Second prototype

src/platform/windows/display_vram.cpp

@@ -425,7 +425,7 @@ namespace platf::dxgi {
           return -1;
         }
 
-        auto draw = [&](auto &input, auto &y_or_yuv_viewports, auto &uv_viewport) {
+        auto draw = [&](auto &input, auto &y_or_yuv_viewports, auto &uv_viewports) {
           device_ctx->PSSetShaderResources(0, 1, &input);
 
           // Draw Y/YUV
@@ -443,20 +443,22 @@ namespace platf::dxgi {
             device_ctx->OMSetRenderTargets(1, &out_UV_rtv, nullptr);
             device_ctx->VSSetShader(convert_UV_vs.get(), nullptr, 0);
             device_ctx->PSSetShader(img.format == DXGI_FORMAT_R16G16B16A16_FLOAT ? convert_UV_fp16_ps.get() : convert_UV_ps.get(), nullptr, 0);
-            device_ctx->RSSetViewports(1, &uv_viewport);
-            device_ctx->Draw(3, 0);
+            viewport_count = recombine_yuv444_into_yuv420 ? 2 : 1;
+            assert(viewport_count <= uv_viewports.size());
+            device_ctx->RSSetViewports(viewport_count, uv_viewports.data());
+            device_ctx->Draw(3 * viewport_count, 0);  // vertex shader will spread vertices across viewports
           }
         };
 
         // Clear render target view(s) once so that the aspect ratio mismatch "bars" appear black
         if (!rtvs_cleared) {
           auto black = create_black_texture_for_rtv_clear();
-          if (black) draw(black, out_Y_or_YUV_viewports_for_clear, out_UV_viewport_for_clear);
+          if (black) draw(black, out_Y_or_YUV_viewports_for_clear, out_UV_viewports_for_clear);
           rtvs_cleared = true;
         }
 
         // Draw captured frame
-        draw(img_ctx.encoder_input_res, out_Y_or_YUV_viewports, out_UV_viewport);
+        draw(img_ctx.encoder_input_res, out_Y_or_YUV_viewports, out_UV_viewports);
 
         // Release encoder mutex to allow capture code to reuse this image
         img_ctx.encoder_mutex->ReleaseSync(0);
@@ -527,6 +529,7 @@ namespace platf::dxgi {
         // VP1-> |UL |VL | <-VP2
         //       |   |   |
         //       +---+---+
+        /*
         out_Y_or_YUV_viewports[0] = { offsetX, offsetY, out_width_f, out_height_f, 0.0f, 1.0f };  // Y plane
         out_Y_or_YUV_viewports[1] = out_Y_or_YUV_viewports[0];  // left half of U plane
         out_Y_or_YUV_viewports[1].TopLeftY += out_height;
@@ -540,12 +543,19 @@ namespace platf::dxgi {
         out_Y_or_YUV_viewports_for_clear[1].Width /= 2;
         out_Y_or_YUV_viewports_for_clear[2] = out_Y_or_YUV_viewports_for_clear[1];  // left half of V plane
         out_Y_or_YUV_viewports_for_clear[2].TopLeftX += out_Y_or_YUV_viewports_for_clear[1].Width;
+        */
 
-        out_UV_viewport = out_Y_or_YUV_viewports[0];
-        out_UV_viewport.TopLeftX = 0;
-        out_UV_viewport.Width = out_width_f - out_Y_or_YUV_viewports[1].Width;
+        out_Y_or_YUV_viewports[0] = { offsetX, offsetY, out_width_f, out_height_f, 0.0f, 1.0f };  // Y plane
+        out_Y_or_YUV_viewports[1] = out_Y_or_YUV_viewports[0];  // left half of U plane
+        out_Y_or_YUV_viewports[1].TopLeftY += out_height;
+
+        out_UV_viewports[0] = { offsetX / 2, offsetY / 2, out_width_f / 2, out_height_f / 2, 0.0f, 1.0f };
+        out_UV_viewports[1] = out_UV_viewports[0];
+        out_UV_viewports[1].TopLeftY += out_height / 2;
 
