#include "VulkanDeferred.h"

#include <array>

#include "VulkanRenderer.h"
#include "VulkanBarrier.h"
#include "VulkanBuffer.h"
#include "VulkanTexture.h"
#include "VulkanDescriptorManager.h"
#include "VulkanPipeline.h"
#include "VulkanState.h"
#include "VulkanDraw.h"
#include "VulkanPostProcessing.h"
#include "gr_vulkan.h"

#include "cmdline/cmdline.h"
#include "graphics/2d.h"
#include "graphics/matrix.h"
#include "graphics/material.h"
#include "graphics/grinternal.h"
#include "graphics/rtao.h"
#include "graphics/shadows.h"
#include "lighting/lighting.h"
#include "mission/missionparse.h"
#include "nebula/neb.h"
#include "nebula/volumetrics.h"
#include "render/3d.h"


namespace graphics::vulkan {

namespace {

bool s_vulkanOverrideFog = false;

} // anonymous namespace

/**
 * @brief Transition all 6 MSAA images between two layout states.
 *
 * Builds a 6-barrier batch (5 color + 1 depth) and submits via pipelineBarrier.
 */
static void transitionMsaaImages(vk::CommandBuffer cmd, VulkanPostProcessor* pp,
                                  vk::ImageLayout colorOldLayout, vk::ImageLayout colorNewLayout,
                                  vk::AccessFlags2 colorSrcAccess, vk::AccessFlags2 colorDstAccess,
                                  vk::ImageLayout depthOldLayout, vk::ImageLayout depthNewLayout,
                                  vk::AccessFlags2 depthSrcAccess, vk::AccessFlags2 depthDstAccess,
                                  vk::PipelineStageFlags2 srcStage, vk::PipelineStageFlags2 dstStage)
{
	std::array<ImageBarrier2, 6> barriers;

	std::array<vk::Image, 5> msaaImages = {
		pp->deferred().msaaColorImage(),
		pp->deferred().msaaPositionImage(),
		pp->deferred().msaaNormalImage(),
		pp->deferred().msaaSpecularImage(),
		pp->deferred().msaaEmissiveImage(),
	};
	for (size_t i = 0; i < msaaImages.size(); ++i) {
		barriers[i].image = msaaImages[i];
		barriers[i].levelCount = 1;
		barriers[i].layerCount = 1;
		barriers[i].oldLayout = colorOldLayout;
		barriers[i].newLayout = colorNewLayout;
		barriers[i].srcStage = srcStage;
		barriers[i].srcAccess = colorSrcAccess;
		barriers[i].dstStage = dstStage;
		barriers[i].dstAccess = colorDstAccess;
	}

	barriers[5].image = pp->deferred().msaaDepthImage();
	barriers[5].aspectMask = imageAspectFromFormat(pp->getDepthFormat());
	barriers[5].levelCount = 1;
	barriers[5].layerCount = 1;
	barriers[5].oldLayout = depthOldLayout;
	barriers[5].newLayout = depthNewLayout;
	barriers[5].srcStage = srcStage;
	barriers[5].srcAccess = depthSrcAccess;
	barriers[5].dstStage = dstStage;
	barriers[5].dstAccess = depthDstAccess;

	cmdImageBarriers(cmd, ArrayView<const ImageBarrier2>(barriers.data(), barriers.size()));
}

// ========== Deferred Lighting ==========

void vulkan_deferred_lighting_begin(bool clearNonColorBufs)
{
	if (!light_deferred_enabled()) {
		return;
	}

	auto* pp = getPostProcessor();
	if (!pp || !pp->deferred().isInitialized()) {
		return;
	}

	auto* renderer = getRendererInstance();
	if (!renderer->isSceneRendering()) {
		return;
	}

	GR_DEBUG_SCOPE("Deferred lighting begin");

	auto* stateTracker = getStateTracker();
	vk::CommandBuffer cmd = stateTracker->getCommandBuffer();

	const bool msaaActive = (Cmdline_msaa_enabled > 0 && pp->deferred().isMsaaInitialized());

	// End the current G-buffer render pass to perform the color→emissive copy.
	// All 6 color attachments transition to eShaderReadOnlyOptimal (finalLayout).
	cmd.endRenderPass();

	// Copy scene color → non-MSAA emissive (pre-deferred content becomes emissive).
	// Skip both post-barriers — conditional MSAA/non-MSAA code below handles transitions.
	copyImageToImage(cmd,
		pp->getSceneColorImage(), vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageLayout::eTransferSrcOptimal,
		pp->deferred().emissiveImage(), vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageLayout::eTransferDstOptimal,
		pp->getSceneExtent());

	if (msaaActive) {
		// --- MSAA path ---
		// Transition scene color: eTransferSrcOptimal → eShaderReadOnlyOptimal
		// (will be sampled inside MSAA pass to fill emissive)
		// Transition non-MSAA emissive: eTransferDstOptimal → eShaderReadOnlyOptimal (preserved for later)
		{
			std::array<ImageBarrier2, 2> barriers;

			barriers[0].image = pp->getSceneColorImage();
			barriers[0].levelCount = 1;
			barriers[0].layerCount = 1;
			barriers[0].oldLayout = vk::ImageLayout::eTransferSrcOptimal;
			barriers[0].newLayout = vk::ImageLayout::eShaderReadOnlyOptimal;
			barriers[0].srcStage = vk::PipelineStageFlagBits2::eTransfer;
			barriers[0].srcAccess = vk::AccessFlagBits2::eTransferRead;
			barriers[0].dstStage = vk::PipelineStageFlagBits2::eFragmentShader;
			barriers[0].dstAccess = vk::AccessFlagBits2::eShaderSampledRead;

			barriers[1].image = pp->deferred().emissiveImage();
			barriers[1].levelCount = 1;
			barriers[1].layerCount = 1;
			barriers[1].oldLayout = vk::ImageLayout::eTransferDstOptimal;
			barriers[1].newLayout = vk::ImageLayout::eShaderReadOnlyOptimal;
			barriers[1].srcStage = vk::PipelineStageFlagBits2::eTransfer;
			barriers[1].srcAccess = vk::AccessFlagBits2::eTransferWrite;
			barriers[1].dstStage = vk::PipelineStageFlagBits2::eFragmentShader;
			barriers[1].dstAccess = {};

			cmdImageBarriers(cmd, ArrayView<const ImageBarrier2>(barriers.data(), barriers.size()));
		}

		// Transition MSAA images to expected initial layouts
		pp->deferred().transitionMsaaForBegin(cmd);

		// Copy scene color → MSAA emissive via dedicated single-attachment render pass.
		// bindGlobalSet=true: this can run before any material draw this frame has
		// bound Set 0. sampleCount matches the MSAA G-buffer this feeds into.
		pp->context().drawFullscreenTriangle(cmd,
			pp->deferred().msaaEmissiveCopyRenderPass(), pp->deferred().msaaEmissiveCopyFramebuffer(),
			pp->getSceneExtent(), SDR_TYPE_COPY,
			pp->getSceneColorView(), pp->getSceneColorSampler(),
			nullptr, 0, ALPHA_BLEND_NONE, 0,
			renderer->getMsaaSampleCount(), true);

