/*
===========================================================================
Copyright (C) 2023-2024 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/>.
===========================================================================
*/
// Cinematic Rendering Pipeline - main interface


#include "crp_local.h"
#include "../client/cl_imgui.h"
#include "shaders/crp/oit.h.hlsli"
#include "shaders/crp/scene_view.h.hlsli"
#include "compshaders/crp/fullscreen.h"
#include "compshaders/crp/blit.h"
#include "compshaders/crp/ui.h"
#include "compshaders/crp/imgui.h"
#include "compshaders/crp/nuklear.h"
#include "compshaders/crp/mip_1.h"
#include "compshaders/crp/mip_2.h"
#include "compshaders/crp/mip_3.h"
#include "compshaders/crp/depth_pyramid.h"


struct SceneViewConst
{
	enum Constants
	{
		MaxViews = 1024,
		LightBytes = sizeof(DynamicLight),
		MaxLights = SCENE_VIEW_MAX_LIGHTS,
		StructBytes = sizeof(SceneView),
		BufferBytes = MaxViews * StructBytes
	};
};

#pragma pack(push, 4)
struct DepthPyramidRC
{
	uint32_t destTextureIndices[7];
	uint32_t depthTextureIndex;
	uint32_t depthSamplerIndex;
};
#pragma pack(pop)


CRP crp;
IRenderPipeline* crpp = &crp;

cvar_t* crp_dof;
cvar_t* crp_dof_overlay;
cvar_t* crp_dof_blades;
cvar_t* crp_dof_angle;
cvar_t* crp_gatherDof_focusNearDist;
cvar_t* crp_gatherDof_focusNearRange;
cvar_t* crp_gatherDof_focusFarDist;
cvar_t* crp_gatherDof_focusFarRange;
cvar_t* crp_gatherDof_brightness;
cvar_t* crp_accumDof_focusDist;
cvar_t* crp_accumDof_radius;
cvar_t* crp_accumDof_samples;
cvar_t* crp_accumDof_preview;
cvar_t* crp_mblur;
cvar_t* crp_mblur_exposure;
cvar_t* crp_sunlight;
cvar_t* crp_volLight;
cvar_t* crp_drawNormals;
cvar_t* crp_updateRTAS;
cvar_t* crp_debug0;
cvar_t* crp_debug1;
cvar_t* crp_debug2;
cvar_t* crp_debug3;

