Optimize silk_warped_autocorrelation_FIX() for ARM NEON
authorLinfeng Zhang <linfengz@google.com>
Wed, 13 Jul 2016 23:40:48 +0000 (16:40 -0700)
committerJean-Marc Valin <jmvalin@jmvalin.ca>
Thu, 13 Apr 2017 22:30:27 +0000 (18:30 -0400)
The optimization is bit exact with C function.

This optimization speeds up fixed-point SILK encoder on NEON about
5% to 8%. (Tested on Acer Chromebook, ARMv7 Processor rev 3 (v7l).)

Change-Id: I582f6f3585b7946149e16a2ad3084ebc0ae79a4f

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
Makefile.am
silk/arm/arm_silk_map.c
silk/fixed/arm/warped_autocorrelation_FIX_arm.h [new file with mode: 0644]
silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c [new file with mode: 0644]
silk/fixed/main_FIX.h
silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h
silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h
silk/fixed/noise_shape_analysis_FIX.c
silk/fixed/warped_autocorrelation_FIX.c
silk_headers.mk
silk_sources.mk

index 018a7c1..67453e4 100644 (file)
@@ -21,6 +21,9 @@ SILK_SOURCES += $(SILK_SOURCES_FIXED)
 if HAVE_SSE4_1
 SILK_SOURCES += $(SILK_SOURCES_SSE4_1) $(SILK_SOURCES_FIXED_SSE4_1)
 endif
+if HAVE_ARM_NEON_INTR
+SILK_SOURCES += $(SILK_SOURCES_FIXED_ARM_NEON_INTR)
+endif
 else
 SILK_SOURCES += $(SILK_SOURCES_FLOAT)
 if HAVE_SSE4_1
@@ -321,7 +324,9 @@ $(SSE4_1_OBJ) $(OPT_UNIT_TEST_OBJ): CFLAGS += $(OPUS_X86_SSE4_1_CFLAGS)
 endif
 
 if HAVE_ARM_NEON_INTR
-ARM_NEON_INTR_OBJ = $(CELT_SOURCES_ARM_NEON_INTR:.c=.lo) $(SILK_SOURCES_ARM_NEON_INTR:.c=.lo)
+ARM_NEON_INTR_OBJ = $(CELT_SOURCES_ARM_NEON_INTR:.c=.lo) \
+                    $(SILK_SOURCES_ARM_NEON_INTR:.c=.lo) \
+                    $(SILK_SOURCES_FIXED_ARM_NEON_INTR:.c=.lo)
 $(ARM_NEON_INTR_OBJ) $(OPT_UNIT_TEST_OBJ): CFLAGS += \
  $(OPUS_ARM_NEON_INTR_CFLAGS)  $(NE10_CFLAGS)
 endif
index 0966cfd..7ac9ad3 100644 (file)
@@ -87,4 +87,23 @@ opus_int32
 
