Handle chroma co-siting in the D3D11 shaders

2026-02-16 02:30:52 +00:00 · 2025-11-01 22:42:51 -05:00
parent 9b3050514b
commit f67272b153
15 changed files with 129 additions and 132 deletions
--- a/app/resources.qrc
+++ b/app/resources.qrc
@@ -91,9 +91,7 @@
        <file alias="egl_overlay.vert">shaders/egl_overlay.vert</file>
        <file alias="d3d11_vertex.fxc">shaders/d3d11_vertex.fxc</file>
        <file alias="d3d11_overlay_pixel.fxc">shaders/d3d11_overlay_pixel.fxc</file>
-        <file alias="d3d11_genyuv_pixel.fxc">shaders/d3d11_genyuv_pixel.fxc</file>
-        <file alias="d3d11_bt601lim_pixel.fxc">shaders/d3d11_bt601lim_pixel.fxc</file>
-        <file alias="d3d11_bt2020lim_pixel.fxc">shaders/d3d11_bt2020lim_pixel.fxc</file>
+        <file alias="d3d11_yuv420_pixel.fxc">shaders/d3d11_yuv420_pixel.fxc</file>
        <file alias="d3d11_ayuv_pixel.fxc">shaders/d3d11_ayuv_pixel.fxc</file>
        <file alias="d3d11_y410_pixel.fxc">shaders/d3d11_y410_pixel.fxc</file>
        <file alias="vt_renderer.metal">shaders/vt_renderer.metal</file>
--- a/app/shaders/build_hlsl.bat
+++ b/app/shaders/build_hlsl.bat
@@ -1,8 +1,6 @@
 fxc /T vs_4_0_level_9_3 /Fo d3d11_vertex.fxc d3d11_vertex.hlsl

 fxc /T ps_4_0_level_9_3 /Fo d3d11_overlay_pixel.fxc d3d11_overlay_pixel.hlsl
-fxc /T ps_4_0_level_9_3 /Fo d3d11_genyuv_pixel.fxc d3d11_genyuv_pixel.hlsl
-fxc /T ps_4_0_level_9_3 /Fo d3d11_bt601lim_pixel.fxc d3d11_bt601lim_pixel.hlsl
-fxc /T ps_4_0_level_9_3 /Fo d3d11_bt2020lim_pixel.fxc d3d11_bt2020lim_pixel.hlsl
+fxc /T ps_4_0_level_9_3 /Fo d3d11_yuv420_pixel.fxc d3d11_yuv420_pixel.hlsl
 fxc /T ps_4_0_level_9_3 /Fo d3d11_ayuv_pixel.fxc d3d11_ayuv_pixel.hlsl
 fxc /T ps_4_0_level_9_3 /Fo d3d11_y410_pixel.fxc d3d11_y410_pixel.hlsl
--- a/app/shaders/d3d11_ayuv_pixel.fxc
+++ b/app/shaders/d3d11_ayuv_pixel.fxc
--- a/app/shaders/d3d11_bt2020lim_pixel.fxc
+++ b/app/shaders/d3d11_bt2020lim_pixel.fxc
--- a/app/shaders/d3d11_bt2020lim_pixel.hlsl
+++ b/app/shaders/d3d11_bt2020lim_pixel.hlsl
@@ -1,15 +0,0 @@
-#include "d3d11_video_pixel_start.hlsli"
-
-static const min16float3x3 cscMatrix =
-{
-    1.1678, 1.1678, 1.1678,
-    0.0, -0.1879, 2.1481,
-    1.6836, -0.6524, 0.0,
-};
-
-static const min16float3 offsets =
-{
-    64.0 / 1023.0, 512.0 / 1023.0, 512.0 / 1023.0
-};
-
-#include "d3d11_video_pixel_end.hlsli"
--- a/app/shaders/d3d11_bt601lim_pixel.fxc
+++ b/app/shaders/d3d11_bt601lim_pixel.fxc
--- a/app/shaders/d3d11_bt601lim_pixel.hlsl
+++ b/app/shaders/d3d11_bt601lim_pixel.hlsl
@@ -1,15 +0,0 @@
-#include "d3d11_video_pixel_start.hlsli"
-
-static const min16float3x3 cscMatrix =
-{
-    1.1644, 1.1644, 1.1644,
-    0.0, -0.3917, 2.0172,
-    1.5960, -0.8129, 0.0,
-};
-
-static const min16float3 offsets =
-{
-    16.0 / 255.0, 128.0 / 255.0, 128.0 / 255.0
-};
-
-#include "d3d11_video_pixel_end.hlsli"
--- a/app/shaders/d3d11_genyuv_pixel.fxc
+++ b/app/shaders/d3d11_genyuv_pixel.fxc
--- a/app/shaders/d3d11_video_pixel_end.hlsli
+++ b/app/shaders/d3d11_video_pixel_end.hlsli
@@ -2,7 +2,7 @@ min16float4 main(ShaderInput input) : SV_TARGET
 {
    // Clamp the chrominance texcoords to avoid sampling the row of texels adjacent to the alignment padding
    min16float3 yuv = min16float3(luminancePlane.Sample(theSampler, input.tex),
-                                  chrominancePlane.Sample(theSampler, min(input.tex, chromaTexMax.rg)));
+                                  chrominancePlane.Sample(theSampler, min(input.tex + chromaOffset, chromaTexMax.rg)));

