wuggy/Dactyloidae
wuggy/Dactyloidae
1
0
Fork 0
mirror of https://repo.dactyloidae.xyz/Dactyloidae/UXP.git synced 2026-09-27 10:57:34 +09:00
Code Issues Projects Releases Packages Wiki Activity

Issue #1890 - Update libopus

Browse source 
This resolves #1890
This commit is contained in:
Moonchild 2022-05-04 17:03:37 +00:00 • committed by roytam1
commit a8b1099a1e
228 changed files with 10990 additions and 5670 deletions
Download patch file Download diff file Expand all files Collapse all files

10
media/libopus/celt/x86/celt_lpc_sse.h
View file

@ -41,12 +41,11 @@ void celt_fir_sse4_1(
opus_val16 *y,
int N,
int ord,
opus_val16 *mem,
int arch);
#if defined(OPUS_X86_PRESUME_SSE4_1)
#define celt_fir(x, num, y, N, ord, mem, arch) \
((void)arch, celt_fir_sse4_1(x, num, y, N, ord, mem, arch))
#define celt_fir(x, num, y, N, ord, arch) \
((void)arch, celt_fir_sse4_1(x, num, y, N, ord, arch))
#else
@ -56,11 +55,10 @@ extern void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])(
opus_val16 *y,
int N,
int ord,
opus_val16 *mem,
int arch);
# define celt_fir(x, num, y, N, ord, mem, arch) \
((*CELT_FIR_IMPL[(arch) & OPUS_ARCHMASK])(x, num, y, N, ord, mem, arch))
# define celt_fir(x, num, y, N, ord, arch) \
((*CELT_FIR_IMPL[(arch) & OPUS_ARCHMASK])(x, num, y, N, ord, arch))
#endif
#endif

57
media/libopus/celt/x86/celt_lpc_sse.c → media/libopus/celt/x86/celt_lpc_sse4_1.c
View file

@ -40,65 +40,23 @@
#if defined(FIXED_POINT)
void celt_fir_sse4_1(const opus_val16 *_x,
void celt_fir_sse4_1(const opus_val16 *x,
const opus_val16 *num,
opus_val16 *_y,
opus_val16 *y,
int N,
int ord,
opus_val16 *mem,
int arch)
{
int i,j;
VARDECL(opus_val16, rnum);
VARDECL(opus_val16, x);
__m128i vecNoA;
opus_int32 noA ;
SAVE_STACK;
ALLOC(rnum, ord, opus_val16);
ALLOC(x, N+ord, opus_val16);
for(i=0;i<ord;i++)
rnum[i] = num[ord-i-1];
for(i=0;i<ord;i++)
x[i] = mem[ord-i-1];
for (i=0;i<N-7;i+=8)
{
x[i+ord ]=_x[i ];
x[i+ord+1]=_x[i+1];
x[i+ord+2]=_x[i+2];
x[i+ord+3]=_x[i+3];
x[i+ord+4]=_x[i+4];
x[i+ord+5]=_x[i+5];
x[i+ord+6]=_x[i+6];
x[i+ord+7]=_x[i+7];
}
for (;i<N-3;i+=4)
{
x[i+ord ]=_x[i ];
x[i+ord+1]=_x[i+1];
x[i+ord+2]=_x[i+2];
x[i+ord+3]=_x[i+3];
}
for (;i<N;i++)
x[i+ord]=_x[i];
for(i=0;i<ord;i++)
mem[i] = _x[N-i-1];
#ifdef SMALL_FOOTPRINT
for (i=0;i<N;i++)
{
opus_val32 sum = SHL32(EXTEND32(_x[i]), SIG_SHIFT);
for (j=0;j<ord;j++)
{
sum = MAC16_16(sum,rnum[j],x[i+j]);
}
_y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT));
}
#else
noA = EXTEND32(1) << SIG_SHIFT >> 1;
vecNoA = _mm_set_epi32(noA, noA, noA, noA);
@ -107,25 +65,24 @@ void celt_fir_sse4_1(const opus_val16 *_x,
opus_val32 sums[4] = {0};
__m128i vecSum, vecX;
xcorr_kernel(rnum, x+i, sums, ord, arch);
xcorr_kernel(rnum, x+i-ord, sums, ord, arch);
vecSum = _mm_loadu_si128((__m128i *)sums);
vecSum = _mm_add_epi32(vecSum, vecNoA);
vecSum = _mm_srai_epi32(vecSum, SIG_SHIFT);
vecX = OP_CVTEPI16_EPI32_M64(_x + i);
vecX = OP_CVTEPI16_EPI32_M64(x + i);
vecSum = _mm_add_epi32(vecSum, vecX);
vecSum = _mm_packs_epi32(vecSum, vecSum);
_mm_storel_epi64((__m128i *)(_y + i), vecSum);
_mm_storel_epi64((__m128i *)(y + i), vecSum);
}
for (;i<N;i++)
{
opus_val32 sum = 0;
for (j=0;j<ord;j++)
sum = MAC16_16(sum, rnum[j], x[i + j]);
_y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT)));
sum = MAC16_16(sum, rnum[j], x[i+j-ord]);
y[i] = SATURATE16(ADD32(EXTEND32(x[i]), PSHR32(sum, SIG_SHIFT)));
}
#endif
RESTORE_STACK;
}