		// Begin MSAA G-buffer render pass (eClear for 0-3+depth, eLoad for emissive [4])
		{
			auto extent = pp->getSceneExtent();
			vk::RenderPassBeginInfo rpBegin;
			rpBegin.renderPass = pp->deferred().msaaRenderPass();
			rpBegin.framebuffer = pp->deferred().msaaFramebuffer();
			rpBegin.renderArea.offset = vk::Offset2D(0, 0);
			rpBegin.renderArea.extent = extent;
			std::array<vk::ClearValue, VulkanDeferredGBuffer::MSAA_COLOR_ATTACHMENT_COUNT + 1> clearValues{};
			clearValues[VulkanDeferredGBuffer::GBUF_ATT_COLOR].color.setFloat32({0.0f, 0.0f, 0.0f, 0.0f});
			clearValues[VulkanDeferredGBuffer::GBUF_ATT_POSITION].color.setFloat32({0.0f, 0.0f, 0.0f, 0.0f});
			clearValues[VulkanDeferredGBuffer::GBUF_ATT_NORMAL].color.setFloat32({0.0f, 0.0f, 0.0f, 0.0f});
			clearValues[VulkanDeferredGBuffer::GBUF_ATT_SPECULAR].color.setFloat32({0.0f, 0.0f, 0.0f, 0.0f});
			clearValues[VulkanDeferredGBuffer::MSAA_COLOR_ATTACHMENT_COUNT].depthStencil = vk::ClearDepthStencilValue(1.0f, 0);
			rpBegin.clearValueCount = static_cast<uint32_t>(clearValues.size());
			rpBegin.pClearValues = clearValues.data();
			cmd.beginRenderPass(rpBegin, vk::SubpassContents::eInline);
			stateTracker->setRenderPass(pp->deferred().msaaRenderPass(), 0);
			stateTracker->setColorAttachmentCount(VulkanDeferredGBuffer::MSAA_COLOR_ATTACHMENT_COUNT);
			stateTracker->setCurrentSampleCount(renderer->getMsaaSampleCount());
		}
	} else {
		// --- Non-MSAA path (original) ---
		// Transition scene color back to eColorAttachmentOptimal.
		// Transition emissive to eShaderReadOnlyOptimal (where transitionGbufForResume expects it).
		{
			std::array<ImageBarrier2, 2> barriers;

			barriers[0].image = pp->getSceneColorImage();
			barriers[0].levelCount = 1;
			barriers[0].layerCount = 1;
			barriers[0].oldLayout = vk::ImageLayout::eTransferSrcOptimal;
			barriers[0].newLayout = vk::ImageLayout::eColorAttachmentOptimal;
			barriers[0].srcStage = vk::PipelineStageFlagBits2::eTransfer;
			barriers[0].srcAccess = vk::AccessFlagBits2::eTransferRead;
			barriers[0].dstStage = vk::PipelineStageFlagBits2::eColorAttachmentOutput;
			// eColorAttachmentRead as well as Write: the resumed non-MSAA G-buffer
			// pass loads attachment 0 (scene color, loadOp=eLoad), a read that must
			// be ordered after this transition -- otherwise a READ_AFTER_WRITE vs
			// the layout transition (flagged by -gr_sync_validation on the deferred
			// forward/RocketUI path).
			barriers[0].dstAccess = vk::AccessFlagBits2::eColorAttachmentWrite
			                      | vk::AccessFlagBits2::eColorAttachmentRead;

			barriers[1].image = pp->deferred().emissiveImage();
			barriers[1].levelCount = 1;
			barriers[1].layerCount = 1;
			barriers[1].oldLayout = vk::ImageLayout::eTransferDstOptimal;
			barriers[1].newLayout = vk::ImageLayout::eShaderReadOnlyOptimal;
			barriers[1].srcStage = vk::PipelineStageFlagBits2::eTransfer;
			barriers[1].srcAccess = vk::AccessFlagBits2::eTransferWrite;
			barriers[1].dstStage = vk::PipelineStageFlagBits2::eColorAttachmentOutput;
			barriers[1].dstAccess = {};

			cmdImageBarriers(cmd, ArrayView<const ImageBarrier2>(barriers.data(), barriers.size()));
		}

		// Transition G-buffer attachments 1-5 from eShaderReadOnlyOptimal → eColorAttachmentOptimal
		pp->deferred().transitionForResume(cmd);

		// Resume G-buffer render pass with eLoad
		{
			auto extent = pp->getSceneExtent();
			vk::RenderPassBeginInfo rpBegin;
			rpBegin.renderPass = pp->deferred().renderPassLoad();
			rpBegin.framebuffer = pp->deferred().framebuffer();
			rpBegin.renderArea.offset = vk::Offset2D(0, 0);
			rpBegin.renderArea.extent = extent;
			std::array<vk::ClearValue, VulkanDeferredGBuffer::GBUF_COLOR_ATTACHMENT_COUNT + 1> clearValues{};
			clearValues[VulkanDeferredGBuffer::GBUF_COLOR_ATTACHMENT_COUNT].depthStencil = vk::ClearDepthStencilValue(1.0f, 0);
			rpBegin.clearValueCount = static_cast<uint32_t>(clearValues.size());
			rpBegin.pClearValues = clearValues.data();
			cmd.beginRenderPass(rpBegin, vk::SubpassContents::eInline);
			stateTracker->setRenderPass(pp->deferred().renderPassLoad(), 0);
			// The G-buffer pass has multiple color attachments; pipelines built while
			// it is bound must match, or deferred-on cockpits render incorrectly.
			stateTracker->setColorAttachmentCount(VulkanDeferredGBuffer::GBUF_COLOR_ATTACHMENT_COUNT);
		}

		// Optionally clear non-color G-buffer attachments
		if (clearNonColorBufs) {
			vk::ClearAttachment clearAtt;
			clearAtt.aspectMask = vk::ImageAspectFlagBits::eColor;
			clearAtt.clearValue.color.setFloat32({0.0f, 0.0f, 0.0f, 0.0f});

			auto extent = pp->getSceneExtent();
			vk::ClearRect clearRect;
			clearRect.rect.offset = vk::Offset2D(0, 0);
			clearRect.rect.extent = extent;
			clearRect.baseArrayLayer = 0;
			clearRect.layerCount = 1;

			for (uint32_t att : {VulkanDeferredGBuffer::GBUF_ATT_POSITION,
		                     VulkanDeferredGBuffer::GBUF_ATT_NORMAL,
		                     VulkanDeferredGBuffer::GBUF_ATT_SPECULAR,
		                     VulkanDeferredGBuffer::GBUF_ATT_COMPOSITE}) {
				clearAtt.colorAttachment = att;
				cmd.clearAttachments(clearAtt, clearRect);
			}
		}
	}

	renderer->setUseGbufRenderPass(true);

