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

GS/HW: Improve detection of clears 

- Detect "normal" target clears and HLE accordingly
 - Rewrite double-half clears to mask Z or FRAME, reducing false target creation.
 - Handle split single-page-wide clears, and attempt to get a real FBW.
 - Propogate clear values between targets, instead of blitting.
This commit is contained in:
Stenzek 2023-06-03 23:52:26 +10:00 committed by Connor McLaughlin
parent ce7c466041
commit 96fad124ac
4 changed files with 596 additions and 340 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

724
pcsx2/GS/Renderers/HW/GSRendererHW.cpp
View File

@ -911,6 +911,92 @@ GSVector4i GSRendererHW::GetSplitTextureShuffleDrawRect() const
return r.insert64<0>(0).ralign<Align_Outside>(frame_psm.pgs); return r.insert64<0>(0).ralign<Align_Outside>(frame_psm.pgs);
} }
GSVector4i GSRendererHW::GetDrawRectForPages(u32 bw, u32 psm, u32 num_pages)
{
const GSVector2i& pgs = GSLocalMemory::m_psm[psm].pgs;
const GSVector2i size = GSVector2i(static_cast<int>(bw) * pgs.x, static_cast<int>(num_pages / bw) * pgs.y);
return GSVector4i::loadh(size);
}
bool GSRendererHW::TryToResolveSinglePageFramebuffer(GIFRegFRAME& FRAME, bool only_next_draw)
{
const u32 start_bp = FRAME.Block();
u32 new_bw = FRAME.FBW;
u32 new_psm = FRAME.PSM;
pxAssert(new_bw <= 1);
if (m_backed_up_ctx >= 0)
{
const GSDrawingContext& next_ctx = m_env.CTXT[m_backed_up_ctx];
if (next_ctx.FRAME.FBW != new_bw)
{
// Using it as a target/Z next (Superman Returns).
if (start_bp == next_ctx.FRAME.Block())
{
GL_INS("TryToResolveSinglePageWidth(): Next FBP is split clear, using FBW of %u", next_ctx.FRAME.FBW);
new_bw = next_ctx.FRAME.FBW;
new_psm = next_ctx.FRAME.PSM;
}
else if (start_bp == next_ctx.ZBUF.Block())
{
GL_INS("TryToResolveSinglePageWidth(): Next ZBP is split clear, using FBW of %u", next_ctx.FRAME.FBW);
new_bw = next_ctx.FRAME.FBW;
}
}
// Might be using it as a texture next (NARC).
if (new_bw <= 1 && next_ctx.TEX0.TBP0 == start_bp && new_bw != next_ctx.TEX0.TBW)
{
GL_INS("TryToResolveSinglePageWidth(): Next texture is using split clear, using FBW of %u", next_ctx.TEX0.TBW);
new_bw = next_ctx.TEX0.TBW;
new_psm = next_ctx.TEX0.PSM;
}
}
if (!only_next_draw)
{
// Try for an exiting target at the start BP. (Tom & Jerry)
if (new_bw <= 1)
{
GSTextureCache::Target* tgt = g_texture_cache->GetTargetWithSharedBits(start_bp, new_psm);
if (!tgt)
{
// Try with Z or FRAME (whichever we're not using).
tgt = g_texture_cache->GetTargetWithSharedBits(start_bp, new_psm ^ 0x30);
}
if (tgt && ((start_bp + (m_split_clear_pages * BLOCKS_PER_PAGE)) - 1) <= tgt->m_end_block)
{
GL_INS("TryToResolveSinglePageWidth(): Using FBW of %u and PSM %s from existing target",
tgt->m_TEX0.PSM, psm_str(tgt->m_TEX0.PSM));
new_bw = tgt->m_TEX0.TBW;
new_psm = tgt->m_TEX0.PSM;
}
}
// Still bad FBW? Fall back to the resolution hack (Brave).
if (new_bw <= 1)
{
// Framebuffer is likely to be read as 16bit later, so we will need to double the width if the write is 32bit.
const bool double_width =
GSLocalMemory::m_psm[new_psm].bpp == 32 && PCRTCDisplays.GetFramebufferBitDepth() == 16;
const GSVector2i fb_size = PCRTCDisplays.GetFramebufferSize(-1);
u32 width =
std::ceil(static_cast<float>(m_split_clear_pages * GSLocalMemory::m_psm[new_psm].pgs.y) / fb_size.y) *
64;
width = std::max((width * (double_width ? 2 : 1)), static_cast<u32>(fb_size.x));
new_bw = (width + 63) / 64;
GL_INS("TryToResolveSinglePageWidth(): Fallback guess target FBW of %u", new_bw);
}
}
if (new_bw <= 1)
return false;
FRAME.FBW = new_bw;
FRAME.PSM = new_psm;
return true;
}
bool GSRendererHW::IsSplitClearActive() const bool GSRendererHW::IsSplitClearActive() const
{ {
return (m_split_clear_pages != 0); return (m_split_clear_pages != 0);
@ -918,9 +1004,12 @@ bool GSRendererHW::IsSplitClearActive() const
bool GSRendererHW::IsStartingSplitClear() bool GSRendererHW::IsStartingSplitClear()
{ {
// Mem clear conditions have already been checked by the caller, except for Z. // Shouldn't have gaps.
// We _could_ handle the split for both colour and depth, but nothing I've seen hits it yet. if (m_vt.m_eq.rgba != 0xFFFF || (!m_cached_ctx.ZBUF.ZMSK && !m_vt.m_eq.z) || !PrimitiveCoversWithoutGaps())
if (!m_cached_ctx.ZBUF.ZMSK) return false;
// Limit to only single page wide tall draws for now. Too many false positives otherwise (e.g. NFSU).
if (m_context->FRAME.FBW > 1 || m_r.height() < 1024)
return false; return false;
u32 pages_covered; u32 pages_covered;
@ -928,19 +1017,21 @@ bool GSRendererHW::IsStartingSplitClear()
return false; return false;
m_split_clear_start = m_cached_ctx.FRAME; m_split_clear_start = m_cached_ctx.FRAME;
m_split_clear_start_Z = m_cached_ctx.ZBUF;
m_split_clear_pages = pages_covered; m_split_clear_pages = pages_covered;
m_split_clear_color = m_vertex.buff[1].RGBAQ.U32[0]; m_split_clear_color = GetConstantDirectWriteMemClearColor();
if (PRIM->ABE && m_context->ALPHA.IsBlack())
m_split_clear_color &= ~0xFF000000;
GL_INS("Starting split clear at FBP %x FBW %u PSM %s with %dx%d rect covering %u pages", GL_INS("Starting split clear at FBP %x FBW %u PSM %s with %dx%d rect covering %u pages",
m_cached_ctx.FRAME.Block(), m_cached_ctx.FRAME.FBW, psm_str(m_cached_ctx.FRAME.PSM), m_cached_ctx.FRAME.Block(), m_cached_ctx.FRAME.FBW, psm_str(m_cached_ctx.FRAME.PSM),
m_r.width(), m_r.height(), pages_covered); m_r.width(), m_r.height(), pages_covered);
// Remove any targets which are directly at the start. // Remove any targets which are directly at the start.
if (IsDiscardingDstColor())
{
const u32 bp = m_cached_ctx.FRAME.Block(); const u32 bp = m_cached_ctx.FRAME.Block();
g_texture_cache->InvalidateVideoMemType(GSTextureCache::RenderTarget, bp); g_texture_cache->InvalidateVideoMemType(GSTextureCache::RenderTarget, bp);
g_texture_cache->InvalidateVideoMemType(GSTextureCache::DepthStencil, bp); g_texture_cache->InvalidateVideoMemType(GSTextureCache::DepthStencil, bp);
}
return true; return true;
} }
@ -951,18 +1042,17 @@ bool GSRendererHW::ContinueSplitClear()
if (!IsConstantDirectWriteMemClear()) if (!IsConstantDirectWriteMemClear())
return false; return false;
// Shouldn't be writing Z, in theory we could track this too and clear both though.
if (!m_cached_ctx.ZBUF.ZMSK)
return false;
// Shouldn't have gaps. // Shouldn't have gaps.
if (m_vt.m_eq.rgba != 0xFFFF || !PrimitiveCoversWithoutGaps()) if (m_vt.m_eq.rgba != 0xFFFF || (!m_cached_ctx.ZBUF.ZMSK && !m_vt.m_eq.z) || !PrimitiveCoversWithoutGaps())
return false; return false;
// Remove any targets which are directly at the start, since we checked this draw in the last. // Remove any targets which are directly at the start, since we checked this draw in the last.
if (IsDiscardingDstColor())
{
const u32 bp = m_cached_ctx.FRAME.Block(); const u32 bp = m_cached_ctx.FRAME.Block();
g_texture_cache->InvalidateVideoMemType(GSTextureCache::RenderTarget, bp); g_texture_cache->InvalidateVideoMemType(GSTextureCache::RenderTarget, bp);
g_texture_cache->InvalidateVideoMemType(GSTextureCache::DepthStencil, bp); g_texture_cache->InvalidateVideoMemType(GSTextureCache::DepthStencil, bp);
}
// Check next draw. // Check next draw.
