From 5d2e0a05b7df5aca4525cc547cd8538f3339b526 Mon Sep 17 00:00:00 2001 From: Ray Essick Date: Wed, 4 Sep 2019 09:09:52 -0700 Subject: improve handling at end of buffer a prior change reduced iterations through the input buffer to avoid the NEON operations from overrunning the end of the locally allocated buffer. While avoiding the overrun, it generated bad results. Here we instead extend the locally allocated buffers enough that the original iteration count won't overrun. Some pre-existing bit-exact issues remain. Bug: 136616344 Test: CTS + bit-exact cross-checks. (cherry picked from commit aae866aed579da4e1c3299a1e9b94a1713a0decb) Merged-In: Ifb790a7d1d09a4ce7da900b43e3fa1f7ab01ac53 Change-Id: Ia4a94c89979d6b3b0ddead135036aec40259f53a --- silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-) diff --git a/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c b/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c index ee06f986..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. */ @@ -172,7 +177,7 @@ void silk_warped_autocorrelation_FIX_neon( state_QS_s32x4[ 0 ][ 1 ] = calc_state( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], state_QS_s32x4[ 1 ][ 1 ], warping_Q16_s32x4 ); state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ]; state_QS_s32x4[ 1 ][ 1 ] = state_QS_s32x4[ 2 ][ 1 ]; - } while( ++n < ( length + order - 3) ); + } while( ++n < ( length + order ) ); in = state; o -= 8; } while( o > 4 ); -- cgit v1.2.3