[WIN32] add nv12 rendering (software and ps)

5 files changed, 128 insertions, 38 deletions

diff --git a/system/shaders/yuv2rgb_d3d.fx b/system/shaders/yuv2rgb_d3d.fx
index 833bbadd43..1f39734781 100644
--- a/system/shaders/yuv2rgb_d3d.fx
+++ b/system/shaders/yuv2rgb_d3d.fx
@@ -87,10 +87,16 @@ struct PS_OUTPUT
 PS_OUTPUT YUV2RGB( VS_OUTPUT In)
 {
   PS_OUTPUT OUT;
+#if defined(XBMC_YV12)
   float4 YUV = float4(tex2D (YSampler, In.TextureY).x
                     , tex2D (USampler, In.TextureU).x
                     , tex2D (VSampler, In.TextureV).x
                     , 1.0);
+#elif defined(XBMC_NV12)
+  float4 YUV = float4(tex2D (YSampler, In.TextureY).x
+                    , tex2D (USampler, In.TextureU).ra
+                    , 1.0);
+#endif
   OUT.RGBColor = mul(YUV, g_ColorMatrix);
   OUT.RGBColor.a = 1.0;
   return OUT;
diff --git a/xbmc/cores/VideoRenderers/VideoShaders/WinVideoFilter.cpp b/xbmc/cores/VideoRenderers/VideoShaders/WinVideoFilter.cpp
index 5a29583e55..e0f53d1e76 100644
--- a/xbmc/cores/VideoRenderers/VideoShaders/WinVideoFilter.cpp
+++ b/xbmc/cores/VideoRenderers/VideoShaders/WinVideoFilter.cpp
@@ -192,7 +192,7 @@ bool CWinShader::Execute()
 
 //==================================================================================
 
-bool CYUV2RGBShader::Create(bool singlepass, unsigned int sourceWidth, unsigned int sourceHeight)
+bool CYUV2RGBShader::Create(bool singlepass, unsigned int sourceWidth, unsigned int sourceHeight, BufferFormat fmt)
 {
   ReleaseInternal();
 
@@ -202,6 +202,13 @@ bool CYUV2RGBShader::Create(bool singlepass, unsigned int sourceWidth, unsigned
   if (singlepass)
     defines["SINGLEPASS"] = "";
 
+  if (fmt == YV12)
+    defines["XBMC_YV12"] = "";
+  else if (fmt == NV12)
+    defines["XBMC_NV12"] = "";
+  else
+    return false;
+
   CStdString effectString = "special://xbmc/system/shaders/yuv2rgb_d3d.fx";
 
   if(!LoadEffect(effectString, &defines))
@@ -213,12 +220,24 @@ bool CYUV2RGBShader::Create(bool singlepass, unsigned int sourceWidth, unsigned
   m_sourceWidth = sourceWidth;
   m_sourceHeight = sourceHeight;
 
-  if(!m_YUVPlanes[0].Create(m_sourceWidth    , m_sourceHeight    , 1, 0, D3DFMT_L8, D3DPOOL_DEFAULT)
-  || !m_YUVPlanes[1].Create(m_sourceWidth / 2, m_sourceHeight / 2, 1, 0, D3DFMT_L8, D3DPOOL_DEFAULT)
-  || !m_YUVPlanes[2].Create(m_sourceWidth / 2, m_sourceHeight / 2, 1, 0, D3DFMT_L8, D3DPOOL_DEFAULT))
+  if (fmt == YV12)
   {
-    CLog::Log(LOGERROR, __FUNCTION__": Failed to create YUV planes.");
-    return false;
+    if(!m_YUVPlanes[0].Create(m_sourceWidth    , m_sourceHeight    , 1, 0, D3DFMT_L8, D3DPOOL_DEFAULT)
+    || !m_YUVPlanes[1].Create(m_sourceWidth / 2, m_sourceHeight / 2, 1, 0, D3DFMT_L8, D3DPOOL_DEFAULT)
+    || !m_YUVPlanes[2].Create(m_sourceWidth / 2, m_sourceHeight / 2, 1, 0, D3DFMT_L8, D3DPOOL_DEFAULT))
+    {
+      CLog::Log(LOGERROR, __FUNCTION__": Failed to create YUV planes.");
+      return false;
+    }
+  }
+  else if (fmt == NV12)
+  {
+    if(!m_YUVPlanes[0].Create(m_sourceWidth    , m_sourceHeight    , 1, 0, D3DFMT_L8, D3DPOOL_DEFAULT)
+    || !m_YUVPlanes[1].Create(m_sourceWidth / 2, m_sourceHeight / 2, 1, 0, D3DFMT_A8L8, D3DPOOL_DEFAULT))
+    {
+      CLog::Log(LOGERROR, __FUNCTION__": Failed to create YUV planes.");
+      return false;
+    }
   }
 
