Merge pull request #17139 from hrydgard/replacement-padding-support

Texture replacement: Improve padding support

Common/MemoryUtil.cpp

@@ -250,7 +250,7 @@ void *AllocateAlignedMemory(size_t size, size_t alignment) {
 #endif
 #endif
 
-	_assert_msg_(ptr != nullptr, "Failed to allocate aligned memory");
+	_assert_msg_(ptr != nullptr, "Failed to allocate aligned memory of size %llu", size);
 	return ptr;
 }
 

Core/FileSystems/MetaFileSystem.cpp

@@ -79,7 +79,6 @@ static bool RealPath(const std::string &currentDirectory, const std::string &inP
 	size_t inLen = inPath.length();
 	if (inLen == 0)
 	{
-		WARN_LOG(FILESYS, "RealPath: inPath is empty");
 		outPath = currentDirectory;
 		return true;
 	}

GPU/Common/ReplacedTexture.cpp

@@ -141,7 +141,7 @@ void ReplacedTexture::PurgeIfNotUsedSinceTime(double t) {
 	alphaStatus_ = ReplacedTextureAlpha::UNKNOWN;
 
 	// This means we have to reload.  If we never purge any, there's no need.
-	SetState(ReplacementState::POPULATED);
+	SetState(ReplacementState::UNLOADED);
 }
 
 // This can only return true if ACTIVE or NOT_FOUND.
@@ -165,13 +165,10 @@ bool ReplacedTexture::IsReady(double budget) {
 		}
 		lastUsed_ = now;
 		return true;
-	case ReplacementState::UNINITIALIZED:
-		// _dbg_assert_(false);
-		return false;
 	case ReplacementState::CANCEL_INIT:
 	case ReplacementState::PENDING:
 		return false;
-	case ReplacementState::POPULATED:
+	case ReplacementState::UNLOADED:
 		// We're gonna need to spawn a task.
 		break;
 	}
@@ -195,6 +192,10 @@ bool ReplacedTexture::IsReady(double budget) {
 	return false;
 }
 
+inline uint32_t RoundUpTo4(uint32_t value) {
+	return (value + 3) & ~3;
+}
+
 void ReplacedTexture::Prepare(VFSBackend *vfs) {
 	this->vfs_ = vfs;
 
@@ -259,16 +260,26 @@ void ReplacedTexture::Prepare(VFSBackend *vfs) {
 		return;
 	}
 
+	// Update the level dimensions.
+	for (auto &level : levels_) {
+		level.fullW = (level.w * desc_.w) / desc_.newW;
+		level.fullH = (level.h * desc_.h) / desc_.newH;
+
+		int blockSize;
+		bool bc = Draw::DataFormatIsBlockCompressed(fmt, &blockSize);
+		if (!bc) {
+			level.fullDataSize = level.fullW * level.fullH * 4;
+		} else {
+			level.fullDataSize = RoundUpTo4(level.fullW) * RoundUpTo4(level.fullH) * blockSize / 16;
+		}
+	}
+
 	SetState(ReplacementState::ACTIVE);
 
 	if (threadWaitable_)
 		threadWaitable_->Notify();
 }
 
-inline uint32_t RoundUpTo4(uint32_t value) {
-	return (value + 3) & ~3;
-}
-
 // Returns true if Prepare should keep calling this to load more levels.
 ReplacedTexture::LoadLevelResult ReplacedTexture::LoadLevelData(VFSFileReference *fileRef, const std::string &filename, int mipLevel, Draw::DataFormat *pixelFormat) {
 	bool good = false;
@@ -393,21 +404,17 @@ ReplacedTexture::LoadLevelResult ReplacedTexture::LoadLevelData(VFSFileReference
 		ERROR_LOG(G3D, "Could not load texture replacement info: %s - unsupported format %s", filename.c_str(), magic.c_str());
 	}
 
-
 	// Already populated from cache. TODO: Move this above the first read, and take level.w/h from the cache.
 	if (!data_[mipLevel].empty()) {
 		vfs_->CloseFile(openFile);
 		*pixelFormat = fmt;
 		return LoadLevelResult::DONE;
 	}
 
