x86: don't use the red zone on win64

[x264] / common / x86 / dct.h

diff --git a/common/x86/dct.h b/common/x86/dct.h
index 030ce76ecadcf6b61e32650da660013f5b713ba8..a16cc2db8dcecad90a590cc3b41bd5fb44b67f2e 100644 (file)
--- a/common/x86/dct.h
+++ b/common/x86/dct.h
@@ -1,7 +1,7 @@
 /*****************************************************************************
  * dct.h: x86 transform and zigzag
  *****************************************************************************
- * Copyright (C) 2003-2011 x264 project
+ * Copyright (C) 2003-2013 x264 project
  *
  * Authors: Loren Merritt <lorenm@u.washington.edu>
  *          Laurent Aimar <fenrir@via.ecp.fr>
@@ -38,17 +38,22 @@ void x264_sub8x8_dct_ssse3  ( int16_t dct[ 4][16], uint8_t *pix1, uint8_t *pix2
 void x264_sub16x16_dct_ssse3( int16_t dct[16][16], uint8_t *pix1, uint8_t *pix2 );
 void x264_sub8x8_dct_avx    ( int16_t dct[ 4][16], uint8_t *pix1, uint8_t *pix2 );
 void x264_sub16x16_dct_avx  ( int16_t dct[16][16], uint8_t *pix1, uint8_t *pix2 );
+void x264_sub8x8_dct_xop    ( int16_t dct[ 4][16], uint8_t *pix1, uint8_t *pix2 );
+void x264_sub16x16_dct_xop  ( int16_t dct[16][16], uint8_t *pix1, uint8_t *pix2 );
 void x264_sub8x8_dct_dc_mmx2( int16_t dct    [ 4], uint8_t *pix1, uint8_t *pix2 );
-void x264_sub8x8_dct_dc_sse2( int16_t dct    [ 4], uint8_t *pix1, uint8_t *pix2 );
+void x264_sub8x8_dct_dc_sse2( dctcoef dct    [ 4], pixel   *pix1, pixel   *pix2 );
+void x264_sub8x16_dct_dc_sse2 ( dctcoef dct  [ 4], pixel   *pix1, pixel   *pix2 );
+void x264_sub8x16_dct_dc_ssse3( int16_t dct  [ 4], uint8_t *pix1, uint8_t *pix2 );
+void x264_sub8x16_dct_dc_avx  ( dctcoef dct  [ 4], pixel   *pix1, pixel   *pix2 );
 
 void x264_add4x4_idct_mmx       ( uint8_t *p_dst, int16_t dct    [16] );
 void x264_add4x4_idct_sse2     ( uint16_t *p_dst, int32_t dct    [16] );
 void x264_add4x4_idct_sse4      ( uint8_t *p_dst, int16_t dct    [16] );
 void x264_add4x4_idct_avx       ( pixel   *p_dst, dctcoef dct    [16] );
 void x264_add8x8_idct_mmx       ( uint8_t *p_dst, int16_t dct[ 4][16] );
-void x264_add8x8_idct_dc_mmx    ( uint8_t *p_dst, int16_t dct    [ 4] );
+void x264_add8x8_idct_dc_mmx2   ( uint8_t *p_dst, int16_t dct    [ 4] );
 void x264_add16x16_idct_mmx     ( uint8_t *p_dst, int16_t dct[16][16] );
-void x264_add16x16_idct_dc_mmx  ( uint8_t *p_dst, int16_t dct    [16] );
+void x264_add16x16_idct_dc_mmx2 ( uint8_t *p_dst, int16_t dct    [16] );
 void x264_add8x8_idct_sse2      ( pixel   *p_dst, dctcoef dct[ 4][16] );
 void x264_add8x8_idct_avx       ( pixel   *p_dst, dctcoef dct[ 4][16] );
 void x264_add16x16_idct_sse2    ( pixel   *p_dst, dctcoef dct[16][16] );
@@ -69,31 +74,36 @@ void x264_idct4x4dc_avx      ( int32_t d[16] );
 
