summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--Source/Core/VideoCommon/Src/DLCache.cpp2
-rw-r--r--Source/Core/VideoCommon/Src/DataReader.h52
-rw-r--r--Source/Core/VideoCommon/Src/OpcodeDecoding.cpp30
-rw-r--r--Source/Core/VideoCommon/Src/VertexLoader_Position.cpp4
4 files changed, 78 insertions, 10 deletions
diff --git a/Source/Core/VideoCommon/Src/DLCache.cpp b/Source/Core/VideoCommon/Src/DLCache.cpp
index 4d76cfb290..8d1a9b829b 100644
--- a/Source/Core/VideoCommon/Src/DLCache.cpp
+++ b/Source/Core/VideoCommon/Src/DLCache.cpp
@@ -305,7 +305,6 @@ u8 AnalyzeAndRunDisplayList(u32 address, int size, CachedDisplayList *dl)
u32 Cmd2 = DataReadU32();
int transfer_size = ((Cmd2 >> 16) & 15) + 1;
u32 xf_address = Cmd2 & 0xFFFF;
- // TODO - speed this up. pshufb?
u32 data_buffer[16];
DataReadU32xFuncs[transfer_size-1](data_buffer);
LoadXFReg(transfer_size, xf_address, data_buffer);
@@ -453,7 +452,6 @@ bool CompileAndRunDisplayList(u32 address, int size, CachedDisplayList *dl)
u32 Cmd2 = DataReadU32();
int transfer_size = ((Cmd2 >> 16) & 15) + 1;
u32 xf_address = Cmd2 & 0xFFFF;
- // TODO - speed this up. pshufb?
ReferencedDataRegion* NewRegion = new ReferencedDataRegion;
NewRegion->MustClean = true;
NewRegion->size = transfer_size * 4;
diff --git a/Source/Core/VideoCommon/Src/DataReader.h b/Source/Core/VideoCommon/Src/DataReader.h
index 933de18849..fb9a6bbe60 100644
--- a/Source/Core/VideoCommon/Src/DataReader.h
+++ b/Source/Core/VideoCommon/Src/DataReader.h
@@ -15,11 +15,17 @@
// Official SVN repository and contact information can be found at
// http://code.google.com/p/dolphin-emu/
+
+
#ifndef _DATAREADER_H
#define _DATAREADER_H
extern u8* g_pVideoData;
+#if _M_SSE >= 0x301 && !(defined __GNUC__ && !defined __SSSE3__)
+#include <tmmintrin.h>
+#endif
+
__forceinline void DataSkip(u32 skip)
{
g_pVideoData += skip;
@@ -64,6 +70,49 @@ __forceinline u32 DataReadU32()
return tmp;
}
+typedef void (*DataReadU32xNfunc)(u32 *buf);
+extern DataReadU32xNfunc DataReadU32xFuncs[16];
+
+#if _M_SSE >= 0x301
+
+const __m128i mask1 = _mm_set_epi8(15,14,13,12,11,10,9,8,7,6,5,4,0,1,2,3);
+const __m128i mask2 = _mm_set_epi8(15,14,13,12,11,10,9,8,4,5,6,7,0,1,2,3);
+const __m128i mask3 = _mm_set_epi8(15,14,13,12,8,9,10,11,4,5,6,7,0,1,2,3);
+const __m128i mask4 = _mm_set_epi8(12,13,14,15,8,9,10,11,4,5,6,7,0,1,2,3);
+
+template<unsigned int N>
+void DataReadU32xN_SSSE3(u32 *bufx16)
+{
+ __m128i* store = (__m128i *)bufx16;
+ __m128i* load = (__m128i *)g_pVideoData;
+ switch(N)
+ {
+ case 13: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 9: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 5: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 1: _mm_storeu_si128(store, _mm_shuffle_epi8(_mm_loadu_si128(load), mask1));
+ break;
+ case 14: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 10: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 6: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 2: _mm_storeu_si128(store, _mm_shuffle_epi8(_mm_loadu_si128(load), mask2));
