thirdparty/ppsspp
2
0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
ppsspp/GPU/Directx9/TextureCacheDX9.cpp
Unknown W. Brackets 1dfecbf2b8 d3d: Actually upload the clut texture.
2015-03-01 11:27:28 -08:00

1798 lines
58 KiB
C++
Raw Blame History

// Copyright (c) 2012- PPSSPP Project.
// This program 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, version 2.0 or later versions.
// This program 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 2.0 for more details.
// A copy of the GPL 2.0 should have been included with the program.
// If not, see http://www.gnu.org/licenses/
// Official git repository and contact information can be found at
// https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
#include <map>
#include <algorithm>
#include <cstring>
#include "Core/MemMap.h"
#include "Core/Reporting.h"
#include "GPU/ge_constants.h"
#include "GPU/GPUState.h"
#include "GPU/Directx9/PixelShaderGeneratorDX9.h"
#include "GPU/Directx9/TextureCacheDX9.h"
#include "GPU/Directx9/FramebufferDX9.h"
#include "GPU/Directx9/ShaderManagerDX9.h"
#include "GPU/Directx9/DepalettizeShaderDX9.h"
#include "GPU/Directx9/helper/dx_state.h"
#include "GPU/Common/FramebufferCommon.h"
#include "GPU/Common/TextureDecoder.h"
#include "Core/Config.h"
#include "Core/Host.h"
#include "ext/xxhash.h"
#include "math/math_util.h"
extern int g_iNumVideos;
namespace DX9 {
#define INVALID_TEX (LPDIRECT3DTEXTURE9)(-1)
// If a texture hasn't been seen for this many frames, get rid of it.
#define TEXTURE_KILL_AGE 200
#define TEXTURE_KILL_AGE_LOWMEM 60
// Not used in lowmem mode.
#define TEXTURE_SECOND_KILL_AGE 100
// Try to be prime to other decimation intervals.
#define TEXCACHE_DECIMATION_INTERVAL 13
// Changes more frequent than this will be considered "frequent" and prevent texture scaling.
#define TEXCACHE_FRAME_CHANGE_FREQUENT 6
#define TEXCACHE_MAX_TEXELS_SCALED (256*256) // Per frame
#define TEXCACHE_MIN_PRESSURE 16 * 1024 * 1024 // Total in VRAM
#define TEXCACHE_SECOND_MIN_PRESSURE 4 * 1024 * 1024
TextureCacheDX9::TextureCacheDX9() : cacheSizeEstimate_(0), secondCacheSizeEstimate_(0), clearCacheNextFrame_(false), lowMemoryMode_(false), clutBuf_(NULL), clutMaxBytes_(0), texelsScaledThisFrame_(0) {
timesInvalidatedAllThisFrame_ = 0;
lastBoundTexture = INVALID_TEX;
decimationCounter_ = TEXCACHE_DECIMATION_INTERVAL;
// This is 5MB of temporary storage. Might be possible to shrink it.
tmpTexBuf32.resize(1024 * 512); // 2MB
tmpTexBuf16.resize(1024 * 512); // 1MB
tmpTexBufRearrange.resize(1024 * 512); // 2MB
// Aren't these way too big?
clutBufConverted_ = (u32 *)AllocateAlignedMemory(4096 * sizeof(u32), 16); // 16KB
clutBufRaw_ = (u32 *)AllocateAlignedMemory(4096 * sizeof(u32), 16); // 16KB
// Zap these so that reads from uninitialized parts of the CLUT look the same in
// release and debug
memset(clutBufConverted_, 0, 4096 * sizeof(u32));
memset(clutBufRaw_, 0, 4096 * sizeof(u32));
D3DCAPS9 pCaps;
ZeroMemory(&pCaps, sizeof(pCaps));
HRESULT result = 0;
if (pD3DdeviceEx) {
result = pD3DdeviceEx->GetDeviceCaps(&pCaps);
} else {
result = pD3Ddevice->GetDeviceCaps(&pCaps);
}
if (FAILED(result)) {
WARN_LOG(G3D, "Failed to get the device caps!");
maxAnisotropyLevel = 16;
} else {
maxAnisotropyLevel = pCaps.MaxAnisotropy;
}
SetupTextureDecoder();
}
TextureCacheDX9::~TextureCacheDX9() {
Clear(true);
FreeAlignedMemory(clutBufConverted_);
FreeAlignedMemory(clutBufRaw_);
}
static u32 EstimateTexMemoryUsage(const TextureCacheDX9::TexCacheEntry *entry) {
const u16 dim = entry->dim;
const u8 dimW = ((dim >> 0) & 0xf);
const u8 dimH = ((dim >> 8) & 0xf);
u32 pixelSize = 2;
switch (entry->format) {
case GE_TFMT_CLUT4:
case GE_TFMT_CLUT8:
case GE_TFMT_CLUT16:
case GE_TFMT_CLUT32:
// We assume cluts always point to 8888 for simplicity.
pixelSize = 4;
break;
case GE_TFMT_4444:
case GE_TFMT_5551:
case GE_TFMT_5650:
break;
case GE_TFMT_8888:
case GE_TFMT_DXT1:
case GE_TFMT_DXT3:
case GE_TFMT_DXT5:
default:
pixelSize = 4;
break;
}
// This in other words multiplies by w and h.
return pixelSize << (dimW + dimH);
}
void TextureCacheDX9::Clear(bool delete_them) {
pD3Ddevice->SetTexture(0, NULL);
lastBoundTexture = INVALID_TEX;
if (delete_them) {
for (TexCache::iterator iter = cache.begin(); iter != cache.end(); ++iter) {
DEBUG_LOG(G3D, "Deleting texture %p", iter->second.texture);
iter->second.ReleaseTexture();
}
for (TexCache::iterator iter = secondCache.begin(); iter != secondCache.end(); ++iter) {
DEBUG_LOG(G3D, "Deleting texture %p", iter->second.texture);
iter->second.ReleaseTexture();
}
}
if (cache.size() + secondCache.size()) {
INFO_LOG(G3D, "Texture cached cleared from %i textures", (int)(cache.size() + secondCache.size()));
cache.clear();
secondCache.clear();
cacheSizeEstimate_ = 0;
secondCacheSizeEstimate_ = 0;
}
fbTexInfo_.clear();
}
void TextureCacheDX9::DeleteTexture(TexCache::iterator it) {
it->second.ReleaseTexture();
auto fbInfo = fbTexInfo_.find(it->second.addr);
if (fbInfo != fbTexInfo_.end()) {
fbTexInfo_.erase(fbInfo);
}
cacheSizeEstimate_ -= EstimateTexMemoryUsage(&it->second);
cache.erase(it);
}
void TextureCacheDX9::ForgetLastTexture() {
lastBoundTexture = INVALID_TEX;
gstate_c.textureChanged |= TEXCHANGE_PARAMSONLY;
}
// Removes old textures.
void TextureCacheDX9::Decimate() {
if (--decimationCounter_ <= 0) {
decimationCounter_ = TEXCACHE_DECIMATION_INTERVAL;
} else {
return;
}
if (cacheSizeEstimate_ >= TEXCACHE_MIN_PRESSURE) {
const u32 had = cacheSizeEstimate_;
pD3Ddevice->SetTexture(0, NULL);
lastBoundTexture = INVALID_TEX;
int killAge = lowMemoryMode_ ? TEXTURE_KILL_AGE_LOWMEM : TEXTURE_KILL_AGE;
for (TexCache::iterator iter = cache.begin(); iter != cache.end(); ) {
if (iter->second.lastFrame + killAge < gpuStats.numFlips) {
DeleteTexture(iter++);
} else {
++iter;
}
}
VERBOSE_LOG(G3D, "Decimated texture cache, saved %d estimated bytes - now %d bytes", had - cacheSizeEstimate_, cacheSizeEstimate_);
}
if (g_Config.bTextureSecondaryCache && secondCacheSizeEstimate_ >= TEXCACHE_SECOND_MIN_PRESSURE) {
const u32 had = secondCacheSizeEstimate_;
for (TexCache::iterator iter = secondCache.begin(); iter != secondCache.end(); ) {
// In low memory mode, we kill them all.
if (lowMemoryMode_ || iter->second.lastFrame + TEXTURE_KILL_AGE < gpuStats.numFlips) {
iter->second.ReleaseTexture();
secondCache.erase(iter++);
} else {
++iter;
}
}
VERBOSE_LOG(G3D, "Decimated second texture cache, saved %d estimated bytes - now %d bytes", had - secondCacheSizeEstimate_, secondCacheSizeEstimate_);
}
}
void TextureCacheDX9::Invalidate(u32 addr, int size, GPUInvalidationType type) {
// If we're hashing every use, without backoff, then this isn't needed.
if (!g_Config.bTextureBackoffCache) {
return;
}
addr &= 0x0FFFFFFF;
u32 addr_end = addr + size;
// They could invalidate inside the texture, let's just give a bit of leeway.
