ref: 44e8e7385037af88994088d677d8319fc9850a48
parent: fdb3b3ac34403890cf62857f30d4f58a294ce142
author: Jonathan Lennox <[email protected]>
date: Tue Dec 22 14:21:44 EST 2015
Add Neon intrinsics for Silk noise shape quantization. Signed-off-by: Timothy B. Terriberry <[email protected]>
--- a/Makefile.am
+++ b/Makefile.am
@@ -49,6 +49,7 @@
if HAVE_ARM_NEON_INTR
CELT_SOURCES += $(CELT_SOURCES_ARM_NEON_INTR)
+SILK_SOURCES += $(SILK_SOURCES_ARM_NEON_INTR)
endif
if HAVE_ARM_NE10
@@ -295,7 +296,7 @@
endif
if HAVE_ARM_NEON_INTR
-CELT_ARM_NEON_INTR_OBJ = $(CELT_SOURCES_ARM_NEON_INTR:.c=.lo)
-$(CELT_ARM_NEON_INTR_OBJ) $(OPT_UNIT_TEST_OBJ): CFLAGS += \
+ARM_NEON_INTR_OBJ = $(CELT_SOURCES_ARM_NEON_INTR:.c=.lo) $(SILK_SOURCES_ARM_NEON_INTR:.c=.lo)
+$(ARM_NEON_INTR_OBJ) $(OPT_UNIT_TEST_OBJ): CFLAGS += \
$(OPUS_ARM_NEON_INTR_CFLAGS) $(NE10_CFLAGS)
endif
--- a/silk/NSQ.c
+++ b/silk/NSQ.c
@@ -31,7 +31,9 @@
#include "main.h"
#include "stack_alloc.h"
+#include "NSQ.h"
+
static OPUS_INLINE void silk_nsq_scale_states(
const silk_encoder_state *psEncC, /* I Encoder State */
silk_nsq_state *NSQ, /* I/O NSQ state */
@@ -66,7 +68,8 @@
opus_int offset_Q10, /* I */
opus_int length, /* I Input length */
opus_int shapingLPCOrder, /* I Noise shaping AR filter order */
- opus_int predictLPCOrder /* I Prediction filter order */
+ opus_int predictLPCOrder, /* I Prediction filter order */
+ int arch /* I Architecture */
);
#endif
@@ -155,7 +158,7 @@
silk_noise_shape_quantizer( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14,
AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ], Lambda_Q10,
- offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder );
+ offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->arch );
x_Q3 += psEncC->subfr_length;
pulses += psEncC->subfr_length;
@@ -198,7 +201,8 @@
opus_int offset_Q10, /* I */
opus_int length, /* I Input length */
opus_int shapingLPCOrder, /* I Noise shaping AR filter order */
- opus_int predictLPCOrder /* I Prediction filter order */
+ opus_int predictLPCOrder, /* I Prediction filter order */
+ int arch /* I Architecture */
)
{opus_int i, j;
@@ -207,6 +211,9 @@
opus_int32 exc_Q14, LPC_exc_Q14, xq_Q14, Gain_Q10;
opus_int32 tmp1, tmp2, sLF_AR_shp_Q14;
opus_int32 *psLPC_Q14, *shp_lag_ptr, *pred_lag_ptr;
+#ifdef silk_short_prediction_create_arch_coef
+ opus_int32 a_Q12_arch[MAX_LPC_ORDER];
+#endif
shp_lag_ptr = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ];
pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ];
@@ -215,32 +222,16 @@
/* Set up short term AR state */
psLPC_Q14 = &NSQ->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 ];
+#ifdef silk_short_prediction_create_arch_coef
+ silk_short_prediction_create_arch_coef(a_Q12_arch, a_Q12, predictLPCOrder);
+#endif
+
for( i = 0; i < length; i++ ) {/* Generate dither */
NSQ->rand_seed = silk_RAND( NSQ->rand_seed );
/* Short-term prediction */
- silk_assert( predictLPCOrder == 10 || predictLPCOrder == 16 );
- /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
- LPC_pred_Q10 = silk_RSHIFT( predictLPCOrder, 1 );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ 0 ], a_Q12[ 0 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -1 ], a_Q12[ 1 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -2 ], a_Q12[ 2 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -3 ], a_Q12[ 3 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -4 ], a_Q12[ 4 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -5 ], a_Q12[ 5 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -6 ], a_Q12[ 6 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -7 ], a_Q12[ 7 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -8 ], a_Q12[ 8 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -9 ], a_Q12[ 9 ] );
- if( predictLPCOrder == 16 ) {- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -10 ], a_Q12[ 10 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -11 ], a_Q12[ 11 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -12 ], a_Q12[ 12 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -13 ], a_Q12[ 13 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -14 ], a_Q12[ 14 ] );
- LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -15 ], a_Q12[ 15 ] );
- }
+ LPC_pred_Q10 = silk_noise_shape_quantizer_short_prediction(psLPC_Q14, a_Q12, a_Q12_arch, predictLPCOrder, arch);
/* Long-term prediction */
if( signalType == TYPE_VOICED ) {--- /dev/null
+++ b/silk/NSQ.h
@@ -1,0 +1,72 @@
+/***********************************************************************
+Copyright (c) 2014 Vidyo.