 # endif
 
+# if defined(FIXED_POINT) && \
+ defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)
+
+void (*const SILK_WARPED_AUTOCORRELATION_FIX_IMPL[OPUS_ARCHMASK + 1])(
+          opus_int32                *corr,                                  /* O    Result [order + 1]                                                          */
+          opus_int                  *scale,                                 /* O    Scaling of the correlation vector                                           */
+    const opus_int16                *input,                                 /* I    Input data to correlate                                                     */
+    const opus_int                  warping_Q16,                            /* I    Warping coefficient                                                         */
+    const opus_int                  length,                                 /* I    Length of input                                                             */
+    const opus_int                  order                                   /* I    Correlation order (even)                                                    */
+) = {
+      silk_warped_autocorrelation_FIX_c,              /* ARMv4 */
+      silk_warped_autocorrelation_FIX_c,              /* EDSP */
+      silk_warped_autocorrelation_FIX_c,              /* Media */
+      MAY_HAVE_NEON(silk_warped_autocorrelation_FIX), /* Neon */
+};
+
+# endif
+
 #endif /* OPUS_HAVE_RTCD */
diff --git a/silk/fixed/arm/warped_autocorrelation_FIX_arm.h b/silk/fixed/arm/warped_autocorrelation_FIX_arm.h
new file mode 100644 (file)
index 0000000..1992e43
--- /dev/null
@@ -0,0 +1,68 @@
+/***********************************************************************
+Copyright (c) 2017 Google Inc.
+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_WARPED_AUTOCORRELATION_FIX_ARM_H
+# define SILK_WARPED_AUTOCORRELATION_FIX_ARM_H
+
+# include "celt/arm/armcpu.h"
+
+# if defined(FIXED_POINT)
+
+#  if defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
+void silk_warped_autocorrelation_FIX_neon(
+          opus_int32                *corr,                                  /* O    Result [order + 1]                                                          */
+          opus_int                  *scale,                                 /* O    Scaling of the correlation vector                                           */
+    const opus_int16                *input,                                 /* I    Input data to correlate                                                     */
+    const opus_int                  warping_Q16,                            /* I    Warping coefficient                                                         */
+    const opus_int                  length,                                 /* I    Length of input                                                             */
+    const opus_int                  order                                   /* I    Correlation order (even)                                                    */
+);
+
+#  if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON)
+#   define OVERRIDE_silk_warped_autocorrelation_FIX (1)
+#   define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \
+    ((void)(arch), PRESUME_NEON(silk_warped_autocorrelation_FIX)(corr, scale, input, warping_Q16, length, order))
+#  endif
+#  endif
+
+#  if !defined(OVERRIDE_silk_warped_autocorrelation_FIX)
+/*Is run-time CPU detection enabled on this platform?*/
+#   if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR))
+extern void (*const SILK_WARPED_AUTOCORRELATION_FIX_IMPL[OPUS_ARCHMASK+1])(opus_int32*, opus_int*, const opus_int16*, const opus_int, const opus_int, const opus_int);
+#    define OVERRIDE_silk_warped_autocorrelation_FIX (1)
+#    define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \
+    ((*SILK_WARPED_AUTOCORRELATION_FIX_IMPL[(arch)&OPUS_ARCHMASK])(corr, scale, input, warping_Q16, length, order))
+#   elif defined(OPUS_ARM_PRESUME_NEON_INTR)
+#    define OVERRIDE_silk_warped_autocorrelation_FIX (1)
+#    define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \
+    ((void)(arch), silk_warped_autocorrelation_FIX_neon(corr, scale, input, warping_Q16, length, order))
+#   endif
+#  endif
+
+# endif /* end FIXED_POINT */
+
+#endif /* end SILK_WARPED_AUTOCORRELATION_FIX_ARM_H */
diff --git a/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c b/silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c
new file mode 100644 (file)
index 0000000..00a70cb
--- /dev/null
@@ -0,0 +1,260 @@
+/***********************************************************************
+Copyright (c) 2017 Google Inc., Jean-Marc Valin
+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>
+#ifdef OPUS_CHECK_ASM
+# include <string.h>
+#endif
+#include "stack_alloc.h"
+#include "main_FIX.h"
+
+static OPUS_INLINE void calc_corr( const opus_int32 *const input_QS, opus_int64 *const corr_QC, const opus_int offset, const int32x4_t state_QS_s32x4 )
+{
+    int64x2_t corr_QC_s64x2[ 2 ], t_s64x2[ 2 ];
+    const int32x4_t input_QS_s32x4 = vld1q_s32( input_QS + offset );