const int LARGEST_TEXTURE_SIZE = 512 * 512 * 4;
const u64 startKey = (u64)(addr - LARGEST_TEXTURE_SIZE) << 32;
u64 endKey = (u64)(addr + size + LARGEST_TEXTURE_SIZE) << 32;
if (endKey < startKey) {
endKey = (u64)-1;
}
for (TexCache::iterator iter = cache.lower_bound(startKey), end = cache.upper_bound(endKey); iter != end; ++iter) {
u32 texAddr = iter->second.addr;
u32 texEnd = iter->second.addr + iter->second.sizeInRAM;
if (texAddr < addr_end && addr < texEnd) {
if (iter->second.GetHashStatus() == TexCacheEntry::STATUS_RELIABLE) {
iter->second.SetHashStatus(TexCacheEntry::STATUS_HASHING);
}
if (type != GPU_INVALIDATE_ALL) {
gpuStats.numTextureInvalidations++;
// Start it over from 0 (unless it's safe.)
iter->second.numFrames = type == GPU_INVALIDATE_SAFE ? 256 : 0;
iter->second.framesUntilNextFullHash = 0;
} else if (!iter->second.framebuffer) {
iter->second.invalidHint++;
}
}
}
}
void TextureCacheDX9::InvalidateAll(GPUInvalidationType /*unused*/) {
// If we're hashing every use, without backoff, then this isn't needed.
if (!g_Config.bTextureBackoffCache) {
return;
}
if (timesInvalidatedAllThisFrame_ > 5) {
return;
}
timesInvalidatedAllThisFrame_++;
for (TexCache::iterator iter = cache.begin(), end = cache.end(); iter != end; ++iter) {
if (iter->second.GetHashStatus() == TexCacheEntry::STATUS_RELIABLE) {
iter->second.SetHashStatus(TexCacheEntry::STATUS_HASHING);
}
if (!iter->second.framebuffer) {
iter->second.invalidHint++;
}
}
}
void TextureCacheDX9::ClearNextFrame() {
clearCacheNextFrame_ = true;
}
void TextureCacheDX9::AttachFramebufferValid(TexCacheEntry *entry, VirtualFramebuffer *framebuffer, const AttachedFramebufferInfo &fbInfo) {
const bool hasInvalidFramebuffer = entry->framebuffer == nullptr || entry->invalidHint == -1;
const bool hasOlderFramebuffer = entry->framebuffer != nullptr && entry->framebuffer->last_frame_render < framebuffer->last_frame_render;
bool hasFartherFramebuffer = false;
if (!hasInvalidFramebuffer && !hasOlderFramebuffer) {
// If it's valid, but the offset is greater, then we still win.
if (fbTexInfo_[entry->addr].yOffset == fbInfo.yOffset)
hasFartherFramebuffer = fbTexInfo_[entry->addr].xOffset > fbInfo.xOffset;
else
hasFartherFramebuffer = fbTexInfo_[entry->addr].yOffset > fbInfo.yOffset;
}
if (hasInvalidFramebuffer || hasOlderFramebuffer || hasFartherFramebuffer) {
if (entry->framebuffer == nullptr) {
cacheSizeEstimate_ -= EstimateTexMemoryUsage(entry);
}
entry->framebuffer = framebuffer;
entry->invalidHint = 0;
entry->status &= ~TextureCacheDX9::TexCacheEntry::STATUS_DEPALETTIZE;
fbTexInfo_[entry->addr] = fbInfo;
framebuffer->last_frame_attached = gpuStats.numFlips;
host->GPUNotifyTextureAttachment(entry->addr);
} else if (entry->framebuffer == framebuffer) {
framebuffer->last_frame_attached = gpuStats.numFlips;
}
}
void TextureCacheDX9::AttachFramebufferInvalid(TexCacheEntry *entry, VirtualFramebuffer *framebuffer, const AttachedFramebufferInfo &fbInfo) {
if (entry->framebuffer == 0 || entry->framebuffer == framebuffer) {
if (entry->framebuffer == nullptr) {
cacheSizeEstimate_ -= EstimateTexMemoryUsage(entry);
}
entry->framebuffer = framebuffer;
entry->invalidHint = -1;
entry->status &= ~TextureCacheDX9::TexCacheEntry::STATUS_DEPALETTIZE;
fbTexInfo_[entry->addr] = fbInfo;
host->GPUNotifyTextureAttachment(entry->addr);
}
}
bool TextureCacheDX9::AttachFramebuffer(TexCacheEntry *entry, u32 address, VirtualFramebuffer *framebuffer, u32 texaddrOffset) {
static const u32 MAX_SUBAREA_Y_OFFSET_SAFE = 32;
AttachedFramebufferInfo fbInfo = {0};
const u64 mirrorMask = 0x00600000;
// Must be in VRAM so | 0x04000000 it is. Also, ignore memory mirrors.
const u32 addr = (address | 0x04000000) & 0x3FFFFFFF & ~mirrorMask;
const u32 texaddr = ((entry->addr + texaddrOffset) & ~mirrorMask);
const bool noOffset = texaddr == addr;
const bool exactMatch = noOffset && entry->format < 4;
const u32 h = 1 << ((entry->dim >> 8) & 0xf);
// 512 on a 272 framebuffer is sane, so let's be lenient.
const u32 minSubareaHeight = h / 4;
// If they match exactly, it's non-CLUT and from the top left.
if (exactMatch) {
// Apply to non-buffered and buffered mode only.
if (!(g_Config.iRenderingMode == FB_NON_BUFFERED_MODE || g_Config.iRenderingMode == FB_BUFFERED_MODE))
return false;
DEBUG_LOG(G3D, "Render to texture detected at %08x!", address);
if (framebuffer->fb_stride != entry->bufw) {
WARN_LOG_REPORT_ONCE(diffStrides1, G3D, "Render to texture with different strides %d != %d", entry->bufw, framebuffer->fb_stride);
}
if (entry->format != framebuffer->format) {
WARN_LOG_REPORT_ONCE(diffFormat1, G3D, "Render to texture with different formats %d != %d", entry->format, framebuffer->format);
// Let's avoid using it when we know the format is wrong. May be a video/etc. updating memory.
// However, some games use a different format to clear the buffer.
if (framebuffer->last_frame_attached + 1 < gpuStats.numFlips) {
DetachFramebuffer(entry, address, framebuffer);
}
} else {
AttachFramebufferValid(entry, framebuffer, fbInfo);
return true;
}
} else {
// Apply to buffered mode only.
if (!(g_Config.iRenderingMode == FB_BUFFERED_MODE))
return false;
const bool clutFormat =
(framebuffer->format == GE_FORMAT_8888 && entry->format == GE_TFMT_CLUT32) ||
(framebuffer->format != GE_FORMAT_8888 && entry->format == GE_TFMT_CLUT16);
const u32 bitOffset = (texaddr - addr) * 8;
const u32 pixelOffset = bitOffset / std::max(1U, (u32)textureBitsPerPixel[entry->format]);
fbInfo.yOffset = pixelOffset / entry->bufw;
fbInfo.xOffset = entry->bufw == 0 ? 0 : pixelOffset % entry->bufw;
if (framebuffer->fb_stride != entry->bufw) {
if (noOffset) {
WARN_LOG_REPORT_ONCE(diffStrides2, G3D, "Render to texture using CLUT with different strides %d != %d", entry->bufw, framebuffer->fb_stride);
} else {
// Assume any render-to-tex with different bufw + offset is a render from ram.
DetachFramebuffer(entry, address, framebuffer);
return false;
}
}
if (fbInfo.yOffset + minSubareaHeight >= framebuffer->height) {
// Can't be inside the framebuffer then, ram. Detach to be safe.
DetachFramebuffer(entry, address, framebuffer);
return false;
}
// Trying to play it safe. Below 0x04110000 is almost always framebuffers.
// TODO: Maybe we can reduce this check and find a better way above 0x04110000?
if (fbInfo.yOffset > MAX_SUBAREA_Y_OFFSET_SAFE && addr > 0x04110000) {
WARN_LOG_REPORT_ONCE(subareaIgnored, G3D, "Ignoring possible render to texture at %08x +%dx%d / %dx%d", address, fbInfo.xOffset, fbInfo.yOffset, framebuffer->width, framebuffer->height);
DetachFramebuffer(entry, address, framebuffer);
return false;
}
// Check for CLUT. The framebuffer is always RGB, but it can be interpreted as a CLUT texture.
// 3rd Birthday (and a bunch of other games) render to a 16 bit clut texture.
if (clutFormat) {
if (!noOffset) {
WARN_LOG_REPORT_ONCE(subareaClut, G3D, "Render to texture using CLUT with offset at %08x +%dx%d", address, fbInfo.xOffset, fbInfo.yOffset);
}
AttachFramebufferValid(entry, framebuffer, fbInfo);
entry->status |= TexCacheEntry::STATUS_DEPALETTIZE;
// We'll validate it compiles later.
return true;
} else if (entry->format == GE_TFMT_CLUT8 || entry->format == GE_TFMT_CLUT4) {
ERROR_LOG_REPORT_ONCE(fourEightBit, G3D, "4 and 8-bit CLUT format not supported for framebuffers");
}
// This is either normal or we failed to generate a shader to depalettize
if (framebuffer->format == entry->format || clutFormat) {
if (framebuffer->format != entry->format) {
WARN_LOG_REPORT_ONCE(diffFormat2, G3D, "Render to texture with different formats %d != %d at %08x", entry->format, framebuffer->format, address);
AttachFramebufferValid(entry, framebuffer, fbInfo);
return true;
} else {
WARN_LOG_REPORT_ONCE(subarea, G3D, "Render to area containing texture at %08x +%dx%d", address, fbInfo.xOffset, fbInfo.yOffset);
// If "AttachFramebufferValid" , God of War Ghost of Sparta/Chains of Olympus will be missing special effect.
