ShuriZma
/
pcsx2
mirror of https://github.com/PCSX2/pcsx2.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

GS-HW: use page aligned rects for inexact matches when possible

This commit is contained in:
refractionpcsx2 2023-02-27 10:58:54 +00:00
parent 500b449422
commit 5cbcf706e9
5 changed files with 336 additions and 245 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
bin/resources/GameIndex.yaml
View File

@ -2117,9 +2117,8 @@ SCED-52049:
clampModes: clampModes:
vuClampMode: 3 # Fixes minor SPS on characters. vuClampMode: 3 # Fixes minor SPS on characters.
gsHWFixes: gsHWFixes:
preloadFrameData: 1 # Fixes bad textures on Jake.
halfPixelOffset: 1 # Fixes double image. halfPixelOffset: 1 # Fixes double image.
cpuCLUTRender: 1 # Fixes Jake going black in Smellovision. preloadFrameData: 1 # Stops Jake getting black spots coming out of Smell-o-vision.
SCED-52051: SCED-52051:
name: "Official PlayStation 2 Magazine Demo 43" name: "Official PlayStation 2 Magazine Demo 43"
region: "PAL-M5" region: "PAL-M5"
@ -3338,9 +3337,8 @@ SCES-51248:
vuClampMode: 3 # Fixes minor SPS on characters. vuClampMode: 3 # Fixes minor SPS on characters.
gsHWFixes: gsHWFixes:
roundSprite: 1 # Fix lines in the sky. roundSprite: 1 # Fix lines in the sky.
preloadFrameData: 1 # Fixes bad textures on Jake.
halfPixelOffset: 1 # Fixes double image. halfPixelOffset: 1 # Fixes double image.
cpuCLUTRender: 1 # Fixes Jake going black in Smellovision. preloadFrameData: 1 # Stops Jake getting black spots coming out of Smell-o-vision.
SCES-51426: SCES-51426:
name: "Getaway, The" name: "Getaway, The"
region: "PAL-M5" region: "PAL-M5"
@ -31364,9 +31362,8 @@ SLPM-65995:
clampModes: clampModes:
vuClampMode: 3 # Fixes minor SPS on characters. vuClampMode: 3 # Fixes minor SPS on characters.
gsHWFixes: gsHWFixes:
preloadFrameData: 1 # Fixes bad textures on Jake.
halfPixelOffset: 1 # Fixes double image. halfPixelOffset: 1 # Fixes double image.
cpuCLUTRender: 1 # Fixes Jake going black in Smellovision. preloadFrameData: 1 # Stops Jake getting black spots coming out of Smell-o-vision.
SLPM-65996: SLPM-65996:
name: "Hametsu no Mars [Limited Edition]" name: "Hametsu no Mars [Limited Edition]"
region: "NTSC-J" region: "NTSC-J"
@ -45989,9 +45986,8 @@ SLUS-21018:
clampModes: clampModes:
vuClampMode: 3 # Fixes minor SPS on characters. vuClampMode: 3 # Fixes minor SPS on characters.
gsHWFixes: gsHWFixes:
preloadFrameData: 1 # Fixes bad textures on Jake. halfPixelOffset: 1 # Fixes double image.
halfPixelOffset: 1 # Fixes double image.gsHWFixes: preloadFrameData: 1 # Stops Jake getting black spots coming out of Smell-o-vision.
cpuCLUTRender: 1 # Fixes Jake going black in Smellovision.
SLUS-21019: SLUS-21019:
name: "Technic Beat" name: "Technic Beat"
region: "NTSC-U" region: "NTSC-U"

44
pcsx2/GS/GSLocalMemory.cpp
View File

@ -566,6 +566,50 @@ void GSLocalMemory::SaveBMP(const std::string& fn, u32 bp, u32 bw, u32 psm, int
_aligned_free(bits); _aligned_free(bits);
} }
bool GSLocalMemory::IsPageAligned(u32 bp, u32 spsm, GSVector4i r, bool bppbw_match)
{
const bool bp_page_aligned_bp = (bp & ~((1 << 5) - 1)) == bp;
const GSVector2i page_size = m_psm[spsm].pgs;
const GSVector4i page_mask(GSVector4i(~(page_size.x - 1), ~(page_size.y - 1)).xyxy());
const GSVector4i masked_rect(r & page_mask);
// If the BPP and BW matches, then just make sure it's starting on the edge of a page.
if (bppbw_match)
{
return bp_page_aligned_bp && masked_rect.xyxy().eq(r.xyxy());
}
else
{
return bp_page_aligned_bp && masked_rect.eq(r);
}
}
GSVector4i GSLocalMemory::TranslateAlignedRectByPage(u32 sbp, u32 spsm, GSVector4i src_r, u32 dbp, u32 dpsm, u32 bw)
{
const GSVector2i src_page_size = m_psm[spsm].pgs;
const GSVector2i dst_page_size = m_psm[dpsm].pgs;
const int page_offset = (static_cast<int>(sbp) - static_cast<int>(dbp)) >> 5;
const int vertical_offset = (page_offset / static_cast<int>(bw)) * dst_page_size.y;
const int horizontal_offset = (page_offset % static_cast<int>(bw)) * dst_page_size.x;
const GSVector4i rect_pages = GSVector4i(src_r.x / src_page_size.x, src_r.y / src_page_size.y, src_r.z / src_page_size.x, src_r.w / src_page_size.y);
GSVector4i new_rect = {};
// If they match, we can cheat and just offset the rect by the number of pages.
