actions/itgmania212121
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Specialize on aligned loads for sse since misaligned loads are much slower.

Browse Source
This commit is contained in:
Steve Checkoway
2006-07-17 00:42:46 +00:00
parent c5635656ef
commit bd7422ef6e
2 changed files with 32 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
stepmania/src/archutils/Darwin/VectorHelper.cpp
+33 -22
View File
@@ -3,7 +3,7 @@
#include "RageUtil.h" #include "RageUtil.h"
#include <sys/sysctl.h> #include <sys/sysctl.h>
#ifdef USE_VEC #if defined(USE_VEC)
#if defined(__VEC__) #if defined(__VEC__)
#include <vecLib/vecLib.h> #include <vecLib/vecLib.h>
#ifndef __VECLIBTYPES__ #ifndef __VECLIBTYPES__
@@ -398,9 +398,9 @@ void Vector::FastSoundRead( float *dest, const int32_t *src, unsigned size )
vec_ste( result, 0, dest++ ); vec_ste( result, 0, dest++ );
} }
} }
#elif defined(__SSE__) #elif defined(__SSE2__)
#include <xmmintrin.h> #include <xmmintrin.h>
// This might even be portable to other sysems since it uses Intel's intrinsics. // This is portable to other sysems since it uses Intel's intrinsics.
bool Vector::CheckForVector() bool Vector::CheckForVector()
{ {
@@ -408,6 +408,31 @@ bool Vector::CheckForVector()
return true; return true;
} }
template<typename load>
static inline void Write( int32_t *&dest, const int16_t *&src,
unsigned &size, __m128i vol ) __attribute__((always_inline));
template<typename T>
inline void Write( T load, int32_t *&dest, const int16_t *&src, unsigned &size, __m128i vol )
{
// There are only 8 XMM registers so no 4x unrolling
while( size >= 8 )
{
__m128i data = load( (__m128i *)src );
__m128i hi = _mm_mulhi_epi16( data, vol );
__m128i low = _mm_mullo_epi16( data, vol );
__m128i result1 = _mm_unpacklo_epi16( low, hi );
__m128i result2 = _mm_unpackhi_epi16( low, hi );
result1 = _mm_add_epi32( result1, *(__m128i *)(dest + 0) );
result2 = _mm_add_epi32( result2, *(__m128i *)(dest + 4) );
_mm_store_si128( (__m128i *)(dest + 0), result1 );
_mm_store_si128( (__m128i *)(dest + 4), result2 );
src += 8;
dest += 8;
size -= 8;
}
}
void Vector::FastSoundWrite( int32_t *dest, const int16_t *src, unsigned size, short volume ) void Vector::FastSoundWrite( int32_t *dest, const int16_t *src, unsigned size, short volume )
{ {
if( size == 0 ) if( size == 0 )
@@ -420,26 +445,12 @@ void Vector::FastSoundWrite( int32_t *dest, const int16_t *src, unsigned size, s
--size; --size;
} }
// There are only 8 XMM registers so no 4x unrolling
__m128i vol = _mm_set1_epi16( volume ); __m128i vol = _mm_set1_epi16( volume );
// Misaligned loads are slower so specialize to aligned loads when possible.
while( size >= 8 ) if( intptr_t(src) & 0xF )
{ Write( _mm_loadu_si128, dest, src, size, vol );
// Aligned stores, possibly misaligned loads. else
__m128i data = _mm_loadu_si128( (__m128i *)src ); Write( _mm_load_si128, dest, src, size, vol );
__m128i hi = _mm_mulhi_epi16( data, vol );
__m128i low = _mm_mullo_epi16( data, vol );
__m128i result1 = _mm_unpacklo_epi16( low, hi );
__m128i result2 = _mm_unpackhi_epi16( low, hi );
result1 = _mm_add_epi32( result1, *(__m128i *)(dest + 0) );
result2 = _mm_add_epi32( result2, *(__m128i *)(dest + 4) );
_mm_store_si128( (__m128i *)(dest + 0), result1 );
_mm_store_si128( (__m128i *)(dest + 4), result2 );
src += 8;
dest += 8;
size -= 8;
}
while( size-- ) while( size-- )
*(dest++) += *(src++) * volume; *(dest++) += *(src++) * volume;
} }
stepmania/src/archutils/Darwin/VectorHelper.h
+1 -1
View File
@@ -1,7 +1,7 @@
#ifndef VECTOR_HELPER_H #ifndef VECTOR_HELPER_H
#define VECTOR_HELPER_H #define VECTOR_HELPER_H
#if defined(__VEC__) || defined(__SSE__) #if ( defined(__VEC__) || (defined(__SSE__) && defined(__SSE2__)) ) && defined(__GNUC__)
namespace Vector namespace Vector
{ {
bool CheckForVector(); bool CheckForVector();