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 __m128i predict_unclipped(const __m128i *input, __m128i alpha_q12,
__m128i alpha_sign, __m128i dc_q0) {
__m128i ac_q3 = _mm_loadu_si128(input);
__m128i ac_sign = _mm_sign_epi16(alpha_sign, ac_q3);
__m128i scaled_luma_q0 = _mm_mulhrs_epi16(_mm_abs_epi16(ac_q3), alpha_q12);
scaled_luma_q0 = _mm_sign_epi16(scaled_luma_q0, ac_sign);
return _mm_add_epi16(scaled_luma_q0, dc_q0);
}
int main(int, char**)
{
volatile __m128i a = _mm_set1_epi32(42);
_mm_abs_epi8(a);
volatile __m128i result = _mm_sign_epi16(a, _mm_set1_epi32(64));
(void)result;
return 0;
}
/* Test the 128-bit form */
static void
ssse3_test_psignw128 (int *i1, int *i2, int *r)
{
/* Assumes incoming pointers are 16-byte aligned */
__m128i t1 = *(__m128i *) i1;
__m128i t2 = *(__m128i *) i2;
*(__m128i *) r = _mm_sign_epi16 (t1, t2);
}
static INLINE void hor_transform_row_avx2(__m128i* row){
__m128i mask_pos = _mm_set1_epi16(1);
__m128i mask_neg = _mm_set1_epi16(-1);
__m128i sign_mask = _mm_unpacklo_epi64(mask_pos, mask_neg);
__m128i temp = _mm_shuffle_epi32(*row, KVZ_PERMUTE(2, 3, 0, 1));
*row = _mm_sign_epi16(*row, sign_mask);
*row = _mm_add_epi16(*row, temp);
sign_mask = _mm_unpacklo_epi32(mask_pos, mask_neg);
temp = _mm_shuffle_epi32(*row, KVZ_PERMUTE(1, 0, 3, 2));
*row = _mm_sign_epi16(*row, sign_mask);
*row = _mm_add_epi16(*row, temp);
sign_mask = _mm_unpacklo_epi16(mask_pos, mask_neg);
temp = _mm_shufflelo_epi16(*row, KVZ_PERMUTE(1,0,3,2));
temp = _mm_shufflehi_epi16(temp, KVZ_PERMUTE(1,0,3,2));
*row = _mm_sign_epi16(*row, sign_mask);
*row = _mm_add_epi16(*row, temp);
}
static WEBP_INLINE int DoQuantizeBlock(int16_t in[16], int16_t out[16],
const uint16_t* const sharpen,
const VP8Matrix* const mtx) {
const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
const __m128i zero = _mm_setzero_si128();
__m128i out0, out8;
__m128i packed_out;
// Load all inputs.
// TODO(cduvivier): Make variable declarations and allocations aligned so that
// we can use _mm_load_si128 instead of _mm_loadu_si128.
