Chris@69: /* Copyright (c) 2014, Cisco Systems, INC
Chris@69:    Written by XiangMingZhu WeiZhou MinPeng YanWang
Chris@69: 
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: 
Chris@69:    - Redistributions of source code must retain the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer.
Chris@69: 
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: 
Chris@69:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
Chris@69:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
Chris@69:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
Chris@69:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
Chris@69:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
Chris@69:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
Chris@69:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
Chris@69:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
Chris@69:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
Chris@69:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
Chris@69:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
Chris@69: */
Chris@69: 
Chris@69: #ifdef HAVE_CONFIG_H
Chris@69: #include "config.h"
Chris@69: #endif
Chris@69: 
Chris@69: #include <xmmintrin.h>
Chris@69: #include <emmintrin.h>
Chris@69: 
Chris@69: #include "macros.h"
Chris@69: #include "celt_lpc.h"
Chris@69: #include "stack_alloc.h"
Chris@69: #include "mathops.h"
Chris@69: #include "pitch.h"
Chris@69: 
Chris@69: #if defined(OPUS_X86_MAY_HAVE_SSE4_1) && defined(FIXED_POINT)
Chris@69: #include <smmintrin.h>
Chris@69: #include "x86cpu.h"
Chris@69: 
Chris@69: opus_val32 celt_inner_prod_sse4_1(const opus_val16 *x, const opus_val16 *y,
Chris@69:       int N)
Chris@69: {
Chris@69:     opus_int  i, dataSize16;
Chris@69:     opus_int32 sum;
Chris@69:     __m128i inVec1_76543210, inVec1_FEDCBA98, acc1;
Chris@69:     __m128i inVec2_76543210, inVec2_FEDCBA98, acc2;
Chris@69:     __m128i inVec1_3210, inVec2_3210;
Chris@69: 
Chris@69:     sum = 0;
Chris@69:     dataSize16 = N & ~15;
Chris@69: 
Chris@69:     acc1 = _mm_setzero_si128();
Chris@69:     acc2 = _mm_setzero_si128();
Chris@69: 
Chris@69:     for (i=0;i<dataSize16;i+=16) {
Chris@69:         inVec1_76543210 = _mm_loadu_si128((__m128i *)(&x[i + 0]));
Chris@69:         inVec2_76543210 = _mm_loadu_si128((__m128i *)(&y[i + 0]));
Chris@69: 
Chris@69:         inVec1_FEDCBA98 = _mm_loadu_si128((__m128i *)(&x[i + 8]));
Chris@69:         inVec2_FEDCBA98 = _mm_loadu_si128((__m128i *)(&y[i + 8]));
Chris@69: 
Chris@69:         inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210);
Chris@69:         inVec1_FEDCBA98 = _mm_madd_epi16(inVec1_FEDCBA98, inVec2_FEDCBA98);
Chris@69: 
Chris@69:         acc1 = _mm_add_epi32(acc1, inVec1_76543210);
Chris@69:         acc2 = _mm_add_epi32(acc2, inVec1_FEDCBA98);
Chris@69:     }
Chris@69: 
Chris@69:     acc1 = _mm_add_epi32(acc1, acc2);
Chris@69: 
Chris@69:     if (N - i >= 8)
Chris@69:     {
Chris@69:         inVec1_76543210 = _mm_loadu_si128((__m128i *)(&x[i + 0]));
Chris@69:         inVec2_76543210 = _mm_loadu_si128((__m128i *)(&y[i + 0]));
Chris@69: 
Chris@69:         inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210);
Chris@69: 
Chris@69:         acc1 = _mm_add_epi32(acc1, inVec1_76543210);
Chris@69:         i += 8;
Chris@69:     }
Chris@69: 
Chris@69:     if (N - i >= 4)
Chris@69:     {
Chris@69:         inVec1_3210 = OP_CVTEPI16_EPI32_M64(&x[i + 0]);
Chris@69:         inVec2_3210 = OP_CVTEPI16_EPI32_M64(&y[i + 0]);
Chris@69: 
Chris@69:         inVec1_3210 = _mm_mullo_epi32(inVec1_3210, inVec2_3210);
Chris@69: 
Chris@69:         acc1 = _mm_add_epi32(acc1, inVec1_3210);
Chris@69:         i += 4;
Chris@69:     }
Chris@69: 
Chris@69:     acc1 = _mm_add_epi32(acc1, _mm_unpackhi_epi64(acc1, acc1));
Chris@69:     acc1 = _mm_add_epi32(acc1, _mm_shufflelo_epi16(acc1, 0x0E));
Chris@69: 
Chris@69:     sum += _mm_cvtsi128_si32(acc1);
Chris@69: 
Chris@69:     for (;i<N;i++)
Chris@69:     {