    // Subtract the YUV offset for limited vs full range
    yuv -= offsets;
--- a/app/shaders/d3d11_y410_pixel.fxc
+++ b/app/shaders/d3d11_y410_pixel.fxc
--- a/app/shaders/d3d11_yuv420_pixel.fxc
+++ b/app/shaders/d3d11_yuv420_pixel.fxc
--- a/app/shaders/d3d11_yuv420_pixel.hlsl
+++ b/app/shaders/d3d11_yuv420_pixel.hlsl
@@ -4,6 +4,7 @@ cbuffer CSC_CONST_BUF : register(b1)
 {
    min16float3x3 cscMatrix;
    min16float3 offsets;
+    min16float2 chromaOffset;
 };

 #include "d3d11_video_pixel_end.hlsli"
--- a/app/shaders/d3d11_yuv444_pixel_start.hlsli
+++ b/app/shaders/d3d11_yuv444_pixel_start.hlsli
@@ -16,4 +16,5 @@ cbuffer CSC_CONST_BUF : register(b1)
 {
    min16float3x3 cscMatrix;
    min16float3 offsets;
+    min16float2 chromaOffset; // Unused for 4:4:4
 };
--- a/app/streaming/video/ffmpeg-renderers/d3d11va.cpp
+++ b/app/streaming/video/ffmpeg-renderers/d3d11va.cpp
@@ -54,16 +54,20 @@ typedef struct _CSC_CONST_BUF
    // YUV offset values
    float offsets[OFFSETS_ELEMENT_COUNT];

-    // Padding float to be a multiple of 16 bytes
+    // Padding float to end 16-byte boundary
    float padding;
+
+    // Chroma offset values
+    float chromaOffset[2];
+
+    // Padding to final 16-byte boundary
+    float padding2[2];
 } CSC_CONST_BUF, *PCSC_CONST_BUF;
 static_assert(sizeof(CSC_CONST_BUF) % 16 == 0, "Constant buffer sizes must be a multiple of 16");

 static const std::array<const char*, D3D11VARenderer::PixelShaders::_COUNT> k_VideoShaderNames =
 {
-    "d3d11_genyuv_pixel.fxc",
-    "d3d11_bt601lim_pixel.fxc",
-    "d3d11_bt2020lim_pixel.fxc",
+    "d3d11_yuv420_pixel.fxc",
    "d3d11_ayuv_pixel.fxc",
    "d3d11_y410_pixel.fxc",
 };
@@ -541,6 +545,8 @@ bool D3D11VARenderer::initialize(PDECODER_PARAMETERS params)
        D3D11_TEXTURE2D_DESC textureDesc;
        d3d11vaFramesContext->texture_infos->texture->GetDesc(&textureDesc);
        m_TextureFormat = textureDesc.Format;
+        m_TextureWidth = textureDesc.Width;
+        m_TextureHeight = textureDesc.Height;

        if (m_BindDecoderOutputTextures) {
            // Create SRVs for all textures in the decoder pool
@@ -705,14 +711,6 @@ void D3D11VARenderer::bindColorConversion(AVFrame* frame)
    bool yuv444 = (m_DecoderParams.videoFormat & VIDEO_FORMAT_MASK_YUV444);
    int bits = (colorspace == COLORSPACE_REC_2020) ? 10 : 8;

