From b29e5146ec63e1b35cee45e01f54c4b69f15938e Mon Sep 17 00:00:00 2001
From: comex <comexk@gmail.com>
Date: Tue, 21 Oct 2014 20:42:55 -0400
Subject: [PATCH] Convert some VideoCommon stuff to BitSet.
Now with a minor performance improvement removed for no reason.
---
Source/Core/VideoCommon/CPMemory.h | 3 +-
.../Core/VideoCommon/VertexLoaderManager.cpp | 18 ++++----
Source/Core/VideoCommon/VertexManagerBase.cpp | 43 +++++++++----------
.../Core/VideoCommon/VertexShaderManager.cpp | 40 +++++------------
4 files changed, 42 insertions(+), 62 deletions(-)
diff --git a/Source/Core/VideoCommon/CPMemory.h b/Source/Core/VideoCommon/CPMemory.h
index ae8ff08303..a256fc6849 100644
--- a/Source/Core/VideoCommon/CPMemory.h
+++ b/Source/Core/VideoCommon/CPMemory.h
@@ -4,6 +4,7 @@
#pragma once
+#include "Common/BitSet.h"
#include "Common/CommonTypes.h"
// Vertex array numbers
@@ -252,7 +253,7 @@ struct CPState final
VAT vtx_attr[8];
// Attributes that actually belong to VertexLoaderManager:
- int attr_dirty; // bitfield
+ BitSet32 attr_dirty;
VertexLoader* vertex_loaders[8];
};
diff --git a/Source/Core/VideoCommon/VertexLoaderManager.cpp b/Source/Core/VideoCommon/VertexLoaderManager.cpp
index ca925d0302..b8132fa806 100644
--- a/Source/Core/VideoCommon/VertexLoaderManager.cpp
+++ b/Source/Core/VideoCommon/VertexLoaderManager.cpp
@@ -100,14 +100,14 @@ void AppendListToString(std::string *dest)
void MarkAllDirty()
{
- g_main_cp_state.attr_dirty = 0xff;
- g_preprocess_cp_state.attr_dirty = 0xff;
+ g_main_cp_state.attr_dirty = BitSet32::AllTrue(8);
+ g_preprocess_cp_state.attr_dirty = BitSet32::AllTrue(8);
}
static VertexLoader* RefreshLoader(int vtx_attr_group, CPState* state)
{
VertexLoader* loader;
- if ((state->attr_dirty >> vtx_attr_group) & 1)
+ if (state->attr_dirty[vtx_attr_group])
{
VertexLoaderUID uid(state->vtx_desc, state->vtx_attr[vtx_attr_group]);
std::lock_guard<std::mutex> lk(s_vertex_loader_map_lock);
@@ -123,7 +123,7 @@ static VertexLoader* RefreshLoader(int vtx_attr_group, CPState* state)
INCSTAT(stats.numVertexLoaders);
}
state->vertex_loaders[vtx_attr_group] = loader;
- state->attr_dirty &= ~(1 << vtx_attr_group);
+ state->attr_dirty[vtx_attr_group] = false;
} else {
loader = state->vertex_loaders[vtx_attr_group];
}
@@ -200,31 +200,31 @@ void LoadCPReg(u32 sub_cmd, u32 value, bool is_preprocess)
case 0x50:
state->vtx_desc.Hex &= ~0x1FFFF; // keep the Upper bits
state->vtx_desc.Hex |= value;
- state->attr_dirty = 0xFF;
+ state->attr_dirty = BitSet32::AllTrue(8);
break;
case 0x60:
state->vtx_desc.Hex &= 0x1FFFF; // keep the lower 17Bits
state->vtx_desc.Hex |= (u64)value << 17;
- state->attr_dirty = 0xFF;
+ state->attr_dirty = BitSet32::AllTrue(8);
break;
case 0x70:
_assert_((sub_cmd & 0x0F) < 8);