-	// Is this really the right place to do it?
-	level.w = (level.w * desc_.w) / desc_.newW;
-	level.h = (level.h * desc_.h) / desc_.newH;
-
 	if (good && mipLevel != 0) {
-		// Check that the mipmap size is correct.  Can't load mips of the wrong size.
-		if (level.w != (levels_[0].w >> mipLevel) || level.h != (levels_[0].h >> mipLevel)) {
+		// If loading a low mip directly (through png most likely), check that the mipmap size is correct.
+		// Can't load mips of the wrong size.
+		if (level.w != std::max(1, (levels_[0].w >> mipLevel)) || level.h != std::max(1, (levels_[0].h >> mipLevel))) {
 			WARN_LOG(G3D, "Replacement mipmap invalid: size=%dx%d, expected=%dx%d (level %d)",
 				level.w, level.h, levels_[0].w >> mipLevel, levels_[0].h >> mipLevel, mipLevel);
 			good = false;
@@ -662,7 +669,7 @@ ReplacedTexture::LoadLevelResult ReplacedTexture::LoadLevelData(VFSFileReference
 	return LoadLevelResult::LOAD_ERROR;
 }
 
-bool ReplacedTexture::CopyLevelTo(int level, void *out, int rowPitch) {
+bool ReplacedTexture::CopyLevelTo(int level, uint8_t *out, size_t outDataSize, int rowPitch) {
 	_assert_msg_((size_t)level < levels_.size(), "Invalid miplevel");
 	_assert_msg_(out != nullptr && rowPitch > 0, "Invalid out/pitch");
 
@@ -671,6 +678,13 @@ bool ReplacedTexture::CopyLevelTo(int level, void *out, int rowPitch) {
 		return false;
 	}
 
+	// We pad the images right here during the copy.
+	// TODO: Add support for the texture cache to scale texture coordinates instead.
+	// It already supports this for render target textures that aren't powers of 2.
+
+	int outW = levels_[level].fullW;
+	int outH = levels_[level].fullH;
+
 	// We probably could avoid this lock, but better to play it safe.
 	std::lock_guard<std::mutex> guard(lock_);
 
@@ -684,9 +698,15 @@ bool ReplacedTexture::CopyLevelTo(int level, void *out, int rowPitch) {
 
 #define PARALLEL_COPY
 
-	if (fmt == Draw::DataFormat::R8G8B8A8_UNORM) {
+	int blockSize;
+	if (!Draw::DataFormatIsBlockCompressed(fmt, &blockSize)) {
+		if (fmt != Draw::DataFormat::R8G8B8A8_UNORM) {
+			ERROR_LOG(G3D, "Unexpected linear data format");
+			return false;
+		}
+
 		if (rowPitch < info.w * 4) {
-			ERROR_LOG(G3D, "Replacement rowPitch=%d, but w=%d (level=%d)", rowPitch, info.w * 4, level);
+			ERROR_LOG(G3D, "Replacement rowPitch=%d, but w=%d (level=%d) (too small)", rowPitch, info.w * 4, level);
 			return false;
 		}
 
