Discussion:
Correctly align buffers for AVX and AVX-512
(too old to reply)
Henrik Gramner
2018-01-18 20:22:54 UTC
Permalink
Raw Message
x264 | branch: master | Henrik Gramner <***@gramner.com> | Mon Jan 15 21:42:59 2018 +0100| [9384a7389b251b59a079ccc3d1af9edd42e3d5e6] | committer: Henrik Gramner

Correctly align buffers for AVX and AVX-512

Fixes segfaults on Windows where the stack is only 16-byte aligned.
http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=9384a7389b251b59a079ccc3d1af9edd42e3d5e6
---

common/common.h | 8 ++++----
encoder/analyse.c | 4 ++--
encoder/ratecontrol.c | 2 +-
3 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/common/common.h b/common/common.h
index 27a56fbd..84d39ce3 100644
--- a/common/common.h
+++ b/common/common.h
@@ -569,16 +569,16 @@ struct x264_t
ALIGNED_64( pixel fdec_buf[54*FDEC_STRIDE] );

/* i4x4 and i8x8 backup data, for skipping the encode stage when possible */
- ALIGNED_16( pixel i4x4_fdec_buf[16*16] );
- ALIGNED_16( pixel i8x8_fdec_buf[16*16] );
+ ALIGNED_32( pixel i4x4_fdec_buf[16*16] );
+ ALIGNED_32( pixel i8x8_fdec_buf[16*16] );
ALIGNED_64( dctcoef i8x8_dct_buf[3][64] );
ALIGNED_64( dctcoef i4x4_dct_buf[15][16] );
uint32_t i4x4_nnz_buf[4];
uint32_t i8x8_nnz_buf[4];

/* Psy trellis DCT data */
- ALIGNED_16( dctcoef fenc_dct8[4][64] );
- ALIGNED_16( dctcoef fenc_dct4[16][16] );
+ ALIGNED_64( dctcoef fenc_dct8[4][64] );
+ ALIGNED_64( dctcoef fenc_dct4[16][16] );

/* Psy RD SATD/SA8D scores cache */
ALIGNED_64( uint32_t fenc_satd_cache[32] );
diff --git a/encoder/analyse.c b/encoder/analyse.c
index ecd6dae1..3577e5bc 100644
--- a/encoder/analyse.c
+++ b/encoder/analyse.c
@@ -558,7 +558,7 @@ static ALWAYS_INLINE const int8_t *predict_4x4_mode_available( int force_intra,
/* For trellis=2, we need to do this for both sizes of DCT, for trellis=1 we only need to use it on the chosen mode. */
static void inline psy_trellis_init( x264_t *h, int do_both_dct )
{
- ALIGNED_16( static pixel zero[16*FDEC_STRIDE] ) = {0};
+ ALIGNED_64( static pixel zero[16*FDEC_STRIDE] ) = {0};

if( do_both_dct || h->mb.b_transform_8x8 )
h->dctf.sub16x16_dct8( h->mb.pic.fenc_dct8, h->mb.pic.p_fenc[0], zero );
@@ -2011,7 +2011,7 @@ static void mb_analyse_inter_b16x16( x264_t *h, x264_mb_analysis_t *a )
}
else
{
- ALIGNED_ARRAY_32( pixel, pixuv, [2],[16*FENC_STRIDE] );
+ ALIGNED_ARRAY_64( pixel, pixuv, [2],[16*FENC_STRIDE] );
int chromapix = h->luma2chroma_pixel[PIXEL_16x16];
int v_shift = CHROMA_V_SHIFT;

diff --git a/encoder/ratecontrol.c b/encoder/ratecontrol.c
index b09c2434..99803718 100644
--- a/encoder/ratecontrol.c
+++ b/encoder/ratecontrol.c
@@ -243,7 +243,7 @@ static ALWAYS_INLINE uint32_t ac_energy_plane( x264_t *h, int mb_x, int mb_y, x2
stride <<= b_field;
if( b_chroma )
{
- ALIGNED_ARRAY_32( pixel, pix,[FENC_STRIDE*16] );
+ ALIGNED_ARRAY_64( pixel, pix,[FENC_STRIDE*16] );
int chromapix = h->luma2chroma_pixel[PIXEL_16x16];
int shift = 7 - CHROMA_V_SHIFT;

Loading...