modify simd switches

Author: taqu

README.md

@@ -57,13 +57,21 @@ An implementaion of adaptive range coder for C++
 |xargs.1|0.628815|13|15|
 
 # Building
-For Visual Studio 15 2017,
+First, make sure you clone this repository.  
+Then, deflate cantrbry.tar.bz2. 
+If you work on Windows, you're going to need to deflate zlib.tar.bz2 and lz4.tar.bz2.
+
 ```
 git clone https://github.com/taqu/cpprcoder.git
 cd cpprcoder/test
-mkdir build_simd
-pushd build_simd
-cmake -G"Visual Studio 15 2017 Win64" -DUSE_SIMD=1 ..
+tar jxvf cantrbry.tar.bz2
+```
+
+Next,
+```
+mkdir build
+pushd build
+cmake -DUSE_SIMD=1 ..
 ```
 
 # License

cpprcoder.h

@@ -109,12 +109,6 @@ namespace cpprcoder
 
 #ifndef CPPRCODER_MAXRANGE_DECODE
 #define CPPRCODER_MAXRANGE_DECODE (0x00FFFFFFU)
-#endif
-
-#if defined(__AVX__)
-#define CPPRCODER_USE_AVX (1)
-#elif defined(__SSE2__)
-#define CPPRCODER_USE_SSE2 (1)
 #endif
 
     enum Status
@@ -277,7 +271,7 @@ namespace cpprcoder
         void countChunks();
 
         u32 total_;
-#if defined(CPPRCODER_USE_SSE2) || defined(CPPRCODER_USE_AVX)
+#if defined(CPPRCODER_USE_SIMD)
         u32* prefix_;
         u32* frequencies_;
         u32 prefix_buffer_[CHUNKS + 4];
@@ -404,12 +398,7 @@ namespace cpprcoder
 
 #ifdef CPPRCODER_IMPLEMENTATION
 
-#ifdef CPPRCODER_USE_SSE2
-#include <immintrin.h>
-#include <emmintrin.h>
-#endif
-
-#ifdef CPPRCODER_USE_AVX
+#ifdef CPPRCODER_USE_SIMD
 #include <immintrin.h>
 #endif
 
@@ -656,7 +645,7 @@ namespace cpprcoder
     FrequencyTable::FrequencyTable()
         :total_(SIZE)
     {
-#if defined(CPPRCODER_USE_SSE2) || defined(CPPRCODER_USE_AVX)
+#if defined(CPPRCODER_USE_SIMD)
         prefix_ = reinterpret_cast<u32*>((reinterpret_cast<uintptr_t>(prefix_buffer_) + ALIGN_MASK) & ~ALIGN_MASK);
         frequencies_ = reinterpret_cast<u32*>((reinterpret_cast<uintptr_t>(frequencies_buffer_) + ALIGN_MASK) & ~ALIGN_MASK);
 
@@ -695,7 +684,7 @@ namespace cpprcoder
 
     void FrequencyTable::update(u8 b)
     {
-#if defined(CPPRCODER_USE_SSE2) || defined(CPPRCODER_USE_AVX)
+#if defined(CPPRCODER_USE_SIMD)
         ++frequencies_[b];
         if(CPPRCODER_MINRANGE<=(++total_)){
             __m128i one = _mm_set1_epi32(1);
@@ -766,7 +755,7 @@ namespace cpprcoder
 
     void FrequencyTable::find(u32& count, u8& code, u32 target) const
     {
-#if 0 //#if defined(CPPRCODER_USE_SSE2) || defined(CPPRCODER_USE_AVX)
+#if 0 //#if defined(CPPRCODER_USE_SIMD)
         __m128i target128 = _mm_set1_epi32(target);
         __m128i one = _mm_set1_epi32(1);
 
@@ -855,7 +844,7 @@ namespace cpprcoder
 
     void BinaryIndexedTree::initialize(u32 value)
     {
-#if defined(CPPRCODER_USE_SSE2) || defined(CPPRCODER_USE_AVX)
+#if defined(CPPRCODER_USE_SIMD)
         __m128i one = _mm_set1_epi32(0);
         __m128i* p = reinterpret_cast<__m128i*>(frequencies_);
         __m128i* end = p + (SIZE/4);

test/CMakeLists.txt

@@ -25,7 +25,7 @@ add_executable(${ProjectName} ${FILES})
 set(CXX_FLAGS_SIMD "")
 if(MSVC)
     if(USE_SIMD)
-        set(CXX_FLAGS_SIMD "/arch:AVX")
+        set(CXX_FLAGS_SIMD "/arch:AVX /DCPPRCODER_USE_SIMD")
     endif()
 
     set(DEFAULT_CXX_FLAGS "/DWIN32 /D_WINDOWS /D_MBCS /W4 /WX- /nologo /fp:precise /Zc:wchar_t /TP /Gd ${CXX_FLAGS_SIMD}")
@@ -41,7 +41,7 @@ if(MSVC)
 
 elseif(UNIX)
     if(USE_SIMD)
-        set(CXX_FLAGS_SIMD "-march=native")
+        set(CXX_FLAGS_SIMD "-march=native -DCPPRCODER_USE_SIMD")
     endif()
     set(DEFAULT_CXX_FLAGS "-Wall -O2 -std=c++11 ${CXX_FLAGS_SIMD}")
     set(CMAKE_CXX_FLAGS "${DEFAULT_CXX_FLAGS}")