u32 pages_covered; u32 pages_covered;
@ -989,30 +1079,45 @@ bool GSRendererHW::CheckNextDrawForSplitClear(const GSVector4i& r, u32* pages_co
// next FBP should point to the end of the rect // next FBP should point to the end of the rect
const GSDrawingContext& next_ctx = m_env.CTXT[m_backed_up_ctx]; const GSDrawingContext& next_ctx = m_env.CTXT[m_backed_up_ctx];
if (next_ctx.FRAME.Block() != ((end_block + 1) % MAX_BLOCKS) || next_ctx.FRAME.FBW != m_cached_ctx.FRAME.FBW || if (next_ctx.FRAME.Block() != ((end_block + 1) % MAX_BLOCKS) ||
next_ctx.FRAME.PSM != m_cached_ctx.FRAME.PSM) m_context->TEX0.U64 != next_ctx.TEX0.U64 ||
m_context->TEX1.U64 != next_ctx.TEX1.U64 || m_context->CLAMP.U64 != next_ctx.CLAMP.U64 ||
m_context->TEST.U64 != next_ctx.TEST.U64 || ((m_context->FRAME.U64 ^ next_ctx.FRAME.U64) & (~0x1FF)) != 0 ||
((m_context->ZBUF.U64 ^ next_ctx.ZBUF.U64) & (~0x1FF)) != 0)
{ {
return false; return false;
} }
// check ZBP if we're doing Z too
if (!m_cached_ctx.ZBUF.ZMSK && m_cached_ctx.FRAME.FBP != m_cached_ctx.ZBUF.ZBP)
{
const u32 end_z_block = GSLocalMemory::GetEndBlockAddress(
m_cached_ctx.ZBUF.Block(), m_cached_ctx.FRAME.FBW, m_cached_ctx.ZBUF.PSM, r);
if (next_ctx.ZBUF.Block() != ((end_z_block + 1) % MAX_BLOCKS))
return false;
}
return true; return true;
} }
void GSRendererHW::FinishSplitClear() void GSRendererHW::FinishSplitClear()
{ {
const GSLocalMemory::psm_t& psm_s = GSLocalMemory::m_psm[m_split_clear_start.PSM]; const u32 start_bp = m_split_clear_start.Block();
const GSOffset clear_off = GSOffset(psm_s.info, m_split_clear_start.Block(), m_split_clear_start.FBW, m_split_clear_start.PSM); GL_INS("FinishSplitClear(): Start %x FBW %u PSM %s, %u pages, %08X color", start_bp, m_split_clear_start.FBW,
const GSVector4i rect = GSVector4i(0, 0, m_split_clear_start.FBW * 64, (m_split_clear_pages * psm_s.pgs.y) / m_split_clear_start.FBW);
GL_INS("FinishSplitClear(): Start %x FBW %u PSM %s, %u pages, %08X color", m_split_clear_start.Block(), m_split_clear_start.FBW,
psm_str(m_split_clear_start.PSM), m_split_clear_pages, m_split_clear_color); psm_str(m_split_clear_start.PSM), m_split_clear_pages, m_split_clear_color);
OI_DoGsMemClear(clear_off, rect, m_split_clear_color); // If this was a tall single-page draw, try to get a better BW from somewhere.
if (m_split_clear_start.FBW <= 1 && m_split_clear_pages >= 16) // 1024 high
// Invalidate any targets in this range. TryToResolveSinglePageFramebuffer(m_split_clear_start, false);
g_texture_cache->InvalidateVideoMem(clear_off, rect, true);
SetNewFRAME(m_split_clear_start.Block(), m_split_clear_start.FBW, m_split_clear_start.PSM);
SetNewZBUF(m_split_clear_start_Z.Block(), m_split_clear_start_Z.PSM);
ReplaceVerticesWithSprite(
GetDrawRectForPages(m_split_clear_start.FBW, m_split_clear_start.PSM, m_split_clear_pages), GSVector2i(1, 1));
GL_INS("FinishSplitClear(): New draw rect is (%d,%d=>%d,%d) with FBW %u and PSM %s", m_r.x, m_r.y, m_r.z, m_r.w,
m_split_clear_start.FBW, psm_str(m_split_clear_start.PSM));
m_split_clear_start.U64 = 0; m_split_clear_start.U64 = 0;
m_split_clear_start_Z.U64 = 0;
m_split_clear_pages = 0; m_split_clear_pages = 0;
m_split_clear_color = 0; m_split_clear_color = 0;
} }
@ -1652,21 +1757,22 @@ void GSRendererHW::Draw()
// 3/ 50cents really draws (0,0,0,128) color and a (0) 24 bits depth // 3/ 50cents really draws (0,0,0,128) color and a (0) 24 bits depth
// Note: FF DoC has both buffer at same location but disable the depth test (write?) with ZTE = 0 // Note: FF DoC has both buffer at same location but disable the depth test (write?) with ZTE = 0
const u32 max_z = (0xFFFFFFFF >> (GSLocalMemory::m_psm[m_cached_ctx.ZBUF.PSM].fmt * 8)); const u32 max_z = (0xFFFFFFFF >> (GSLocalMemory::m_psm[m_cached_ctx.ZBUF.PSM].fmt * 8));
const bool no_rt = (context->ALPHA.IsCd() && PRIM->ABE && (m_cached_ctx.FRAME.PSM == 1)) bool no_rt = (context->ALPHA.IsCd() && PRIM->ABE && (m_cached_ctx.FRAME.PSM == 1))
|| (!m_cached_ctx.TEST.DATE && (m_cached_ctx.FRAME.FBMSK & GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].fmsk) == GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].fmsk); || (!m_cached_ctx.TEST.DATE && (m_cached_ctx.FRAME.FBMSK & GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].fmsk) == GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].fmsk);
const bool no_ds = ( const bool all_depth_tests_pass =
// Depth is always pass/fail (no read) and write are discarded. // Depth is always pass/fail (no read) and write are discarded.
(zm != 0 && (!m_cached_ctx.TEST.ZTE || m_cached_ctx.TEST.ZTST <= ZTST_ALWAYS)) || (!m_cached_ctx.TEST.ZTE || m_cached_ctx.TEST.ZTST <= ZTST_ALWAYS) ||
// Depth test will always pass // Depth test will always pass
(zm != 0 && m_cached_ctx.TEST.ZTST == ZTST_GEQUAL && m_vt.m_eq.z && std::min(m_vertex.buff[0].XYZ.Z, max_z) == max_z) || (m_cached_ctx.TEST.ZTST == ZTST_GEQUAL && m_vt.m_eq.z && std::min(m_vertex.buff[0].XYZ.Z, max_z) == max_z);
bool no_ds = (zm != 0 && all_depth_tests_pass) ||
// Depth will be written through the RT // Depth will be written through the RT
(!no_rt && m_cached_ctx.FRAME.FBP == m_cached_ctx.ZBUF.ZBP && !PRIM->TME && zm == 0 && (fm & fm_mask) == 0 && m_cached_ctx.TEST.ZTE)); (!no_rt && m_cached_ctx.FRAME.FBP == m_cached_ctx.ZBUF.ZBP && !PRIM->TME && zm == 0 && (fm & fm_mask) == 0 && m_cached_ctx.TEST.ZTE);
// No Z test if no z buffer. // No Z test if no z buffer.
if (no_ds) if (no_ds || all_depth_tests_pass)
{ {
if (m_cached_ctx.TEST.ZTST != ZTST_ALWAYS) if (m_cached_ctx.TEST.ZTST != ZTST_ALWAYS)
GL_CACHE("Disabling Z buffer because all tests will pass."); GL_CACHE("Disabling Z tests because all tests will pass.");
m_cached_ctx.TEST.ZTST = ZTST_ALWAYS; m_cached_ctx.TEST.ZTST = ZTST_ALWAYS;
} }
@ -1706,11 +1812,24 @@ void GSRendererHW::Draw()
m_r = m_r.blend8(m_r + GSVector4i::cxpr(0, 0, 1, 1), (m_r.xyxy() == m_r.zwzw())); m_r = m_r.blend8(m_r + GSVector4i::cxpr(0, 0, 1, 1), (m_r.xyxy() == m_r.zwzw()));
m_r = m_r.rintersect(context->scissor.in); m_r = m_r.rintersect(context->scissor.in);
// We want to fix up the context if we're doing a double half clear, regardless of whether we do the CPU fill.
const bool is_possible_mem_clear = IsConstantDirectWriteMemClear();
if (!GSConfig.UserHacks_DisableSafeFeatures && is_possible_mem_clear)
{
if (!DetectStripedDoubleClear(no_rt, no_ds))
DetectDoubleHalfClear(no_rt, no_ds);
}
const bool process_texture = PRIM->TME && !(PRIM->ABE && m_context->ALPHA.IsBlack() && !m_cached_ctx.TEX0.TCC); const bool process_texture = PRIM->TME && !(PRIM->ABE && m_context->ALPHA.IsBlack() && !m_cached_ctx.TEX0.TCC);
const u32 frame_end_bp = GSLocalMemory::GetEndBlockAddress(m_cached_ctx.FRAME.Block(), m_cached_ctx.FRAME.FBW, m_cached_ctx.FRAME.PSM, m_r); const u32 frame_end_bp = GSLocalMemory::GetEndBlockAddress(m_cached_ctx.FRAME.Block(), m_cached_ctx.FRAME.FBW, m_cached_ctx.FRAME.PSM, m_r);
bool preserve_rt_color =
((!no_rt && (!IsDiscardingDstColor() || !PrimitiveCoversWithoutGaps() || !all_depth_tests_pass)) || // Using Dst Color or draw has gaps
(process_texture && m_cached_ctx.TEX0.TBP0 >= m_cached_ctx.FRAME.Block() &&
m_cached_ctx.TEX0.TBP0 < frame_end_bp)); // Tex is RT
bool preserve_depth = preserve_rt_color || (!no_ds && (!all_depth_tests_pass || !m_cached_ctx.DepthWrite()));
// SW CLUT Render enable. // SW CLUT Render enable.
bool force_preload = GSConfig.PreloadFrameWithGSData; bool force_preload = GSConfig.PreloadFrameWithGSData;
bool preload_uploads = false;
if (GSConfig.UserHacks_CPUCLUTRender > 0 || GSConfig.UserHacks_GPUTargetCLUTMode != GSGPUTargetCLUTMode::Disabled) if (GSConfig.UserHacks_CPUCLUTRender > 0 || GSConfig.UserHacks_GPUTargetCLUTMode != GSGPUTargetCLUTMode::Disabled)
{ {
const CLUTDrawTestResult result = (GSConfig.UserHacks_CPUCLUTRender == 2) ? PossibleCLUTDrawAggressive() : PossibleCLUTDraw(); const CLUTDrawTestResult result = (GSConfig.UserHacks_CPUCLUTRender == 2) ? PossibleCLUTDrawAggressive() : PossibleCLUTDraw();
@ -1727,21 +1846,11 @@ void GSRendererHW::Draw()
{ {
// Force enable preloading if any of the existing data is needed. // Force enable preloading if any of the existing data is needed.