static const cvarTableItem_t crp_cvars[] =
{
	{
		&crp_dof, "crp_dof", "1", CVAR_ARCHIVE, CVART_INTEGER, "0", "2",
			"enables depth of field\n"
			S_COLOR_VAL "    0 " S_COLOR_HELP "= Disabled\n"
			S_COLOR_VAL "    1 " S_COLOR_HELP "= Gather (fast, more flexible, issues with transparency)\n"
			S_COLOR_VAL "    2 " S_COLOR_HELP "= Accumulation (slow, less flexible, great IQ)\n",
		"DoF mode", CVARCAT_GRAPHICS, "Depth of field mode", "",
		CVAR_GUI_VALUE("0", "Disabled", "")
		CVAR_GUI_VALUE("1", "Gather", "Fast, lower IQ")
		CVAR_GUI_VALUE("2", "Accumulation", "Very slow, great IQ")
	},
	{
		&crp_dof_overlay, "crp_dof_overlay", "0", CVAR_ARCHIVE, CVART_INTEGER, "0", "2",
			"debug overlay mode\n"
			S_COLOR_VAL "    0 " S_COLOR_HELP "= Disabled\n"
			S_COLOR_VAL "    1 " S_COLOR_HELP "= Colorized Blur\n"
			S_COLOR_VAL "    2 " S_COLOR_HELP "= Focus Plane",
		"DoF overlay mode", CVARCAT_GRAPHICS, "Debug overlay mode", "",
		CVAR_GUI_VALUE("0", "Disabled", "")
		CVAR_GUI_VALUE("1", "Colorized Blur", "")
		CVAR_GUI_VALUE("2", "Focus Plane", "")
	},
	{
		&crp_dof_blades, "crp_dof_blades", "6", CVAR_ARCHIVE, CVART_FLOAT, "0", "16",
			"aperture blade count\n"
			"Set to less than 3 for a disk shape.",
		"DoF blade count", CVARCAT_GRAPHICS, "Aperture blade count", "Set to less than 3 for a disk shape."
	},
	{
		&crp_dof_angle, "crp_dof_angle", "20", CVAR_ARCHIVE, CVART_FLOAT, "0", "360", "aperture angle, in degrees",
		"DoF aperture angle", CVARCAT_GRAPHICS, "Aperture angle, in degrees", ""
	},
	{
		&crp_accumDof_focusDist, "crp_accumDof_focusDist", "256", CVAR_ARCHIVE, CVART_FLOAT, "2", "2048", "focus distance",
		"Accum DoF focus distance", CVARCAT_GRAPHICS, "Focus distance", ""
	},
	{
		&crp_accumDof_radius, "crp_accumDof_blurRadius", "0.1", CVAR_ARCHIVE, CVART_FLOAT, "0.001", "20", "aperture radius in world units",
		"Accum DoF aperture radius", CVARCAT_GRAPHICS, "Aperture radius in world units", ""
	},
	{
		&crp_accumDof_samples, "crp_accumDof_samples", "2", CVAR_ARCHIVE, CVART_INTEGER, "1", "12",
			"per-axis sampling density\n"
			"Density N means (2N + 1)(2N + 1) scene renders in total.",
		"Accum DoF sample count", CVARCAT_GRAPHICS, "Per-axis sampling density", "Density N means (2N + 1)^2 scene renders in total."
	},
	{
		&crp_accumDof_preview, "crp_accumDof_preview", "0", CVAR_ARCHIVE, CVART_INTEGER, "0", "2",
			"low-res preview mode\n"
			S_COLOR_VAL "    0 " S_COLOR_HELP "= Disabled\n"
			S_COLOR_VAL "    1 " S_COLOR_HELP "= 1/4 pixel count, 9 samples total\n"
			S_COLOR_VAL "    2 " S_COLOR_HELP "= 1/16 pixel count, 25 samples total",
		"Accum DoF preview mode", CVARCAT_GRAPHICS, "Low-resolution preview modes", "",
		CVAR_GUI_VALUE("0", "Disabled", "")
		CVAR_GUI_VALUE("1", "1/4 pixel count", "9 samples total")
		CVAR_GUI_VALUE("2", "1/16 pixel count", "25 samples total")
	},
	{
		&crp_gatherDof_focusNearDist, "crp_gatherDof_focusNearDist", "192", CVAR_ARCHIVE, CVART_FLOAT, "1", "2048", "near focus distance",
		"Gather DoF near focus distance", CVARCAT_GRAPHICS, "Near focus distance", ""
	},
	{
		&crp_gatherDof_focusNearRange, "crp_gatherDof_focusNearRange", "256", CVAR_ARCHIVE, CVART_FLOAT, "1", "2048", "near focus range",
		"Gather DoF near focus range", CVARCAT_GRAPHICS, "Near focus range", ""
	},
	{
		&crp_gatherDof_focusFarDist, "crp_gatherDof_focusFarDist", "512", CVAR_ARCHIVE, CVART_FLOAT, "1", "2048", "far focus distance",
		"Gather DoF far focus distance", CVARCAT_GRAPHICS, "Far focus distance", ""
	},
	{
		&crp_gatherDof_focusFarRange, "crp_gatherDof_focusFarRange", "384", CVAR_ARCHIVE, CVART_FLOAT, "1", "2048", "far focus range",
		"Gather DoF far focus range", CVARCAT_GRAPHICS, "Far focus range", ""
	},
	{
		&crp_gatherDof_brightness, "crp_gatherDof_brightness", "2", CVAR_ARCHIVE, CVART_FLOAT, "0", "8", "blur brightness weight",
		"Gather DoF bokeh brightness", CVARCAT_GRAPHICS, "Blur brightness weight", ""
	},
	{
		&crp_mblur, "crp_mblur", "0", CVAR_ARCHIVE, CVART_INTEGER, "0", "3",
			"motion blur mode\n"
			S_COLOR_VAL "    0 " S_COLOR_HELP "= Disabled\n"
			S_COLOR_VAL "    1 " S_COLOR_HELP "= Camera only\n"
			S_COLOR_VAL "    2 " S_COLOR_HELP "= Object only\n"
			S_COLOR_VAL "    3 " S_COLOR_HELP "= Camera + Object",
		"Motion blur mode", CVARCAT_GRAPHICS, "", "",
		CVAR_GUI_VALUE("0", "Disabled", "")
		CVAR_GUI_VALUE("1", "Camera only", "")
		CVAR_GUI_VALUE("2", "Object only", "")
		CVAR_GUI_VALUE("3", "Camera + Object", "")
	},
	{
		&crp_mblur_exposure, "crp_mblur_exposure", "0.5", CVAR_ARCHIVE, CVART_FLOAT, "0", "1",
			"motion blur scale\n"
			"This is the exposure time in percentage of frame time.",
		"Motion blur exposure", CVARCAT_GRAPHICS, "Exposure time in percentage of frame time", ""
	},
	{
		&crp_sunlight, "crp_sunlight", "1", CVAR_ARCHIVE, CVART_BOOL, NULL, NULL, "sunlight",
		"Sunlight", CVARCAT_GRAPHICS, "Sunlight on non-lightmapped surfaces", ""
	},
	{
		&crp_volLight, "crp_volLight", "1", CVAR_ARCHIVE, CVART_BOOL, NULL, NULL, "volumetric light",
		"Volumetric light", CVARCAT_GRAPHICS, "Sunlight scattering through the air", ""
	},
	{
		&crp_drawNormals, "crp_drawNormals", "0", CVAR_TEMP, CVART_BOOL, NULL, NULL, "draws vertex normals",
		"Draw vertex normals", CVARCAT_GRAPHICS | CVARCAT_DEBUGGING, "", ""
	},
	{
		&crp_updateRTAS, "crp_updateRTAS", "1", CVAR_TEMP, CVART_BOOL, NULL, NULL, "enables RTAS builds every frame",
		"Enable RTAS builds", CVARCAT_GRAPHICS | CVARCAT_DEBUGGING, "Allows raytracing acceleration structure updates", ""
	}
#if defined(_DEBUG)
	,
	{
		&crp_debug0, "crp_debug0", "0", CVAR_TEMP, CVART_FLOAT, "0", "1", "debug value 0",
		"Debug value 0", CVARCAT_GRAPHICS | CVARCAT_DEBUGGING, "", ""
	},
	{
		&crp_debug1, "crp_debug1", "0", CVAR_TEMP, CVART_FLOAT, "0", "1", "debug value 1",
		"Debug value 1", CVARCAT_GRAPHICS | CVARCAT_DEBUGGING, "", ""
	},
	{
		&crp_debug2, "crp_debug2", "0", CVAR_TEMP, CVART_FLOAT, "0", "1", "debug value 2",
		"Debug value 2", CVARCAT_GRAPHICS | CVARCAT_DEBUGGING, "", ""
	},
	{
		&crp_debug3, "crp_debug3", "0", CVAR_TEMP, CVART_FLOAT, "0", "1", "debug value 3",
		"Debug value 3", CVARCAT_GRAPHICS | CVARCAT_DEBUGGING, "", ""
	}
#endif
};

static void FreezeFrame_f();
static void FreezeFrameMB_f();

