cnq3/code/renderer/grp_main.cpp

/*
===========================================================================
Copyright (C) 2022-2023 Gian 'myT' Schellenbaum

This file is part of Challenge Quake 3 (CNQ3).

Challenge Quake 3 is free software; you can redistribute it
and/or modify it under the terms of the GNU General Public License as
published by the Free Software Foundation; either version 2 of the License,
or (at your option) any later version.

Challenge Quake 3 is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with Challenge Quake 3. If not, see <https://www.gnu.org/licenses/>.
===========================================================================
*/
// Gameplay Rendering Pipeline - main interface


#include "grp_local.h"
#include "uber_shaders.h"
#include "hlsl/uber_shader.h"
#include "hlsl/complete_uber_vs.h"
#include "hlsl/complete_uber_ps.h"
#include "../client/cl_imgui.h"


GRP grp;

static const ShaderByteCode vertexShaderByteCodes[8] =
{
	ShaderByteCode(g_vs_1),
	ShaderByteCode(g_vs_2),
	ShaderByteCode(g_vs_3),
	ShaderByteCode(g_vs_4),
	ShaderByteCode(g_vs_5),
	ShaderByteCode(g_vs_6),
	ShaderByteCode(g_vs_7),
	ShaderByteCode(g_vs_8)
};

#define PS(Data) #Data,
static const char* uberPixelShaderStateStrings[] =
{
	UBER_SHADER_PS_LIST(PS)
};
#undef PS

#define PS(Data) ShaderByteCode(g_ps_##Data),
static const ShaderByteCode uberPixelShaderByteCodes[] =
{
	UBER_SHADER_PS_LIST(PS)
};
#undef PS

#define PS(Data) 1 +
static const uint32_t uberPixelShaderCacheSize = UBER_SHADER_PS_LIST(PS) 0;
#undef PS

static UberPixelShaderState uberPixelShaderStates[uberPixelShaderCacheSize];


static ImPlotPoint FrameTimeGetter(int index, void*)
{
	const FrameStats& fs = grp.frameStats;
	const int realIndex = (fs.frameIndex + index) % fs.frameCount;
	const float value = fs.p2pMS[realIndex];

	ImPlotPoint p;
	p.x = index;
	p.y = value;

	return p;
}

static void UpdateAnimatedImage(image_t* image, int w, int h, const byte* data, qbool dirty)
{
	if(w != image->width || h != image->height)
	{
		// @TODO: ?
		/*image->width = w;
		image->height = h;
		CreateTexture(&d3d.textures[image->texnum], image, 1, w, h);
		GAL_UpdateTexture(image, 0, 0, 0, w, h, data);*/
	}
	else if(dirty)
	{
		// @TODO: ?
		//GAL_UpdateTexture(image, 0, 0, 0, w, h, data);
	}
}

const image_t* GetBundleImage(const textureBundle_t& bundle)
{
	return R_UpdateAndGetBundleImage(&bundle, &UpdateAnimatedImage);
}

uint32_t GetSamplerIndex(textureWrap_t wrap, TextureFilter::Id filter, uint32_t minLOD)
{
	Q_assert((uint32_t)wrap < TW_COUNT);
	Q_assert((uint32_t)filter < TextureFilter::Count);

	const uint32_t index =
		(uint32_t)filter +
		(uint32_t)TextureFilter::Count * (uint32_t)wrap +
		(uint32_t)TextureFilter::Count * (uint32_t)TW_COUNT * minLOD;

	return index;
}

uint32_t GetSamplerIndex(const image_t* image)
{
	TextureFilter::Id filter = TextureFilter::Anisotropic;
	if(r_lego->integer &&
		grp.renderMode == RenderMode::World &&
		(image->flags & (IMG_LMATLAS | IMG_EXTLMATLAS | IMG_NOPICMIP)) == 0)
	{
		filter = TextureFilter::Point;
	}
	else if((image->flags & IMG_NOAF) != 0 ||
		grp.renderMode != RenderMode::World)
	{
		filter = TextureFilter::Linear;
	}

	int minLOD = 0;
	if(grp.renderMode == RenderMode::World &&
		(image->flags & IMG_NOPICMIP) == 0)
	{
		minLOD = Com_ClampInt(0, MaxTextureMips - 1, r_picmip->integer);
	}