state->vtx_attr[sub_cmd & 7].g0.Hex = value;
- state->attr_dirty |= 1 << (sub_cmd & 7);
+ state->attr_dirty[sub_cmd & 7] = true;
break;
case 0x80:
_assert_((sub_cmd & 0x0F) < 8);
state->vtx_attr[sub_cmd & 7].g1.Hex = value;
- state->attr_dirty |= 1 << (sub_cmd & 7);
+ state->attr_dirty[sub_cmd & 7] = true;
break;
case 0x90:
_assert_((sub_cmd & 0x0F) < 8);
state->vtx_attr[sub_cmd & 7].g2.Hex = value;
- state->attr_dirty |= 1 << (sub_cmd & 7);
+ state->attr_dirty[sub_cmd & 7] = true;
break;
// Pointers to vertex arrays in GC RAM
diff --git a/Source/Core/VideoCommon/VertexManagerBase.cpp b/Source/Core/VideoCommon/VertexManagerBase.cpp
index d637ad9017..7a18ba435b 100644
--- a/Source/Core/VideoCommon/VertexManagerBase.cpp
+++ b/Source/Core/VideoCommon/VertexManagerBase.cpp
@@ -180,39 +180,36 @@ void VertexManager::Flush()
(int)bpmem.genMode.numtexgens, (u32)bpmem.dstalpha.enable, (bpmem.alpha_test.hex>>16)&0xff);
#endif
- u32 usedtextures = 0;
+ BitSet32 usedtextures;
for (u32 i = 0; i < bpmem.genMode.numtevstages + 1u; ++i)
if (bpmem.tevorders[i / 2].getEnable(i & 1))
- usedtextures |= 1 << bpmem.tevorders[i/2].getTexMap(i & 1);
+ usedtextures[bpmem.tevorders[i/2].getTexMap(i & 1)] = true;
if (bpmem.genMode.numindstages > 0)
for (unsigned int i = 0; i < bpmem.genMode.numtevstages + 1u; ++i)
if (bpmem.tevind[i].IsActive() && bpmem.tevind[i].bt < bpmem.genMode.numindstages)
- usedtextures |= 1 << bpmem.tevindref.getTexMap(bpmem.tevind[i].bt);
+ usedtextures[bpmem.tevindref.getTexMap(bpmem.tevind[i].bt)] = true;
- for (unsigned int i = 0; i < 8; i++)
+ for (unsigned int i : usedtextures)
{
- if (usedtextures & (1 << i))
- {
- g_renderer->SetSamplerState(i & 3, i >> 2);
- const FourTexUnits &tex = bpmem.tex[i >> 2];
- const TextureCache::TCacheEntryBase* tentry = TextureCache::Load(i,
- (tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5,
- tex.texImage0[i&3].width + 1, tex.texImage0[i&3].height + 1,
- tex.texImage0[i&3].format, tex.texTlut[i&3].tmem_offset<<9,
- tex.texTlut[i&3].tlut_format,
- ((tex.texMode0[i&3].min_filter & 3) != 0),
- (tex.texMode1[i&3].max_lod + 0xf) / 0x10,
- (tex.texImage1[i&3].image_type != 0));
+ g_renderer->SetSamplerState(i & 3, i >> 2);
+ const FourTexUnits &tex = bpmem.tex[i >> 2];
+ const TextureCache::TCacheEntryBase* tentry = TextureCache::Load(i,
+ (tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5,
+ tex.texImage0[i&3].width + 1, tex.texImage0[i&3].height + 1,
+ tex.texImage0[i&3].format, tex.texTlut[i&3].tmem_offset<<9,
+ tex.texTlut[i&3].tlut_format,
+ ((tex.texMode0[i&3].min_filter & 3) != 0),
+ (tex.texMode1[i&3].max_lod + 0xf) / 0x10,
+ (tex.texImage1[i&3].image_type != 0));
- if (tentry)
- {
- // 0s are probably for no manual wrapping needed.
