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

gsdx: add some notes of potential optimization

This commit is contained in:
Gregory Hainaut 2016-05-23 19:23:04 +02:00
parent 5c02636dd9
commit 1f4439a89e
2 changed files with 29 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
plugins/GSdx/GSRendererHW.cpp
View File

@ -52,6 +52,24 @@ void GSRendererHW::SetScaling()
{ {
GSVector2i crtc_size(GetDisplayRect().width(), GetDisplayRect().height()); GSVector2i crtc_size(GetDisplayRect().width(), GetDisplayRect().height());
// Details of (potential) perf impact of a big framebuffer
// 1/ extra memory
// 2/ texture cache framebuffer rescaling/copy
// 3/ upload of framebuffer (preload hack)
// 4/ framebuffer clear (color/depth/stencil)
// 5/ read back of the frambuffer
// 6/ MSAA
//
// With the solution
// 1/ Nothing to do.Except the texture cache bug (channel shuffle effect)
// most of the market is 1GB of VRAM (and soon 2GB)
// 2/ limit rescaling/copy to the valid data of the framebuffer
// 3/ ??? no solution so far
// 4a/ stencil can be limited to valid data.
// 4b/ is it useful to clear color? depth? (in any case, it ought to be few operation)
// 5/ limit the read to the valid data
// 6/ not support on openGL
// Framebuffer width is always a multiple of 64 so at certain cases it can't cover some weird width values. // Framebuffer width is always a multiple of 64 so at certain cases it can't cover some weird width values.
// 480P , 576P use width as 720 which is not referencable by FBW * 64. so it produces 704 ( the closest value multiple by 64). // 480P , 576P use width as 720 which is not referencable by FBW * 64. so it produces 704 ( the closest value multiple by 64).
// In such cases, let's just use the CRTC width. // In such cases, let's just use the CRTC width.
@ -848,6 +866,7 @@ void GSRendererHW::OI_GsMemClear()
// FIXME: loop can likely be optimized with AVX/SSE. Pixels aren't // FIXME: loop can likely be optimized with AVX/SSE. Pixels aren't
// linear but the value will be done for all pixels of a block. // linear but the value will be done for all pixels of a block.
// FIXME: maybe we could limit the write to the top and bottom row page.
if (format == 0) { if (format == 0) {
// Based on WritePixel32 // Based on WritePixel32
for(int y = r.top; y < r.bottom; y++) for(int y = r.top; y < r.bottom; y++)

10
plugins/GSdx/GSRendererOGL.cpp
View File

@ -674,6 +674,10 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
// vertex list (it will interact with PrimitiveOverlap and accurate // vertex list (it will interact with PrimitiveOverlap and accurate
// blending) // blending)
// //
//
// Uncomment to disable HLE emulation (allow to trace the draw call)
// m_channel_shuffle = false;
// First let's check we really have a channel shuffle effect // First let's check we really have a channel shuffle effect
if (m_channel_shuffle) { if (m_channel_shuffle) {
if (m_game.title == CRC::GT4 || m_game.title == CRC::GT3 || m_game.title == CRC::GTConcept || m_game.title == CRC::TouristTrophy) { if (m_game.title == CRC::GT4 || m_game.title == CRC::GT3 || m_game.title == CRC::GTConcept || m_game.title == CRC::TouristTrophy) {
@ -861,7 +865,13 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
DATE = false; DATE = false;
} }
} }
} else if (!om_csel.wa && (!m_context->TEST.ATE || m_context->TEST.ATST == ATST_ALWAYS)) {
// TODO: is it legal ? Likely but it need to be tested carefully
// DATE_GL45 = true;
// m_unsafe_fbmask = true; << replace it with a cheap barrier
} }
// Will save my life ! // Will save my life !
ASSERT(!(DATE_GL45 && DATE_one)); ASSERT(!(DATE_GL45 && DATE_one));
ASSERT(!(DATE_GL42 && DATE_one)); ASSERT(!(DATE_GL42 && DATE_one));