if (m_psm[spsm].bpp == m_psm[dpsm].bpp)
{
new_rect = src_r;
}
else
{
new_rect = GSVector4i(rect_pages.x * dst_page_size.x, rect_pages.y * dst_page_size.y, rect_pages.z * dst_page_size.x, rect_pages.w * dst_page_size.y);
}
new_rect = (new_rect + GSVector4i(0, vertical_offset).xyxy()).max_i32(GSVector4i(0));
new_rect = (new_rect + GSVector4i(horizontal_offset, 0).xyxy()).max_i32(GSVector4i(0));
return new_rect;
}
// GSOffset // GSOffset
namespace namespace

3
pcsx2/GS/GSLocalMemory.h
View File

@ -1133,6 +1133,9 @@ public:
// //
void SaveBMP(const std::string& fn, u32 bp, u32 bw, u32 psm, int w, int h); void SaveBMP(const std::string& fn, u32 bp, u32 bw, u32 psm, int w, int h);
static bool IsPageAligned(u32 bp, u32 spsm, GSVector4i r, bool bp_match);
static GSVector4i TranslateAlignedRectByPage(u32 sbp, u32 spsm, GSVector4i src_r, u32 dbp, u32 dpsm, u32 bw);
}; };
constexpr inline GSOffset GSOffset::fromKnownPSM(u32 bp, u32 bw, GS_PSM psm) constexpr inline GSOffset GSOffset::fromKnownPSM(u32 bp, u32 bw, GS_PSM psm)

38
pcsx2/GS/Renderers/HW/GSRendererHW.cpp
View File

@ -970,47 +970,19 @@ void GSRendererHW::InvalidateLocalMem(const GIFRegBITBLTBUF& BITBLTBUF, const GS
if (clut) if (clut)
return; // FIXME return; // FIXME
const u32 incoming_end = GSLocalMemory::m_psm[BITBLTBUF.SPSM].info.bn(r.z - 1, r.w - 1, BITBLTBUF.SBP, BITBLTBUF.SBW);
std::vector<GSState::GSUploadQueue>::iterator iter = GSRendererHW::GetInstance()->m_draw_transfers.end(); std::vector<GSState::GSUploadQueue>::iterator iter = GSRendererHW::GetInstance()->m_draw_transfers.end();
bool skip = false; bool skip = false;
// If the EE write overlaps the readback and was done since the last draw, there's no need to read it back. // If the EE write overlaps the readback and was done since the last draw, there's no need to read it back.
// Dog's life and Ratchet Gladiator do this. // Dog's life does this.
while (iter != GSRendererHW::GetInstance()->m_draw_transfers.begin()) while (iter != GSRendererHW::GetInstance()->m_draw_transfers.begin())
{ {
--iter; --iter;
if (!GSUtil::HasSharedBits(iter->blit.DPSM, BITBLTBUF.SPSM) || iter->draw != s_n) if (!(iter->draw == s_n && BITBLTBUF.SBP == iter->blit.DBP && iter->blit.DPSM == BITBLTBUF.SPSM && r.eq(iter->rect)))
continue; continue;
m_tc->InvalidateVideoMem(m_mem.GetOffset(BITBLTBUF.SBP, BITBLTBUF.SBW, BITBLTBUF.SPSM), r);
// Make sure write covers the read area. skip = true;
const u32 ee_write_end = GSLocalMemory::m_psm[iter->blit.DPSM].info.bn(iter->rect.z - 1, iter->rect.w - 1, iter->blit.DBP, iter->blit.DBW); break;
if (!(iter->blit.DBP < incoming_end && ee_write_end > BITBLTBUF.SBP))
continue;
GSTextureCache::SurfaceOffsetKey sok;
sok.elems[0].bp = BITBLTBUF.SBP;
sok.elems[0].bw = BITBLTBUF.SBW;
sok.elems[0].psm = BITBLTBUF.SPSM;
sok.elems[0].rect = r;
sok.elems[1].bp = iter->blit.DBP;
sok.elems[1].bw = iter->blit.DBW;
sok.elems[1].psm = iter->blit.DPSM;
sok.elems[1].rect = iter->rect;
// Calculate the rect offset if the BP doesn't match.
const GSVector4i targetr = GSUtil::HasCompatibleBits(iter->blit.DPSM, BITBLTBUF.SPSM) ? r : m_tc->ComputeSurfaceOffset(sok).b2a_offset;
// Possibly incompatible or missed, we don't know, so let's assume it's a fail.
if (targetr.rempty())
continue;
//u32 ee_write_end = GSLocalMemory::m_psm[iter->blit.DPSM].info.bn(iter->rect.z - 1, iter->rect.w - 1, iter->blit.DBP, iter->blit.DBW);
// If the format, and location doesn't match, but also the upload is at least the size of the target, don't preload.