-    // We have purpose-built shaders for the common Rec 601 (SDR) and Rec 2020 (HDR) YUV 4:2:0 cases
-    if (!yuv444 && !fullRange && colorspace == COLORSPACE_REC_601) {
-        m_DeviceContext->PSSetShader(m_VideoPixelShaders[PixelShaders::BT_601_LIMITED_YUV_420].Get(), nullptr, 0);
-    }
-    else if (!yuv444 && !fullRange && colorspace == COLORSPACE_REC_2020) {
-        m_DeviceContext->PSSetShader(m_VideoPixelShaders[PixelShaders::BT_2020_LIMITED_YUV_420].Get(), nullptr, 0);
-    }
-    else {
    if (yuv444) {
        // We'll need to use one of the 4:4:4 shaders for this pixel format
        switch (m_TextureFormat)
@@ -737,13 +735,6 @@ void D3D11VARenderer::bindColorConversion(AVFrame* frame)
        return;
    }

-        if (!yuv444) {
-            SDL_LogWarn(SDL_LOG_CATEGORY_APPLICATION,
-                        "Falling back to generic video pixel shader for %d (%s range)",
-                        colorspace,
-                        fullRange ? "full" : "limited");
-        }
-
    D3D11_BUFFER_DESC constDesc = {};
    constDesc.ByteWidth = sizeof(CSC_CONST_BUF);
    constDesc.Usage = D3D11_USAGE_IMMUTABLE;
@@ -792,6 +783,45 @@ void D3D11VARenderer::bindColorConversion(AVFrame* frame)
        }
    }

+    switch (frame->chroma_location) {
+    default:
+    case AVCHROMA_LOC_LEFT:
+        constBuf.chromaOffset[0] = 0;
+        constBuf.chromaOffset[1] = 0.5;
+        break;
+    case AVCHROMA_LOC_CENTER:
+        constBuf.chromaOffset[0] = 0.5;
+        constBuf.chromaOffset[1] = 0.5;
+        break;
+    case AVCHROMA_LOC_TOPLEFT:
+        constBuf.chromaOffset[0] = 0;
+        constBuf.chromaOffset[1] = 0;
+        break;
+    case AVCHROMA_LOC_TOP:
+        constBuf.chromaOffset[0] = 0.5;
+        constBuf.chromaOffset[1] = 0;
+        break;
+    case AVCHROMA_LOC_BOTTOMLEFT:
+        constBuf.chromaOffset[0] = 0;
+        constBuf.chromaOffset[1] = 1.0;
+        break;
+    case AVCHROMA_LOC_BOTTOM:
+        constBuf.chromaOffset[0] = 0.5;
+        constBuf.chromaOffset[1] = 1.0;
+        break;
+    }
+
+    if (yuv444) {
+        // 4:4:4 has no subsampling
+        constBuf.chromaOffset[0] = 0;
+        constBuf.chromaOffset[1] = 0;
+    }
+    else {
+        // 4:2:0 has 2x2 subsampling
+        constBuf.chromaOffset[0] /= m_TextureWidth / 2;
+        constBuf.chromaOffset[1] /= m_TextureHeight / 2;
+    }
+
    D3D11_SUBRESOURCE_DATA constData = {};
    constData.pSysMem = &constBuf;

@@ -806,7 +836,6 @@ void D3D11VARenderer::bindColorConversion(AVFrame* frame)
                     hr);
        return;
    }
-    }

    m_LastColorSpace = colorspace;
    m_LastFullRange = fullRange;
--- a/app/streaming/video/ffmpeg-renderers/d3d11va.h
+++ b/app/streaming/video/ffmpeg-renderers/d3d11va.h
@@ -28,8 +28,6 @@ public:

    enum PixelShaders {
        GENERIC_YUV_420,
-        BT_601_LIMITED_YUV_420,
-        BT_2020_LIMITED_YUV_420,
        GENERIC_AYUV,
        GENERIC_Y410,
        _COUNT
@@ -72,6 +70,8 @@ private:
    DECODER_PARAMETERS m_DecoderParams;
    int m_TextureAlignment;
    DXGI_FORMAT m_TextureFormat;
+    UINT m_TextureWidth;
+    UINT m_TextureHeight;
    int m_DisplayWidth;
    int m_DisplayHeight;
    int m_LastColorSpace;