__m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
__m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]);
const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]);
const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]);
const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]);
// coeff = abs(in)
__m128i coeff0 = _mm_abs_epi16(in0);
__m128i coeff8 = _mm_abs_epi16(in8);
// coeff = abs(in) + sharpen
if (sharpen != NULL) {
const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]);
const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]);
coeff0 = _mm_add_epi16(coeff0, sharpen0);
coeff8 = _mm_add_epi16(coeff8, sharpen8);
}
// out = (coeff * iQ + B) >> QFIX
{
// doing calculations with 32b precision (QFIX=17)
// out = (coeff * iQ)
const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
__m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
__m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
// out = (coeff * iQ + B)
const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]);
const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]);
const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]);
const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]);
out_00 = _mm_add_epi32(out_00, bias_00);
out_04 = _mm_add_epi32(out_04, bias_04);
out_08 = _mm_add_epi32(out_08, bias_08);
out_12 = _mm_add_epi32(out_12, bias_12);
// out = QUANTDIV(coeff, iQ, B, QFIX)
out_00 = _mm_srai_epi32(out_00, QFIX);
out_04 = _mm_srai_epi32(out_04, QFIX);
out_08 = _mm_srai_epi32(out_08, QFIX);
out_12 = _mm_srai_epi32(out_12, QFIX);
// pack result as 16b
out0 = _mm_packs_epi32(out_00, out_04);
out8 = _mm_packs_epi32(out_08, out_12);
// if (coeff > 2047) coeff = 2047
out0 = _mm_min_epi16(out0, max_coeff_2047);
out8 = _mm_min_epi16(out8, max_coeff_2047);
}
// put sign back
out0 = _mm_sign_epi16(out0, in0);
out8 = _mm_sign_epi16(out8, in8);
// in = out * Q
in0 = _mm_mullo_epi16(out0, q0);
in8 = _mm_mullo_epi16(out8, q8);
_mm_storeu_si128((__m128i*)&in[0], in0);
_mm_storeu_si128((__m128i*)&in[8], in8);
// zigzag the output before storing it. The re-ordering is:
// 0 1 2 3 4 5 6 7 | 8 9 10 11 12 13 14 15
// -> 0 1 4[8]5 2 3 6 | 9 12 13 10 [7]11 14 15
// There's only two misplaced entries ([8] and [7]) that are crossing the
// reg's boundaries.
// We use pshufb instead of pshuflo/pshufhi.
{
const __m128i kCst_lo = PSHUFB_CST(0, 1, 4, -1, 5, 2, 3, 6);
const __m128i kCst_7 = PSHUFB_CST(-1, -1, -1, -1, 7, -1, -1, -1);
const __m128i tmp_lo = _mm_shuffle_epi8(out0, kCst_lo);
const __m128i tmp_7 = _mm_shuffle_epi8(out0, kCst_7); // extract #7
const __m128i kCst_hi = PSHUFB_CST(1, 4, 5, 2, -1, 3, 6, 7);
const __m128i kCst_8 = PSHUFB_CST(-1, -1, -1, 0, -1, -1, -1, -1);
const __m128i tmp_hi = _mm_shuffle_epi8(out8, kCst_hi);
const __m128i tmp_8 = _mm_shuffle_epi8(out8, kCst_8); // extract #8
const __m128i out_z0 = _mm_or_si128(tmp_lo, tmp_8);
const __m128i out_z8 = _mm_or_si128(tmp_hi, tmp_7);
_mm_storeu_si128((__m128i*)&out[0], out_z0);
_mm_storeu_si128((__m128i*)&out[8], out_z8);