   return true;
@@ -233,7 +252,7 @@ void CYUV2RGBShader::Render(CRect sourceRect, CRect destRect,
   PrepareParameters(sourceRect, destRect,
                     contrast, brightness, flags);
   UploadToGPU(YUVbuf);
-  SetShaderParameters();
+  SetShaderParameters(YUVbuf);
   Execute();
 }
 
@@ -299,13 +318,14 @@ void CYUV2RGBShader::PrepareParameters(CRect sourceRect,
                          flags);
 }
 
-void CYUV2RGBShader::SetShaderParameters()
+void CYUV2RGBShader::SetShaderParameters(YUVBuffer* YUVbuf)
 {
   m_effect.SetMatrix( "g_ColorMatrix", m_matrix.Matrix());
   m_effect.SetTechnique( "YUV2RGB_T" );
   m_effect.SetTexture( "g_YTexture",  m_YUVPlanes[0] ) ;
   m_effect.SetTexture( "g_UTexture",  m_YUVPlanes[1] ) ;
-  m_effect.SetTexture( "g_VTexture",  m_YUVPlanes[2] ) ;
+  if (YUVbuf->GetActivePlanes() == 3)
+    m_effect.SetTexture( "g_VTexture",  m_YUVPlanes[2] ) ;
 }
 
 void CYUV2RGBShader::ReleaseInternal()
@@ -323,7 +343,7 @@ bool CYUV2RGBShader::UploadToGPU(YUVBuffer* YUVbuf)
   const RECT recthalf = { 0, 0, m_sourceWidth / 2, m_sourceHeight / 2};
   const POINT point = { 0, 0 };
 
-  for (int i = 0; i<3; i++)
+  for (unsigned int i = 0; i<YUVbuf->GetActivePlanes(); i++)
   {
     IDirect3DSurface9 *src, *dest;
     if(FAILED(YUVbuf->planes[i].texture.Get()->GetSurfaceLevel(0, &src)))
diff --git a/xbmc/cores/VideoRenderers/VideoShaders/WinVideoFilter.h b/xbmc/cores/VideoRenderers/VideoShaders/WinVideoFilter.h
index d74fad4bbb..ea3d992da6 100644
--- a/xbmc/cores/VideoRenderers/VideoShaders/WinVideoFilter.h
+++ b/xbmc/cores/VideoRenderers/VideoShaders/WinVideoFilter.h
@@ -76,7 +76,7 @@ private:
 class CYUV2RGBShader : public CWinShader
 {
 public:
-  virtual bool Create(bool singlepass, unsigned int sourceWidth, unsigned int sourceHeight);
+  virtual bool Create(bool singlepass, unsigned int sourceWidth, unsigned int sourceHeight, BufferFormat fmt);
   virtual void Render(CRect sourceRect,
                       CRect destRect,
                       float contrast,
@@ -90,7 +90,7 @@ protected:
                                  float contrast,
                                  float brightness,
                                  unsigned int flags);
-  virtual void SetShaderParameters();
+  virtual void SetShaderParameters(YUVBuffer* YUVbuf);
   virtual void ReleaseInternal();
   virtual bool UploadToGPU(YUVBuffer* YUVbuf);
 
diff --git a/xbmc/cores/VideoRenderers/WinRenderer.cpp b/xbmc/cores/VideoRenderers/WinRenderer.cpp
index 104d283437..357fe341b6 100644
--- a/xbmc/cores/VideoRenderers/WinRenderer.cpp
+++ b/xbmc/cores/VideoRenderers/WinRenderer.cpp
@@ -527,6 +527,10 @@ void CWinRenderer::UpdatePSVideoFilter()
 {
   SAFE_RELEASE(m_scalerShader)
 