// e.g. NFSMW only writes the alpha channel, and needs the RGB preloaded. // e.g. NFSMW only writes the alpha channel, and needs the RGB preloaded.
if (((fm & fm_mask) != 0 && ((fm & fm_mask) != fm_mask)) || // Some channels masked force_preload |= preserve_rt_color;
!IsOpaque()) // Blending enabled if (preserve_rt_color)
{
GL_INS("Forcing preload due to partial/blended CLUT draw"); GL_INS("Forcing preload due to partial/blended CLUT draw");
force_preload = true;
} }
} }
}
if (((fm & fm_mask) != 0) || // Some channels masked
!IsDiscardingDstColor() || !PrimitiveCoversWithoutGaps() || // Using Dst Color or draw has gaps
(process_texture && m_cached_ctx.TEX0.TBP0 >= m_cached_ctx.FRAME.Block() && m_cached_ctx.TEX0.TBP0 < frame_end_bp)) // Tex is RT
{
preload_uploads = true;
}
if (!m_channel_shuffle && m_cached_ctx.FRAME.Block() == m_cached_ctx.TEX0.TBP0 && if (!m_channel_shuffle && m_cached_ctx.FRAME.Block() == m_cached_ctx.TEX0.TBP0 &&
IsPossibleChannelShuffle()) IsPossibleChannelShuffle())
@ -1785,18 +1894,15 @@ void GSRendererHW::Draw()
// Fudge FRAME and TEX0 to point to the start of the shuffle. // Fudge FRAME and TEX0 to point to the start of the shuffle.
m_cached_ctx.TEX0.TBP0 = m_split_texture_shuffle_start_TBP; m_cached_ctx.TEX0.TBP0 = m_split_texture_shuffle_start_TBP;
m_cached_ctx.FRAME.FBP = m_split_texture_shuffle_start_FBP; SetNewFRAME(m_split_texture_shuffle_start_FBP << 5, m_cached_ctx.FRAME.FBW, m_cached_ctx.FRAME.PSM);
m_context->offset.fb = GSOffset(GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].info, m_cached_ctx.FRAME.Block(),
m_cached_ctx.FRAME.FBW, m_cached_ctx.FRAME.PSM);
} }
const auto cleanup_draw = [&]() { const auto cleanup_draw = [&]() {
// Restore offsets. // Restore offsets.
if ((m_context->FRAME.U32[0] ^ m_cached_ctx.FRAME.U32[0]) & 0x3f3f01ff) if ((m_context->FRAME.U32[0] ^ m_cached_ctx.FRAME.U32[0]) & 0x3f3f01ff)
{ m_context->offset.fb = m_mem.GetOffset(m_context->FRAME.Block(), m_context->FRAME.FBW, m_context->FRAME.PSM);
m_context->offset.fb = GSOffset(GSLocalMemory::m_psm[m_context->FRAME.PSM].info, if ((m_context->ZBUF.U32[0] ^ m_cached_ctx.ZBUF.U32[0]) & 0x3f0001ff)
m_context->FRAME.Block(), m_context->FRAME.FBW, m_context->FRAME.PSM); m_context->offset.zb = m_mem.GetOffset(m_context->ZBUF.Block(), m_context->FRAME.FBW, m_context->ZBUF.PSM);
}
}; };
const auto cleanup_cancelled_draw = [&]() { const auto cleanup_cancelled_draw = [&]() {
@ -1805,83 +1911,69 @@ void GSRendererHW::Draw()
cleanup_draw(); cleanup_draw();
}; };
const bool is_possible_mem_clear = IsConstantDirectWriteMemClear(); if (!GSConfig.UserHacks_DisableSafeFeatures && is_possible_mem_clear)
if (!GSConfig.UserHacks_DisableSafeFeatures)
{ {
if (is_possible_mem_clear) GL_INS("WARNING: Possible mem clear.");
{
// Likely doing a huge single page width clear, which never goes well. (Superman)
// Burnout 3 does a 32x1024 double width clear on its reflection targets.
const bool clear_height_valid = (m_r.w >= 1024);
if (clear_height_valid && m_cached_ctx.FRAME.FBW == 1)
{
GSVector2i fb_size = PCRTCDisplays.GetFramebufferSize(-1);
u32 width = ceil(static_cast<float>(m_r.w) / fb_size.y) * 64;
// Framebuffer is likely to be read as 16bit later, so we will need to double the width if the write is 32bit.
const bool double_width = GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].bpp == 32 && PCRTCDisplays.GetFramebufferBitDepth() == 16;
m_r.x = 0;
m_r.y = 0;
m_r.w = fb_size.y;
m_r.z = std::max((width * (double_width ? 2 : 1)), static_cast<u32>(fb_size.x));
m_cached_ctx.FRAME.FBW = (m_r.z + 63) / 64;
m_context->scissor.in.z = m_cached_ctx.FRAME.FBW * 64;
GSVertex* s = &m_vertex.buff[0]; // We'll finish things off later.
s[0].XYZ.X = static_cast<u16>(m_context->XYOFFSET.OFX + 0); if (IsStartingSplitClear())
s[1].XYZ.X = static_cast<u16>(m_context->XYOFFSET.OFX + 16384); {
s[0].XYZ.Y = static_cast<u16>(m_context->XYOFFSET.OFY + 0); cleanup_draw();
s[1].XYZ.Y = static_cast<u16>(m_context->XYOFFSET.OFY + 16384); return;
m_vertex.head = m_vertex.tail = m_vertex.next = 2;
m_index.tail = 2;
} }
// Superman does a clear to white, not black, on its depth buffer. // Try to fix large single-page-wide draws.
// Since we don't preload depth, OI_GsMemClear() won't work here, since we invalidate the target later bool clear_height_valid = m_r.w >= 1024;
// on. So, instead, let the draw go through with the expanded rectangle, and copy color->depth. if (clear_height_valid && m_cached_ctx.FRAME.FBW <= 1 &&
const bool is_zero_clear = (((GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].fmt == 0) ? TryToResolveSinglePageFramebuffer(m_cached_ctx.FRAME, true))
m_vertex.buff[1].RGBAQ.U32[0] : {
(m_vertex.buff[1].RGBAQ.U32[0] & ~0xFF000000)) == 0) && m_cached_ctx.FRAME.FBMSK == 0 && IsDiscardingDstColor(); const GSVector2i& pgs = GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].pgs;
ReplaceVerticesWithSprite(
GetDrawRectForPages(m_cached_ctx.FRAME.FBW, m_cached_ctx.FRAME.PSM, (m_r.w + (pgs.y - 1)) / pgs.y),
GSVector2i(1, 1));
clear_height_valid = false;
}
const bool is_zero_clear = (GetConstantDirectWriteMemClearColor() == 0 && !preserve_rt_color);
const bool req_z = m_cached_ctx.FRAME.FBP != m_cached_ctx.ZBUF.ZBP && !m_cached_ctx.ZBUF.ZMSK; const bool req_z = m_cached_ctx.FRAME.FBP != m_cached_ctx.ZBUF.ZBP && !m_cached_ctx.ZBUF.ZMSK;
bool no_target_found = false; bool no_target_found = false;
// This is behind the if just to reduce lookups. // This is behind the if just to reduce lookups.
if (is_zero_clear && !clear_height_valid) if (is_zero_clear && !clear_height_valid)
{ {
const u32 target_end = GSLocalMemory::GetEndBlockAddress(m_cached_ctx.FRAME.Block(), m_cached_ctx.FRAME.FBW, m_cached_ctx.FRAME.PSM, m_r); const u32 fbw = m_cached_ctx.FRAME.FBW;
const u32 frame_start = m_cached_ctx.FRAME.Block();
no_target_found = !g_texture_cache->GetExactTarget(m_cached_ctx.FRAME.Block(), m_cached_ctx.FRAME.FBW, GSTextureCache::RenderTarget, target_end) && const u32 frame_end = GSLocalMemory::GetEndBlockAddress(frame_start, fbw, m_cached_ctx.FRAME.PSM, m_r);
!g_texture_cache->GetExactTarget(m_cached_ctx.FRAME.Block(), m_cached_ctx.FRAME.FBW, GSTextureCache::DepthStencil, target_end); no_target_found =
!g_texture_cache->GetExactTarget(frame_start, fbw, GSTextureCache::RenderTarget, frame_end) &&
!g_texture_cache->GetExactTarget(frame_start, fbw, GSTextureCache::DepthStencil, frame_end);
} }
if (is_zero_clear && OI_GsMemClear() && (clear_height_valid || (!req_z && no_target_found))) // If it's an invalid-sized draw, do the mem clear on the CPU, we don't want to create huge targets.