AttachFramebufferInvalid(entry, framebuffer, fbInfo);
return true;
}
} else {
WARN_LOG_REPORT_ONCE(diffFormat2, G3D, "Render to texture with incompatible formats %d != %d at %08x", entry->format, framebuffer->format, address);
}
}
return false;
}
inline void TextureCacheDX9::DetachFramebuffer(TexCacheEntry *entry, u32 address, VirtualFramebuffer *framebuffer) {
if (entry->framebuffer == framebuffer) {
cacheSizeEstimate_ += EstimateTexMemoryUsage(entry);
entry->framebuffer = 0;
host->GPUNotifyTextureAttachment(entry->addr);
}
}
void TextureCacheDX9::NotifyFramebuffer(u32 address, VirtualFramebuffer *framebuffer, FramebufferNotification msg) {
// Must be in VRAM so | 0x04000000 it is. Also, ignore memory mirrors.
// These checks are mainly to reduce scanning all textures.
const u32 addr = (address | 0x04000000) & 0x3F9FFFFF;
const u32 bpp = framebuffer->format == GE_FORMAT_8888 ? 4 : 2;
const u64 cacheKey = (u64)addr << 32;
// If it has a clut, those are the low 32 bits, so it'll be inside this range.
// Also, if it's a subsample of the buffer, it'll also be within the FBO.
const u64 cacheKeyEnd = cacheKey + ((u64)(framebuffer->fb_stride * framebuffer->height * bpp) << 32);
// The first mirror starts at 0x04200000 and there are 3. We search all for framebuffers.
const u64 mirrorCacheKey = (u64)0x04200000 << 32;
const u64 mirrorCacheKeyEnd = (u64)0x04800000 << 32;
switch (msg) {
case NOTIFY_FB_CREATED:
case NOTIFY_FB_UPDATED:
// Ensure it's in the framebuffer cache.
if (std::find(fbCache_.begin(), fbCache_.end(), framebuffer) == fbCache_.end()) {
fbCache_.push_back(framebuffer);
}
for (auto it = cache.lower_bound(cacheKey), end = cache.upper_bound(cacheKeyEnd); it != end; ++it) {
AttachFramebuffer(&it->second, addr, framebuffer);
}
// Let's assume anything in mirrors is fair game to check.
for (auto it = cache.lower_bound(mirrorCacheKey), end = cache.upper_bound(mirrorCacheKeyEnd); it != end; ++it) {
AttachFramebuffer(&it->second, addr, framebuffer);
}
break;
case NOTIFY_FB_DESTROYED:
fbCache_.erase(std::remove(fbCache_.begin(), fbCache_.end(), framebuffer), fbCache_.end());
for (auto it = cache.lower_bound(cacheKey), end = cache.upper_bound(cacheKeyEnd); it != end; ++it) {
DetachFramebuffer(&it->second, addr, framebuffer);
}
for (auto it = cache.lower_bound(mirrorCacheKey), end = cache.upper_bound(mirrorCacheKeyEnd); it != end; ++it) {
DetachFramebuffer(&it->second, addr, framebuffer);
}
break;
}
}
void *TextureCacheDX9::UnswizzleFromMem(const u8 *texptr, u32 bufw, u32 bytesPerPixel, u32 level) {
const u32 rowWidth = (bytesPerPixel > 0) ? (bufw * bytesPerPixel) : (bufw / 2);
const u32 pitch = rowWidth / 4;
const int bxc = rowWidth / 16;
int byc = (gstate.getTextureHeight(level) + 7) / 8;
if (byc == 0)
byc = 1;
u32 ydest = 0;
if (rowWidth >= 16) {
u32 *ydestp = tmpTexBuf32.data();
// The most common one, so it gets an optimized implementation.
DoUnswizzleTex16(texptr, ydestp, bxc, byc, pitch, rowWidth);
} else if (rowWidth == 8) {
const u32 *src = (const u32 *) texptr;
for (int by = 0; by < byc; by++) {
for (int n = 0; n < 8; n++, ydest += 2) {
tmpTexBuf32[ydest + 0] = *src++;
tmpTexBuf32[ydest + 1] = *src++;
src += 2; // skip two u32
}
}
} else if (rowWidth == 4) {
const u32 *src = (const u32 *) texptr;
for (int by = 0; by < byc; by++) {
for (int n = 0; n < 8; n++, ydest++) {
tmpTexBuf32[ydest] = *src++;
src += 3;
}
}
} else if (rowWidth == 2) {
const u16 *src = (const u16 *) texptr;
for (int by = 0; by < byc; by++) {
for (int n = 0; n < 4; n++, ydest++) {
u16 n1 = src[0];
u16 n2 = src[8];
tmpTexBuf32[ydest] = (u32)n1 | ((u32)n2 << 16);
src += 16;
}
}
} else if (rowWidth == 1) {
const u8 *src = (const u8 *) texptr;
for (int by = 0; by < byc; by++) {
for (int n = 0; n < 2; n++, ydest++) {
u8 n1 = src[ 0];
u8 n2 = src[16];
u8 n3 = src[32];
u8 n4 = src[48];
tmpTexBuf32[ydest] = (u32)n1 | ((u32)n2 << 8) | ((u32)n3 << 16) | ((u32)n4 << 24);
src += 64;
}
}
}
return tmpTexBuf32.data();
}
void *TextureCacheDX9::ReadIndexedTex(int level, const u8 *texptr, int bytesPerIndex, u32 dstFmt, int bufw) {
int w = gstate.getTextureWidth(level);
int h = gstate.getTextureHeight(level);
int length = bufw * h;
void *buf = NULL;
switch (gstate.getClutPaletteFormat()) {
case GE_CMODE_16BIT_BGR5650:
case GE_CMODE_16BIT_ABGR5551:
case GE_CMODE_16BIT_ABGR4444:
{
tmpTexBuf16.resize(std::max(bufw, w) * h);
tmpTexBufRearrange.resize(std::max(bufw, w) * h);
const u16 *clut = GetCurrentClut<u16>();
if (!gstate.isTextureSwizzled()) {
switch (bytesPerIndex) {
case 1:
DeIndexTexture(tmpTexBuf16.data(), (const u8 *)texptr, length, clut);
break;
case 2:
DeIndexTexture(tmpTexBuf16.data(), (const u16_le *)texptr, length, clut);
break;
case 4:
DeIndexTexture(tmpTexBuf16.data(), (const u32_le *)texptr, length, clut);
break;
}
} else {
tmpTexBuf32.resize(std::max(bufw, w) * h);
UnswizzleFromMem(texptr, bufw, bytesPerIndex, level);
switch (bytesPerIndex) {
case 1:
DeIndexTexture(tmpTexBuf16.data(), (u8 *) tmpTexBuf32.data(), length, clut);
break;
case 2:
DeIndexTexture(tmpTexBuf16.data(), (u16 *) tmpTexBuf32.data(), length, clut);
break;
case 4:
DeIndexTexture(tmpTexBuf16.data(), (u32 *) tmpTexBuf32.data(), length, clut);
break;
}
}
buf = tmpTexBuf16.data();
}
break;
case GE_CMODE_32BIT_ABGR8888:
{
tmpTexBuf32.resize(std::max(bufw, w) * h);
tmpTexBufRearrange.resize(std::max(bufw, w) * h);
const u32 *clut = GetCurrentClut<u32>();
if (!gstate.isTextureSwizzled()) {
switch (bytesPerIndex) {
case 1:
DeIndexTexture(tmpTexBuf32.data(), (const u8 *)texptr, length, clut);
break;
case 2:
DeIndexTexture(tmpTexBuf32.data(), (const u16_le *)texptr, length, clut);
break;
case 4:
DeIndexTexture(tmpTexBuf32.data(), (const u32_le *)texptr, length, clut);
break;
}
buf = tmpTexBuf32.data();
} else {
UnswizzleFromMem(texptr, bufw, bytesPerIndex, level);
// Since we had to unswizzle to tmpTexBuf32, let's output to tmpTexBuf16.
tmpTexBuf16.resize(std::max(bufw, w) * h * 2);
u32 *dest32 = (u32 *) tmpTexBuf16.data();
switch (bytesPerIndex) {
case 1:
DeIndexTexture(dest32, (u8 *) tmpTexBuf32.data(), length, clut);
buf = dest32;
break;
case 2:
DeIndexTexture(dest32, (u16 *) tmpTexBuf32.data(), length, clut);
buf = dest32;
break;
case 4:
// TODO: If a game actually uses this mode, check if using dest32 or tmpTexBuf32 is faster.