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+#ifndef SILK_NSQ_H
+#define SILK_NSQ_H
+
+#include "SigProc_FIX.h"
+
+#undef silk_short_prediction_create_arch_coef
+
+static OPUS_INLINE opus_int32 silk_noise_shape_quantizer_short_prediction_c(const opus_int32 *buf32, const opus_int16 *coef16, opus_int order)
+{+ opus_int32 out;
+ silk_assert( order == 10 || order == 16 );
+
+ /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+ out = silk_RSHIFT( order, 1 );
+ out = silk_SMLAWB( out, buf32[ 0 ], coef16[ 0 ] );
+ out = silk_SMLAWB( out, buf32[ -1 ], coef16[ 1 ] );
+ out = silk_SMLAWB( out, buf32[ -2 ], coef16[ 2 ] );
+ out = silk_SMLAWB( out, buf32[ -3 ], coef16[ 3 ] );
+ out = silk_SMLAWB( out, buf32[ -4 ], coef16[ 4 ] );
+ out = silk_SMLAWB( out, buf32[ -5 ], coef16[ 5 ] );
+ out = silk_SMLAWB( out, buf32[ -6 ], coef16[ 6 ] );
+ out = silk_SMLAWB( out, buf32[ -7 ], coef16[ 7 ] );
+ out = silk_SMLAWB( out, buf32[ -8 ], coef16[ 8 ] );
+ out = silk_SMLAWB( out, buf32[ -9 ], coef16[ 9 ] );
+
+ if( order == 16 )
+ {+ out = silk_SMLAWB( out, buf32[ -10 ], coef16[ 10 ] );
+ out = silk_SMLAWB( out, buf32[ -11 ], coef16[ 11 ] );
+ out = silk_SMLAWB( out, buf32[ -12 ], coef16[ 12 ] );
+ out = silk_SMLAWB( out, buf32[ -13 ], coef16[ 13 ] );
+ out = silk_SMLAWB( out, buf32[ -14 ], coef16[ 14 ] );
+ out = silk_SMLAWB( out, buf32[ -15 ], coef16[ 15 ] );
+ }
+ return out;
+}
+
+#define silk_noise_shape_quantizer_short_prediction(in, coef, coefRev, order, arch) ((void)arch,silk_noise_shape_quantizer_short_prediction_c(in, coef, order))
+
+
+#if defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
+#include "arm/NSQ_neon.h"
+#endif
+
+#endif /* SILK_NSQ_H */
--- /dev/null
+++ b/silk/arm/NSQ_neon.c
@@ -1,0 +1,69 @@
+/***********************************************************************
+Copyright (C) 2014 Vidyo
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <arm_neon.h>
+#include "main.h"
+#include "stack_alloc.h"
+#include "NSQ.h"
+#include "celt/cpu_support.h"
+#include "celt/arm/armcpu.h"
+
+opus_int32 silk_noise_shape_quantizer_short_prediction_neon(const opus_int32 *buf32, const opus_int32 *coef32, opus_int order)
+{+ int32x4_t coef0 = vld1q_s32(coef32);
+ int32x4_t coef1 = vld1q_s32(coef32 + 4);
+ int32x4_t coef2 = vld1q_s32(coef32 + 8);
+ int32x4_t coef3 = vld1q_s32(coef32 + 12);
+
+ int32x4_t a0 = vld1q_s32(buf32 - 15);
+ int32x4_t a1 = vld1q_s32(buf32 - 11);
+ int32x4_t a2 = vld1q_s32(buf32 - 7);
+ int32x4_t a3 = vld1q_s32(buf32 - 3);
+
+ int32x4_t b0 = vqdmulhq_s32(coef0, a0);
+ int32x4_t b1 = vqdmulhq_s32(coef1, a1);
+ int32x4_t b2 = vqdmulhq_s32(coef2, a2);
+ int32x4_t b3 = vqdmulhq_s32(coef3, a3);
+
+ int32x4_t c0 = vaddq_s32(b0, b1);
+ int32x4_t c1 = vaddq_s32(b2, b3);
+
+ int32x4_t d = vaddq_s32(c0, c1);
+
+ int64x2_t e = vpaddlq_s32(d);
+
+ int64x1_t f = vadd_s64(vget_low_s64(e), vget_high_s64(e));
+
+ opus_int32 out = vget_lane_s32(vreinterpret_s32_s64(f), 0);
+
+ out += silk_RSHIFT( order, 1 );
+
+ return out;
+}
--- /dev/null
+++ b/silk/arm/NSQ_neon.h
@@ -1,0 +1,97 @@
+/***********************************************************************
+Copyright (C) 2014 Vidyo
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+#ifndef SILK_NSQ_NEON_H
+#define SILK_NSQ_NEON_H
+
+#undef silk_short_prediction_create_arch_coef
+/* For vectorized calc, reverse a_Q12 coefs, convert to 32-bit, and shift for vqdmulhq_s32. */
+static OPUS_INLINE void silk_short_prediction_create_arch_coef_neon(opus_int32 *out, const opus_int16 *in, opus_int order)