if (iter->rect.rintersect(targetr).eq(targetr))
{
skip = true;
}
} }
if(!skip) if(!skip)

482
pcsx2/GS/Renderers/HW/GSTextureCache.cpp
View File

@ -398,14 +398,65 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const GIFRegTEX0& TEX0, con
// //
// Solution: consider the RT as 32 bits if the alpha was used in the past // Solution: consider the RT as 32 bits if the alpha was used in the past
const u32 t_psm = (t->m_dirty_alpha) ? t->m_TEX0.PSM & ~0x1 : t->m_TEX0.PSM; const u32 t_psm = (t->m_dirty_alpha) ? t->m_TEX0.PSM & ~0x1 : t->m_TEX0.PSM;
bool rect_clean = psm == t_psm;
if (rect_clean && bp >= t->m_TEX0.TBP0 && bp < t->m_end_block && bw == t->m_TEX0.TBW && bp <= t->m_end_block && !t->m_dirty.empty())
{
const GSVector2i src_size = GSVector2i(r.z, r.w);
const GSVector4i src_rect = GSVector4i(0, 0, src_size.x, src_size.y);
GSVector4i rect = r;
const bool page_aligned = GSLocalMemory::IsPageAligned(bp, psm, r, true);
const bool bpp_match = GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == GSLocalMemory::m_psm[psm].bpp;
// If it's compatible and page aligned, then handle it this way.
// It's quicker, and Surface Offsets can get it wrong.
// Example doing PSMT8H to C32, BP 0x1c80, TBP 0x1d80, incoming rect 0,128 -> 128,256
// Surface offsets translates it to 0, 128 -> 128, 128, not 0, 0 -> 128, 128.
if (bp > t->m_TEX0.TBP0)
{
if (page_aligned && bpp_match && bw == t->m_TEX0.TBW)
{
rect = GSLocalMemory::TranslateAlignedRectByPage(bp, psm, r, t->m_TEX0.TBP0, t->m_TEX0.PSM, t->m_TEX0.TBW);
}
else
{
SurfaceOffsetKey sok;
sok.elems[0].bp = bp;
sok.elems[0].bw = bw;
sok.elems[0].psm = psm;
sok.elems[0].rect = r;
sok.elems[1].bp = t->m_TEX0.TBP0;
sok.elems[1].bw = t->m_TEX0.TBW;
sok.elems[1].psm = t->m_TEX0.PSM;
sok.elems[1].rect = t->m_valid;
const u32 channels = t->m_dirty.GetDirtyChannels() & GSUtil::GetChannelMask(psm); const SurfaceOffset so = ComputeSurfaceOffset(sok);
// If not all channels are clean/dirty, we need to update the target. if (so.is_valid)
if (channels != 0 && channels != GSUtil::GetChannelMask(psm)) {
t->Update(false); rect = so.b2a_offset;
const bool t_clean = (t->m_dirty.GetDirtyChannels() & GSUtil::GetChannelMask(psm)) == 0; }
}
}
for (auto& dirty : t->m_dirty)
{
GSVector4i dirty_rect = dirty.GetDirtyRect(t->m_TEX0);
if (!dirty_rect.rintersect(rect).rempty())
{
rect_clean = false;
break;
}
}
const u32 channel_mask = GSUtil::GetChannelMask(psm);
const u32 channels = t->m_dirty.GetDirtyChannels() & channel_mask;
// If not all channels are clean/dirty, we need to update the target.
if ((channels & channel_mask) != 0 && !rect_clean)
t->Update(false);
}
else
rect_clean = t->m_dirty.empty();
const bool t_clean = ((t->m_dirty.GetDirtyChannels() & GSUtil::GetChannelMask(psm)) == 0) || rect_clean;
const bool t_wraps = t->m_end_block > GSTextureCache::MAX_BP; const bool t_wraps = t->m_end_block > GSTextureCache::MAX_BP;
// Match if we haven't already got a tex in rt // Match if we haven't already got a tex in rt
if (t_clean && GSUtil::HasSharedBits(bp, psm, t->m_TEX0.TBP0, t_psm)) if (t_clean && GSUtil::HasSharedBits(bp, psm, t->m_TEX0.TBP0, t_psm))
{ {
@ -1483,23 +1534,37 @@ void GSTextureCache::InvalidateVideoMem(const GSOffset& off, const GSVector4i& r
} }
else if (check_inside_target && t->Overlaps(bp, bw, psm, rect) && GSUtil::HasSharedBits(psm, t->m_TEX0.PSM)) else if (check_inside_target && t->Overlaps(bp, bw, psm, rect) && GSUtil::HasSharedBits(psm, t->m_TEX0.PSM))
{ {
SurfaceOffsetKey sok; const bool bpp_match = GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == GSLocalMemory::m_psm[psm].bpp;
sok.elems[0].bp = bp; const bool page_aligned = GSLocalMemory::IsPageAligned(bp, psm, r, bpp_match && bw == t->m_TEX0.TBW);
sok.elems[0].bw = bw; // If it's compatible and page aligned, then handle it this way.
