Added CPUID fetch assembly functions (x86 and x64)

build/VS2010/HEVC_encoder.vcxproj

@@ -48,6 +48,7 @@
   </PropertyGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
   <ImportGroup Label="ExtensionSettings">
+    <Import Project="..\..\..\..\..\yasm\vsyasm.props" />
   </ImportGroup>
   <ImportGroup Label="PropertySheets" Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
@@ -99,7 +100,7 @@
       </PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
-      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>WIN32;X64;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <CompileAs>CompileAsC</CompileAs>
     </ClCompile>
     <Link>
@@ -145,6 +146,7 @@
       <InlineFunctionExpansion>Default</InlineFunctionExpansion>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
       <AdditionalOptions>/MP %(AdditionalOptions)</AdditionalOptions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -179,8 +181,15 @@
     <ClInclude Include="..\..\src\nal.h" />
     <ClInclude Include="..\..\src\picture.h" />
     <ClInclude Include="..\..\src\transform.h" />
+    <ClInclude Include="..\..\src\x64\test64.h" />
+    <ClInclude Include="..\..\src\x86\test.h" />
+  </ItemGroup>
+  <ItemGroup>
+    <YASM Include="..\..\src\x64\test64.asm" />
+    <YASM Include="..\..\src\x86\test.asm" />
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
+    <Import Project="..\..\..\..\..\yasm\vsyasm.targets" />
   </ImportGroup>
 </Project>

build/VS2010/HEVC_encoder.vcxproj.filters

@@ -13,6 +13,18 @@
       <UniqueIdentifier>{67DA6AB6-F800-4c08-8B7A-83BB121AAD01}</UniqueIdentifier>
       <Extensions>rc;ico;cur;bmp;dlg;rc2;rct;bin;rgs;gif;jpg;jpeg;jpe;resx;tiff;tif;png;wav;mfcribbon-ms</Extensions>
     </Filter>
+    <Filter Include="Source Files\x86">
+      <UniqueIdentifier>{9e551421-7564-43be-a2b6-1ffb16d744a4}</UniqueIdentifier>
+    </Filter>
+    <Filter Include="Header Files\x86">
+      <UniqueIdentifier>{d1533c70-cb9a-419d-9a38-5161894ec359}</UniqueIdentifier>
+    </Filter>
+    <Filter Include="Header Files\x64">
+      <UniqueIdentifier>{cbe787b6-15a0-4c72-b658-6f444735fd73}</UniqueIdentifier>
+    </Filter>
+    <Filter Include="Source Files\x64">
+      <UniqueIdentifier>{39fb72d7-bfcd-4505-9a00-7ed6bf67d9ad}</UniqueIdentifier>
+    </Filter>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="..\..\src\encmain.c">
@@ -83,5 +95,19 @@
     <ClInclude Include="..\..\src\filter.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\src\x86\test.h">
+      <Filter>Header Files\x86</Filter>
+    </ClInclude>
+    <ClInclude Include="..\..\src\x64\test64.h">
+      <Filter>Header Files\x86</Filter>
+    </ClInclude>
+  </ItemGroup>
+  <ItemGroup>
+    <YASM Include="..\..\src\x86\test.asm">
+      <Filter>Source Files\x86</Filter>
+    </YASM>
+    <YASM Include="..\..\src\x64\test64.asm">
+      <Filter>Source Files\x64</Filter>
+    </YASM>
   </ItemGroup>
 </Project>

src/encmain.c

@@ -45,6 +45,12 @@
  #include "picture.h"
  #include "transform.h"
  
+ /* Assembly optimizations */
+#ifndef X64
+ #include "x86/test.h"
+#else
+ #include "x64/test64.h"
+#endif
  
  /*!
      \brief Program main function.
@@ -54,6 +60,8 @@
  */
   int main(int argc, char* argv[])
   {    
+    int ecx = 0,edx =0;
+    enum { BIT_SSSE3 = 9, BIT_SSE41 = 19, BIT_SSE42 = 20, BIT_MMX = 24, BIT_SSE = 25, BIT_SSE2 = 26};
     uint32_t curFrame = 0;
     config *cfg  = NULL;       /* Global configuration */
     FILE *input  = NULL;
@@ -64,6 +72,25 @@
     #endif
     encoder_control* encoder = (encoder_control*)malloc(sizeof(encoder_control));;
 