// If clearing to zero, don't bother creating the target. Games tend to clear more than they use, wasting VRAM/bandwidth.
if ((is_zero_clear || clear_height_valid) && TryGSMemClear(no_rt, no_ds) &&
(clear_height_valid || (!req_z && no_target_found)))
{ {
GL_INS("Clear draw with mem clear and valid clear height, invalidating."); GL_INS("Skipping (%d,%d=>%d,%d) draw at FBP %x/ZBP %x due to invalid height or zero clear.", m_r.x, m_r.y,
m_r.z, m_r.w, m_cached_ctx.FRAME.Block(), m_cached_ctx.ZBUF.Block());
g_texture_cache->InvalidateVideoMem(context->offset.fb, m_r, true); if (!no_rt)
{
g_texture_cache->InvalidateVideoMemType(GSTextureCache::RenderTarget, m_cached_ctx.FRAME.Block()); g_texture_cache->InvalidateVideoMemType(GSTextureCache::RenderTarget, m_cached_ctx.FRAME.Block());
if(no_target_found)
g_texture_cache->InvalidateVideoMemType(GSTextureCache::DepthStencil, m_cached_ctx.FRAME.Block()); g_texture_cache->InvalidateVideoMemType(GSTextureCache::DepthStencil, m_cached_ctx.FRAME.Block());
g_texture_cache->InvalidateVideoMem(m_context->offset.fb, m_r, true);
}
if (m_cached_ctx.ZBUF.ZMSK == 0) if (!no_ds && m_cached_ctx.ZBUF.ZMSK == 0)
{ {
g_texture_cache->InvalidateVideoMem(context->offset.zb, m_r, false); g_texture_cache->InvalidateVideoMemType(GSTextureCache::RenderTarget, m_cached_ctx.ZBUF.Block());
g_texture_cache->InvalidateVideoMemType(GSTextureCache::DepthStencil, m_cached_ctx.ZBUF.Block()); g_texture_cache->InvalidateVideoMemType(GSTextureCache::DepthStencil, m_cached_ctx.ZBUF.Block());
g_texture_cache->InvalidateVideoMem(m_context->offset.zb, m_r, true);
} }
cleanup_draw(); cleanup_draw();
return; return;
} }
else if (!clear_height_valid && is_zero_clear && IsStartingSplitClear())
{
// Currently we only allow this for zero, because what we clear won't get preloaded back to the RT.
// But, if we did, then we could allow it for non-zero clears, also same for the above case.
cleanup_draw();
return;
}
}
} }
GIFRegTEX0 TEX0 = {}; GIFRegTEX0 TEX0 = {};
@ -2019,9 +2111,6 @@ void GSRendererHW::Draw()
// Ensure draw rect is clamped to framebuffer size. Necessary for updating valid area. // Ensure draw rect is clamped to framebuffer size. Necessary for updating valid area.
m_r = m_r.rintersect(t_size_rect); m_r = m_r.rintersect(t_size_rect);
// Ensure areas not drawn to are filled in by preloads. Test case: Okami
preload_uploads |= !m_r.eq(t_size_rect);
float target_scale = GetTextureScaleFactor(); float target_scale = GetTextureScaleFactor();
// This upscaling hack is for games which construct P8 textures by drawing a bunch of small sprites in C32, // This upscaling hack is for games which construct P8 textures by drawing a bunch of small sprites in C32,
@ -2050,7 +2139,7 @@ void GSRendererHW::Draw()
const bool is_square = (t_size.y == t_size.x) && m_r.w >= 1023 && PrimitiveCoversWithoutGaps(); const bool is_square = (t_size.y == t_size.x) && m_r.w >= 1023 && PrimitiveCoversWithoutGaps();
const bool is_clear = is_possible_mem_clear && is_square; const bool is_clear = is_possible_mem_clear && is_square;
rt = g_texture_cache->LookupTarget(FRAME_TEX0, t_size, target_scale, GSTextureCache::RenderTarget, true, rt = g_texture_cache->LookupTarget(FRAME_TEX0, t_size, target_scale, GSTextureCache::RenderTarget, true,
fm, false, force_preload, preload_uploads); fm, false, force_preload, preserve_rt_color, m_r);
// Draw skipped because it was a clear and there was no target. // Draw skipped because it was a clear and there was no target.
if (!rt) if (!rt)
@ -2059,12 +2148,12 @@ void GSRendererHW::Draw()
{ {
GL_INS("Clear draw with no target, skipping."); GL_INS("Clear draw with no target, skipping.");
cleanup_cancelled_draw(); cleanup_cancelled_draw();
OI_GsMemClear(); TryGSMemClear(no_rt, no_ds);
return; return;
} }
rt = g_texture_cache->CreateTarget(FRAME_TEX0, t_size, target_scale, GSTextureCache::RenderTarget, true, rt = g_texture_cache->CreateTarget(FRAME_TEX0, t_size, target_scale, GSTextureCache::RenderTarget, true,
fm, false, force_preload, preload_uploads); fm, false, force_preload, preserve_rt_color, m_r);
if (unlikely(!rt)) if (unlikely(!rt))
{ {
GL_INS("ERROR: Failed to create FRAME target, skipping."); GL_INS("ERROR: Failed to create FRAME target, skipping.");
@ -2084,11 +2173,11 @@ void GSRendererHW::Draw()
ZBUF_TEX0.PSM = m_cached_ctx.ZBUF.PSM; ZBUF_TEX0.PSM = m_cached_ctx.ZBUF.PSM;
ds = g_texture_cache->LookupTarget(ZBUF_TEX0, t_size, target_scale, GSTextureCache::DepthStencil, ds = g_texture_cache->LookupTarget(ZBUF_TEX0, t_size, target_scale, GSTextureCache::DepthStencil,
m_cached_ctx.DepthWrite(), 0, false, force_preload); m_cached_ctx.DepthWrite(), 0, false, force_preload, preserve_depth, m_r);
if (!ds) if (!ds)
{ {
ds = g_texture_cache->CreateTarget(ZBUF_TEX0, t_size, target_scale, GSTextureCache::DepthStencil, ds = g_texture_cache->CreateTarget(ZBUF_TEX0, t_size, target_scale, GSTextureCache::DepthStencil,
m_cached_ctx.DepthWrite(), 0, false, force_preload); m_cached_ctx.DepthWrite(), 0, false, force_preload, preserve_depth, m_r);
if (unlikely(!ds)) if (unlikely(!ds))
{ {
GL_INS("ERROR: Failed to create ZBUF target, skipping."); GL_INS("ERROR: Failed to create ZBUF target, skipping.");
@ -2325,6 +2414,13 @@ void GSRendererHW::Draw()
const int new_w = std::max(t_size.x, std::max(rt ? rt->m_unscaled_size.x : 0, ds ? ds->m_unscaled_size.x : 0)); const int new_w = std::max(t_size.x, std::max(rt ? rt->m_unscaled_size.x : 0, ds ? ds->m_unscaled_size.x : 0));
const int new_h = std::max(t_size.y, std::max(rt ? rt->m_unscaled_size.y : 0, ds ? ds->m_unscaled_size.y : 0)); const int new_h = std::max(t_size.y, std::max(rt ? rt->m_unscaled_size.y : 0, ds ? ds->m_unscaled_size.y : 0));
// Expanding the target means we can't fast clear anymore.
const bool growing = (new_w > m_r.z || new_h > m_r.w);
if ((!preserve_rt_color || !preserve_depth) && growing)
GL_INS("Preserving target due to draw being %dx%d and target being %dx%d", m_r.z, m_r.w, new_w, new_h);
preserve_rt_color |= growing;
preserve_depth |= growing;
if (rt) if (rt)
{ {
const u32 old_end_block = rt->m_end_block; const u32 old_end_block = rt->m_end_block;
@ -2465,8 +2561,9 @@ void GSRendererHW::Draw()
return; return;
} }
if (!GSConfig.UserHacks_DisableSafeFeatures && is_possible_mem_clear && IsDiscardingDstColor()) bool skip_draw = false;
OI_DoubleHalfClear(rt, ds); if (!GSConfig.UserHacks_DisableSafeFeatures && is_possible_mem_clear)
skip_draw = TryTargetClear(rt, ds, preserve_rt_color, preserve_depth);
// A couple of hack to avoid upscaling issue. So far it seems to impacts mostly sprite // A couple of hack to avoid upscaling issue. So far it seems to impacts mostly sprite
// Note: first hack corrects both position and texture coordinate // Note: first hack corrects both position and texture coordinate
@ -2519,6 +2616,7 @@ void GSRendererHW::Draw()
// //
if (!skip_draw)
DrawPrims(rt, ds, src, tmm); DrawPrims(rt, ds, src, tmm);
// //
@ -5124,192 +5222,214 @@ bool GSRendererHW::CanUseSwPrimRender(bool no_rt, bool no_ds, bool draw_sprite_t
return true; return true;
} }
// Trick to do a fast clear on the GS void GSRendererHW::SetNewFRAME(u32 bp, u32 bw, u32 psm)
// Set frame buffer pointer on the start of the buffer. Set depth buffer pointer on the half buffer
// FB + depth write will fill the full buffer.