packed_out = _mm_packs_epi16(out_z0, out_z8);
}
// detect if all 'out' values are zeroes or not
return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff);
}
void ulsch_channel_compensation_alamouti(int **rxdataF_ext, // For Distributed Alamouti Combining
int **ul_ch_estimates_ext_0,
int **ul_ch_estimates_ext_1,
int **ul_ch_mag_0,
int **ul_ch_magb_0,
int **ul_ch_mag_1,
int **ul_ch_magb_1,
int **rxdataF_comp_0,
int **rxdataF_comp_1,
LTE_DL_FRAME_PARMS *frame_parms,
unsigned char symbol,
unsigned char Qm,
unsigned short nb_rb,
unsigned char output_shift_0,
unsigned char output_shift_1) {
unsigned short rb;
__m128i *ul_ch128_0,*ul_ch128_1,*ul_ch_mag128_0,*ul_ch_mag128_1,*ul_ch_mag128b_0,*ul_ch_mag128b_1,*rxdataF128,*rxdataF_comp128_0,*rxdataF_comp128_1;
unsigned char aarx;//,symbol_mod;
// symbol_mod = (symbol>=(7-frame_parms->Ncp)) ? symbol-(7-frame_parms->Ncp) : symbol;
#ifndef __SSE3__
zeroU = _mm_xor_si128(zeroU,zeroU);
#endif
// printf("comp: symbol %d\n",symbol);
if (Qm == 4) {
QAM_amp128U_0 = _mm_set1_epi16(QAM16_n1);
QAM_amp128U_1 = _mm_set1_epi16(QAM16_n1);
}
else if (Qm == 6) {
QAM_amp128U_0 = _mm_set1_epi16(QAM64_n1);
QAM_amp128bU_0 = _mm_set1_epi16(QAM64_n2);
QAM_amp128U_1 = _mm_set1_epi16(QAM64_n1);
QAM_amp128bU_1 = _mm_set1_epi16(QAM64_n2);
}
for (aarx=0;aarx<frame_parms->nb_antennas_rx;aarx++) {
ul_ch128_0 = (__m128i *)&ul_ch_estimates_ext_0[aarx][symbol*frame_parms->N_RB_DL*12];
ul_ch_mag128_0 = (__m128i *)&ul_ch_mag_0[aarx][symbol*frame_parms->N_RB_DL*12];
ul_ch_mag128b_0 = (__m128i *)&ul_ch_magb_0[aarx][symbol*frame_parms->N_RB_DL*12];
ul_ch128_1 = (__m128i *)&ul_ch_estimates_ext_1[aarx][symbol*frame_parms->N_RB_DL*12];
ul_ch_mag128_1 = (__m128i *)&ul_ch_mag_1[aarx][symbol*frame_parms->N_RB_DL*12];
ul_ch_mag128b_1 = (__m128i *)&ul_ch_magb_1[aarx][symbol*frame_parms->N_RB_DL*12];
rxdataF128 = (__m128i *)&rxdataF_ext[aarx][symbol*frame_parms->N_RB_DL*12];
rxdataF_comp128_0 = (__m128i *)&rxdataF_comp_0[aarx][symbol*frame_parms->N_RB_DL*12];
rxdataF_comp128_1 = (__m128i *)&rxdataF_comp_1[aarx][symbol*frame_parms->N_RB_DL*12];
for (rb=0;rb<nb_rb;rb++) {
// printf("comp: symbol %d rb %d\n",symbol,rb);
if (Qm>2) {
// get channel amplitude if not QPSK
mmtmpU0 = _mm_madd_epi16(ul_ch128_0[0],ul_ch128_0[0]);
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_0);
mmtmpU1 = _mm_madd_epi16(ul_ch128_0[1],ul_ch128_0[1]);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift_0);
mmtmpU0 = _mm_packs_epi32(mmtmpU0,mmtmpU1);
ul_ch_mag128_0[0] = _mm_unpacklo_epi16(mmtmpU0,mmtmpU0);
ul_ch_mag128b_0[0] = ul_ch_mag128_0[0];
ul_ch_mag128_0[0] = _mm_mulhi_epi16(ul_ch_mag128_0[0],QAM_amp128U_0);
ul_ch_mag128_0[0] = _mm_slli_epi16(ul_ch_mag128_0[0],2); // 2 to compensate the scale channel estimate
ul_ch_mag128_0[1] = _mm_unpackhi_epi16(mmtmpU0,mmtmpU0);
ul_ch_mag128b_0[1] = ul_ch_mag128_0[1];
ul_ch_mag128_0[1] = _mm_mulhi_epi16(ul_ch_mag128_0[1],QAM_amp128U_0);
ul_ch_mag128_0[1] = _mm_slli_epi16(ul_ch_mag128_0[1],2); // 2 to scale compensate the scale channel estimate
mmtmpU0 = _mm_madd_epi16(ul_ch128_0[2],ul_ch128_0[2]);