+ break;
+ case 15: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 11: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 7: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 3: _mm_storeu_si128(store, _mm_shuffle_epi8(_mm_loadu_si128(load), mask3));
+ break;
+ case 16: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 12: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 8: _mm_storeu_si128(store++, _mm_shuffle_epi8(_mm_loadu_si128(load++), mask4));
+ case 4: _mm_storeu_si128(store, _mm_shuffle_epi8(_mm_loadu_si128(load), mask4));
+ break;
+ }
+ g_pVideoData += (sizeof(u32) * N);
+}
+
+#endif
+
template<unsigned int N>
void DataReadU32xN(u32 *bufx16)
{
@@ -87,9 +136,6 @@ void DataReadU32xN(u32 *bufx16)
g_pVideoData += (sizeof(u32) * N);
}
-typedef void (*DataReadU32xNfunc)(u32 *buf);
-extern DataReadU32xNfunc DataReadU32xFuncs[16];
-
__forceinline u32 DataReadU32Unswapped()
{
u32 tmp = *(u32*)g_pVideoData;
diff --git a/Source/Core/VideoCommon/Src/OpcodeDecoding.cpp b/Source/Core/VideoCommon/Src/OpcodeDecoding.cpp
index ff4d6d1d97..3555f1fba1 100644
--- a/Source/Core/VideoCommon/Src/OpcodeDecoding.cpp
+++ b/Source/Core/VideoCommon/Src/OpcodeDecoding.cpp
@@ -30,6 +30,7 @@
#include "Profiler.h"
#include "OpcodeDecoding.h"
#include "CommandProcessor.h"
+#include "CPUDetect.h"
#include "VertexLoaderManager.h"
@@ -47,6 +48,27 @@
#include "VideoConfig.h"
u8* g_pVideoData = 0;
+#if _M_SSE >= 0x301
+DataReadU32xNfunc DataReadU32xFuncs_SSSE3[16] = {
+ DataReadU32xN_SSSE3<1>,
+ DataReadU32xN_SSSE3<2>,
+ DataReadU32xN_SSSE3<3>,
+ DataReadU32xN_SSSE3<4>,
+ DataReadU32xN_SSSE3<5>,
+ DataReadU32xN_SSSE3<6>,
+ DataReadU32xN_SSSE3<7>,
+ DataReadU32xN_SSSE3<8>,
+ DataReadU32xN_SSSE3<9>,
+ DataReadU32xN_SSSE3<10>,
+ DataReadU32xN_SSSE3<11>,
+ DataReadU32xN_SSSE3<12>,
+ DataReadU32xN_SSSE3<13>,
+ DataReadU32xN_SSSE3<14>,
+ DataReadU32xN_SSSE3<15>,
+ DataReadU32xN_SSSE3<16>
+};
+#endif
+
DataReadU32xNfunc DataReadU32xFuncs[16] = {
DataReadU32xN<1>,
DataReadU32xN<2>,
@@ -250,7 +272,6 @@ static void Decode()
u32 Cmd2 = DataReadU32();
int transfer_size = ((Cmd2 >> 16) & 15) + 1;
u32 xf_address = Cmd2 & 0xFFFF;
- // TODO - speed this up. pshufb?
u32 data_buffer[16];
DataReadU32xFuncs[transfer_size-1](data_buffer);
@@ -401,6 +422,13 @@ void OpcodeDecoder_Init()
{
g_pVideoData = FAKE_GetFifoStartPtr();
+#if _M_SSE >= 0x301
+ if (cpu_info.bSSSE3)
+ {
+ *DataReadU32xFuncs = *DataReadU32xFuncs_SSSE3;
+ }
+#endif
+
if (g_Config.bEnableOpenCL)
{
OpenCL::Initialize();
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
index 6f5d7e680c..b8483b119c 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
@@ -22,10 +22,6 @@
#include "VertexManagerBase.h"
#include "CPUDetect.h"
-#if _M_SSE >= 0x301 && !(defined __GNUC__ && !defined __SSSE3__)
-#include <tmmintrin.h>
-#endif
-
extern float posScale;
extern TVtxAttr *pVtxAttr;
generated by cgit v1.2.3 (git 2.25.1) at 2026-09-07 16:29:33 +0000