static const cmdTableItem_t crp_cmds[] =
{
	{ "cin_freezeFrame", FreezeFrame_f, NULL, "toggles freeze frame mode" },
	{ "cin_freezeFrameMB", FreezeFrameMB_f, NULL, "toggles freeze frame mode for MB" }
};

static void FreezeFrame_f()
{
	if(crp.freezeFrame == FreezeFrame::Active ||
		crp.freezeFrame == FreezeFrame::ActiveBeforeMB)
	{
		crp.freezeFrame = FreezeFrame::Inactive;
	}
	else
	{
		crp.freezeFrame = FreezeFrame::Pending;
	}
}

static void FreezeFrameMB_f()
{
	if(crp.freezeFrame == FreezeFrame::Active ||
		crp.freezeFrame == FreezeFrame::ActiveBeforeMB)
	{
		crp.freezeFrame = FreezeFrame::Inactive;
	}
	else
	{
		crp.freezeFrame = FreezeFrame::PendingBeforeMB;
	}
}

static HTexture LoadTexture(const char* name, int flags, textureWrap_t glWrapClampMode, const image_t* onFail = tr.defaultImage)
{
	image_t* const image = R_FindImageFile(name, flags, glWrapClampMode);
	if(image == NULL)
	{
		return onFail->texture;
	}

	return image->texture;
}

static void SunToZMatrix(matrix3x3_t rot)
{
	float sc[2];
	DirectionToAzimuthInclination(sc, crp.sunlightData.direction);
	const float azi = -(sc[0] + M_PI / 2.0f);
	const float ele = M_PI - sc[1];
	const float rol = 0.0f;
	rot[0] = cosf(rol) * cosf(azi) - sinf(rol) * cosf(ele) * sinf(azi);
	rot[3] = sinf(rol) * cosf(azi) + cosf(rol) * cosf(ele) * sinf(azi);
	rot[6] = sinf(ele) * sinf(azi);
	rot[1] = -cosf(rol) * sinf(azi) - sinf(rol) * cosf(ele) * cosf(azi);
	rot[4] = -sinf(rol) * sinf(azi) + cosf(rol) * cosf(ele) * cosf(azi);
	rot[7] = sinf(ele) * cosf(azi);
	rot[2] = sinf(rol) * sinf(ele);
	rot[5] = -cosf(rol) * sinf(ele);
	rot[8] = cosf(ele);
}

HPipeline CreateComputePipeline(const char* name, const ShaderByteCode& shader)
{
	ComputePipelineDesc desc(name);
	desc.shortLifeTime = true;
	desc.shader = shader;

	return CreateComputePipeline(desc);
}

void MakeFullScreenPipeline(GraphicsPipelineDesc& desc, const ShaderByteCode& pixelShader)
{
	desc.shortLifeTime = true;
	desc.vertexShader = ShaderByteCode(g_fullscreen_vs);
	desc.pixelShader = pixelShader;
	desc.SetPostProcessState();
}

void PSOCache::Init(Entry* entries_, uint32_t maxEntryCount_)
{
	entries = entries_;
	maxEntryCount = maxEntryCount_;
	entryCount = 1; // we treat index 0 as invalid
}

int PSOCache::AddPipeline(const GraphicsPipelineDesc& desc, const char* name)
{
	// we treat index 0 as invalid, so start at 1
	for(uint32_t i = 1; i < entryCount; ++i)
	{
		Entry& entry = entries[i];
		if(memcmp(&entry.desc, &desc, sizeof(desc)) == 0)
		{
			return (int)i;
		}
	}

	ASSERT_OR_DIE(entryCount < maxEntryCount, "Not enough entries in the PSO cache");

	GraphicsPipelineDesc namedDesc = desc;
	namedDesc.name = name;

	// @NOTE: we keep the original desc and its padding bytes for proper comparison results
	const uint32_t index = entryCount++;
	Entry& entry = entries[index];
	memcpy(&entry.desc, &desc, sizeof(entry.desc));
	entry.handle = CreateGraphicsPipeline(namedDesc);

	return (int)index;
}

void CRP::Init()
{
	ri.Cvar_RegisterTable(crp_cvars, ARRAY_LEN(crp_cvars));
	ri.Cmd_RegisterTable(crp_cmds, ARRAY_LEN(crp_cmds));

	InitDesc initDesc;
	initDesc.directDescriptorHeapIndexing = true;
	srp.firstInit = RHI::Init(initDesc);
	srp.psoStatsValid = false;

	if(srp.firstInit)
	{
		srp.CreateShaderTraceBuffers();

		for(uint32_t f = 0; f < FrameCount; ++f)
		{
			// the doubled index count is for the depth pre-pass
			const int MaxDynamicVertexCount = 16 << 20;
			const int MaxDynamicIndexCount = MaxDynamicVertexCount * 4;
			GeoBuffers& db = dynBuffers[f];
			db.Create(va("world #%d", f + 1), MaxDynamicVertexCount, MaxDynamicIndexCount);
		}
	}

	// 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 = GetBaseSamplerIndex(wrap, filter, m);
				SamplerDesc desc(wrap, filter, (float)m);
				desc.shortLifeTime = true;
				samplers[s] = CreateSampler(desc);
				samplerIndices[s] = RHI::GetSamplerIndex(samplers[s]);
			}
		}
	}

	{
		renderTargetFormat = TextureFormat::R16G16B16A16_Float;

		TextureDesc desc("render target #1", glConfig.vidWidth, glConfig.vidHeight);
		desc.initialState = ResourceStates::RenderTargetBit;
		desc.allowedState = ResourceStates::RenderTargetBit | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.SetClearColor(vec4_zero);
		desc.committedResource = true;
		desc.format = renderTargetFormat;
		desc.shortLifeTime = true;
		renderTargets[0] = RHI::CreateTexture(desc);
		desc.name = "render target #2";
		renderTargets[1] = RHI::CreateTexture(desc);
		renderTargetIndex = 0;
		renderTarget = renderTargets[0];
	}