	Deferred_lighting = true;
}

void vulkan_deferred_lighting_msaa()
{
	if (!Deferred_lighting) {
		return;
	}
	if (Cmdline_msaa_enabled <= 0) {
		return;
	}

	auto* pp = getPostProcessor();
	if (!pp || !pp->deferred().isMsaaInitialized()) {
		return;
	}

	GR_DEBUG_SCOPE("MSAA Pass");

	auto* stateTracker = getStateTracker();
	vk::CommandBuffer cmd = stateTracker->getCommandBuffer();

	// End MSAA G-buffer render pass.
	// With finalLayout == subpass layout, all attachments stay in their subpass layouts:
	// colors remain eColorAttachmentOptimal, depth remains eDepthStencilAttachmentOptimal.
	cmd.endRenderPass();

	// Reset sample count to 1x (resolve and subsequent passes are non-MSAA)
	stateTracker->setCurrentSampleCount(vk::SampleCountFlagBits::e1);

	// Transition all 6 MSAA images to eShaderReadOnlyOptimal for the resolve shader.
	// We use explicit barriers instead of render pass finalLayout transitions to
	// ensure the validation layer tracks the layout changes correctly.
	transitionMsaaImages(cmd, pp,
		vk::ImageLayout::eColorAttachmentOptimal, vk::ImageLayout::eShaderReadOnlyOptimal,
		vk::AccessFlagBits2::eColorAttachmentWrite, vk::AccessFlagBits2::eShaderSampledRead,
		vk::ImageLayout::eDepthStencilAttachmentOptimal, vk::ImageLayout::eShaderReadOnlyOptimal,
		vk::AccessFlagBits2::eDepthStencilAttachmentWrite, vk::AccessFlagBits2::eShaderSampledRead,
		vk::PipelineStageFlagBits2::eColorAttachmentOutput | vk::PipelineStageFlagBits2::eLateFragmentTests,
		vk::PipelineStageFlagBits2::eFragmentShader);

	// Begin resolve render pass (non-MSAA, writes to standard G-buffer images)
	{
		auto extent = pp->getSceneExtent();
		vk::RenderPassBeginInfo rpBegin;
		rpBegin.renderPass = pp->deferred().msaaResolveRenderPass();
		rpBegin.framebuffer = pp->deferred().msaaResolveFramebuffer();
		rpBegin.renderArea.offset = vk::Offset2D(0, 0);
		rpBegin.renderArea.extent = extent;
		// 6 attachments: 5 color + depth. loadOp=eDontCare for all (fully overwritten).
		std::array<vk::ClearValue, VulkanDeferredGBuffer::MSAA_COLOR_ATTACHMENT_COUNT + 1> clearValues{};
		clearValues[VulkanDeferredGBuffer::MSAA_COLOR_ATTACHMENT_COUNT].depthStencil = vk::ClearDepthStencilValue(1.0f, 0);
		rpBegin.clearValueCount = static_cast<uint32_t>(clearValues.size());
		rpBegin.pClearValues = clearValues.data();
		cmd.beginRenderPass(rpBegin, vk::SubpassContents::eInline);

		auto* pipelineMgr = getPipelineManager();
		auto* descriptorMgr = getDescriptorManager();
		auto* texMgr = getTextureManager();

		PipelineConfig config;
		config.shaderType = SDR_TYPE_MSAA_RESOLVE;
		switch (getRendererInstance()->getMsaaSampleCount()){
			case vk::SampleCountFlagBits::e4:
				config.shaderFlags = SDR_FLAG_MSAA_SAMPLES_4;
				break;
			case vk::SampleCountFlagBits::e8:
				config.shaderFlags = SDR_FLAG_MSAA_SAMPLES_8;
				break;
			case vk::SampleCountFlagBits::e16:
				config.shaderFlags = SDR_FLAG_MSAA_SAMPLES_16;
				break;
			default:
				config.shaderFlags = 0; //Dynamic safety fallback
				break;
		}
		config.primitiveType = PRIM_TYPE_TRIS;
		config.depthMode = ZBUFFER_TYPE_WRITE;
		config.blendMode = ALPHA_BLEND_NONE;
		config.cullEnabled = false;
		config.depthWriteEnabled = true;
		config.renderPass = pp->deferred().msaaResolveRenderPass();
		config.colorAttachmentCount = 5;

		vertex_layout emptyLayout;
		vk::Pipeline pipeline = pipelineMgr->getPipeline(config, emptyLayout);
		if (pipeline) {
			cmd.bindPipeline(vk::PipelineBindPoint::eGraphics, pipeline);

			vk::Viewport viewport;
			viewport.x = 0.0f;
			viewport.y = 0.0f;
			viewport.width = static_cast<float>(extent.width);
			viewport.height = static_cast<float>(extent.height);
			viewport.minDepth = 0.0f;
			viewport.maxDepth = 1.0f;
			cmd.setViewport(0, viewport);
			vk::Rect2D scissor;
			scissor.offset = vk::Offset2D(0, 0);
			scissor.extent = extent;
			cmd.setScissor(0, scissor);

			DescriptorWriter writer;
			writer.reset(descriptorMgr->getDevice(), descriptorMgr->getFallbacks());

			// Global set (fallback — resolve shader doesn't use global bindings)
			vk::DescriptorSet globalSet = descriptorMgr->allocateFrameSet(DescriptorSetIndex::Global);
			Assert(globalSet);
			writer.writeSet(DescriptorSetIndex::Global, globalSet);

			// Material set: All 6 MSAA textures in binding 1 array (elements 0-5)
			// [0]=color, [1]=position, [2]=normal, [3]=specular, [4]=emissive, [5]=depth
			vk::DescriptorSet materialSet = descriptorMgr->allocateFrameSet(DescriptorSetIndex::Material);
			Assert(materialSet);
			writer.writeSet(DescriptorSetIndex::Material, materialSet);

			// Build texture array: elements 0-5 are MSAA textures, 6-15 are fallback
			vk::Sampler nearestSampler = texMgr->getSampler(
				vk::Filter::eNearest, vk::Filter::eNearest,
				vk::SamplerAddressMode::eClampToEdge, false, 0.0f, false);

			std::array<vk::DescriptorImageInfo, VulkanDescriptorManager::MAX_TEXTURE_BINDINGS> texImages;
			// MSAA textures at slots 0-5
			// Fill all slots with MSAA color view (validation checks ALL elements
			// even though the shader only accesses 0-5 — sample count must match).
			texImages.fill({nearestSampler, pp->deferred().msaaColorView(), vk::ImageLayout::eShaderReadOnlyOptimal});
			texImages[1] = {nearestSampler, pp->deferred().msaaPositionView(), vk::ImageLayout::eShaderReadOnlyOptimal};
			texImages[2] = {nearestSampler, pp->deferred().msaaNormalView(), vk::ImageLayout::eShaderReadOnlyOptimal};
			texImages[3] = {nearestSampler, pp->deferred().msaaSpecularView(), vk::ImageLayout::eShaderReadOnlyOptimal};
			texImages[4] = {nearestSampler, pp->deferred().msaaEmissiveView(), vk::ImageLayout::eShaderReadOnlyOptimal};
			texImages[5] = {nearestSampler, pp->deferred().msaaDepthView(), vk::ImageLayout::eShaderReadOnlyOptimal};
			writer.setImageArray(MaterialBinding::TextureArray, texImages);