sok.elems[0].psm = psm; // It's quicker, and Surface Offsets can get it wrong.
sok.elems[0].rect = r; // Example doing PSMT8H to C32, BP 0x1c80, TBP 0x1d80, incoming rect 0,128 -> 128,256
sok.elems[1].bp = t->m_TEX0.TBP0; // Surface offsets translates it to 0, 128 -> 128, 128, not 0, 0 -> 128, 128.
sok.elems[1].bw = t->m_TEX0.TBW; if (page_aligned && bpp_match && bw == t->m_TEX0.TBW)
sok.elems[1].psm = t->m_TEX0.PSM;
sok.elems[1].rect = t->m_valid;
const SurfaceOffset so = ComputeSurfaceOffset(sok);
if (so.is_valid)
{ {
if (eewrite) GSVector4i aligned_rect = GSLocalMemory::TranslateAlignedRectByPage(bp, psm, r, t->m_TEX0.TBP0, t->m_TEX0.PSM, t->m_TEX0.TBW);
t->m_age = 0; AddDirtyRectTarget(t, aligned_rect, t->m_TEX0.PSM, t->m_TEX0.TBW, rgba);
}
else
{
SurfaceOffsetKey sok;
sok.elems[0].bp = bp;
sok.elems[0].bw = bw;
sok.elems[0].psm = psm;
sok.elems[0].rect = r;
sok.elems[1].bp = t->m_TEX0.TBP0;
sok.elems[1].bw = t->m_TEX0.TBW;
sok.elems[1].psm = t->m_TEX0.PSM;
sok.elems[1].rect = t->m_valid;
AddDirtyRectTarget(t, so.b2a_offset, t->m_TEX0.PSM, t->m_TEX0.TBW, rgba); const SurfaceOffset so = ComputeSurfaceOffset(sok);
if (so.is_valid)
{
if (eewrite)
t->m_age = 0;
AddDirtyRectTarget(t, so.b2a_offset, t->m_TEX0.PSM, t->m_TEX0.TBW, rgba);
}
} }
} }
#endif #endif
@ -1516,8 +1581,7 @@ void GSTextureCache::InvalidateLocalMem(const GSOffset& off, const GSVector4i& r
const u32 psm = off.psm(); const u32 psm = off.psm();
[[maybe_unused]] const u32 bw = off.bw(); [[maybe_unused]] const u32 bw = off.bw();
const u32 read_start = GSLocalMemory::m_psm[psm].info.bn(r.x, r.y, bp, bw); const u32 read_start = GSLocalMemory::m_psm[psm].info.bn(r.x, r.y, bp, bw);
const u32 read_end = GSLocalMemory::m_psm[psm].info.bn(r.x, r.y, bp, bw); const u32 read_end = GSLocalMemory::m_psm[psm].info.bn(r.z - 1, r.w - 1, bp, bw);
const bool read_paltex = GSLocalMemory::m_psm[psm].pal > 0;
GL_CACHE("TC: InvalidateLocalMem off(0x%x, %u, %s) r(%d, %d => %d, %d)", GL_CACHE("TC: InvalidateLocalMem off(0x%x, %u, %s) r(%d, %d => %d, %d)",
bp, bp,
@ -1538,6 +1602,8 @@ void GSTextureCache::InvalidateLocalMem(const GSOffset& off, const GSVector4i& r
return; return;
} }
bool z_found = false;
if (!GSConfig.UserHacks_DisableDepthSupport) if (!GSConfig.UserHacks_DisableDepthSupport)
{ {
auto& dss = m_dst[DepthStencil]; auto& dss = m_dst[DepthStencil];
@ -1546,215 +1612,224 @@ void GSTextureCache::InvalidateLocalMem(const GSOffset& off, const GSVector4i& r
Target* t = *it; Target* t = *it;
// Check the offset of the read, if they're not pointing at or inside this texture, it's probably not what we want. // Check the offset of the read, if they're not pointing at or inside this texture, it's probably not what we want.
const bool expecting_this_tex = ((bp < t->m_TEX0.TBP0 && read_start >= t->m_TEX0.TBP0) || bp >= t->m_TEX0.TBP0) && read_end <= t->m_end_block; //const bool expecting_this_tex = ((bp <= t->m_TEX0.TBP0 && read_start >= t->m_TEX0.TBP0) || bp >= t->m_TEX0.TBP0) && read_end <= t->m_end_block;
if (!expecting_this_tex || !t->Overlaps(bp, bw, psm, r) || !GSUtil::HasSharedBits(psm, t->m_TEX0.PSM) || ((bp != t->m_TEX0.TBP0) && !GSUtil::HasCompatibleBits(psm, t->m_TEX0.PSM)))
continue;
const bool bpp_match = GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == GSLocalMemory::m_psm[psm].bpp; const bool bpp_match = GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == GSLocalMemory::m_psm[psm].bpp;
const u32 page_mask = ((1 << 5) - 1);
const bool expecting_this_tex = bpp_match && (((read_start & ~page_mask) == t->m_TEX0.TBP0) || (bp >= t->m_TEX0.TBP0 && ((read_end + page_mask) & ~page_mask) <= ((t->m_end_block + page_mask) & ~page_mask)));
if (!expecting_this_tex)
continue;
z_found = true;
if (t->m_drawn_since_read.eq(GSVector4i::zero()))
return;
t->readbacks_since_draw++;
const bool format_match = (bp == t->m_TEX0.TBP0 && bw == t->m_TEX0.TBW && bpp_match); const bool format_match = (bp == t->m_TEX0.TBP0 && bw == t->m_TEX0.TBW && bpp_match);
// Calculate the rect offset if the BP doesn't match. // Calculate the rect offset if the BP doesn't match.