-        out_UV_viewport_for_clear = { 0, 0, (float) out_width / 2, (float) out_height, 0.0f, 1.0f };
+        // TODO: clear viewports
+        out_Y_or_YUV_viewports_for_clear = out_Y_or_YUV_viewports;
+        out_UV_viewports_for_clear = out_UV_viewports;
       }
       else {
         out_Y_or_YUV_viewports[0] = { offsetX, offsetY, out_width_f, out_height_f, 0.0f, 1.0f };  // Y plane
@@ -560,8 +570,8 @@ namespace platf::dxgi {
         out_Y_or_YUV_viewports_for_clear[2] = out_Y_or_YUV_viewports_for_clear[1];  // V plane
         out_Y_or_YUV_viewports_for_clear[2].TopLeftY += out_height;
 
-        out_UV_viewport = { offsetX / 2, offsetY / 2, out_width_f / 2, out_height_f / 2, 0.0f, 1.0f };
-        out_UV_viewport_for_clear = { 0, 0, (float) out_width / 2, (float) out_height / 2, 0.0f, 1.0f };
+        out_UV_viewports[0] = { offsetX / 2, offsetY / 2, out_width_f / 2, out_height_f / 2, 0.0f, 1.0f };
+        out_UV_viewports_for_clear[0] = { 0, 0, (float) out_width / 2, (float) out_height / 2, 0.0f, 1.0f };
       }
 
       float subsample_offset_in[16 / sizeof(float)] { 1.0f / (float) out_width_f, 1.0f / (float) out_height_f };  // aligned to 16-byte
@@ -1024,7 +1034,7 @@ namespace platf::dxgi {
     ps_t convert_UV_fp16_ps;
 
     std::array<D3D11_VIEWPORT, 3> out_Y_or_YUV_viewports, out_Y_or_YUV_viewports_for_clear;
-    D3D11_VIEWPORT out_UV_viewport, out_UV_viewport_for_clear;
+    std::array<D3D11_VIEWPORT, 2> out_UV_viewports, out_UV_viewports_for_clear;
 
     DXGI_FORMAT format;
 

src_assets/windows/assets/shaders/directx/convert_yuv444in420_nv12_uv_ps.hlsl

@@ -1,3 +1,4 @@
 #include "include/convert_base.hlsl"
 
+#define PROTOTYPE_UV_SAMPLING
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_nv12_uv_ps_linear.hlsl

@@ -1,3 +1,4 @@
 #include "include/convert_linear_base.hlsl"
 
+#define PROTOTYPE_UV_SAMPLING
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_nv12_y_ps.hlsl

@@ -1,4 +1,3 @@
 #include "include/convert_base.hlsl"
 
-#define PLANAR_VIEWPORTS
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_nv12_y_ps_linear.hlsl

@@ -1,4 +1,3 @@
 #include "include/convert_linear_base.hlsl"
 
-#define PLANAR_VIEWPORTS
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_p010_uv_ps.hlsl

@@ -1,4 +1,5 @@
 #include "include/convert_base.hlsl"
 
 #define P010
+#define PROTOTYPE_UV_SAMPLING
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_p010_uv_ps_linear.hlsl

@@ -1,4 +1,5 @@
 #include "include/convert_linear_base.hlsl"
 
 #define P010
+#define PROTOTYPE_UV_SAMPLING
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_p010_uv_ps_perceptual_quantizer.hlsl

@@ -1,4 +1,5 @@
 #include "include/convert_perceptual_quantizer_base.hlsl"
 
 #define P010
+#define PROTOTYPE_UV_SAMPLING
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_p010_y_ps.hlsl

@@ -1,5 +1,4 @@
 #include "include/convert_base.hlsl"
 
-#define PLANAR_VIEWPORTS
 #define P010
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_p010_y_ps_linear.hlsl

@@ -1,5 +1,4 @@
 #include "include/convert_linear_base.hlsl"
 