			// PerDraw set: GenericData UBO with {samples, fov} at binding 0
			vk::DescriptorSet perDrawSet = descriptorMgr->allocateFrameSet(DescriptorSetIndex::PerDraw);
			Assert(perDrawSet);
			writer.writeSet(DescriptorSetIndex::PerDraw, perDrawSet);

			struct MsaaResolveData {
				int samples;
				float fov;
			} resolveData;
			resolveData.samples = Cmdline_msaa_enabled;
			resolveData.fov = g3_get_hfov(Proj_fov);

			auto& ring = pp->context().scratchRing;
			vk::DeviceSize slotOffset =
				ring.alloc(descriptorMgr->getCurrentFrame(), &resolveData, sizeof(resolveData));
			writer.setBuffer(PerDrawBinding::GenericData, {ring.buffer(), slotOffset, ring.slotSize()});
			writer.flush();
			const vk::DescriptorSet sets[] = {globalSet, materialSet, perDrawSet};
			writer.bindSets(cmd, pipelineMgr->getPipelineLayout(), DescriptorSetIndex::Global, sets);

			cmd.draw(3, 1, 0, 0);
		}

		cmd.endRenderPass();
	}

	// Restore MSAA images to their resting layout (attachment-optimal) so they
	// match the validation layer's tracking state for the next frame.
	transitionMsaaImages(cmd, pp,
		vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageLayout::eColorAttachmentOptimal,
		vk::AccessFlagBits2::eShaderSampledRead, vk::AccessFlagBits2::eColorAttachmentWrite,
		vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageLayout::eDepthStencilAttachmentOptimal,
		vk::AccessFlagBits2::eShaderSampledRead, vk::AccessFlagBits2::eDepthStencilAttachmentWrite,
		vk::PipelineStageFlagBits2::eFragmentShader,
		vk::PipelineStageFlagBits2::eColorAttachmentOutput | vk::PipelineStageFlagBits2::eEarlyFragmentTests);

	// After resolve, the non-MSAA G-buffer has properly resolved data.
	// Color attachments 0-4 are in eShaderReadOnlyOptimal (from resolve pass finalLayout).
	// Depth is in eDepthStencilAttachmentOptimal.
	// Subsequent deferred_lighting_end/finish operate on the non-MSAA G-buffer unchanged.

	// Transition scene color from eShaderReadOnlyOptimal → eColorAttachmentOptimal
	// (deferred_lighting_end resumes the non-MSAA gbuf pass and needs scene color writable)
	{
		ImageBarrier2 barrier;
		barrier.image = pp->getSceneColorImage();
		barrier.levelCount = 1;
		barrier.layerCount = 1;
		barrier.oldLayout = vk::ImageLayout::eShaderReadOnlyOptimal;
		barrier.newLayout = vk::ImageLayout::eColorAttachmentOptimal;
		barrier.srcStage = vk::PipelineStageFlagBits2::eColorAttachmentOutput;
		barrier.srcAccess = vk::AccessFlagBits2::eColorAttachmentWrite;
		barrier.dstStage = vk::PipelineStageFlagBits2::eColorAttachmentOutput;
		// eColorAttachmentRead as well as Write: the resumed non-MSAA G-buffer
		// pass loads attachment 0 (scene color, loadOp=eLoad), a read that must
		// be ordered after this transition (else READ_AFTER_WRITE vs the layout
		// transition, flagged by -gr_sync_validation).
		barrier.dstAccess = vk::AccessFlagBits2::eColorAttachmentWrite
		                  | vk::AccessFlagBits2::eColorAttachmentRead;

		cmdImageBarrier(cmd, barrier);
	}

	// Composite is not part of the resolve framebuffer, so its layout is
	// indeterminate (UNDEFINED on first frame, eTransferSrcOptimal from
	// previous frame's composite→scene copy, etc.). Use oldLayout=eUndefined
	// to transition it regardless of current state — content will be fully
	// overwritten by emissive→composite copy in deferred_lighting_finish().
	{
		ImageBarrier2 barrier;
		barrier.image = pp->deferred().compositeImage();
		barrier.levelCount = 1;
		barrier.layerCount = 1;
		barrier.oldLayout = vk::ImageLayout::eUndefined;
		barrier.newLayout = vk::ImageLayout::eShaderReadOnlyOptimal;
		barrier.srcStage = vk::PipelineStageFlagBits2::eTopOfPipe;
		barrier.srcAccess = {};
		barrier.dstStage = vk::PipelineStageFlagBits2::eColorAttachmentOutput;
		barrier.dstAccess = {};

		cmdImageBarrier(cmd, barrier);
	}

	// Transition G-buffer attachments 1-5 for resume
	// (all now in eShaderReadOnlyOptimal: 1-4 from resolve finalLayout, 5 from above)
	pp->deferred().transitionForResume(cmd);

	// Resume the non-MSAA G-buffer render pass with eLoad
	{
		auto extent = pp->getSceneExtent();
		vk::RenderPassBeginInfo rpBegin;
		rpBegin.renderPass = pp->deferred().renderPassLoad();
		rpBegin.framebuffer = pp->deferred().framebuffer();
		rpBegin.renderArea.offset = vk::Offset2D(0, 0);
		rpBegin.renderArea.extent = extent;
		std::array<vk::ClearValue, 7> clearValues{};
		clearValues[6].depthStencil = vk::ClearDepthStencilValue(1.0f, 0);
		rpBegin.clearValueCount = static_cast<uint32_t>(clearValues.size());
		rpBegin.pClearValues = clearValues.data();
		cmd.beginRenderPass(rpBegin, vk::SubpassContents::eInline);
		stateTracker->setRenderPass(pp->deferred().renderPassLoad(), 0);
		stateTracker->setColorAttachmentCount(VulkanDeferredGBuffer::GBUF_COLOR_ATTACHMENT_COUNT);
	}
}

void vulkan_deferred_lighting_end()
{
	if (!Deferred_lighting) {
		return;
	}

	Deferred_lighting = false;

