[d3d9] Implement NV12 format conversion

src/d3d9/d3d9_common_texture.cpp

@@ -190,7 +190,10 @@ namespace dxvk {
     const VkExtent3D blockCount = util::computeBlockCount(
       mipExtent, formatInfo.blockSize);
 
+    const uint32_t planeCount = m_mapping.ConversionFormatInfo.MacroPixelSize.depth;
+
     return formatInfo.elementSize
+         * planeCount
          * blockCount.width
          * blockCount.height
          * blockCount.depth;

src/d3d9/d3d9_format.cpp

@@ -406,6 +406,15 @@ namespace dxvk {
         { VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_R,
           VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_R }};
 
+      case D3D9Format::NV12: return {
+        VK_FORMAT_R8_UNORM,
+        VK_FORMAT_UNDEFINED,
+        VK_IMAGE_ASPECT_COLOR_BIT,
+        { VK_COMPONENT_SWIZZLE_IDENTITY, VK_COMPONENT_SWIZZLE_IDENTITY,
+          VK_COMPONENT_SWIZZLE_IDENTITY, VK_COMPONENT_SWIZZLE_IDENTITY },
+        { D3D9ConversionFormat_NV12, { 1u, 1u, 2u }, VK_FORMAT_B8G8R8A8_UNORM }
+      };
+
       case D3D9Format::RAWZ: return {}; // Unsupported
 
       default:

src/d3d9/d3d9_format.h

@@ -134,12 +134,13 @@ namespace dxvk {
     D3D9ConversionFormat_L6V5U5,
     D3D9ConversionFormat_X8L8V8U8,
     D3D9ConversionFormat_A2W10V10U10,
+    D3D9ConversionFormat_NV12,
     D3D9ConversionFormat_Count
   };
 
   struct D3D9_CONVERSION_FORMAT_INFO {
     D3D9ConversionFormat FormatType     = D3D9ConversionFormat_None;
-    VkExtent2D           MacroPixelSize = { 1u, 1u };
+    VkExtent3D           MacroPixelSize = { 1u, 1u, 1u };
     VkFormat             FormatColor    = VK_FORMAT_UNDEFINED;
     VkFormat             FormatSrgb     = VK_FORMAT_UNDEFINED;
   };

src/d3d9/d3d9_format_helpers.cpp

@@ -4,6 +4,7 @@
 #include <d3d9_convert_l6v5u5.h>
 #include <d3d9_convert_x8l8v8u8.h>
 #include <d3d9_convert_a2w10v10u10.h>
+#include <d3d9_convert_nv12.h>
 
 namespace dxvk {
 
@@ -35,6 +36,10 @@ namespace dxvk {
         break;
       }
 
+      case D3D9ConversionFormat_NV12:
+        ConvertGenericFormat(conversionFormat, dstImage, dstSubresource, srcBuffer, VK_FORMAT_R8_UINT, 0);
+        break;
+
       case D3D9ConversionFormat_L6V5U5:
         ConvertGenericFormat(conversionFormat, dstImage, dstSubresource, srcBuffer, VK_FORMAT_R16_UINT, 0);
         break;
@@ -108,6 +113,7 @@ namespace dxvk {
     m_shaders[D3D9ConversionFormat_L6V5U5] = InitShader(d3d9_convert_l6v5u5);
     m_shaders[D3D9ConversionFormat_X8L8V8U8] = InitShader(d3d9_convert_x8l8v8u8);
     m_shaders[D3D9ConversionFormat_A2W10V10U10] = InitShader(d3d9_convert_a2w10v10u10);
+    m_shaders[D3D9ConversionFormat_NV12] = InitShader(d3d9_convert_nv12);
   }
 
 

src/d3d9/meson.build

@@ -6,7 +6,8 @@ d3d9_shaders = files([
   'shaders/d3d9_convert_yuy2_uyvy.comp',
   'shaders/d3d9_convert_l6v5u5.comp',
   'shaders/d3d9_convert_x8l8v8u8.comp',
-  'shaders/d3d9_convert_a2w10v10u10.comp'
+  'shaders/d3d9_convert_a2w10v10u10.comp',
+  'shaders/d3d9_convert_nv12.comp'
 ])
 
 d3d9_src = [

src/d3d9/shaders/d3d9_convert_common.h

@@ -26,4 +26,14 @@ vec4 convertYUV(vec3 yuv) {
   vec3 value = vec4(yuv, 1 / 255.0) * g_yuv_to_rgb;
 
   return vec4(clamp(value, 0, 1), 1);
-}
+}
+
+mat3x3 g_bt703_to_rgb = {
+  { 1.164,  0,          1.793    },
+  { 1.164, -0.213,     -0.533    },
+  { 1.164,  2.112,      0        }
+};
+
+vec4 convertBT_703(vec3 cde) {
+  return vec4(clamp(cde * g_bt703_to_rgb, 0, 1), 1);
+}

src/d3d9/shaders/d3d9_convert_nv12.comp

@@ -0,0 +1,56 @@
+#version 450
+#extension GL_GOOGLE_include_directive : enable
+
+#include "d3d9_convert_common.h"
+
+layout(
+  local_size_x = 8,
+  local_size_y = 8,
+  local_size_z = 1) in;
+
+layout(binding = 0)
+writeonly uniform image2D dst;
+
+layout(binding = 1) uniform usamplerBuffer src;
+
+layout(push_constant)
+uniform u_info_t {
+  uvec2 extent;
+} u_info;
+
+float fetchUnorm(usamplerBuffer source, uint offset) {
+  return unpackUnorm(texelFetch(src, int(offset)).r);
+}
+
+void main() {
+  ivec3 thread_id = ivec3(gl_GlobalInvocationID);
+
+  if (all(lessThan(thread_id.xy, u_info.extent))) {
+    uvec2 pitch = uvec2(u_info.extent.x, u_info.extent.y);
+
+    // Format is:
+    // YYYYYYYYYYYYYYY...
+    // UVUVUVUVUVUVUVU...
+    uint offset = thread_id.x
+                + thread_id.y * pitch.x;
+
+    float c0 = fetchUnorm(src, offset) - (16 / 255.0);
+
+    // Floor .x to the nearest 2, because
+    // UV data is in WORDs, and we want to get the color
+    // for this pixel.
+    // Then divide thread_id.y by 2 because the macropixel
+    // layout for chroma data is [2, 2].
+    offset = (thread_id.x / 2) * 2
+           + thread_id.y / 2 * pitch.x
+           + pitch.x * pitch.y;
+
+    float u = fetchUnorm(src, offset)     - (128 / 255.0);
+    float v = fetchUnorm(src, offset + 1) - (128 / 255.0);
+
+    // The NV12 format seems to use the BT.703 color space.
+    vec4 color0 = convertBT_703(vec3(c0, u, v));
+    
+    imageStore(dst, thread_id.xy, color0);
+  }
+}