+    corr_QC_s64x2[ 0 ] = vld1q_s64( corr_QC + offset + 0 );
+    corr_QC_s64x2[ 1 ] = vld1q_s64( corr_QC + offset + 2 );
+    t_s64x2[ 0 ] = vmull_s32( vget_low_s32( state_QS_s32x4 ), vget_low_s32( input_QS_s32x4 ) );
+    t_s64x2[ 1 ] = vmull_s32( vget_high_s32( state_QS_s32x4 ), vget_high_s32( input_QS_s32x4 ) );
+    corr_QC_s64x2[ 0 ] = vsraq_n_s64( corr_QC_s64x2[ 0 ], t_s64x2[ 0 ], 2 * QS - QC );
+    corr_QC_s64x2[ 1 ] = vsraq_n_s64( corr_QC_s64x2[ 1 ], t_s64x2[ 1 ], 2 * QS - QC );
+    vst1q_s64( corr_QC + offset + 0, corr_QC_s64x2[ 0 ] );
+    vst1q_s64( corr_QC + offset + 2, corr_QC_s64x2[ 1 ] );
+}
+
+static OPUS_INLINE int32x4_t calc_state( const int32x4_t state_QS0_s32x4, const int32x4_t state_QS0_1_s32x4, const int32x4_t state_QS1_1_s32x4, const int32x4_t warping_Q16_s32x4 )
+{
+    int32x4_t t_s32x4 = vsubq_s32( state_QS0_s32x4, state_QS0_1_s32x4 );
+    t_s32x4 = vqdmulhq_s32( t_s32x4, warping_Q16_s32x4 );
+    return vaddq_s32( state_QS1_1_s32x4, t_s32x4 );
+}
+
+void silk_warped_autocorrelation_FIX_neon(
+          opus_int32                *corr,                                  /* O    Result [order + 1]                                                          */
+          opus_int                  *scale,                                 /* O    Scaling of the correlation vector                                           */
+    const opus_int16                *input,                                 /* I    Input data to correlate                                                     */
+    const opus_int                  warping_Q16,                            /* I    Warping coefficient                                                         */
+    const opus_int                  length,                                 /* I    Length of input                                                             */
+    const opus_int                  order                                   /* I    Correlation order (even)                                                    */
+)
+{
+    if( ( MAX_SHAPE_LPC_ORDER > 24 ) || ( order < 6 ) ) {
+        silk_warped_autocorrelation_FIX_c( corr, scale, input, warping_Q16, length, order );
+    } else {
+        opus_int       n, i, lsh;
+        opus_int64     corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; /* In reverse order */
+        opus_int64     corr_QC_orderT;
+        int64x2_t      lsh_s64x2;
+        const opus_int orderT = ( order + 3 ) & ~3;
+        opus_int64     *corr_QCT;
+        opus_int32     *input_QS;
+        VARDECL( opus_int32, input_QST );
+        VARDECL( opus_int32, state );
+        SAVE_STACK;
+
+        /* Order must be even */
+        silk_assert( ( order & 1 ) == 0 );
+        silk_assert( 2 * QS - QC >= 0 );
+
+        ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER, opus_int32 );
+
+        input_QS = input_QST;
+        /* input_QS has zero paddings in the beginning and end. */
+        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 += 4;
+        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 += 4;
+
+        /* Loop over samples */
+        for( n = 0; n < length - 7; n += 8, input_QS += 8 ) {
+            const int16x8_t t0_s16x4 = vld1q_s16( input + n );
+            vst1q_s32( input_QS + 0, vshll_n_s16( vget_low_s16( t0_s16x4 ), QS ) );
+            vst1q_s32( input_QS + 4, vshll_n_s16( vget_high_s16( t0_s16x4 ), QS ) );
+        }
+        for( ; n < length; n++, input_QS++ ) {
+            input_QS[ 0 ] = silk_LSHIFT32( (opus_int32)input[ n ], QS );
+        }
+        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 += 4;
+        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.                  */
+        /* The zero paddings in input_QS guarantee corr_QC's correctness even with the extra epilogues.       */
+        /* The values of state_QS will be polluted by the extra epilogues, however they are temporary values. */
+
+        /* Keep the C code here to help understand the intrinsics optimization. */
+        /*
+        {
+            opus_int32 state_QS[ 2 ][ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };
+            opus_int32 *state_QST[ 3 ];
+            state_QST[ 0 ] = state_QS[ 0 ];
+            state_QST[ 1 ] = state_QS[ 1 ];
+            for( n = 0; n < length + order; n++, input_QS++ ) {
+                state_QST[ 0 ][ orderT ] = input_QS[ orderT ];
+                for( i = 0; i < orderT; i++ ) {
+                    corr_QC[ i ] += silk_RSHIFT64( silk_SMULL( state_QST[ 0 ][ i ], input_QS[ i ] ), 2 * QS - QC );
+                    state_QST[ 1 ][ i ] = silk_SMLAWB( state_QST[ 1 ][ i + 1 ], state_QST[ 0 ][ i ] - state_QST[ 0 ][ i + 1 ], warping_Q16 );
+                }
+                state_QST[ 2 ] = state_QST[ 0 ];
+                state_QST[ 0 ] = state_QST[ 1 ];
+                state_QST[ 1 ] = state_QST[ 2 ];
+            }
+        }
+        */
+
+        {
+            const int32x4_t warping_Q16_s32x4 = vdupq_n_s32( warping_Q16 << 15 );
+            const opus_int32 *in = input_QS + orderT;
+            opus_int o = orderT;
+            int32x4_t state_QS_s32x4[ 3 ][ 2 ];