 void x264_sub8x8_dct8_mmx    ( int16_t dct   [64], uint8_t *pix1, uint8_t *pix2 );
 void x264_sub16x16_dct8_mmx  ( int16_t dct[4][64], uint8_t *pix1, uint8_t *pix2 );
-void x264_sub8x8_dct8_sse2   ( int16_t dct   [64], uint8_t *pix1, uint8_t *pix2 );
-void x264_sub16x16_dct8_sse2 ( int16_t dct[4][64], uint8_t *pix1, uint8_t *pix2 );
+void x264_sub8x8_dct8_sse2   ( dctcoef dct   [64], pixel *pix1, pixel *pix2 );
+void x264_sub16x16_dct8_sse2 ( dctcoef dct[4][64], pixel *pix1, pixel *pix2 );
 void x264_sub8x8_dct8_ssse3  ( int16_t dct   [64], uint8_t *pix1, uint8_t *pix2 );
 void x264_sub16x16_dct8_ssse3( int16_t dct[4][64], uint8_t *pix1, uint8_t *pix2 );
-void x264_sub8x8_dct8_avx    ( int16_t dct   [64], uint8_t *pix1, uint8_t *pix2 );
-void x264_sub16x16_dct8_avx  ( int16_t dct[4][64], uint8_t *pix1, uint8_t *pix2 );
+void x264_sub8x8_dct8_sse4   ( int32_t dct   [64], uint16_t *pix1, uint16_t *pix2 );
+void x264_sub16x16_dct8_sse4 ( int32_t dct[4][64], uint16_t *pix1, uint16_t *pix2 );
+void x264_sub8x8_dct8_avx    ( dctcoef dct   [64], pixel *pix1, pixel *pix2 );
+void x264_sub16x16_dct8_avx  ( dctcoef dct[4][64], pixel *pix1, pixel *pix2 );
 
 
 void x264_add8x8_idct8_mmx   ( uint8_t *dst, int16_t dct   [64] );
 void x264_add16x16_idct8_mmx ( uint8_t *dst, int16_t dct[4][64] );
-void x264_add8x8_idct8_sse2  ( uint8_t *dst, int16_t dct   [64] );
-void x264_add16x16_idct8_sse2( uint8_t *dst, int16_t dct[4][64] );
-void x264_add8x8_idct8_avx   ( uint8_t *dst, int16_t dct   [64] );
-void x264_add16x16_idct8_avx ( uint8_t *dst, int16_t dct[4][64] );
+void x264_add8x8_idct8_sse2  ( pixel *dst, dctcoef dct   [64] );
+void x264_add16x16_idct8_sse2( pixel *dst, dctcoef dct[4][64] );
+void x264_add8x8_idct8_avx   ( pixel *dst, dctcoef dct   [64] );
+void x264_add16x16_idct8_avx ( pixel *dst, dctcoef dct[4][64] );
 
+void x264_zigzag_scan_8x8_frame_xop  ( int16_t level[64], int16_t dct[64] );
 void x264_zigzag_scan_8x8_frame_avx  ( dctcoef level[64], dctcoef dct[64] );
 void x264_zigzag_scan_8x8_frame_ssse3( int16_t level[64], int16_t dct[64] );
 void x264_zigzag_scan_8x8_frame_sse2 ( dctcoef level[64], dctcoef dct[64] );
 void x264_zigzag_scan_8x8_frame_mmx2 ( int16_t level[64], int16_t dct[64] );
+void x264_zigzag_scan_4x4_frame_xop  ( dctcoef level[16], dctcoef dct[16] );
 void x264_zigzag_scan_4x4_frame_avx  ( dctcoef level[16], dctcoef dct[16] );
 void x264_zigzag_scan_4x4_frame_ssse3( int16_t level[16], int16_t dct[16] );
 void x264_zigzag_scan_4x4_frame_sse2 ( int32_t level[16], int32_t dct[16] );
 void x264_zigzag_scan_4x4_frame_mmx  ( int16_t level[16], int16_t dct[16] );
 void x264_zigzag_scan_4x4_field_sse2 ( int32_t level[16], int32_t dct[16] );
 void x264_zigzag_scan_4x4_field_mmx2 ( int16_t level[16], int16_t dct[16] );
+void x264_zigzag_scan_8x8_field_xop  ( int16_t level[64], int16_t dct[64] );
 void x264_zigzag_scan_8x8_field_avx  ( int32_t level[64], int32_t dct[64] );
 void x264_zigzag_scan_8x8_field_sse4 ( int32_t level[64], int32_t dct[64] );
 void x264_zigzag_scan_8x8_field_mmx2 ( int16_t level[64], int16_t dct[64] );