2
media/libopus/celt/x86/pitch_sse.h
View file

@ -91,7 +91,7 @@ opus_val32 celt_inner_prod_sse2(
int N);
#endif
#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT)
#if defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT)
opus_val32 celt_inner_prod_sse(
const opus_val16 *x,
const opus_val16 *y,

51
media/libopus/celt/x86/pitch_sse4_1.c
View file

@ -117,6 +117,14 @@ void xcorr_kernel_sse4_1(const opus_val16 * x, const opus_val16 * y, opus_val32
__m128i sum0, sum1, sum2, sum3, vecSum;
__m128i initSum;
#ifdef OPUS_CHECK_ASM
opus_val32 sum_c[4];
for (j=0;j<4;j++) {
sum_c[j] = sum[j];
}
xcorr_kernel_c(x, y, sum_c, len);
#endif
celt_assert(len >= 3);
sum0 = _mm_setzero_si128();
@ -177,19 +185,56 @@ void xcorr_kernel_sse4_1(const opus_val16 * x, const opus_val16 * y, opus_val32
vecSum = _mm_add_epi32(vecSum, sum2);
}
for (;j<len;j++)
vecX = OP_CVTEPI16_EPI32_M64(&x[len - 4]);
if (len - j == 3)
{
vecX = OP_CVTEPI16_EPI32_M64(&x[j + 0]);
vecX0 = _mm_shuffle_epi32(vecX, 0x00);
vecX0 = _mm_shuffle_epi32(vecX, 0x55);
vecX1 = _mm_shuffle_epi32(vecX, 0xaa);
vecX2 = _mm_shuffle_epi32(vecX, 0xff);
vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
vecY1 = OP_CVTEPI16_EPI32_M64(&y[j + 1]);
vecY2 = OP_CVTEPI16_EPI32_M64(&y[j + 2]);
sum0 = _mm_mullo_epi32(vecX0, vecY0);
sum1 = _mm_mullo_epi32(vecX1, vecY1);
sum2 = _mm_mullo_epi32(vecX2, vecY2);
vecSum = _mm_add_epi32(vecSum, sum0);
vecSum = _mm_add_epi32(vecSum, sum1);
vecSum = _mm_add_epi32(vecSum, sum2);
}
else if (len - j == 2)
{
vecX0 = _mm_shuffle_epi32(vecX, 0xaa);
vecX1 = _mm_shuffle_epi32(vecX, 0xff);
vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
vecY1 = OP_CVTEPI16_EPI32_M64(&y[j + 1]);
sum0 = _mm_mullo_epi32(vecX0, vecY0);
sum1 = _mm_mullo_epi32(vecX1, vecY1);
vecSum = _mm_add_epi32(vecSum, sum0);
vecSum = _mm_add_epi32(vecSum, sum1);
}
else if (len - j == 1)
{
vecX0 = _mm_shuffle_epi32(vecX, 0xff);
vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
sum0 = _mm_mullo_epi32(vecX0, vecY0);
vecSum = _mm_add_epi32(vecSum, sum0);
}
initSum = _mm_loadu_si128((__m128i *)(&sum[0]));
initSum = _mm_add_epi32(initSum, vecSum);
_mm_storeu_si128((__m128i *)sum, initSum);
#ifdef OPUS_CHECK_ASM
celt_assert(!memcmp(sum_c, sum, sizeof(sum_c)));
#endif
}
#endif