-#define PLANAR_VIEWPORTS
 #define P010
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_p010_y_ps_perceptual_quantizer.hlsl

@@ -1,5 +1,4 @@
 #include "include/convert_perceptual_quantizer_base.hlsl"
 
-#define PLANAR_VIEWPORTS
 #define P010
 #include "include/convert_yuv444in420_ps_base.hlsl"

src_assets/windows/assets/shaders/directx/convert_yuv444in420_uv_vs.hlsl

@@ -1 +1,29 @@
-#include "include/convert_yuv444in420_vs_base.hlsl"
+cbuffer subsample_offset_cbuffer : register(b0) {
+    float2 subsample_offset;
+};
+
+cbuffer rotate_texture_steps_cbuffer : register(b1) {
+    int rotate_texture_steps;
+};
+
+cbuffer color_matrix_cbuffer : register(b3) {
+    float4 color_vec_y;
+    float4 color_vec_u;
+    float4 color_vec_v;
+    float2 range_y;
+    float2 range_uv;
+};
+
+#define PLANAR_VIEWPORTS
+#define PROTOTYPE_UV_SAMPLING
+#include "include/base_vs.hlsl"
+
+vertex_t main_vs(uint vertex_id : SV_VertexID)
+{
+    vertex_t output = generate_fullscreen_triangle_vertex(vertex_id % 3, subsample_offset, vertex_id / 3, rotate_texture_steps);
+
+    output.viewport = vertex_id / 3;
+    output.color_vec = color_vec_v;
+
+    return output;
+}

src_assets/windows/assets/shaders/directx/convert_yuv444in420_y_vs.hlsl

@@ -1,2 +1,30 @@
+cbuffer rotate_texture_steps_cbuffer : register(b1) {
+    int rotate_texture_steps;
+};
+
+cbuffer color_matrix_cbuffer : register(b3) {
+    float4 color_vec_y;
+    float4 color_vec_u;
+    float4 color_vec_v;
+    float2 range_y;
+    float2 range_uv;
+};
+
 #define PLANAR_VIEWPORTS
-#include "include/convert_yuv444in420_vs_base.hlsl"
+#include "include/base_vs.hlsl"
+
+vertex_t main_vs(uint vertex_id : SV_VertexID)
+{
+    vertex_t output = generate_fullscreen_triangle_vertex(vertex_id % 3, rotate_texture_steps);
+
+    output.viewport = vertex_id / 3;
+
+    if (output.viewport == 0) {
+        output.color_vec = color_vec_y;
+    }
+    else {
+        output.color_vec = color_vec_u;
+    }
+
+    return output;
+}

src_assets/windows/assets/shaders/directx/include/base_vs.hlsl

@@ -4,6 +4,8 @@
 vertex_t generate_fullscreen_triangle_vertex(uint vertex_id, float subsample_offset, int rotate_texture_steps)
 #elif defined(TOPLEFT_SUBSAMPLING)
 vertex_t generate_fullscreen_triangle_vertex(uint vertex_id, float2 subsample_offset, int rotate_texture_steps)
+#elif defined(PROTOTYPE_UV_SAMPLING)
+vertex_t generate_fullscreen_triangle_vertex(uint vertex_id, float2 subsample_offset, bool bottom_half, int rotate_texture_steps)
 #else
 vertex_t generate_fullscreen_triangle_vertex(uint vertex_id, int rotate_texture_steps)
 #endif
@@ -37,6 +39,11 @@ vertex_t generate_fullscreen_triangle_vertex(uint vertex_id, int rotate_texture_
 #elif defined (TOPLEFT_SUBSAMPLING)
     output.tex_right_left_top = float3(tex_coord.x, tex_coord.x - subsample_offset.x, tex_coord.y - subsample_offset.y);
     output.tex_right_left_bottom = float3(tex_coord.x, tex_coord.x - subsample_offset.x, tex_coord.y);
+#elif defined(PROTOTYPE_UV_SAMPLING)
+    output.tex_right_left_center = float3(tex_coord.x + subsample_offset.x / 2,
+                                          tex_coord.x - subsample_offset.x / 2,
+                                          bottom_half ? tex_coord.y + subsample_offset.y / 2 :
+                                                        tex_coord.y - subsample_offset.y / 2);
 #else
     output.tex_coord = tex_coord;
 #endif