void GSRendererHW::OI_DoubleHalfClear(GSTextureCache::Target*& rt, GSTextureCache::Target*& ds)
{ {
// Note gs mem clear must be tested before calling this function m_cached_ctx.FRAME.FBP = bp >> 5;
m_cached_ctx.FRAME.FBW = bw;
m_cached_ctx.FRAME.PSM = psm;
m_context->offset.fb = m_mem.GetOffset(bp, bw, psm);
}
// If the mask is on any of the colour or partial alpha, keep the whole channel void GSRendererHW::SetNewZBUF(u32 bp, u32 psm)
const bool keep_a = (m_cached_ctx.FRAME.FBMSK >> 24) & 0xFF;
const bool keep_b = (m_cached_ctx.FRAME.FBMSK >> 16) & 0xFF;
const bool keep_g = (m_cached_ctx.FRAME.FBMSK >> 8) & 0xFF;
const bool keep_r = m_cached_ctx.FRAME.FBMSK & 0xFF;
// Limit further to unmask Z write
if (!m_cached_ctx.ZBUF.ZMSK && rt && ds)
{ {
const GSVertex* v = &m_vertex.buff[0]; m_cached_ctx.ZBUF.ZBP = bp >> 5;
const GSLocalMemory::psm_t& frame_psm = GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM]; m_cached_ctx.ZBUF.PSM = psm;
//const GSLocalMemory::psm_t& depth_psm = GSLocalMemory::m_psm[m_ZBUF.PSM]; m_context->offset.zb = m_mem.GetOffset(bp, m_cached_ctx.FRAME.FBW, psm);
}
bool GSRendererHW::DetectStripedDoubleClear(bool& no_rt, bool& no_ds)
{
const bool ZisFrame = m_cached_ctx.FRAME.FBP == m_cached_ctx.ZBUF.ZBP && !m_cached_ctx.ZBUF.ZMSK &&
(m_cached_ctx.FRAME.PSM & 0x30) != (m_cached_ctx.ZBUF.PSM & 0x30) &&
(m_cached_ctx.FRAME.PSM & 0xF) == (m_cached_ctx.ZBUF.PSM & 0xF) && m_vt.m_eq.z == 1 &&
m_vertex.buff[1].XYZ.Z == m_vertex.buff[1].RGBAQ.U32[0];
// Z and color must be constant and the same // Z and color must be constant and the same
if (m_vt.m_eq.rgba != 0xFFFF || !m_vt.m_eq.z || v[1].XYZ.Z != v[1].RGBAQ.U32[0]) if (!ZisFrame || m_vt.m_eq.rgba != 0xFFFF)
return; return false;
// Format doesn't have the same size. It smells fishy (xmen...) // Half a page extra width is written through Z.
//if (frame_psm.trbpp != depth_psm.trbpp) m_r.z += 32;
// return; m_context->scissor.in = m_r;
GL_INS("DetectStripedDoubleClear(): %d,%d => %d,%d @ FBP %x FBW %u ZBP %x", m_r.x, m_r.y, m_r.z, m_r.w,
m_cached_ctx.FRAME.Block(), m_cached_ctx.FRAME.FBW, m_cached_ctx.ZBUF.Block());
// And replace the vertex with a fullscreen quad.
GSVertex* const v = &m_vertex.buff[0];
const GSVector4i fpr = m_r.sll32(4);
v[0].XYZ.X = static_cast<u16>(m_context->XYOFFSET.OFX + fpr.x);
v[0].XYZ.Y = static_cast<u16>(m_context->XYOFFSET.OFY + fpr.y);
v[1].XYZ.X = static_cast<u16>(m_context->XYOFFSET.OFX + fpr.z);
v[1].XYZ.Y = static_cast<u16>(m_context->XYOFFSET.OFY + fpr.w);
m_vertex.head = m_vertex.tail = m_vertex.next = 2;
m_index.tail = 2;
m_primitive_covers_without_gaps = true;
// Remove Z, we'll write it through colour.
m_cached_ctx.ZBUF.ZMSK = true;
no_rt = false;
no_ds = true;
return true;
}
bool GSRendererHW::DetectDoubleHalfClear(bool& no_rt, bool& no_ds)
{
if (m_cached_ctx.FRAME.FBMSK != 0 || m_cached_ctx.TEST.ZTST != ZTST_ALWAYS || m_cached_ctx.ZBUF.ZMSK)
return false;
// Z and color must be constant and the same
// TODO: move covers check to caller
GSVertex* v = &m_vertex.buff[0];
if (m_vt.m_eq.rgba != 0xFFFF || !m_vt.m_eq.z || v[1].XYZ.Z != v[1].RGBAQ.U32[0])
return false;
// Frame and depth pointer can be inverted
const bool clear_depth = (m_cached_ctx.FRAME.FBP > m_cached_ctx.ZBUF.ZBP);
const u32 base = clear_depth ? m_cached_ctx.ZBUF.ZBP : m_cached_ctx.FRAME.FBP;
const u32 half = clear_depth ? m_cached_ctx.FRAME.FBP : m_cached_ctx.ZBUF.ZBP;
// Size of the current draw // Size of the current draw
const GSLocalMemory::psm_t& frame_psm = GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM];
const u32 w_pages = static_cast<u32>(roundf(m_vt.m_max.p.x / frame_psm.pgs.x)); const u32 w_pages = static_cast<u32>(roundf(m_vt.m_max.p.x / frame_psm.pgs.x));
const u32 h_pages = static_cast<u32>(roundf(m_vt.m_max.p.y / frame_psm.pgs.y)); const u32 h_pages = static_cast<u32>(roundf(m_vt.m_max.p.y / frame_psm.pgs.y));
const u32 written_pages = w_pages * h_pages; const u32 written_pages = w_pages * h_pages;
// Frame and depth pointer can be inverted
u32 base = 0, half = 0;
if (m_cached_ctx.FRAME.FBP > m_cached_ctx.ZBUF.ZBP)
{
base = m_cached_ctx.ZBUF.ZBP;
half = m_cached_ctx.FRAME.FBP;
}
else
{
base = m_cached_ctx.FRAME.FBP;
half = m_cached_ctx.ZBUF.ZBP;
}
// If both buffers are side by side we can expect a fast clear in on-going // If both buffers are side by side we can expect a fast clear in on-going
if (half <= (base + written_pages)) if (half != (base + written_pages))
{
// Take the vertex colour, but check if the blending would make it black.
u32 vert_color = v[1].RGBAQ.U32[0];
if (PRIM->ABE && m_context->ALPHA.IsBlack())
vert_color &= 0xFF000000;
const u32 color = vert_color;
const bool clear_depth = (m_cached_ctx.FRAME.FBP > m_cached_ctx.ZBUF.ZBP);
GL_INS("OI_DoubleHalfClear:%s: base %x half %x. w_pages %d h_pages %d fbw %d. Color %x",
clear_depth ? "depth" : "target", base << 5, half << 5, w_pages, h_pages, m_cached_ctx.FRAME.FBW, color);
// If some of the channels are masked, we need to keep them.
if (!clear_depth && m_cached_ctx.FRAME.FBMSK != 0)
{
GSTextureCache::Target* target = clear_depth ? ds : rt;
GSTexture* tex = g_gs_device->CreateRenderTarget(target->m_texture->GetWidth(), target->m_texture->GetHeight(), target->m_texture->GetFormat(), false);
if (!tex)
return;
if (clear_depth)
{
// Only pure clear are supported for depth
ASSERT(color == 0);
g_gs_device->ClearDepth(tex, 0.0f);
}
else
{
g_gs_device->ClearRenderTarget(tex, color);
}
const GSVector4 drect = GSVector4(target->GetUnscaledRect()) * GSVector4(target->m_scale);
// Copy channels being masked.
g_gs_device->StretchRect(target->m_texture, GSVector4(0.0f,0.0f,1.0f,1.0f), tex, drect, keep_r, keep_g, keep_b, keep_a);
g_perfmon.Put(GSPerfMon::TextureCopies, 1);
delete target->m_texture;
target->m_texture = tex;
target->UpdateValidity(target->GetUnscaledRect());
}
else
{
if (clear_depth)
{
// Only pure clear are supported for depth
ASSERT(color == 0);
g_gs_device->ClearDepth(ds->m_texture, 0.0f);
ds->UpdateValidity(ds->GetUnscaledRect());
}
else
{
g_gs_device->ClearRenderTarget(rt->m_texture, color);
rt->UpdateValidity(rt->GetUnscaledRect());
}
}
}
}
// Striped double clear done by Powerdrome and Snoopy Vs Red Baron, it will clear in 32 pixel stripes half done by the Z and half done by the FRAME
else if (rt && !ds && m_cached_ctx.FRAME.FBP == m_cached_ctx.ZBUF.ZBP && (m_cached_ctx.FRAME.PSM & 0x30) != (m_cached_ctx.ZBUF.PSM & 0x30)
&& (m_cached_ctx.FRAME.PSM & 0xF) == (m_cached_ctx.ZBUF.PSM & 0xF) && m_vt.m_eq.z == 1)
{
const GSVertex* v = &m_vertex.buff[0];
// Z and color must be constant and the same
if (m_vt.m_eq.rgba != 0xFFFF || !m_vt.m_eq.z || v[1].XYZ.Z != v[1].RGBAQ.U32[0])
return;
// If both buffers are side by side we can expect a fast clear in on-going
const u32 color = v[1].RGBAQ.U32[0];
// If some of the channels are masked, we need to keep them.
if (m_cached_ctx.FRAME.FBMSK != 0)
{
GSTextureCache::Target* target = rt;
GSTexture* tex = g_gs_device->CreateRenderTarget(target->m_texture->GetWidth(), target->m_texture->GetHeight(), target->m_texture->GetFormat(), true);
if (!tex)
return;
g_gs_device->ClearRenderTarget(tex, color);
const GSVector4 drect = GSVector4(target->GetUnscaledRect()) * GSVector4(target->m_scale);
// Copy channels being masked.
g_gs_device->StretchRect(target->m_texture, GSVector4(0, 0, 1, 1), tex, drect, keep_r, keep_g, keep_b, keep_a);
g_perfmon.Put(GSPerfMon::TextureCopies, 1);
delete target->m_texture;
target->m_texture = tex;
target->UpdateValidity(target->GetUnscaledRect());
}
else
{
g_gs_device->ClearRenderTarget(rt->m_texture, color);
rt->UpdateValidity(rt->GetUnscaledRect());
}
}
}
// Note: hack is safe, but it could impact the perf a little (normally games do only a couple of clear by frame)
bool GSRendererHW::OI_GsMemClear()
{
// Note gs mem clear must be tested before calling this function
// Striped double clear done by Powerdrome and Snoopy Vs Red Baron, it will clear in 32 pixel stripes half done by the Z and half done by the FRAME
const bool ZisFrame = m_cached_ctx.FRAME.FBP == m_cached_ctx.ZBUF.ZBP && !m_cached_ctx.ZBUF.ZMSK && (m_cached_ctx.FRAME.PSM & 0x30) != (m_cached_ctx.ZBUF.PSM & 0x30)
&& (m_cached_ctx.FRAME.PSM & 0xF) == (m_cached_ctx.ZBUF.PSM & 0xF) && m_vt.m_eq.z == 1 && m_vertex.buff[1].XYZ.Z == m_vertex.buff[1].RGBAQ.U32[0];
// Limit it further to a full screen 0 write
if ((PrimitiveCoversWithoutGaps() || ZisFrame) && m_vt.m_eq.rgba == 0xFFFF)
{
const GSOffset& off = m_context->offset.fb;
GSVector4i r = GSVector4i(m_vt.m_min.p.upld(m_vt.m_max.p)).rintersect(m_context->scissor.in);
if (r.width() == 32 && ZisFrame)
r.z += 32;
// Limit the hack to a single full buffer clear. Some games might use severals column to clear a screen
// but hopefully it will be enough.
