int32_t dot_product(int16_t *x,
int16_t *y,
uint32_t N, //must be a multiple of 8
uint8_t output_shift)
{
uint32_t n;
#if defined(__x86_64__) || defined(__i386__)
__m128i *x128,*y128,mmtmp1,mmtmp2,mmtmp3,mmcumul,mmcumul_re,mmcumul_im;
__m64 mmtmp7;
__m128i minus_i = _mm_set_epi16(-1,1,-1,1,-1,1,-1,1);
int32_t result;
x128 = (__m128i*) x;
y128 = (__m128i*) y;
mmcumul_re = _mm_setzero_si128();
mmcumul_im = _mm_setzero_si128();
for (n=0; n<(N>>2); n++) {
//printf("n=%d, x128=%p, y128=%p\n",n,x128,y128);
// print_shorts("x",&x128[0]);
// print_shorts("y",&y128[0]);
// this computes Re(z) = Re(x)*Re(y) + Im(x)*Im(y)
mmtmp1 = _mm_madd_epi16(x128[0],y128[0]);
// print_ints("re",&mmtmp1);
// mmtmp1 contains real part of 4 consecutive outputs (32-bit)
// shift and accumulate results
mmtmp1 = _mm_srai_epi32(mmtmp1,output_shift);
mmcumul_re = _mm_add_epi32(mmcumul_re,mmtmp1);
// print_ints("re",&mmcumul_re);
// this computes Im(z) = Re(x)*Im(y) - Re(y)*Im(x)
mmtmp2 = _mm_shufflelo_epi16(y128[0],_MM_SHUFFLE(2,3,0,1));
// print_shorts("y",&mmtmp2);
mmtmp2 = _mm_shufflehi_epi16(mmtmp2,_MM_SHUFFLE(2,3,0,1));
// print_shorts("y",&mmtmp2);
mmtmp2 = _mm_sign_epi16(mmtmp2,minus_i);
// print_shorts("y",&mmtmp2);
mmtmp3 = _mm_madd_epi16(x128[0],mmtmp2);
// print_ints("im",&mmtmp3);
// mmtmp3 contains imag part of 4 consecutive outputs (32-bit)
// shift and accumulate results
mmtmp3 = _mm_srai_epi32(mmtmp3,output_shift);
mmcumul_im = _mm_add_epi32(mmcumul_im,mmtmp3);
// print_ints("im",&mmcumul_im);
x128++;
y128++;
}
// this gives Re Re Im Im
mmcumul = _mm_hadd_epi32(mmcumul_re,mmcumul_im);
// print_ints("cumul1",&mmcumul);
// this gives Re Im Re Im
mmcumul = _mm_hadd_epi32(mmcumul,mmcumul);
// print_ints("cumul2",&mmcumul);
//mmcumul = _mm_srai_epi32(mmcumul,output_shift);
// extract the lower half
mmtmp7 = _mm_movepi64_pi64(mmcumul);
// print_ints("mmtmp7",&mmtmp7);
// pack the result
mmtmp7 = _mm_packs_pi32(mmtmp7,mmtmp7);
// print_shorts("mmtmp7",&mmtmp7);
// convert back to integer
result = _mm_cvtsi64_si32(mmtmp7);
_mm_empty();
_m_empty();
return(result);
#elif defined(__arm__)
int16x4_t *x_128=(int16x4_t*)x;
int16x4_t *y_128=(int16x4_t*)y;
int32x4_t tmp_re,tmp_im;
int32x4_t tmp_re1,tmp_im1;
int32x4_t re_cumul,im_cumul;
int32x2_t re_cumul2,im_cumul2;
int32x4_t shift = vdupq_n_s32(-output_shift);
int32x2x2_t result2;
int16_t conjug[4]__attribute__((aligned(16))) = {-1,1,-1,1} ;
re_cumul = vdupq_n_s32(0);
im_cumul = vdupq_n_s32(0);
for (n=0; n<(N>>2); n++) {
tmp_re = vmull_s16(*x_128++, *y_128++);
//tmp_re = [Re(x[0])Re(y[0]) Im(x[0])Im(y[0]) Re(x[1])Re(y[1]) Im(x[1])Im(y[1])]
tmp_re1 = vmull_s16(*x_128++, *y_128++);
//tmp_re1 = [Re(x1[1])Re(x2[1]) Im(x1[1])Im(x2[1]) Re(x1[1])Re(x2[2]) Im(x1[1])Im(x2[2])]
tmp_re = vcombine_s32(vpadd_s32(vget_low_s32(tmp_re),vget_high_s32(tmp_re)),
vpadd_s32(vget_low_s32(tmp_re1),vget_high_s32(tmp_re1)));
//tmp_re = [Re(ch[0])Re(rx[0])+Im(ch[0])Im(ch[0]) Re(ch[1])Re(rx[1])+Im(ch[1])Im(ch[1]) Re(ch[2])Re(rx[2])+Im(ch[2]) Im(ch[2]) Re(ch[3])Re(rx[3])+Im(ch[3])Im(ch[3])]
tmp_im = vmull_s16(vrev32_s16(vmul_s16(*x_128++,*(int16x4_t*)conjug)),*y_128++);
//tmp_im = [-Im(ch[0])Re(rx[0]) Re(ch[0])Im(rx[0]) -Im(ch[1])Re(rx[1]) Re(ch[1])Im(rx[1])]
tmp_im1 = vmull_s16(vrev32_s16(vmul_s16(*x_128++,*(int16x4_t*)conjug)),*y_128++);
//tmp_im1 = [-Im(ch[2])Re(rx[2]) Re(ch[2])Im(rx[2]) -Im(ch[3])Re(rx[3]) Re(ch[3])Im(rx[3])]
tmp_im = vcombine_s32(vpadd_s32(vget_low_s32(tmp_im),vget_high_s32(tmp_im)),
vpadd_s32(vget_low_s32(tmp_im1),vget_high_s32(tmp_im1)));
//tmp_im = [-Im(ch[0])Re(rx[0])+Re(ch[0])Im(rx[0]) -Im(ch[1])Re(rx[1])+Re(ch[1])Im(rx[1]) -Im(ch[2])Re(rx[2])+Re(ch[2])Im(rx[2]) -Im(ch[3])Re(rx[3])+Re(ch[3])Im(rx[3])]
re_cumul = vqaddq_s32(re_cumul,vqshlq_s32(tmp_re,shift));
im_cumul = vqaddq_s32(im_cumul,vqshlq_s32(tmp_im,shift));
}
re_cumul2 = vpadd_s32(vget_low_s32(re_cumul),vget_high_s32(re_cumul));
im_cumul2 = vpadd_s32(vget_low_s32(im_cumul),vget_high_s32(im_cumul));
re_cumul2 = vpadd_s32(re_cumul2,re_cumul2);
im_cumul2 = vpadd_s32(im_cumul2,im_cumul2);
result2 = vzip_s32(re_cumul2,im_cumul2);
return(vget_lane_s32(result2.val[0],0));
#endif
}
static inline void silk_biquad_alt_stride2_kernel( const int32x4_t A_L_s32x4, const int32x4_t A_U_s32x4, const int32x4_t B_Q28_s32x4, const int32x2_t t_s32x2, const int32x4_t in_s32x4, int32x4_t *S_s32x4, int32x2_t *out32_Q14_s32x2 )
{
int32x4_t t_s32x4, out32_Q14_s32x4;
*out32_Q14_s32x2 = vadd_s32( vget_low_s32( *S_s32x4 ), t_s32x2 ); /* silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} ) */
*S_s32x4 = vcombine_s32( vget_high_s32( *S_s32x4 ), vdup_n_s32( 0 ) ); /* S{0,1} = S{2,3}; S{2,3} = 0; */
*out32_Q14_s32x2 = vshl_n_s32( *out32_Q14_s32x2, 2 ); /* out32_Q14_{0,1} = silk_LSHIFT( silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} ), 2 ); */
out32_Q14_s32x4 = vcombine_s32( *out32_Q14_s32x2, *out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} */
t_s32x4 = vqdmulhq_s32( out32_Q14_s32x4, A_L_s32x4 ); /* silk_SMULWB( out32_Q14_{0,1,0,1}, A{0,0,1,1}_L_Q28 ) */
*S_s32x4 = vrsraq_n_s32( *S_s32x4, t_s32x4, 14 ); /* S{0,1} = S{2,3} + silk_RSHIFT_ROUND(); S{2,3} = silk_RSHIFT_ROUND(); */
t_s32x4 = vqdmulhq_s32( out32_Q14_s32x4, A_U_s32x4 ); /* silk_SMULWB( out32_Q14_{0,1,0,1}, A{0,0,1,1}_U_Q28 ) */
*S_s32x4 = vaddq_s32( *S_s32x4, t_s32x4 ); /* S0 = silk_SMLAWB( S{0,1,2,3}, out32_Q14_{0,1,0,1}, A{0,0,1,1}_U_Q28 ); */
t_s32x4 = vqdmulhq_s32( in_s32x4, B_Q28_s32x4 ); /* silk_SMULWB( B_Q28[ {1,1,2,2} ], in{0,1,0,1} ) */
*S_s32x4 = vaddq_s32( *S_s32x4, t_s32x4 ); /* S0 = silk_SMLAWB( S0, B_Q28[ {1,1,2,2} ], in{0,1,0,1} ); */
}
inline int v_signmask(const v_uint32x4& a)
{
int32x2_t m0 = vcreate_s32(CV_BIG_UINT(0x0000000100000000));
uint32x4_t v0 = vshlq_u32(vshrq_n_u32(a.val, 31), vcombine_s32(m0, m0));
uint64x2_t v1 = vpaddlq_u32(v0);
return (int)vgetq_lane_u64(v1, 0) + ((int)vgetq_lane_u64(v1, 1) << 2);
}
static INLINE void iadst_half_butterfly_bd12_neon(int32x4_t *const x,
const int32x2_t c) {
const int32x4_t sum = vaddq_s32(x[0], x[1]);
const int32x4_t sub = vsubq_s32(x[0], x[1]);
const int64x2_t t0_lo = vmull_lane_s32(vget_low_s32(sum), c, 0);
