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

GS/HW: More alterations for new RT in RT system

This commit is contained in:
refractionpcsx2 2025-01-11 01:29:08 +00:00
parent 8857736b62
commit 12cdc86984
2 changed files with 90 additions and 57 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
pcsx2/GS/Renderers/HW/GSRendererHW.cpp
View File

@ -2594,20 +2594,23 @@ void GSRendererHW::Draw()
GIFRegTEX0 FRAME_TEX0; GIFRegTEX0 FRAME_TEX0;
bool shuffle_target = false; bool shuffle_target = false;
if (!no_rt && GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].bpp == 16 && GSLocalMemory::m_psm[m_cached_ctx.TEX0.PSM].bpp >= 16) if (!no_rt && GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].bpp == 16 && GSLocalMemory::m_psm[m_cached_ctx.TEX0.PSM].bpp >= 16 &&
(m_vt.m_primclass == GS_SPRITE_CLASS || (m_vt.m_primclass == GS_TRIANGLE_CLASS && (m_index.tail % 6) == 0 && TrianglesAreQuads(true))))
{ {
if (!shuffle_target && GSLocalMemory::m_psm[m_cached_ctx.TEX0.PSM].bpp == 16) if (!shuffle_target && GSLocalMemory::m_psm[m_cached_ctx.TEX0.PSM].bpp == 16)
{ {
const GSVertex* v = &m_vertex.buff[0]; const GSVertex* v = &m_vertex.buff[0];
const int first_x = std::abs(static_cast<int>(((v[0].XYZ.X - m_context->XYOFFSET.OFX) + 8))) >> 4; const int first_x = std::clamp((static_cast<int>(((v[0].XYZ.X - m_context->XYOFFSET.OFX) + 8))) >> 4, 0, 2048);
const int first_u = PRIM->FST ? ((v[0].U + 9) >> 4) : static_cast<int>(((1 << m_cached_ctx.TEX0.TW) * (v[0].ST.S / v[1].RGBAQ.Q))); const bool offset_last = PRIM->FST ? (v[1].U > v[0].U) : ((v[1].ST.S / v[1].RGBAQ.Q) > (v[0].ST.S / v[1].RGBAQ.Q));
const int second_u = PRIM->FST ? ((v[1].U + 9) >> 4) : static_cast<int>(((1 << m_cached_ctx.TEX0.TW) * (v[1].ST.S / v[1].RGBAQ.Q)) + 0.6f); const int first_u = PRIM->FST ? ((v[0].U + (offset_last ? 0 : 9)) >> 4) : std::clamp(static_cast<int>(((1 << m_cached_ctx.TEX0.TW) * (v[0].ST.S / v[1].RGBAQ.Q)) + (offset_last ? 0.0f : 0.6f)), 0, 2048);
const int second_u = PRIM->FST ? ((v[1].U + (offset_last ? 9 : 0)) >> 4) : std::clamp(static_cast<int>(((1 << m_cached_ctx.TEX0.TW) * (v[1].ST.S / v[1].RGBAQ.Q)) + (offset_last ? 0.6f : 0.0f)), 0, 2048);
// offset coordinates swap around RG/BA. (Ace Combat) // offset coordinates swap around RG/BA. (Ace Combat)
const u32 minv = m_cached_ctx.CLAMP.MINV; const u32 minv = m_cached_ctx.CLAMP.MINV;
const u32 minu = m_cached_ctx.CLAMP.MINU; const u32 minu = m_cached_ctx.CLAMP.MINU;
const bool rgba_shuffle = ((m_cached_ctx.CLAMP.WMS == m_cached_ctx.CLAMP.WMT && m_cached_ctx.CLAMP.WMS == CLAMP_REGION_REPEAT) && (minu && minv)); const bool rgba_shuffle = ((m_cached_ctx.CLAMP.WMS == m_cached_ctx.CLAMP.WMT && m_cached_ctx.CLAMP.WMS == CLAMP_REGION_REPEAT) && (minu && minv));
const bool shuffle_coords = ((first_x ^ first_u) & 8) || rgba_shuffle; const bool shuffle_coords = ((first_x ^ first_u) & 0xF) == 8 || rgba_shuffle;
// Round up half of second coord, it can sometimes be slightly under. // Round up half of second coord, it can sometimes be slightly under.
const int draw_width = std::abs(v[1].XYZ.X + 9 - v[0].XYZ.X) >> 4; const int draw_width = std::abs(v[1].XYZ.X + 9 - v[0].XYZ.X) >> 4;
const int read_width = std::abs(second_u - first_u); const int read_width = std::abs(second_u - first_u);
@ -3038,14 +3041,13 @@ void GSRendererHW::Draw()
// Don't resize if the BPP don't match. // Don't resize if the BPP don't match.