	return GetSamplerIndex(image->wrapClampMode, filter, (uint32_t)minLOD);
}

static bool IsCommutativeBlendState(unsigned int stateBits)
{
	const unsigned int blendStates[] =
	{
		GLS_SRCBLEND_ONE | GLS_DSTBLEND_ONE, // additive
		GLS_SRCBLEND_DST_COLOR | GLS_DSTBLEND_ZERO, // modulate
		GLS_SRCBLEND_ZERO | GLS_DSTBLEND_SRC_COLOR, // modulate
		GLS_SRCBLEND_SRC_ALPHA | GLS_DSTBLEND_ONE // pre-multiplied alpha blend
	};

	const unsigned int blendBits = stateBits & GLS_BLEND_BITS;
	for(int b = 0; b < ARRAY_LEN(blendStates); ++b)
	{
		if(blendBits == blendStates[b])
		{
			return true;
		}
	}

	return false;
}

static cullType_t GetMirrorredCullType(cullType_t cullType)
{
	switch(cullType)
	{
		case CT_BACK_SIDED: return CT_FRONT_SIDED;
		case CT_FRONT_SIDED: return CT_BACK_SIDED;
		default: return CT_TWO_SIDED;
	}
}


void FrameStats::EndFrame()
{
	frameCount = min(frameCount + 1, (int)MaxFrames);
	frameIndex = (frameIndex + 1) % MaxFrames;
	Com_StatsFromArray(p2pMS, frameCount, temp, &p2pStats);
}


void RenderPassStats::EndFrame(uint32_t cpu, uint32_t gpu)
{
	static uint32_t tempSamples[MaxStatsFrameCount];
	samplesCPU[index] = cpu;
	samplesGPU[index] = gpu;
	count = min(count + 1, (uint32_t)MaxStatsFrameCount);
	index = (index + 1) % MaxStatsFrameCount;
	Com_StatsFromArray((const int*)samplesCPU, count, (int*)tempSamples, &statsCPU);
	Com_StatsFromArray((const int*)samplesGPU, count, (int*)tempSamples, &statsGPU);
}


void GRP::Init()
{
	firstInit = RHI::Init();

	if(firstInit)
	{
		RootSignatureDesc desc("main");
		desc.usingVertexBuffers = true;
		desc.samplerCount = ARRAY_LEN(samplers);
		desc.samplerVisibility = ShaderStages::PixelBit;
		desc.genericVisibility = ShaderStages::PixelBit;
		desc.AddRange(DescriptorType::Texture, 0, MAX_DRAWIMAGES * 2);
		desc.AddRange(DescriptorType::RWBuffer, MAX_DRAWIMAGES * 2, 1);
		rootSignatureDesc = desc;
		rootSignature = CreateRootSignature(desc);

		descriptorTable = CreateDescriptorTable(DescriptorTableDesc("game textures", rootSignature));

		desc.name = "world";
		desc.usingVertexBuffers = true;
		desc.constants[ShaderStage::Vertex].byteCount = sizeof(WorldVertexRC);
		desc.constants[ShaderStage::Pixel].byteCount = sizeof(WorldPixelRC);
		desc.samplerVisibility = ShaderStages::PixelBit;
		desc.genericVisibility = ShaderStages::VertexBit | ShaderStages::PixelBit;
		uberRootSignature = CreateRootSignature(desc);

		for(uint32_t i = 0; i < uberPixelShaderCacheSize; ++i)
		{
			if(!ParseUberPixelShaderState(uberPixelShaderStates[i], uberPixelShaderStateStrings[i]))
			{
				Q_assert(!"ParseUberPixelShaderState failed!");
			}
		}
	}

	// we recreate the samplers on every vid_restart to create the right level
	// of anisotropy based on the latched CVar
	for(uint32_t w = 0; w < TW_COUNT; ++w)
	{
		for(uint32_t f = 0; f < TextureFilter::Count; ++f)
		{
			for(uint32_t m = 0; m < MaxTextureMips; ++m)
			{
				const textureWrap_t wrap = (textureWrap_t)w;
				const TextureFilter::Id filter = (TextureFilter::Id)f;
				const uint32_t s = GetSamplerIndex(wrap, filter, m);
				SamplerDesc desc(wrap, filter, (float)m);
				desc.shortLifeTime = true;
				samplers[s] = CreateSampler(desc);
			}
		}
	}