+
+            ALLOC( state, length + orderT, opus_int32 );
+            state_QS_s32x4[ 2 ][ 1 ] = vdupq_n_s32( 0 );
+
+            /* Calculate 8 taps of all inputs in each loop. */
+            do {
+                state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 0 ][ 1 ] =
+                state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 1 ][ 1 ] = vdupq_n_s32( 0 );
+                n = 0;
+                do {
+                    calc_corr( input_QS + n, corr_QC, o - 8, state_QS_s32x4[ 0 ][ 0 ] );
+                    calc_corr( input_QS + n, corr_QC, o - 4, state_QS_s32x4[ 0 ][ 1 ] );
+                    state_QS_s32x4[ 2 ][ 1 ] = vld1q_s32( in + n );
+                    vst1q_lane_s32( state + n, state_QS_s32x4[ 0 ][ 0 ], 0 );
+                    state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 0 ][ 1 ], 1 );
+                    state_QS_s32x4[ 2 ][ 1 ] = vextq_s32( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], 1 );
+                    state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
+                    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 ) );
+                in = state;
+                o -= 8;
+            } while( o > 4 );
+
+            if( o ) {
+                /* Calculate the last 4 taps of all inputs. */
+                opus_int32 *stateT = state;
+                silk_assert( o == 4 );
+                state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 1 ][ 0 ] = vdupq_n_s32( 0 );
+                n = length + order;
+                do {
+                    calc_corr( input_QS, corr_QC, 0, state_QS_s32x4[ 0 ][ 0 ] );
+                    state_QS_s32x4[ 2 ][ 0 ] = vld1q_s32( stateT );
+                    vst1q_lane_s32( stateT, state_QS_s32x4[ 0 ][ 0 ], 0 );
+                    state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], 1 );
+                    state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
+                    state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
+                    input_QS++;
+                    stateT++;
+                } while( --n );
+            }
+        }
+
+        {
+            const opus_int16 *inputT = input;
+            int32x4_t t_s32x4;
+            int64x1_t t_s64x1;
+            int64x2_t t_s64x2 = vdupq_n_s64( 0 );
+            for( n = 0; n <= length - 8; n += 8 ) {
+                int16x8_t input_s16x8 = vld1q_s16( inputT );
+                t_s32x4 = vmull_s16( vget_low_s16( input_s16x8 ), vget_low_s16( input_s16x8 ) );
+                t_s32x4 = vmlal_s16( t_s32x4, vget_high_s16( input_s16x8 ), vget_high_s16( input_s16x8 ) );
+                t_s64x2 = vaddw_s32( t_s64x2, vget_low_s32( t_s32x4 ) );
+                t_s64x2 = vaddw_s32( t_s64x2, vget_high_s32( t_s32x4 ) );
+                inputT += 8;
+            }
+            t_s64x1 = vadd_s64( vget_low_s64( t_s64x2 ), vget_high_s64( t_s64x2 ) );
+            corr_QC_orderT = vget_lane_s64( t_s64x1, 0 );
+            for( ; n < length; n++ ) {
+                corr_QC_orderT += silk_SMULL( input[ n ], input[ n ] );
+            }
+            corr_QC_orderT = silk_LSHIFT64( corr_QC_orderT, QC );
+            corr_QC[ orderT ] = corr_QC_orderT;
+        }
+
+        corr_QCT = corr_QC + orderT - order;
+        lsh = silk_CLZ64( corr_QC_orderT ) - 35;
+        lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC );
+        *scale = -( QC + lsh );
+        silk_assert( *scale >= -30 && *scale <= 12 );
+        lsh_s64x2 = vdupq_n_s64( lsh );
+        for( i = 0; i <= order - 3; i += 4 ) {
+            int32x4_t corr_s32x4;
+            int64x2_t corr_QC0_s64x2, corr_QC1_s64x2;
+            corr_QC0_s64x2 = vld1q_s64( corr_QCT + i );
+            corr_QC1_s64x2 = vld1q_s64( corr_QCT + i + 2 );
+            corr_QC0_s64x2 = vshlq_s64( corr_QC0_s64x2, lsh_s64x2 );
+            corr_QC1_s64x2 = vshlq_s64( corr_QC1_s64x2, lsh_s64x2 );
+            corr_s32x4     = vcombine_s32( vmovn_s64( corr_QC1_s64x2 ), vmovn_s64( corr_QC0_s64x2 ) );
+            corr_s32x4     = vrev64q_s32( corr_s32x4 );
+            vst1q_s32( corr + order - i - 3, corr_s32x4 );
+        }
+        if( lsh >= 0 ) {
+            for( ; i < order + 1; i++ ) {
+                corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QCT[ i ], lsh ) );
+            }
+        } else {
+            for( ; i < order + 1; i++ ) {
+                corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QCT[ i ], -lsh ) );
+            }
+        }
+        silk_assert( corr_QCT[ order ] >= 0 ); /* If breaking, decrease QC*/
+        RESTORE_STACK;
+    }
+
+#ifdef OPUS_CHECK_ASM
+    {
+        opus_int32 corr_c[ MAX_SHAPE_LPC_ORDER + 1 ];
+        opus_int   scale_c;
+        silk_warped_autocorrelation_FIX_c( corr_c, &scale_c, input, warping_Q16, length, order );
+        silk_assert( !memcmp( corr_c, corr, sizeof( corr_c[ 0 ] ) * ( order + 1 ) ) );
+        silk_assert( scale_c == *scale );
+    }
+#endif
+}
index ddbf377..780afa3 100644 (file)
@@ -36,6 +36,11 @@ POSSIBILITY OF SUCH DAMAGE.
 #include "debug.h"
 #include "entenc.h"
 