if (frame_psm.bpp == GSLocalMemory::m_psm[rt->m_TEX0.PSM].bpp) if (frame_psm.bpp == GSLocalMemory::m_psm[rt->m_TEX0.PSM].bpp)
{ {
if (m_r.w > rt->m_unscaled_size.y) if (m_r.w > rt->m_unscaled_size.y || m_r.z > rt->m_unscaled_size.x)
{ {
u32 new_height = m_r.w; u32 new_height = std::max(m_r.w, rt->m_unscaled_size.y);
u32 new_width = std::max(m_r.z, rt->m_unscaled_size.x);
if (possible_shuffle && std::abs(static_cast<s16>(GSLocalMemory::m_psm[rt->m_TEX0.PSM].bpp - GSLocalMemory::m_psm[TEX0.PSM].bpp)) == 16)
new_height /= 2;
//DevCon.Warning("Resizing texture %d x %d draw %d", rt->m_unscaled_size.x, new_height, s_n); //DevCon.Warning("Resizing texture %d x %d draw %d", rt->m_unscaled_size.x, new_height, s_n);
rt->ResizeTexture(rt->m_unscaled_size.x, new_height); rt->ResizeTexture(new_height, new_height);
const bool frame_masked = ((m_cached_ctx.FRAME.FBMSK & frame_psm.fmsk) == frame_psm.fmsk) || (m_cached_ctx.TEST.ATE && m_cached_ctx.TEST.ATST == ATST_NEVER && !(m_cached_ctx.TEST.AFAIL & AFAIL_FB_ONLY)); const bool frame_masked = ((m_cached_ctx.FRAME.FBMSK & frame_psm.fmsk) == frame_psm.fmsk) || (m_cached_ctx.TEST.ATE && m_cached_ctx.TEST.ATST == ATST_NEVER && !(m_cached_ctx.TEST.AFAIL & AFAIL_FB_ONLY));
@ -3650,7 +3652,7 @@ void GSRendererHW::Draw()
if (rt && GSConfig.SaveRT && s_n >= GSConfig.SaveN) if (rt && GSConfig.SaveRT && s_n >= GSConfig.SaveN)
{ {
s = GetDrawDumpPath("%05d_f%lld_rt0_%05x_%s.bmp", s_n, frame, m_cached_ctx.FRAME.Block(), psm_str(m_cached_ctx.FRAME.PSM)); s = GetDrawDumpPath("%05d_f%lld_rt0_%05x_(%05x)_%s.bmp", s_n, frame, m_cached_ctx.FRAME.Block(), rt->m_TEX0.TBP0, psm_str(m_cached_ctx.FRAME.PSM));
if (rt->m_texture) if (rt->m_texture)
rt->m_texture->Save(s); rt->m_texture->Save(s);
@ -3658,7 +3660,7 @@ void GSRendererHW::Draw()
if (ds && GSConfig.SaveDepth && s_n >= GSConfig.SaveN) if (ds && GSConfig.SaveDepth && s_n >= GSConfig.SaveN)
{ {
s = GetDrawDumpPath("%05d_f%lld_rz0_%05x_%s.bmp", s_n, frame, m_cached_ctx.ZBUF.Block(), psm_str(m_cached_ctx.ZBUF.PSM)); s = GetDrawDumpPath("%05d_f%lld_rz0_%05x_(%05x)_%s.bmp", s_n, frame, m_cached_ctx.ZBUF.Block(), ds->m_TEX0.TBP0, psm_str(m_cached_ctx.ZBUF.PSM));
if (g_texture_cache->GetTemporaryZ()) if (g_texture_cache->GetTemporaryZ())
g_texture_cache->GetTemporaryZ()->Save(s); g_texture_cache->GetTemporaryZ()->Save(s);

81
pcsx2/GS/Renderers/HW/GSTextureCache.cpp
View File

@ -149,8 +149,7 @@ void GSTextureCache::AddDirtyRectTarget(Target* target, GSVector4i rect, u32 psm
if (rect.rempty()) if (rect.rempty())
return; return;
if (rect.w > 2048)
DevCon.Warning("BAd");
std::vector<GSDirtyRect>::iterator it = target->m_dirty.end(); std::vector<GSDirtyRect>::iterator it = target->m_dirty.end();
while (it != target->m_dirty.begin()) while (it != target->m_dirty.begin())
{ {
@ -336,10 +335,24 @@ GSVector4i GSTextureCache::TranslateAlignedRectByPage(u32 tbp, u32 tebp, u32 tbw
{ {
// Results won't be square, if it's not invalidation, it's a texture, which is problematic to translate, so let's not (FIFA 2005). // Results won't be square, if it's not invalidation, it's a texture, which is problematic to translate, so let's not (FIFA 2005).