const int64x2_t t1_lo = vmull_lane_s32(vget_low_s32(sub), c, 0);
const int64x2_t t0_hi = vmull_lane_s32(vget_high_s32(sum), c, 0);
const int64x2_t t1_hi = vmull_lane_s32(vget_high_s32(sub), c, 0);
const int32x2_t out0_lo = vrshrn_n_s64(t0_lo, DCT_CONST_BITS);
const int32x2_t out1_lo = vrshrn_n_s64(t1_lo, DCT_CONST_BITS);
const int32x2_t out0_hi = vrshrn_n_s64(t0_hi, DCT_CONST_BITS);
const int32x2_t out1_hi = vrshrn_n_s64(t1_hi, DCT_CONST_BITS);
x[0] = vcombine_s32(out0_lo, out0_hi);
x[1] = vcombine_s32(out1_lo, out1_hi);
}
static INLINE int32x4_t sub_dct_const_round_shift_low_8_bd12(
const int64x2_t *const in0, const int64x2_t *const in1) {
const int64x2_t sub_lo = vsubq_s64(in0[0], in1[0]);
const int64x2_t sub_hi = vsubq_s64(in0[1], in1[1]);
const int32x2_t out_lo = vrshrn_n_s64(sub_lo, DCT_CONST_BITS);
const int32x2_t out_hi = vrshrn_n_s64(sub_hi, DCT_CONST_BITS);
return vcombine_s32(out_lo, out_hi);
}
int32x4_t
bar (int64_t x)
{
int32x2_t i = vcreate_s32 (x);
int32x2_t zeroes = vcreate_s32 (0l);
int32x4_t ret = vcombine_s32 (i, zeroes);
return ret;
}
int32x4_t
foo (int32x2_t *x)
{
int32x2_t i = *x;
int32x2_t zeroes = vcreate_s32 (0l);
int32x4_t ret = vcombine_s32 (i, zeroes);
return ret;
}
void test_vcombines32 (void)
{
int32x4_t out_int32x4_t;
int32x2_t arg0_int32x2_t;
int32x2_t arg1_int32x2_t;
out_int32x4_t = vcombine_s32 (arg0_int32x2_t, arg1_int32x2_t);
}
static INLINE void idct4x4_16_kernel_bd12(const int32x4_t cospis,
int32x4_t *const a0,
int32x4_t *const a1,
int32x4_t *const a2,
int32x4_t *const a3) {
int32x4_t b0, b1, b2, b3;
int64x2_t c0, c1, c2, c3, c4, c5, c6, c7, c8, c9, c10, c11;
transpose_s32_4x4(a0, a1, a2, a3);
b0 = vaddq_s32(*a0, *a2);
b1 = vsubq_s32(*a0, *a2);
c0 = vmull_lane_s32(vget_low_s32(b0), vget_high_s32(cospis), 0);
c1 = vmull_lane_s32(vget_high_s32(b0), vget_high_s32(cospis), 0);
c2 = vmull_lane_s32(vget_low_s32(b1), vget_high_s32(cospis), 0);
c3 = vmull_lane_s32(vget_high_s32(b1), vget_high_s32(cospis), 0);
c4 = vmull_lane_s32(vget_low_s32(*a1), vget_high_s32(cospis), 1);
c5 = vmull_lane_s32(vget_high_s32(*a1), vget_high_s32(cospis), 1);
c6 = vmull_lane_s32(vget_low_s32(*a1), vget_low_s32(cospis), 1);
c7 = vmull_lane_s32(vget_high_s32(*a1), vget_low_s32(cospis), 1);
c8 = vmull_lane_s32(vget_low_s32(*a3), vget_low_s32(cospis), 1);
c9 = vmull_lane_s32(vget_high_s32(*a3), vget_low_s32(cospis), 1);
c10 = vmull_lane_s32(vget_low_s32(*a3), vget_high_s32(cospis), 1);
c11 = vmull_lane_s32(vget_high_s32(*a3), vget_high_s32(cospis), 1);
c4 = vsubq_s64(c4, c8);
c5 = vsubq_s64(c5, c9);
c6 = vaddq_s64(c6, c10);
c7 = vaddq_s64(c7, c11);
b0 = vcombine_s32(vrshrn_n_s64(c0, DCT_CONST_BITS),
vrshrn_n_s64(c1, DCT_CONST_BITS));
b1 = vcombine_s32(vrshrn_n_s64(c2, DCT_CONST_BITS),
vrshrn_n_s64(c3, DCT_CONST_BITS));
b2 = vcombine_s32(vrshrn_n_s64(c4, DCT_CONST_BITS),
vrshrn_n_s64(c5, DCT_CONST_BITS));
b3 = vcombine_s32(vrshrn_n_s64(c6, DCT_CONST_BITS),
vrshrn_n_s64(c7, DCT_CONST_BITS));
*a0 = vaddq_s32(b0, b3);
*a1 = vaddq_s32(b1, b2);
*a2 = vsubq_s32(b1, b2);
*a3 = vsubq_s32(b0, b3);
}
void dotProd_i16_neon(const float *dataf, const float *weightsf, float *vals, const int n, const int len, const float *istd) {
const int16_t *data = (const int16_t *)dataf;
const int16_t *weights = (const int16_t *)weightsf;
weightsf += n * len / 2; // sizeof(float) / sizeof(int16_t)
for (int i = 0; i < n; i += 4) {
int32x4_t accum0 = { 0, 0, 0, 0 };
int32x4_t accum1 = accum0;
int32x4_t accum2 = accum0;
int32x4_t accum3 = accum0;
for (int j = 0; j < len; j += 8) {
int16x4x2_t d0 = vld2_s16(data + j);
int16x4x2_t w0 = vld2_s16(weights);
int16x4x2_t w1 = vld2_s16(weights + 8);
int16x4x2_t w2 = vld2_s16(weights + 16);
int16x4x2_t w3 = vld2_s16(weights + 24);
accum0 = vmlal_s16(accum0, d0.val[0], w0.val[0]);
accum0 = vmlal_s16(accum0, d0.val[1], w0.val[1]);
accum1 = vmlal_s16(accum1, d0.val[0], w1.val[0]);
accum1 = vmlal_s16(accum1, d0.val[1], w1.val[1]);
accum2 = vmlal_s16(accum2, d0.val[0], w2.val[0]);
accum2 = vmlal_s16(accum2, d0.val[1], w2.val[1]);
accum3 = vmlal_s16(accum3, d0.val[0], w3.val[0]);
accum3 = vmlal_s16(accum3, d0.val[1], w3.val[1]);
weights += 32;
}
int32x2_t sum0 = vpadd_s32(vget_low_s32(accum0), vget_high_s32(accum0));
int32x2_t sum1 = vpadd_s32(vget_low_s32(accum1), vget_high_s32(accum1));
int32x2_t sum2 = vpadd_s32(vget_low_s32(accum2), vget_high_s32(accum2));
int32x2_t sum3 = vpadd_s32(vget_low_s32(accum3), vget_high_s32(accum3));
sum0 = vpadd_s32(sum0, sum1);
sum1 = vpadd_s32(sum2, sum3);
int32x4_t sum = vcombine_s32(sum0, sum1);
float32x4_t val = vcvtq_f32_s32(sum);
val = vmulq_f32(val, vld1q_f32(weightsf + i*2));
val = vmulq_n_f32(val, istd[0]);
val = vaddq_f32(val, vld1q_f32(weightsf + i*2 + 4));
vst1q_f32(vals + i, val);
}
}
int
main (void)
{
int64_t arg1;
int32_t arg2;
int32x4_t arg3;
int64_t actual;
int64_t expected;
arg1 = -9223182289494545592LL;
arg2 = 32768;
arg3 = vcombine_s32 (vcreate_s32 (0xffff7fff8000ffffULL),
vcreate_s32 (0x80000000ffff0000ULL));
actual = vqdmlals_laneq_s32 (arg1, arg2, arg3, 3);
expected = -9223323026982900920LL;
if (expected != actual)
abort ();
return 0;
}
int
main (void)
{
int64_t arg1;
int32_t arg2;
int32x4_t arg3;
int64_t actual;
int64_t expected;
arg1 = 140733193453567LL;
arg2 = 25544;
arg3 = vcombine_s32 (vcreate_s32 (0x417b8000ffff8397LL),
vcreate_s32 (0x7fffffff58488000LL));
actual = vqdmlsls_laneq_s32 (arg1, arg2, arg3, 3);
expected = 31022548895631LL;
if (expected != actual)
abort ();
return 0;
}
int rotate_cpx_vector(int16_t *x,
int16_t *alpha,
int16_t *y,
uint32_t N,
uint16_t output_shift)
{
// Multiply elementwise two complex vectors of N elements
// x - input 1 in the format |Re0 Im0 |,......,|Re(N-1) Im(N-1)|
// We assume x1 with a dynamic of 15 bit maximum
//
// alpha - input 2 in the format |Re0 Im0|
// We assume x2 with a dynamic of 15 bit maximum
//
// y - output in the format |Re0 Im0|,......,|Re(N-1) Im(N-1)|
//
// N - the size f the vectors (this function does N cpx mpy. WARNING: N>=4;
//
// log2_amp - increase the output amplitude by a factor 2^log2_amp (default is 0)
// WARNING: log2_amp>0 can cause overflow!!