	// After this, rendering goes back to writing only attachment 0.
	// The pipeline still has 6 blend states (matching the G-buffer render pass)
	// but the shader only outputs to location 0. Attachments 1-5 are untouched.
}

void vulkan_deferred_lighting_finish()
{
	if (!light_deferred_enabled()) {
		return;
	}

	auto* pp = getPostProcessor();
	if (!pp || !pp->deferred().isInitialized()) {
		return;
	}

	auto* renderer = getRendererInstance();
	if (!renderer->isSceneRendering()) {
		return;
	}

	GR_DEBUG_SCOPE("Deferred lighting finish");

	auto* stateTracker = getStateTracker();
	vk::CommandBuffer cmd = stateTracker->getCommandBuffer();

	// TLAS for RTAO: when shadow rendering is off (Shadow_quality Disabled, or a
	// frame that skips shadows_render_all entirely), nothing has built this
	// frame's TLAS yet -- request it here. No-op when the shadow path already
	// built it (buildTlas()'s per-frame guard). Acceleration-structure builds
	// must be recorded outside a render pass, so this ends the G-buffer pass and
	// clears it from the state tracker -- which is why step 1 below only ends
	// the pass when it is still active.
	if (rtao_enabled()) {
		vulkan_build_shadow_tlas();
	}

	// 1. End G-buffer render pass (unless the RTAO TLAS build above already did)
	// All 6 color attachments → eShaderReadOnlyOptimal
	// Depth → eDepthStencilAttachmentOptimal
	if (stateTracker->getCurrentRenderPass()) {
		cmd.endRenderPass();
	}

	// 2. Copy emissive → composite (the emissive data becomes the base for light accumulation)
	// Emissive → eShaderReadOnlyOptimal (done), composite → eColorAttachmentOptimal (for light accum)
	copyImageToImage(cmd,
		pp->deferred().emissiveImage(), vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageLayout::eShaderReadOnlyOptimal,
		pp->deferred().compositeImage(), vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageLayout::eColorAttachmentOptimal,
		pp->getSceneExtent());

	// 3. Render deferred lights (begins + ends light accum render pass internally)
	// After this, composite is in eShaderReadOnlyOptimal
	pp->renderDeferredLights(cmd);

	// 4. Fog rendering (between light accumulation and forward rendering)
	// Matches OpenGL flow in opengl_deferred_lighting_finish()
	bool bDrawFullNeb = The_mission.flags[Mission::Mission_Flags::Fullneb]
		&& Neb2_render_mode != NEB2_RENDER_NONE && !s_vulkanOverrideFog;
	bool bDrawNebVolumetrics = The_mission.volumetrics
		&& The_mission.volumetrics->get_enabled() && !s_vulkanOverrideFog;

	bool fogRendered = false;
	if (bDrawFullNeb) {
		// Scene fog reads composite + depth → writes scene color
		pp->renderSceneFog(cmd);
		fogRendered = true;

		if (bDrawNebVolumetrics) {
			// Copy scene color → composite so volumetric reads the fogged result
			copyImageToImage(cmd,
				pp->getSceneColorImage(), vk::ImageLayout::eColorAttachmentOptimal, vk::ImageLayout::eColorAttachmentOptimal,
				pp->deferred().compositeImage(), vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageLayout::eShaderReadOnlyOptimal,
				pp->getSceneExtent());
		}
	}
	if (bDrawNebVolumetrics) {
		// Volumetric fog reads composite + emissive + depth + 3D volumes → writes scene color
		pp->renderVolumetricFog(cmd);
		fogRendered = true;
	}

	if (!fogRendered) {
		// No fog — copy composite → scene color (existing behavior)
		// Skip src post-barrier (composite not used again in this path)
		copyImageToImage(cmd,
			pp->deferred().compositeImage(), vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageLayout::eTransferSrcOptimal,
			pp->getSceneColorImage(), vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageLayout::eColorAttachmentOptimal,
			pp->getSceneExtent());
	}

	// 5. Switch to scene render pass for forward transparent objects
	// After light accumulation, use the 2-attachment scene render pass instead
	// of the 6-attachment G-buffer pass. Forward-rendered transparent objects
	// only write to fragOut0 — using the G-buffer pass would leave undefined
	// values at attachment locations 1-5.
	renderer->setUseGbufRenderPass(false);
	stateTracker->setColorAttachmentCount(1);

	// Resume scene render pass (loadOp=eLoad) with depth preserved
	{
		auto extent = pp->getSceneExtent();
		vk::RenderPassBeginInfo rpBegin;
		rpBegin.renderPass = pp->getSceneRenderPassLoad();
		rpBegin.framebuffer = pp->getSceneFramebuffer();
		rpBegin.renderArea.offset = vk::Offset2D(0, 0);
		rpBegin.renderArea.extent = extent;
		std::array<vk::ClearValue, 2> clearValues;
		clearValues[0].color.setFloat32({0.0f, 0.0f, 0.0f, 1.0f});
		clearValues[1].depthStencil = vk::ClearDepthStencilValue(1.0f, 0);
		rpBegin.clearValueCount = static_cast<uint32_t>(clearValues.size());
		rpBegin.pClearValues = clearValues.data();
		cmd.beginRenderPass(rpBegin, vk::SubpassContents::eInline);
		stateTracker->setRenderPass(pp->getSceneRenderPassLoad(), 0);
	}
}

void vulkan_override_fog(bool set_override) {
	s_vulkanOverrideFog = set_override;
}

// ========== Shadow Map Rendering ==========

} // namespace graphics::vulkan



extern bool gr_htl_projection_matrix_set;


namespace graphics::vulkan {

namespace {
bool Glowpoint_override_save = false;
bool Restore_swapchain_after_shadow_pass = false;
// Non-null when the shadow pass was started while rendering into an off-screen render target
// (e.g. SCPUI's 3D ship/weapon select preview). In that case the shadow pass must resume the
// render target's pass afterwards, not the swapchain, or subsequent draws leak onto the screen.
tcache_slot_vulkan* Restore_render_target_after_shadow_pass = nullptr;
bool Shadow_pass_active = false;
} // anonymous namespace

void vulkan_shadow_map_start(matrix4* shadow_view_matrix, const matrix* light_matrix, vec3d* eye_pos, bool first_pass)
{
	if (Shadow_quality == ShadowQuality::Disabled || !getRendererInstance()->supportsShaderViewportLayerOutput()) {
		return;
	}

	auto* pp = getPostProcessor();
	if (!pp) {
		return;
	}

	// Lazy-init shadow resources (framebuffer/render pass). The CPU-side cascade
	// bookkeeping (Shadow_frustums/Shadow_proj_matrix/Shadow_cascade_distances)
	// is sized eagerly at Vulkan startup via shadow_cascade_params_init() (see
	// graphics::vulkan::initialize() in gr_vulkan.cpp) -- shadows_start_render()
	// indexes those vectors before this function is ever called, so that can't
	// be deferred to first use here.
	if (!pp->shadow().isInitialized()) {
		if (!pp->initShadowPass()) {
			return;
		}
	}

	auto* stateTracker = getStateTracker();
	vk::CommandBuffer cmd = stateTracker->getCommandBuffer();

	if (first_pass) {
		// End the current render pass but keep track if we need to resume the swapchain render pass or the scene render pass afterwards.
		Restore_swapchain_after_shadow_pass = !getRendererInstance()->isSceneRendering();