@@ -702,32 +722,62 @@ bool ReplacedTexture::CopyLevelTo(int level, void *out, int rowPitch) {
 #ifdef PARALLEL_COPY
 			const int MIN_LINES_PER_THREAD = 4;
 			ParallelRangeLoop(&g_threadManager, [&](int l, int h) {
+				int extraPixels = outW - info.w;
 				for (int y = l; y < h; ++y) {
 					memcpy((uint8_t *)out + rowPitch * y, data.data() + info.w * 4 * y, info.w * 4);
+					// Fill the rest of the line with black.
+					memset((uint8_t *)out + rowPitch * y + info.w * 4, 0, extraPixels * 4);
 				}
 				}, 0, info.h, MIN_LINES_PER_THREAD);
 #else
 			for (int y = 0; y < info.h; ++y) {
 				memcpy((uint8_t *)out + rowPitch * y, data.data() + info.w * 4 * y, info.w * 4);
 			}
 #endif
+			// Memset the rest of the padding to avoid leaky edge pixels. Guess we could parallelize this too, but meh.
+			for (int y = info.h; y < outH; y++) {
+				uint8_t *dest = (uint8_t *)out + rowPitch * y;
+				memset(dest, 0, outW * 4);
+			}
 		}
 	} else {
 #ifdef PARALLEL_COPY
-		// TODO: Add sanity checks here for other formats?
-		ParallelMemcpy(&g_threadManager, out, data.data(), data.size());
-#else
-		memcpy(out, data.data(), data.size());
+		// Only parallel copy in the simple case for now.
+		if (info.w == outW && info.h == outH) {
+			// TODO: Add sanity checks here for other formats?
+			ParallelMemcpy(&g_threadManager, out, data.data(), data.size());
+			return true;
+		}
 #endif
+		// Alright, so careful copying of blocks it is, padding with zero-blocks as needed.
+		int inBlocksW = (info.w + 3) / 4;
+		int inBlocksH = (info.h + 3) / 4;
+		int outBlocksW = (info.fullW + 3) / 4;
+		int outBlocksH = (info.fullH + 3) / 4;
+
+		int paddingBlocksX = outBlocksW - inBlocksW;
+
+		// Copy all the known blocks, and zero-fill out the lines.
+		for (int y = 0; y < inBlocksH; y++) {
+			const uint8_t *input = data.data() + y * inBlocksW * blockSize;
+			uint8_t *output = (uint8_t *)out + y * outBlocksW * blockSize;
+			memcpy(output, input, inBlocksW * blockSize);
+			memset(output + inBlocksW * blockSize, 0, paddingBlocksX * blockSize);
+		}
+
+		// Vertical zero-padding.
+		for (int y = inBlocksH; y < outBlocksH; y++) {
+			uint8_t *output = (uint8_t *)out + y * outBlocksW * blockSize;
+			memset(output, 0, outBlocksW * blockSize);
+		}
 	}
 
 	return true;
 }
 
 const char *StateString(ReplacementState state) {
 	switch (state) {
-	case ReplacementState::UNINITIALIZED: return "UNINITIALIZED";
-	case ReplacementState::POPULATED: return "PREPARED";
+	case ReplacementState::UNLOADED: return "PREPARED";
 	case ReplacementState::PENDING: return "PENDING";
 	case ReplacementState::NOT_FOUND: return "NOT_FOUND";
 	case ReplacementState::ACTIVE: return "ACTIVE";

GPU/Common/ReplacedTexture.h

@@ -52,8 +52,7 @@ enum class ReplacedImageType {
 static const int MAX_REPLACEMENT_MIP_LEVELS = 12;  // 12 should be plenty, 8 is the max mip levels supported by the PSP.
 
 enum class ReplacementState : uint32_t {
-	UNINITIALIZED,
-	POPULATED,  // We located the texture files but have not started the thread.
+	UNLOADED,
 	PENDING,
 	NOT_FOUND,  // Also used on error loading the images.
 	ACTIVE,
@@ -94,8 +93,14 @@ struct ReplacedTextureRef {
 
 // Metadata about a given texture level.
 struct ReplacedTextureLevel {
+	// Data dimensions
 	int w = 0;
 	int h = 0;
+	// PSP texture dimensions
+	int fullW = 0;
+	int fullH = 0;
+
+	int fullDataSize = 0;
 
 	// To be able to reload, we need to be able to reopen, unfortunate we can't use zip_file_t.
 	// TODO: This really belongs on the level in the cache, not in the individual ReplacedTextureLevel objects.
@@ -122,13 +127,13 @@ class ReplacedTexture {
 	void GetSize(int level, int *w, int *h) const {
 		_dbg_assert_(State() == ReplacementState::ACTIVE);
 		_dbg_assert_(level < levels_.size());
-		*w = levels_[level].w;
-		*h = levels_[level].h;
+		*w = levels_[level].fullW;
+		*h = levels_[level].fullH;
 	}
 