src_assets/windows/assets/shaders/directx/include/base_vs_types.hlsl

@@ -1,7 +1,7 @@
 struct vertex_t
 {
     float4 viewpoint_pos : SV_Position;
-#if defined(LEFT_SUBSAMPLING)
+#if defined(LEFT_SUBSAMPLING) || defined(PROTOTYPE_UV_SAMPLING)
     float3 tex_right_left_center : TEXCOORD;
 #elif defined (TOPLEFT_SUBSAMPLING)
     float3 tex_right_left_top : TEXCOORD;

src_assets/windows/assets/shaders/directx/include/convert_yuv444in420_ps_base.hlsl

@@ -1,51 +1,42 @@
 Texture2D image : register(t0);
 SamplerState def_sampler : register(s0);
 
-#ifndef PLANAR_VIEWPORTS
-cbuffer color_matrix_cbuffer : register(b0) {
-    float4 color_vec_y;
-    float4 color_vec_u;
-    float4 color_vec_v;
-    float2 range_y;
-    float2 range_uv;
-};
-#endif
-
+#define PLANAR_VIEWPORTS
 #include "include/base_vs_types.hlsl"
 
-#ifdef PLANAR_VIEWPORTS
-uint main_ps(vertex_t input) : SV_Target
-#else
+#ifdef PROTOTYPE_UV_SAMPLING
 uint2 main_ps(vertex_t input) : SV_Target
+#else
+uint main_ps(vertex_t input) : SV_Target
 #endif
 {
-    //       Y         U     V
-    //       +-------+ +---+ +---+
-    //       |       | |   | |   |
-    // VP0-> |   Y   | |UR | |VR |
-    //       |       | |   | |   |
-    //       +---+---+ +---+ +---+
-    //       |   |   |
-    // VP1-> |UL |VL | <-VP2
-    //       |   |   |
-    //       +---+---+
-
-    float3 rgb = CONVERT_FUNCTION(image.Sample(def_sampler, input.tex_coord, 0).rgb);
-
-#ifdef PLANAR_VIEWPORTS
-    float y = dot(input.color_vec.xyz, rgb) + input.color_vec.w;
+    //     Y       U     V
+    // +-------+ +---+ +---+
+    // |       | |V0 | |V1 |
+    // |   Y   | +---+ +---+
+    // |       | |V2 | |V3 |
+    // +-------+ +---+ +---+
+    // |       |
+    // |   U   |
+    // |       |
+    // +-------+
+#ifdef PROTOTYPE_UV_SAMPLING
+    float3 rgb_left = CONVERT_FUNCTION(image.Sample(def_sampler, input.tex_right_left_center.yz).rgb);
+    float3 rgb_right = CONVERT_FUNCTION(image.Sample(def_sampler, input.tex_right_left_center.xz).rgb);
+    uint2 vv = uint2(dot(input.color_vec.xyz, rgb_left) + input.color_vec.w,
+                     dot(input.color_vec.xyz, rgb_right) + input.color_vec.w);
 #ifdef P010
-    return uint(y) << 6;
+    return vv << 6;
 #else
-    return uint(y);
+    return vv;
 #endif
 #else
-    float u = dot(color_vec_u.xyz, rgb) + color_vec_u.w;
-    float v = dot(color_vec_v.xyz, rgb) + color_vec_v.w;
+    float3 rgb = CONVERT_FUNCTION(image.Sample(def_sampler, input.tex_coord, 0).rgb);
+    uint yu = dot(input.color_vec.xyz, rgb) + input.color_vec.w;
 #ifdef P010
-    return uint2(u, v) << 6;
+    return yu << 6;
 #else
-    return uint2(u, v);
+    return yu;
 #endif
 #endif
 }