		// If shadows are being rendered while an off-screen render target is bound (SCPUI 3D
		// ship/weapon select preview renders a rotating model via draw_model_rotating into an RTT),
		// remember that target so vulkan_shadow_map_end() resumes ITS pass instead of the swapchain.
		Restore_render_target_after_shadow_pass = nullptr;
		if (getRendererInstance()->isRenderTargetActive()) {
			auto* texMgr = getTextureManager();
			const int rtHandle = texMgr ? texMgr->getCurrentRenderTarget() : -1;
			if (texMgr && rtHandle >= 0) {
				Restore_render_target_after_shadow_pass = texMgr->getTextureSlot(rtHandle);
			}
		}

		// vulkan_build_shadow_tlas() (called just before this, every frame shadows
		// are enabled) already ends whatever render pass was active -- TLAS builds
		// are illegal inside a render pass instance -- and nulls out the tracked
		// render pass. Only end it here if that didn't already happen.
		if (stateTracker->getCurrentRenderPass()) {
			cmd.endRenderPass();
		}

		// Shadow render pass is always non-MSAA (1x sample count)
		stateTracker->setCurrentSampleCount(vk::SampleCountFlagBits::e1);

		// Begin shadow render pass (eClear, depth-only)
		{
			int shadowSize = pp->shadow().textureSize();
			vk::RenderPassBeginInfo rpBegin;
			rpBegin.renderPass = pp->shadow().renderPass();
			rpBegin.framebuffer = pp->shadow().framebuffer();
			rpBegin.renderArea.offset = vk::Offset2D(0, 0);
			rpBegin.renderArea.extent = vk::Extent2D(static_cast<uint32_t>(shadowSize), static_cast<uint32_t>(shadowSize));

			vk::ClearValue clearValue;
			clearValue.depthStencil = vk::ClearDepthStencilValue(1.0f, 0);
			rpBegin.clearValueCount = 1;
			rpBegin.pClearValues = &clearValue;

			cmd.beginRenderPass(rpBegin, vk::SubpassContents::eInline);
			stateTracker->setRenderPass(pp->shadow().renderPass(), 0);
			stateTracker->setColorAttachmentCount(0);
		}

		// Set viewport and scissor to shadow texture size
		{
			const int shadowSize = pp->shadow().textureSize();

			stateTracker->setViewport(0.0f, 0.0f, static_cast<float>(shadowSize), static_cast<float>(shadowSize), 0.0f, 1.0f);
			stateTracker->setScissor(0, 0, static_cast<uint32_t>(shadowSize), static_cast<uint32_t>(shadowSize));
		}

		Shadow_pass_active = true;
		Glowpoint_override_save = Glowpoint_override;
		Glowpoint_override = true;

		gr_htl_projection_matrix_set = true;
	} else {
		gr_end_view_matrix();
	}

	gr_set_view_matrix(eye_pos, light_matrix);

	*shadow_view_matrix = gr_view_matrix;
}

void vulkan_shadow_map_end()
{
	if (!Shadow_pass_active) {
		return;
	}

	auto* pp = getPostProcessor();
	auto* stateTracker = getStateTracker();
	auto* renderer = getRendererInstance();
	vk::CommandBuffer cmd = stateTracker->getCommandBuffer();

	gr_end_view_matrix();
	Shadow_pass_active = false;

	gr_zbuffer_set(ZBUFFER_TYPE_FULL);

	Glowpoint_override = Glowpoint_override_save;
	gr_htl_projection_matrix_set = false;

	// End shadow render pass (color transitions to eShaderReadOnlyOptimal via finalLayout)
	cmd.endRenderPass();

	if ( Restore_render_target_after_shadow_pass ) {
		// Shadows were rendered inside an off-screen render target; resume its pass (loadOp=eLoad
		// so the pre-shadow contents are preserved) rather than the swapchain, otherwise the
		// following model draws would land on the screen at the RTT's viewport (top-left corner).
		renderer->resumeRenderTargetPass(Restore_render_target_after_shadow_pass);
		Restore_render_target_after_shadow_pass = nullptr;
	} else if ( Restore_swapchain_after_shadow_pass ) {
		renderer->resumeSwapChainPass();
	} else if ( renderer->isUsingGbufRenderPass() ) {
		const bool msaaActive = (Cmdline_msaa_enabled > 0 && pp->deferred().isMsaaInitialized());

		if ( msaaActive ) {
			// Resume MSAA G-buffer render pass
			pp->deferred().transitionMsaaForResume(cmd);

			auto extent = pp->getSceneExtent();
			vk::RenderPassBeginInfo rpBegin;
			rpBegin.renderPass = pp->deferred().msaaRenderPassLoad();
			rpBegin.framebuffer = pp->deferred().msaaFramebuffer();
			rpBegin.renderArea.offset = vk::Offset2D(0, 0);
			rpBegin.renderArea.extent = extent;
			std::array<vk::ClearValue, VulkanDeferredGBuffer::MSAA_COLOR_ATTACHMENT_COUNT + 1> clearValues{};
			clearValues[VulkanDeferredGBuffer::MSAA_COLOR_ATTACHMENT_COUNT].depthStencil = vk::ClearDepthStencilValue(1.0f, 0);
			rpBegin.clearValueCount = static_cast<uint32_t>(clearValues.size());
			rpBegin.pClearValues = clearValues.data();
			cmd.beginRenderPass(rpBegin, vk::SubpassContents::eInline);
			stateTracker->setRenderPass(pp->deferred().msaaRenderPassLoad(), 0);
			stateTracker->setColorAttachmentCount(VulkanDeferredGBuffer::MSAA_COLOR_ATTACHMENT_COUNT);
			stateTracker->setCurrentSampleCount(getRendererInstance()->getMsaaSampleCount());
		} else {
			// Transition scene color: eShaderReadOnlyOptimal → eColorAttachmentOptimal
			// (Scene color was in eShaderReadOnlyOptimal from ending G-buffer pass before shadow start)
			{
				ImageBarrier2 barrier;
				barrier.image = pp->getSceneColorImage();
				barrier.levelCount = 1;
				barrier.layerCount = 1;
				barrier.oldLayout = vk::ImageLayout::eShaderReadOnlyOptimal;
				barrier.newLayout = vk::ImageLayout::eColorAttachmentOptimal;
				barrier.srcStage = vk::PipelineStageFlagBits2::eTopOfPipe;
				barrier.srcAccess = {};
				barrier.dstStage = vk::PipelineStageFlagBits2::eColorAttachmentOutput;
				// eColorAttachmentRead too: the resumed G-buffer pass (renderPassLoad,
				// loadOp=eLoad) reads this attachment; order the read after the transition.
				barrier.dstAccess = vk::AccessFlagBits2::eColorAttachmentWrite
				                  | vk::AccessFlagBits2::eColorAttachmentRead;

				cmdImageBarrier(cmd, barrier);
			}

			// Transition G-buffer attachments 1-5 for resume
			pp->deferred().transitionForResume(cmd);

