Merge pull request #15884 from unknownbrackets/texreplace-io-async

Replacement: Read files only within time budget

Core/TextureReplacer.cpp

@@ -49,7 +49,8 @@ static const int VERSION = 1;
 static const int MAX_MIP_LEVELS = 12;  // 12 should be plenty, 8 is the max mip levels supported by the PSP.
 
 TextureReplacer::TextureReplacer() {
-	none_.alphaStatus_ = ReplacedTextureAlpha::UNKNOWN;
+	none_.initDone_ = true;
+	none_.prepareDone_ = true;
 }
 
 TextureReplacer::~TextureReplacer() {
@@ -373,7 +374,7 @@ u32 TextureReplacer::ComputeHash(u32 addr, int bufw, int w, int h, GETextureForm
 	}
 }
 
-ReplacedTexture &TextureReplacer::FindReplacement(u64 cachekey, u32 hash, int w, int h) {
+ReplacedTexture &TextureReplacer::FindReplacement(u64 cachekey, u32 hash, int w, int h, double budget) {
 	// Only actually replace if we're replacing.  We might just be saving.
 	if (!Enabled() || !g_Config.bReplaceTextures) {
 		return none_;
@@ -382,13 +383,18 @@ ReplacedTexture &TextureReplacer::FindReplacement(u64 cachekey, u32 hash, int w,
 	ReplacementCacheKey replacementKey(cachekey, hash);
 	auto it = cache_.find(replacementKey);
 	if (it != cache_.end()) {
+		if (!it->second.prepareDone_ && budget > 0.0) {
+			// We don't do this on a thread, but we only do it while within budget.
+			PopulateReplacement(&it->second, cachekey, hash, w, h);
+		}
 		return it->second;
 	}
 
 	// Okay, let's construct the result.
 	ReplacedTexture &result = cache_[replacementKey];
-	result.alphaStatus_ = ReplacedTextureAlpha::UNKNOWN;
-	PopulateReplacement(&result, cachekey, hash, w, h);
+	if (!g_Config.bReplaceTexturesAllowLate || budget > 0.0) {
+		PopulateReplacement(&result, cachekey, hash, w, h);
+	}
 	return result;
 }
 
@@ -433,7 +439,7 @@ void TextureReplacer::PopulateReplacement(ReplacedTexture *result, u64 cachekey,
 			break;
 	}
 
-	result->alphaStatus_ = ReplacedTextureAlpha::UNKNOWN;
+	result->prepareDone_ = true;
 }
 
 enum class ReplacedImageType {
@@ -815,11 +821,13 @@ bool ReplacedTexture::IsReady(double budget) {
 	}
 
 	// Loaded already, or not yet on a thread?
-	if (!levelData_.empty())
+	if (initDone_ && !levelData_.empty())
 		return true;
 	// Let's not even start a new texture if we're already behind.
 	if (budget < 0.0)
 		return false;
+	if (!prepareDone_)
+		return false;
 
 	if (g_Config.bReplaceTexturesAllowLate) {
 		if (threadWaitable_)
@@ -829,10 +837,11 @@ bool ReplacedTexture::IsReady(double budget) {
 
 		if (threadWaitable_->WaitFor(budget)) {
 			// If we finished all the levels, we're done.
-			return !levelData_.empty();
+			return initDone_ && !levelData_.empty();
 		}
 	} else {
 		Prepare();
+		_assert_(initDone_);
 		return true;
 	}
 
@@ -842,16 +851,19 @@ bool ReplacedTexture::IsReady(double budget) {
 
 void ReplacedTexture::Prepare() {
 	std::unique_lock<std::mutex> lock(mutex_);
-	if (cancelPrepare_)
+	if (cancelPrepare_) {
+		initDone_ = true;
 		return;
+	}
 
-	levelData_.resize(NumLevels());
-	for (int i = 0; i < NumLevels(); ++i) {
+	levelData_.resize(levels_.size());
+	for (int i = 0; i < (int)levels_.size(); ++i) {
 		if (cancelPrepare_)
 			break;
 		PrepareData(i);
 	}
 
+	initDone_ = true;
 	if (!cancelPrepare_ && threadWaitable_)
 		threadWaitable_->Notify();
 }
@@ -975,6 +987,8 @@ bool ReplacedTexture::Load(int level, void *out, int rowPitch) {
 	_assert_msg_((size_t)level < levels_.size(), "Invalid miplevel");
 	_assert_msg_(out != nullptr && rowPitch > 0, "Invalid out/pitch");
 