	{
		TextureDesc desc("frozen frame", glConfig.vidWidth, glConfig.vidHeight);
		desc.initialState = ResourceStates::RenderTargetBit;
		desc.allowedState = ResourceStates::RenderTargetBit | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.format = renderTargetFormat;
		desc.shortLifeTime = true;
		frozenTexture = RHI::CreateTexture(desc);

		freezeFrame = FreezeFrame::Inactive;
	}

	{
		TextureDesc desc("readback 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 = TextureFormat::R8G8B8A8_UNorm;
		desc.shortLifeTime = true;
		readbackRenderTarget = RHI::CreateTexture(desc);
	}

	{
		TextureDesc desc("OIT index", glConfig.vidWidth, glConfig.vidHeight);
		desc.initialState = ResourceStates::UnorderedAccessBit;
		desc.allowedState = ResourceStates::UnorderedAccessBit | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.committedResource = true;
		desc.format = TextureFormat::R32_UInt;
		desc.shortLifeTime = true;
		oitIndexTexture = RHI::CreateTexture(desc);
	}

	uint32_t oitMaxFragmentCount = 0;
	{
		const int byteCountPerFragment = sizeof(OIT_Fragment);
		const int fragmentCount = glConfig.vidWidth * glConfig.vidHeight * OIT_AVG_FRAGMENTS_PER_PIXEL;
		const int byteCount = byteCountPerFragment * fragmentCount;
		oitMaxFragmentCount = fragmentCount;

		BufferDesc desc("OIT fragment", byteCount, ResourceStates::UnorderedAccessBit);
		desc.committedResource = true;
		desc.memoryUsage = MemoryUsage::GPU;
		desc.structureByteCount = byteCountPerFragment;
		desc.shortLifeTime = true;
		oitFragmentBuffer = CreateBuffer(desc);
	}

	{
		const int byteCount = sizeof(OIT_Counter);

		{
			BufferDesc desc("OIT counter", byteCount, ResourceStates::UnorderedAccessBit);
			desc.committedResource = true;
			desc.memoryUsage = MemoryUsage::GPU;
			desc.structureByteCount = byteCount;
			desc.shortLifeTime = true;
			oitCounterBuffer = CreateBuffer(desc);
		}
		{
			BufferDesc desc("OIT counter staging", byteCount, ResourceStates::Common);
			desc.committedResource = false;
			desc.memoryUsage = MemoryUsage::Upload;
			desc.structureByteCount = byteCount;
			desc.shortLifeTime = true;
			oitCounterStagingBuffer = CreateBuffer(desc);

			uint32_t* dst = (uint32_t*)MapBuffer(oitCounterStagingBuffer);
			dst[0] = 1; // fragment index 0 is the end-of-list value
			dst[1] = oitMaxFragmentCount;
			dst[2] = 0;
			UnmapBuffer(oitCounterStagingBuffer);
		}
	}

	{
		TextureDesc desc("depth buffer", glConfig.vidWidth, glConfig.vidHeight);
		desc.committedResource = true;
		desc.shortLifeTime = true;
		desc.initialState = ResourceStates::DepthWriteBit;
		desc.allowedState = ResourceStates::DepthAccessBits | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.format = TextureFormat::Depth32_Float;
		desc.SetClearDepthStencil(0.0f, 0);
		depthTexture = RHI::CreateTexture(desc);
	}

	{
		TextureDesc desc("depth pyramid", glConfig.vidWidth, glConfig.vidHeight, 7);
		desc.shortLifeTime = true;
		desc.initialState = ResourceStates::UnorderedAccessBit;
		desc.allowedState = ResourceStates::UnorderedAccessBit | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.format = TextureFormat::R32G32_Float;
		depthMinMaxTexture = RHI::CreateTexture(desc);
	}

	{
		TextureDesc desc("GBuffer normals", glConfig.vidWidth, glConfig.vidHeight);
		desc.committedResource = true;
		desc.shortLifeTime = true;
		desc.initialState = ResourceStates::RenderTargetBit;
		desc.allowedState = ResourceStates::RenderTargetBit | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.format = TextureFormat::R16G16_SNorm;
		desc.SetClearColor(vec4_zero);
		normalTexture = RHI::CreateTexture(desc);
	}

	{
		TextureDesc desc("GBuffer raw motion vectors", glConfig.vidWidth, glConfig.vidHeight);
		desc.committedResource = true;
		desc.shortLifeTime = true;
		desc.initialState = ResourceStates::RenderTargetBit;
		desc.allowedState = ResourceStates::RenderTargetBit | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.format = TextureFormat::R16G16_Float;
		desc.SetClearColor(vec4_zero);
		motionVectorTexture = RHI::CreateTexture(desc);
	}

	{
		TextureDesc desc("GBuffer MB motion vectors", glConfig.vidWidth, glConfig.vidHeight);
		desc.committedResource = true;
		desc.shortLifeTime = true;
		desc.initialState = ResourceStates::RenderTargetBit;
		desc.allowedState = ResourceStates::RenderTargetBit | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.format = TextureFormat::R16G16_Float;
		desc.SetClearColor(vec4_zero);
		motionVectorMBTexture = RHI::CreateTexture(desc);
	}

	{
		TextureDesc desc("GBuffer direct light", glConfig.vidWidth, glConfig.vidHeight);
		desc.committedResource = true;
		desc.shortLifeTime = true;
		desc.initialState = ResourceStates::RenderTargetBit;
		desc.allowedState = ResourceStates::RenderTargetBit | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.format = TextureFormat::R16G16B16A16_Float;
		desc.SetClearColor(colorBlack);
		lightTexture = RHI::CreateTexture(desc);
		desc.name = "GBuffer raw direct light";
		sunlightTexture = RHI::CreateTexture(desc);
	}