	// update our descriptor table with the new sampler descriptors
	{
		DescriptorTableUpdate update;
		update.SetSamplers(ARRAY_LEN(samplers), samplers);
		UpdateDescriptorTable(descriptorTable, update);
	}

	textureIndex = 0;
	psoCount = 1; // we treat index 0 as invalid

	{
		switch(r_rtColorFormat->integer)
		{
			case RTCF_R10G10B10A2:
				renderTargetFormat = TextureFormat::R10G10B10A2_UNorm;
				break;
			case RTCF_R16G16B16A16:
				renderTargetFormat = TextureFormat::RGBA64_Float;
				break;
			case RTCF_R8G8B8A8:
			default:
				renderTargetFormat = TextureFormat::RGBA32_UNorm;
				break;
		}

		TextureDesc desc("render target", glConfig.vidWidth, glConfig.vidHeight);
		desc.initialState = ResourceStates::RenderTargetBit;
		desc.allowedState = ResourceStates::RenderTargetBit | ResourceStates::PixelShaderAccessBit;
		Vector4Clear(desc.clearColor);
		desc.usePreferredClearValue = true;
		desc.committedResource = true;
		desc.format = renderTargetFormat;
		desc.shortLifeTime = true;
		renderTarget = RHI::CreateTexture(desc);
	}

	ui.Init();
	world.Init();
	mipMapGen.Init();
	imgui.Init();
	post.Init();
	post.SetToneMapInput(renderTarget);
	smaa.Init(); // must be after post

	firstInit = false;
}

void GRP::ShutDown(bool fullShutDown)
{
	RHI::ShutDown(fullShutDown);
}

void GRP::BeginFrame()
{
	renderPasses[tr.frameCount % FrameCount].count = 0;

	R_SetColorMappings();

	smaa.Update();

	// have it be first to we can use ImGUI in the other components too
	grp.imgui.SafeBeginFrame();

	RHI::BeginFrame();
	ui.BeginFrame();
	world.BeginFrame();

	const float clearColor[4] = { 0.0f, 0.0f, 0.0f, 0.0f };
	const TextureBarrier barrier(renderTarget, ResourceStates::RenderTargetBit);
	CmdBarrier(1, &barrier);
	CmdClearColorTarget(renderTarget, clearColor);

	// nothing is bound to the command list yet!
	renderMode = RenderMode::None;

	frameSeed = (float)rand() / (float)RAND_MAX;
}

void GRP::EndFrame()
{
	DrawGUI();
	R_DrawGUI();
	imgui.Draw();
	post.Draw();
	world.EndFrame();
	RHI::EndFrame();

	if(rhie.presentToPresentUS > 0)
	{
		frameStats.p2pMS[frameStats.frameIndex] = (float)rhie.presentToPresentUS / 1000.0f;
		frameStats.EndFrame();
	}
	else
	{
		frameStats.skippedFrames++;
	}

	if(backEnd.renderFrame)
	{
		Sys_V_EndFrame();
	}
}

void GRP::CreateTexture(image_t* image, int mipCount, int width, int height)
{
	TextureDesc desc(image->name, width, height, mipCount);
	desc.committedResource = width * height >= (1 << 20);
	desc.shortLifeTime = true;
	if(mipCount > 1)
	{
		desc.allowedState |= ResourceStates::UnorderedAccessBit; // for mip-map generation
	}

	image->texture = ::RHI::CreateTexture(desc);
	image->textureIndex = RegisterTexture(image->texture);
}

void GRP::UpoadTextureAndGenerateMipMaps(image_t* image, const byte* data)
{
	MappedTexture texture;
	RHI::BeginTextureUpload(texture, image->texture);
	for(uint32_t r = 0; r < texture.rowCount; ++r)
	{
		memcpy(texture.mappedData + r * texture.dstRowByteCount, data + r * texture.srcRowByteCount, texture.srcRowByteCount);
	}
	RHI::EndTextureUpload();

	mipMapGen.GenerateMipMaps(image->texture);
}

void GRP::BeginTextureUpload(MappedTexture& mappedTexture, image_t* image)
{
	RHI::BeginTextureUpload(mappedTexture, image->texture);
}

void GRP::EndTextureUpload()
{
	RHI::EndTextureUpload();
}

void GRP::ProcessWorld(world_t& world_)
{
	world.ProcessWorld(world_);
}

void GRP::ProcessModel(model_t& model)
{
	// @TODO: !!!
	//__debugbreak();
}

void GRP::ProcessShader(shader_t& shader)
{
	shader.numPipelines = 0;
	if(shader.numStages < 1)
	{
		return;
	}