-	int GetLevelDataSize(int level) const {
-		_dbg_assert_(State() == ReplacementState::ACTIVE);
-		return (int)data_[level].size();
+	int GetLevelDataSizeAfterCopy(int level) const {
+		// Includes padding etc.
+		return levels_[level].fullDataSize;
 	}
 
 	size_t GetTotalDataSize() const {
@@ -157,7 +162,7 @@ class ReplacedTexture {
 	}
 
 	bool IsReady(double budget);
-	bool CopyLevelTo(int level, void *out, int rowPitch);
+	bool CopyLevelTo(int level, uint8_t *out, size_t outDataSize, int rowPitch);
 
 	std::string logId_;
 
@@ -182,7 +187,7 @@ class ReplacedTexture {
 	ReplacedTextureAlpha alphaStatus_ = ReplacedTextureAlpha::UNKNOWN;
 	double lastUsed = 0.0;
 
-	std::atomic<ReplacementState> state_ = ReplacementState::POPULATED;
+	std::atomic<ReplacementState> state_ = ReplacementState::UNLOADED;
 
 	VFSBackend *vfs_ = nullptr;
 	ReplacementDesc desc_;

GPU/Common/TextureCacheCommon.cpp

@@ -1545,9 +1545,8 @@ ReplacedTexture *TextureCacheCommon::FindReplacement(TexCacheEntry *entry, int &
 	}
 
 	switch (replaced->State()) {
-	case ReplacementState::POPULATED:
+	case ReplacementState::UNLOADED:
 	case ReplacementState::PENDING:
-	case ReplacementState::UNINITIALIZED:
 		// Make sure we keep polling.
 		entry->status |= TexCacheEntry::STATUS_TO_REPLACE;
 		break;
@@ -2883,7 +2882,8 @@ bool TextureCacheCommon::PrepareBuildTexture(BuildTexturePlan &plan, TexCacheEnt
 	return true;
 }
 
-void TextureCacheCommon::LoadTextureLevel(TexCacheEntry &entry, uint8_t *data, int stride, BuildTexturePlan &plan, int srcLevel, Draw::DataFormat dstFmt, TexDecodeFlags texDecFlags) {
+// Passing 0 into dataSize will disable checking.
+void TextureCacheCommon::LoadTextureLevel(TexCacheEntry &entry, uint8_t *data, size_t dataSize, int stride, BuildTexturePlan &plan, int srcLevel, Draw::DataFormat dstFmt, TexDecodeFlags texDecFlags) {
 	int w = gstate.getTextureWidth(srcLevel);
 	int h = gstate.getTextureHeight(srcLevel);
 
@@ -2892,7 +2892,7 @@ void TextureCacheCommon::LoadTextureLevel(TexCacheEntry &entry, uint8_t *data, i
 	if (plan.replaceValid) {
 		plan.replaced->GetSize(srcLevel, &w, &h);
 		double replaceStart = time_now_d();
-		plan.replaced->CopyLevelTo(srcLevel, data, stride);
+		plan.replaced->CopyLevelTo(srcLevel, data, dataSize, stride);
 		replacementTimeThisFrame_ += time_now_d() - replaceStart;
 	} else {
 		GETextureFormat tfmt = (GETextureFormat)entry.format;

GPU/Common/TextureCacheCommon.h

@@ -180,6 +180,7 @@ struct TexCacheEntry {
 	u32 fullhash;
 	u32 cluthash;
 	u16 maxSeenV;
+	ReplacedTexture *replacedTexture;
 
 	TexStatus GetHashStatus() {
 		return TexStatus(status & STATUS_MASK);
@@ -385,7 +386,7 @@ class TextureCacheCommon {
 	ReplacedTexture *FindReplacement(TexCacheEntry *entry, int &w, int &h, int &d);
 