DeIndexTexture(tmpTexBuf32.data(), tmpTexBuf32.data(), length, clut);
buf = tmpTexBuf32.data();
break;
}
}
}
break;
default:
ERROR_LOG_REPORT(G3D, "Unhandled clut texture mode %d!!!", (gstate.clutformat & 3));
break;
}
return buf;
}
D3DFORMAT getClutDestFormat(GEPaletteFormat format) {
switch (format) {
case GE_CMODE_16BIT_ABGR4444:
return D3DFMT_A4R4G4B4;
case GE_CMODE_16BIT_ABGR5551:
return D3DFMT_A1R5G5B5;
case GE_CMODE_16BIT_BGR5650:
return D3DFMT_R5G6B5;
case GE_CMODE_32BIT_ABGR8888:
return D3DFMT_A8R8G8B8;
}
// Should never be here !
return D3DFMT_A8R8G8B8;
}
static const u8 texByteAlignMap[] = {2, 2, 2, 4};
static const u8 MinFilt[8] = {
D3DTEXF_POINT,
D3DTEXF_LINEAR,
D3DTEXF_POINT,
D3DTEXF_LINEAR,
D3DTEXF_POINT, // GL_NEAREST_MIPMAP_NEAREST,
D3DTEXF_LINEAR, // GL_LINEAR_MIPMAP_NEAREST,
D3DTEXF_POINT, // GL_NEAREST_MIPMAP_LINEAR,
D3DTEXF_LINEAR, // GL_LINEAR_MIPMAP_LINEAR,
};
static const u8 MipFilt[8] = {
D3DTEXF_POINT,
D3DTEXF_LINEAR,
D3DTEXF_POINT,
D3DTEXF_LINEAR,
D3DTEXF_POINT, // GL_NEAREST_MIPMAP_NEAREST,
D3DTEXF_POINT, // GL_LINEAR_MIPMAP_NEAREST,
D3DTEXF_LINEAR, // GL_NEAREST_MIPMAP_LINEAR,
D3DTEXF_LINEAR, // GL_LINEAR_MIPMAP_LINEAR,
};
static const u8 MagFilt[2] = {
D3DTEXF_POINT,
D3DTEXF_LINEAR
};
void TextureCacheDX9::GetSamplingParams(int &minFilt, int &magFilt, bool &sClamp, bool &tClamp, float &lodBias, int maxLevel) {
minFilt = gstate.texfilter & 0x7;
magFilt = (gstate.texfilter>>8) & 1;
sClamp = gstate.isTexCoordClampedS();
tClamp = gstate.isTexCoordClampedT();
bool noMip = (gstate.texlevel & 0xFFFFFF) == 0x000001 || (gstate.texlevel & 0xFFFFFF) == 0x100001 ; // Fix texlevel at 0
if (maxLevel == 0) {
// Enforce no mip filtering, for safety.
minFilt &= 1; // no mipmaps yet
lodBias = 0.0f;
} else {
// Texture lod bias should be signed.
lodBias = (float)(int)(s8)((gstate.texlevel >> 16) & 0xFF) / 16.0f;
}
if (g_Config.iTexFiltering == LINEARFMV && g_iNumVideos > 0 && (gstate.getTextureDimension(0) & 0xF) >= 9) {
magFilt |= 1;
minFilt |= 1;
}
if (g_Config.iTexFiltering == LINEAR && (!gstate.isColorTestEnabled() || IsColorTestTriviallyTrue())) {
// TODO: IsAlphaTestTriviallyTrue() is unsafe here. vertexFullAlpha is not calculated yet.
if (!gstate.isAlphaTestEnabled() || IsAlphaTestTriviallyTrue()) {
magFilt |= 1;
minFilt |= 1;
}
}
bool forceNearest = g_Config.iTexFiltering == NEAREST;
// Force Nearest when color test enabled and rendering resolution greater than 480x272
if ((gstate.isColorTestEnabled() && !IsColorTestTriviallyTrue()) && g_Config.iInternalResolution != 1 && gstate.isModeThrough()) {
// Some games use 0 as the color test color, which won't be too bad if it bleeds.
// Fuchsia and green, etc. are the problem colors.
if (gstate.getColorTestRef() != 0) {
forceNearest = true;
}
}
if (forceNearest) {
magFilt &= ~1;
minFilt &= ~1;
}
if (!g_Config.bMipMap || noMip) {
minFilt &= 1;
}
}
void TextureCacheDX9::UpdateSamplingParams(TexCacheEntry &entry, bool force) {
int minFilt;
int magFilt;
bool sClamp;
bool tClamp;
float lodBias;
GetSamplingParams(minFilt, magFilt, sClamp, tClamp, lodBias, entry.maxLevel);
if (entry.maxLevel != 0) {
GETexLevelMode mode = gstate.getTexLevelMode();
switch (mode) {
case GE_TEXLEVEL_MODE_AUTO:
// TODO
break;
case GE_TEXLEVEL_MODE_CONST:
dxstate.texMipLodBias.set(lodBias);
// TODO
break;
case GE_TEXLEVEL_MODE_SLOPE:
// TODO
break;
}
entry.lodBias = lodBias;
}
D3DTEXTUREFILTERTYPE minf = (D3DTEXTUREFILTERTYPE)MinFilt[minFilt];
D3DTEXTUREFILTERTYPE mipf = (D3DTEXTUREFILTERTYPE)MipFilt[minFilt];
D3DTEXTUREFILTERTYPE magf = (D3DTEXTUREFILTERTYPE)MagFilt[magFilt];
if (g_Config.iAnisotropyLevel > 0 && minf == D3DTEXF_LINEAR) {
minf = D3DTEXF_ANISOTROPIC;
}
dxstate.texMinFilter.set(minf);
dxstate.texMipFilter.set(mipf);
dxstate.texMagFilter.set(magf);
dxstate.texAddressU.set(sClamp ? D3DTADDRESS_CLAMP : D3DTADDRESS_WRAP);
dxstate.texAddressV.set(tClamp ? D3DTADDRESS_CLAMP : D3DTADDRESS_WRAP);
}
void TextureCacheDX9::SetFramebufferSamplingParams(u16 bufferWidth, u16 bufferHeight) {
int minFilt;
int magFilt;
bool sClamp;
bool tClamp;
float lodBias;
GetSamplingParams(minFilt, magFilt, sClamp, tClamp, lodBias, 0);
dxstate.texMinFilter.set(MinFilt[minFilt]);
dxstate.texMipFilter.set(MipFilt[minFilt]);
dxstate.texMagFilter.set(MagFilt[magFilt]);
// Often the framebuffer will not match the texture size. We'll wrap/clamp in the shader in that case.
// This happens whether we have OES_texture_npot or not.
int w = gstate.getTextureWidth(0);
int h = gstate.getTextureHeight(0);
if (w != bufferWidth || h != bufferHeight) {
return;
}
dxstate.texAddressU.set(sClamp ? D3DTADDRESS_CLAMP : D3DTADDRESS_WRAP);
dxstate.texAddressV.set(tClamp ? D3DTADDRESS_CLAMP : D3DTADDRESS_WRAP);
}
void TextureCacheDX9::StartFrame() {
lastBoundTexture = INVALID_TEX;
timesInvalidatedAllThisFrame_ = 0;
if (texelsScaledThisFrame_) {
// INFO_LOG(G3D, "Scaled %i texels", texelsScaledThisFrame_);
}
texelsScaledThisFrame_ = 0;
if (clearCacheNextFrame_) {
Clear(true);
clearCacheNextFrame_ = false;
} else {
Decimate();
}
DWORD anisotropyLevel = (DWORD)g_Config.iAnisotropyLevel > maxAnisotropyLevel ? maxAnisotropyLevel : g_Config.iAnisotropyLevel;
pD3Ddevice->SetSamplerState(0, D3DSAMP_MAXANISOTROPY, anisotropyLevel);
}
static inline u32 MiniHash(const u32 *ptr) {
return ptr[0];
}
static inline u32 QuickTexHash(u32 addr, int bufw, int w, int h, GETextureFormat format) {
const u32 sizeInRAM = (textureBitsPerPixel[format] * bufw * h) / 8;
const u32 *checkp = (const u32 *) Memory::GetPointer(addr);
return DoQuickTexHash(checkp, sizeInRAM);
}
inline bool TextureCacheDX9::TexCacheEntry::Matches(u16 dim2, u8 format2, int maxLevel2) {
return dim == dim2 && format == format2 && maxLevel == maxLevel2;
}
void TextureCacheDX9::LoadClut() {
u32 clutAddr = gstate.getClutAddress();
if (Memory::IsValidAddress(clutAddr)) {
#ifdef _M_SSE
int numBlocks = gstate.getClutLoadBlocks();
clutTotalBytes_ = numBlocks * 32;
const __m128i *source = (const __m128i *)Memory::GetPointerUnchecked(clutAddr);
__m128i *dest = (__m128i *)clutBufRaw_;
for (int i = 0; i < numBlocks; i++, source += 2, dest += 2) {
__m128i data1 = _mm_loadu_si128(source);
__m128i data2 = _mm_loadu_si128(source + 1);
_mm_store_si128(dest, data1);
_mm_store_si128(dest + 1, data2);
}
#else
clutTotalBytes_ = gstate.getClutLoadBytes();
Memory::MemcpyUnchecked(clutBufRaw_, clutAddr, clutTotalBytes_);
#endif
} else {
clutTotalBytes_ = gstate.getClutLoadBytes();
memset(clutBufRaw_, 0xFF, clutTotalBytes_);
}
// Reload the clut next time.
clutLastFormat_ = 0xFFFFFFFF;
clutMaxBytes_ = std::max(clutMaxBytes_, clutTotalBytes_);
}
void TextureCacheDX9::UpdateCurrentClut() {
const GEPaletteFormat clutFormat = gstate.getClutPaletteFormat();
const u32 clutBase = gstate.getClutIndexStartPos();
const u32 clutBaseBytes = clutBase * (clutFormat == GE_CMODE_32BIT_ABGR8888 ? sizeof(u32) : sizeof(u16));
// Technically, these extra bytes weren't loaded, but hopefully it was loaded earlier.