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_0);
mmtmpU1 = _mm_packs_epi32(mmtmpU0,mmtmpU0);
ul_ch_mag128_0[2] = _mm_unpacklo_epi16(mmtmpU1,mmtmpU1);
ul_ch_mag128b_0[2] = ul_ch_mag128_0[2];
ul_ch_mag128_0[2] = _mm_mulhi_epi16(ul_ch_mag128_0[2],QAM_amp128U_0);
ul_ch_mag128_0[2] = _mm_slli_epi16(ul_ch_mag128_0[2],2); // 2 to scale compensate the scale channel estimat
ul_ch_mag128b_0[0] = _mm_mulhi_epi16(ul_ch_mag128b_0[0],QAM_amp128bU_0);
ul_ch_mag128b_0[0] = _mm_slli_epi16(ul_ch_mag128b_0[0],2); // 2 to scale compensate the scale channel estima
ul_ch_mag128b_0[1] = _mm_mulhi_epi16(ul_ch_mag128b_0[1],QAM_amp128bU_0);
ul_ch_mag128b_0[1] = _mm_slli_epi16(ul_ch_mag128b_0[1],2); // 2 to scale compensate the scale channel estima
ul_ch_mag128b_0[2] = _mm_mulhi_epi16(ul_ch_mag128b_0[2],QAM_amp128bU_0);
ul_ch_mag128b_0[2] = _mm_slli_epi16(ul_ch_mag128b_0[2],2); // 2 to scale compensate the scale channel estima
mmtmpU0 = _mm_madd_epi16(ul_ch128_1[0],ul_ch128_1[0]);
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_1);
mmtmpU1 = _mm_madd_epi16(ul_ch128_1[1],ul_ch128_1[1]);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift_1);
mmtmpU0 = _mm_packs_epi32(mmtmpU0,mmtmpU1);
ul_ch_mag128_1[0] = _mm_unpacklo_epi16(mmtmpU0,mmtmpU0);
ul_ch_mag128b_1[0] = ul_ch_mag128_1[0];
ul_ch_mag128_1[0] = _mm_mulhi_epi16(ul_ch_mag128_1[0],QAM_amp128U_1);
ul_ch_mag128_1[0] = _mm_slli_epi16(ul_ch_mag128_1[0],2); // 2 to compensate the scale channel estimate
ul_ch_mag128_1[1] = _mm_unpackhi_epi16(mmtmpU0,mmtmpU0);
ul_ch_mag128b_1[1] = ul_ch_mag128_1[1];
ul_ch_mag128_1[1] = _mm_mulhi_epi16(ul_ch_mag128_1[1],QAM_amp128U_1);
ul_ch_mag128_1[1] = _mm_slli_epi16(ul_ch_mag128_1[1],2); // 2 to scale compensate the scale channel estimate
mmtmpU0 = _mm_madd_epi16(ul_ch128_1[2],ul_ch128_1[2]);
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_1);
mmtmpU1 = _mm_packs_epi32(mmtmpU0,mmtmpU0);
ul_ch_mag128_1[2] = _mm_unpacklo_epi16(mmtmpU1,mmtmpU1);
ul_ch_mag128b_1[2] = ul_ch_mag128_1[2];
ul_ch_mag128_1[2] = _mm_mulhi_epi16(ul_ch_mag128_1[2],QAM_amp128U_0);
ul_ch_mag128_1[2] = _mm_slli_epi16(ul_ch_mag128_1[2],2); // 2 to scale compensate the scale channel estimat
ul_ch_mag128b_1[0] = _mm_mulhi_epi16(ul_ch_mag128b_1[0],QAM_amp128bU_1);
ul_ch_mag128b_1[0] = _mm_slli_epi16(ul_ch_mag128b_1[0],2); // 2 to scale compensate the scale channel estima
ul_ch_mag128b_1[1] = _mm_mulhi_epi16(ul_ch_mag128b_1[1],QAM_amp128bU_1);
ul_ch_mag128b_1[1] = _mm_slli_epi16(ul_ch_mag128b_1[1],2); // 2 to scale compensate the scale channel estima
ul_ch_mag128b_1[2] = _mm_mulhi_epi16(ul_ch_mag128b_1[2],QAM_amp128bU_1);
ul_ch_mag128b_1[2] = _mm_slli_epi16(ul_ch_mag128b_1[2],2); // 2 to scale compensate the scale channel estima
}
/************************For Computing (y)*(h0*)********************************************/
// multiply by conjugated channel
mmtmpU0 = _mm_madd_epi16(ul_ch128_0[0],rxdataF128[0]);
// print_ints("re",&mmtmpU0);
// mmtmpU0 contains real part of 4 consecutive outputs (32-bit)