uint32_t i; // loop counter
simd_q15_t *y_128,alpha_128;
int32_t *xd=(int32_t *)x;
#if defined(__x86_64__) || defined(__i386__)
__m128i shift = _mm_cvtsi32_si128(output_shift);
register simd_q15_t m0,m1,m2,m3;
((int16_t *)&alpha_128)[0] = alpha[0];
((int16_t *)&alpha_128)[1] = -alpha[1];
((int16_t *)&alpha_128)[2] = alpha[1];
((int16_t *)&alpha_128)[3] = alpha[0];
((int16_t *)&alpha_128)[4] = alpha[0];
((int16_t *)&alpha_128)[5] = -alpha[1];
((int16_t *)&alpha_128)[6] = alpha[1];
((int16_t *)&alpha_128)[7] = alpha[0];
#elif defined(__arm__)
int32x4_t shift;
int32x4_t ab_re0,ab_re1,ab_im0,ab_im1,re32,im32;
int16_t reflip[8] __attribute__((aligned(16))) = {1,-1,1,-1,1,-1,1,-1};
int32x4x2_t xtmp;
((int16_t *)&alpha_128)[0] = alpha[0];
((int16_t *)&alpha_128)[1] = alpha[1];
((int16_t *)&alpha_128)[2] = alpha[0];
((int16_t *)&alpha_128)[3] = alpha[1];
((int16_t *)&alpha_128)[4] = alpha[0];
((int16_t *)&alpha_128)[5] = alpha[1];
((int16_t *)&alpha_128)[6] = alpha[0];
((int16_t *)&alpha_128)[7] = alpha[1];
int16x8_t bflip = vrev32q_s16(alpha_128);
int16x8_t bconj = vmulq_s16(alpha_128,*(int16x8_t *)reflip);
shift = vdupq_n_s32(-output_shift);
#endif
y_128 = (simd_q15_t *) y;
for(i=0; i<N>>2; i++) {
#if defined(__x86_64__) || defined(__i386__)
m0 = _mm_setr_epi32(xd[0],xd[0],xd[1],xd[1]);
m1 = _mm_setr_epi32(xd[2],xd[2],xd[3],xd[3]);
m2 = _mm_madd_epi16(m0,alpha_128); //complex multiply. result is 32bit [Re Im Re Im]
m3 = _mm_madd_epi16(m1,alpha_128); //complex multiply. result is 32bit [Re Im Re Im]
m2 = _mm_sra_epi32(m2,shift); // shift right by shift in order to compensate for the input amplitude
m3 = _mm_sra_epi32(m3,shift); // shift right by shift in order to compensate for the input amplitude
y_128[0] = _mm_packs_epi32(m2,m3); // pack in 16bit integers with saturation [re im re im re im re im]
#elif defined(__arm__)
ab_re0 = vmull_s16(((int16x4_t*)xd)[0],((int16x4_t*)&bconj)[0]);
ab_re1 = vmull_s16(((int16x4_t*)xd)[1],((int16x4_t*)&bconj)[1]);
ab_im0 = vmull_s16(((int16x4_t*)xd)[0],((int16x4_t*)&bflip)[0]);
ab_im1 = vmull_s16(((int16x4_t*)xd)[1],((int16x4_t*)&bflip)[1]);
re32 = vshlq_s32(vcombine_s32(vpadd_s32(((int32x2_t*)&ab_re0)[0],((int32x2_t*)&ab_re0)[1]),
vpadd_s32(((int32x2_t*)&ab_re1)[0],((int32x2_t*)&ab_re1)[1])),
shift);
im32 = vshlq_s32(vcombine_s32(vpadd_s32(((int32x2_t*)&ab_im0)[0],((int32x2_t*)&ab_im0)[1]),
vpadd_s32(((int32x2_t*)&ab_im1)[0],((int32x2_t*)&ab_im1)[1])),
shift);
xtmp = vzipq_s32(re32,im32);
y_128[0] = vcombine_s16(vmovn_s32(xtmp.val[0]),vmovn_s32(xtmp.val[1]));
#endif
xd+=4;
y_128+=1;
}
_mm_empty();
_m_empty();
return(0);
}
inline int32x4_t vcombine(const int32x2_t & v0, const int32x2_t & v1) { return vcombine_s32(v0, v1); }
void silk_biquad_alt_stride2_neon(
const opus_int16 *in, /* I input signal */
const opus_int32 *B_Q28, /* I MA coefficients [3] */
const opus_int32 *A_Q28, /* I AR coefficients [2] */
opus_int32 *S, /* I/O State vector [4] */
opus_int16 *out, /* O output signal */
const opus_int32 len /* I signal length (must be even) */
)
{
/* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */
opus_int k = 0;
const int32x2_t offset_s32x2 = vdup_n_s32( (1<<14) - 1 );
const int32x4_t offset_s32x4 = vcombine_s32( offset_s32x2, offset_s32x2 );
int16x4_t in_s16x4 = vdup_n_s16( 0 );
int16x4_t out_s16x4;
int32x2_t A_Q28_s32x2, A_L_s32x2, A_U_s32x2, B_Q28_s32x2, t_s32x2;
int32x4_t A_L_s32x4, A_U_s32x4, B_Q28_s32x4, S_s32x4, out32_Q14_s32x4;
int32x2x2_t t0_s32x2x2, t1_s32x2x2, t2_s32x2x2, S_s32x2x2;
#ifdef OPUS_CHECK_ASM
opus_int32 S_c[ 4 ];
VARDECL( opus_int16, out_c );
SAVE_STACK;
ALLOC( out_c, 2 * len, opus_int16 );
silk_memcpy( &S_c, S, sizeof( S_c ) );
silk_biquad_alt_stride2_c( in, B_Q28, A_Q28, S_c, out_c, len );
#endif
/* Negate A_Q28 values and split in two parts */
A_Q28_s32x2 = vld1_s32( A_Q28 );
A_Q28_s32x2 = vneg_s32( A_Q28_s32x2 );
A_L_s32x2 = vshl_n_s32( A_Q28_s32x2, 18 ); /* ( -A_Q28[] & 0x00003FFF ) << 18 */
A_L_s32x2 = vreinterpret_s32_u32( vshr_n_u32( vreinterpret_u32_s32( A_L_s32x2 ), 3 ) ); /* ( -A_Q28[] & 0x00003FFF ) << 15 */
A_U_s32x2 = vshr_n_s32( A_Q28_s32x2, 14 ); /* silk_RSHIFT( -A_Q28[], 14 ) */
A_U_s32x2 = vshl_n_s32( A_U_s32x2, 16 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 16 (Clip two leading bits to conform to C function.) */
A_U_s32x2 = vshr_n_s32( A_U_s32x2, 1 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 15 */
B_Q28_s32x2 = vld1_s32( B_Q28 );
t_s32x2 = vld1_s32( B_Q28 + 1 );
t0_s32x2x2 = vzip_s32( A_L_s32x2, A_L_s32x2 );
t1_s32x2x2 = vzip_s32( A_U_s32x2, A_U_s32x2 );
t2_s32x2x2 = vzip_s32( t_s32x2, t_s32x2 );
A_L_s32x4 = vcombine_s32( t0_s32x2x2.val[ 0 ], t0_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_L_Q28 */
A_U_s32x4 = vcombine_s32( t1_s32x2x2.val[ 0 ], t1_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_U_Q28 */
B_Q28_s32x4 = vcombine_s32( t2_s32x2x2.val[ 0 ], t2_s32x2x2.val[ 1 ] ); /* B_Q28[ {1,1,2,2} ] */
S_s32x4 = vld1q_s32( S ); /* S0 = S[ 0 ]; S3 = S[ 3 ]; */
S_s32x2x2 = vtrn_s32( vget_low_s32( S_s32x4 ), vget_high_s32( S_s32x4 ) ); /* S2 = S[ 1 ]; S1 = S[ 2 ]; */
S_s32x4 = vcombine_s32( S_s32x2x2.val[ 0 ], S_s32x2x2.val[ 1 ] );
for( ; k < len - 1; k += 2 ) {
int32x4_t in_s32x4[ 2 ], t_s32x4;
int32x2_t out32_Q14_s32x2[ 2 ];
/* S[ 2 * i + 0 ], S[ 2 * i + 1 ], S[ 2 * i + 2 ], S[ 2 * i + 3 ]: Q12 */
in_s16x4 = vld1_s16( &in[ 2 * k ] ); /* in{0,1,2,3} = in[ 2 * k + {0,1,2,3} ]; */
in_s32x4[ 0 ] = vshll_n_s16( in_s16x4, 15 ); /* in{0,1,2,3} << 15 */
t_s32x4 = vqdmulhq_lane_s32( in_s32x4[ 0 ], B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1,2,3} ) */
in_s32x4[ 1 ] = vcombine_s32( vget_high_s32( in_s32x4[ 0 ] ), vget_high_s32( in_s32x4[ 0 ] ) ); /* in{2,3,2,3} << 15 */
in_s32x4[ 0 ] = vcombine_s32( vget_low_s32 ( in_s32x4[ 0 ] ), vget_low_s32 ( in_s32x4[ 0 ] ) ); /* in{0,1,0,1} << 15 */
silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_low_s32 ( t_s32x4 ), in_s32x4[ 0 ], &S_s32x4, &out32_Q14_s32x2[ 0 ] );
silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_high_s32( t_s32x4 ), in_s32x4[ 1 ], &S_s32x4, &out32_Q14_s32x2[ 1 ] );
/* Scale back to Q0 and saturate */
out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2[ 0 ], out32_Q14_s32x2[ 1 ] ); /* out32_Q14_{0,1,2,3} */
out32_Q14_s32x4 = vaddq_s32( out32_Q14_s32x4, offset_s32x4 ); /* out32_Q14_{0,1,2,3} + (1<<14) - 1 */