 	// Return value is mapData normally, but could be another buffer allocated with AllocateAlignedMemory.
-	void LoadTextureLevel(TexCacheEntry &entry, uint8_t *mapData, int mapRowPitch, BuildTexturePlan &plan, int srcLevel, Draw::DataFormat dstFmt, TexDecodeFlags texDecFlags);
+	void LoadTextureLevel(TexCacheEntry &entry, uint8_t *mapData, size_t dataSize, int mapRowPitch, BuildTexturePlan &plan, int srcLevel, Draw::DataFormat dstFmt, TexDecodeFlags texDecFlags);
 
 	template <typename T>
 	inline const T *GetCurrentClut() {

GPU/Common/TextureReplacer.cpp

@@ -904,12 +904,14 @@ bool TextureReplacer::GenerateIni(const std::string &gameID, Path &generatedFile
 
 		// Let's also write some defaults.
 		fprintf(f, R"(# This file is optional and describes your textures.
-# Some information on syntax available here:
-# https://github.com/hrydgard/ppsspp/wiki/Texture-replacement-ini-syntax
+# Documentation about the options and syntax is available here:
+# https://www.ppsspp.org/docs/reference/texture-replacement
 [options]
 version = 1
 hash = quick
-ignoreMipmap = false
+ignoreMipmap = false  # Set to true to avoid dumping mipmaps. Instead use basisu to generate them, see docs.
+reduceHash = false  # Usually a good idea to use.
+allowVideo = false
 
 [games]
 # Used to make it easier to install, and override settings for other regions.
@@ -921,8 +923,11 @@ ignoreMipmap = false
 # See wiki for more info.
 
 [hashranges]
+# See the documentation.
+# Example: 08b31020,512,512 = 480,272
 
 [filtering]
+# You can enforce specific filtering modes with this. See the docs.
 
 [reducehashranges]
 )", gameID.c_str(), INI_FILENAME.c_str());

GPU/Common/TextureReplacer.h

@@ -138,18 +138,20 @@ class TextureReplacer {
 	bool allowVideo_ = false;
 	bool ignoreAddress_ = false;
 	bool reduceHash_ = false;
+	bool ignoreMipmap_ = false;
+
 	float reduceHashSize = 1.0f; // default value with reduceHash to false
 	float reduceHashGlobalValue = 0.5f; // Global value for textures dump pngs of all sizes, 0.5 by default but can be set in textures.ini
 
 	double lastTextureCacheSizeGB_ = 0.0;
-	bool ignoreMipmap_ = false;
 	std::string gameID_;
 	Path basePath_;
 	Path newTextureDir_;
 	ReplacedTextureHash hash_ = ReplacedTextureHash::QUICK;
 
 	VFSBackend *vfs_ = nullptr;
 	bool vfsIsZip_ = false;
+
 	GPUFormatSupport formatSupport_{};
 
 	typedef std::pair<int, int> WidthHeightPair;

GPU/D3D11/TextureCacheD3D11.cpp

@@ -301,8 +301,8 @@ void TextureCacheD3D11::BuildTexture(TexCacheEntry *const entry) {
 		if (plan.replaceValid) {
 			int blockSize = 0;
 			if (Draw::DataFormatIsBlockCompressed(plan.replaced->Format(), &blockSize)) {
-				stride = ((mipWidth + 3) & ~3) * blockSize / 4;  // This stride value doesn't quite make sense to me, but it works?
-				dataSize = plan.replaced->GetLevelDataSize(i);
+				stride = ((mipWidth + 3) & ~3) * blockSize / 4;  // Number of blocks * 4 * Size of a block / 4
+				dataSize = plan.replaced->GetLevelDataSizeAfterCopy(i);
 			} else {
 				int bpp = (int)Draw::DataFormatSizeInBytes(plan.replaced->Format());
 				stride = std::max(mipWidth * bpp, 16);
@@ -338,7 +338,7 @@ void TextureCacheD3D11::BuildTexture(TexCacheEntry *const entry) {
 			return;
 		}
 
-		LoadTextureLevel(*entry, data, stride, plan, srcLevel, texFmt, TexDecodeFlags{});
+		LoadTextureLevel(*entry, data, 0, stride, plan, srcLevel, texFmt, TexDecodeFlags{});
 	}
 
 	int tw;