Fix some code and remove old SSE code

This commit is contained in:
2018-10-14 22:13:41 +01:00
parent 8408b7ba2b
commit a473abc498
4 changed files with 6 additions and 159 deletions
+2 -2
View File
@@ -660,8 +660,8 @@ void VideoHandle::playEvent()
{
bool bEndOfSequence = false;
#if( _CONTROL_PLAYING_TIME_ == 1 )
m_dAverageFps = Double( m_dAverageFps * m_uiNumberPlayedFrames + m_pcPlayControlTimer->elapsed() ) /
Double( m_uiNumberPlayedFrames + 1 );
m_dAverageFps = double( m_dAverageFps * m_uiNumberPlayedFrames + m_pcPlayControlTimer->elapsed() ) /
double( m_uiNumberPlayedFrames + 1 );
m_uiNumberPlayedFrames++;
m_pcPlayControlTimer->restart();
#endif
+2 -2
View File
@@ -34,7 +34,7 @@
#include <QVector>
#include <QWidget>
#define _CONTROL_PLAYING_TIME_ 0
#define _CONTROL_PLAYING_TIME_ 1
class QToolBar;
class QDockWidget;
@@ -124,7 +124,7 @@ private:
#if( _CONTROL_PLAYING_TIME_ == 1 )
unsigned int m_uiNumberPlayedFrames;
Double m_dAverageFps;
double m_dAverageFps;
QElapsedTimer* m_pcPlayControlTimer;
#endif
+2 -1
View File
@@ -669,6 +669,7 @@ void CalypFrame::frameToBuffer( ClpByte* output_buffer, int iEndianness )
}
}
void CalypFrame::fillRGBBuffer()
{
#define PEL_ARGB( a, r, g, b ) ( ( a & 0xff ) << 24 ) | ( ( r & 0xff ) << 16 ) | ( ( g & 0xff ) << 8 ) | ( b & 0xff )
@@ -694,7 +695,7 @@ void CalypFrame::fillRGBBuffer()
unsigned char finalPel;
for( unsigned int i = 0; i < d->m_uiHeight * d->m_uiWidth; i++ )
{
finalPel = (*pY++) >> shiftBits;
finalPel = ( *pY++ ) >> shiftBits;
*pARGB++ = PEL_RGB( finalPel, finalPel, finalPel );
}
}
-154
View File
@@ -30,11 +30,6 @@
#ifdef USE_FFMPEG
#include "StreamHandlerLibav.h"
#endif
#ifdef USE_SSE
#include <emmintrin.h>
#include <limits.h>
#include <smmintrin.h>
#endif
/*
**************************************************************
@@ -49,155 +44,6 @@
#define ADD_FFMPEG_PEL_FMT( fmt ) 0
#endif
#if 0
void fillARGB32bufferYUV420p( Pel*** in, UChar* out, unsigned int width, unsigned int height, unsigned int depth )
{
#ifdef USE_SSE
if( ( width % 16 ) == 0 )
{
uint8_t *yp = in[LUMA][0];
uint8_t *up = in[CHROMA_U][0];
uint8_t *vp = in[CHROMA_V][0];
uint32_t suv = width >> 1;
uint32_t *rgb = ( uint32_t* )out;
__m128i y0r0, y0r1, u0, v0;
__m128i y00r0, y01r0, y00r1, y01r1;
__m128i u00, u01, v00, v01;
__m128i rv00, rv01, gu00, gu01, gv00, gv01, bu00, bu01;
__m128i r00, r01, g00, g01, b00, b01;
__m128i rgb0123, rgb4567, rgb89ab, rgbcdef;
__m128i gbgb;
__m128i ysub, uvsub;
__m128i zero, facy, facrv, facgu, facgv, facbu;
__m128i *srcy128r0, *srcy128r1;
__m128i *dstrgb128r0, *dstrgb128r1;
__m64 *srcu64, *srcv64;
unsigned int x, y;
ysub = _mm_set1_epi32( 0x00100010 );
uvsub = _mm_set1_epi32( 0x00800080 );
facy = _mm_set1_epi32( 0x004a004a );
facrv = _mm_set1_epi32( 0x00660066 );
facgu = _mm_set1_epi32( 0x00190019 );
facgv = _mm_set1_epi32( 0x00340034 );
facbu = _mm_set1_epi32( 0x00810081 );
zero = _mm_set1_epi32( 0x00000000 );
for( y = 0; y < height; y += 2 )
{
srcy128r0 = ( __m128i * )( yp + width * y );
srcy128r1 = ( __m128i * )( yp + width * y + width );
srcu64 = ( __m64 * )( up + suv * ( y / 2 ) );
srcv64 = ( __m64 * )( vp + suv * ( y / 2 ) );
dstrgb128r0 = ( __m128i * )( rgb + width * y );
dstrgb128r1 = ( __m128i * )( rgb + width * y + width );
for( x = 0; x < width; x += 16 )
{
u0 = _mm_loadl_epi64( ( __m128i * )srcu64 );
srcu64++;
v0 = _mm_loadl_epi64( ( __m128i * )srcv64 );
srcv64++;
y0r0 = _mm_load_si128( srcy128r0++ );
y0r1 = _mm_load_si128( srcy128r1++ );
// constant y factors
y00r0 = _mm_mullo_epi16( _mm_sub_epi16( _mm_unpacklo_epi8( y0r0, zero ), ysub ), facy );
y01r0 = _mm_mullo_epi16( _mm_sub_epi16( _mm_unpackhi_epi8( y0r0, zero ), ysub ), facy );
y00r1 = _mm_mullo_epi16( _mm_sub_epi16( _mm_unpacklo_epi8( y0r1, zero ), ysub ), facy );
y01r1 = _mm_mullo_epi16( _mm_sub_epi16( _mm_unpackhi_epi8( y0r1, zero ), ysub ), facy );
// expand u and v so they're aligned with y values
u0 = _mm_unpacklo_epi8( u0, zero );
u00 = _mm_sub_epi16( _mm_unpacklo_epi16( u0, u0 ), uvsub );
u01 = _mm_sub_epi16( _mm_unpackhi_epi16( u0, u0 ), uvsub );
v0 = _mm_unpacklo_epi8( v0, zero );
v00 = _mm_sub_epi16( _mm_unpacklo_epi16( v0, v0 ), uvsub );
v01 = _mm_sub_epi16( _mm_unpackhi_epi16( v0, v0 ), uvsub );
// common factors on both rows.