			// Resume G-buffer render pass with eLoad
			auto extent = pp->getSceneExtent();
			vk::RenderPassBeginInfo rpBegin;
			rpBegin.renderPass = pp->deferred().renderPassLoad();
			rpBegin.framebuffer = pp->deferred().framebuffer();
			rpBegin.renderArea.offset = vk::Offset2D(0, 0);
			rpBegin.renderArea.extent = extent;

			std::array<vk::ClearValue, 7> clearValues{};
			clearValues[6].depthStencil = vk::ClearDepthStencilValue(1.0f, 0);
			rpBegin.clearValueCount = static_cast<uint32_t>(clearValues.size());
			rpBegin.pClearValues = clearValues.data();

			cmd.beginRenderPass(rpBegin, vk::SubpassContents::eInline);
			stateTracker->setRenderPass(pp->deferred().renderPassLoad(), 0);
			stateTracker->setColorAttachmentCount(VulkanDeferredGBuffer::GBUF_COLOR_ATTACHMENT_COUNT);
		}
	} else {
		renderer->resumeSceneRendering();
	}

	// Restore viewport and scissor to scene size
	{
		vk::Viewport viewport;
		viewport.x = static_cast<float>(gr_screen.offset_x);
		viewport.y = static_cast<float>(gr_screen.offset_y);
		viewport.width = static_cast<float>(gr_screen.clip_width);
		viewport.height = static_cast<float>(gr_screen.clip_height);
		viewport.minDepth = 0.0f;
		viewport.maxDepth = 1.0f;
		cmd.setViewport(0, viewport);

		vk::Rect2D scissor;
		scissor.offset = vk::Offset2D(gr_screen.offset_x, gr_screen.offset_y);
		scissor.extent = vk::Extent2D(static_cast<uint32_t>(gr_screen.clip_width), static_cast<uint32_t>(gr_screen.clip_height));
		cmd.setScissor(0, scissor);
	}
}

// ========== Decal Pass ==========

void vulkan_start_decal_pass()
{
	auto* renderer = getRendererInstance();
	auto* pp = getPostProcessor();
	auto* stateTracker = getStateTracker();

	if (!renderer->isSceneRendering() || !pp || !pp->deferred().isInitialized()) {
		return;
	}

	vk::CommandBuffer cmd = stateTracker->getCommandBuffer();

	// End the G-buffer render pass (transitions all color attachments to eShaderReadOnlyOptimal)
	cmd.endRenderPass();

	// Copy scene depth → samplable depth copy (for fragment depth reconstruction)
	pp->copySceneDepth(cmd);

	// Copy G-buffer normal → samplable normal copy (for angle rejection)
	pp->deferred().copyNormal(cmd);

	// Transition scene color: eShaderReadOnlyOptimal → eColorAttachmentOptimal
	{
		ImageBarrier2 barrier;
		barrier.image = pp->getSceneColorImage();
		barrier.levelCount = 1;
		barrier.layerCount = 1;
		barrier.oldLayout = vk::ImageLayout::eShaderReadOnlyOptimal;
		barrier.newLayout = vk::ImageLayout::eColorAttachmentOptimal;
		barrier.srcStage = vk::PipelineStageFlagBits2::eTopOfPipe;
		barrier.srcAccess = {};
		barrier.dstStage = vk::PipelineStageFlagBits2::eColorAttachmentOutput;
		// eColorAttachmentRead too: the resumed G-buffer pass (renderPassLoad,
		// loadOp=eLoad) reads this attachment; order the read after the transition.
		barrier.dstAccess = vk::AccessFlagBits2::eColorAttachmentWrite
		                  | vk::AccessFlagBits2::eColorAttachmentRead;

		cmdImageBarrier(cmd, barrier);
	}

	// Transition G-buffer attachments 1-5 for render pass resume
	pp->deferred().transitionForResume(cmd);

	// Resume G-buffer render pass with eLoad
	{
		auto extent = pp->getSceneExtent();
		vk::RenderPassBeginInfo rpBegin;
		rpBegin.renderPass = pp->deferred().renderPassLoad();
		rpBegin.framebuffer = pp->deferred().framebuffer();
		rpBegin.renderArea.offset = vk::Offset2D(0, 0);
		rpBegin.renderArea.extent = extent;

		std::array<vk::ClearValue, 7> clearValues{};
		clearValues[6].depthStencil = vk::ClearDepthStencilValue(1.0f, 0);
		rpBegin.clearValueCount = static_cast<uint32_t>(clearValues.size());
		rpBegin.pClearValues = clearValues.data();

		cmd.beginRenderPass(rpBegin, vk::SubpassContents::eInline);
		stateTracker->setRenderPass(pp->deferred().renderPassLoad(), 0);
		stateTracker->setColorAttachmentCount(VulkanDeferredGBuffer::GBUF_COLOR_ATTACHMENT_COUNT);
	}

	// Restore viewport (Y-flipped for Vulkan scene rendering)
	auto extent = pp->getSceneExtent();
	stateTracker->setViewport(0.0f,
		static_cast<float>(extent.height),
		static_cast<float>(extent.width),
		-static_cast<float>(extent.height));
}

void vulkan_stop_decal_pass()
{
	// No-op — decals draw within the resumed G-buffer render pass
}

void vulkan_render_decals(decal_material* material_info,
                          primitive_type prim_type,
                          vertex_layout* layout,
                          int num_elements,
                          const indexed_vertex_source& buffers,
                          const gr_buffer_handle& instance_buffer,
                          int num_instances)
{
	if (!material_info || !layout || num_instances <= 0) {
		return;
	}

	auto* stateTracker = getStateTracker();
	auto* pipelineManager = getPipelineManager();
	auto* descManager = getDescriptorManager();
	auto* bufferManager = getBufferManager();
	auto* drawManager = getDrawManager();
	auto* texManager = getTextureManager();
	auto* pp = getPostProcessor();

	// Set up matrices
	gr_matrix_set_uniforms();

	// Build pipeline config for decal rendering
	PipelineConfig config;
	config.shaderType = SDR_TYPE_DECAL;
	config.shaderFlags = material_info->get_shader_flags();
	config.primitiveType = prim_type;
	config.depthMode = material_info->get_depth_mode();
	config.depthWriteEnabled = false;
	config.cullEnabled = false;
	config.frontFaceCW = false;
	config.blendMode = material_info->get_blend_mode();
	config.renderPass = stateTracker->getCurrentRenderPass();
	config.colorAttachmentCount = stateTracker->getColorAttachmentCount();

