cannam@154: /* Copyright (c) 2014-2015 Xiph.Org Foundation
cannam@154:    Written by Viswanath Puttagunta */
cannam@154: /**
cannam@154:    @file celt_neon_intr.c
cannam@154:    @brief ARM Neon Intrinsic optimizations for celt
cannam@154:  */
cannam@154: 
cannam@154: /*
cannam@154:    Redistribution and use in source and binary forms, with or without
cannam@154:    modification, are permitted provided that the following conditions
cannam@154:    are met:
cannam@154: 
cannam@154:    - Redistributions of source code must retain the above copyright
cannam@154:    notice, this list of conditions and the following disclaimer.
cannam@154: 
cannam@154:    - Redistributions in binary form must reproduce the above copyright
cannam@154:    notice, this list of conditions and the following disclaimer in the
cannam@154:    documentation and/or other materials provided with the distribution.
cannam@154: 
cannam@154:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
cannam@154:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
cannam@154:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
cannam@154:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
cannam@154:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
cannam@154:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
cannam@154:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
cannam@154:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
cannam@154:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
cannam@154:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
cannam@154:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
cannam@154: */
cannam@154: 
cannam@154: #ifdef HAVE_CONFIG_H
cannam@154: #include "config.h"
cannam@154: #endif
cannam@154: 
cannam@154: #include <arm_neon.h>
cannam@154: #include "../pitch.h"
cannam@154: 
cannam@154: #if defined(FIXED_POINT)
cannam@154: void xcorr_kernel_neon_fixed(const opus_val16 * x, const opus_val16 * y, opus_val32 sum[4], int len)
cannam@154: {
cannam@154:    int j;
cannam@154:    int32x4_t a = vld1q_s32(sum);
cannam@154:    /* Load y[0...3] */
cannam@154:    /* This requires len>0 to always be valid (which we assert in the C code). */
cannam@154:    int16x4_t y0 = vld1_s16(y);
cannam@154:    y += 4;
cannam@154: 
cannam@154:    for (j = 0; j + 8 <= len; j += 8)
cannam@154:    {
cannam@154:       /* Load x[0...7] */
cannam@154:       int16x8_t xx = vld1q_s16(x);
cannam@154:       int16x4_t x0 = vget_low_s16(xx);
cannam@154:       int16x4_t x4 = vget_high_s16(xx);
cannam@154:       /* Load y[4...11] */
cannam@154:       int16x8_t yy = vld1q_s16(y);
cannam@154:       int16x4_t y4 = vget_low_s16(yy);
cannam@154:       int16x4_t y8 = vget_high_s16(yy);
cannam@154:       int32x4_t a0 = vmlal_lane_s16(a, y0, x0, 0);
cannam@154:       int32x4_t a1 = vmlal_lane_s16(a0, y4, x4, 0);
cannam@154: 
cannam@154:       int16x4_t y1 = vext_s16(y0, y4, 1);
cannam@154:       int16x4_t y5 = vext_s16(y4, y8, 1);
cannam@154:       int32x4_t a2 = vmlal_lane_s16(a1, y1, x0, 1);
cannam@154:       int32x4_t a3 = vmlal_lane_s16(a2, y5, x4, 1);
cannam@154: 
cannam@154:       int16x4_t y2 = vext_s16(y0, y4, 2);
cannam@154:       int16x4_t y6 = vext_s16(y4, y8, 2);
cannam@154:       int32x4_t a4 = vmlal_lane_s16(a3, y2, x0, 2);
cannam@154:       int32x4_t a5 = vmlal_lane_s16(a4, y6, x4, 2);
cannam@154: 
cannam@154:       int16x4_t y3 = vext_s16(y0, y4, 3);
cannam@154:       int16x4_t y7 = vext_s16(y4, y8, 3);
cannam@154:       int32x4_t a6 = vmlal_lane_s16(a5, y3, x0, 3);
cannam@154:       int32x4_t a7 = vmlal_lane_s16(a6, y7, x4, 3);
cannam@154: 
cannam@154:       y0 = y8;
cannam@154:       a = a7;
cannam@154:       x += 8;
cannam@154:       y += 8;
cannam@154:    }
cannam@154: 
cannam@154:    for (; j < len; j++)
cannam@154:    {
cannam@154:       int16x4_t x0 = vld1_dup_s16(x);  /* load next x */
cannam@154:       int32x4_t a0 = vmlal_s16(a, y0, x0);
cannam@154: 
cannam@154:       int16x4_t y4 = vld1_dup_s16(y);  /* load next y */
cannam@154:       y0 = vext_s16(y0, y4, 1);
cannam@154:       a = a0;
cannam@154:       x++;
cannam@154:       y++;
cannam@154:    }
cannam@154: 
cannam@154:    vst1q_s32(sum, a);
cannam@154: }
cannam@154: 
cannam@154: #else
cannam@154: /*
cannam@154:  * Function: xcorr_kernel_neon_float
cannam@154:  * ---------------------------------
cannam@154:  * Computes 4 correlation values and stores them in sum[4]
cannam@154:  */
cannam@154: static void xcorr_kernel_neon_float(const float32_t *x, const float32_t *y,