Chris@69:         sum = silk_SMLABB(sum, x[i], y[i]);
Chris@69:     }
Chris@69: 
Chris@69:     return sum;
Chris@69: }
Chris@69: 
Chris@69: void xcorr_kernel_sse4_1(const opus_val16 * x, const opus_val16 * y, opus_val32 sum[ 4 ], int len)
Chris@69: {
Chris@69:     int j;
Chris@69: 
Chris@69:     __m128i vecX, vecX0, vecX1, vecX2, vecX3;
Chris@69:     __m128i vecY0, vecY1, vecY2, vecY3;
Chris@69:     __m128i sum0, sum1, sum2, sum3, vecSum;
Chris@69:     __m128i initSum;
Chris@69: 
Chris@69:     celt_assert(len >= 3);
Chris@69: 
Chris@69:     sum0 = _mm_setzero_si128();
Chris@69:     sum1 = _mm_setzero_si128();
Chris@69:     sum2 = _mm_setzero_si128();
Chris@69:     sum3 = _mm_setzero_si128();
Chris@69: 
Chris@69:     for (j=0;j<(len-7);j+=8)
Chris@69:     {
Chris@69:         vecX = _mm_loadu_si128((__m128i *)(&x[j + 0]));
Chris@69:         vecY0 = _mm_loadu_si128((__m128i *)(&y[j + 0]));
Chris@69:         vecY1 = _mm_loadu_si128((__m128i *)(&y[j + 1]));
Chris@69:         vecY2 = _mm_loadu_si128((__m128i *)(&y[j + 2]));
Chris@69:         vecY3 = _mm_loadu_si128((__m128i *)(&y[j + 3]));
Chris@69: 
Chris@69:         sum0 = _mm_add_epi32(sum0, _mm_madd_epi16(vecX, vecY0));
Chris@69:         sum1 = _mm_add_epi32(sum1, _mm_madd_epi16(vecX, vecY1));
Chris@69:         sum2 = _mm_add_epi32(sum2, _mm_madd_epi16(vecX, vecY2));
Chris@69:         sum3 = _mm_add_epi32(sum3, _mm_madd_epi16(vecX, vecY3));
Chris@69:     }
Chris@69: 
Chris@69:     sum0 = _mm_add_epi32(sum0, _mm_unpackhi_epi64( sum0, sum0));
Chris@69:     sum0 = _mm_add_epi32(sum0, _mm_shufflelo_epi16( sum0, 0x0E));
Chris@69: 
Chris@69:     sum1 = _mm_add_epi32(sum1, _mm_unpackhi_epi64( sum1, sum1));
Chris@69:     sum1 = _mm_add_epi32(sum1, _mm_shufflelo_epi16( sum1, 0x0E));
Chris@69: 
Chris@69:     sum2 = _mm_add_epi32(sum2, _mm_unpackhi_epi64( sum2, sum2));
Chris@69:     sum2 = _mm_add_epi32(sum2, _mm_shufflelo_epi16( sum2, 0x0E));
Chris@69: 
Chris@69:     sum3 = _mm_add_epi32(sum3, _mm_unpackhi_epi64( sum3, sum3));
Chris@69:     sum3 = _mm_add_epi32(sum3, _mm_shufflelo_epi16( sum3, 0x0E));
Chris@69: 
Chris@69:     vecSum = _mm_unpacklo_epi64(_mm_unpacklo_epi32(sum0, sum1),
Chris@69:           _mm_unpacklo_epi32(sum2, sum3));
Chris@69: 
Chris@69:     for (;j<(len-3);j+=4)
Chris@69:     {
Chris@69:         vecX = OP_CVTEPI16_EPI32_M64(&x[j + 0]);
Chris@69:         vecX0 = _mm_shuffle_epi32(vecX, 0x00);
Chris@69:         vecX1 = _mm_shuffle_epi32(vecX, 0x55);
Chris@69:         vecX2 = _mm_shuffle_epi32(vecX, 0xaa);
Chris@69:         vecX3 = _mm_shuffle_epi32(vecX, 0xff);
Chris@69: 
Chris@69:         vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
Chris@69:         vecY1 = OP_CVTEPI16_EPI32_M64(&y[j + 1]);
Chris@69:         vecY2 = OP_CVTEPI16_EPI32_M64(&y[j + 2]);
Chris@69:         vecY3 = OP_CVTEPI16_EPI32_M64(&y[j + 3]);
Chris@69: 
Chris@69:         sum0 = _mm_mullo_epi32(vecX0, vecY0);
Chris@69:         sum1 = _mm_mullo_epi32(vecX1, vecY1);
Chris@69:         sum2 = _mm_mullo_epi32(vecX2, vecY2);
Chris@69:         sum3 = _mm_mullo_epi32(vecX3, vecY3);
Chris@69: 
Chris@69:         sum0 = _mm_add_epi32(sum0, sum1);
Chris@69:         sum2 = _mm_add_epi32(sum2, sum3);
Chris@69:         vecSum = _mm_add_epi32(vecSum, sum0);
Chris@69:         vecSum = _mm_add_epi32(vecSum, sum2);
Chris@69:     }
Chris@69: 
Chris@69:     for (;j<len;j++)
Chris@69:     {
Chris@69:         vecX = OP_CVTEPI16_EPI32_M64(&x[j + 0]);
Chris@69:         vecX0 = _mm_shuffle_epi32(vecX, 0x00);
Chris@69: 
Chris@69:         vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
Chris@69: 
Chris@69:         sum0 = _mm_mullo_epi32(vecX0, vecY0);
Chris@69:         vecSum = _mm_add_epi32(vecSum, sum0);
Chris@69:     }
Chris@69: 
Chris@69:     initSum = _mm_loadu_si128((__m128i *)(&sum[0]));
Chris@69:     initSum = _mm_add_epi32(initSum, vecSum);
Chris@69:     _mm_storeu_si128((__m128i *)sum, initSum);
Chris@69: }
Chris@69: #endif