if (m_r.width() < ((static_cast<int>(m_cached_ctx.FRAME.FBW) - 1) * 64) || r.height() <= 128)
return false; return false;
// Take the vertex colour, but check if the blending would make it black. // Try peeking ahead to confirm whether this is a "normal" clear, where the two buffers just happen to be
u32 vert_color = m_vertex.buff[1].RGBAQ.U32[0]; // bang up next to each other, or a double half clear. The two are really difficult to differentiate.
if (PRIM->ABE && m_context->ALPHA.IsBlack()) if (m_backed_up_ctx >= 0 && m_env.CTXT[m_backed_up_ctx].FRAME.FBW == m_cached_ctx.FRAME.FBW &&
vert_color &= 0xFF000000; ((m_env.CTXT[m_backed_up_ctx].FRAME.FBP == base && m_env.CTXT[m_backed_up_ctx].ZBUF.ZBP != half) ||
(m_env.CTXT[m_backed_up_ctx].ZBUF.ZBP == base && m_env.CTXT[m_backed_up_ctx].FRAME.FBP != half)))
{
// Needed for Spider-Man 2 (target was previously half size, double half cleared at new size).
GL_INS("Confirmed double-half clear by next FBP/ZBP");
}
else
{
// Check for a target matching the starting point. It might be in Z or FRAME...
GSTextureCache::Target* tgt = g_texture_cache->GetTargetWithSharedBits(
base << 5, clear_depth ? m_cached_ctx.ZBUF.PSM : m_cached_ctx.FRAME.PSM);
if (!tgt)
{
tgt = g_texture_cache->GetTargetWithSharedBits(
base << 5, clear_depth ? m_cached_ctx.FRAME.PSM : m_cached_ctx.ZBUF.PSM);
}
// Are we clearing over the middle of this target?
if (!tgt || (((half + written_pages) << 5) - 1) > tgt->m_end_block)
return false;
}
GL_INS("DetectDoubleHalfClear(): Clearing %s, fbp=%x, zbp=%x, pages=%u, base=%x, half=%x, rect=(%d,%d=>%d,%d)",
clear_depth ? "depth" : "color", m_cached_ctx.FRAME.Block(), m_cached_ctx.ZBUF.Block(), written_pages,
base << 5, half << 5, m_r.x, m_r.y, m_r.z, m_r.w);
// Double the clear rect.
m_r.w += m_r.y + m_r.height();
ReplaceVerticesWithSprite(m_r, GSVector2i(1, 1));
// Prevent wasting time looking up and creating the target which is getting blown away.
if (!clear_depth)
{
SetNewFRAME(base << 5, m_cached_ctx.FRAME.FBW, m_cached_ctx.FRAME.PSM);
m_cached_ctx.ZBUF.ZMSK = true;
no_rt = false;
no_ds = true;
}
else
{
SetNewZBUF(base << 5, m_cached_ctx.ZBUF.PSM);
m_cached_ctx.FRAME.FBMSK = 0xFFFFFFFF;
no_rt = true;
no_ds = false;
}
// Remove any targets at the half-buffer point, they're getting overwritten.
g_texture_cache->InvalidateVideoMemType(GSTextureCache::RenderTarget, half);
g_texture_cache->InvalidateVideoMemType(GSTextureCache::DepthStencil, half);
// TODO HACK: REMOVE ME
m_primitive_covers_without_gaps = true;
OI_DoGsMemClear(off, r, vert_color);
return true; return true;
} }
bool GSRendererHW::TryTargetClear(GSTextureCache::Target* rt, GSTextureCache::Target* ds, bool preserve_rt_color, bool preserve_depth)
{
if (m_vt.m_eq.rgba != 0xFFFF || !m_vt.m_eq.z)
return false;
bool skip = true;
if (rt)
{
if (!preserve_rt_color && !IsReallyDithered() && m_r.rintersect(rt->m_valid).eq(rt->m_valid))
{
const u32 c = GetConstantDirectWriteMemClearColor();
GL_INS("TryTargetClear(): RT at %x <= %08X", rt->m_TEX0.TBP0, c);
g_gs_device->ClearRenderTarget(rt->m_texture, c);
}
else
{
skip = false;
}
}
if (ds)
{
if (ds && !preserve_depth && m_r.rintersect(ds->m_valid).eq(ds->m_valid))
{
const u32 max_z = 0xFFFFFFFF >> (GSLocalMemory::m_psm[m_cached_ctx.ZBUF.PSM].fmt * 8);
const u32 z = std::min(max_z, m_vertex.buff[1].XYZ.Z);
const float d = static_cast<float>(z) * (g_gs_device->Features().clip_control ? 0x1p-32f : 0x1p-24f);
GL_INS("TryTargetClear(): DS at %x <= %f", ds->m_TEX0.TBP0, d);
g_gs_device->ClearDepth(ds->m_texture, d);
}
else
{
skip = false;
}
}
return skip;
}
bool GSRendererHW::TryGSMemClear(bool no_rt, bool no_ds)
{
if (!PrimitiveCoversWithoutGaps())
return false;
// Limit the hack to a single full buffer clear. Some games might use severals column to clear a screen
// but hopefully it will be enough.
if (m_r.width() < ((static_cast<int>(m_cached_ctx.FRAME.FBW) - 1) * 64))
return false;
// Don't mem clear one of frame or z, only do both.
const u32 fbmsk = (m_cached_ctx.FRAME.FBMSK & GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].fmsk);
if ((!no_rt && (fbmsk != 0 || m_vt.m_eq.rgba != 0xFFFF)) ||
(!no_ds && (m_cached_ctx.ZBUF.ZMSK != 0 || !m_vt.m_eq.z)))
{
return false; return false;
} }
void GSRendererHW::OI_DoGsMemClear(const GSOffset& off, const GSVector4i& r, u32 vert_color) if (!no_rt)
ClearGSLocalMemory(m_context->offset.fb, m_r, GetConstantDirectWriteMemClearColor());
if (!no_ds)
ClearGSLocalMemory(m_context->offset.zb, m_r, m_vertex.buff[1].XYZ.Z);
return true;
}
void GSRendererHW::ClearGSLocalMemory(const GSOffset& off, const GSVector4i& r, u32 vert_color)
{ {
GL_INS("OI_DoGsMemClear (%d,%d => %d,%d)", r.x, r.y, r.z, r.w); GL_INS(
"ClearGSLocalMemory(): %08X %d,%d => %d,%d @ BP %x BW %u", vert_color, r.x, r.y, r.z, r.w, off.bp(), off.bw());
const int format = GSLocalMemory::m_psm[off.psm()].fmt; const int format = GSLocalMemory::m_psm[off.psm()].fmt;
const u32 color = (format == 0) ? vert_color : (vert_color & ~0xFF000000);
const int left = r.left; const int left = r.left;
const int right = r.right; const int right = r.right;
@ -5330,7 +5450,7 @@ void GSRendererHW::OI_DoGsMemClear(const GSOffset& off, const GSVector4i& r, u32
if (format == 0) if (format == 0)
{ {
const GSVector4i vcolor = GSVector4i(color); const GSVector4i vcolor = GSVector4i(vert_color);
const u32 iterations_per_page = (pages_wide * pixels_per_page) / 4; const u32 iterations_per_page = (pages_wide * pixels_per_page) / 4;
pxAssert((off.bp() & (BLOCKS_PER_PAGE - 1)) == 0); pxAssert((off.bp() & (BLOCKS_PER_PAGE - 1)) == 0);
for (u32 current_page = off.bp() >> 5; top < page_aligned_bottom; top += pgs.y, current_page += fbw) for (u32 current_page = off.bp() >> 5; top < page_aligned_bottom; top += pgs.y, current_page += fbw)
@ -5344,7 +5464,7 @@ void GSRendererHW::OI_DoGsMemClear(const GSOffset& off, const GSVector4i& r, u32
else if (format == 1) else if (format == 1)
{ {
const GSVector4i mask = GSVector4i::xff000000(); const GSVector4i mask = GSVector4i::xff000000();
const GSVector4i vcolor = GSVector4i(color & 0x00ffffffu); const GSVector4i vcolor = GSVector4i(vert_color & 0x00ffffffu);
const u32 iterations_per_page = (pages_wide * pixels_per_page) / 4; const u32 iterations_per_page = (pages_wide * pixels_per_page) / 4;
pxAssert((off.bp() & (BLOCKS_PER_PAGE - 1)) == 0); pxAssert((off.bp() & (BLOCKS_PER_PAGE - 1)) == 0);
for (u32 current_page = off.bp() >> 5; top < page_aligned_bottom; top += pgs.y, current_page += fbw) for (u32 current_page = off.bp() >> 5; top < page_aligned_bottom; top += pgs.y, current_page += fbw)
@ -5383,23 +5503,19 @@ void GSRendererHW::OI_DoGsMemClear(const GSOffset& off, const GSVector4i& r, u32
auto pa = off.assertSizesMatch(GSLocalMemory::swizzle32).paMulti(m_mem.vm32(), 0, y); auto pa = off.assertSizesMatch(GSLocalMemory::swizzle32).paMulti(m_mem.vm32(), 0, y);
for (int x = left; x < right; x++) for (int x = left; x < right; x++)
{ *pa.value(x) = vert_color;
*pa.value(x) = color; // Here the constant color
}
} }
} }
else if (format == 1) else if (format == 1)
{ {
// Based on WritePixel24 // Based on WritePixel24