	// @TODO: GLS_POLYMODE_LINE

	const bool clampDepth = r_depthClamp->integer != 0 || shader.isSky;

	if(shader.isOpaque)
	{
		Q_assert(IsDepthFadeEnabled(shader) == false);

		// @TODO: fix up cache.stageStateBits[0] based on depth state from follow-up states
		CachedPSO cache = {};
		cache.desc.depthFade = false;
		cache.desc.polygonOffset = !!shader.polygonOffset;
		cache.desc.clampDepth = clampDepth;
		cache.stageStateBits[0] = shader.stages[0]->stateBits & (~GLS_POLYMODE_LINE);
		for(int s = 1; s < shader.numStages; ++s)
		{
			cache.stageStateBits[s] = shader.stages[s]->stateBits & (GLS_BLEND_BITS | GLS_ATEST_BITS);
		}
		cache.stageCount = shader.numStages;

		cache.desc.cullType = shader.cullType;
		shader.pipelines[0].pipeline = CreatePSO(cache, shader.name);
		cache.desc.cullType = GetMirrorredCullType(shader.cullType);
		shader.pipelines[0].mirrorPipeline = CreatePSO(cache, va("%s mirror", shader.name));
		shader.pipelines[0].firstStage = 0;
		shader.pipelines[0].numStages = shader.numStages;
		shader.numPipelines = 1;
	}
	else
	{
		CachedPSO cache = {};
		
		cache.desc.depthFade = IsDepthFadeEnabled(shader);
		cache.desc.polygonOffset = !!shader.polygonOffset;
		cache.desc.clampDepth = clampDepth;
		cache.stageCount = 0;

		unsigned int prevStateBits = 0xFFFFFFFF;
		int firstStage = 0;
		for(int s = 0; s < shader.numStages; ++s)
		{
			const unsigned int currStateBits = shader.stages[s]->stateBits & (~GLS_POLYMODE_LINE);
			if(cache.stageCount > 0)
			{
				// we could combine AT/DW in some circumstances, but we don't care to for now
				const bool cantCombine = (shader.stages[s]->stateBits & (GLS_ATEST_BITS | GLS_DEPTHMASK_TRUE)) != 0;
				if(currStateBits == prevStateBits &&
					!cantCombine &&
					IsCommutativeBlendState(currStateBits))
				{
					cache.stageStateBits[cache.stageCount++] = currStateBits;
				}
				else
				{
					pipeline_t& p = shader.pipelines[shader.numPipelines++];
					cache.desc.cullType = shader.cullType;
					p.pipeline = CreatePSO(cache, va("%s #%d", shader.name, shader.numPipelines));
					cache.desc.cullType = GetMirrorredCullType(shader.cullType);
					p.mirrorPipeline = CreatePSO(cache, va("%s #%d mirror", shader.name, shader.numPipelines));
					p.firstStage = firstStage;
					p.numStages = cache.stageCount;
					cache.stageStateBits[0] = currStateBits;
					cache.stageCount = 1;
					firstStage = s;
				}
			}
			else
			{
				cache.stageStateBits[0] = currStateBits;
				cache.stageCount = 1;
			}
			prevStateBits = currStateBits;
		}

		if(cache.stageCount > 0)
		{
			pipeline_t& p = shader.pipelines[shader.numPipelines++];
			cache.desc.cullType = shader.cullType;
			p.pipeline = CreatePSO(cache, va("%s #%d", shader.name, shader.numPipelines));
			cache.desc.cullType = GetMirrorredCullType(shader.cullType);
			p.mirrorPipeline = CreatePSO(cache, va("%s #%d mirror", shader.name, shader.numPipelines));
			p.firstStage = firstStage;
			p.numStages = cache.stageCount;
		}
	}
}

uint32_t GRP::RegisterTexture(HTexture htexture)
{
	const uint32_t index = textureIndex++;

	DescriptorTableUpdate update;
	update.SetTextures(1, &htexture, index);
	UpdateDescriptorTable(descriptorTable, update);

	return index;
}

uint32_t GRP::BeginRenderPass(const char* name, float r, float g, float b)
{
	RenderPassFrame& f = renderPasses[tr.frameCount % FrameCount];
	if(f.count >= ARRAY_LEN(f.passes))
	{
		Q_assert(0);
		return UINT32_MAX;
	}