	{
		TextureDesc desc("GBuffer shading position", glConfig.vidWidth, glConfig.vidHeight);
		desc.committedResource = true;
		desc.shortLifeTime = true;
		desc.initialState = ResourceStates::RenderTargetBit;
		desc.allowedState = ResourceStates::RenderTargetBit | ResourceStates::PixelShaderAccessBit | ResourceStates::ComputeShaderAccessBit;
		desc.format = TextureFormat::R32G32B32A32_Float;
		desc.SetClearColor(vec4_zero);
		shadingPositionTexture = RHI::CreateTexture(desc);
	}

	{
		GraphicsPipelineDesc desc("blit LDR");
		MakeFullScreenPipeline(desc, ShaderByteCode(g_blit_ps));
		desc.AddRenderTarget(0, TextureFormat::R8G8B8A8_UNorm);
		blitPipelineLDR = CreateGraphicsPipeline(desc);
		desc.name = "blit HDR";
		desc.renderTargets[0].format = TextureFormat::R16G16B16A16_Float;
		blitPipelineHDR = CreateGraphicsPipeline(desc);
	}

	depthPyramidPipeline = CreateComputePipeline("Depth Pyramid", ShaderByteCode(g_depth_pyramid_cs));

	{
		BufferDesc desc("scene view upload #1", SceneViewConst::BufferBytes, ResourceStates::ShaderAccessBits);
		desc.shortLifeTime = true;
		desc.memoryUsage = MemoryUsage::Upload;
		desc.structureByteCount = SceneViewConst::StructBytes;
		sceneViewUploadBuffers[0] = CreateBuffer(desc);
		desc.name = "scene view upload #2";
		sceneViewUploadBuffers[1] = CreateBuffer(desc);
	}

	{
		BufferDesc desc("scene view", SceneViewConst::StructBytes, ResourceStates::ShaderAccessBits);
		desc.shortLifeTime = true;
		desc.structureByteCount = SceneViewConst::StructBytes;
		desc.useSrvIndex0 = true; // the one and only buffer allowed to be there
		sceneViewBuffer = CreateBuffer(desc);
	}

	raytracing.Init();
	ui.Init(true, ShaderByteCode(g_ui_vs), ShaderByteCode(g_ui_ps), renderTargetFormat, RHI_MAKE_NULL_HANDLE(), NULL);
	imgui.Init(true, ShaderByteCode(g_imgui_vs), ShaderByteCode(g_imgui_ps), renderTargetFormat, RHI_MAKE_NULL_HANDLE(), NULL);
	nuklear.Init(true, ShaderByteCode(g_nuklear_vs), ShaderByteCode(g_nuklear_ps), renderTargetFormat, RHI_MAKE_NULL_HANDLE(), NULL);
	mipMapGen.Init(true, ShaderByteCode(g_mip_1_cs), ShaderByteCode(g_mip_2_cs), ShaderByteCode(g_mip_3_cs));
	prepass.Init();
	opaque.Init();
	transp.Init();
	transpResolve.Init();
	toneMap.Init();
	gatherDof.Init();
	accumDof.Init();
	motionBlur.Init();
	magnifier.Init();
	dynamicLights.Init();
	sunlight.Init();
	volumetricLight.Init();
	gbufferViz.Init();
	sunlightEditor.Init();

	srp.firstInit = false;
}

void CRP::LoadResources()
{
	const int flags = IMG_NOPICMIP | IMG_NOMIPMAP | IMG_NOIMANIP | IMG_NOAF;
	blueNoise2D = LoadTexture("textures/stbn_2d.tga", flags, TW_REPEAT);
}

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

void CRP::BeginFrame()
{
	renderTargetIndex = 0;
	renderTarget = renderTargets[0];
	sceneViewIndex = 0;

	srp.BeginFrame();

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

	// must be run outside of the RHI::BeginFrame/RHI::EndFrame pair
	const bool rtasUpdate =
		dynamicLights.WantRTASUpdate(tr.rtRefdef) ||
		sunlight.WantRTASUpdate(tr.rtRefdef) ||
		volumetricLight.WantRTASUpdate(tr.rtRefdef);
	raytracing.BeginFrame(rtasUpdate);

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

	CmdBeginBarrier();
	CmdTextureBarrier(renderTarget, ResourceStates::RenderTargetBit);
	CmdEndBarrier();

	const float clearColor[4] = { 0.0f, 0.5f, 0.0f, 0.0f };
	CmdClearColorTarget(renderTarget, clearColor);

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

	dynBuffers[GetFrameIndex()].Rewind();
}

void CRP::EndFrame()
{
	if(freezeFrame == FreezeFrame::Pending)
	{
		crp.SwapRenderTargets();
		CmdBeginBarrier();
		CmdTextureBarrier(GetReadRenderTarget(), ResourceStates::PixelShaderAccessBit);
		CmdTextureBarrier(frozenTexture, ResourceStates::RenderTargetBit);
		CmdEndBarrier();
		Blit(frozenTexture, GetReadRenderTarget(), "Blit Frozen Frame", true, vec2_one, vec2_zero);
		CmdBeginBarrier();
		CmdTextureBarrier(frozenTexture, ResourceStates::PixelShaderAccessBit);
		CmdEndBarrier();
		srp.EndFrame();
		freezeFrame = FreezeFrame::Active;
		return;
	}

	srp.DrawGUI();
	gbufferViz.DrawGUI();
	magnifier.DrawGUI();
	sunlightEditor.DrawGUI();
	volumetricLight.DrawGUI();
	imgui.Draw(renderTarget);
	toneMap.DrawToneMap();
	magnifier.Draw();
	BlitRenderTarget(GetSwapChainTexture(), "Blit to Swap Chain");
	BlitRenderTarget(readbackRenderTarget, "Blit to Readback Texture");
	srp.EndFrame();
}

void CRP::Blit(HTexture destination, HTexture source, const char* passName, bool hdr, const vec2_t tcScale, const vec2_t tcBias)
{
	SCOPED_DEBUG_LABEL(passName, 0.125f, 0.125f, 0.5f);