50
media/libopus/celt/x86/vq_sse.h Normal file
View file

@ -0,0 +1,50 @@
/* Copyright (c) 2016 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.
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 VQ_SSE_H
#define VQ_SSE_H
#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT)
#define OVERRIDE_OP_PVQ_SEARCH
opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch);
#if defined(OPUS_X86_PRESUME_SSE2)
#define op_pvq_search(x, iy, K, N, arch) \
(op_pvq_search_sse2(x, iy, K, N, arch))
#else
extern opus_val16 (*const OP_PVQ_SEARCH_IMPL[OPUS_ARCHMASK + 1])(
celt_norm *_X, int *iy, int K, int N, int arch);
# define op_pvq_search(X, iy, K, N, arch) \
((*OP_PVQ_SEARCH_IMPL[(arch) & OPUS_ARCHMASK])(X, iy, K, N, arch))
#endif
#endif
#endif

217
media/libopus/celt/x86/vq_sse2.c Normal file
View file

@ -0,0 +1,217 @@
/* Copyright (c) 2007-2008 CSIRO
Copyright (c) 2007-2009 Xiph.Org Foundation
Copyright (c) 2007-2016 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.
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 <xmmintrin.h>
#include <emmintrin.h>
#include "celt_lpc.h"
#include "stack_alloc.h"
#include "mathops.h"
#include "vq.h"
#include "x86cpu.h"
#ifndef FIXED_POINT
opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch)
{
int i, j;
int pulsesLeft;
float xy, yy;
VARDECL(celt_norm, y);
VARDECL(celt_norm, X);
VARDECL(float, signy);
__m128 signmask;
__m128 sums;
__m128i fours;
SAVE_STACK;
(void)arch;
/* All bits set to zero, except for the sign bit. */
signmask = _mm_set_ps1(-0.f);
fours = _mm_set_epi32(4, 4, 4, 4);
ALLOC(y, N+3, celt_norm);
ALLOC(X, N+3, celt_norm);
ALLOC(signy, N+3, float);
OPUS_COPY(X, _X, N);
X[N] = X[N+1] = X[N+2] = 0;
sums = _mm_setzero_ps();
for (j=0;j<N;j+=4)
{
__m128 x4, s4;
x4 = _mm_loadu_ps(&X[j]);
s4 = _mm_cmplt_ps(x4, _mm_setzero_ps());
/* Get rid of the sign */
x4 = _mm_andnot_ps(signmask, x4);
sums = _mm_add_ps(sums, x4);
/* Clear y and iy in case we don't do the projection. */
_mm_storeu_ps(&y[j], _mm_setzero_ps());
_mm_storeu_si128((__m128i*)&iy[j], _mm_setzero_si128());
_mm_storeu_ps(&X[j], x4);
_mm_storeu_ps(&signy[j], s4);
}
sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(1, 0, 3, 2)));
sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(2, 3, 0, 1)));
xy = yy = 0;
pulsesLeft = K;
/* Do a pre-search by projecting on the pyramid */
if (K > (N>>1))
{
__m128i pulses_sum;
__m128 yy4, xy4;
__m128 rcp4;
opus_val32 sum = _mm_cvtss_f32(sums);
/* If X is too small, just replace it with a pulse at 0 */
/* Prevents infinities and NaNs from causing too many pulses
to be allocated. 64 is an approximation of infinity here. */
if (!(sum > EPSILON && sum < 64))
{
X[0] = QCONST16(1.f,14);
j=1; do
X[j]=0;
while (++j<N);
sums = _mm_set_ps1(1.f);
}
/* Using K+e with e < 1 guarantees we cannot get more than K pulses. */
rcp4 = _mm_mul_ps(_mm_set_ps1((float)(K+.8)), _mm_rcp_ps(sums));
xy4 = yy4 = _mm_setzero_ps();
pulses_sum = _mm_setzero_si128();
for (j=0;j<N;j+=4)
{
__m128 rx4, x4, y4;
__m128i iy4;
x4 = _mm_loadu_ps(&X[j]);
rx4 = _mm_mul_ps(x4, rcp4);
iy4 = _mm_cvttps_epi32(rx4);
pulses_sum = _mm_add_epi32(pulses_sum, iy4);
_mm_storeu_si128((__m128i*)&iy[j], iy4);