mmtmpU1 = _mm_shufflelo_epi16(ul_ch128_0[0],_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_shufflehi_epi16(mmtmpU1,_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_sign_epi16(mmtmpU1,*(__m128i*)&conjugate[0]);
// print_ints("im",&mmtmpU1);
mmtmpU1 = _mm_madd_epi16(mmtmpU1,rxdataF128[0]);
// mmtmpU1 contains imag part of 4 consecutive outputs (32-bit)
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_0);
// print_ints("re(shift)",&mmtmpU0);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift_0);
// print_ints("im(shift)",&mmtmpU1);
mmtmpU2 = _mm_unpacklo_epi32(mmtmpU0,mmtmpU1);
mmtmpU3 = _mm_unpackhi_epi32(mmtmpU0,mmtmpU1);
// print_ints("c0",&mmtmpU2);
// print_ints("c1",&mmtmpU3);
rxdataF_comp128_0[0] = _mm_packs_epi32(mmtmpU2,mmtmpU3);
// print_shorts("rx:",rxdataF128[0]);
// print_shorts("ch:",ul_ch128_0[0]);
// print_shorts("pack:",rxdataF_comp128_0[0]);
// multiply by conjugated channel
mmtmpU0 = _mm_madd_epi16(ul_ch128_0[1],rxdataF128[1]);
// mmtmpU0 contains real part of 4 consecutive outputs (32-bit)
mmtmpU1 = _mm_shufflelo_epi16(ul_ch128_0[1],_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_shufflehi_epi16(mmtmpU1,_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_sign_epi16(mmtmpU1,*(__m128i*)conjugate);
mmtmpU1 = _mm_madd_epi16(mmtmpU1,rxdataF128[1]);
// mmtmpU1 contains imag part of 4 consecutive outputs (32-bit)
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_0);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift_0);
mmtmpU2 = _mm_unpacklo_epi32(mmtmpU0,mmtmpU1);
mmtmpU3 = _mm_unpackhi_epi32(mmtmpU0,mmtmpU1);
rxdataF_comp128_0[1] = _mm_packs_epi32(mmtmpU2,mmtmpU3);
// print_shorts("rx:",rxdataF128[1]);
// print_shorts("ch:",ul_ch128_0[1]);
// print_shorts("pack:",rxdataF_comp128_0[1]);
// multiply by conjugated channel
mmtmpU0 = _mm_madd_epi16(ul_ch128_0[2],rxdataF128[2]);
// mmtmpU0 contains real part of 4 consecutive outputs (32-bit)
mmtmpU1 = _mm_shufflelo_epi16(ul_ch128_0[2],_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_shufflehi_epi16(mmtmpU1,_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_sign_epi16(mmtmpU1,*(__m128i*)conjugate);
mmtmpU1 = _mm_madd_epi16(mmtmpU1,rxdataF128[2]);
// mmtmpU1 contains imag part of 4 consecutive outputs (32-bit)
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_0);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift_0);
mmtmpU2 = _mm_unpacklo_epi32(mmtmpU0,mmtmpU1);
mmtmpU3 = _mm_unpackhi_epi32(mmtmpU0,mmtmpU1);
rxdataF_comp128_0[2] = _mm_packs_epi32(mmtmpU2,mmtmpU3);
// print_shorts("rx:",rxdataF128[2]);
// print_shorts("ch:",ul_ch128_0[2]);
// print_shorts("pack:",rxdataF_comp128_0[2]);
/*************************For Computing (y*)*(h1)************************************/
// multiply by conjugated signal
mmtmpU0 = _mm_madd_epi16(ul_ch128_1[0],rxdataF128[0]);
// print_ints("re",&mmtmpU0);
// mmtmpU0 contains real part of 4 consecutive outputs (32-bit)
mmtmpU1 = _mm_shufflelo_epi16(rxdataF128[0],_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_shufflehi_epi16(mmtmpU1,_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_sign_epi16(mmtmpU1,*(__m128i*)&conjugate[0]);