- PixelShaderManager::SetTexDims(i, tentry->native_width, tentry->native_height, 0, 0);
- }
- else
- ERROR_LOG(VIDEO, "error loading texture");
+ if (tentry)
+ {
+ // 0s are probably for no manual wrapping needed.
+ PixelShaderManager::SetTexDims(i, tentry->native_width, tentry->native_height, 0, 0);
}
+ else
+ ERROR_LOG(VIDEO, "error loading texture");
}
// set global constants
diff --git a/Source/Core/VideoCommon/VertexShaderManager.cpp b/Source/Core/VideoCommon/VertexShaderManager.cpp
index 022bf7683d..93f969b0b5 100644
--- a/Source/Core/VideoCommon/VertexShaderManager.cpp
+++ b/Source/Core/VideoCommon/VertexShaderManager.cpp
@@ -5,6 +5,7 @@
#include <cmath>
#include <sstream>
+#include "Common/BitSet.h"
#include "Common/CommonTypes.h"
#include "Common/MathUtil.h"
#include "VideoCommon/BPMemory.h"
@@ -22,7 +23,7 @@ static float GC_ALIGNED16(g_fProjectionMatrix[16]);
// track changes
static bool bTexMatricesChanged[2], bPosNormalMatrixChanged, bProjectionChanged, bViewportChanged;
-static int nMaterialsChanged;
+static BitSet32 nMaterialsChanged;
static int nTransformMatricesChanged[2]; // min,max
static int nNormalMatricesChanged[2]; // min,max
static int nPostTransformMatricesChanged[2]; // min,max
@@ -202,7 +203,7 @@ void VertexShaderManager::Dirty()
bProjectionChanged = true;
- nMaterialsChanged = 15;
+ nMaterialsChanged = BitSet32::AllTrue(4);
dirty = true;
}
@@ -295,35 +296,16 @@ void VertexShaderManager::SetConstants()
nLightsChanged[0] = nLightsChanged[1] = -1;
}
- if (nMaterialsChanged)
+ for (int i : nMaterialsChanged)
{
- for (int i = 0; i < 2; ++i)
- {
- if (nMaterialsChanged & (1 << i))
- {
- u32 data = xfmem.ambColor[i];
- constants.materials[i][0] = (data >> 24) & 0xFF;
- constants.materials[i][1] = (data >> 16) & 0xFF;
- constants.materials[i][2] = (data >> 8) & 0xFF;
- constants.materials[i][3] = data & 0xFF;
- }
- }
-
- for (int i = 0; i < 2; ++i)
- {
- if (nMaterialsChanged & (1 << (i + 2)))
- {
- u32 data = xfmem.matColor[i];
- constants.materials[i+2][0] = (data >> 24) & 0xFF;
- constants.materials[i+2][1] = (data >> 16) & 0xFF;
- constants.materials[i+2][2] = (data >> 8) & 0xFF;
- constants.materials[i+2][3] = data & 0xFF;
- }
- }
+ u32 data = i >= 2 ? xfmem.matColor[i - 2] : xfmem.ambColor[i];
+ constants.materials[i][0] = (data >> 24) & 0xFF;
+ constants.materials[i][1] = (data >> 16) & 0xFF;
+ constants.materials[i][2] = (data >> 8) & 0xFF;
+ constants.materials[i][3] = data & 0xFF;
dirty = true;
-
- nMaterialsChanged = 0;
}
+ nMaterialsChanged = BitSet32(0);
if (bPosNormalMatrixChanged)
{
@@ -660,7 +642,7 @@ void VertexShaderManager::SetProjectionChanged()
void VertexShaderManager::SetMaterialColorChanged(int index, u32 color)
{
- nMaterialsChanged |= (1 << index);
+ nMaterialsChanged[index] = true;
}
void VertexShaderManager::TranslateView(float x, float y, float z)