out_s16x4 = vqshrn_n_s32( out32_Q14_s32x4, 14 ); /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) ) */
vst1_s16( &out[ 2 * k ], out_s16x4 ); /* out[ 2 * k + {0,1,2,3} ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) ); */
}
/* Process leftover. */
if( k < len ) {
int32x4_t in_s32x4;
int32x2_t out32_Q14_s32x2;
/* S[ 2 * i + 0 ], S[ 2 * i + 1 ]: Q12 */
in_s16x4 = vld1_lane_s16( &in[ 2 * k + 0 ], in_s16x4, 0 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */
in_s16x4 = vld1_lane_s16( &in[ 2 * k + 1 ], in_s16x4, 1 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */
in_s32x4 = vshll_n_s16( in_s16x4, 15 ); /* in{0,1} << 15 */
t_s32x2 = vqdmulh_lane_s32( vget_low_s32( in_s32x4 ), B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1} ) */
in_s32x4 = vcombine_s32( vget_low_s32( in_s32x4 ), vget_low_s32( in_s32x4 ) ); /* in{0,1,0,1} << 15 */
silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, t_s32x2, in_s32x4, &S_s32x4, &out32_Q14_s32x2 );
/* Scale back to Q0 and saturate */
out32_Q14_s32x2 = vadd_s32( out32_Q14_s32x2, offset_s32x2 ); /* out32_Q14_{0,1} + (1<<14) - 1 */
out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2, out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} + (1<<14) - 1 */
out_s16x4 = vqshrn_n_s32( out32_Q14_s32x4, 14 ); /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,0,1} + (1<<14) - 1, 14 ) ) */
vst1_lane_s16( &out[ 2 * k + 0 ], out_s16x4, 0 ); /* out[ 2 * k + 0 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_0 + (1<<14) - 1, 14 ) ); */
vst1_lane_s16( &out[ 2 * k + 1 ], out_s16x4, 1 ); /* out[ 2 * k + 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_1 + (1<<14) - 1, 14 ) ); */
}
vst1q_lane_s32( &S[ 0 ], S_s32x4, 0 ); /* S[ 0 ] = S0; */
vst1q_lane_s32( &S[ 1 ], S_s32x4, 2 ); /* S[ 1 ] = S2; */
vst1q_lane_s32( &S[ 2 ], S_s32x4, 1 ); /* S[ 2 ] = S1; */
vst1q_lane_s32( &S[ 3 ], S_s32x4, 3 ); /* S[ 3 ] = S3; */
#ifdef OPUS_CHECK_ASM
silk_assert( !memcmp( S_c, S, sizeof( S_c ) ) );
silk_assert( !memcmp( out_c, out, 2 * len * sizeof( opus_int16 ) ) );
RESTORE_STACK;
#endif
}
void qt_blend_argb32_on_argb32_neon(uchar *destPixels, int dbpl,
const uchar *srcPixels, int sbpl,
int w, int h,
int const_alpha)
{
const uint *src = (const uint *) srcPixels;
uint *dst = (uint *) destPixels;
int16x8_t half = vdupq_n_s16(0x80);
int16x8_t full = vdupq_n_s16(0xff);
if (const_alpha == 256) {
for (int y = 0; y < h; ++y) {
int x = 0;
for (; x < w-3; x += 4) {
int32x4_t src32 = vld1q_s32((int32_t *)&src[x]);
if ((src[x] & src[x+1] & src[x+2] & src[x+3]) >= 0xff000000) {
// all opaque
vst1q_s32((int32_t *)&dst[x], src32);
} else if (src[x] | src[x+1] | src[x+2] | src[x+3]) {
int32x4_t dst32 = vld1q_s32((int32_t *)&dst[x]);
const uint8x16_t src8 = vreinterpretq_u8_s32(src32);
const uint8x16_t dst8 = vreinterpretq_u8_s32(dst32);
const uint8x8_t src8_low = vget_low_u8(src8);
const uint8x8_t dst8_low = vget_low_u8(dst8);
const uint8x8_t src8_high = vget_high_u8(src8);
const uint8x8_t dst8_high = vget_high_u8(dst8);
const int16x8_t src16_low = vreinterpretq_s16_u16(vmovl_u8(src8_low));
const int16x8_t dst16_low = vreinterpretq_s16_u16(vmovl_u8(dst8_low));
const int16x8_t src16_high = vreinterpretq_s16_u16(vmovl_u8(src8_high));
const int16x8_t dst16_high = vreinterpretq_s16_u16(vmovl_u8(dst8_high));
const int16x8_t result16_low = qvsource_over_s16(src16_low, dst16_low, half, full);
const int16x8_t result16_high = qvsource_over_s16(src16_high, dst16_high, half, full);
const int32x2_t result32_low = vreinterpret_s32_s8(vmovn_s16(result16_low));
const int32x2_t result32_high = vreinterpret_s32_s8(vmovn_s16(result16_high));
vst1q_s32((int32_t *)&dst[x], vcombine_s32(result32_low, result32_high));
}
}
for (; x<w; ++x) {
uint s = src[x];
if (s >= 0xff000000)
dst[x] = s;
else if (s != 0)
dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
}
dst = (quint32 *)(((uchar *) dst) + dbpl);
src = (const quint32 *)(((const uchar *) src) + sbpl);
}
} else if (const_alpha != 0) {
const_alpha = (const_alpha * 255) >> 8;
int16x8_t const_alpha16 = vdupq_n_s16(const_alpha);
for (int y = 0; y < h; ++y) {
int x = 0;
for (; x < w-3; x += 4) {
if (src[x] | src[x+1] | src[x+2] | src[x+3]) {
int32x4_t src32 = vld1q_s32((int32_t *)&src[x]);
int32x4_t dst32 = vld1q_s32((int32_t *)&dst[x]);
const uint8x16_t src8 = vreinterpretq_u8_s32(src32);
const uint8x16_t dst8 = vreinterpretq_u8_s32(dst32);
const uint8x8_t src8_low = vget_low_u8(src8);
const uint8x8_t dst8_low = vget_low_u8(dst8);
const uint8x8_t src8_high = vget_high_u8(src8);
const uint8x8_t dst8_high = vget_high_u8(dst8);
const int16x8_t src16_low = vreinterpretq_s16_u16(vmovl_u8(src8_low));
const int16x8_t dst16_low = vreinterpretq_s16_u16(vmovl_u8(dst8_low));
const int16x8_t src16_high = vreinterpretq_s16_u16(vmovl_u8(src8_high));
const int16x8_t dst16_high = vreinterpretq_s16_u16(vmovl_u8(dst8_high));
const int16x8_t srcalpha16_low = qvbyte_mul_s16(src16_low, const_alpha16, half);
const int16x8_t srcalpha16_high = qvbyte_mul_s16(src16_high, const_alpha16, half);
const int16x8_t result16_low = qvsource_over_s16(srcalpha16_low, dst16_low, half, full);
const int16x8_t result16_high = qvsource_over_s16(srcalpha16_high, dst16_high, half, full);
const int32x2_t result32_low = vreinterpret_s32_s8(vmovn_s16(result16_low));
const int32x2_t result32_high = vreinterpret_s32_s8(vmovn_s16(result16_high));
vst1q_s32((int32_t *)&dst[x], vcombine_s32(result32_low, result32_high));
}
}
for (; x<w; ++x) {
uint s = src[x];
if (s != 0) {
s = BYTE_MUL(s, const_alpha);
dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
}
}
dst = (quint32 *)(((uchar *) dst) + dbpl);
src = (const quint32 *)(((const uchar *) src) + sbpl);
}
}
static inline int32_t TransformAndFindMaxNeon(int16_t* inre,
int16_t* inim,
int32_t* outre,
int32_t* outim) {
int k;
int16_t* inre1 = inre;
int16_t* inre2 = &inre[FRAMESAMPLES/2 - 4];
int16_t* inim1 = inim;
int16_t* inim2 = &inim[FRAMESAMPLES/2 - 4];
int32_t* outre1 = outre;
int32_t* outre2 = &outre[FRAMESAMPLES/2 - 4];
int32_t* outim1 = outim;
int32_t* outim2 = &outim[FRAMESAMPLES/2 - 4];
const int16_t* kSinTab1 = &WebRtcIsacfix_kSinTab2[0];
const int16_t* kSinTab2 = &WebRtcIsacfix_kSinTab2[FRAMESAMPLES/4 - 4];
uint32x4_t max_r = vdupq_n_u32(0);
uint32x4_t max_i = vdupq_n_u32(0);
// Use ">> 5", instead of "<< 9" and then ">> 14" as in the C code.