// If not, we're going to hash random data, which hopefully doesn't cause a performance issue.
const u32 clutExtendedBytes = clutTotalBytes_ + clutBaseBytes;
clutHash_ = DoReliableHash32((const char *)clutBufRaw_, clutExtendedBytes, 0xC0108888);
clutBuf_ = clutBufRaw_;
// Special optimization: fonts typically draw clut4 with just alpha values in a single color.
clutAlphaLinear_ = false;
clutAlphaLinearColor_ = 0;
if (gstate.getClutPaletteFormat() == GE_CMODE_16BIT_ABGR4444 && gstate.isClutIndexSimple()) {
const u16_le *clut = GetCurrentClut<u16_le>();
clutAlphaLinear_ = true;
clutAlphaLinearColor_ = clut[15] & 0x0FFF;
for (int i = 0; i < 16; ++i) {
if ((clut[i] >> 12) != i) {
clutAlphaLinear_ = false;
break;
}
// Alpha 0 doesn't matter.
// TODO: Well, depending on blend mode etc, it can actually matter, although unlikely.
if (i != 0 && (clut[i] >> 12) != clutAlphaLinearColor_) {
clutAlphaLinear_ = false;
break;
}
}
}
clutLastFormat_ = gstate.clutformat;
}
template <typename T>
inline const T *TextureCacheDX9::GetCurrentClut() {
return (const T *)clutBuf_;
}
inline u32 TextureCacheDX9::GetCurrentClutHash() {
return clutHash_;
}
void TextureCacheDX9::SetTextureFramebuffer(TexCacheEntry *entry, VirtualFramebuffer *framebuffer) {
_dbg_assert_msg_(G3D, framebuffer != nullptr, "Framebuffer must not be null.");
framebuffer->usageFlags |= FB_USAGE_TEXTURE;
bool useBufferedRendering = g_Config.iRenderingMode != FB_NON_BUFFERED_MODE;
if (useBufferedRendering) {
LPDIRECT3DPIXELSHADER9 pshader = nullptr;
if ((entry->status & TexCacheEntry::STATUS_DEPALETTIZE) && !g_Config.bDisableSlowFramebufEffects) {
pshader = depalShaderCache_->GetDepalettizePixelShader(framebuffer->drawnFormat);
}
if (pshader) {
LPDIRECT3DTEXTURE9 clutTexture = depalShaderCache_->GetClutTexture(clutHash_, clutBuf_);
FBO *depalFBO = framebufferManager_->GetTempFBO(framebuffer->renderWidth, framebuffer->renderHeight, FBO_8888);
fbo_bind_as_render_target(depalFBO);
static const float pos[12 + 8] = {
-1, -1, -1, 0, 0,
1, -1, -1, 1, 0,
1, 1, -1, 1, 1,
-1, 1, -1, 0, 1,
};
static const u16 indices[4] = { 0, 1, 3, 2 };
shaderManager_->DirtyLastShader();
pD3Ddevice->SetPixelShader(pshader);
pD3Ddevice->SetVertexShader(depalShaderCache_->GetDepalettizeVertexShader());
pD3Ddevice->SetTexture(1, clutTexture);
framebufferManager_->BindFramebufferColor(0, framebuffer, true);
pD3Ddevice->SetRenderState(D3DRS_ZENABLE, FALSE);
pD3Ddevice->SetRenderState(D3DRS_STENCILENABLE, FALSE);
pD3Ddevice->SetRenderState(D3DRS_SCISSORTESTENABLE, FALSE);
pD3Ddevice->SetRenderState(D3DRS_CULLMODE, D3DCULL_NONE);
pD3Ddevice->SetFVF(D3DFVF_XYZ | D3DFVF_TEX0);
D3DVIEWPORT9 vp;
vp.MinZ = 0;
vp.MaxZ = 1;
vp.X = 0;
vp.Y = 0;
vp.Width = framebuffer->renderWidth;
vp.Height = framebuffer->renderHeight;
pD3Ddevice->SetViewport(&vp);
pD3Ddevice->DrawIndexedPrimitiveUP(D3DPT_TRIANGLEFAN, 0, 4, 2, indices, D3DFMT_INDEX16, pos, (3 + 2) * sizeof(float));
fbo_bind_color_as_texture(depalFBO, 0);
dxstate.Restore();
framebufferManager_->RebindFramebuffer();
const GEPaletteFormat clutFormat = gstate.getClutPaletteFormat();
const u32 clutBase = gstate.getClutIndexStartPos();
const u32 bytesPerColor = clutFormat == GE_CMODE_32BIT_ABGR8888 ? sizeof(u32) : sizeof(u16);
const u32 clutExtendedColors = (clutTotalBytes_ / bytesPerColor) + clutBase;
TexCacheEntry::Status alphaStatus = CheckAlpha(clutBuf_, getClutDestFormat(gstate.getClutPaletteFormat()), clutExtendedColors, clutExtendedColors, 1);
gstate_c.textureFullAlpha = alphaStatus == TexCacheEntry::STATUS_ALPHA_FULL;
gstate_c.textureSimpleAlpha = alphaStatus == TexCacheEntry::STATUS_ALPHA_SIMPLE;
} else {
entry->status &= ~TexCacheEntry::STATUS_DEPALETTIZE;
framebufferManager_->BindFramebufferColor(0, framebuffer);
gstate_c.textureFullAlpha = gstate.getTextureFormat() == GE_TFMT_5650;
gstate_c.textureSimpleAlpha = gstate_c.textureFullAlpha;
}
// Keep the framebuffer alive.
framebuffer->last_frame_used = gpuStats.numFlips;
// We need to force it, since we may have set it on a texture before attaching.
gstate_c.curTextureWidth = framebuffer->bufferWidth;
gstate_c.curTextureHeight = framebuffer->bufferHeight;
gstate_c.flipTexture = false;
gstate_c.bgraTexture = false;
gstate_c.curTextureXOffset = fbTexInfo_[entry->addr].xOffset;
gstate_c.curTextureYOffset = fbTexInfo_[entry->addr].yOffset;
gstate_c.needShaderTexClamp = gstate_c.curTextureWidth != (u32)gstate.getTextureWidth(0) || gstate_c.curTextureHeight != (u32)gstate.getTextureHeight(0);
if (gstate_c.curTextureXOffset != 0 || gstate_c.curTextureYOffset != 0) {
gstate_c.needShaderTexClamp = true;
}
SetFramebufferSamplingParams(framebuffer->bufferWidth, framebuffer->bufferHeight);
} else {
if (framebuffer->fbo) {
fbo_destroy(framebuffer->fbo);
framebuffer->fbo = 0;
}
pD3Ddevice->SetTexture(0, NULL);
gstate_c.needShaderTexClamp = false;
}
}
bool TextureCacheDX9::SetOffsetTexture(u32 offset) {
if (g_Config.iRenderingMode != FB_BUFFERED_MODE) {
return false;
}
u32 texaddr = gstate.getTextureAddress(0);
if (!Memory::IsValidAddress(texaddr) || !Memory::IsValidAddress(texaddr + offset)) {
return false;
}
u64 cachekey = (u64)(texaddr & 0x3FFFFFFF) << 32;
TexCache::iterator iter = cache.find(cachekey);
if (iter == cache.end()) {
return false;
}
TexCacheEntry *entry = &iter->second;
bool success = false;
for (size_t i = 0, n = fbCache_.size(); i < n; ++i) {
auto framebuffer = fbCache_[i];
if (AttachFramebuffer(entry, framebuffer->fb_address, framebuffer, offset)) {
success = true;
}
}
if (success && entry->framebuffer) {
SetTextureFramebuffer(entry, entry->framebuffer);
lastBoundTexture = INVALID_TEX;
entry->lastFrame = gpuStats.numFlips;
return true;
}
return false;
}
void TextureCacheDX9::SetTexture(bool force) {
#ifdef DEBUG_TEXTURES
if (SetDebugTexture()) {
// A different texture was bound, let's rebind next time.
lastBoundTexture = -1;
return;
}
#endif
if (force) {
lastBoundTexture = INVALID_TEX;
}
u32 texaddr = gstate.getTextureAddress(0);
if (!Memory::IsValidAddress(texaddr)) {
// Bind a null texture and return.
pD3Ddevice->SetTexture(0, NULL);
lastBoundTexture = INVALID_TEX;
return;
}
int w = gstate.getTextureWidth(0);
int h = gstate.getTextureHeight(0);
GETextureFormat format = gstate.getTextureFormat();
if (format >= 11) {
ERROR_LOG_REPORT(G3D, "Unknown texture format %i", format);
// TODO: Better assumption?
format = GE_TFMT_5650;
}
bool hasClut = gstate.isTextureFormatIndexed();
// Ignore uncached/kernel when caching.
u64 cachekey = (u64)(texaddr & 0x3FFFFFFF) << 32;
u32 cluthash;
if (hasClut) {
if (clutLastFormat_ != gstate.clutformat) {
// We update here because the clut format can be specified after the load.