+#if ((defined(OPUS_ARM_ASM) && defined(FIXED_POINT)) \
+   || defined(OPUS_ARM_MAY_HAVE_NEON_INTR))
+#include "fixed/arm/warped_autocorrelation_FIX_arm.h"
+#endif
+
 #ifndef FORCE_CPP_BUILD
 #ifdef __cplusplus
 extern "C"
@@ -47,6 +52,9 @@ extern "C"
 #define silk_encode_do_VAD_Fxx      silk_encode_do_VAD_FIX
 #define silk_encode_frame_Fxx       silk_encode_frame_FIX
 
+#define QC  10
+#define QS  13
+
 /*********************/
 /* Encoder Functions */
 /*********************/
@@ -99,7 +107,7 @@ void silk_noise_shape_analysis_FIX(
 );
 
 /* Autocorrelations for a warped frequency axis */
-void silk_warped_autocorrelation_FIX(
+void silk_warped_autocorrelation_FIX_c(
           opus_int32                *corr,                                  /* O    Result [order + 1]                                                          */
           opus_int                  *scale,                                 /* O    Scaling of the correlation vector                                           */
     const opus_int16                *input,                                 /* I    Input data to correlate                                                     */
@@ -108,6 +116,11 @@ void silk_warped_autocorrelation_FIX(
     const opus_int                  order                                   /* I    Correlation order (even)                                                    */
 );
 