y4 = _mm_cvtepi32_ps(iy4);
xy4 = _mm_add_ps(xy4, _mm_mul_ps(x4, y4));
yy4 = _mm_add_ps(yy4, _mm_mul_ps(y4, y4));
/* double the y[] vector so we don't have to do it in the search loop. */
_mm_storeu_ps(&y[j], _mm_add_ps(y4, y4));
}
pulses_sum = _mm_add_epi32(pulses_sum, _mm_shuffle_epi32(pulses_sum, _MM_SHUFFLE(1, 0, 3, 2)));
pulses_sum = _mm_add_epi32(pulses_sum, _mm_shuffle_epi32(pulses_sum, _MM_SHUFFLE(2, 3, 0, 1)));
pulsesLeft -= _mm_cvtsi128_si32(pulses_sum);
xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(1, 0, 3, 2)));
xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(2, 3, 0, 1)));
xy = _mm_cvtss_f32(xy4);
yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(1, 0, 3, 2)));
yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(2, 3, 0, 1)));
yy = _mm_cvtss_f32(yy4);
}
X[N] = X[N+1] = X[N+2] = -100;
y[N] = y[N+1] = y[N+2] = 100;
celt_sig_assert(pulsesLeft>=0);
/* This should never happen, but just in case it does (e.g. on silence)
we fill the first bin with pulses. */
if (pulsesLeft > N+3)
{
opus_val16 tmp = (opus_val16)pulsesLeft;
yy = MAC16_16(yy, tmp, tmp);
yy = MAC16_16(yy, tmp, y[0]);
iy[0] += pulsesLeft;
pulsesLeft=0;
}
for (i=0;i<pulsesLeft;i++)
{
int best_id;
__m128 xy4, yy4;
__m128 max, max2;
__m128i count;
__m128i pos;
/* The squared magnitude term gets added anyway, so we might as well
add it outside the loop */
yy = ADD16(yy, 1);
xy4 = _mm_load1_ps(&xy);
yy4 = _mm_load1_ps(&yy);
max = _mm_setzero_ps();
pos = _mm_setzero_si128();
count = _mm_set_epi32(3, 2, 1, 0);
for (j=0;j<N;j+=4)
{
__m128 x4, y4, r4;
x4 = _mm_loadu_ps(&X[j]);
y4 = _mm_loadu_ps(&y[j]);
x4 = _mm_add_ps(x4, xy4);
y4 = _mm_add_ps(y4, yy4);
y4 = _mm_rsqrt_ps(y4);
r4 = _mm_mul_ps(x4, y4);
/* Update the index of the max. */
pos = _mm_max_epi16(pos, _mm_and_si128(count, _mm_castps_si128(_mm_cmpgt_ps(r4, max))));
/* Update the max. */
max = _mm_max_ps(max, r4);
/* Update the indices (+4) */
count = _mm_add_epi32(count, fours);
}
/* Horizontal max */
max2 = _mm_max_ps(max, _mm_shuffle_ps(max, max, _MM_SHUFFLE(1, 0, 3, 2)));
max2 = _mm_max_ps(max2, _mm_shuffle_ps(max2, max2, _MM_SHUFFLE(2, 3, 0, 1)));
/* Now that max2 contains the max at all positions, look at which value(s) of the
partial max is equal to the global max. */
pos = _mm_and_si128(pos, _mm_castps_si128(_mm_cmpeq_ps(max, max2)));
pos = _mm_max_epi16(pos, _mm_unpackhi_epi64(pos, pos));
pos = _mm_max_epi16(pos, _mm_shufflelo_epi16(pos, _MM_SHUFFLE(1, 0, 3, 2)));
best_id = _mm_cvtsi128_si32(pos);
/* Updating the sums of the new pulse(s) */
xy = ADD32(xy, EXTEND32(X[best_id]));
/* We're multiplying y[j] by two so we don't have to do it here */
yy = ADD16(yy, y[best_id]);
/* Only now that we've made the final choice, update y/iy */
/* Multiplying y[j] by 2 so we don't have to do it everywhere else */
y[best_id] += 2;
iy[best_id]++;
}
/* Put the original sign back */
for (j=0;j<N;j+=4)
{
__m128i y4;
__m128i s4;
y4 = _mm_loadu_si128((__m128i*)&iy[j]);
s4 = _mm_castps_si128(_mm_loadu_ps(&signy[j]));
y4 = _mm_xor_si128(_mm_add_epi32(y4, s4), s4);
_mm_storeu_si128((__m128i*)&iy[j], y4);
}
RESTORE_STACK;
return yy;
}
#endif

