Chris@69: /***********************************************************************
Chris@69: Copyright (C) 2014 Vidyo
Chris@69: Redistribution and use in source and binary forms, with or without
Chris@69: modification, are permitted provided that the following conditions
Chris@69: are met:
Chris@69: - Redistributions of source code must retain the above copyright notice,
Chris@69: this list of conditions and the following disclaimer.
Chris@69: - Redistributions in binary form must reproduce the above copyright
Chris@69: notice, this list of conditions and the following disclaimer in the
Chris@69: documentation and/or other materials provided with the distribution.
Chris@69: - Neither the name of Internet Society, IETF or IETF Trust, nor the
Chris@69: names of specific contributors, may be used to endorse or promote
Chris@69: products derived from this software without specific prior written
Chris@69: permission.
Chris@69: THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
Chris@69: AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
Chris@69: IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
Chris@69: ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
Chris@69: LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
Chris@69: CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
Chris@69: SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
Chris@69: INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
Chris@69: CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
Chris@69: ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
Chris@69: POSSIBILITY OF SUCH DAMAGE.
Chris@69: ***********************************************************************/
Chris@69: #ifdef HAVE_CONFIG_H
Chris@69: #include "config.h"
Chris@69: #endif
Chris@69: 
Chris@69: #include <arm_neon.h>
Chris@69: #include "main.h"
Chris@69: #include "stack_alloc.h"
Chris@69: #include "NSQ.h"
Chris@69: #include "celt/cpu_support.h"
Chris@69: #include "celt/arm/armcpu.h"
Chris@69: 
Chris@69: opus_int32 silk_noise_shape_quantizer_short_prediction_neon(const opus_int32 *buf32, const opus_int32 *coef32, opus_int order)
Chris@69: {
Chris@69:     int32x4_t coef0 = vld1q_s32(coef32);
Chris@69:     int32x4_t coef1 = vld1q_s32(coef32 + 4);
Chris@69:     int32x4_t coef2 = vld1q_s32(coef32 + 8);
Chris@69:     int32x4_t coef3 = vld1q_s32(coef32 + 12);
Chris@69: 
Chris@69:     int32x4_t a0 = vld1q_s32(buf32 - 15);
Chris@69:     int32x4_t a1 = vld1q_s32(buf32 - 11);
Chris@69:     int32x4_t a2 = vld1q_s32(buf32 - 7);
Chris@69:     int32x4_t a3 = vld1q_s32(buf32 - 3);
Chris@69: 
Chris@69:     int32x4_t b0 = vqdmulhq_s32(coef0, a0);
Chris@69:     int32x4_t b1 = vqdmulhq_s32(coef1, a1);
Chris@69:     int32x4_t b2 = vqdmulhq_s32(coef2, a2);
Chris@69:     int32x4_t b3 = vqdmulhq_s32(coef3, a3);
Chris@69: 
Chris@69:     int32x4_t c0 = vaddq_s32(b0, b1);
Chris@69:     int32x4_t c1 = vaddq_s32(b2, b3);
Chris@69: 
Chris@69:     int32x4_t d = vaddq_s32(c0, c1);
Chris@69: 
Chris@69:     int64x2_t e = vpaddlq_s32(d);
Chris@69: 
Chris@69:     int64x1_t f = vadd_s64(vget_low_s64(e), vget_high_s64(e));
Chris@69: 
Chris@69:     opus_int32 out = vget_lane_s32(vreinterpret_s32_s64(f), 0);
Chris@69: 
Chris@69:     out += silk_RSHIFT( order, 1 );
Chris@69: 
Chris@69:     return out;
Chris@69: }
Chris@69: 
Chris@69: 
Chris@69: opus_int32 silk_NSQ_noise_shape_feedback_loop_neon(const opus_int32 *data0, opus_int32 *data1, const opus_int16 *coef, opus_int order)
Chris@69: {
Chris@69:     opus_int32 out;
Chris@69:     if (order == 8)
Chris@69:     {
Chris@69:         int32x4_t a00 = vdupq_n_s32(data0[0]);
Chris@69:         int32x4_t a01 = vld1q_s32(data1);  /* data1[0] ... [3] */
Chris@69: 
Chris@69:         int32x4_t a0 = vextq_s32 (a00, a01, 3); /* data0[0] data1[0] ...[2] */
Chris@69:         int32x4_t a1 = vld1q_s32(data1 + 3);  /* data1[3] ... [6] */
Chris@69: 
Chris@69:         /*TODO: Convert these once in advance instead of once per sample, like
Chris@69:           silk_noise_shape_quantizer_short_prediction_neon() does.*/
Chris@69:         int16x8_t coef16 = vld1q_s16(coef);
Chris@69:         int32x4_t coef0 = vmovl_s16(vget_low_s16(coef16));
Chris@69:         int32x4_t coef1 = vmovl_s16(vget_high_s16(coef16));
Chris@69: 
Chris@69:         /*This is not bit-exact with the C version, since we do not drop the
Chris@69:           lower 16 bits of each multiply, but wait until the end to truncate
Chris@69:           precision. This is an encoder-specific calculation (and unlike
Chris@69:           silk_noise_shape_quantizer_short_prediction_neon(), is not meant to
Chris@69:           simulate what the decoder will do). We still could use vqdmulhq_s32()
Chris@69:           like silk_noise_shape_quantizer_short_prediction_neon() and save
Chris@69:           half the multiplies, but the speed difference is not large, since we
Chris@69:           then need two extra adds.*/
Chris@69:         int64x2_t b0 = vmull_s32(vget_low_s32(a0), vget_low_s32(coef0));
Chris@69:         int64x2_t b1 = vmlal_s32(b0, vget_high_s32(a0), vget_high_s32(coef0));
Chris@69:         int64x2_t b2 = vmlal_s32(b1, vget_low_s32(a1), vget_low_s32(coef1));
Chris@69:         int64x2_t b3 = vmlal_s32(b2, vget_high_s32(a1), vget_high_s32(coef1));
Chris@69: 
Chris@69:         int64x1_t c = vadd_s64(vget_low_s64(b3), vget_high_s64(b3));
Chris@69:         int64x1_t cS = vrshr_n_s64(c, 15);
Chris@69:         int32x2_t d = vreinterpret_s32_s64(cS);
Chris@69: 
Chris@69:         out = vget_lane_s32(d, 0);
Chris@69:         vst1q_s32(data1, a0);
Chris@69:         vst1q_s32(data1 + 4, a1);
Chris@69:         return out;
Chris@69:     }
Chris@69:     return silk_NSQ_noise_shape_feedback_loop_c(data0, data1, coef, order);
Chris@69: }