cannam@154:       float32_t sum[4], int len) {
cannam@154:    float32x4_t YY[3];
cannam@154:    float32x4_t YEXT[3];
cannam@154:    float32x4_t XX[2];
cannam@154:    float32x2_t XX_2;
cannam@154:    float32x4_t SUMM;
cannam@154:    const float32_t *xi = x;
cannam@154:    const float32_t *yi = y;
cannam@154: 
cannam@154:    celt_assert(len>0);
cannam@154: 
cannam@154:    YY[0] = vld1q_f32(yi);
cannam@154:    SUMM = vdupq_n_f32(0);
cannam@154: 
cannam@154:    /* Consume 8 elements in x vector and 12 elements in y
cannam@154:     * vector. However, the 12'th element never really gets
cannam@154:     * touched in this loop. So, if len == 8, then we only
cannam@154:     * must access y[0] to y[10]. y[11] must not be accessed
cannam@154:     * hence make sure len > 8 and not len >= 8
cannam@154:     */
cannam@154:    while (len > 8) {
cannam@154:       yi += 4;
cannam@154:       YY[1] = vld1q_f32(yi);
cannam@154:       yi += 4;
cannam@154:       YY[2] = vld1q_f32(yi);
cannam@154: 
cannam@154:       XX[0] = vld1q_f32(xi);
cannam@154:       xi += 4;
cannam@154:       XX[1] = vld1q_f32(xi);
cannam@154:       xi += 4;
cannam@154: 
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YY[0], vget_low_f32(XX[0]), 0);
cannam@154:       YEXT[0] = vextq_f32(YY[0], YY[1], 1);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YEXT[0], vget_low_f32(XX[0]), 1);
cannam@154:       YEXT[1] = vextq_f32(YY[0], YY[1], 2);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[0]), 0);
cannam@154:       YEXT[2] = vextq_f32(YY[0], YY[1], 3);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1);
cannam@154: 
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YY[1], vget_low_f32(XX[1]), 0);
cannam@154:       YEXT[0] = vextq_f32(YY[1], YY[2], 1);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YEXT[0], vget_low_f32(XX[1]), 1);
cannam@154:       YEXT[1] = vextq_f32(YY[1], YY[2], 2);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[1]), 0);
cannam@154:       YEXT[2] = vextq_f32(YY[1], YY[2], 3);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[1]), 1);
cannam@154: 
cannam@154:       YY[0] = YY[2];
cannam@154:       len -= 8;
cannam@154:    }
cannam@154: 
cannam@154:    /* Consume 4 elements in x vector and 8 elements in y
cannam@154:     * vector. However, the 8'th element in y never really gets
cannam@154:     * touched in this loop. So, if len == 4, then we only
cannam@154:     * must access y[0] to y[6]. y[7] must not be accessed
cannam@154:     * hence make sure len>4 and not len>=4
cannam@154:     */
cannam@154:    if (len > 4) {
cannam@154:       yi += 4;
cannam@154:       YY[1] = vld1q_f32(yi);
cannam@154: 
cannam@154:       XX[0] = vld1q_f32(xi);
cannam@154:       xi += 4;
cannam@154: 
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YY[0], vget_low_f32(XX[0]), 0);
cannam@154:       YEXT[0] = vextq_f32(YY[0], YY[1], 1);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YEXT[0], vget_low_f32(XX[0]), 1);
cannam@154:       YEXT[1] = vextq_f32(YY[0], YY[1], 2);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[0]), 0);
cannam@154:       YEXT[2] = vextq_f32(YY[0], YY[1], 3);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1);
cannam@154: 
cannam@154:       YY[0] = YY[1];
cannam@154:       len -= 4;
cannam@154:    }
cannam@154: 
cannam@154:    while (--len > 0) {
cannam@154:       XX_2 = vld1_dup_f32(xi++);
cannam@154:       SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0);
cannam@154:       YY[0]= vld1q_f32(++yi);
cannam@154:    }
cannam@154: 
cannam@154:    XX_2 = vld1_dup_f32(xi);
cannam@154:    SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0);
cannam@154: 
cannam@154:    vst1q_f32(sum, SUMM);
cannam@154: }
cannam@154: 
cannam@154: void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y,
cannam@154:                         opus_val32 *xcorr, int len, int max_pitch, int arch) {
cannam@154:    int i;
cannam@154:    (void)arch;
cannam@154:    celt_assert(max_pitch > 0);
cannam@154:    celt_sig_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0);
cannam@154: 
cannam@154:    for (i = 0; i < (max_pitch-3); i += 4) {
cannam@154:       xcorr_kernel_neon_float((const float32_t *)_x, (const float32_t *)_y+i,
cannam@154:             (float32_t *)xcorr+i, len);
cannam@154:    }
cannam@154: 
cannam@154:    /* In case max_pitch isn't a multiple of 4, do non-unrolled version. */
cannam@154:    for (; i < max_pitch; i++) {
cannam@154:       xcorr[i] = celt_inner_prod_neon(_x, _y+i, len);
cannam@154:    }
cannam@154: }
cannam@154: #endif