if (!is_invalidation) if (!is_invalidation)
{
if (sbp != tbp)
{
// Just take the start page, as this is likely tex in rt, and that's all we care about.
const u32 start_page = (in_rect.y / src_page_size.y) + (in_rect.x / src_page_size.x);
in_rect.x = (start_page % dst_pgw) * dst_page_size.x;
in_rect.y = (start_page / dst_pgw) * dst_page_size.y;
in_rect.z = in_rect.x + dst_page_size.x;
in_rect.w = in_rect.y + dst_page_size.y;
return in_rect;
}
else
{ {
DevCon.Warning("Uneven pages mess up sbp %x dbp %x spgw %d dpgw %d", sbp, tbp, src_pgw, dst_pgw); DevCon.Warning("Uneven pages mess up sbp %x dbp %x spgw %d dpgw %d", sbp, tbp, src_pgw, dst_pgw);
return GSVector4i::zero(); return GSVector4i::zero();
} }
}
//TODO: Maybe control dirty blocks directly and add them page at a time for better granularity. //TODO: Maybe control dirty blocks directly and add them page at a time for better granularity.
const u32 start_y_page = (rect_pages.y * src_pgw) / dst_pgw; const u32 start_y_page = (rect_pages.y * src_pgw) / dst_pgw;
@ -1465,13 +1478,16 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const bool is_color, const
if (!t->HasValidBitsForFormat(psm, req_color, req_alpha) && !(possible_shuffle && GSLocalMemory::m_psm[psm].bpp == 16 && GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == 32)) if (!t->HasValidBitsForFormat(psm, req_color, req_alpha) && !(possible_shuffle && GSLocalMemory::m_psm[psm].bpp == 16 && GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == 32))
continue; continue;
if (GSLocalMemory::m_psm[color_psm].bpp == 16 && GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp && (t->m_TEX0.TBW != bw && (t->m_TEX0.TBW * 2) != bw)) u32 horz_page_offset = ((bp - t->m_TEX0.TBP0) >> 5) % t->m_TEX0.TBW;
if (GSLocalMemory::m_psm[color_psm].bpp == 16 && GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == 32 && bw != 1 &&
((t->m_TEX0.TBW < (horz_page_offset + ((block_boundary_rect.z + GSLocalMemory::m_psm[psm].pgs.x - 1) / GSLocalMemory::m_psm[psm].pgs.x)) ||
(t->m_TEX0.TBW != bw && block_boundary_rect.w > GSLocalMemory::m_psm[psm].pgs.y))))
{ {
DevCon.Warning("BP %x - 16bit bad match for target bp %x bw %d src %d format %d", bp, t->m_TEX0.TBP0, t->m_TEX0.TBW, bw, t->m_TEX0.PSM); DevCon.Warning("BP %x - 16bit bad match for target bp %x bw %d src %d format %d", bp, t->m_TEX0.TBP0, t->m_TEX0.TBW, bw, t->m_TEX0.PSM);
continue; continue;
} }
else if (!possible_shuffle && (GSLocalMemory::m_psm[color_psm].bpp == 8 && GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == 32 && else if (!possible_shuffle && (GSLocalMemory::m_psm[color_psm].bpp == 8 && GSLocalMemory::m_psm[t->m_TEX0.PSM].bpp == 32 && bw != 1 &&
!((t->m_TEX0.TBW == (bw / 2)) || (t->m_TEX0.TBW >= (bw / 2) && (req_rect.w < GSLocalMemory::m_psm[psm].pgs.y))))) !((t->m_TEX0.TBW == (bw / 2)) || (t->m_TEX0.TBW >= (bw / 2) && (block_boundary_rect.w <= GSLocalMemory::m_psm[psm].pgs.y)))))
{ {
DevCon.Warning("BP %x - 8bit bad match for target bp %x bw %d src %d format %d", bp, t->m_TEX0.TBP0, t->m_TEX0.TBW, bw, t->m_TEX0.PSM); DevCon.Warning("BP %x - 8bit bad match for target bp %x bw %d src %d format %d", bp, t->m_TEX0.TBP0, t->m_TEX0.TBW, bw, t->m_TEX0.PSM);
continue; continue;
@ -1935,21 +1951,17 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