	CmdBeginBarrier();
	CmdTextureBarrier(source, ResourceStates::PixelShaderAccessBit);
	CmdTextureBarrier(destination, ResourceStates::RenderTargetBit);
	CmdEndBarrier();

#pragma pack(push, 4)
	struct BlitRC
	{
		uint32_t textureIndex;
		uint32_t samplerIndex;
		float tcScale[2];
		float tcBias[2];
	};
#pragma pack(pop)

	BlitRC rc;
	rc.textureIndex = GetTextureIndexSRV(source);
	rc.samplerIndex = GetSamplerIndex(TW_CLAMP_TO_EDGE, TextureFilter::Linear);
	rc.tcScale[0] = tcScale[0];
	rc.tcScale[1] = tcScale[1];
	rc.tcBias[0] = tcBias[0];
	rc.tcBias[1] = tcBias[0];
	CmdSetViewportAndScissor(0, 0, glConfig.vidWidth, glConfig.vidHeight);
	CmdBindRenderTargets(1, &destination, NULL);
	CmdBindPipeline(hdr ? blitPipelineHDR : blitPipelineLDR);
	CmdSetGraphicsRootConstants(0, sizeof(rc), &rc);
	CmdDraw(3, 0);
}

void CRP::BlitRenderTarget(HTexture destination, const char* passName)
{
	Blit(destination, crp.renderTarget, passName, false, vec2_one, vec2_zero);
}

void CRP::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 = GetTextureIndexSRV(image->texture);
}

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

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

void CRP::ProcessWorld(world_t& world)
{
	raytracing.ProcessWorld(world);
	sunlightEditor.ProcessWorld(world);
	volumetricLight.ProcessWorld(world);
}

void CRP::ProcessModel(model_t&)
{
}

void CRP::ProcessShader(shader_t& shader)
{
	if(shader.isOpaque)
	{
		prepass.ProcessShader(shader);
		opaque.ProcessShader(shader);
	}
	else
	{
		transp.ProcessShader(shader);
	}
}

void CRP::ExecuteRenderCommands(const byte* data, bool /*readbackRequested*/)
{
	// @NOTE: readbackRequested is unused because
	// the CRP always blits the final result to the readback texture

	if(freezeFrame == FreezeFrame::Active ||
		freezeFrame == FreezeFrame::ActiveBeforeMB)
	{
		BeginFrame();
		Blit(GetWriteRenderTarget(), frozenTexture, "Blit Frozen Frame", true, vec2_one, vec2_zero);
		if(freezeFrame == FreezeFrame::ActiveBeforeMB)
		{
			motionBlur.Draw();
		}
		EndFrame();
		return;
	}

	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:
				ui.CmdSetColor(*(const uiSetColorCommand_t*)data);
				break;
			case RC_UI_DRAW_QUAD:
				ui.CmdDrawQuad(*(const uiDrawQuadCommand_t*)data);
				break;
			case RC_UI_DRAW_TRIANGLE:
				ui.CmdDrawTriangle(*(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(renderTarget);
				break;
			case RC_END_UI:
				ui.End();
				break;
			case RC_BEGIN_3D:
				// @TODO:
				srp.renderMode = RenderMode::None;
				break;
			case RC_END_3D:
				// @TODO:
				srp.renderMode = RenderMode::None;
				break;
			case RC_END_SCENE:
				// @TODO: post-processing
				break;
			case RC_BEGIN_NK:
				nuklear.Begin(renderTarget);
				break;
			case RC_END_NK:
				nuklear.End();
				break;
			case RC_NK_UPLOAD:
				nuklear.Upload(*(const nuklearUploadCommand_t*)data);
				break;
			case RC_NK_DRAW:
				nuklear.Draw(*(const nuklearDrawCommand_t*)data);
				break;
			default:
				Q_assert(!"Unsupported render command type");
				return;
		}

		data += renderCommandSizes[commandId];
	}
}

void CRP::TessellationOverflow()
{
	switch(tess.tessellator)
	{
		case Tessellator::Prepass: prepass.TessellationOverflow(); break;
		case Tessellator::Opaque: opaque.TessellationOverflow(); break;
		case Tessellator::Transp: transp.TessellationOverflow(); break;
		default: break;
	}
	tess.numIndexes = 0;
	tess.numVertexes = 0;
}

void CRP::DrawSceneView3D(const drawSceneViewCommand_t& cmd)
{
	const int lightCount = backEnd.refdef.num_dlights;

	prepass.Draw(cmd);
	BuildDepthPyramid();
	dynamicLights.DrawBegin();
	if(volumetricLight.ShouldDraw())
	{
		volumetricLight.DrawBegin();
		if(raytracing.CanRaytrace())
		{
			{
				SCOPED_RENDER_PASS("VL/DL Point Lights", 1.0f, 1.0f, 1.0f);
				for(int i = 0; i < lightCount; i++)
				{
					volumetricLight.DrawPointLight(backEnd.refdef.dlights[i]);
					dynamicLights.DrawPointLight(backEnd.refdef.dlights[i]);
				}
			}
			volumetricLight.DrawSunlight();
		}
		volumetricLight.DrawEnd();
	}
	else
	{
		SCOPED_RENDER_PASS("DL Point Lights", 1.0f, 1.0f, 1.0f);
		for(int i = 0; i < lightCount; i++)
		{
			dynamicLights.DrawPointLight(backEnd.refdef.dlights[i]);
		}
	}
	sunlight.Draw();
	opaque.Draw(cmd);
	volumetricLight.DrawDebug();
	transp.Draw(cmd);
	transpResolve.Draw(cmd);
}

