cannam@154: /* Copyright (c) 2014, Cisco Systems, INC
cannam@154:    Written by XiangMingZhu WeiZhou MinPeng YanWang
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 <xmmintrin.h>
cannam@154: #include <emmintrin.h>
cannam@154: 
cannam@154: #include "macros.h"
cannam@154: #include "celt_lpc.h"
cannam@154: #include "stack_alloc.h"
cannam@154: #include "mathops.h"
cannam@154: #include "pitch.h"
cannam@154: 
cannam@154: #if defined(OPUS_X86_MAY_HAVE_SSE4_1) && defined(FIXED_POINT)
cannam@154: #include <smmintrin.h>
cannam@154: #include "x86cpu.h"
cannam@154: 
cannam@154: opus_val32 celt_inner_prod_sse4_1(const opus_val16 *x, const opus_val16 *y,
cannam@154:       int N)
cannam@154: {
cannam@154:     opus_int  i, dataSize16;
cannam@154:     opus_int32 sum;
cannam@154:     __m128i inVec1_76543210, inVec1_FEDCBA98, acc1;
cannam@154:     __m128i inVec2_76543210, inVec2_FEDCBA98, acc2;
cannam@154:     __m128i inVec1_3210, inVec2_3210;
cannam@154: 
cannam@154:     sum = 0;
cannam@154:     dataSize16 = N & ~15;
cannam@154: 
cannam@154:     acc1 = _mm_setzero_si128();
cannam@154:     acc2 = _mm_setzero_si128();
cannam@154: 
cannam@154:     for (i=0;i<dataSize16;i+=16) {
cannam@154:         inVec1_76543210 = _mm_loadu_si128((__m128i *)(&x[i + 0]));
cannam@154:         inVec2_76543210 = _mm_loadu_si128((__m128i *)(&y[i + 0]));
cannam@154: 
cannam@154:         inVec1_FEDCBA98 = _mm_loadu_si128((__m128i *)(&x[i + 8]));
cannam@154:         inVec2_FEDCBA98 = _mm_loadu_si128((__m128i *)(&y[i + 8]));
cannam@154: 
cannam@154:         inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210);
cannam@154:         inVec1_FEDCBA98 = _mm_madd_epi16(inVec1_FEDCBA98, inVec2_FEDCBA98);
cannam@154: 
cannam@154:         acc1 = _mm_add_epi32(acc1, inVec1_76543210);
cannam@154:         acc2 = _mm_add_epi32(acc2, inVec1_FEDCBA98);
cannam@154:     }
cannam@154: 
cannam@154:     acc1 = _mm_add_epi32(acc1, acc2);
cannam@154: 
cannam@154:     if (N - i >= 8)
cannam@154:     {
cannam@154:         inVec1_76543210 = _mm_loadu_si128((__m128i *)(&x[i + 0]));
cannam@154:         inVec2_76543210 = _mm_loadu_si128((__m128i *)(&y[i + 0]));
cannam@154: 
cannam@154:         inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210);
cannam@154: 
cannam@154:         acc1 = _mm_add_epi32(acc1, inVec1_76543210);
cannam@154:         i += 8;
cannam@154:     }
cannam@154: 
cannam@154:     if (N - i >= 4)
cannam@154:     {
cannam@154:         inVec1_3210 = OP_CVTEPI16_EPI32_M64(&x[i + 0]);
cannam@154:         inVec2_3210 = OP_CVTEPI16_EPI32_M64(&y[i + 0]);
cannam@154: 
cannam@154:         inVec1_3210 = _mm_mullo_epi32(inVec1_3210, inVec2_3210);
cannam@154: 
cannam@154:         acc1 = _mm_add_epi32(acc1, inVec1_3210);
cannam@154:         i += 4;
cannam@154:     }
cannam@154: 
cannam@154:     acc1 = _mm_add_epi32(acc1, _mm_unpackhi_epi64(acc1, acc1));
cannam@154:     acc1 = _mm_add_epi32(acc1, _mm_shufflelo_epi16(acc1, 0x0E));
cannam@154: 
cannam@154:     sum += _mm_cvtsi128_si32(acc1);
cannam@154: 
cannam@154:     for (;i<N;i++)
cannam@154:     {
cannam@154:         sum = silk_SMLABB(sum, x[i], y[i]);
cannam@154:     }
cannam@154: 
cannam@154:     return sum;
cannam@154: }
cannam@154: 
cannam@154: void xcorr_kernel_sse4_1(const opus_val16 * x, const opus_val16 * y, opus_val32 sum[ 4 ], int len)
cannam@154: {
cannam@154:     int j;
cannam@154: 
cannam@154:     __m128i vecX, vecX0, vecX1, vecX2, vecX3;
cannam@154:     __m128i vecY0, vecY1, vecY2, vecY3;
cannam@154:     __m128i sum0, sum1, sum2, sum3, vecSum;
cannam@154:     __m128i initSum;
cannam@154: 