+    /* CPU id */
+    
+    printf("Checking for CPU features...\r\n");
+    #ifndef X64
+    cpuId(&ecx,&edx);
+    #else
+    cpuId64(&ecx,&edx);
+    #endif
+    //printf("CPUID ECX: %X EDX: %X\r\n", ecx, edx);
+    printf("CPU features enabled: ");
+    if(edx & (1<<BIT_MMX)) printf("MMX ");
+    if(edx & (1<<BIT_SSE)) printf("SSE ");
+    if(edx & (1<<BIT_SSE2)) printf("SSE2 ");
+    if(ecx & (1<<BIT_SSSE3)) printf("SSSE3 ");
+    if(ecx & (1<<BIT_SSE41)) printf("SSE4.1 ");
+    if(ecx & (1<<BIT_SSE42)) printf("SSE4.2 ");
+    printf("\r\n");
+    
+ 
     /* Handle configuration */
     cfg = config_alloc();
     

src/encoder.c

@@ -128,7 +128,7 @@ void initSigLastScan(uint32_t* pBuffD, uint32_t* pBuffH, uint32_t* pBuffV, int32
     {
       for(blkY=0; blkY < numBlkSide; blkY++)
       {
-        uint32_t offset    = blkY * 4 * iWidth + blkX * 4;
+        uint32_t offset = blkY * 4 * iWidth + blkX * 4;
         for(x=0; x < 4; x++)
         {
           for(y=0; y < 4; y++)
@@ -640,7 +640,7 @@ void encode_slice_data(encoder_control* encoder)
 
 void encode_coding_tree(encoder_control* encoder,uint16_t xCtb,uint16_t yCtb, uint8_t depth)
 {    
-  uint8_t split_flag = (depth<3)?1:0; /* ToDo: get from CU data */
+  uint8_t split_flag = (depth<1)?1:0; /* ToDo: get from CU data */
   uint8_t split_model = 0;
 
   /* Check for slice border */
@@ -707,7 +707,7 @@ void encode_coding_tree(encoder_control* encoder,uint16_t xCtb,uint16_t yCtb, ui
     if(cur_CU->type == CU_INTRA)
     {
       uint8_t intraPredMode = 1;
-      uint8_t intraPredModeChroma =36; /* 36 = Chroma derived from luma */
+      uint8_t intraPredModeChroma = 1; /* 36 = Chroma derived from luma */
       int8_t intraPreds[3] = {-1, -1, -1};
       int8_t mpmPred = -1;
       int i;
@@ -747,8 +747,14 @@ void encode_coding_tree(encoder_control* encoder,uint16_t xCtb,uint16_t yCtb, ui
         intra_DCPredFiltering(recShift,(LCU_WIDTH>>(depth))*2+8,pred,width,LCU_WIDTH>>depth,LCU_WIDTH>>depth);
       }
 
+      
       /* ToDo: separate chroma prediction(?) */
       /* intraPredModeChroma = 1; */
+
+      if(intraPredModeChroma != 36 && intraPredModeChroma == intraPredMode)
+      {
+        intraPredModeChroma = 36;
+      }
       intra_buildReferenceBorder(&encoder->in.cur_pic, xCtb, yCtb,(LCU_WIDTH>>(depth+1))*2+8, rec, (LCU_WIDTH>>(depth+1))*2+8, 1);
       intra_recon(recShiftU,(LCU_WIDTH>>(depth+1))*2+8,xCtb*(LCU_WIDTH>>(MAX_DEPTH+1)),yCtb*(LCU_WIDTH>>(MAX_DEPTH+1)),width>>1,predU,width>>1,intraPredModeChroma!=36?intraPredModeChroma:intraPredMode,1);
       intra_buildReferenceBorder(&encoder->in.cur_pic, xCtb, yCtb,(LCU_WIDTH>>(depth+1))*2+8, rec, (LCU_WIDTH>>(depth+1))*2+8, 2);

src/intra.c

@@ -290,7 +290,7 @@ int16_t intra_prediction(uint8_t* orig,int32_t origstride,int16_t* rec,int32_t r
     if(MIN(abs(i-26),abs(i-10)) <= threshold)
     {
       intra_getAngularPred(rec,recstride,pred, width,width,width,i, xpos?1:0, ypos?1:0, filter);
-      //CHECK_FOR_BEST(i);
+      CHECK_FOR_BEST(i);
     }
   } 
   

src/picture.h

@@ -34,7 +34,7 @@ enum { CU_NOTSET = 0,CU_PCM, CU_SKIP, CU_SPLIT, CU_INTRA, CU_INTER };
 #define SET_SPLITDATA(CU,flag) { (CU)->split=(flag); }
 
 /*!
-    \brief Struct for CU info
+    \brief Struct for CU intra info
 */
 typedef struct
 {
@@ -42,6 +42,16 @@ typedef struct
   uint32_t cost;
 } CU_info_intra;
 
+/*!
+    \brief Struct for CU inter info
+*/
+typedef struct
+{
+  uint8_t mode;
+  uint32_t cost;
+  int16_t mv[2];
+} CU_info_inter;
+
 