UpdateCurrentClut();
}
cluthash = GetCurrentClutHash() ^ gstate.clutformat;
cachekey |= cluthash;
} else {
cluthash = 0;
}
int bufw = GetTextureBufw(0, texaddr, format);
int maxLevel = gstate.getTextureMaxLevel();
u32 texhash = MiniHash((const u32 *)Memory::GetPointer(texaddr));
u32 fullhash = 0;
TexCache::iterator iter = cache.find(cachekey);
TexCacheEntry *entry = NULL;
gstate_c.flipTexture = false;
gstate_c.needShaderTexClamp = false;
gstate_c.bgraTexture = true;
gstate_c.skipDrawReason &= ~SKIPDRAW_BAD_FB_TEXTURE;
bool useBufferedRendering = g_Config.iRenderingMode != FB_NON_BUFFERED_MODE;
bool replaceImages = false;
if (iter != cache.end()) {
entry = &iter->second;
// Validate the texture still matches the cache entry.
u16 dim = gstate.getTextureDimension(0);
bool match = entry->Matches(dim, format, maxLevel);
// Check for FBO - slow!
if (entry->framebuffer) {
if (match) {
SetTextureFramebuffer(entry, entry->framebuffer);
lastBoundTexture = INVALID_TEX;
entry->lastFrame = gpuStats.numFlips;
return;
} else {
// Make sure we re-evaluate framebuffers.
DetachFramebuffer(entry, texaddr, entry->framebuffer);
match = false;
}
}
bool rehash = entry->GetHashStatus() == TexCacheEntry::STATUS_UNRELIABLE;
bool doDelete = true;
// First let's see if another texture with the same address had a hashfail.
if (entry->status & TexCacheEntry::STATUS_CLUT_RECHECK) {
// Always rehash in this case, if one changed the rest all probably did.
rehash = true;
entry->status &= ~TexCacheEntry::STATUS_CLUT_RECHECK;
} else if ((gstate_c.textureChanged & TEXCHANGE_UPDATED) == 0) {
// Okay, just some parameter change - the data didn't change, no need to rehash.
rehash = false;
}
if (match) {
if (entry->lastFrame != gpuStats.numFlips) {
u32 diff = gpuStats.numFlips - entry->lastFrame;
entry->numFrames++;
if (entry->framesUntilNextFullHash < diff) {
// Exponential backoff up to 512 frames. Textures are often reused.
if (entry->numFrames > 32) {
// Also, try to add some "randomness" to avoid rehashing several textures the same frame.
entry->framesUntilNextFullHash = std::min(512, entry->numFrames) + ((intptr_t)entry->texture & 15);
} else {
entry->framesUntilNextFullHash = entry->numFrames;
}
rehash = true;
} else {
entry->framesUntilNextFullHash -= diff;
}
}
// If it's not huge or has been invalidated many times, recheck the whole texture.
if (entry->invalidHint > 180 || (entry->invalidHint > 15 && (dim >> 8) < 9 && (dim & 0xF) < 9)) {
entry->invalidHint = 0;
rehash = true;
}
bool hashFail = false;
if (texhash != entry->hash) {
fullhash = QuickTexHash(texaddr, bufw, w, h, format);
hashFail = true;
rehash = false;
}
if (rehash && entry->GetHashStatus() != TexCacheEntry::STATUS_RELIABLE) {
fullhash = QuickTexHash(texaddr, bufw, w, h, format);
if (fullhash != entry->fullhash) {
hashFail = true;
} else if (entry->GetHashStatus() != TexCacheEntry::STATUS_HASHING && entry->numFrames > TexCacheEntry::FRAMES_REGAIN_TRUST) {
// Reset to STATUS_HASHING.
if (g_Config.bTextureBackoffCache) {
entry->SetHashStatus(TexCacheEntry::STATUS_HASHING);
}
entry->status &= ~TexCacheEntry::STATUS_CHANGE_FREQUENT;
}
}
if (hashFail) {
match = false;
entry->status |= TexCacheEntry::STATUS_UNRELIABLE;
if (entry->numFrames < TEXCACHE_FRAME_CHANGE_FREQUENT) {
entry->status |= TexCacheEntry::STATUS_CHANGE_FREQUENT;
}
entry->numFrames = 0;
// Don't give up just yet. Let's try the secondary cache if it's been invalidated before.
// If it's failed a bunch of times, then the second cache is just wasting time and VRAM.
if (g_Config.bTextureSecondaryCache) {
if (entry->numInvalidated > 2 && entry->numInvalidated < 128 && !lowMemoryMode_) {
u64 secondKey = fullhash | (u64)cluthash << 32;
TexCache::iterator secondIter = secondCache.find(secondKey);
if (secondIter != secondCache.end()) {
TexCacheEntry *secondEntry = &secondIter->second;
if (secondEntry->Matches(dim, format, maxLevel)) {
// Reset the numInvalidated value lower, we got a match.
if (entry->numInvalidated > 8) {
--entry->numInvalidated;
}
entry = secondEntry;
match = true;
}
} else {
secondKey = entry->fullhash | ((u64)entry->cluthash << 32);
secondCacheSizeEstimate_ += EstimateTexMemoryUsage(entry);
secondCache[secondKey] = *entry;
doDelete = false;
}
}
}
}
}
if (match && (entry->status & TexCacheEntry::STATUS_TO_SCALE) && g_Config.iTexScalingLevel != 1 && texelsScaledThisFrame_ < TEXCACHE_MAX_TEXELS_SCALED) {
// INFO_LOG(G3D, "Reloading texture to do the scaling we skipped..");
match = false;
}
if (match) {
// TODO: Mark the entry reliable if it's been safe for long enough?
//got one!
entry->lastFrame = gpuStats.numFlips;
if (entry->texture != lastBoundTexture) {
pD3Ddevice->SetTexture(0, entry->texture);
lastBoundTexture = entry->texture;
gstate_c.textureFullAlpha = entry->GetAlphaStatus() == TexCacheEntry::STATUS_ALPHA_FULL;
gstate_c.textureSimpleAlpha = entry->GetAlphaStatus() != TexCacheEntry::STATUS_ALPHA_UNKNOWN;
}
UpdateSamplingParams(*entry, false);
VERBOSE_LOG(G3D, "Texture at %08x Found in Cache, applying", texaddr);
return; //Done!
} else {
cacheSizeEstimate_ -= EstimateTexMemoryUsage(entry);
entry->numInvalidated++;
gpuStats.numTextureInvalidations++;
DEBUG_LOG(G3D, "Texture different or overwritten, reloading at %08x", texaddr);
if (doDelete) {
if (entry->maxLevel == maxLevel && entry->dim == gstate.getTextureDimension(0) && entry->format == format && g_Config.iTexScalingLevel == 1) {
// Actually, if size and number of levels match, let's try to avoid deleting and recreating.
// Instead, let's use glTexSubImage to replace the images.
replaceImages = true;
} else {
if (entry->texture == lastBoundTexture) {
lastBoundTexture = INVALID_TEX;
}
entry->ReleaseTexture();
}
}
// Clear the reliable bit if set.
if (entry->GetHashStatus() == TexCacheEntry::STATUS_RELIABLE) {
entry->SetHashStatus(TexCacheEntry::STATUS_HASHING);
}
// Also, mark any textures with the same address but different clut. They need rechecking.
if (cluthash != 0) {
const u64 cachekeyMin = (u64)(texaddr & 0x3FFFFFFF) << 32;
const u64 cachekeyMax = cachekeyMin + (1ULL << 32);
for (auto it = cache.lower_bound(cachekeyMin), end = cache.upper_bound(cachekeyMax); it != end; ++it) {
if (it->second.cluthash != cluthash) {
it->second.status |= TexCacheEntry::STATUS_CLUT_RECHECK;
}
}
}
}
} else {
VERBOSE_LOG(G3D, "No texture in cache, decoding...");
TexCacheEntry entryNew = {0};
cache[cachekey] = entryNew;
entry = &cache[cachekey];
if (g_Config.bTextureBackoffCache) {
entry->status = TexCacheEntry::STATUS_HASHING;
} else {
entry->status = TexCacheEntry::STATUS_UNRELIABLE;
}
}
if ((bufw == 0 || (gstate.texbufwidth[0] & 0xf800) != 0) && texaddr >= PSP_GetKernelMemoryEnd()) {
ERROR_LOG_REPORT(G3D, "Texture with unexpected bufw (full=%d)", gstate.texbufwidth[0] & 0xffff);
}
// We have to decode it, let's setup the cache entry first.
entry->addr = texaddr;
entry->hash = texhash;
entry->format = format;
entry->lastFrame = gpuStats.numFlips;
entry->framebuffer = 0;
entry->maxLevel = maxLevel;
entry->lodBias = 0.0f;
entry->dim = gstate.getTextureDimension(0);
entry->bufw = bufw;
// This would overestimate the size in many case so we underestimate instead
// to avoid excessive clearing caused by cache invalidations.
entry->sizeInRAM = (textureBitsPerPixel[format] * bufw * h / 2) / 8;
entry->fullhash = fullhash == 0 ? QuickTexHash(texaddr, bufw, w, h, format) : fullhash;
entry->cluthash = cluthash;
entry->status &= ~TexCacheEntry::STATUS_ALPHA_MASK;
gstate_c.curTextureWidth = w;
gstate_c.curTextureHeight = h;
// For the estimate, we assume cluts always point to 8888 for simplicity.
cacheSizeEstimate_ += EstimateTexMemoryUsage(entry);
// TODO: If a framebuffer is attached here, might end up with a bad entry.texture.
// Should just always create one here or something (like GLES.)