for (k = 0; k < FRAMESAMPLES/4; k += 4) {
int16x4_t tmpi = vld1_s16(kSinTab1);
kSinTab1 += 4;
int16x4_t tmpr = vld1_s16(kSinTab2);
kSinTab2 -= 4;
int16x4_t inre_0 = vld1_s16(inre1);
inre1 += 4;
int16x4_t inre_1 = vld1_s16(inre2);
inre2 -= 4;
int16x4_t inim_0 = vld1_s16(inim1);
inim1 += 4;
int16x4_t inim_1 = vld1_s16(inim2);
inim2 -= 4;
tmpr = vneg_s16(tmpr);
inre_1 = vrev64_s16(inre_1);
inim_1 = vrev64_s16(inim_1);
tmpr = vrev64_s16(tmpr);
int32x4_t xr = vmull_s16(tmpr, inre_0);
int32x4_t xi = vmull_s16(tmpr, inim_0);
int32x4_t yr = vmull_s16(tmpr, inim_1);
int32x4_t yi = vmull_s16(tmpi, inim_1);
xr = vmlal_s16(xr, tmpi, inim_0);
xi = vmlsl_s16(xi, tmpi, inre_0);
yr = vmlal_s16(yr, tmpi, inre_1);
yi = vmlsl_s16(yi, tmpr, inre_1);
yr = vnegq_s32(yr);
xr = vshrq_n_s32(xr, 5);
xi = vshrq_n_s32(xi, 5);
yr = vshrq_n_s32(yr, 5);
yi = vshrq_n_s32(yi, 5);
int32x4_t outr0 = vsubq_s32(xr, yi);
int32x4_t outr1 = vaddq_s32(xr, yi);
int32x4_t outi0 = vaddq_s32(xi, yr);
int32x4_t outi1 = vsubq_s32(yr, xi);
// Find the absolute maximum in the vectors.
int32x4_t tmp0 = vabsq_s32(outr0);
int32x4_t tmp1 = vabsq_s32(outr1);
int32x4_t tmp2 = vabsq_s32(outi0);
int32x4_t tmp3 = vabsq_s32(outi1);
// vabs doesn't change the value of 0x80000000.
// Use u32 so we don't lose the value 0x80000000.
max_r = vmaxq_u32(max_r, vreinterpretq_u32_s32(tmp0));
max_i = vmaxq_u32(max_i, vreinterpretq_u32_s32(tmp2));
max_r = vmaxq_u32(max_r, vreinterpretq_u32_s32(tmp1));
max_i = vmaxq_u32(max_i, vreinterpretq_u32_s32(tmp3));
// Store the vectors.
outr1 = vrev64q_s32(outr1);
outi1 = vrev64q_s32(outi1);
int32x4_t outr_1 = vcombine_s32(vget_high_s32(outr1), vget_low_s32(outr1));
int32x4_t outi_1 = vcombine_s32(vget_high_s32(outi1), vget_low_s32(outi1));
vst1q_s32(outre1, outr0);
outre1 += 4;
vst1q_s32(outim1, outi0);
outim1 += 4;
vst1q_s32(outre2, outr_1);
outre2 -= 4;
vst1q_s32(outim2, outi_1);
outim2 -= 4;
}
max_r = vmaxq_u32(max_r, max_i);
#if defined(WEBRTC_ARCH_ARM64)
uint32_t maximum = vmaxvq_u32(max_r);
#else
uint32x2_t max32x2_r = vmax_u32(vget_low_u32(max_r), vget_high_u32(max_r));
max32x2_r = vpmax_u32(max32x2_r, max32x2_r);
uint32_t maximum = vget_lane_u32(max32x2_r, 0);
#endif
return (int32_t)maximum;
}
// CHECK-LABEL: define <4 x i32> @test_vcombine_s32(<2 x i32> %low, <2 x i32> %high) #0 {
// CHECK: [[SHUFFLE_I:%.*]] = shufflevector <2 x i32> %low, <2 x i32> %high, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
// CHECK: ret <4 x i32> [[SHUFFLE_I]]
int32x4_t test_vcombine_s32(int32x2_t low, int32x2_t high) {
return vcombine_s32(low, high);
}
static inline void PostShiftAndDivideAndDemodulateNeon(int16_t* inre,
int16_t* inim,
int32_t* outre1,
int32_t* outre2,
int32_t sh) {
int k;
int16_t* p_inre = inre;
int16_t* p_inim = inim;
int32_t* p_outre1 = outre1;
int32_t* p_outre2 = outre2;
const int16_t* kCosTab = &WebRtcIsacfix_kCosTab1[0];
const int16_t* kSinTab = &WebRtcIsacfix_kSinTab1[0];
int32x4_t shift = vdupq_n_s32(-sh - 16);
// Divide through by the normalizing constant:
// scale all values with 1/240, i.e. with 273 in Q16.
// 273/65536 ~= 0.0041656
// 1/240 ~= 0.0041666
int16x8_t scale = vdupq_n_s16(273);
// Sqrt(240) in Q11 is round(15.49193338482967 * 2048) = 31727.
int factQ19 = 31727 << 16;
int32x4_t fact = vdupq_n_s32(factQ19);
for (k = 0; k < FRAMESAMPLES/2; k += 8) {
int16x8_t inre16x8 = vld1q_s16(p_inre);
int16x8_t inim16x8 = vld1q_s16(p_inim);
p_inre += 8;
p_inim += 8;
int16x8_t tmpr = vld1q_s16(kCosTab);
int16x8_t tmpi = vld1q_s16(kSinTab);
kCosTab += 8;
kSinTab += 8;
// By vshl and vmull, we effectively did "<< (-sh - 16)",
// instead of "<< (-sh)" and ">> 16" as in the C code.
int32x4_t outre1_0 = vmull_s16(vget_low_s16(inre16x8), vget_low_s16(scale));
int32x4_t outre2_0 = vmull_s16(vget_low_s16(inim16x8), vget_low_s16(scale));
#if defined(WEBRTC_ARCH_ARM64)
int32x4_t outre1_1 = vmull_high_s16(inre16x8, scale);
int32x4_t outre2_1 = vmull_high_s16(inim16x8, scale);
#else
int32x4_t outre1_1 = vmull_s16(vget_high_s16(inre16x8),
vget_high_s16(scale));
int32x4_t outre2_1 = vmull_s16(vget_high_s16(inim16x8),
vget_high_s16(scale));
#endif
outre1_0 = vshlq_s32(outre1_0, shift);
outre1_1 = vshlq_s32(outre1_1, shift);
outre2_0 = vshlq_s32(outre2_0, shift);
outre2_1 = vshlq_s32(outre2_1, shift);
// Demodulate and separate.