+	if (!initDone_)
+		return false;
 	if (levelData_.empty())
 		return false;
 

Core/TextureReplacer.h

@@ -118,10 +118,20 @@ struct ReplacedTexture {
 	~ReplacedTexture();
 
 	inline bool Valid() const {
+		if (!initDone_)
+			return false;
 		return !levels_.empty();
 	}
 
+	inline bool IsInvalid() const {
+		if (!initDone_)
+			return false;
+		return levels_.empty();
+	}
+
 	bool GetSize(int level, int &w, int &h) const {
+		if (!initDone_)
+			return false;
 		if ((size_t)level < levels_.size()) {
 			w = levels_[level].w;
 			h = levels_[level].h;
@@ -131,12 +141,16 @@ struct ReplacedTexture {
 	}
 
 	int NumLevels() const {
+		if (!initDone_)
+			return 0;
 		return (int)levels_.size();
 	}
 
 	Draw::DataFormat Format(int level) const {
-		if ((size_t)level < levels_.size()) {
-			return levels_[level].fmt;
+		if (initDone_) {
+			if ((size_t)level < levels_.size()) {
+				return levels_[level].fmt;
+			}
 		}
 		return Draw::DataFormat::R8G8B8A8_UNORM;
 	}
@@ -156,11 +170,13 @@ struct ReplacedTexture {
 
 	std::vector<ReplacedTextureLevel> levels_;
 	std::vector<std::vector<uint8_t>> levelData_;
-	ReplacedTextureAlpha alphaStatus_;
+	ReplacedTextureAlpha alphaStatus_ = ReplacedTextureAlpha::UNKNOWN;
 	double lastUsed_ = 0.0;
 	LimitedWaitable *threadWaitable_ = nullptr;
 	std::mutex mutex_;
 	bool cancelPrepare_ = false;
+	bool initDone_ = false;
+	bool prepareDone_ = false;
 
 	friend TextureReplacer;
 	friend ReplacedTextureTask;
@@ -196,7 +212,7 @@ class TextureReplacer {
 
 	u32 ComputeHash(u32 addr, int bufw, int w, int h, GETextureFormat fmt, u16 maxSeenV);
 
-	ReplacedTexture &FindReplacement(u64 cachekey, u32 hash, int w, int h);
+	ReplacedTexture &FindReplacement(u64 cachekey, u32 hash, int w, int h, double budget);
 	bool FindFiltering(u64 cachekey, u32 hash, TextureFiltering *forceFiltering);
 	ReplacedTexture &FindNone() {
 		return none_;

GPU/Common/TextureCacheCommon.cpp

@@ -482,7 +482,14 @@ TexCacheEntry *TextureCacheCommon::SetTexture() {
 			int h0 = gstate.getTextureHeight(0);
 			int d0 = 1;
 			ReplacedTexture &replaced = FindReplacement(entry, w0, h0, d0);
-			if (replaced.Valid()) {
+			if (replaced.IsInvalid()) {
+				entry->status &= ~TexCacheEntry::STATUS_TO_REPLACE;
+				if (g_Config.bSaveNewTextures) {
+					// Load once more to actually save.
+					match = false;
+					reason = "replacing";
+				}
+			} else {
 				match = false;
 				reason = "replacing";
 			}
@@ -1303,22 +1310,22 @@ ReplacedTexture &TextureCacheCommon::FindReplacement(TexCacheEntry *entry, int &
 	constexpr double MAX_BUDGET_PER_TEX = 0.25 / 60.0;
 