const u32 write_color = vert_color & 0xffffffu;
for (int y = top; y < bottom; y++) for (int y = top; y < bottom; y++)
{ {
auto pa = off.assertSizesMatch(GSLocalMemory::swizzle32).paMulti(m_mem.vm32(), 0, y); auto pa = off.assertSizesMatch(GSLocalMemory::swizzle32).paMulti(m_mem.vm32(), 0, y);
for (int x = left; x < right; x++) for (int x = left; x < right; x++)
{ *pa.value(x) = (*pa.value(x) & 0xff000000u) | write_color;
*pa.value(x) &= 0xff000000; // Clear the color
*pa.value(x) |= color; // OR in our constant
}
} }
} }
else if (format == 2) else if (format == 2)
@ -5412,9 +5528,7 @@ void GSRendererHW::OI_DoGsMemClear(const GSOffset& off, const GSVector4i& r, u32
auto pa = off.assertSizesMatch(GSLocalMemory::swizzle16).paMulti(m_mem.vm16(), 0, y); auto pa = off.assertSizesMatch(GSLocalMemory::swizzle16).paMulti(m_mem.vm16(), 0, y);
for (int x = left; x < right; x++) for (int x = left; x < right; x++)
{ *pa.value(x) = converted_color;
*pa.value(x) = converted_color; // Here the constant color
}
} }
} }
} }
@ -5483,8 +5597,11 @@ bool GSRendererHW::OI_BlitFMV(GSTextureCache::Target* _rt, GSTextureCache::Sourc
bool GSRendererHW::IsDiscardingDstColor() bool GSRendererHW::IsDiscardingDstColor()
{ {
return (!PRIM->ABE || IsOpaque() || m_context->ALPHA.IsBlack()) && !m_cached_ctx.TEST.ATE && return (
!m_cached_ctx.TEST.DATE; (!PRIM->ABE || IsOpaque() || m_context->ALPHA.IsBlack()) && // no blending or writing black
!m_cached_ctx.TEST.ATE && // not testing alpha (might discard some pixels)
!m_cached_ctx.TEST.DATE && // not reading alpha
(m_cached_ctx.FRAME.FBMSK & GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].fmsk) == 0); // no channels masked
} }
bool GSRendererHW::PrimitiveCoversWithoutGaps() bool GSRendererHW::PrimitiveCoversWithoutGaps()
@ -5585,6 +5702,41 @@ bool GSRendererHW::IsConstantDirectWriteMemClear()
return false; return false;
} }
u32 GSRendererHW::GetConstantDirectWriteMemClearColor() const
{
// Take the vertex colour, but check if the blending would make it black.
u32 vert_color = m_vertex.buff[1].RGBAQ.U32[0];
if (PRIM->ABE && m_context->ALPHA.IsBlack())
vert_color &= 0xFF000000u;
// 24-bit format? Otherwise, FBA sets the high bit in alpha.
const u32 cfmt = GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].fmt;
if (cfmt == 1)
vert_color &= 0xFFFFFFu;
else
vert_color |= m_context->FBA.FBA << 31;
// Apply mask for 16-bit formats.
if (cfmt == 2)
vert_color &= 0x80F8F8F8u;
return vert_color;
}
bool GSRendererHW::IsReallyDithered() const
{
// Must have dither on, not disabled in config, and using 16-bit.
const GSDrawingEnvironment* env = m_draw_env;
if (!env->DTHE.DTHE || GSConfig.Dithering == 0 || GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].fmt != 2)
return false;
// Dithering is still on, but if the matrix is all-zero, it has no effect.
if ((env->DIMX.U64 & UINT64_C(0x7777777777777777)) == 0)
return false;
return true;
}
void GSRendererHW::ReplaceVerticesWithSprite(const GSVector4i& unscaled_rect, const GSVector4i& unscaled_uv_rect, void GSRendererHW::ReplaceVerticesWithSprite(const GSVector4i& unscaled_rect, const GSVector4i& unscaled_uv_rect,
const GSVector2i& unscaled_size, const GSVector4i& scissor) const GSVector2i& unscaled_size, const GSVector4i& scissor)
{ {

16
pcsx2/GS/Renderers/HW/GSRendererHW.h
View File

@ -44,9 +44,13 @@ private:
// Require special argument // Require special argument
bool OI_BlitFMV(GSTextureCache::Target* _rt, GSTextureCache::Source* t, const GSVector4i& r_draw); bool OI_BlitFMV(GSTextureCache::Target* _rt, GSTextureCache::Source* t, const GSVector4i& r_draw);
bool OI_GsMemClear(); // always on bool TryGSMemClear(bool no_rt, bool no_ds);
void OI_DoGsMemClear(const GSOffset& off, const GSVector4i& r, u32 vert_color); void ClearGSLocalMemory(const GSOffset& off, const GSVector4i& r, u32 vert_color);
void OI_DoubleHalfClear(GSTextureCache::Target*& rt, GSTextureCache::Target*& ds); // always on bool DetectDoubleHalfClear(bool& no_rt, bool& no_ds);
bool DetectStripedDoubleClear(bool& no_rt, bool& no_ds);
bool TryTargetClear(GSTextureCache::Target* rt, GSTextureCache::Target* ds, bool preserve_rt_color, bool preserve_depth);
void SetNewFRAME(u32 bp, u32 bw, u32 psm);
void SetNewZBUF(u32 bp, u32 psm);
u16 Interpolate_UV(float alpha, int t0, int t1); u16 Interpolate_UV(float alpha, int t0, int t1);
float alpha0(int L, int X0, int X1); float alpha0(int L, int X0, int X1);
@ -54,6 +58,8 @@ private:
void SwSpriteRender(); void SwSpriteRender();
bool CanUseSwSpriteRender(); bool CanUseSwSpriteRender();
bool IsConstantDirectWriteMemClear(); bool IsConstantDirectWriteMemClear();
u32 GetConstantDirectWriteMemClearColor() const;
bool IsReallyDithered() const;
bool IsDiscardingDstColor(); bool IsDiscardingDstColor();
bool PrimitiveCoversWithoutGaps(); bool PrimitiveCoversWithoutGaps();
@ -97,6 +103,9 @@ private:
bool IsSplitTextureShuffle(); bool IsSplitTextureShuffle();
GSVector4i GetSplitTextureShuffleDrawRect() const; GSVector4i GetSplitTextureShuffleDrawRect() const;
static GSVector4i GetDrawRectForPages(u32 bw, u32 psm, u32 num_pages);
bool TryToResolveSinglePageFramebuffer(GIFRegFRAME& FRAME, bool only_next_draw);
bool IsSplitClearActive() const; bool IsSplitClearActive() const;
bool CheckNextDrawForSplitClear(const GSVector4i& r, u32* pages_covered_by_this_draw) const; bool CheckNextDrawForSplitClear(const GSVector4i& r, u32* pages_covered_by_this_draw) const;
bool IsStartingSplitClear(); bool IsStartingSplitClear();
@ -144,6 +153,7 @@ private:
u32 m_last_channel_shuffle_fbmsk = 0; u32 m_last_channel_shuffle_fbmsk = 0;
GIFRegFRAME m_split_clear_start = {}; GIFRegFRAME m_split_clear_start = {};
GIFRegZBUF m_split_clear_start_Z = {};
u32 m_split_clear_pages = 0; // if zero, inactive u32 m_split_clear_pages = 0; // if zero, inactive
u32 m_split_clear_color = 0; u32 m_split_clear_color = 0;

92
pcsx2/GS/Renderers/HW/GSTextureCache.cpp
View File

@ -1199,7 +1199,7 @@ GSTextureCache::Target* GSTextureCache::FindTargetOverlap(u32 bp, u32 end_block,
} }
GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale, int type, GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale, int type,
bool used, u32 fbmask, bool is_frame, bool preload, bool preload_uploads) bool used, u32 fbmask, bool is_frame, bool preload, bool preserve_target, const GSVector4i draw_rect)
{ {
const GSLocalMemory::psm_t& psm_s = GSLocalMemory::m_psm[TEX0.PSM]; const GSLocalMemory::psm_t& psm_s = GSLocalMemory::m_psm[TEX0.PSM];
const u32 bp = TEX0.TBP0; const u32 bp = TEX0.TBP0;
@ -1354,6 +1354,20 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
dst->m_scale = scale; dst->m_scale = scale;
dst->m_unscaled_size = new_size; dst->m_unscaled_size = new_size;
} }
// Drop dirty rect if we're overwriting the whole target.
if (!preserve_target && draw_rect.rintersect(dst->m_valid).eq(dst->m_valid))
{
if (!dst->m_dirty.empty())
{
GL_INS("TC: Clearing dirty list for %s[%x] because we're overwriting the whole target.", to_string(type), dst->m_TEX0.TBP0);
dst->m_dirty.clear();
}
// And invalidate the target, we're drawing over it so we don't care what's there.