int32x4_t tmpr_0 = vmovl_s16(vget_low_s16(tmpr));
int32x4_t tmpi_0 = vmovl_s16(vget_low_s16(tmpi));
#if defined(WEBRTC_ARCH_ARM64)
int32x4_t tmpr_1 = vmovl_high_s16(tmpr);
int32x4_t tmpi_1 = vmovl_high_s16(tmpi);
#else
int32x4_t tmpr_1 = vmovl_s16(vget_high_s16(tmpr));
int32x4_t tmpi_1 = vmovl_s16(vget_high_s16(tmpi));
#endif
int64x2_t xr0 = vmull_s32(vget_low_s32(tmpr_0), vget_low_s32(outre1_0));
int64x2_t xi0 = vmull_s32(vget_low_s32(tmpr_0), vget_low_s32(outre2_0));
int64x2_t xr2 = vmull_s32(vget_low_s32(tmpr_1), vget_low_s32(outre1_1));
int64x2_t xi2 = vmull_s32(vget_low_s32(tmpr_1), vget_low_s32(outre2_1));
xr0 = vmlsl_s32(xr0, vget_low_s32(tmpi_0), vget_low_s32(outre2_0));
xi0 = vmlal_s32(xi0, vget_low_s32(tmpi_0), vget_low_s32(outre1_0));
xr2 = vmlsl_s32(xr2, vget_low_s32(tmpi_1), vget_low_s32(outre2_1));
xi2 = vmlal_s32(xi2, vget_low_s32(tmpi_1), vget_low_s32(outre1_1));
#if defined(WEBRTC_ARCH_ARM64)
int64x2_t xr1 = vmull_high_s32(tmpr_0, outre1_0);
int64x2_t xi1 = vmull_high_s32(tmpr_0, outre2_0);
int64x2_t xr3 = vmull_high_s32(tmpr_1, outre1_1);
int64x2_t xi3 = vmull_high_s32(tmpr_1, outre2_1);
xr1 = vmlsl_high_s32(xr1, tmpi_0, outre2_0);
xi1 = vmlal_high_s32(xi1, tmpi_0, outre1_0);
xr3 = vmlsl_high_s32(xr3, tmpi_1, outre2_1);
xi3 = vmlal_high_s32(xi3, tmpi_1, outre1_1);
#else
int64x2_t xr1 = vmull_s32(vget_high_s32(tmpr_0), vget_high_s32(outre1_0));
int64x2_t xi1 = vmull_s32(vget_high_s32(tmpr_0), vget_high_s32(outre2_0));
int64x2_t xr3 = vmull_s32(vget_high_s32(tmpr_1), vget_high_s32(outre1_1));
int64x2_t xi3 = vmull_s32(vget_high_s32(tmpr_1), vget_high_s32(outre2_1));
xr1 = vmlsl_s32(xr1, vget_high_s32(tmpi_0), vget_high_s32(outre2_0));
xi1 = vmlal_s32(xi1, vget_high_s32(tmpi_0), vget_high_s32(outre1_0));
xr3 = vmlsl_s32(xr3, vget_high_s32(tmpi_1), vget_high_s32(outre2_1));
xi3 = vmlal_s32(xi3, vget_high_s32(tmpi_1), vget_high_s32(outre1_1));
#endif
outre1_0 = vcombine_s32(vshrn_n_s64(xr0, 10), vshrn_n_s64(xr1, 10));
outre2_0 = vcombine_s32(vshrn_n_s64(xi0, 10), vshrn_n_s64(xi1, 10));
outre1_1 = vcombine_s32(vshrn_n_s64(xr2, 10), vshrn_n_s64(xr3, 10));
outre2_1 = vcombine_s32(vshrn_n_s64(xi2, 10), vshrn_n_s64(xi3, 10));
outre1_0 = vqdmulhq_s32(outre1_0, fact);
outre2_0 = vqdmulhq_s32(outre2_0, fact);
outre1_1 = vqdmulhq_s32(outre1_1, fact);
outre2_1 = vqdmulhq_s32(outre2_1, fact);
vst1q_s32(p_outre1, outre1_0);
p_outre1 += 4;
vst1q_s32(p_outre1, outre1_1);
p_outre1 += 4;
vst1q_s32(p_outre2, outre2_0);
p_outre2 += 4;
vst1q_s32(p_outre2, outre2_1);
p_outre2 += 4;
}
}
void silk_warped_autocorrelation_FIX_neon(
opus_int32 *corr, /* O Result [order + 1] */
opus_int *scale, /* O Scaling of the correlation vector */
const opus_int16 *input, /* I Input data to correlate */
const opus_int warping_Q16, /* I Warping coefficient */
const opus_int length, /* I Length of input */
const opus_int order /* I Correlation order (even) */
)
{
if( ( MAX_SHAPE_LPC_ORDER > 24 ) || ( order < 6 ) ) {
silk_warped_autocorrelation_FIX_c( corr, scale, input, warping_Q16, length, order );
} else {
opus_int n, i, lsh;
opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; /* In reverse order */
opus_int64 corr_QC_orderT;
int64x2_t lsh_s64x2;
const opus_int orderT = ( order + 3 ) & ~3;
opus_int64 *corr_QCT;
opus_int32 *input_QS;
VARDECL( opus_int32, input_QST );
VARDECL( opus_int32, state );
SAVE_STACK;
/* Order must be even */
silk_assert( ( order & 1 ) == 0 );
silk_assert( 2 * QS - QC >= 0 );
ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER, opus_int32 );
input_QS = input_QST;
/* input_QS has zero paddings in the beginning and end. */
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
/* Loop over samples */
for( n = 0; n < length - 7; n += 8, input_QS += 8 ) {
const int16x8_t t0_s16x4 = vld1q_s16( input + n );
vst1q_s32( input_QS + 0, vshll_n_s16( vget_low_s16( t0_s16x4 ), QS ) );
vst1q_s32( input_QS + 4, vshll_n_s16( vget_high_s16( t0_s16x4 ), QS ) );
}
for( ; n < length; n++, input_QS++ ) {
input_QS[ 0 ] = silk_LSHIFT32( (opus_int32)input[ n ], QS );
}
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS = input_QST + MAX_SHAPE_LPC_ORDER - orderT;
/* The following loop runs ( length + order ) times, with ( order ) extra epilogues. */
/* The zero paddings in input_QS guarantee corr_QC's correctness even with the extra epilogues. */
/* The values of state_QS will be polluted by the extra epilogues, however they are temporary values. */
/* Keep the C code here to help understand the intrinsics optimization. */
/*
{
opus_int32 state_QS[ 2 ][ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };
opus_int32 *state_QST[ 3 ];
state_QST[ 0 ] = state_QS[ 0 ];
state_QST[ 1 ] = state_QS[ 1 ];
for( n = 0; n < length + order; n++, input_QS++ ) {
state_QST[ 0 ][ orderT ] = input_QS[ orderT ];
for( i = 0; i < orderT; i++ ) {
corr_QC[ i ] += silk_RSHIFT64( silk_SMULL( state_QST[ 0 ][ i ], input_QS[ i ] ), 2 * QS - QC );
state_QST[ 1 ][ i ] = silk_SMLAWB( state_QST[ 1 ][ i + 1 ], state_QST[ 0 ][ i ] - state_QST[ 0 ][ i + 1 ], warping_Q16 );
}
state_QST[ 2 ] = state_QST[ 0 ];
state_QST[ 0 ] = state_QST[ 1 ];
state_QST[ 1 ] = state_QST[ 2 ];
}
}
*/
{
const int32x4_t warping_Q16_s32x4 = vdupq_n_s32( warping_Q16 << 15 );
const opus_int32 *in = input_QS + orderT;
opus_int o = orderT;
int32x4_t state_QS_s32x4[ 3 ][ 2 ];
ALLOC( state, length + orderT, opus_int32 );
state_QS_s32x4[ 2 ][ 1 ] = vdupq_n_s32( 0 );
/* Calculate 8 taps of all inputs in each loop. */
do {
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 0 ][ 1 ] =
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 1 ][ 1 ] = vdupq_n_s32( 0 );
n = 0;
do {
calc_corr( input_QS + n, corr_QC, o - 8, state_QS_s32x4[ 0 ][ 0 ] );
calc_corr( input_QS + n, corr_QC, o - 4, state_QS_s32x4[ 0 ][ 1 ] );
state_QS_s32x4[ 2 ][ 1 ] = vld1q_s32( in + n );
vst1q_lane_s32( state + n, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 0 ][ 1 ], 1 );
state_QS_s32x4[ 2 ][ 1 ] = vextq_s32( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 0 ][ 1 ] = calc_state( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], state_QS_s32x4[ 1 ][ 1 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
