half3 scale = (m_CLUTsize - 1.0) / m_CLUTsize;
-  half3 offset = 1.0 / (2.0 * m_CLUTsize);
-  color.rgb = m_CLUT.Sample(LutSampler, color.rgb*scale + offset).rgb;
+  half3 scale = m_CLUTParams.x;
+  half3 offset = m_CLUTParams.y;
+  float3 lutRGB = m_CLUT.Sample(LutSampler, color.rgb*scale + offset).rgb;
+  color.rgb = scale.x ? lutRGB : color.rgb;
 #endif
 #if defined(KODI_DITHER)
   half2 ditherpos  = uv * m_ditherParams.xy;
@@ -70,7 +71,7 @@ float3 m_params; // 0 - range (0 - full, 1 - limited), 1 - contrast, 2 - brightn
 float4 OUTPUT_PS(VS_OUTPUT In) : SV_TARGET
 {
   float4 color = g_Texture.Sample(KernelSampler, In.TextureUV);
-  if (m_params.x)
+  [flatten] if (m_params.x)
     color = saturate(0.0625 + color * 219.0 / 255.0);
 
   color *= m_params.y * 2.0;
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/WinVideoFilter.cpp b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/WinVideoFilter.cpp
index e9063be2a2..9128d24273 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/WinVideoFilter.cpp
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/WinVideoFilter.cpp
@@ -255,10 +255,16 @@ COutputShader::~COutputShader()
 
 void COutputShader::ApplyEffectParameters(CD3DEffect &effect, unsigned sourceWidth, unsigned sourceHeight)
 {
-  if (HasCLUT())
+  if (m_useCLUT)
   {
+    float clut_params[2] = { 0.0f, 0.0f };
+    if (HasCLUT())
+    {
+      clut_params[0] = (m_clutSize - 1) / static_cast<float>(m_clutSize);
+      clut_params[1] = 0.5f / static_cast<float>(m_clutSize);
+    };
     effect.SetResources("m_CLUT", &m_pCLUTView, 1);
-    effect.SetScalar("m_CLUTsize", m_clutSize);
+    effect.SetFloatArray("m_CLUTParams", clut_params, 2);
   }
   if (m_useDithering)
   {
@@ -275,7 +281,7 @@ void COutputShader::ApplyEffectParameters(CD3DEffect &effect, unsigned sourceWid
 
 void COutputShader::GetDefines(DefinesMap& map) const
 {
-  if (HasCLUT())
+  if (m_useCLUT)
   {
     map["KODI_3DLUT"] = "";
   }
@@ -285,10 +291,9 @@ void COutputShader::GetDefines(DefinesMap& map) const
   }
 }
 
-bool COutputShader::Create(int clutSize, ID3D11ShaderResourceView* pCLUTView, bool useDithering, int ditherDepth)
+bool COutputShader::Create(bool useCLUT, bool useDithering, int ditherDepth)
 {
-  m_clutSize = clutSize;
-  m_pCLUTView = pCLUTView;
+  m_useCLUT = useCLUT;
   m_ditherDepth = ditherDepth;
 
   CWinShader::CreateVertexBuffer(4, sizeof(CUSTOMVERTEX));
@@ -338,6 +343,12 @@ void COutputShader::Render(CD3DTexture &sourceTexture, unsigned sourceWidth, uns
   Render(sourceTexture, sourceWidth, sourceHeight, sourceRect, points, range, contrast, brightness);
 }
 
+void COutputShader::SetCLUT(int clutSize, ID3D11ShaderResourceView* pCLUTView)
+{
+  m_clutSize = clutSize;
+  m_pCLUTView = pCLUTView;
+}
+
 bool COutputShader::CreateCLUTView(int clutSize, uint16_t* clutData, ID3D11ShaderResourceView** ppCLUTView)
 {
   if (!clutSize || !clutData)
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/WinVideoFilter.h b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/WinVideoFilter.h
index be3c6a8754..4fbbd4bbe4 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/WinVideoFilter.h
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/VideoShaders/WinVideoFilter.h
@@ -80,11 +80,12 @@ public:
 
   void ApplyEffectParameters(CD3DEffect &effect, unsigned sourceWidth, unsigned sourceHeight);
   void GetDefines(DefinesMap &map) const;
-  bool Create(int clutSize, ID3D11ShaderResourceView *pCLUTView, bool useDithering, int ditherDepth);
+  bool Create(bool useCLUT, bool useDithering, int ditherDepth);
   void Render(CD3DTexture &sourceTexture, unsigned sourceWidth, unsigned sourceHeight, CRect sourceRect, const CPoint points[4]
             , unsigned range = 0, float contrast = 0.5f, float brightness = 0.5f);
   void Render(CD3DTexture &sourceTexture, unsigned sourceWidth, unsigned sourceHeight, CRect sourceRect, CRect destRect
             , unsigned range = 0, float contrast = 0.5f, float brightness = 0.5f);
+  void SetCLUT(int clutSize, ID3D11ShaderResourceView *pCLUTView);
 
   static bool CreateCLUTView(int clutSize, uint16_t* clutData, ID3D11ShaderResourceView** ppCLUTView);
 
