diff options
Diffstat (limited to 'silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c')
-rw-r--r-- | silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c | 9 |
1 files changed, 7 insertions, 2 deletions
diff --git a/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c b/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c index 00a70cb5..6f3be025 100644 --- a/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c +++ b/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c @@ -84,7 +84,9 @@ void silk_warped_autocorrelation_FIX_neon( silk_assert( ( order & 1 ) == 0 ); silk_assert( 2 * QS - QC >= 0 ); - ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER, opus_int32 ); + /* The additional +4 is to ensure a later vld1q_s32 call does not overflow. */ + /* Strictly, only +3 is needed but +4 simplifies initialization using the 4x32 neon load. */ + ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER + 4, opus_int32 ); input_QS = input_QST; /* input_QS has zero paddings in the beginning and end. */ @@ -121,6 +123,8 @@ void silk_warped_autocorrelation_FIX_neon( vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); input_QS += 4; vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); + input_QS += 4; + vst1q_s32( input_QS, vdupq_n_s32( 0 ) ); input_QS = input_QST + MAX_SHAPE_LPC_ORDER - orderT; /* The following loop runs ( length + order ) times, with ( order ) extra epilogues. */ @@ -153,7 +157,8 @@ void silk_warped_autocorrelation_FIX_neon( opus_int o = orderT; int32x4_t state_QS_s32x4[ 3 ][ 2 ]; - ALLOC( state, length + orderT, opus_int32 ); + /* The additional +4 is to ensure a later vld1q_s32 call does not overflow. */ + ALLOC( state, length + order + 4, opus_int32 ); state_QS_s32x4[ 2 ][ 1 ] = vdupq_n_s32( 0 ); /* Calculate 8 taps of all inputs in each loop. */ |