	CmdBeginDebugLabel(name, r, g, b);

	const uint32_t index = f.count++;
	RenderPassQueries& q = f.passes[index];
	Q_strncpyz(q.name, name, sizeof(q.name));
	q.cpuStartUS = Sys_Microseconds();
	q.queryIndex = CmdBeginDurationQuery();

	return index;
}

void GRP::EndRenderPass(uint32_t index)
{
	RenderPassFrame& f = renderPasses[tr.frameCount % FrameCount];
	if(index >= f.count)
	{
		Q_assert(0);
		return;
	}

	CmdEndDebugLabel();

	RenderPassQueries& q = f.passes[index];
	q.cpuDurationUS = (uint32_t)(Sys_Microseconds() - q.cpuStartUS);
	CmdEndDurationQuery(q.queryIndex);
}

void GRP::DrawGUI()
{
	uint32_t durations[MaxDurationQueries];
	GetDurations(durations);

	wholeFrameStats.EndFrame(rhie.renderToPresentUS, durations[0]);

	const RenderPassFrame& currFrame = renderPasses[(tr.frameCount % FrameCount) ^ 1];
	RenderPassFrame& tempFrame = tempRenderPasses;

	// see if the render pass list is the same as the previous frame's
	bool sameRenderPass = true;
	if(currFrame.count == tempRenderPasses.count)
	{
		for(uint32_t p = 0; p < currFrame.count; ++p)
		{
			if(Q_stricmp(currFrame.passes[p].name, tempRenderPasses.passes[p].name) != 0)
			{
				sameRenderPass = false;
				break;
			}
		}
	}
	else
	{
		sameRenderPass = false;
	}

	// write out the displayed timings into the temp buffer
	tempFrame.count = currFrame.count;
	if(sameRenderPass)
	{
		for(uint32_t p = 0; p < currFrame.count; ++p)
		{
			const uint32_t index = currFrame.passes[p].queryIndex;
			if(index < MaxDurationQueries)
			{
				renderPassStats[p].EndFrame(currFrame.passes[p].cpuDurationUS, durations[index]);
				tempFrame.passes[p].gpuDurationUS = renderPassStats[p].statsGPU.median;
				tempFrame.passes[p].cpuDurationUS = renderPassStats[p].statsCPU.median;
			}
		}
	}
	else
	{
		for(uint32_t p = 0; p < currFrame.count; ++p)
		{
			const uint32_t index = currFrame.passes[p].queryIndex;
			if(index < MaxDurationQueries)
			{
				tempFrame.passes[p].gpuDurationUS = durations[index];
				tempFrame.passes[p].cpuDurationUS = currFrame.passes[p].cpuDurationUS;
			}
		}
	}

	static bool breakdownActive = false;
	ToggleBooleanWithShortcut(breakdownActive, ImGuiKey_F);
	GUI_AddMainMenuItem(GUI_MainMenu::Perf, "Frame breakdown", "Ctrl+F", &breakdownActive);
	if(breakdownActive)
	{
		if(ImGui::Begin("Frame breakdown", &breakdownActive, ImGuiWindowFlags_AlwaysAutoResize))
		{
			if(BeginTable("Frame breakdown", 3))
			{
				TableHeader(3, "Pass", "GPU [us]", "CPU [us]");

				TableRow(3, "Whole frame",
					va("%d", (int)wholeFrameStats.statsGPU.median),
					va("%d", (int)wholeFrameStats.statsCPU.median));

				for(uint32_t p = 0; p < currFrame.count; ++p)
				{
					const RenderPassQueries& rp = tempFrame.passes[p];
					if(rp.queryIndex < MaxDurationQueries)
					{
						TableRow(3, rp.name,
							va("%d", (int)rp.gpuDurationUS),
							va("%d", (int)rp.cpuDurationUS));
					}
				}

				ImGui::EndTable();
			}

			ImGui::Text("PSO count: %d", (int)grp.psoCount);
			ImGui::Text("PSO changes: %d", (int)grp.world.psoChangeCount);
		}
		ImGui::End();
	}

	// save the current render pass list in the temp buffer
	memcpy(&tempFrame, &currFrame, sizeof(tempFrame));

	static bool frameTimeActive = false;
	GUI_AddMainMenuItem(GUI_MainMenu::Perf, "Frame stats", NULL, &frameTimeActive);
	if(frameTimeActive)
	{
		if(ImGui::Begin("Frame stats", &frameTimeActive, ImGuiWindowFlags_AlwaysAutoResize))
		{
			if(BeginTable("Frame stats", 2))
			{
				const FrameStats& fs = frameStats;
				const stats_t& s = fs.p2pStats;
				TableRow2("Skipped frames", fs.skippedFrames);
				TableRow2("Frame time target", rhie.targetFrameDurationMS);
				TableRow2("Frame time average", s.average);
				TableRow2("Frame time std dev.", s.stdDev);
				TableRow2("Input to render", (float)rhie.inputToRenderUS / 1000.0f);
				TableRow2("Input to present", (float)rhie.inputToPresentUS / 1000.0f);

