From b592a7a8a52bc2b61108570afff943b7e9afdc3d Mon Sep 17 00:00:00 2001 From: Nils Wallménius Date: Fri, 30 Aug 2013 23:27:12 +0200 Subject: Put two hot silk arrays on real stack (iram) Speeds up decoding of 16kbps test file by 16.7MHz on H300. Change-Id: I39c90e3b423ae8e2ee5c2b88c5dcec8d48807f77 --- lib/rbcodec/codecs/libopus/silk/decode_core.c | 5 +++-- lib/rbcodec/codecs/libopus/silk/resampler_private_IIR_FIR.c | 11 +++++++---- 2 files changed, 10 insertions(+), 6 deletions(-) diff --git a/lib/rbcodec/codecs/libopus/silk/decode_core.c b/lib/rbcodec/codecs/libopus/silk/decode_core.c index 693c9faacf..b298389f07 100644 --- a/lib/rbcodec/codecs/libopus/silk/decode_core.c +++ b/lib/rbcodec/codecs/libopus/silk/decode_core.c @@ -49,7 +49,7 @@ void silk_decode_core( opus_int32 LTP_pred_Q13, LPC_pred_Q10, Gain_Q10, inv_gain_Q31, gain_adj_Q16, rand_seed, offset_Q10; opus_int32 *pred_lag_ptr, *pexc_Q14, *pres_Q14; VARDECL( opus_int32, res_Q14 ); - VARDECL( opus_int32, sLPC_Q14 ); +/* VARDECL( opus_int32, sLPC_Q14 ); */ SAVE_STACK; silk_assert( psDec->prev_gain_Q16 != 0 ); @@ -57,7 +57,8 @@ void silk_decode_core( ALLOC( sLTP, psDec->ltp_mem_length, opus_int16 ); ALLOC( sLTP_Q15, psDec->ltp_mem_length + psDec->frame_length, opus_int32 ); ALLOC( res_Q14, psDec->subfr_length, opus_int32 ); - ALLOC( sLPC_Q14, psDec->subfr_length + MAX_LPC_ORDER, opus_int32 ); +/* ALLOC( sLPC_Q14, psDec->subfr_length + MAX_LPC_ORDER, opus_int32 ); */ + opus_int32 sLPC_Q14[psDec->subfr_length + MAX_LPC_ORDER]; /* worst case is 80 + 16 */ offset_Q10 = silk_Quantization_Offsets_Q10[ psDec->indices.signalType >> 1 ][ psDec->indices.quantOffsetType ]; diff --git a/lib/rbcodec/codecs/libopus/silk/resampler_private_IIR_FIR.c b/lib/rbcodec/codecs/libopus/silk/resampler_private_IIR_FIR.c index 04246257e7..ce9aff0907 100644 --- a/lib/rbcodec/codecs/libopus/silk/resampler_private_IIR_FIR.c +++ b/lib/rbcodec/codecs/libopus/silk/resampler_private_IIR_FIR.c @@ -72,10 +72,13 @@ void silk_resampler_private_IIR_FIR( silk_resampler_state_struct *S = (silk_resampler_state_struct *)SS; opus_int32 nSamplesIn; opus_int32 max_index_Q16, index_increment_Q16; - VARDECL( opus_int16, buf ); - SAVE_STACK; +/* VARDECL( opus_int16, buf ); + SAVE_STACK; */ - ALLOC( buf, 2 * S->batchSize + RESAMPLER_ORDER_FIR_12, opus_int16 ); +/* ALLOC( buf, 2 * S->batchSize + RESAMPLER_ORDER_FIR_12, opus_int16 ); */ + + /* worst case = 2*16*10+8 = 328 * 2 = 656bytes */ + opus_int16 buf[2 * S->batchSize + RESAMPLER_ORDER_FIR_12]; /* Copy buffered samples to start of buffer */ silk_memcpy( buf, S->sFIR.i16, RESAMPLER_ORDER_FIR_12 * sizeof( opus_int16 ) ); @@ -103,5 +106,5 @@ void silk_resampler_private_IIR_FIR( /* Copy last part of filtered signal to the state for the next call */ silk_memcpy( S->sFIR.i16, &buf[ nSamplesIn << 1 ], RESAMPLER_ORDER_FIR_12 * sizeof( opus_int16 ) ); - RESTORE_STACK; +/* RESTORE_STACK; */ } -- cgit v1.2.3