gpu: make cull mode and front face dynamic state

src/graphics/host_gpu/renderer/dynamicState.cpp

@@ -125,6 +125,14 @@ void DynamicState::Commit(const GraphicContext& graphics, vk::CommandBuffer cmdb
 			}
 		}
 	}
+	if (dirty_state.cull_mode) {
+		dirty_state.cull_mode = false;
+		cmdbuf.setCullMode(cull_mode);
+	}
+	if (dirty_state.front_face) {
+		dirty_state.front_face = false;
+		cmdbuf.setFrontFace(front_face);
+	}
 	if (dirty_state.blend_constants) {
 		dirty_state.blend_constants = false;
 		cmdbuf.setBlendConstants(blend_constants.data());

src/graphics/host_gpu/renderer/dynamicState.h

@@ -52,6 +52,9 @@ struct DynamicState {
 		bool stencil_back_write_mask    : 1;
 		bool stencil_back_compare_mask  : 1;
 
+		bool cull_mode  : 1;
+		bool front_face : 1;
+
 		bool blend_constants       : 1;
 		bool color_write_enables   : 1;
 		bool line_width            : 1;
@@ -86,6 +89,9 @@ struct DynamicState {
 	uint32_t   stencil_back_write_mask {};
 	uint32_t   stencil_back_compare_mask {};
 
+	vk::CullModeFlags cull_mode {};
+	vk::FrontFace     front_face {};
+
 	std::array<float, 4>                               blend_constants {};
 	std::array<vk::Bool32, RENDER_COLOR_ATTACHMENTS_MAX> color_write_enables {};
 	uint32_t                                           num_color_write_enables {};
@@ -221,6 +227,20 @@ struct DynamicState {
 		}
 	}
 
+	void SetCullMode(const vk::CullModeFlags cull_mode_) {
+		if (cull_mode != cull_mode_) {
+			cull_mode             = cull_mode_;
+			dirty_state.cull_mode = true;
+		}
+	}
+
+	void SetFrontFace(const vk::FrontFace front_face_) {
+		if (front_face != front_face_) {
+			front_face             = front_face_;
+			dirty_state.front_face = true;
+		}
+	}
+
 	void SetBlendConstants(const std::array<float, 4> blend_constants_) {
 		if (blend_constants != blend_constants_) {
 			blend_constants             = blend_constants_;

src/graphics/host_gpu/renderer/pipeline/pipelineCache.cpp

@@ -1003,15 +1003,11 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(const CommandBuffer&
 	}
 	static_params.alpha_to_coverage_enable = ps_active && ctx.GetShaderRegisters().AlphaToMask();
 	const bool rect_list = Prospero::IsRectList(command.GetUserConfig().GetPrimType());
-	static_params.cull_back  = !rect_list && mc.cull_back;
-	static_params.cull_front = !rect_list && mc.cull_front;
-	static_params.face       = mc.face;
 	static_params.provoking_vtx_last = mc.provoking_vtx_last;
 	static_params.polygon_mode =
-	    ResolvePolygonMode(mc, static_params.cull_front, static_params.cull_back);
+	    ResolvePolygonMode(mc, !rect_list && mc.cull_front, !rect_list && mc.cull_back);
 	if (tessellation) {
-		static_params.patch_control_points =
+		static_params.patch_control_points = vs_input_info.tess.input_control_points;
-		    static_cast<uint8_t>(vs_input_info.tess.input_control_points);
 	} else if (!mesh && static_params.topology == vk::PrimitiveTopology::ePatchList) {
 		// Rectangle lists expand through generated tessellation stages.
 		static_params.patch_control_points = 3;

src/graphics/host_gpu/renderer/pipeline/pipelineCache.h

@@ -54,15 +54,12 @@ struct PipelineStaticParameters {
 	bool                       depth_clip_enable        = true;
 	vk::PrimitiveTopology      topology                 = vk::PrimitiveTopology::ePointList;
 	// 0 without tessellation, 3 for a rectangle list, else the input control points.
-	uint8_t                    patch_control_points     = 0;
+	uint32_t                   patch_control_points     = 0;
 	bool                       primitive_restart_enable = false;
 	uint32_t                   samples                  = 1;
 	bool                       sample_shading_enable    = false;
 	bool                       alpha_to_coverage_enable = false;
 	uint32_t                   color_mask[RENDER_COLOR_ATTACHMENTS_MAX]           = {};
-	bool                       cull_front                                         = false;
-	bool                       cull_back                                          = false;
-	bool                       face                                               = false;
 	bool                       provoking_vtx_last                                 = false;
 	vk::PolygonMode            polygon_mode                                       = vk::PolygonMode::eFill;
 	uint8_t                    color_srcblend[RENDER_COLOR_ATTACHMENTS_MAX]       = {};

src/graphics/host_gpu/renderer/pipeline/shaders.cpp

@@ -219,17 +219,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	vk::PipelineViewportStateCreateInfo viewport_state {};
 	viewport_state.pNext = &depth_clip_control;
 
-	vk::CullModeFlags cull_mode = vk::CullModeFlagBits::eNone;
-	if (static_params.cull_back) {
-		cull_mode |= vk::CullModeFlagBits::eBack;
-	}
-	if (static_params.cull_front) {
-		cull_mode |= vk::CullModeFlagBits::eFront;
-	}
-
-	vk::FrontFace front_face =
-	    (static_params.face ? vk::FrontFace::eClockwise : vk::FrontFace::eCounterClockwise);
-
 	vk::PipelineRasterizationDepthClipStateCreateInfoEXT clip_ext {};
 	clip_ext.depthClipEnable = static_params.depth_clip_enable ? VK_TRUE : VK_FALSE;
 