14
media/libopus/celt/x86/x86_celt_map.c
View file

@ -33,6 +33,7 @@
#include "celt_lpc.h"
#include "pitch.h"
#include "pitch_sse.h"
#include "vq.h"
#if defined(OPUS_HAVE_RTCD)
@ -46,7 +47,6 @@ void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])(
opus_val16 *y,
int N,
int ord,
opus_val16 *mem,
int arch
) = {
celt_fir_c, /* non-sse */
@ -151,5 +151,17 @@ void (*const COMB_FILTER_CONST_IMPL[OPUS_ARCHMASK + 1])(
#endif
#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(OPUS_X86_PRESUME_SSE2)
opus_val16 (*const OP_PVQ_SEARCH_IMPL[OPUS_ARCHMASK + 1])(
celt_norm *_X, int *iy, int K, int N, int arch
) = {
op_pvq_search_c, /* non-sse */
op_pvq_search_c,
MAY_HAVE_SSE2(op_pvq_search),
MAY_HAVE_SSE2(op_pvq_search),
MAY_HAVE_SSE2(op_pvq_search)
};
#endif
#endif
#endif

32
media/libopus/celt/x86/x86cpu.h
View file

@ -56,38 +56,10 @@
int opus_select_arch(void);
# endif
/*gcc appears to emit MOVDQA's to load the argument of an _mm_cvtepi8_epi32()
or _mm_cvtepi16_epi32() when optimizations are disabled, even though the
actual PMOVSXWD instruction takes an m32 or m64. Unlike a normal memory
reference, these require 16-byte alignment and load a full 16 bytes (instead
of 4 or 8), possibly reading out of bounds.
We can insert an explicit MOVD or MOVQ using _mm_cvtsi32_si128() or
_mm_loadl_epi64(), which should have the same semantics as an m32 or m64
reference in the PMOVSXWD instruction itself, but gcc is not smart enough to
optimize this out when optimizations ARE enabled.
Clang, in contrast, requires us to do this always for _mm_cvtepi8_epi32
(which is fair, since technically the compiler is always allowed to do the
dereference before invoking the function implementing the intrinsic).
However, it is smart enough to eliminate the extra MOVD instruction.
For _mm_cvtepi16_epi32, it does the right thing, though does *not* optimize out
the extra MOVQ if it's specified explicitly */
# if defined(__clang__) || !defined(__OPTIMIZE__)
# define OP_CVTEPI8_EPI32_M32(x) \
#define OP_CVTEPI8_EPI32_M32(x) \
(_mm_cvtepi8_epi32(_mm_cvtsi32_si128(*(int *)(x))))
# else
# define OP_CVTEPI8_EPI32_M32(x) \
(_mm_cvtepi8_epi32(*(__m128i *)(x)))
#endif
# if !defined(__OPTIMIZE__)
# define OP_CVTEPI16_EPI32_M64(x) \
#define OP_CVTEPI16_EPI32_M64(x) \
(_mm_cvtepi16_epi32(_mm_loadl_epi64((__m128i *)(x))))
# else
# define OP_CVTEPI16_EPI32_M64(x) \
(_mm_cvtepi16_epi32(*(__m128i *)(x)))
# endif
#endif