+  BufferFormat format;
+  if (CONF_FLAGS_FORMAT_MASK(m_flags) == CONF_FLAGS_FORMAT_YV12)      format = YV12;
+  else if (CONF_FLAGS_FORMAT_MASK(m_flags) == CONF_FLAGS_FORMAT_NV12) format = NV12;
+
   if (m_bUseHQScaler)
   {
     m_scalerShader = new CConvolutionShader();
@@ -552,7 +556,7 @@ void CWinRenderer::UpdatePSVideoFilter()
   if (m_bUseHQScaler)
   {
     m_colorShader = new CYUV2RGBShader();
-    if (!m_colorShader->Create(false, m_sourceWidth, m_sourceHeight))
+    if (!m_colorShader->Create(false, m_sourceWidth, m_sourceHeight, format))
     {
       m_IntermediateTarget.Release();
       SAFE_RELEASE(m_scalerShader)
@@ -564,7 +568,7 @@ void CWinRenderer::UpdatePSVideoFilter()
   if (!m_bUseHQScaler) //fallback from HQ scalers and multipass creation above
   {
     m_colorShader = new CYUV2RGBShader();
-    if (!m_colorShader->Create(true, m_sourceWidth, m_sourceHeight))
+    if (!m_colorShader->Create(true, m_sourceWidth, m_sourceHeight, format))
       SAFE_RELEASE(m_colorShader);
     // we're in big trouble - should fallback on D3D accelerated or sw method
   }
@@ -674,35 +678,46 @@ void CWinRenderer::Render(DWORD flags)
 
 void CWinRenderer::RenderSW()
 {
+  enum PixelFormat format;
+
+  if (CONF_FLAGS_FORMAT_MASK(m_flags) == CONF_FLAGS_FORMAT_YV12)      format = PIX_FMT_YUV420P;
+  else if (CONF_FLAGS_FORMAT_MASK(m_flags) == CONF_FLAGS_FORMAT_NV12) format = PIX_FMT_NV12;
+
   // 1. convert yuv to rgb
   m_sw_scale_ctx = m_dllSwScale->sws_getCachedContext(m_sw_scale_ctx,
-                                                      m_sourceWidth, m_sourceHeight, PIX_FMT_YUV420P,
+                                                      m_sourceWidth, m_sourceHeight, format,
                                                       m_sourceWidth, m_sourceHeight, PIX_FMT_BGRA,
                                                       SWS_FAST_BILINEAR | SwScaleCPUFlags(), NULL, NULL, NULL);
 
   YUVBuffer* buf = (YUVBuffer*)m_VideoBuffers[m_iYV12RenderBuffer];
 
-  D3DLOCKED_RECT srclr[3];
-  if(!(buf->planes[0].texture.LockRect(0, &srclr[0], NULL, D3DLOCK_READONLY))
-  || !(buf->planes[1].texture.LockRect(0, &srclr[1], NULL, D3DLOCK_READONLY))
-  || !(buf->planes[2].texture.LockRect(0, &srclr[2], NULL, D3DLOCK_READONLY)))
-    CLog::Log(LOGERROR, __FUNCTION__" - failed to lock yuv textures into memory");
+  D3DLOCKED_RECT   srclr[MAX_PLANES];
+  uint8_t         *src[MAX_PLANES];
+  int              srcStride[MAX_PLANES];
+
+  for (unsigned int idx = 0; idx < buf->GetActivePlanes(); idx++)
+  {
+    if(!(buf->planes[idx].texture.LockRect(0, &srclr[idx], NULL, D3DLOCK_READONLY)))
+      CLog::Log(LOGERROR, __FUNCTION__" - failed to lock yuv textures into memory");
+    else
+    {
+      src[idx] = (uint8_t*)srclr[idx].pBits;
+      srcStride[idx] = srclr[idx].Pitch;
+    }
+  }
   
   D3DLOCKED_RECT destlr = {0,0};
   if (!m_SWTarget.LockRect(0, &destlr, NULL, D3DLOCK_DISCARD))
     CLog::Log(LOGERROR, __FUNCTION__" - failed to lock swtarget texture into memory");
 