void CRP::DrawSceneView(const drawSceneViewCommand_t& cmd)
{
	const viewParms_t& vp = cmd.viewParms;
	if(cmd.shouldClearColor)
	{
		const Rect rect(vp.viewportX, vp.viewportY, vp.viewportWidth, vp.viewportHeight);
		CmdBeginBarrier();
		CmdTextureBarrier(renderTarget, ResourceStates::RenderTargetBit);
		CmdEndBarrier();
		CmdClearColorTarget(renderTarget, cmd.clearColor, &rect);
	}

	if(cmd.numDrawSurfs <= 0 || !cmd.shouldDrawScene)
	{
		return;
	}

	if(crp_dof->integer == DOFMethod::Accumulation &&
		IsViewportFullscreen(vp))
	{
		const Rect rect(0, 0, glConfig.vidWidth, glConfig.vidHeight);
		accumDof.Begin(cmd);
		const uint32_t sampleCount = accumDof.GetSampleCount();
		for(uint32_t y = 0; y < sampleCount; y++)
		{
			for(uint32_t x = 0; x < sampleCount; x++)
			{
				srp.enableRenderPassQueries = x == 0 && y == 0;
				drawSceneViewCommand_t newCmd;
				accumDof.FixCommand(newCmd, cmd, x, y);

				backEnd.refdef = newCmd.refdef;
				backEnd.viewParms = newCmd.viewParms;
				UploadSceneViewData();

				CmdBeginBarrier();
				CmdTextureBarrier(renderTarget, ResourceStates::RenderTargetBit);
				CmdEndBarrier();
				CmdClearColorTarget(renderTarget, cmd.clearColor, &rect);
				DrawSceneView3D(newCmd);
				accumDof.Accumulate();

				// geometry allocation is a linear allocation instead of a ring buffer
				// we force a CPU-GPU sync point after every full scene render
				// that way, we can keep the buffer sizes at least somewhat reasonable
				SubmitAndContinue();
				dynBuffers[GetFrameIndex()].Rewind();
			}
		}
		CmdSetViewportAndScissor(backEnd.viewParms);
		srp.enableRenderPassQueries = true;
		accumDof.Normalize();
		backEnd.viewParms = cmd.viewParms;
		backEnd.refdef = cmd.refdef;
		accumDof.DrawDebug();
	}
	else
	{
		backEnd.refdef = cmd.refdef;
		backEnd.viewParms = cmd.viewParms;
		UploadSceneViewData();

		DrawSceneView3D(cmd);
		CmdSetViewportAndScissor(vp.viewportX, vp.viewportY, vp.viewportWidth, vp.viewportHeight);
		gatherDof.Draw();
		if(freezeFrame == FreezeFrame::PendingBeforeMB &&
			IsViewportFullscreen(backEnd.viewParms))
		{
			crp.SwapRenderTargets();
			CmdBeginBarrier();
			CmdTextureBarrier(GetReadRenderTarget(), ResourceStates::PixelShaderAccessBit);
			CmdTextureBarrier(frozenTexture, ResourceStates::RenderTargetBit);
			CmdEndBarrier();
			Blit(frozenTexture, GetReadRenderTarget(), "Blit Frozen Frame", true, vec2_one, vec2_zero);
			CmdBeginBarrier();
			CmdTextureBarrier(frozenTexture, ResourceStates::PixelShaderAccessBit);
			CmdEndBarrier();
			freezeFrame = FreezeFrame::ActiveBeforeMB;
		}
		motionBlur.Draw();
	}

	sunlightEditor.DrawOverlay();
}

void CRP::UploadSceneViewData()
{
	Q_assert(sceneViewIndex < SceneViewConst::MaxViews);
	if(sceneViewIndex >= SceneViewConst::MaxViews)
	{
		return;	
	}

	SCOPED_DEBUG_LABEL("Scene View Upload", 1.0f, 1.0f, 1.0f);

	const viewParms_t& vp = backEnd.viewParms;
	const HBuffer uploadBuffer = sceneViewUploadBuffers[GetFrameIndex()];
	const uint32_t uploadByteOffset = sceneViewIndex * SceneViewConst::StructBytes;

	if(!vp.isPortal && IsViewportFullscreen(vp))
	{
		Q_assert(tr.currZFar == vp.zFar);
		Q_assert(tr.currZNear == vp.zNear);
	}

	SceneView scene = {};

#if defined(_DEBUG)
	scene.debug[0] = crp_debug0->value;
	scene.debug[1] = crp_debug1->value;
	scene.debug[2] = crp_debug2->value;
	scene.debug[3] = crp_debug3->value;
#endif

	scene.frameSeed = (float)rand() / (float)RAND_MAX;

	// @NOTE: yes, world.modelMatrix is actually the view matrix
	// it's the model-view matrix for the world entity, thus the view matrix
	memcpy(scene.projectionMatrix, vp.projectionMatrix, sizeof(scene.projectionMatrix));
	R_InvMatrix(scene.projectionMatrix, scene.invProjectionMatrix);
	memcpy(scene.viewMatrix, vp.world.modelMatrix, sizeof(scene.viewMatrix));
	R_InvMatrix(scene.viewMatrix, scene.invViewMatrix);

	memcpy(scene.prevViewProjMatrix, tr.prevViewProjMatrix, sizeof(scene.prevViewProjMatrix));
	memcpy(scene.prevViewMatrix, tr.prevViewMatrix, sizeof(scene.prevViewMatrix));
	memcpy(scene.prevProjectionMatrix, tr.prevProjMatrix, sizeof(scene.prevProjectionMatrix));