				ImGui::EndTable();
			}
		}
		ImGui::End();
	}

	static bool graphsActive = false;
	ToggleBooleanWithShortcut(graphsActive, ImGuiKey_G);
	GUI_AddMainMenuItem(GUI_MainMenu::Perf, "Frame time graphs", "Ctrl+G", &graphsActive);
	if(graphsActive)
	{
		const int windowFlags =
			ImGuiWindowFlags_NoDecoration |
			ImGuiWindowFlags_NoBackground |
			ImGuiWindowFlags_NoMove;
		ImGui::SetNextWindowSize(ImVec2(glConfig.vidWidth, glConfig.vidHeight / 2), ImGuiCond_Always);
		ImGui::SetNextWindowPos(ImVec2(0, glConfig.vidHeight / 2), ImGuiCond_Always);
		if(ImGui::Begin("Frame time graphs", &graphsActive, windowFlags))
		{
			const FrameStats& fs = frameStats;
			const double target = (double)rhie.targetFrameDurationMS;

			static bool autoFit = false;
			ImGui::Checkbox("Auto-fit", &autoFit);

			if(ImPlot::BeginPlot("Frame Times", ImVec2(-1, -1), ImPlotFlags_NoInputs))
			{
				const int axisFlags = 0; // ImPlotAxisFlags_NoTickLabels
				const int axisFlagsY = axisFlags | (autoFit ? ImPlotAxisFlags_AutoFit : 0);
				ImPlot::SetupAxes(NULL, NULL, axisFlags, axisFlagsY);
				ImPlot::SetupAxisLimits(ImAxis_X1, 0, FrameStats::MaxFrames, ImGuiCond_Always);
				if(!autoFit)
				{
					ImPlot::SetupAxisLimits(ImAxis_Y1, max(target - 2.0, 0.0), target + 2.0, ImGuiCond_Always);
				}

				ImPlot::SetNextFillStyle(IMPLOT_AUTO_COL, 1.0f);
				ImPlot::SetNextLineStyle(IMPLOT_AUTO_COL, 1.0f);
				ImPlot::PlotInfLines("Target", &target, 1, ImPlotInfLinesFlags_Horizontal);

				ImPlot::SetNextFillStyle(IMPLOT_AUTO_COL, 1.0f);
				ImPlot::SetNextLineStyle(IMPLOT_AUTO_COL, 1.0f);
				ImPlot::PlotLineG("Frame Time", &FrameTimeGetter, NULL, fs.frameCount, ImPlotLineFlags_None);

				ImPlot::EndPlot();
			}
		}
		ImGui::End();
	}

	GUI_DrawMainMenu();
}

uint32_t GRP::CreatePSO(CachedPSO& cache, const char* name)
{
	Q_assert(cache.stageCount > 0);

	const uint32_t pixelShaderStateBits = GLS_BLEND_BITS | GLS_ATEST_BITS;

	for(uint32_t p = 1; p < psoCount; ++p)
	{
		if(cache.stageCount == psos[p].stageCount &&
			memcmp(&cache.desc, &psos[p].desc, sizeof(cache.desc)) == 0 &&
			memcmp(&cache.stageStateBits, &psos[p].stageStateBits, cache.stageCount * sizeof(cache.stageStateBits[0])) == 0)
		{
			return p;
		}
	}

	Q_assert(psoCount < ARRAY_LEN(psos));

#if defined(_DEBUG)
	Q_strncpyz(cache.name, name, sizeof(cache.name));
#endif

	int uberPixelShaderIndex = -1;
	for(uint32_t i = 0; i < uberPixelShaderCacheSize; ++i)
	{
		const UberPixelShaderState& state = uberPixelShaderStates[i];
		const int dither = (state.globalState & UBERPS_DITHER_BIT) != 0 ? 1 : 0;
		const bool depthFade = (state.globalState & UBERPS_DEPTHFADE_BIT) != 0;
		if(cache.stageCount != (uint32_t)state.stageCount ||
			r_dither->integer != dither ||
			cache.desc.depthFade != depthFade)
		{
			continue;
		}