state_QS_s32x4[ 1 ][ 1 ] = state_QS_s32x4[ 2 ][ 1 ];
} while( ++n < ( length + order ) );
in = state;
o -= 8;
} while( o > 4 );
if( o ) {
/* Calculate the last 4 taps of all inputs. */
opus_int32 *stateT = state;
silk_assert( o == 4 );
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 1 ][ 0 ] = vdupq_n_s32( 0 );
n = length + order;
do {
calc_corr( input_QS, corr_QC, 0, state_QS_s32x4[ 0 ][ 0 ] );
state_QS_s32x4[ 2 ][ 0 ] = vld1q_s32( stateT );
vst1q_lane_s32( stateT, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
input_QS++;
stateT++;
} while( --n );
}
}
{
const opus_int16 *inputT = input;
int32x4_t t_s32x4;
int64x1_t t_s64x1;
int64x2_t t_s64x2 = vdupq_n_s64( 0 );
for( n = 0; n <= length - 8; n += 8 ) {
int16x8_t input_s16x8 = vld1q_s16( inputT );
t_s32x4 = vmull_s16( vget_low_s16( input_s16x8 ), vget_low_s16( input_s16x8 ) );
t_s32x4 = vmlal_s16( t_s32x4, vget_high_s16( input_s16x8 ), vget_high_s16( input_s16x8 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_low_s32( t_s32x4 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_high_s32( t_s32x4 ) );
inputT += 8;
}
t_s64x1 = vadd_s64( vget_low_s64( t_s64x2 ), vget_high_s64( t_s64x2 ) );
corr_QC_orderT = vget_lane_s64( t_s64x1, 0 );
for( ; n < length; n++ ) {
corr_QC_orderT += silk_SMULL( input[ n ], input[ n ] );
}
corr_QC_orderT = silk_LSHIFT64( corr_QC_orderT, QC );
corr_QC[ orderT ] = corr_QC_orderT;
}
corr_QCT = corr_QC + orderT - order;
lsh = silk_CLZ64( corr_QC_orderT ) - 35;
lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC );
*scale = -( QC + lsh );
silk_assert( *scale >= -30 && *scale <= 12 );
lsh_s64x2 = vdupq_n_s64( lsh );
for( i = 0; i <= order - 3; i += 4 ) {
int32x4_t corr_s32x4;
int64x2_t corr_QC0_s64x2, corr_QC1_s64x2;
corr_QC0_s64x2 = vld1q_s64( corr_QCT + i );
corr_QC1_s64x2 = vld1q_s64( corr_QCT + i + 2 );
corr_QC0_s64x2 = vshlq_s64( corr_QC0_s64x2, lsh_s64x2 );
corr_QC1_s64x2 = vshlq_s64( corr_QC1_s64x2, lsh_s64x2 );
corr_s32x4 = vcombine_s32( vmovn_s64( corr_QC1_s64x2 ), vmovn_s64( corr_QC0_s64x2 ) );
corr_s32x4 = vrev64q_s32( corr_s32x4 );
vst1q_s32( corr + order - i - 3, corr_s32x4 );
}
if( lsh >= 0 ) {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QCT[ i ], lsh ) );
}
} else {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QCT[ i ], -lsh ) );
}
}
silk_assert( corr_QCT[ order ] >= 0 ); /* If breaking, decrease QC*/
RESTORE_STACK;
}
#ifdef OPUS_CHECK_ASM
{
opus_int32 corr_c[ MAX_SHAPE_LPC_ORDER + 1 ];
opus_int scale_c;
silk_warped_autocorrelation_FIX_c( corr_c, &scale_c, input, warping_Q16, length, order );
silk_assert( !memcmp( corr_c, corr, sizeof( corr_c[ 0 ] ) * ( order + 1 ) ) );
silk_assert( scale_c == *scale );
}
#endif
}
void computeNetwork0new_neon(const float *dataf, const float *weightsf, uint8_t *d) {
const int16_t *data = (const int16_t *)dataf;
const int16_t *weights = (const int16_t *)weightsf;
int32x4_t accum0 = { 0, 0, 0, 0 };
int32x4_t accum1 = accum0;
int32x4_t accum2 = accum0;
int32x4_t accum3 = accum0;
for (int i = 0; i < 128/2; i += 8) {
int16x4x2_t d0 = vld2_s16(data + i);
int16x4x2_t w0 = vld2_s16(weights + i * 4);
int16x4x2_t w1 = vld2_s16(weights + i * 4 + 8);
int16x4x2_t w2 = vld2_s16(weights + i * 4 + 16);
int16x4x2_t w3 = vld2_s16(weights + i * 4 + 24);
accum0 = vmlal_s16(accum0, d0.val[0], w0.val[0]);
accum0 = vmlal_s16(accum0, d0.val[1], w0.val[1]);
accum1 = vmlal_s16(accum1, d0.val[0], w1.val[0]);
accum1 = vmlal_s16(accum1, d0.val[1], w1.val[1]);
accum2 = vmlal_s16(accum2, d0.val[0], w2.val[0]);
accum2 = vmlal_s16(accum2, d0.val[1], w2.val[1]);
accum3 = vmlal_s16(accum3, d0.val[0], w3.val[0]);
accum3 = vmlal_s16(accum3, d0.val[1], w3.val[1]);
}
int32x2_t sum0 = vpadd_s32(vget_low_s32(accum0), vget_high_s32(accum0));
int32x2_t sum1 = vpadd_s32(vget_low_s32(accum1), vget_high_s32(accum1));
int32x2_t sum2 = vpadd_s32(vget_low_s32(accum2), vget_high_s32(accum2));
int32x2_t sum3 = vpadd_s32(vget_low_s32(accum3), vget_high_s32(accum3));
sum0 = vpadd_s32(sum0, sum1);
sum1 = vpadd_s32(sum2, sum3);
int32x4_t sum = vcombine_s32(sum0, sum1);
float32x4_t m0 = vcvtq_f32_s32(sum);
m0 = vmulq_f32(m0, vld1q_f32(weightsf + 512/4));
m0 = vaddq_f32(m0, vld1q_f32(weightsf + 528/4));
float32x4_t m1, m2, m3, m4;
m1 = m0;
m0 = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(m0), sign_bits_f));
m0 = vaddq_f32(m0, ones_f);
m0 = vmulq_f32(reciprocal(m0), m1);
m1 = vdupq_lane_f32(vget_low_f32(m0), 0);
m2 = vdupq_lane_f32(vget_low_f32(m0), 1);
m3 = vdupq_lane_f32(vget_high_f32(m0), 0);
m4 = vdupq_lane_f32(vget_high_f32(m0), 1);
m1 = vmulq_f32(m1, vld1q_f32(weightsf + 544/4));
m2 = vmulq_f32(m2, vld1q_f32(weightsf + 560/4));
m3 = vmulq_f32(m3, vld1q_f32(weightsf + 576/4));
m4 = vmulq_f32(m4, vld1q_f32(weightsf + 592/4));
m1 = vaddq_f32(m1, m2);
m3 = vaddq_f32(m3, m4);
m1 = vaddq_f32(m1, m3);
m1 = vaddq_f32(m1, vld1q_f32(weightsf + 608/4));
uint32x4_t gte = vcgeq_f32(m1, zeroes_f);
uint16x4_t gte_u16 = vmovn_u32(gte);
uint8x8_t gte_u8 = vmovn_u16(vcombine_u16(gte_u16, vget_low_u16(vreinterpretq_u16_u32(sign_bits_f))));
gte_u8 = vshr_n_u8(gte_u8, 7);
vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(gte_u8), 0);
}
void computeNetwork0_i16_neon(const float *inputf, const float *weightsf, uint8_t *d) {
const int16_t *input = (const int16_t *)inputf;
const int16_t *weights = (const int16_t *)weightsf;
int32x4_t accum0 = { 0, 0, 0, 0 };
int32x4_t accum1 = accum0;
int32x4_t accum2 = accum0;
int32x4_t accum3 = accum0;
for (int i = 0; i < 96/2; i += 8) {
int16x4x2_t d0 = vld2_s16(input + i);
int16x4x2_t w0 = vld2_s16(weights + i * 4);
int16x4x2_t w1 = vld2_s16(weights + i * 4 + 8);
int16x4x2_t w2 = vld2_s16(weights + i * 4 + 16);
int16x4x2_t w3 = vld2_s16(weights + i * 4 + 24);
accum0 = vmlal_s16(accum0, d0.val[0], w0.val[0]);
accum0 = vmlal_s16(accum0, d0.val[1], w0.val[1]);
accum1 = vmlal_s16(accum1, d0.val[0], w1.val[0]);
accum1 = vmlal_s16(accum1, d0.val[1], w1.val[1]);
accum2 = vmlal_s16(accum2, d0.val[0], w2.val[0]);
accum2 = vmlal_s16(accum2, d0.val[1], w2.val[1]);
accum3 = vmlal_s16(accum3, d0.val[0], w3.val[0]);