-  uint8_t *src[]  = { (uint8_t*)srclr[0].pBits, (uint8_t*)srclr[1].pBits, (uint8_t*)srclr[2].pBits, 0 };
-  int srcStride[] = { srclr[0].Pitch, srclr[1].Pitch, srclr[2].Pitch, 0 };
   uint8_t *dst[]  = { (uint8_t*) destlr.pBits, 0, 0, 0 };
   int dstStride[] = { destlr.Pitch, 0, 0, 0 };
 
   m_dllSwScale->sws_scale(m_sw_scale_ctx, src, srcStride, 0, m_sourceHeight, dst, dstStride);
 
-  if(!(buf->planes[0].texture.UnlockRect(0))
-  || !(buf->planes[1].texture.UnlockRect(0))
-  || !(buf->planes[2].texture.UnlockRect(0)))
-    CLog::Log(LOGERROR, __FUNCTION__" - failed to unlock yuv textures");
+  for (unsigned int idx = 0; idx < buf->GetActivePlanes(); idx++)
+    if(!(buf->planes[idx].texture.UnlockRect(0)))
+      CLog::Log(LOGERROR, __FUNCTION__" - failed to unlock yuv textures");
 
   if (!m_SWTarget.UnlockRect(0))
     CLog::Log(LOGERROR, __FUNCTION__" - failed to unlock swtarget texture");
@@ -990,7 +1005,11 @@ bool CWinRenderer::CreateYV12Texture(int index)
   {
     YUVBuffer *buf = new YUVBuffer();
 
-    if (!buf->Create(m_sourceWidth, m_sourceHeight))
+    BufferFormat format;
+    if (CONF_FLAGS_FORMAT_MASK(m_flags) == CONF_FLAGS_FORMAT_YV12)      format = YV12;
+    else if (CONF_FLAGS_FORMAT_MASK(m_flags) == CONF_FLAGS_FORMAT_NV12) format = NV12;
+
+    if (!buf->Create(format, m_sourceWidth, m_sourceHeight))
     {
       CLog::Log(LOGERROR, __FUNCTION__" - Unable to create YV12 video texture %i", index);
       return false;
@@ -1085,8 +1104,9 @@ YUVBuffer::~YUVBuffer()
   Release();
 }
 
-bool YUVBuffer::Create(unsigned int width, unsigned int height)
+bool YUVBuffer::Create(BufferFormat format, unsigned int width, unsigned int height)
 {
+  m_format = format;
   m_width = width;
   m_height = height;
 
@@ -1095,17 +1115,36 @@ bool YUVBuffer::Create(unsigned int width, unsigned int height)
   // - Dynamic + D3DPOOL_DEFAULT caused trouble for Intel i3 and some IGP. Bad sync/locking in the driver  I suppose
   // and Present failed every second time for the second video played.
   // - this is what D3D9 does behind the scenes anyway
-  if ( !planes[PLANE_Y].texture.Create(m_width    , m_height    , 1, 0, D3DFMT_L8, D3DPOOL_SYSTEMMEM)
-    || !planes[PLANE_U].texture.Create(m_width / 2, m_height / 2, 1, 0, D3DFMT_L8, D3DPOOL_SYSTEMMEM)
-    || !planes[PLANE_V].texture.Create(m_width / 2, m_height / 2, 1, 0, D3DFMT_L8, D3DPOOL_SYSTEMMEM))
+  switch(m_format)
+  {
+  case YV12:
+    {
+      if ( !planes[PLANE_Y].texture.Create(m_width    , m_height    , 1, 0, D3DFMT_L8, D3DPOOL_SYSTEMMEM)
+        || !planes[PLANE_U].texture.Create(m_width / 2, m_height / 2, 1, 0, D3DFMT_L8, D3DPOOL_SYSTEMMEM)
+        || !planes[PLANE_V].texture.Create(m_width / 2, m_height / 2, 1, 0, D3DFMT_L8, D3DPOOL_SYSTEMMEM))
+        return false;
+      m_activeplanes = 3;
+      break;
+    }
+  case NV12:
+    {
+      if ( !planes[PLANE_Y].texture.Create(m_width    , m_height    , 1, 0, D3DFMT_L8, D3DPOOL_SYSTEMMEM)
+        || !planes[PLANE_UV].texture.Create(m_width / 2, m_height / 2, 1, 0, D3DFMT_A8L8, D3DPOOL_SYSTEMMEM))
+        return false;
+      m_activeplanes = 2;
+      break;
+    }
+  default:
+    m_activeplanes = 0;
     return false;
+  }
 