const GSVector4i targetr = (format_match) ? r.rintersect(t->m_valid) : ComputeSurfaceOffset(bp, bw, psm, r, t).b2a_offset; const GSVector4i targetr = (format_match) ? r.rintersect(t->m_valid) : ComputeSurfaceOffset(bp, bw, psm, r, t).b2a_offset;
const GSVector4i draw_rect = (t->readbacks_since_draw > 0) ? t->m_drawn_since_read : targetr.rintersect(t->m_drawn_since_read); const GSVector4i draw_rect = (t->readbacks_since_draw > 1) ? t->m_drawn_since_read : targetr.rintersect(t->m_drawn_since_read);
Read(t, draw_rect); Read(t, draw_rect);
t->readbacks_since_draw++;
if (draw_rect.rintersect(t->m_drawn_since_read).eq(t->m_drawn_since_read)) if (draw_rect.rintersect(t->m_drawn_since_read).eq(t->m_drawn_since_read))
t->m_drawn_since_read = GSVector4i::zero(); t->m_drawn_since_read = GSVector4i::zero();
} }
} }
return; if (z_found)
return;
} }
// This is a shorter but potentially slower version of the below, commented out code. // Games of note that use this for various effects/transfers which may cause problems.
// It works for all the games mentioned below and fixes a couple of other ones as well // Silent Hill Shattered Memories
// (Busen0: Wizardry and Chaos Legion). // Chaos Legion
// Also in a few games the below code ran the Grandia3 case when it shouldn't :p // Busin 0: Wizardry Alternative
// Kingdom Hearts 2
// Final Fantasy X
// Dark Cloud 2
// Dog's Life
// SOCOM 2
// Fatal Frame series
auto& rts = m_dst[RenderTarget]; auto& rts = m_dst[RenderTarget];
const bool read_is_depth = (psm & 0x30) == 0x30;
// Check exact match first
for (auto it = rts.rbegin(); it != rts.rend(); it++) // Iterate targets from LRU to MRU. for (auto it = rts.rbegin(); it != rts.rend(); it++) // Iterate targets from LRU to MRU.
{ {
Target* t = *it; Target* t = *it;
if (t->m_TEX0.PSM != PSM_PSMZ32 && t->m_TEX0.PSM != PSM_PSMZ24 && t->m_TEX0.PSM != PSM_PSMZ16 && t->m_TEX0.PSM != PSM_PSMZ16S)
if (t->m_32_bits_fmt && t->m_TEX0.PSM > PSM_PSMCT24)
t->m_TEX0.PSM = PSM_PSMCT32;
const bool bpp_match = GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == GSLocalMemory::m_psm[psm].bpp;
const u32 page_mask = ((1 << 5) - 1);
const bool expecting_this_tex = bpp_match && bw == t->m_TEX0.TBW && (((read_start & ~page_mask) == t->m_TEX0.TBP0) || (bp >= t->m_TEX0.TBP0 && ((read_end + page_mask) & ~page_mask) <= ((t->m_end_block + page_mask) & ~page_mask)));
if (!expecting_this_tex)
continue;
if (t->m_drawn_since_read.eq(GSVector4i::zero()))
return;
t->readbacks_since_draw++;
// Calculate the rect offset if the BP doesn't match.
GSVector4i targetr = {};
const bool tex_is_depth = (t->m_TEX0.PSM & 0x30) == 0x30;
const bool page_aligned = GSLocalMemory::IsPageAligned(bp, psm, r, bpp_match && bw == t->m_TEX0.TBW);
if (t->readbacks_since_draw > 1 || (!page_aligned && (tex_is_depth != read_is_depth)))
targetr = t->m_drawn_since_read;
else if (page_aligned && bw == t->m_TEX0.TBW)
{ {
// propagate the format from the result of a channel effect targetr = GSLocalMemory::TranslateAlignedRectByPage(bp, psm, r, t->m_TEX0.TBP0, t->m_TEX0.PSM, t->m_TEX0.TBW);
// texture is 16/8 bit but the real data is 32
// common use for shuffling is moving data into the alpha channel
// the game can then draw using 8H format
// in the case of silent hill blit 8H -> 8P
// this will matter later when the data ends up in GS memory in the wrong format
// Be careful to avoid 24 bit textures which are technically 32bit, as you could lose alpha (8H) data.
if (t->m_32_bits_fmt && t->m_TEX0.PSM > PSM_PSMCT24)
t->m_TEX0.PSM = PSM_PSMCT32;
// Check the offset of the read, if they're not pointing at or inside this texture, it's probably not what we want.
const bool expecting_this_tex = ((bp < t->m_TEX0.TBP0 && read_start >= t->m_TEX0.TBP0) || bp >= t->m_TEX0.TBP0) && read_end <= t->m_end_block;
const bool target_paltex = GSLocalMemory::m_psm[t->m_TEX0.PSM].pal > 0;
// Only allow an indexed format on a 32bit colour, if it's alpha channel.
const bool alpha_read = t->m_TEX0.PSM == PSM_PSMCT32 && psm >= PSM_PSMT8H;
// Okay this is a nightmare of a check, so these are the conditions:
// 1. Check if it's expecting this texture, so the read must be inside this texture, even if the BP doesn't match.