	// we want the first Z slice to be closest to the sun to simplify ray marching
	vec3_t zDown;
	VectorSet(zDown, 0, 0, -1);
	SunToZMatrix(scene.sunToZMatrix);
	R_InvMatrix3x3(scene.sunToZMatrix, scene.zToSunMatrix);

	RB_CreateClipPlane(scene.clipPlane);
	VectorCopy(vp.world.viewOrigin, scene.cameraPosition);
	VectorCopy(tr.prevCameraPosition, scene.prevCameraPosition);
	VectorCopy(vp.orient.axis[0], scene.cameraForward);
	VectorCopy(vp.orient.axis[1], scene.cameraLeft);
	VectorCopy(vp.orient.axis[2], scene.cameraUp);
	scene.sceneViewIndex = sceneViewIndex;
	scene.frameIndex = tr.frameCount;
	scene.depthTextureIndex = GetTextureIndexSRV(depthTexture);
	scene.depthMinMaxTextureIndex = GetTextureIndexSRV(depthMinMaxTexture);
	scene.normalTextureIndex = GetTextureIndexSRV(normalTexture);
	scene.shadingPositionTextureIndex = GetTextureIndexSRV(shadingPositionTexture);
	scene.motionVectorTextureIndex = GetTextureIndexSRV(motionVectorTexture);
	scene.motionVectorMBTextureIndex = GetTextureIndexSRV(motionVectorMBTexture);
	scene.lightTextureIndex = GetTextureIndexSRV(lightTexture);
	scene.sunlightTextureIndex = GetTextureIndexSRV(sunlightTexture);
	scene.tlasBufferIndex = raytracing.GetTLASBufferIndex();
	scene.tlasInstanceBufferIndex = raytracing.GetInstanceBufferIndex();
	RB_LinearDepthConstants(scene.linearDepthConstants);
	scene.zNear = vp.zNear;
	scene.zFar = vp.zFar;
	scene.prevZNear = tr.prevZNear;
	scene.prevZFar = tr.prevZFar;

	VectorCopy(sunlightData.direction, scene.sunDirection);
	VectorCopy(sunlightData.color, scene.sunColor);
	scene.sunIntensity = sunlightData.intensity;

	VectorCopy(volumetricLight.ambientColor, scene.ambientColor);
	scene.ambientIntensity = volumetricLight.ambientIntensity;

	Vector4Copy(volumetricLight.extinctionVolumeScale, scene.extinctionWorldScale);
	for(int c = 0; c < 4; c++)
	{
		scene.extinctionTextureIndices[c] = GetTextureIndexSRV(volumetricLight.extinctionTextures[c]);
	}

	Vector4Copy(volumetricLight.sunShadowVolumeScale, scene.sunVShadowWorldScale);
	for(int c = 0; c < 4; c++)
	{
		scene.sunVShadowTextureIndices[c] = GetTextureIndexSRV(volumetricLight.sunShadowTextures[c]);
	}

	scene.linearClampSamplerIndex = GetSamplerIndex(TW_CLAMP_TO_EDGE, TextureFilter::Linear);

	SceneView* const mappedScene = (SceneView*)(MapBuffer(uploadBuffer) + uploadByteOffset);
	memcpy(mappedScene, &scene, sizeof(scene));
	UnmapBuffer(uploadBuffer);

	CmdBeginBarrier();
	CmdBufferBarrier(uploadBuffer, ResourceStates::CopySourceBit);
	CmdBufferBarrier(sceneViewBuffer, ResourceStates::CopyDestinationBit);
	CmdEndBarrier();
	CmdCopyBuffer(sceneViewBuffer, 0, uploadBuffer, uploadByteOffset, SceneViewConst::StructBytes);
	CmdBeginBarrier();
	CmdBufferBarrier(sceneViewBuffer, ResourceStates::ShaderAccessBits);
	CmdEndBarrier();

	sceneViewIndex++;
}

void CRP::BuildDepthPyramid()
{
	SCOPED_RENDER_PASS("Depth Pyramid", 1.0f, 1.0f, 1.0f);

	CmdBeginBarrier();
	CmdTextureBarrier(depthTexture, ResourceStates::ComputeShaderAccessBit);
	CmdTextureBarrier(depthMinMaxTexture, ResourceStates::UnorderedAccessBit);
	CmdEndBarrier();

	DepthPyramidRC rc = {};
	for(uint32_t i = 0; i < ARRAY_LEN(rc.destTextureIndices); i++)
	{
		rc.destTextureIndices[i] = GetTextureIndexUAV(depthMinMaxTexture, i);
	}
	rc.depthTextureIndex = GetTextureIndexSRV(depthTexture);
	rc.depthSamplerIndex = GetSamplerIndex(TW_CLAMP_TO_EDGE, TextureFilter::Point);

	const int w = glConfig.vidWidth / 2;
	const int h = glConfig.vidHeight / 2;
	CmdBindPipeline(depthPyramidPipeline);
	CmdSetComputeRootConstants(0, sizeof(rc), &rc);
	CmdDispatch((w + 7) / 8, (h + 7) / 8, 1);
}

void CRP::ReadPixels(int w, int h, int alignment, colorSpace_t colorSpace, void* outPixels)
{
	ReadTextureImage(outPixels, readbackRenderTarget, w, h, alignment, colorSpace);
}

uint32_t CRP::GetSamplerDescriptorIndexFromBaseIndex(uint32_t baseIndex)
{
	Q_assert(baseIndex < ARRAY_LEN(samplerIndices));

	return samplerIndices[baseIndex];
}

HTexture CRP::GetReadRenderTarget()
{
	return renderTargets[renderTargetIndex ^ 1];
}

HTexture CRP::GetWriteRenderTarget()
{
	return renderTargets[renderTargetIndex];
}

void CRP::SwapRenderTargets()
{
	renderTargetIndex ^= 1;
	renderTarget = GetWriteRenderTarget();
}