}*/ }*/
const u32 widthpage_offset = (std::abs(static_cast<int>(bp - t->m_TEX0.TBP0)) >> 5) % std::max(t->m_TEX0.TBW, 1U); const u32 widthpage_offset = (std::abs(static_cast<int>(bp - t->m_TEX0.TBP0)) >> 5) % std::max(t->m_TEX0.TBW, 1U);
const bool is_aligned_ok = widthpage_offset == 0 || (t->m_TEX0.TBW == TEX0.TBW && ((min_rect.z >> 6) + widthpage_offset) <= TEX0.TBW) || ((widthpage_offset + TEX0.TBW) <= t->m_TEX0.TBW) || min_rect.width() <= 64 || (widthpage_offset == (t->m_TEX0.TBW >> 1) && (static_cast<u32>(min_rect.width()) <= (widthpage_offset * 64))); const bool is_aligned_ok = widthpage_offset == 0 || (t->m_TEX0.TBW == TEX0.TBW &&
if ((!dst || ((GSState::s_n - dst->m_last_draw) < (GSState::s_n - t->m_last_draw))) && is_aligned_ok && (t->m_TEX0.TBW == TEX0.TBW || (TEX0.TBW == 1 && t->m_TEX0.TBW > 1)) && t->Overlaps(bp, TEX0.TBW, TEX0.PSM, min_rect)) ((((min_rect.z + 63) >> 6) + widthpage_offset) <= TEX0.TBW) ||
{ /*TEX0.TBP0 == ((((t->UnwrappedEndBlock() + 1) - t->m_TEX0.TBP0) >> 1) + t->m_TEX0.TBP0)*/ ((widthpage_offset + TEX0.TBW) <= t->m_TEX0.TBW) ||
// If it's too old, it's probably not a real target to jump in to anymore. min_rect.width() <= 64 || (widthpage_offset == (t->m_TEX0.TBW >> 1) &&
/*if ((GSState::s_n - t->m_last_draw) > 10 && (!t->m_dirty.empty() || (!is_shuffle && (static_cast<u32>(min_rect.width()) <= (widthpage_offset * 64))));
!(widthpage_offset == 0 || min_rect.width() <= 64 || if ((!dst || ((GSState::s_n - dst->m_last_draw) < (GSState::s_n - t->m_last_draw))) && is_aligned_ok && (t->m_TEX0.TBW == TEX0.TBW || TEX0.TBW == 1) && t->Overlaps(bp, TEX0.TBW, TEX0.PSM, min_rect))
(widthpage_offset == (t->m_TEX0.TBW >> 1) && min_rect.width() == widthpage_offset * 64)))))
{ {
GL_INS("TC: Deleting RT BP 0x%x BW %d PSM %s due to change in target", t->m_TEX0.TBP0, t->m_TEX0.TBW, psm_str(t->m_TEX0.PSM)); const GSLocalMemory::psm_t& s_psm = GSLocalMemory::m_psm[TEX0.PSM];
InvalidateSourcesFromTarget(t);
i = list.erase(i); if (!is_shuffle && (!GSUtil::HasSameSwizzleBits(t->m_TEX0.PSM, TEX0.PSM) ||
delete t; (widthpage_offset % std::max(t->m_TEX0.TBW, 1U)) != 0 && ((widthpage_offset + (min_rect.width() + (s_psm.pgs.x - 1)) / s_psm.pgs.x)) > t->m_TEX0.TBW))
}
else*/
if (!is_shuffle && !GSUtil::HasSameSwizzleBits(t->m_TEX0.PSM, TEX0.PSM))
{ {
GL_INS("TC: Deleting RT BP 0x%x BW %d PSM %s due to change in target", t->m_TEX0.TBP0, t->m_TEX0.TBW, psm_str(t->m_TEX0.PSM)); GL_INS("TC: Deleting RT BP 0x%x BW %d PSM %s due to change in target", t->m_TEX0.TBP0, t->m_TEX0.TBW, psm_str(t->m_TEX0.PSM));
InvalidateSourcesFromTarget(t); InvalidateSourcesFromTarget(t);
@ -1958,7 +1970,7 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
continue; continue;
} }
else else if (t->m_dirty.empty())
{ {
//DevCon.Warning("Here draw %d wanted %x PSM %x got %x PSM %x offset of %d pages width %d pages draw width %d", GSState::s_n, bp, TEX0.PSM, t->m_TEX0.TBP0, t->m_TEX0.PSM, (bp - t->m_TEX0.TBP0) >> 5, t->m_TEX0.TBW, draw_rect.width()); //DevCon.Warning("Here draw %d wanted %x PSM %x got %x PSM %x offset of %d pages width %d pages draw width %d", GSState::s_n, bp, TEX0.PSM, t->m_TEX0.TBP0, t->m_TEX0.PSM, (bp - t->m_TEX0.TBP0) >> 5, t->m_TEX0.TBW, draw_rect.width());
dst = t; dst = t;