// 2. It must overlap (okay maybe redundant).
// 3. If it's a paltex (indexed format), the target must also be paltex, unless it's only reading the alpha channel.
// 4. They share bits in some capacity.
// 5. If the BP doesn't match, make sure the formats are compatible, at very least (like CT32 + CT24 for example).
if (!expecting_this_tex || !t->Overlaps(bp, bw, psm, r) || (read_paltex != target_paltex && !alpha_read) || !GSUtil::HasSharedBits(psm, t->m_TEX0.PSM) || (bp != t->m_TEX0.TBP0 && !GSUtil::HasCompatibleBits(psm, t->m_TEX0.PSM)))
continue;
const bool bpp_match = GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == GSLocalMemory::m_psm[psm].bpp;
const bool format_match = (bp == t->m_TEX0.TBP0 && bw == t->m_TEX0.TBW && bpp_match);
SurfaceOffsetKey sok;
sok.elems[0].bp = bp;
sok.elems[0].bw = bw;
sok.elems[0].psm = psm;
sok.elems[0].rect = r;
sok.elems[1].bp = t->m_TEX0.TBP0;
sok.elems[1].bw = t->m_TEX0.TBW;
sok.elems[1].psm = t->m_TEX0.PSM;
sok.elems[1].rect = t->m_valid;
// Calculate the rect offset if the BP doesn't match.
const GSVector4i targetr = GSVector4i((format_match) ? r.rintersect(t->m_valid) : ComputeSurfaceOffset(sok).b2a_offset).rintersect(t->m_drawn_since_read);
if (!targetr.rempty())
{
// GH Note: Read will do a StretchRect and then will sizzle data to the GS memory
// t->m_valid will do the full target texture whereas r.intersect(t->m_valid) will be limited
// to the useful part for the transfer.
// 1/ Logically intersect must be enough, except if we miss some call to InvalidateLocalMem
// or it need the depth part too
// 2/ Read function is slow but I suspect the swizzle part to be costly. Maybe a compute shader
// that do the swizzle at the same time of the Stretching could save CPU computation.
// note: r.rintersect breaks Wizardry and Chaos Legion
// Read(t, t->m_valid) works in all tested games but is very slow in GUST titles ><
// Update: 18/02/2023: Chaos legion breaks because it reads the width at half of the real width.
// Surface offset deals with this.
// If the game has been spamming downloads, we've already read the whole texture back at this point.
if (t->m_drawn_since_read.rempty() || !t->m_dirty.empty())
continue;
if (GSConfig.HWDownloadMode != GSHardwareDownloadMode::Enabled)
{
const GSVector4i rb_rc((!GSConfig.UserHacks_DisablePartialInvalidation && targetr.x == 0 && targetr.y == 0) ? t->m_valid : targetr.rintersect(t->m_valid));
DevCon.Error("Skipping depth readback of %ux%u @ %u,%u", rb_rc.width(), rb_rc.height(), rb_rc.left, rb_rc.top);
}
else
{
// If it's a download to the EE, or we've done multiple reads of the same texture between draws, just get it all.
if (!GSConfig.UserHacks_DisablePartialInvalidation && t->readbacks_since_draw > 0)
{
Read(t, t->m_drawn_since_read);
t->m_drawn_since_read = GSVector4i::zero();
t->readbacks_since_draw++;
}
else if(!targetr.rintersect(t->m_drawn_since_read).rempty()) // Block level read?
{
// Read the width of the draw, reading too much could wipe out dirty memory.
Read(t, targetr);
// After reading, try to cut down our "dirty" rect.
if (targetr.rintersect(t->m_drawn_since_read).eq(t->m_drawn_since_read))
t->m_drawn_since_read = GSVector4i::zero();
else
{
// Try to cut down how much we read next, if we can.