 	double replaceStart = time_now_d();
+	double budget = std::min(MAX_BUDGET_PER_TEX, replacementFrameBudget_ - replacementTimeThisFrame_);
 	u64 cachekey = replacer_.Enabled() ? entry->CacheKey() : 0;
-	ReplacedTexture &replaced = replacer_.FindReplacement(cachekey, entry->fullhash, w, h);
-	if (replaced.IsReady(std::min(MAX_BUDGET_PER_TEX, replacementFrameBudget_ - replacementTimeThisFrame_))) {
+	ReplacedTexture &replaced = replacer_.FindReplacement(cachekey, entry->fullhash, w, h, budget);
+	if (replaced.IsReady(budget)) {
 		if (replaced.GetSize(0, w, h)) {
-			replacementTimeThisFrame_ += time_now_d() - replaceStart;
-
-			// Consider it already "scaled" and remove any delayed replace flag.
+			// Consider it already "scaled."
 			entry->status |= TexCacheEntry::STATUS_IS_SCALED;
-			entry->status &= ~TexCacheEntry::STATUS_TO_REPLACE;
-			return replaced;
 		}
-	} else if (replaced.Valid()) {
+
+		// Remove the flag, even if it was invalid.
+		entry->status &= ~TexCacheEntry::STATUS_TO_REPLACE;
+	} else if (!replaced.IsInvalid()) {
 		entry->status |= TexCacheEntry::STATUS_TO_REPLACE;
 	}
 	replacementTimeThisFrame_ += time_now_d() - replaceStart;
-	return replacer_.FindNone();
+	return replaced;
 }
 
 // This is only used in the GLES backend, where we don't point these to video memory.
@@ -2375,7 +2382,7 @@ void TextureCacheCommon::LoadTextureLevel(TexCacheEntry &entry, uint8_t *data, i
 			}
 		}
 
-		if (replacer_.Enabled()) {
+		if (replacer_.Enabled() && replaced.IsInvalid()) {
 			ReplacedTextureDecodeInfo replacedInfo;
 			replacedInfo.cachekey = entry.CacheKey();
 			replacedInfo.hash = entry.fullhash;

GPU/Vulkan/TextureCacheVulkan.cpp

@@ -464,7 +464,7 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 	// Any texture scaling is gonna move away from the original 16-bit format, if any.
 	VkFormat actualFmt = plan.scaleFactor > 1 ? VULKAN_8888_FORMAT : dstFmt;
 	if (plan.replaced->Valid()) {
-		actualFmt = ToVulkanFormat(plan.replaced->Format(0));
+		actualFmt = ToVulkanFormat(plan.replaced->Format(plan.baseLevelSrc));
 	}
 
 	bool computeUpload = false;
@@ -536,7 +536,7 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 
 	ReplacedTextureDecodeInfo replacedInfo;
 	bool willSaveTex = false;
-	if (replacer_.Enabled() && !plan.replaced->Valid() && plan.depth == 1) {
+	if (replacer_.Enabled() && plan.replaced->IsInvalid() && plan.depth == 1) {
 		replacedInfo.cachekey = entry->CacheKey();
 		replacedInfo.hash = entry->fullhash;
 		replacedInfo.addr = entry->addr;
@@ -565,7 +565,7 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 		int mipWidth = mipUnscaledWidth * plan.scaleFactor;
 		int mipHeight = mipUnscaledHeight * plan.scaleFactor;
 		if (plan.replaced->Valid()) {
-			plan.replaced->GetSize(i, mipWidth, mipHeight);
+			plan.replaced->GetSize(plan.baseLevelSrc + i, mipWidth, mipHeight);
 		}
 
 		int bpp = actualFmt == VULKAN_8888_FORMAT ? 4 : 2;  // output bpp
@@ -595,7 +595,7 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 			// Directly load the replaced image.
 			data = drawEngine_->GetPushBufferForTextureData()->PushAligned(size, &bufferOffset, &texBuf, pushAlignment);
 			double replaceStart = time_now_d();
-			plan.replaced->Load(i, data, stride);  // if it fails, it'll just be garbage data... OK for now.
+			plan.replaced->Load(plan.baseLevelSrc + i, data, stride);  // if it fails, it'll just be garbage data... OK for now.
 			replacementTimeThisFrame_ += time_now_d() - replaceStart;
 			VK_PROFILE_BEGIN(vulkan, cmdInit, VK_PIPELINE_STAGE_TRANSFER_BIT,
 				"Copy Upload (replaced): %dx%d", mipWidth, mipHeight);
@@ -636,7 +636,7 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 				int w = dataScaled ? mipWidth : mipUnscaledWidth;
 				int h = dataScaled ? mipHeight : mipUnscaledHeight;
 				// At this point, data should be saveData, and not slow.
-				replacer_.NotifyTextureDecoded(replacedInfo, data, stride, i, w, h);
+				replacer_.NotifyTextureDecoded(replacedInfo, data, stride, plan.baseLevelSrc + i, w, h);
 			}
 		}
 	}