@@ -94,6 +95,7 @@ private:
   void SetShaderParameters(CD3DTexture &sourceTexture, unsigned range, float contrast, float brightness);
   void CreateDitherView();
 
+  bool m_useCLUT{ false };
   unsigned m_sourceWidth{ 0 };
   unsigned m_sourceHeight{ 0 };
   CRect m_sourceRect{ 0.f, 0.f, 0.f, 0.f };
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/WinRenderer.cpp b/xbmc/cores/VideoPlayer/VideoRenderers/WinRenderer.cpp
index 4664d65857..182841066e 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/WinRenderer.cpp
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/WinRenderer.cpp
@@ -30,7 +30,6 @@
 #include "settings/MediaSettings.h"
 #include "settings/Settings.h"
 #include "threads/SingleLock.h"
-#include "utils/CPUInfo.h"
 #include "utils/log.h"
 #include "utils/win32/gpu_memcpy_sse4.h"
 #include "VideoShaders/WinVideoFilter.h"
@@ -101,7 +100,7 @@ CWinRenderer::CWinRenderer()
   m_processor = nullptr;
   m_neededBuffers = 0;
   m_colorManager.reset(new CColorManager());
-  m_outputShader = nullptr;
+  m_outputShader.reset();
   m_useDithering = CServiceBroker::GetSettings().GetBool("videoscreen.dither");
   m_ditherDepth = CServiceBroker::GetSettings().GetInt("videoscreen.ditherdepth");
 
@@ -362,7 +361,7 @@ void CWinRenderer::UnInit()
     SAFE_DELETE(m_processor);
   }
   SAFE_RELEASE(m_pCLUTView);
-  SAFE_DELETE(m_outputShader);
+  m_outputShader.reset();
 }
 
 void CWinRenderer::Flush()
@@ -541,7 +540,7 @@ void CWinRenderer::UpdatePSVideoFilter()
     // First try the more efficient two pass convolution scaler
     m_scalerShader = new CConvolutionShaderSeparable();
 
-    if (!m_scalerShader->Create(m_scalingMethod, m_outputShader))
+    if (!m_scalerShader->Create(m_scalingMethod, m_outputShader.get()))
     {
       SAFE_DELETE(m_scalerShader);
       CLog::Log(LOGNOTICE, "%s: two pass convolution shader init problem, falling back to one pass.", __FUNCTION__);
@@ -552,7 +551,7 @@ void CWinRenderer::UpdatePSVideoFilter()
     {
       m_scalerShader = new CConvolutionShader1Pass();
 
-      if (!m_scalerShader->Create(m_scalingMethod, m_outputShader))
+      if (!m_scalerShader->Create(m_scalingMethod, m_outputShader.get()))
       {
         SAFE_DELETE(m_scalerShader);
         CGUIDialogKaiToast::QueueNotification(CGUIDialogKaiToast::Error, g_localizeStrings.Get(34400), g_localizeStrings.Get(34401));
@@ -579,7 +578,7 @@ void CWinRenderer::UpdatePSVideoFilter()
   }
 
   m_colorShader = new CYUV2RGBShader();
-  if (!m_colorShader->Create(m_bufferFormat, m_bUseHQScaler ? nullptr : m_outputShader))
+  if (!m_colorShader->Create(m_bufferFormat, m_bUseHQScaler ? nullptr : m_outputShader.get()))
   {
     if (m_bUseHQScaler)
     {
@@ -603,7 +602,9 @@ void CWinRenderer::UpdatePSVideoFilter()
 void CWinRenderer::UpdateVideoFilter()
 {
   bool cmsChanged = m_cmsOn != m_colorManager->IsEnabled()
-                    || m_cmsOn && !m_colorManager->CheckConfiguration(m_cmsToken, m_iFlags);
+                 || m_cmsOn && !m_colorManager->CheckConfiguration(m_cmsToken, m_iFlags);
+  cmsChanged &= m_clutLoaded;
+
   if (m_scalingMethodGui == CMediaSettings::GetInstance().GetCurrentVideoSettings().m_ScalingMethod
    && m_bFilterInitialized && !cmsChanged)
     return;
@@ -623,13 +624,14 @@ void CWinRenderer::UpdateVideoFilter()
 