// Fatal Frame reads in vertical strips, SOCOM 2 does horizontal, so we can handle that below.
if (targetr.width() == t->m_drawn_since_read.width()
&& targetr.w >= t->m_drawn_since_read.y)
{
if (targetr.y <= t->m_drawn_since_read.y)
t->m_drawn_since_read.y = targetr.w;
else if (targetr.w >= t->m_drawn_since_read.w)
t->m_drawn_since_read.w = targetr.y;
}
else if (targetr.height() == t->m_drawn_since_read.height()
&& targetr.z >= t->m_drawn_since_read.x)
{
if (targetr.x <= t->m_drawn_since_read.x)
t->m_drawn_since_read.x = targetr.z;
else if (targetr.z >= t->m_drawn_since_read.z)
t->m_drawn_since_read.z = targetr.x;
}
}
t->readbacks_since_draw++;
}
}
}
} }
else else
{ {
GL_INS("ERROR: InvalidateLocalMem target is a depth format"); const bool format_match = (bp == t->m_TEX0.TBP0 && bw == t->m_TEX0.TBW && bpp_match);
if (!format_match)
{
SurfaceOffsetKey sok;
sok.elems[0].bp = bp;
sok.elems[0].bw = bw;
sok.elems[0].psm = psm;
sok.elems[0].rect = r;
sok.elems[1].bp = t->m_TEX0.TBP0;
sok.elems[1].bw = t->m_TEX0.TBW;
sok.elems[1].psm = t->m_TEX0.PSM;
sok.elems[1].rect = t->m_valid;
SurfaceOffset so = ComputeSurfaceOffset(sok);
targetr = so.b2a_offset;
}
else
{
targetr = r;
}
}
targetr = targetr.rintersect(t->m_drawn_since_read);
if (t->m_drawn_since_read.rempty())
{
// we found the exact one but it's already been read back.
if (targetr.rintersect(t->m_valid).eq(targetr))
return;
}
if (!targetr.rempty())
{
Read(t, targetr);
// Try to cut down how much we read next, if we can.
// Fatal Frame reads in vertical strips, SOCOM 2 does horizontal, so we can handle that below.
if (t->m_drawn_since_read.rintersect(targetr).eq(t->m_drawn_since_read))
{
t->m_drawn_since_read = GSVector4i::zero();
}
else if (targetr.width() == t->m_drawn_since_read.width()
&& targetr.w >= t->m_drawn_since_read.y)
{
if (targetr.y <= t->m_drawn_since_read.y)
t->m_drawn_since_read.y = targetr.w;
else if (targetr.w >= t->m_drawn_since_read.w)
t->m_drawn_since_read.w = targetr.y;
}
else if (targetr.height() == t->m_drawn_since_read.height()
&& targetr.z >= t->m_drawn_since_read.x)
{
if (targetr.x <= t->m_drawn_since_read.x)
t->m_drawn_since_read.x = targetr.z;
else if (targetr.z >= t->m_drawn_since_read.z)
t->m_drawn_since_read.z = targetr.x;
}
if (targetr.rintersect(t->m_valid).eq(targetr))
return;
} }
} }
// Check loose matches if we still haven't got all the data.
for (auto it = rts.rbegin(); it != rts.rend(); it++) // Iterate targets from LRU to MRU.
{
Target* t = *it;
//GSTextureCache::Target* rt2 = NULL; if (t->m_32_bits_fmt && t->m_TEX0.PSM > PSM_PSMCT24)
//int ymin = INT_MAX; t->m_TEX0.PSM = PSM_PSMCT32;
//for(auto i = m_dst[RenderTarget].begin(); i != m_dst[RenderTarget].end(); )
//{
// auto j = i++;
// Target* t = *j; const bool bpp_mismatch = GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp != GSLocalMemory::m_psm[psm].bpp;
const bool expecting_this_tex = t->Overlaps(bp, bw, psm, r);
if (!expecting_this_tex || !GSUtil::HasSharedBits(psm, t->m_TEX0.PSM))
continue;
// if (t->m_TEX0.PSM != PSM_PSMZ32 && t->m_TEX0.PSM != PSM_PSMZ24 && t->m_TEX0.PSM != PSM_PSMZ16 && t->m_TEX0.PSM != PSM_PSMZ16S) if (t->m_drawn_since_read.eq(GSVector4i::zero()))
// { return;
// if (GSUtil::HasSharedBits(bp, psm, t->m_TEX0.TBP0, t->m_TEX0.PSM))
// {
// if (GSUtil::HasCompatibleBits(psm, t->m_TEX0.PSM))
// {
// Read(t, r.rintersect(t->m_valid));
// return;
// }
// else if (psm == PSM_PSMCT32 && (t->m_TEX0.PSM == PSM_PSMCT16 || t->m_TEX0.PSM == PSM_PSMCT16S))
// {
// // ffx-2 riku changing to her default (shoots some reflecting glass at the end), 16-bit rt read as 32-bit
// Read(t, GSVector4i(r.left, r.top, r.right, r.top + (r.bottom - r.top) * 2).rintersect(t->m_valid));
// return;
// }
// else
// {
// if (psm == PSM_PSMT4HH && t->m_TEX0.PSM == PSM_PSMCT32)
// {
// // Silent Hill Origins shadows: Read 8 bit using only the HIGH bits (4 bit) texture as 32 bit.