rv00 = _mm_mullo_epi16( facrv, v00 );
rv01 = _mm_mullo_epi16( facrv, v01 );
gu00 = _mm_mullo_epi16( facgu, u00 );
gu01 = _mm_mullo_epi16( facgu, u01 );
gv00 = _mm_mullo_epi16( facgv, v00 );
gv01 = _mm_mullo_epi16( facgv, v01 );
bu00 = _mm_mullo_epi16( facbu, u00 );
bu01 = _mm_mullo_epi16( facbu, u01 );
// row 0
r00 = _mm_srai_epi16( _mm_add_epi16( y00r0, rv00 ), 6 );
r01 = _mm_srai_epi16( _mm_add_epi16( y01r0, rv01 ), 6 );
g00 = _mm_srai_epi16( _mm_sub_epi16( _mm_sub_epi16( y00r0, gu00 ), gv00 ), 6 );
g01 = _mm_srai_epi16( _mm_sub_epi16( _mm_sub_epi16( y01r0, gu01 ), gv01 ), 6 );
b00 = _mm_srai_epi16( _mm_add_epi16( y00r0, bu00 ), 6 );
b01 = _mm_srai_epi16( _mm_add_epi16( y01r0, bu01 ), 6 );
r00 = _mm_packus_epi16( r00, r01 );// rrrr.. saturated
g00 = _mm_packus_epi16( g00, g01 );// gggg.. saturated
b00 = _mm_packus_epi16( b00, b01 );// bbbb.. saturated
r01 = _mm_unpacklo_epi8( r00, zero );// 0r0r..
gbgb = _mm_unpacklo_epi8( b00, g00 );// gbgb..
rgb0123 = _mm_unpacklo_epi16( gbgb, r01 );// 0rgb0rgb..
rgb4567 = _mm_unpackhi_epi16( gbgb, r01 );// 0rgb0rgb..
r01 = _mm_unpackhi_epi8( r00, zero );
gbgb = _mm_unpackhi_epi8( b00, g00 );
rgb89ab = _mm_unpacklo_epi16( gbgb, r01 );
rgbcdef = _mm_unpackhi_epi16( gbgb, r01 );
_mm_store_si128( dstrgb128r0++, rgb0123 );
_mm_store_si128( dstrgb128r0++, rgb4567 );
_mm_store_si128( dstrgb128r0++, rgb89ab );
_mm_store_si128( dstrgb128r0++, rgbcdef );
// row 1
r00 = _mm_srai_epi16( _mm_add_epi16( y00r1, rv00 ), 6 );
r01 = _mm_srai_epi16( _mm_add_epi16( y01r1, rv01 ), 6 );
g00 = _mm_srai_epi16( _mm_sub_epi16( _mm_sub_epi16( y00r1, gu00 ), gv00 ), 6 );
g01 = _mm_srai_epi16( _mm_sub_epi16( _mm_sub_epi16( y01r1, gu01 ), gv01 ), 6 );
b00 = _mm_srai_epi16( _mm_add_epi16( y00r1, bu00 ), 6 );
b01 = _mm_srai_epi16( _mm_add_epi16( y01r1, bu01 ), 6 );
r00 = _mm_packus_epi16( r00, r01 );// rrrr.. saturated
g00 = _mm_packus_epi16( g00, g01 );// gggg.. saturated
b00 = _mm_packus_epi16( b00, b01 );// bbbb.. saturated
r01 = _mm_unpacklo_epi8( r00, zero );// 0r0r..
gbgb = _mm_unpacklo_epi8( b00, g00 );// gbgb..
rgb0123 = _mm_unpacklo_epi16( gbgb, r01 );// 0rgb0rgb..
rgb4567 = _mm_unpackhi_epi16( gbgb, r01 );// 0rgb0rgb..
r01 = _mm_unpackhi_epi8( r00, zero );
gbgb = _mm_unpackhi_epi8( b00, g00 );
rgb89ab = _mm_unpacklo_epi16( gbgb, r01 );
rgbcdef = _mm_unpackhi_epi16( gbgb, r01 );
_mm_store_si128( dstrgb128r1++, rgb0123 );
_mm_store_si128( dstrgb128r1++, rgb4567 );
_mm_store_si128( dstrgb128r1++, rgb89ab );
_mm_store_si128( dstrgb128r1++, rgbcdef );
}
}
}
else
#endif
{
fillARGB32bufferYUV( in, out, width, height, depth, 1, 1 );
}
}
#endif
const std::map<int, CalypPixelFormatDescriptor> g_CalypPixFmtDescriptorsMap = {
{
CLP_YUV420P,