// Before we go reading the texture from memory, let's check for render-to-texture.
for (size_t i = 0, n = fbCache_.size(); i < n; ++i) {
auto framebuffer = fbCache_[i];
AttachFramebuffer(entry, framebuffer->fb_address, framebuffer);
}
// If we ended up with a framebuffer, attach it - no texture decoding needed.
if (entry->framebuffer) {
SetTextureFramebuffer(entry, entry->framebuffer);
lastBoundTexture = INVALID_TEX;
entry->lastFrame = gpuStats.numFlips;
return;
}
// Adjust maxLevel to actually present levels..
bool badMipSizes = false;
for (int i = 0; i <= maxLevel; i++) {
// If encountering levels pointing to nothing, adjust max level.
u32 levelTexaddr = gstate.getTextureAddress(i);
if (!Memory::IsValidAddress(levelTexaddr)) {
maxLevel = i - 1;
break;
}
if (i > 0) {
int tw = gstate.getTextureWidth(i);
int th = gstate.getTextureHeight(i);
if (tw != 1 && tw != (gstate.getTextureWidth(i - 1) >> 1))
badMipSizes = true;
else if (th != 1 && th != (gstate.getTextureHeight(i - 1) >> 1))
badMipSizes = true;
}
}
// In addition, simply don't load more than level 0 if g_Config.bMipMap is false.
if (!g_Config.bMipMap) {
maxLevel = 0;
}
// If GLES3 is available, we can preallocate the storage, which makes texture loading more efficient.
D3DFORMAT dstFmt = GetDestFormat(format, gstate.getClutPaletteFormat());
int scaleFactor;
// Auto-texture scale upto 5x rendering resolution
if (g_Config.iTexScalingLevel == 0) {
scaleFactor = g_Config.iInternalResolution;
if (scaleFactor == 0) {
scaleFactor = (PSP_CoreParameter().renderWidth + 479) / 480;
}
scaleFactor = std::min(4, scaleFactor);
if (scaleFactor == 3) {
scaleFactor = 2;
}
} else {
scaleFactor = g_Config.iTexScalingLevel;
}
// Don't scale the PPGe texture.
if (entry->addr > 0x05000000 && entry->addr < 0x08800000)
scaleFactor = 1;
if (scaleFactor != 1 && (entry->status & TexCacheEntry::STATUS_CHANGE_FREQUENT) == 0) {
if (texelsScaledThisFrame_ >= TEXCACHE_MAX_TEXELS_SCALED) {
entry->status |= TexCacheEntry::STATUS_TO_SCALE;
scaleFactor = 1;
// INFO_LOG(G3D, "Skipped scaling for now..");
} else {
entry->status &= ~TexCacheEntry::STATUS_TO_SCALE;
texelsScaledThisFrame_ += w * h;
}
}
if (replaceImages) {
// Make sure it's not currently set.
pD3Ddevice->SetTexture(0, NULL);
}
// Seems to cause problems in Tactics Ogre.
if (badMipSizes) {
maxLevel = 0;
}
LoadTextureLevel(*entry, 0, maxLevel, replaceImages, scaleFactor, dstFmt);
if (!entry->texture) {
return;
}
// Mipmapping is only enabled when texture scaling is disabled.
if (maxLevel > 0 && g_Config.iTexScalingLevel == 1) {
for (int i = 1; i <= maxLevel; i++) {
LoadTextureLevel(*entry, i, maxLevel, replaceImages, scaleFactor, dstFmt);
}
}
pD3Ddevice->SetTexture(0, entry->texture);
lastBoundTexture = entry->texture;
gstate_c.textureFullAlpha = entry->GetAlphaStatus() == TexCacheEntry::STATUS_ALPHA_FULL;
gstate_c.textureSimpleAlpha = entry->GetAlphaStatus() != TexCacheEntry::STATUS_ALPHA_UNKNOWN;
UpdateSamplingParams(*entry, true);
}
D3DFORMAT TextureCacheDX9::GetDestFormat(GETextureFormat format, GEPaletteFormat clutFormat) const {
switch (format) {
case GE_TFMT_CLUT4:
case GE_TFMT_CLUT8:
case GE_TFMT_CLUT16:
case GE_TFMT_CLUT32:
return getClutDestFormat(clutFormat);
case GE_TFMT_4444:
return D3DFMT_A4R4G4B4;
case GE_TFMT_5551:
return D3DFMT_A1R5G5B5;
case GE_TFMT_5650:
return D3DFMT_R5G6B5;
case GE_TFMT_8888:
case GE_TFMT_DXT1:
case GE_TFMT_DXT3:
case GE_TFMT_DXT5:
default:
return D3DFMT_A8R8G8B8;
}
}
void *TextureCacheDX9::DecodeTextureLevel(GETextureFormat format, GEPaletteFormat clutformat, int level, u32 &texByteAlign, u32 &dstFmt, int *bufwout) {
void *finalBuf = NULL;
u32 texaddr = gstate.getTextureAddress(level);
if (texaddr & 0x00600000 && Memory::IsVRAMAddress(texaddr)) {
// This means it's in a mirror, possibly a swizzled mirror. Let's report.
WARN_LOG_REPORT_ONCE(texmirror, G3D, "Decoding texture from VRAM mirror at %08x swizzle=%d", texaddr, gstate.isTextureSwizzled() ? 1 : 0);
}
int bufw = GetTextureBufw(level, texaddr, format);
if (bufwout)
*bufwout = bufw;
int w = gstate.getTextureWidth(level);
int h = gstate.getTextureHeight(level);
const u8 *texptr = Memory::GetPointer(texaddr);
switch (format) {
case GE_TFMT_CLUT4:
{
const bool mipmapShareClut = gstate.isClutSharedForMipmaps();
const int clutSharingOffset = mipmapShareClut ? 0 : level * 16;
switch (clutformat) {
case GE_CMODE_16BIT_BGR5650:
case GE_CMODE_16BIT_ABGR5551:
case GE_CMODE_16BIT_ABGR4444:
{
tmpTexBuf16.resize(std::max(bufw, w) * h);
tmpTexBufRearrange.resize(std::max(bufw, w) * h);
const u16 *clut = GetCurrentClut<u16>() + clutSharingOffset;
texByteAlign = 2;
if (!gstate.isTextureSwizzled()) {
if (clutAlphaLinear_ && mipmapShareClut) {
DeIndexTexture4Optimal(tmpTexBuf16.data(), texptr, bufw * h, clutAlphaLinearColor_);
} else {
DeIndexTexture4(tmpTexBuf16.data(), texptr, bufw * h, clut);
}
} else {
tmpTexBuf32.resize(std::max(bufw, w) * h);
UnswizzleFromMem(texptr, bufw, 0, level);
if (clutAlphaLinear_ && mipmapShareClut) {
DeIndexTexture4Optimal(tmpTexBuf16.data(), (const u8 *)tmpTexBuf32.data(), bufw * h, clutAlphaLinearColor_);
} else {
DeIndexTexture4(tmpTexBuf16.data(), (const u8 *)tmpTexBuf32.data(), bufw * h, clut);
}
}
finalBuf = tmpTexBuf16.data();
}
break;
case GE_CMODE_32BIT_ABGR8888:
{
tmpTexBuf32.resize(std::max(bufw, w) * h);
tmpTexBufRearrange.resize(std::max(bufw, w) * h);
const u32 *clut = GetCurrentClut<u32>() + clutSharingOffset;
if (!gstate.isTextureSwizzled()) {
DeIndexTexture4(tmpTexBuf32.data(), texptr, bufw * h, clut);
finalBuf = tmpTexBuf32.data();
} else {
UnswizzleFromMem(texptr, bufw, 0, level);
// Let's reuse tmpTexBuf16, just need double the space.
tmpTexBuf16.resize(std::max(bufw, w) * h * 2);
DeIndexTexture4((u32 *)tmpTexBuf16.data(), (u8 *)tmpTexBuf32.data(), bufw * h, clut);
finalBuf = tmpTexBuf16.data();
}
}
break;
default:
ERROR_LOG_REPORT(G3D, "Unknown CLUT4 texture mode %d", gstate.getClutPaletteFormat());
return NULL;
}
}
break;
case GE_TFMT_CLUT8:
texByteAlign = texByteAlignMap[gstate.getClutPaletteFormat()];
finalBuf = ReadIndexedTex(level, texptr, 1, dstFmt, bufw);
break;
case GE_TFMT_CLUT16:
texByteAlign = texByteAlignMap[gstate.getClutPaletteFormat()];
finalBuf = ReadIndexedTex(level, texptr, 2, dstFmt, bufw);
break;
case GE_TFMT_CLUT32:
texByteAlign = texByteAlignMap[gstate.getClutPaletteFormat()];
finalBuf = ReadIndexedTex(level, texptr, 4, dstFmt, bufw);
break;
case GE_TFMT_4444:
case GE_TFMT_5551:
case GE_TFMT_5650:
texByteAlign = 2;
if (!gstate.isTextureSwizzled()) {
int len = std::max(bufw, w) * h;
tmpTexBuf16.resize(len);
tmpTexBufRearrange.resize(len);
Memory::Memcpy(tmpTexBuf16.data(), texaddr, len * sizeof(u16));
finalBuf = tmpTexBuf16.data();
}
else {
tmpTexBuf32.resize(std::max(bufw, w) * h);
finalBuf = UnswizzleFromMem(texptr, bufw, 2, level);
}
break;
case GE_TFMT_8888:
if (!gstate.isTextureSwizzled()) {
// Special case: if we don't need to deal with packing, we don't need to copy.