GL_INS("TC: Invalidating target %s[%x] because it's completely overwritten.", to_string(type), dst->m_TEX0.TBP0);
g_gs_device->InvalidateRenderTarget(dst->m_texture);
}
} }
else if (!is_frame && !GSConfig.UserHacks_DisableDepthSupport) else if (!is_frame && !GSConfig.UserHacks_DisableDepthSupport)
{ {
@ -1407,6 +1421,12 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
shader = (fmt_16_bits) ? ShaderConvert::FLOAT16_TO_RGB5A1 : ShaderConvert::FLOAT32_TO_RGBA8; shader = (fmt_16_bits) ? ShaderConvert::FLOAT16_TO_RGB5A1 : ShaderConvert::FLOAT32_TO_RGBA8;
} }
if (!preserve_target && draw_rect.rintersect(dst_match->m_valid).eq(dst_match->m_valid))
{
GL_INS("TC: Not converting existing %s[%x] because it's fully overwritten.", to_string(!type), dst->m_TEX0.TBP0);
}
else
{
// The old target's going to get invalidated (at least until we handle concurrent frame+depth at the same BP), // The old target's going to get invalidated (at least until we handle concurrent frame+depth at the same BP),
// so just move the dirty rects across. // so just move the dirty rects across.
dst->m_dirty = std::move(dst_match->m_dirty); dst->m_dirty = std::move(dst_match->m_dirty);
@ -1415,15 +1435,37 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
// Don't bother copying the old target in if the whole thing is dirty. // Don't bother copying the old target in if the whole thing is dirty.
if (dst->m_dirty.empty() || (~dst->m_dirty.GetDirtyChannels() & GSUtil::GetChannelMask(TEX0.PSM)) != 0 || if (dst->m_dirty.empty() || (~dst->m_dirty.GetDirtyChannels() & GSUtil::GetChannelMask(TEX0.PSM)) != 0 ||
!dst->m_dirty.GetDirtyRect(0, TEX0, dst->GetUnscaledRect()).eq(dst->GetUnscaledRect())) !dst->m_dirty.GetDirtyRect(0, TEX0, dst->GetUnscaledRect()).eq(dst->GetUnscaledRect()))
{
// If the old target was cleared, simply propagate that through.
if (dst_match->m_texture->GetState() == GSTexture::State::Cleared)
{
if (type == DepthStencil)
{
const u32 cc = dst_match->m_texture->GetClearColor();
const float cd = ConvertColorToDepth(cc, shader);
GL_INS("TC: Convert clear color[%08X] to depth[%f]", cc, cd);
g_gs_device->ClearDepth(dst->m_texture, cd);
}
else
{
const float cd = dst_match->m_texture->GetClearDepth();
const u32 cc = ConvertDepthToColor(cd, shader);
GL_INS("TC: Convert clear depth[%f] to color[%08X]", cd, cc);
g_gs_device->ClearRenderTarget(dst->m_texture, cc);
}
}
else if (dst_match->m_texture->GetState() == GSTexture::State::Dirty)
{ {
g_gs_device->StretchRect(dst_match->m_texture, sRect, dst->m_texture, dRect, shader, false); g_gs_device->StretchRect(dst_match->m_texture, sRect, dst->m_texture, dRect, shader, false);
g_perfmon.Put(GSPerfMon::TextureCopies, 1); g_perfmon.Put(GSPerfMon::TextureCopies, 1);
} }
}
// Now pull in any dirty areas in the new format. // Now pull in any dirty areas in the new format.
dst->Update(); dst->Update();
} }
} }
}
if (dst) if (dst)
{ {
@ -1441,7 +1483,7 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
} }
GSTextureCache::Target* GSTextureCache::CreateTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale, int type, GSTextureCache::Target* GSTextureCache::CreateTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale, int type,
bool used, u32 fbmask, bool is_frame, bool preload, bool preload_uploads) bool used, u32 fbmask, bool is_frame, bool preload, bool preserve_target, const GSVector4i draw_rect)
{ {
const u32 bp = TEX0.TBP0; const u32 bp = TEX0.TBP0;
const GSLocalMemory::psm_t& psm_s = GSLocalMemory::m_psm[TEX0.PSM]; const GSLocalMemory::psm_t& psm_s = GSLocalMemory::m_psm[TEX0.PSM];
@ -1485,7 +1527,7 @@ GSTextureCache::Target* GSTextureCache::CreateTarget(GIFRegTEX0 TEX0, const GSVe
if (!is_frame && !forced_preload && !preload) if (!is_frame && !forced_preload && !preload)
{ {
if (preload_uploads) if (preserve_target || !draw_rect.eq(dst->m_valid))
{ {
std::vector<GSState::GSUploadQueue>::iterator iter; std::vector<GSState::GSUploadQueue>::iterator iter;
GSVector4i eerect = GSVector4i::zero(); GSVector4i eerect = GSVector4i::zero();
@ -1680,6 +1722,50 @@ void GSTextureCache::ScaleTargetForDisplay(Target* t, const GIFRegTEX0& dispfb,
GetTargetSize(t->m_TEX0.TBP0, t->m_TEX0.TBW, t->m_TEX0.PSM, 0, static_cast<u32>(needed_height)); GetTargetSize(t->m_TEX0.TBP0, t->m_TEX0.TBW, t->m_TEX0.PSM, 0, static_cast<u32>(needed_height));
} }
float GSTextureCache::ConvertColorToDepth(u32 c, ShaderConvert convert)
{
const float mult = std::exp2(g_gs_device->Features().clip_control ? -32.0f : -24.0f);
switch (convert)
{
case ShaderConvert::RGB5A1_TO_FLOAT16:
return static_cast<float>(((c & 0xF8u) >> 3) | (((c >> 8) & 0xF8u) << 2) | (((c >> 16) & 0xF8u) << 7) |
(((c >> 24) & 0x80u) << 8)) *
mult;
case ShaderConvert::RGBA8_TO_FLOAT16:
return static_cast<float>(c & 0x0000FFFF) * mult;
case ShaderConvert::RGBA8_TO_FLOAT24:
return static_cast<float>(c & 0x00FFFFFF) * mult;
case ShaderConvert::RGBA8_TO_FLOAT32:
default:
return static_cast<float>(c) * mult;
}
}
u32 GSTextureCache::ConvertDepthToColor(float d, ShaderConvert convert)
{
const float mult = std::exp2(g_gs_device->Features().clip_control ? 32.0f : 24.0f);
switch (convert)
{
case ShaderConvert::FLOAT16_TO_RGB5A1:
{
const u32 cc = static_cast<u32>(d * mult);
// Truely awful.
const GSVector4i vcc = GSVector4i(
GSVector4(GSVector4i(cc & 0x1Fu, (cc >> 5) & 0x1Fu, (cc >> 10) & 0x1Fu, (cc >> 15) & 0x01u)) *
GSVector4::cxpr(255.0f / 31.0f));
return (vcc.r | (vcc.g << 8) | (vcc.b << 16) | (vcc.a << 24));
}
case ShaderConvert::FLOAT32_TO_RGBA8:
default:
return static_cast<u32>(d * mult);
}
}
bool GSTextureCache::PrepareDownloadTexture(u32 width, u32 height, GSTexture::Format format, std::unique_ptr<GSDownloadTexture>* tex) bool GSTextureCache::PrepareDownloadTexture(u32 width, u32 height, GSTexture::Format format, std::unique_ptr<GSDownloadTexture>* tex)
{ {
GSDownloadTexture* ctex = tex->get(); GSDownloadTexture* ctex = tex->get();

12
pcsx2/GS/Renderers/HW/GSTextureCache.h
View File

@ -457,9 +457,11 @@ public:
Target* FindTargetOverlap(u32 bp, u32 end_block, int type, int psm); Target* FindTargetOverlap(u32 bp, u32 end_block, int type, int psm);
Target* LookupTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale, int type, bool used = true, u32 fbmask = 0, Target* LookupTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale, int type, bool used = true, u32 fbmask = 0,
bool is_frame = false, bool preload = GSConfig.PreloadFrameWithGSData, bool preload_uploads = true); bool is_frame = false, bool preload = GSConfig.PreloadFrameWithGSData, bool preserve_target = true,
const GSVector4i draw_rc = GSVector4i::zero());
Target* CreateTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale, int type, bool used = true, u32 fbmask = 0, Target* CreateTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale, int type, bool used = true, u32 fbmask = 0,
bool is_frame = false, bool preload = GSConfig.PreloadFrameWithGSData, bool preload_uploads = true); bool is_frame = false, bool preload = GSConfig.PreloadFrameWithGSData, bool preserve_target = true,
const GSVector4i draw_rc = GSVector4i::zero());
Target* LookupDisplayTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale); Target* LookupDisplayTarget(GIFRegTEX0 TEX0, const GSVector2i& size, float scale);
/// Looks up a target in the cache, and only returns it if the BP/BW match exactly. /// Looks up a target in the cache, and only returns it if the BP/BW match exactly.
@ -481,6 +483,12 @@ public:
void ReplaceSourceTexture(Source* s, GSTexture* new_texture, float new_scale, const GSVector2i& new_unscaled_size, void ReplaceSourceTexture(Source* s, GSTexture* new_texture, float new_scale, const GSVector2i& new_unscaled_size,
HashCacheEntry* hc_entry, bool new_texture_is_shared); HashCacheEntry* hc_entry, bool new_texture_is_shared);
/// Converts single color value to depth using the specified shader expression.
static float ConvertColorToDepth(u32 c, ShaderConvert convert);
/// Converts single depth value to colour using the specified shader expression.
static u32 ConvertDepthToColor(float d, ShaderConvert convert);
bool 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 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 ShuffleMove(u32 BP, u32 BW, u32 PSM, int sx, int sy, int dx, int dy, int w, int h); bool ShuffleMove(u32 BP, u32 BW, u32 PSM, int sx, int sy, int dx, int dy, int w, int h);
bool PageMove(u32 SBP, u32 DBP, u32 BW, u32 PSM, int sx, int sy, int dx, int dy, int w, int h); bool PageMove(u32 SBP, u32 DBP, u32 BW, u32 PSM, int sx, int sy, int dx, int dy, int w, int h);