		bool found = true;
		for(uint32_t s = 0; s < cache.stageCount; ++s)
		{
			const uint32_t psoCacheState = cache.stageStateBits[s] & pixelShaderStateBits;
			const uint32_t psCacheState = (uint32_t)state.stageStates[s] & pixelShaderStateBits;
			if(psoCacheState != psCacheState)
			{
				found = false;
				break;
			}
		}

		if(found)
		{
			uberPixelShaderIndex = (int)i;
			break;
		}
	}

	HShader pixelShader = RHI_MAKE_NULL_HANDLE();
	ShaderByteCode pixelShaderByteCode;
	if(uberPixelShaderIndex < 0)
	{
		uint32_t macroCount = 0;
		ShaderMacro macros[64];
		macros[macroCount].name = "STAGE_COUNT";
		macros[macroCount].value = va("%d", cache.stageCount);
		macroCount++;
		if(r_dither->integer)
		{
			macros[macroCount].name = "DITHER";
			macros[macroCount].value = "1";
			macroCount++;
		}
		if(cache.desc.depthFade)
		{
			macros[macroCount].name = "DEPTH_FADE";
			macros[macroCount].value = "1";
			macroCount++;
		}
		for(int s = 0; s < cache.stageCount; ++s)
		{
			macros[macroCount].name = va("STAGE%d_BITS", s);
			macros[macroCount].value = va("%d", (int)cache.stageStateBits[s] & pixelShaderStateBits);
			macroCount++;
		}
		Q_assert(macroCount <= ARRAY_LEN(macros));

		pixelShader = CreateShader(ShaderDesc(ShaderStage::Pixel, sizeof(uber_shader_string), uber_shader_string, "ps", macroCount, macros));
		pixelShaderByteCode = GetShaderByteCode(pixelShader);
	}
	else
	{
		pixelShaderByteCode = uberPixelShaderByteCodes[uberPixelShaderIndex];
	}

	// important missing entries can be copy-pasted into UBER_SHADER_PS_LIST
#if 0
	if(uberPixelShaderIndex < 0)
	{
		unsigned int flags = 0;
		if(r_dither->integer)
		{
			flags |= UBERPS_DITHER_BIT;
		}
		if(cache.desc.depthFade)
		{
			flags |= UBERPS_DEPTHFADE_BIT;
		}
		Sys_DebugPrintf("\tshader: %s\n", name);
		ri.Printf(PRINT_ALL, "^2    shader: %s\n", name);
		Sys_DebugPrintf("\tPS(%d_%X", (int)cache.stageCount, flags);
		ri.Printf(PRINT_ALL, "     PS(%d_%X", (int)cache.stageCount, flags);
		for(int s = 0; s < cache.stageCount; ++s)
		{
			Sys_DebugPrintf("_%X", (unsigned int)(cache.stageStateBits[s] & pixelShaderStateBits));
			ri.Printf(PRINT_ALL, "_%X", (unsigned int)(cache.stageStateBits[s] & pixelShaderStateBits));
		}
		Sys_DebugPrintf(") \\\n");
		ri.Printf(PRINT_ALL, ") \\\n");
	}
#endif

	uint32_t a = 0;
	GraphicsPipelineDesc desc(name, uberRootSignature);
	desc.shortLifeTime = true; // the PSO cache is only valid for this map!
	desc.vertexShader = vertexShaderByteCodes[cache.stageCount - 1];
	desc.pixelShader = pixelShaderByteCode;
	desc.vertexLayout.AddAttribute(a++, ShaderSemantic::Position, DataType::Float32, 3, 0);
	desc.vertexLayout.AddAttribute(a++, ShaderSemantic::Normal, DataType::Float32, 2, 0);
	for(int s = 0; s < cache.stageCount; ++s)
	{
		desc.vertexLayout.AddAttribute(a++, ShaderSemantic::TexCoord, DataType::Float32, 2, 0);
		desc.vertexLayout.AddAttribute(a++, ShaderSemantic::Color, DataType::UNorm8, 4, 0);
	}
	if(cache.desc.depthFade)
	{
		desc.depthStencil.DisableDepth();
	}
	else
	{
		desc.depthStencil.depthStencilFormat = TextureFormat::Depth32_Float;
		desc.depthStencil.depthComparison =
			(cache.stageStateBits[0] & GLS_DEPTHFUNC_EQUAL) != 0 ?
			ComparisonFunction::Equal :
			ComparisonFunction::GreaterEqual;
		desc.depthStencil.enableDepthTest = (cache.stageStateBits[0] & GLS_DEPTHTEST_DISABLE) == 0;
		desc.depthStencil.enableDepthWrites = (cache.stageStateBits[0] & GLS_DEPTHMASK_TRUE) != 0;
	}
	desc.rasterizer.cullMode = cache.desc.cullType;
	desc.rasterizer.polygonOffset = cache.desc.polygonOffset;
	desc.rasterizer.clampDepth = cache.desc.clampDepth;
	desc.AddRenderTarget(cache.stageStateBits[0] & GLS_BLEND_BITS, renderTargetFormat);
	cache.pipeline = CreateGraphicsPipeline(desc);