@ -2136,7 +2148,8 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
new_scaled_size = ScaleRenderTargetSize(dst->m_unscaled_size, scale); new_scaled_size = ScaleRenderTargetSize(dst->m_unscaled_size, scale);
dRect = (GSVector4(GSVector4i::loadh(dst->m_unscaled_size)) * GSVector4(scale)).ceil(); dRect = (GSVector4(GSVector4i::loadh(dst->m_unscaled_size)) * GSVector4(scale)).ceil();
if (!is_shuffle || GSLocalMemory::m_psm[dst->m_TEX0.PSM].bpp == 16)
{
if (scale_down) if (scale_down)
{ {
if ((new_size.y * 2) < 1024) if ((new_size.y * 2) < 1024)
@ -2158,6 +2171,7 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
dst->m_valid.y /= 2; dst->m_valid.y /= 2;
dst->m_valid.w /= 2; dst->m_valid.w /= 2;
} }
}
if (!is_shuffle) if (!is_shuffle)
{ {
GL_INS("TC Convert to 16bit: %dx%d: %dx%d @ %f -> %dx%d @ %f", dst->m_unscaled_size.x, dst->m_unscaled_size.y, GL_INS("TC Convert to 16bit: %dx%d: %dx%d @ %f -> %dx%d @ %f", dst->m_unscaled_size.x, dst->m_unscaled_size.y,
@ -2188,9 +2202,16 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
dst->m_texture = tex; dst->m_texture = tex;
dst->m_unscaled_size = new_size; dst->m_unscaled_size = new_size;
} }
// New format or doing a shuffle to a 32bit target that used to be 16bit
dst->m_TEX0.PSM = TEX0.PSM;
// New format or doing a shuffle to a 32bit target that used to be 16bit
if (!is_shuffle)
dst->m_TEX0.PSM = TEX0.PSM;
// LEGO Dome Racers does a copy to a target as 8bit in alpha only, this doesn't really work great for us, so let's make it 32bit with invalid RGB.
else if (dst->m_TEX0.PSM == PSMT8H)
{
//dst->m_TEX0.PSM = PSMCT32;
dst->m_valid_rgb = false;
}
} }
// If our RGB was invalidated, we need to pull it from depth. // If our RGB was invalidated, we need to pull it from depth.
@ -2315,7 +2336,15 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
{ {
continue; continue;
} }
// If the format is completely different, but it's the same location, it's likely just overwriting it, so get rid.
if (!is_shuffle && t->m_TEX0.TBW != TEX0.TBW && TEX0.TBW != 1 && !preserve_rgb && min_rect.w > GSLocalMemory::m_psm[t->m_TEX0.PSM].pgs.y)
{
DevCon.Warning("Deleting Z draw %d", GSState::s_n);
InvalidateSourcesFromTarget(t);
i = rev_list.erase(i);
delete t;
continue;
}
const GSLocalMemory::psm_t& t_psm_s = GSLocalMemory::m_psm[t->m_TEX0.PSM]; const GSLocalMemory::psm_t& t_psm_s = GSLocalMemory::m_psm[t->m_TEX0.PSM];
if (t_psm_s.bpp != psm_s.bpp) if (t_psm_s.bpp != psm_s.bpp)
{ {
@ -2486,7 +2515,8 @@ GSTextureCache::Target* GSTextureCache::LookupTarget(GIFRegTEX0 TEX0, const GSVe
pxAssert(dst && dst->m_texture && dst->m_scale == scale); pxAssert(dst && dst->m_texture && dst->m_scale == scale);
} }
if (dst && dst->m_TEX0.TBP0 == 0x3f80 && dst->m_TEX0.PSM == 0)
DevCon.Warning("It's 32bit on draw %d", GSState::s_n);
return dst; return dst;
} }
@ -2819,6 +2849,7 @@ bool GSTextureCache::PreloadTarget(GIFRegTEX0 TEX0, const GSVector2i& size, cons
GL_INS("RT double buffer copy from FBP 0x%x, %dx%d => %d,%d", t->m_TEX0.TBP0, copy_width, copy_height, 0, dst_offset_scaled_height); GL_INS("RT double buffer copy from FBP 0x%x, %dx%d => %d,%d", t->m_TEX0.TBP0, copy_width, copy_height, 0, dst_offset_scaled_height);
// Clear the dirty first // Clear the dirty first
t->Update(); t->Update();
dst->Update(); dst->Update();