   if (cmsChanged || !m_outputShader)
   {
-    SAFE_DELETE(m_outputShader);
-    m_outputShader = new COutputShader();
-    if (!m_outputShader->Create(m_CLUTSize, m_pCLUTView, m_useDithering, m_ditherDepth))
+    m_outputShader.reset(new COutputShader());
+    if (!m_outputShader->Create(m_cmsOn, m_useDithering, m_ditherDepth))
     {
       CLog::Log(LOGDEBUG, "%s: Unable to create output shader.", __FUNCTION__);
-      SAFE_DELETE(m_outputShader);
+      m_outputShader.reset();
     }
+    else if (m_pCLUTView && m_CLUTSize)
+      m_outputShader->SetCLUT(m_CLUTSize, m_pCLUTView);
   }
 
   RESOLUTION_INFO res = g_graphicsContext.GetResInfo();
@@ -1159,20 +1161,38 @@ void CWinRenderer::ColorManagmentUpdate()
 
 bool CWinRenderer::LoadCLUT()
 {
+  if (m_outputShader)
+    m_outputShader->SetCLUT(0, nullptr);
+  m_CLUTSize = 0;
   SAFE_RELEASE(m_pCLUTView);
-  // load 3DLUT data
-  uint16_t* pCLUT;
-  if (!m_colorManager->GetVideo3dLut(m_iFlags, &m_cmsToken, &m_CLUTSize, &pCLUT))
-  {
-    CLog::Log(LOGERROR, "%s: Error loading the 3DLUT data.", __FUNCTION__);
-    return false;
-  }
-  // create 3DLUT shader view
-  if (!COutputShader::CreateCLUTView(m_CLUTSize, pCLUT, &m_pCLUTView))
-  {
-    CLog::Log(LOGERROR, "%s: Unable to create 3DLUT texture.", __FUNCTION__);
-    SAFE_DELETE(m_pCLUTView);
-  }
-  free(pCLUT);
+  m_clutLoaded = false;
+
+  auto loadLutTask = Concurrency::create_task([this]{
+    // load 3DLUT data
+    uint16_t* pCLUT;
+    int clutSize;
+    if (!m_colorManager->GetVideo3dLut(m_iFlags, &m_cmsToken, &clutSize, &pCLUT))
+    {
+      CLog::Log(LOGERROR, "%s: Error loading the 3DLUT data.", __FUNCTION__);
+      return 0;
+    }
+    // create 3DLUT shader view
+    bool success = COutputShader::CreateCLUTView(clutSize, pCLUT, &m_pCLUTView);
+    free(pCLUT);
+    if (!success)
+    {
+      CLog::Log(LOGERROR, "%s: Unable to create 3DLUT texture.", __FUNCTION__);
+      SAFE_RELEASE(m_pCLUTView);
+      return 0;
+    }
+    return clutSize;
+  });
+
+  loadLutTask.then([this](int clutSize){
+    m_CLUTSize = clutSize;
+    if (m_outputShader)
+        m_outputShader->SetCLUT(m_CLUTSize, m_pCLUTView);
+    m_clutLoaded = true;
+  });
   return true;
 }
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/WinRenderer.h b/xbmc/cores/VideoPlayer/VideoRenderers/WinRenderer.h
index 8c10a508c7..5f0cae05a9 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/WinRenderer.h
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/WinRenderer.h
@@ -146,6 +146,7 @@ protected:
   bool m_bUseHQScaler;
   bool m_bFilterInitialized;
   bool m_cmsOn{ false };
+  bool m_clutLoaded{ true };
   bool m_useDithering;
 
   unsigned int m_destWidth;
@@ -173,7 +174,7 @@ protected:
   struct SwsContext *m_sw_scale_ctx;
   CYUV2RGBShader* m_colorShader;
   CConvolutionShader* m_scalerShader;
-  COutputShader* m_outputShader;
+  std::unique_ptr<COutputShader> m_outputShader;
   CRenderCapture* m_capture = nullptr;
   std::unique_ptr<CColorManager> m_colorManager;
   ID3D11ShaderResourceView *m_pCLUTView{ nullptr };