	if(uberPixelShaderIndex < 0)
	{
		DestroyShader(pixelShader);
	}

	const uint32_t index = psoCount++;
	psos[index] = cache;

	return index;
}

void GRP::ExecuteRenderCommands(const byte* data)
{
	for(;;)
	{
		const int commandId = ((const renderCommandBase_t*)data)->commandId;

		if(commandId < 0 || commandId >= RC_COUNT)
		{
			assert(!"Invalid render command type");
			return;
		}

		if(commandId == RC_END_OF_LIST)
		{
			return;
		}

		switch(commandId)
		{
			case RC_UI_SET_COLOR:
				UISetColor(*(const uiSetColorCommand_t*)data);
				break;
			case RC_UI_DRAW_QUAD:
				UIDrawQuad(*(const uiDrawQuadCommand_t*)data);
				break;
			case RC_UI_DRAW_TRIANGLE:
				UIDrawTriangle(*(const uiDrawTriangleCommand_t*)data);
				break;
			case RC_DRAW_SCENE_VIEW:
				DrawSceneView(*(const drawSceneViewCommand_t*)data);
				break;
			case RC_BEGIN_FRAME:
				BeginFrame();
				break;
			case RC_SWAP_BUFFERS:
				EndFrame();
				break;
			case RC_BEGIN_UI:
				ui.Begin();
				break;
			case RC_END_UI:
				ui.End();
				break;
			case RC_BEGIN_3D:
				world.Begin();
				break;
			case RC_END_3D:
				world.End();
				break;
			case RC_END_SCENE:
				smaa.Draw(((const endSceneCommand_t*)data)->viewParms);
				break;
			case RC_SCREENSHOT:
				RB_TakeScreenshotCmd((const screenshotCommand_t*)data);
				break;
			case RC_VIDEOFRAME:
				RB_TakeVideoFrameCmd((const videoFrameCommand_t*)data);
				break;

			default:
				Q_assert(!"Unsupported render command type");
				return;
		}

		data += renderCommandSizes[commandId];
	}
}

void GRP::ReadPixels(int w, int h, int alignment, colorSpace_t colorSpace, void* outPixels)
{
	const HTexture sourceTexture = GetSwapChainTexture();

	MappedTexture mapped;
	BeginTextureReadback(mapped, sourceTexture);

	byte* const out0 = (byte*)outPixels;
	const byte* const in0 = mapped.mappedData;

	if(colorSpace == CS_RGBA)
	{
		const int dstRowSizeNoPadding = w * 4;
		mapped.dstRowByteCount = AlignUp(dstRowSizeNoPadding, alignment);

		for(int y = 0; y < mapped.rowCount; ++y)
		{
			byte* out = out0 + (mapped.rowCount - 1 - y) * mapped.dstRowByteCount;
			const byte* in = in0 + y * mapped.srcRowByteCount;
			memcpy(out, in, dstRowSizeNoPadding);
		}
	}
	else if(colorSpace == CS_BGR)
	{
		mapped.dstRowByteCount = AlignUp(w * 3, alignment);

		for(int y = 0; y < mapped.rowCount; ++y)
		{
			byte* out = out0 + (mapped.rowCount - 1 - y) * mapped.dstRowByteCount;
			const byte* in = in0 + y * mapped.srcRowByteCount;
			for(int x = 0; x < mapped.columnCount; ++x)
			{
				out[2] = in[0];
				out[1] = in[1];
				out[0] = in[2];
				out += 3;
				in += 4;
			}
		}
	}
	else
	{
		Q_assert(!"Unsupported color space");
	}

	EndTextureReadback();
}

// @TODO: move out once the cinematic render pipeline is added
IRenderPipeline* renderPipeline = &grp;