//if (w == bufw) {
// finalBuf = Memory::GetPointer(texaddr);
//} else
{
int len = bufw * h;
tmpTexBuf32.resize(std::max(bufw, w) * h);
tmpTexBufRearrange.resize(std::max(bufw, w) * h);
Memory::Memcpy(tmpTexBuf32.data(), texaddr, len * sizeof(u32));
finalBuf = tmpTexBuf32.data();
}
} else {
tmpTexBuf32.resize(std::max(bufw, w) * h);
finalBuf = UnswizzleFromMem(texptr, bufw, 4, level);
}
break;
case GE_TFMT_DXT1:
{
int minw = std::min(bufw, w);
tmpTexBuf32.resize(std::max(bufw, w) * h);
tmpTexBufRearrange.resize(std::max(bufw, w) * h);
u32 *dst = tmpTexBuf32.data();
DXT1Block *src = (DXT1Block*)texptr;
for (int y = 0; y < h; y += 4) {
u32 blockIndex = (y / 4) * (bufw / 4);
for (int x = 0; x < minw; x += 4) {
DecodeDXT1Block(dst + bufw * y + x, src + blockIndex, bufw);
blockIndex++;
}
}
finalBuf = tmpTexBuf32.data();
w = (w + 3) & ~3;
}
break;
case GE_TFMT_DXT3:
{
int minw = std::min(bufw, w);
tmpTexBuf32.resize(std::max(bufw, w) * h);
tmpTexBufRearrange.resize(std::max(bufw, w) * h);
u32 *dst = tmpTexBuf32.data();
DXT3Block *src = (DXT3Block*)texptr;
for (int y = 0; y < h; y += 4) {
u32 blockIndex = (y / 4) * (bufw / 4);
for (int x = 0; x < minw; x += 4) {
DecodeDXT3Block(dst + bufw * y + x, src + blockIndex, bufw);
blockIndex++;
}
}
w = (w + 3) & ~3;
finalBuf = tmpTexBuf32.data();
}
break;
case GE_TFMT_DXT5:
{
int minw = std::min(bufw, w);
tmpTexBuf32.resize(std::max(bufw, w) * h);
tmpTexBufRearrange.resize(std::max(bufw, w) * h);
u32 *dst = tmpTexBuf32.data();
DXT5Block *src = (DXT5Block*)texptr;
for (int y = 0; y < h; y += 4) {
u32 blockIndex = (y / 4) * (bufw / 4);
for (int x = 0; x < minw; x += 4) {
DecodeDXT5Block(dst + bufw * y + x, src + blockIndex, bufw);
blockIndex++;
}
}
w = (w + 3) & ~3;
finalBuf = tmpTexBuf32.data();
}
break;
default:
ERROR_LOG_REPORT(G3D, "Unknown Texture Format %d!!!", format);
return NULL;
}
if (!finalBuf) {
ERROR_LOG_REPORT(G3D, "NO finalbuf! Will crash!");
}
if (w != bufw) {
int pixelSize;
switch (dstFmt) {
case D3DFMT_A4R4G4B4:
case D3DFMT_A1R5G5B5:
case D3DFMT_R5G6B5:
pixelSize = 2;
break;
default:
pixelSize = 4;
break;
}
// Need to rearrange the buffer to simulate GL_UNPACK_ROW_LENGTH etc.
int inRowBytes = bufw * pixelSize;
int outRowBytes = w * pixelSize;
const u8 *read = (const u8 *)finalBuf;
u8 *write = 0;
if (w > bufw) {
write = (u8 *)tmpTexBufRearrange.data();
finalBuf = tmpTexBufRearrange.data();
} else {
write = (u8 *)finalBuf;
}
for (int y = 0; y < h; y++) {
memmove(write, read, outRowBytes);
read += inRowBytes;
write += outRowBytes;
}
}
return finalBuf;
}
TextureCacheDX9::TexCacheEntry::Status TextureCacheDX9::CheckAlpha(const u32 *pixelData, u32 dstFmt, int stride, int w, int h) {
// TODO: Could probably be optimized more.
u32 hitZeroAlpha = 0;
u32 hitSomeAlpha = 0;
switch (dstFmt) {
case D3DFMT_A4R4G4B4:
{
const u32 *p = pixelData;
for (int y = 0; y < h && hitSomeAlpha == 0; ++y) {
for (int i = 0; i < (w + 1) / 2; ++i) {
u32 a = p[i] & 0xF000F000;
hitZeroAlpha |= a ^ 0xF000F000;
if (a != 0xF000F000 && a != 0xF0000000 && a != 0x0000F000 && a != 0) {
hitSomeAlpha = 1;
break;
}
}
p += stride/2;
}
}
break;
case D3DFMT_A1R5G5B5:
{
const u32 *p = pixelData;
for (int y = 0; y < h; ++y) {
for (int i = 0; i < (w + 1) / 2; ++i) {
u32 a = p[i] & 0x80008000;
hitZeroAlpha |= a ^ 0x80008000;
}
p += stride/2;
}
}
break;
case D3DFMT_R5G6B5:
{
// Never has any alpha.
}
break;
default:
{
const u32 *p = pixelData;
for (int y = 0; y < h && hitSomeAlpha == 0; ++y) {
for (int i = 0; i < w; ++i) {
u32 a = p[i] & 0xFF000000;
hitZeroAlpha |= a ^ 0xFF000000;
if (a != 0xFF000000 && a != 0) {
hitSomeAlpha = 1;
break;
}
}
p += stride;
}
}
break;
}
if (hitSomeAlpha != 0)
return TexCacheEntry::STATUS_ALPHA_UNKNOWN;
else if (hitZeroAlpha != 0)
return TexCacheEntry::STATUS_ALPHA_SIMPLE;
else
return TexCacheEntry::STATUS_ALPHA_FULL;
}
static inline void copyTexture(int xoffset, int yoffset, int w, int h, int pitch, int srcfmt, int fmt, void * pSrc, void * pDst) {
int y;
switch(fmt) {
case D3DFMT_R5G6B5:
case D3DFMT_A4R4G4B4:
case D3DFMT_A1R5G5B5:
for(y = 0; y < h; y++) {
const u16 *src = (const u16 *)((u8*)pSrc + (w*2) * y);
u16 *dst = (u16*)((u8*)pDst + pitch * y);
memcpy(dst, src, w * sizeof(u16));
}
break;
// 32 bit texture
case D3DFMT_A8R8G8B8:
for(y = 0; y < h; y++) {
const u32 *src = (const u32 *)((u8*)pSrc + (w*4) * y);
u32 *dst = (u32*)((u8*)pDst + pitch * y);
memcpy(dst, src, w * sizeof(u32));
}
break;
}
}
void TextureCacheDX9::LoadTextureLevel(TexCacheEntry &entry, int level, int maxLevel, bool replaceImages, int scaleFactor, u32 dstFmt) {
// TODO: only do this once
u32 texByteAlign = 1;
GEPaletteFormat clutformat = gstate.getClutPaletteFormat();
int bufw;
void *finalBuf = DecodeTextureLevel(GETextureFormat(entry.format), clutformat, level, texByteAlign, dstFmt, &bufw);
if (finalBuf == NULL) {
return;
}
int w = gstate.getTextureWidth(level);
int h = gstate.getTextureHeight(level);
gpuStats.numTexturesDecoded++;
u32 *pixelData = (u32 *)finalBuf;
if (scaleFactor > 1 && (entry.status & TexCacheEntry::STATUS_CHANGE_FREQUENT) == 0)
scaler.Scale(pixelData, dstFmt, w, h, scaleFactor);
if ((entry.status & TexCacheEntry::STATUS_CHANGE_FREQUENT) == 0) {
TexCacheEntry::Status alphaStatus = CheckAlpha(pixelData, dstFmt, w, w, h);
entry.SetAlphaStatus(alphaStatus, level);
} else {
entry.SetAlphaStatus(TexCacheEntry::STATUS_ALPHA_UNKNOWN);
}
if (level == 0 && (!replaceImages || entry.texture == nullptr)) {
// Create texture
D3DPOOL pool = D3DPOOL_MANAGED;
int usage = 0;
if (pD3DdeviceEx) {
pool = D3DPOOL_DEFAULT;
usage = D3DUSAGE_DYNAMIC; // TODO: Switch to using a staging texture?
}
int levels = g_Config.iTexScalingLevel == 1 ? maxLevel + 1 : 1;
HRESULT hr = pD3Ddevice->CreateTexture(w, h, levels, usage, (D3DFORMAT)D3DFMT(dstFmt), pool, &entry.texture, NULL);
if (FAILED(hr)) {
INFO_LOG(G3D, "Failed to create D3D texture");
entry.ReleaseTexture();
return;
}
}
D3DLOCKED_RECT rect;
entry.texture->LockRect(level, &rect, NULL, 0);
copyTexture(0, 0, w, h, rect.Pitch, entry.format, dstFmt, pixelData, rect.pBits);
entry.texture->UnlockRect(level);
}
bool TextureCacheDX9::DecodeTexture(u8 *output, const GPUgstate &state)
{
OutputDebugStringA("TextureCache::DecodeTexture : FixMe\r\n");
return true;
}
};
Reference in a new issue View git blame Copy permalink