+{+ out[15] = in[0] << 15;
+ out[14] = in[1] << 15;
+ out[13] = in[2] << 15;
+ out[12] = in[3] << 15;
+ out[11] = in[4] << 15;
+ out[10] = in[5] << 15;
+ out[9] = in[6] << 15;
+ out[8] = in[7] << 15;
+ out[7] = in[8] << 15;
+ out[6] = in[9] << 15;
+
+ if (order == 16)
+ {+ out[5] = in[10] << 15;
+ out[4] = in[11] << 15;
+ out[3] = in[12] << 15;
+ out[2] = in[13] << 15;
+ out[1] = in[14] << 15;
+ out[0] = in[15] << 15;
+ }
+ else
+ {+ out[5] = 0;
+ out[4] = 0;
+ out[3] = 0;
+ out[2] = 0;
+ out[1] = 0;
+ out[0] = 0;
+ }
+}
+
+#if defined(OPUS_ARM_PRESUME_NEON_INTR)
+
+#define silk_short_prediction_create_arch_coef(out, in, order) \
+ (silk_short_prediction_create_arch_coef_neon(out, in, order))
+
+#elif defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
+
+#define silk_short_prediction_create_arch_coef(out, in, order) \
+ do { if (arch == OPUS_ARCH_ARM_NEON) { silk_short_prediction_create_arch_coef_neon(out, in, order); } } while (0)+
+#endif
+
+opus_int32 silk_noise_shape_quantizer_short_prediction_neon(const opus_int32 *buf32, const opus_int32 *coef32, opus_int order);
+
+#if defined(OPUS_ARM_PRESUME_NEON_INTR)
+#undef silk_noise_shape_quantizer_short_prediction
+#define silk_noise_shape_quantizer_short_prediction(in, coef, coefRev, order, arch) \
+ ((void)arch,silk_noise_shape_quantizer_short_prediction_neon(in, coefRev, order))
+
+#elif defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
+
+/* silk_noise_shape_quantizer_short_prediction implementations take different parameters based on arch
+ (coef vs. coefRev) so can't use the usual IMPL table implementation */
+#undef silk_noise_shape_quantizer_short_prediction
+#define silk_noise_shape_quantizer_short_prediction(in, coef, coefRev, order, arch) \
+ (arch == OPUS_ARCH_ARM_NEON ? \
+ silk_noise_shape_quantizer_short_prediction_neon(in, coefRev, order) : \
+ silk_noise_shape_quantizer_short_prediction_c(in, coef, order))
+
+
+#endif
+
+#endif /* SILK_NSQ_NEON_H */
--- a/silk/x86/NSQ_sse.c
+++ b/silk/x86/NSQ_sse.c
@@ -221,7 +221,7 @@
{silk_noise_shape_quantizer( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14,
AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ], Lambda_Q10,
- offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder );
+ offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->arch );
}
x_Q3 += psEncC->subfr_length;
--- a/silk/x86/main_sse.h
+++ b/silk/x86/main_sse.h
@@ -207,7 +207,8 @@
opus_int offset_Q10, /* I */
opus_int length, /* I Input length */
opus_int shapingLPCOrder, /* I Noise shaping AR filter order */
- opus_int predictLPCOrder /* I Prediction filter order */
+ opus_int predictLPCOrder, /* I Prediction filter order */
+ int arch /* I Architecture */
);
/**************************/
--- a/silk_headers.mk
+++ b/silk_headers.mk
@@ -15,6 +15,7 @@
silk/MacroCount.h \
silk/MacroDebug.h \
silk/macros.h \
+silk/NSQ.h \
silk/pitch_est_defines.h \
silk/resampler_private.h \
silk/resampler_rom.h \
@@ -25,6 +26,7 @@
silk/arm/macros_armv5e.h \
silk/arm/SigProc_FIX_armv4.h \
silk/arm/SigProc_FIX_armv5e.h \
+silk/arm/NSQ_neon.h \
silk/fixed/main_FIX.h \
silk/fixed/structs_FIX.h \
silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h \
--- a/silk_sources.mk
+++ b/silk_sources.mk
@@ -82,6 +82,8 @@
silk/x86/VAD_sse.c \
silk/x86/VQ_WMat_EC_sse.c
+SILK_SOURCES_ARM_NEON_INTR = silk/arm/NSQ_neon.c
+
SILK_SOURCES_FIXED = \
silk/fixed/LTP_analysis_filter_FIX.c \
silk/fixed/LTP_scale_ctrl_FIX.c \