cannam@154:     celt_assert(len >= 3);
cannam@154: 
cannam@154:     sum0 = _mm_setzero_si128();
cannam@154:     sum1 = _mm_setzero_si128();
cannam@154:     sum2 = _mm_setzero_si128();
cannam@154:     sum3 = _mm_setzero_si128();
cannam@154: 
cannam@154:     for (j=0;j<(len-7);j+=8)
cannam@154:     {
cannam@154:         vecX = _mm_loadu_si128((__m128i *)(&x[j + 0]));
cannam@154:         vecY0 = _mm_loadu_si128((__m128i *)(&y[j + 0]));
cannam@154:         vecY1 = _mm_loadu_si128((__m128i *)(&y[j + 1]));
cannam@154:         vecY2 = _mm_loadu_si128((__m128i *)(&y[j + 2]));
cannam@154:         vecY3 = _mm_loadu_si128((__m128i *)(&y[j + 3]));
cannam@154: 
cannam@154:         sum0 = _mm_add_epi32(sum0, _mm_madd_epi16(vecX, vecY0));
cannam@154:         sum1 = _mm_add_epi32(sum1, _mm_madd_epi16(vecX, vecY1));
cannam@154:         sum2 = _mm_add_epi32(sum2, _mm_madd_epi16(vecX, vecY2));
cannam@154:         sum3 = _mm_add_epi32(sum3, _mm_madd_epi16(vecX, vecY3));
cannam@154:     }
cannam@154: 
cannam@154:     sum0 = _mm_add_epi32(sum0, _mm_unpackhi_epi64( sum0, sum0));
cannam@154:     sum0 = _mm_add_epi32(sum0, _mm_shufflelo_epi16( sum0, 0x0E));
cannam@154: 
cannam@154:     sum1 = _mm_add_epi32(sum1, _mm_unpackhi_epi64( sum1, sum1));
cannam@154:     sum1 = _mm_add_epi32(sum1, _mm_shufflelo_epi16( sum1, 0x0E));
cannam@154: 
cannam@154:     sum2 = _mm_add_epi32(sum2, _mm_unpackhi_epi64( sum2, sum2));
cannam@154:     sum2 = _mm_add_epi32(sum2, _mm_shufflelo_epi16( sum2, 0x0E));
cannam@154: 
cannam@154:     sum3 = _mm_add_epi32(sum3, _mm_unpackhi_epi64( sum3, sum3));
cannam@154:     sum3 = _mm_add_epi32(sum3, _mm_shufflelo_epi16( sum3, 0x0E));
cannam@154: 
cannam@154:     vecSum = _mm_unpacklo_epi64(_mm_unpacklo_epi32(sum0, sum1),
cannam@154:           _mm_unpacklo_epi32(sum2, sum3));
cannam@154: 
cannam@154:     for (;j<(len-3);j+=4)
cannam@154:     {
cannam@154:         vecX = OP_CVTEPI16_EPI32_M64(&x[j + 0]);
cannam@154:         vecX0 = _mm_shuffle_epi32(vecX, 0x00);
cannam@154:         vecX1 = _mm_shuffle_epi32(vecX, 0x55);
cannam@154:         vecX2 = _mm_shuffle_epi32(vecX, 0xaa);
cannam@154:         vecX3 = _mm_shuffle_epi32(vecX, 0xff);
cannam@154: 
cannam@154:         vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
cannam@154:         vecY1 = OP_CVTEPI16_EPI32_M64(&y[j + 1]);
cannam@154:         vecY2 = OP_CVTEPI16_EPI32_M64(&y[j + 2]);
cannam@154:         vecY3 = OP_CVTEPI16_EPI32_M64(&y[j + 3]);
cannam@154: 
cannam@154:         sum0 = _mm_mullo_epi32(vecX0, vecY0);
cannam@154:         sum1 = _mm_mullo_epi32(vecX1, vecY1);
cannam@154:         sum2 = _mm_mullo_epi32(vecX2, vecY2);
cannam@154:         sum3 = _mm_mullo_epi32(vecX3, vecY3);
cannam@154: 
cannam@154:         sum0 = _mm_add_epi32(sum0, sum1);
cannam@154:         sum2 = _mm_add_epi32(sum2, sum3);
cannam@154:         vecSum = _mm_add_epi32(vecSum, sum0);
cannam@154:         vecSum = _mm_add_epi32(vecSum, sum2);
cannam@154:     }
cannam@154: 
cannam@154:     for (;j<len;j++)
cannam@154:     {
cannam@154:         vecX = OP_CVTEPI16_EPI32_M64(&x[j + 0]);
cannam@154:         vecX0 = _mm_shuffle_epi32(vecX, 0x00);
cannam@154: 
cannam@154:         vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
cannam@154: 
cannam@154:         sum0 = _mm_mullo_epi32(vecX0, vecY0);
cannam@154:         vecSum = _mm_add_epi32(vecSum, sum0);
cannam@154:     }
cannam@154: 
cannam@154:     initSum = _mm_loadu_si128((__m128i *)(&sum[0]));
cannam@154:     initSum = _mm_add_epi32(initSum, vecSum);
cannam@154:     _mm_storeu_si128((__m128i *)sum, initSum);
cannam@154: }
cannam@154: #endif