+#if !defined(OVERRIDE_silk_warped_autocorrelation_FIX)
+#define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \
+        ((void)(arch), silk_warped_autocorrelation_FIX_c(corr, scale, input, warping_Q16, length, order))
+#endif
+
 /* Calculation of LTP state scaling */
 void silk_LTP_scale_ctrl_FIX(
     silk_encoder_state_FIX          *psEnc,                                 /* I/O  encoder state                                                               */
index f9b5473..3999b5b 100644 (file)
@@ -169,7 +169,7 @@ void silk_noise_shape_analysis_FIX(
 
         if( psEnc->sCmn.warping_Q16 > 0 ) {
             /* Calculate warped auto correlation */
-            silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder );
+            silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder, arch );
         } else {
             /* Calculate regular auto correlation */
             silk_autocorr( auto_corr, &scale, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1, arch );
index e803ef0..fcbd96c 100644 (file)
@@ -48,7 +48,8 @@ void silk_warped_autocorrelation_FIX(
     const opus_int16                *input,                                 /* I    Input data to correlate                                                     */
     const opus_int                  warping_Q16,                            /* I    Warping coefficient                                                         */
     const opus_int                  length,                                 /* I    Length of input                                                             */
-    const opus_int                  order                                   /* I    Correlation order (even)                                                    */
+    const opus_int                  order,                                  /* I    Correlation order (even)                                                    */
+    int                             arch                                    /* I    Run-time architecture                                                       */
 )
 {
     opus_int   n, i, lsh;
index 8fe2377..85fea0b 100644 (file)
@@ -262,7 +262,7 @@ void silk_noise_shape_analysis_FIX(
 
         if( psEnc->sCmn.warping_Q16 > 0 ) {
             /* Calculate warped auto correlation */
-            silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder );
+            silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder, arch );
         } else {
             /* Calculate regular auto correlation */
             silk_autocorr( auto_corr, &scale, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1, arch );
index af16402..994c299 100644 (file)
@@ -31,17 +31,13 @@ POSSIBILITY OF SUCH DAMAGE.
 
 #include "main_FIX.h"
 
-#define QC  10
-#define QS  13
-
 #if defined(MIPSr1_ASM)
 #include "mips/warped_autocorrelation_FIX_mipsr1.h"
 #endif
 
 
-#ifndef OVERRIDE_silk_warped_autocorrelation_FIX
 /* Autocorrelations for a warped frequency axis */
-void silk_warped_autocorrelation_FIX(
+void silk_warped_autocorrelation_FIX_c(
           opus_int32                *corr,                                  /* O    Result [order + 1]                                                          */
           opus_int                  *scale,                                 /* O    Scaling of the correlation vector                                           */
     const opus_int16                *input,                                 /* I    Input data to correlate                                                     */
@@ -92,4 +88,3 @@ void silk_warped_autocorrelation_FIX(
     }
     silk_assert( corr_QC[ 0 ] >= 0 ); /* If breaking, decrease QC*/
 }
-#endif /* OVERRIDE_silk_warped_autocorrelation_FIX */
index cb10491..278500b 100644 (file)
@@ -32,6 +32,7 @@ silk/arm/NSQ_del_dec_arm.h \
 silk/arm/NSQ_neon.h \
 silk/fixed/main_FIX.h \
 silk/fixed/structs_FIX.h \
+silk/fixed/arm/warped_autocorrelation_FIX_arm.h \
 silk/fixed/mips/noise_shape_analysis_FIX_mipsr1.h \
 silk/fixed/mips/warped_autocorrelation_FIX_mipsr1.h \
 silk/float/main_FLP.h \
index 0dcf671..c031214 100644 (file)
@@ -117,6 +117,9 @@ silk/fixed/schur_FIX.c
 SILK_SOURCES_FIXED_SSE4_1 = silk/fixed/x86/vector_ops_FIX_sse.c \
 silk/fixed/x86/burg_modified_FIX_sse.c
 
+SILK_SOURCES_FIXED_ARM_NEON_INTR = \
+silk/fixed/arm/warped_autocorrelation_FIX_neon_intr.c
+
 SILK_SOURCES_FLOAT = \
 silk/float/apply_sine_window_FLP.c \
 silk/float/corrMatrix_FLP.c \