// Read(t, r.rintersect(t->m_valid));
// return;
// }
// else
// {
// //printf("Trashing render target. We have a %d type texture and we are trying to write into a %d type texture\n", t->m_TEX0.PSM, psm);
// m_dst[RenderTarget].erase(j);
// delete t;
// }
// }
// }
// // Grandia3, FFX, FFX-2 pause menus. t->m_TEX0.TBP0 magic number checks because otherwise kills xs2 videos t->readbacks_since_draw++;
// if ((GSUtil::HasSharedBits(psm, t->m_TEX0.PSM) && (bp > t->m_TEX0.TBP0))
// && ((t->m_TEX0.TBP0 == 0) || (t->m_TEX0.TBP0==3328) || (t->m_TEX0.TBP0==3584)))
// {
// //printf("first : %d-%d child : %d-%d\n", psm, bp, t->m_TEX0.PSM, t->m_TEX0.TBP0);
// u32 rowsize = bw * 8192;
// u32 offset = (u32)((bp - t->m_TEX0.TBP0) * 256);
// if (rowsize > 0 && offset % rowsize == 0) // Calculate the rect offset if the BP doesn't match.
// { GSVector4i targetr = {};
// int y = GSLocalMemory::m_psm[psm].pgs.y * offset / rowsize; const bool tex_is_depth = (t->m_TEX0.PSM & 0x30) == 0x30;
const bool page_aligned = GSLocalMemory::IsPageAligned(bp, psm, r, !bpp_mismatch && bw == t->m_TEX0.TBW);
// if (y < ymin && y < 512) if (t->readbacks_since_draw > 1 || !page_aligned || (tex_is_depth != read_is_depth))
// { targetr = t->m_drawn_since_read;
// rt2 = t; else
// ymin = y; {
// } if (page_aligned && bw == t->m_TEX0.TBW)
// } {
// } targetr = GSLocalMemory::TranslateAlignedRectByPage(bp, psm, r, t->m_TEX0.TBP0, t->m_TEX0.PSM, t->m_TEX0.TBW);
// } }
//} else
//if (rt2) {
//{ const bool format_match = (bp == t->m_TEX0.TBP0 && bw == t->m_TEX0.TBW && !bpp_mismatch);
// Read(rt2, GSVector4i(r.left, r.top + ymin, r.right, r.bottom + ymin)); SurfaceOffsetKey sok;
//} sok.elems[0].bp = bp;
sok.elems[0].bw = bw;
sok.elems[0].psm = psm;
sok.elems[0].rect = r;
sok.elems[1].bp = t->m_TEX0.TBP0;
sok.elems[1].bw = t->m_TEX0.TBW;
sok.elems[1].psm = t->m_TEX0.PSM;
sok.elems[1].rect = t->m_valid;
SurfaceOffset so = ComputeSurfaceOffset(sok);
targetr = GSVector4i((format_match) ? r.rintersect(t->m_drawn_since_read) : so.b2a_offset);
}
}
if (!targetr.rempty())
{
Read(t, targetr);
// TODO: ds // Try to cut down how much we read next, if we can.
// Fatal Frame reads in vertical strips, SOCOM 2 does horizontal, so we can handle that below.
if (t->m_drawn_since_read.rintersect(targetr).eq(t->m_drawn_since_read))
{
t->m_drawn_since_read = GSVector4i::zero();
}
else if (targetr.width() == t->m_drawn_since_read.width()
&& targetr.w >= t->m_drawn_since_read.y)
{
if (targetr.y <= t->m_drawn_since_read.y)
t->m_drawn_since_read.y = targetr.w;
else if (targetr.w >= t->m_drawn_since_read.w)
t->m_drawn_since_read.w = targetr.y;
}
else if (targetr.height() == t->m_drawn_since_read.height()
&& targetr.z >= t->m_drawn_since_read.x)
{
if (targetr.x <= t->m_drawn_since_read.x)
t->m_drawn_since_read.x = targetr.z;
else if (targetr.z >= t->m_drawn_since_read.z)
t->m_drawn_since_read.z = targetr.x;
}
if (targetr.rintersect(t->m_valid).eq(targetr))
return;
}
}
} }
bool GSTextureCache::Move(u32 SBP, u32 SBW, u32 SPSM, int sx, int sy, u32 DBP, u32 DBW, u32 DPSM, int dx, int dy, int w, int h) bool GSTextureCache::Move(u32 SBP, u32 SBW, u32 SPSM, int sx, int sy, u32 DBP, u32 DBW, u32 DPSM, int dx, int dy, int w, int h)
@ -1861,6 +1936,7 @@ bool GSTextureCache::Move(u32 SBP, u32 SBW, u32 SPSM, int sx, int sy, u32 DBP, u
GSVector4i(scaled_sx, scaled_sy, scaled_sx + scaled_w, scaled_sy + scaled_h), GSVector4i(scaled_sx, scaled_sy, scaled_sx + scaled_w, scaled_sy + scaled_h),
scaled_dx, scaled_dy); scaled_dx, scaled_dy);
dst->UpdateValidity(GSVector4i(dx, dy, dx + w, dy + h));
// Invalidate any sources that overlap with the target (since they're now stale). // Invalidate any sources that overlap with the target (since they're now stale).
InvalidateVideoMem(g_gs_renderer->m_mem.GetOffset(DBP, DBW, DPSM), GSVector4i(dx, dy, dx + w, dy + h), false, false); InvalidateVideoMem(g_gs_renderer->m_mem.GetOffset(DBP, DBW, DPSM), GSVector4i(dx, dy, dx + w, dy + h), false, false);
return true; return true;