accum3 = vmlal_s16(accum3, d0.val[1], w3.val[1]);
}
int32x2_t sum0 = vpadd_s32(vget_low_s32(accum0), vget_high_s32(accum0));
int32x2_t sum1 = vpadd_s32(vget_low_s32(accum1), vget_high_s32(accum1));
int32x2_t sum2 = vpadd_s32(vget_low_s32(accum2), vget_high_s32(accum2));
int32x2_t sum3 = vpadd_s32(vget_low_s32(accum3), vget_high_s32(accum3));
sum0 = vpadd_s32(sum0, sum1);
sum1 = vpadd_s32(sum2, sum3);
int32x4_t sum = vcombine_s32(sum0, sum1);
float32x4_t m0 = vcvtq_f32_s32(sum);
m0 = vmulq_f32(m0, vld1q_f32(weightsf + 384/4));
m0 = vaddq_f32(m0, vld1q_f32(weightsf + 400/4));
float32x4_t m1, m2, m3, m4, m5, m6, m7;
m1 = m0;
m0 = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(m0), sign_bits_f_zero_l));
m0 = vaddq_f32(m0, ones_f);
m0 = vmulq_f32(reciprocal(m0), m1);
m1 = vdupq_lane_f32(vget_low_f32(m0), 0);
m2 = vdupq_lane_f32(vget_low_f32(m0), 1);
m3 = vdupq_lane_f32(vget_high_f32(m0), 0);
m4 = vdupq_lane_f32(vget_high_f32(m0), 1);
m1 = vmulq_f32(m1, vld1q_f32(weightsf + 416/4));
m2 = vmulq_f32(m2, vld1q_f32(weightsf + (416+16)/4));
m3 = vmulq_f32(m3, vld1q_f32(weightsf + (416+32)/4));
m4 = vmulq_f32(m4, vld1q_f32(weightsf + (416+48)/4));
m1 = vaddq_f32(m1, m2);
m3 = vaddq_f32(m3, m4);
m1 = vaddq_f32(m1, m3);
m1 = vaddq_f32(m1, vld1q_f32(weightsf + (416+64)/4));
m7 = m1;
m1 = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(m1), sign_bits_f));
m1 = vaddq_f32(m1, ones_f);
m7 = vmulq_f32(reciprocal(m1), m7);
m3 = m0;
m0 = vdupq_lane_f32(vget_low_f32(m0), 0);
m1 = vdupq_lane_f32(vget_low_f32(m3), 1);
m2 = vdupq_lane_f32(vget_high_f32(m3), 0);
m3 = vdupq_lane_f32(vget_high_f32(m3), 1);
m0 = vmulq_f32(m0, vld1q_f32(weightsf + 496/4));
m1 = vmulq_f32(m1, vld1q_f32(weightsf + (496+16)/4));
m2 = vmulq_f32(m2, vld1q_f32(weightsf + (496+32)/4));
m3 = vmulq_f32(m3, vld1q_f32(weightsf + (496+48)/4));
m4 = vdupq_lane_f32(vget_low_f32(m7), 0);
m5 = vdupq_lane_f32(vget_low_f32(m7), 1);
m6 = vdupq_lane_f32(vget_high_f32(m7), 0);
m7 = vdupq_lane_f32(vget_high_f32(m7), 1);
m4 = vmulq_f32(m4, vld1q_f32(weightsf + (496+64)/4));
m5 = vmulq_f32(m5, vld1q_f32(weightsf + (496+80)/4));
m6 = vmulq_f32(m6, vld1q_f32(weightsf + (496+96)/4));
m7 = vmulq_f32(m7, vld1q_f32(weightsf + (496+112)/4));
m0 = vaddq_f32(m0, m1);
m2 = vaddq_f32(m2, m3);
m4 = vaddq_f32(m4, m5);
m6 = vaddq_f32(m6, m7);
m0 = vaddq_f32(m0, m2);
m4 = vaddq_f32(m4, m6);
m0 = vaddq_f32(m0, m4);
m0 = vaddq_f32(m0, vld1q_f32(weightsf + (496+128)/4));
float32x2_t maximum = vmax_f32(vget_low_f32(m0), vget_high_f32(m0));
d[0] = (vget_lane_f32(maximum, 1) <= vget_lane_f32(maximum, 0));
}
int mult_cpx_conj_vector(int16_t *x1,
int16_t *x2,
int16_t *y,
uint32_t N,
int output_shift,
int madd)
{
// Multiply elementwise the complex conjugate of x1 with x2.
// x1 - input 1 in the format |Re0 Im0 Re1 Im1|,......,|Re(N-2) Im(N-2) Re(N-1) Im(N-1)|
// We assume x1 with a dinamic of 15 bit maximum
//
// x2 - input 2 in the format |Re0 Im0 Re1 Im1|,......,|Re(N-2) Im(N-2) Re(N-1) Im(N-1)|
// We assume x2 with a dinamic of 14 bit maximum
///
// y - output in the format |Re0 Im0 Re1 Im1|,......,|Re(N-2) Im(N-2) Re(N-1) Im(N-1)|
//
// N - the size f the vectors (this function does N cpx mpy. WARNING: N>=4;
//
// output_shift - shift to be applied to generate output
//
// madd - add the output to y
uint32_t i; // loop counter
simd_q15_t *x1_128;
simd_q15_t *x2_128;
simd_q15_t *y_128;
#if defined(__x86_64__) || defined(__i386__)
simd_q15_t tmp_re,tmp_im;
simd_q15_t tmpy0,tmpy1;
#elif defined(__arm__)
int32x4_t tmp_re,tmp_im;
int32x4_t tmp_re1,tmp_im1;
int16x4x2_t tmpy;
int32x4_t shift = vdupq_n_s32(-output_shift);
#endif
x1_128 = (simd_q15_t *)&x1[0];
x2_128 = (simd_q15_t *)&x2[0];
y_128 = (simd_q15_t *)&y[0];
// we compute 4 cpx multiply for each loop
for(i=0; i<(N>>2); i++) {
#if defined(__x86_64__) || defined(__i386__)
tmp_re = _mm_madd_epi16(*x1_128,*x2_128);
tmp_im = _mm_shufflelo_epi16(*x1_128,_MM_SHUFFLE(2,3,0,1));
tmp_im = _mm_shufflehi_epi16(tmp_im,_MM_SHUFFLE(2,3,0,1));
tmp_im = _mm_sign_epi16(tmp_im,*(__m128i*)&conjug[0]);
tmp_im = _mm_madd_epi16(tmp_im,*x2_128);
tmp_re = _mm_srai_epi32(tmp_re,output_shift);
tmp_im = _mm_srai_epi32(tmp_im,output_shift);
tmpy0 = _mm_unpacklo_epi32(tmp_re,tmp_im);
tmpy1 = _mm_unpackhi_epi32(tmp_re,tmp_im);
if (madd==0)
*y_128 = _mm_packs_epi32(tmpy0,tmpy1);
else
*y_128 += _mm_packs_epi32(tmpy0,tmpy1);
#elif defined(__arm__)
tmp_re = vmull_s16(((simdshort_q15_t *)x1_128)[0], ((simdshort_q15_t*)x2_128)[0]);
//tmp_re = [Re(x1[0])Re(x2[0]) Im(x1[0])Im(x2[0]) Re(x1[1])Re(x2[1]) Im(x1[1])Im(x2[1])]
tmp_re1 = vmull_s16(((simdshort_q15_t *)x1_128)[1], ((simdshort_q15_t*)x2_128)[1]);
//tmp_re1 = [Re(x1[1])Re(x2[1]) Im(x1[1])Im(x2[1]) Re(x1[1])Re(x2[2]) Im(x1[1])Im(x2[2])]
tmp_re = vcombine_s32(vpadd_s32(vget_low_s32(tmp_re),vget_high_s32(tmp_re)),
vpadd_s32(vget_low_s32(tmp_re1),vget_high_s32(tmp_re1)));
//tmp_re = [Re(ch[0])Re(rx[0])+Im(ch[0])Im(ch[0]) Re(ch[1])Re(rx[1])+Im(ch[1])Im(ch[1]) Re(ch[2])Re(rx[2])+Im(ch[2]) Im(ch[2]) Re(ch[3])Re(rx[3])+Im(ch[3])Im(ch[3])]
tmp_im = vmull_s16(vrev32_s16(vmul_s16(((simdshort_q15_t*)x2_128)[0],*(simdshort_q15_t*)conjug)), ((simdshort_q15_t*)x1_128)[0]);
//tmp_im = [-Im(ch[0])Re(rx[0]) Re(ch[0])Im(rx[0]) -Im(ch[1])Re(rx[1]) Re(ch[1])Im(rx[1])]
tmp_im1 = vmull_s16(vrev32_s16(vmul_s16(((simdshort_q15_t*)x2_128)[1],*(simdshort_q15_t*)conjug)), ((simdshort_q15_t*)x1_128)[1]);
//tmp_im1 = [-Im(ch[2])Re(rx[2]) Re(ch[2])Im(rx[2]) -Im(ch[3])Re(rx[3]) Re(ch[3])Im(rx[3])]
tmp_im = vcombine_s32(vpadd_s32(vget_low_s32(tmp_im),vget_high_s32(tmp_im)),
vpadd_s32(vget_low_s32(tmp_im1),vget_high_s32(tmp_im1)));
//tmp_im = [-Im(ch[0])Re(rx[0])+Re(ch[0])Im(rx[0]) -Im(ch[1])Re(rx[1])+Re(ch[1])Im(rx[1]) -Im(ch[2])Re(rx[2])+Re(ch[2])Im(rx[2]) -Im(ch[3])Re(rx[3])+Re(ch[3])Im(rx[3])]
tmp_re = vqshlq_s32(tmp_re,shift);
tmp_im = vqshlq_s32(tmp_im,shift);
tmpy = vzip_s16(vmovn_s32(tmp_re),vmovn_s32(tmp_im));
if (madd==0)
*y_128 = vcombine_s16(tmpy.val[0],tmpy.val[1]);
else
*y_128 += vcombine_s16(tmpy.val[0],tmpy.val[1]);
#endif
x1_128++;
x2_128++;
y_128++;
}
_mm_empty();
_m_empty();
return(0);
}