@@ -250,8 +239,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		provoking_vertex.pNext = rasterizer.pNext;
 		rasterizer.pNext = &provoking_vertex;
 	}
-	rasterizer.cullMode  = cull_mode;
-	rasterizer.frontFace = front_face;
 	rasterizer.polygonMode = static_params.polygon_mode;
 	rasterizer.lineWidth = 1.0f;
 
@@ -347,6 +334,8 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	    vk::DynamicState::eViewportWithCount,
 	    vk::DynamicState::eScissorWithCount,
 	    vk::DynamicState::eLineWidth,
+	    vk::DynamicState::eCullMode,
+	    vk::DynamicState::eFrontFace,
 	    vk::DynamicState::eDepthTestEnable,
 	    vk::DynamicState::eDepthWriteEnable,
 	    vk::DynamicState::eDepthCompareOp,

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -403,6 +403,20 @@ static void UpdateGraphicsDynamicState(const CommandBuffer& buffer, DynamicState
 		dynamic_state.SetDepthBias(constant_factor, poly_offset.clamp, slope_factor);
 	}
 
+	// Rect lists are drawn without culling.
+	vk::CullModeFlags cull_mode {};
+	if (!Prospero::IsRectList(buffer.GetUserConfig().GetPrimType())) {
+		if (mode.cull_front) {
+			cull_mode |= vk::CullModeFlagBits::eFront;
+		}
+		if (mode.cull_back) {
+			cull_mode |= vk::CullModeFlagBits::eBack;
+		}
+	}
+	dynamic_state.SetCullMode(cull_mode);
+	dynamic_state.SetFrontFace(mode.face ? vk::FrontFace::eClockwise
+	                                     : vk::FrontFace::eCounterClockwise);
+
 	dynamic_state.SetStencilTestEnabled(depth.stencil_test_enable);
 	if (depth.stencil_test_enable) {
 		const auto& front = depth.stencil_front;

tests/ShaderRecompilerComputeTests.cpp

@@ -18826,6 +18826,10 @@ public:
       cmd.setViewportWithCount(1, &viewport);
       cmd.setScissorWithCount(1, &scissor);
       cmd.setLineWidth(1);
+      cmd.setCullMode(vk::CullModeFlagBits::eNone);
+      cmd.setFrontFace(registers.GetModeControl().face
+                           ? vk::FrontFace::eClockwise
+                           : vk::FrontFace::eCounterClockwise);
       cmd.setDepthTestEnable(depth.depth_test_enable);
       cmd.setDepthWriteEnable(depth.depth_write_enable);
       cmd.setDepthCompareOp(depth.depth_compare_op);
@@ -19023,6 +19027,9 @@ public:
       const std::array stencil_pipelines{
           pipeline(true, 2, 2, false, true).pipeline,
           pipeline(true, 2, 2, false, false).pipeline};
+      Require(name, "dynamic winding pipeline reuse",
+              stencil_pipelines[0] == stencil_pipelines[1],
+              "the winding changed the graphics pipeline");
       const auto check_stencil = [&](const char *label, bool back_face,
                                      uint8_t initial, uint8_t written) {
         vk::ClearValue clear{};
@@ -19655,6 +19662,22 @@ public:
                   "native vertex or instance offset was applied twice");
         }
       }
+      // With back faces culled, the triangle draws for exactly one winding.
+      const auto draws_with_back_cull = [&](bool clockwise) {
+        registers.SetModeControl({.cull_back = true, .face = clockwise});
+        TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), sparse_color.image_id,
+            {vk::ImageAspectFlagBits::eColor, 0, 1, 0, 1}, {});
+        RenderExecutorTestAccess::DrawAuto(executor, scheduler.Current(),
+            {.vertex_count = 3, .instance_count = 1, .first_vertex = 7, .first_instance = 5,
+             .offset_source = DrawOffsetSource::IndirectArgs});
+        const auto pixels = ReadCachedTexel(name, context, sparse_color.image_id,
+                                            {}, {extent, extent, 1});
+        return std::ranges::any_of(pixels, [](u32 word) { return word != 0; });
+      };
+      Require("DynamicCullMode", "back faces culled by winding",
+              draws_with_back_cull(false) != draws_with_back_cull(true),
+              "the draw's cull mode or front face did not reach the host GPU");
+      registers.SetModeControl({});
 
       // Indirect records come from the GPU copy; the CPU backing holds zeros. In each pair the
       // first record draws nothing and the second covers the target, through vertex offset 7
@@ -19800,9 +19823,10 @@ public:
       registers.SetPsInputEna(2);
       registers.SetPsInputAddr(2);
       registers.SetPsInputSettings(0, 0);
+      // Rect lists are drawn without culling.
       const auto check_rect = [&](Prospero::PrimitiveType primitive, const char *rect_name) {
         user_config.SetPrimitiveType(primitive);
-        registers.SetModeControl({});
+        registers.SetModeControl({.cull_front = true, .cull_back = true});
         TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), sparse_color.image_id,
             {vk::ImageAspectFlagBits::eColor, 0, 1, 0, 1}, {});
         RenderExecutorTestAccess::DrawAuto(