 /*!
     \brief Struct for CU info
@@ -49,7 +59,9 @@ typedef struct
 typedef struct
 {
   uint8_t type;
+  int8_t coded;
   CU_info_intra intra;
+  CU_info_inter inter;
   uint8_t split;
 } CU_info;
 

src/transform.c

@@ -249,10 +249,10 @@ void scalinglist_processDec( int32_t *coeff, int32_t *dequantcoeff, int32_t invQ
 
 void scalinglist_set(int32_t *coeff, uint32_t listId, uint32_t sizeId, uint32_t qp)
 {
-  uint32_t width = g_scalingListSizeX[sizeId];
+  uint32_t width  = g_scalingListSizeX[sizeId];
   uint32_t height = g_scalingListSizeX[sizeId];
-  uint32_t ratio = g_scalingListSizeX[sizeId]/MIN(8,g_scalingListSizeX[sizeId]);
+  uint32_t ratio  = g_scalingListSizeX[sizeId]/MIN(8,g_scalingListSizeX[sizeId]);
-  int32_t *quantcoeff = g_quant_coeff[sizeId][listId][qp];
+  int32_t *quantcoeff   = g_quant_coeff[sizeId][listId][qp];
   int32_t *dequantcoeff = g_de_quant_coeff[sizeId][listId][qp];
 
   scalinglist_processEnc(coeff,quantcoeff,g_quantScales[qp]<<4,height,width,ratio,MIN(8,g_scalingListSizeX[sizeId]),/*SCALING_LIST_DC*/16, 0);
@@ -667,13 +667,13 @@ void transform2d(int16_t *block,int16_t *coeff, int8_t blockSize, int32_t uiMode
   int16_t tmp[LCU_WIDTH*LCU_WIDTH];
   
   if(blockSize== 4)
-  {
+  {/*
     if (uiMode != 65535)
     {
       fastForwardDst(block,tmp,shift_1st); // Forward DST BY FAST ALGORITHM, block input, tmp output
       fastForwardDst(tmp,coeff,shift_2nd); // Forward DST BY FAST ALGORITHM, tmp input, coeff output
     }
-    else
+    else*/
     {
       partialButterfly4(block, tmp, shift_1st, blockSize);
       partialButterfly4(tmp, coeff, shift_2nd, blockSize);
@@ -718,13 +718,13 @@ void itransform2d(int16_t *block,int16_t *coeff, int8_t blockSize, int32_t uiMod
   int16_t tmp[LCU_WIDTH*LCU_WIDTH];
 
   if( blockSize == 4)
-  {
+  {/*
     if (uiMode != 65535)
     {
       fastInverseDst(coeff,tmp,shift_1st);    // Inverse DST by FAST Algorithm, coeff input, tmp output
       fastInverseDst(tmp,block,shift_2nd); // Inverse DST by FAST Algorithm, tmp input, coeff output
     }
-    else
+    else*/
     {
       partialButterflyInverse4(coeff,tmp,shift_1st,blockSize);
       partialButterflyInverse4(tmp,block,shift_2nd,blockSize);

src/x64/test64.asm

@@ -0,0 +1,15 @@
+; Function to get CPUID for identifying CPU capabilities
+bits 64
+section .code
+global cpuId64
+
+cpuId64:
+    mov  r8, rcx ; pointer to ecx-output
+    mov  r9, rdx ; pointer to edx-output
+
+    mov eax,1
+    cpuid
+    mov dword [r8], ecx
+    mov dword [r9], edx
+    mov eax,0
+    ret

src/x64/test64.h

@@ -0,0 +1,19 @@
+/**
+ *  Part of HEVC Encoder
+ *  By Marko Viitanen ( fador at iki.fi ), Tampere University of Technology, Department of Computer Systems.
+ */
+
+/*! \file test.h
+    \brief test header
+    \author Marko Viitanen
+    \date 2013-04
+    
+*/
+
+#ifndef _TEST64_H_
+#define _TEST64_H_
+
+void __cdecl cpuId64(int* ecx, int *edx );
+
+
+#endif

src/x86/test.asm

@@ -0,0 +1,13 @@
+; Function to get CPUID for identifying CPU capabilities
+bits 32
+global _cpuId
+
+_cpuId:    
+    mov eax,1
+    cpuid
+    mov eax, dword [esp+4]
+    mov dword [eax], ecx
+    mov eax, dword [esp+8]
+    mov dword [eax], edx
+    mov eax,0
+    ret

src/x86/test.h

@@ -0,0 +1,19 @@
+/**
+ *  Part of HEVC Encoder
+ *  By Marko Viitanen ( fador at iki.fi ), Tampere University of Technology, Department of Computer Systems.
+ */
+
+/*! \file test.h
+    \brief test header
+    \author Marko Viitanen
+    \date 2013-04
+    
+*/
+
+#ifndef _TEST_H_
+#define _TEST_H_
+
+void __cdecl cpuId(int* ecx, int *edx );
+
+
+#endif