   return true;
 }
 
 void YUVBuffer::Release()
 {
-  for(unsigned i = 0; i < MAX_PLANES; i++)
+  for(unsigned i = 0; i < m_activeplanes; i++)
   {
     planes[i].texture.Release();
     memset(&planes[i].rect, 0, sizeof(planes[i].rect));
@@ -1114,7 +1153,7 @@ void YUVBuffer::Release()
 
 void YUVBuffer::StartRender()
 {
-  for(unsigned i = 0; i < MAX_PLANES; i++)
+  for(unsigned i = 0; i < m_activeplanes; i++)
   {
     if(planes[i].texture.Get() && planes[i].rect.pBits)
       if (!planes[i].texture.UnlockRect(0))
@@ -1125,7 +1164,7 @@ void YUVBuffer::StartRender()
 
 void YUVBuffer::StartDecode()
 {
-  for(unsigned i = 0; i < MAX_PLANES; i++)
+  for(unsigned i = 0; i < m_activeplanes; i++)
   {
     if(planes[i].texture.Get()
     && planes[i].texture.LockRect(0, &planes[i].rect, NULL, D3DLOCK_DISCARD) == false)
@@ -1138,9 +1177,22 @@ void YUVBuffer::StartDecode()
 
 void YUVBuffer::Clear()
 {
-    memset(planes[PLANE_Y].rect.pBits, 0,   planes[PLANE_Y].rect.Pitch *  m_height);
-    memset(planes[PLANE_U].rect.pBits, 128, planes[PLANE_U].rect.Pitch * (m_height/2));
-    memset(planes[PLANE_V].rect.pBits, 128, planes[PLANE_V].rect.Pitch * (m_height/2));
+  switch(m_format)
+  {
+  case YV12:
+    {
+      memset(planes[PLANE_Y].rect.pBits, 0,   planes[PLANE_Y].rect.Pitch *  m_height);
+      memset(planes[PLANE_U].rect.pBits, 128, planes[PLANE_U].rect.Pitch * (m_height/2));
+      memset(planes[PLANE_V].rect.pBits, 128, planes[PLANE_V].rect.Pitch * (m_height/2));
+      break;
+    }
+  case NV12:
+    {
+      memset(planes[PLANE_Y].rect.pBits, 0,   planes[PLANE_Y].rect.Pitch *  m_height);
+      memset(planes[PLANE_UV].rect.pBits, 128, planes[PLANE_U].rect.Pitch * (m_height/2));
+      break;
+    }
+  }
 }
 
 //==================================
diff --git a/xbmc/cores/VideoRenderers/WinRenderer.h b/xbmc/cores/VideoRenderers/WinRenderer.h
index 28c2f15a1d..2f98184c06 100644
--- a/xbmc/cores/VideoRenderers/WinRenderer.h
+++ b/xbmc/cores/VideoRenderers/WinRenderer.h
@@ -115,11 +115,20 @@ enum RenderMethod
 #define PLANE_Y 0
 #define PLANE_U 1
 #define PLANE_V 2
+#define PLANE_UV 1
 
 #define FIELD_FULL 0
 #define FIELD_TOP 1
 #define FIELD_BOT 2
 
+enum BufferFormat
+{
+  YV12,
+  NV12,
+  YUY2,
+  UYVY
+};
+
 struct SVideoBuffer
 {
   virtual ~SVideoBuffer() {}
@@ -139,17 +148,20 @@ struct SVideoPlane
 struct YUVBuffer : SVideoBuffer
 {
   ~YUVBuffer();
-  bool Create(unsigned int width, unsigned int height);
+  bool Create(BufferFormat format, unsigned int width, unsigned int height);
   virtual void Release();
   virtual void StartDecode();
   virtual void StartRender();
   virtual void Clear();
+  unsigned int GetActivePlanes() { return m_activeplanes; }
 
   SVideoPlane planes[MAX_PLANES];
 
 private:
   unsigned int     m_width;
   unsigned int     m_height;
+  BufferFormat     m_format;
+  unsigned int     m_activeplanes;
 };
 
 struct DXVABuffer : SVideoBuffer