	// Per-attachment blend: active attachments (color, normal, emissive) get
	// the material's blend mode with RGB-only write mask. Inactive attachments get
	// write mask = 0 to avoid corrupting G-buffer data.
	config.perAttachmentBlendEnabled = true;
	for (uint32_t i = 0; i < config.colorAttachmentCount; ++i) {
		config.attachmentBlends[i].blendMode = ALPHA_BLEND_NONE;
		config.attachmentBlends[i].writeMask = {false, false, false, false};
	}
	// Color/diffuse — use material blend mode 0
	config.attachmentBlends[VulkanDeferredGBuffer::GBUF_ATT_COLOR].blendMode = material_info->get_blend_mode(0);
	config.attachmentBlends[VulkanDeferredGBuffer::GBUF_ATT_COLOR].writeMask = {true, true, true, false};
	// Normal — always additive
	config.attachmentBlends[VulkanDeferredGBuffer::GBUF_ATT_NORMAL].blendMode = ALPHA_BLEND_ADDITIVE;
	config.attachmentBlends[VulkanDeferredGBuffer::GBUF_ATT_NORMAL].writeMask = {true, true, true, false};
	// Emissive — use material blend mode 2
	config.attachmentBlends[VulkanDeferredGBuffer::GBUF_ATT_EMISSIVE].blendMode = material_info->get_blend_mode(2);
	config.attachmentBlends[VulkanDeferredGBuffer::GBUF_ATT_EMISSIVE].writeMask = {true, true, true, false};

	// Get or create pipeline
	vk::Pipeline pipeline = pipelineManager->getPipeline(config, *layout);
	if (!pipeline) {
		nprintf(("vulkan", "vulkan_render_decals: Failed to get pipeline!\n"));
		return;
	}

	stateTracker->bindPipeline(pipeline, pipelineManager->getPipelineLayout());

	// Helper: get DescriptorBufferInfo from pending binding (null buffer = fallback)
	auto getPendingBufInfo = [&](size_t blockIdx) -> vk::DescriptorBufferInfo {
		const auto& pending = drawManager->getPendingUniformBinding(blockIdx);
		if (pending.valid) {
			vk::Buffer buf = bufferManager->getVkBuffer(pending.bufferHandle);
			if (buf) {
				return {buf, pending.offset, pending.size};
			}
		}
		return {};
	};

	DescriptorWriter writer;
	writer.reset(descManager->getDevice(), descManager->getFallbacks());

	// Set 0: Global (all fallback)
	vk::DescriptorSet globalSet = descManager->allocateFrameSet(DescriptorSetIndex::Global);
	Assert(globalSet);
	writer.writeSet(DescriptorSetIndex::Global, globalSet);

	// Set 1: Material
	vk::DescriptorSet materialSet = descManager->allocateFrameSet(DescriptorSetIndex::Material);
	Assert(materialSet);
	writer.writeSet(DescriptorSetIndex::Material, materialSet);

	// Binding 2: DecalGlobals UBO
	writer.setBuffer(MaterialBinding::DecalGlobals,
		getPendingBufInfo(static_cast<size_t>(uniform_block_type::DecalGlobals)));

	// Binding 1: decal textures (diffuse, glow, normal as texture array)
	drawManager->bindMaterialTextures(material_info, &writer);

	// Binding 4: scene depth copy (for fragment depth reconstruction)
	{
		vk::Sampler nearestSampler = texManager->getSampler(
			vk::Filter::eNearest, vk::Filter::eNearest,
			vk::SamplerAddressMode::eClampToEdge, false, 0.0f, false);
		vk::ImageView depthView = pp->getSceneDepthCopyView();
		if (depthView && nearestSampler) {
			writer.setImage(MaterialBinding::DepthMap, {nearestSampler, depthView, vk::ImageLayout::eShaderReadOnlyOptimal});
		}
	}

	// Binding 6: G-buffer normal copy (for angle rejection)
	{
		vk::Sampler nearestSampler = texManager->getSampler(
			vk::Filter::eNearest, vk::Filter::eNearest,
			vk::SamplerAddressMode::eClampToEdge, false, 0.0f, false);
		vk::ImageView normalView = pp->deferred().normalCopyView();
		if (normalView && nearestSampler) {
			writer.setImage(MaterialBinding::DistortionMap, {nearestSampler, normalView, vk::ImageLayout::eShaderReadOnlyOptimal});
		}
	}

	// Set 2: PerDraw
	vk::DescriptorSet perDrawSet = descManager->allocateFrameSet(DescriptorSetIndex::PerDraw);
	Assert(perDrawSet);
	writer.writeSet(DescriptorSetIndex::PerDraw, perDrawSet);
	writer.setBuffer(PerDrawBinding::Matrices,
		getPendingBufInfo(static_cast<size_t>(uniform_block_type::Matrices)));
	writer.setBuffer(PerDrawBinding::DecalInfo,
		getPendingBufInfo(static_cast<size_t>(uniform_block_type::DecalInfo)));
	writer.flush();
	stateTracker->bindDescriptorSet(DescriptorSetIndex::Global, globalSet, nullptr);
	stateTracker->bindDescriptorSet(DescriptorSetIndex::Material,
		materialSet,
		writer.dynamicOffsets(DescriptorSetIndex::Material));
	stateTracker->bindDescriptorSet(DescriptorSetIndex::PerDraw,
		perDrawSet,
		writer.dynamicOffsets(DescriptorSetIndex::PerDraw));

	// Bind vertex buffers: binding 0 = box VBO, binding 1 = instance buffer
	vk::Buffer boxVBO = bufferManager->getVkBuffer(buffers.Vbuffer_handle);
	vk::Buffer boxIBO = bufferManager->getVkBuffer(buffers.Ibuffer_handle);
	vk::Buffer instBuf = bufferManager->getVkBuffer(instance_buffer);

	if (!boxVBO || !boxIBO || !instBuf) {
		nprintf(("vulkan", "vulkan_render_decals: Missing buffer(s)!\n"));
		return;
	}

	stateTracker->bindVertexBuffer(0, boxVBO, 0);

	// Instance buffer needs frame base offset for streaming buffers
	size_t instFrameOffset = bufferManager->getFrameBaseOffset(instance_buffer);
	stateTracker->bindVertexBuffer(1, instBuf, static_cast<vk::DeviceSize>(instFrameOffset));

	stateTracker->bindIndexBuffer(boxIBO, 0, vk::IndexType::eUint32);

	// Flush dynamic state and draw
	stateTracker->applyDynamicState();

	auto cmdBuffer = stateTracker->getCommandBuffer();
	cmdBuffer.drawIndexed(
		static_cast<uint32_t>(num_elements),  // index count
		static_cast<uint32_t>(num_instances), // instance count
		0,                                     // first index
		0,                                     // vertex offset
		0                                      // first instance
	);
}

} // namespace graphics::vulkan