// print_ints("im",&mmtmpU1);
mmtmpU1 = _mm_madd_epi16(mmtmpU1,ul_ch128_1[0]);
// mmtmpU1 contains imag part of 4 consecutive outputs (32-bit)
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_1);
// print_ints("re(shift)",&mmtmpU0);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift_1);
// print_ints("im(shift)",&mmtmpU1);
mmtmpU2 = _mm_unpacklo_epi32(mmtmpU0,mmtmpU1);
mmtmpU3 = _mm_unpackhi_epi32(mmtmpU0,mmtmpU1);
// print_ints("c0",&mmtmpU2);
// print_ints("c1",&mmtmpU3);
rxdataF_comp128_1[0] = _mm_packs_epi32(mmtmpU2,mmtmpU3);
// print_shorts("rx:",rxdataF128[0]);
// print_shorts("ch_conjugate:",ul_ch128_1[0]);
// print_shorts("pack:",rxdataF_comp128_1[0]);
// multiply by conjugated signal
mmtmpU0 = _mm_madd_epi16(ul_ch128_1[1],rxdataF128[1]);
// mmtmpU0 contains real part of 4 consecutive outputs (32-bit)
mmtmpU1 = _mm_shufflelo_epi16(rxdataF128[1],_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_shufflehi_epi16(mmtmpU1,_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_sign_epi16(mmtmpU1,*(__m128i*)conjugate);
mmtmpU1 = _mm_madd_epi16(mmtmpU1,ul_ch128_1[1]);
// mmtmpU1 contains imag part of 4 consecutive outputs (32-bit)
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_1);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift_1);
mmtmpU2 = _mm_unpacklo_epi32(mmtmpU0,mmtmpU1);
mmtmpU3 = _mm_unpackhi_epi32(mmtmpU0,mmtmpU1);
rxdataF_comp128_1[1] = _mm_packs_epi32(mmtmpU2,mmtmpU3);
// print_shorts("rx:",rxdataF128[1]);
// print_shorts("ch_conjugate:",ul_ch128_1[1]);
// print_shorts("pack:",rxdataF_comp128_1[1]);
// multiply by conjugated signal
mmtmpU0 = _mm_madd_epi16(ul_ch128_1[2],rxdataF128[2]);
// mmtmpU0 contains real part of 4 consecutive outputs (32-bit)
mmtmpU1 = _mm_shufflelo_epi16(rxdataF128[2],_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_shufflehi_epi16(mmtmpU1,_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_sign_epi16(mmtmpU1,*(__m128i*)conjugate);
mmtmpU1 = _mm_madd_epi16(mmtmpU1,ul_ch128_1[2]);
// mmtmpU1 contains imag part of 4 consecutive outputs (32-bit)
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift_1);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift_1);
mmtmpU2 = _mm_unpacklo_epi32(mmtmpU0,mmtmpU1);
mmtmpU3 = _mm_unpackhi_epi32(mmtmpU0,mmtmpU1);
rxdataF_comp128_1[2] = _mm_packs_epi32(mmtmpU2,mmtmpU3);
// print_shorts("rx:",rxdataF128[2]);
// print_shorts("ch_conjugate:",ul_ch128_0[2]);
// print_shorts("pack:",rxdataF_comp128_1[2]);
ul_ch128_0+=3;
ul_ch_mag128_0+=3;
ul_ch_mag128b_0+=3;
ul_ch128_1+=3;
ul_ch_mag128_1+=3;
ul_ch_mag128b_1+=3;
rxdataF128+=3;
rxdataF_comp128_0+=3;
rxdataF_comp128_1+=3;
}
}
_mm_empty();
_m_empty();
}
void ulsch_channel_compensation(int **rxdataF_ext,
int **ul_ch_estimates_ext,
int **ul_ch_mag,
int **ul_ch_magb,
int **rxdataF_comp,
LTE_DL_FRAME_PARMS *frame_parms,
unsigned char symbol,
unsigned char Qm,
unsigned short nb_rb,
unsigned char output_shift) {
unsigned short rb;
__m128i *ul_ch128,*ul_ch_mag128,*ul_ch_mag128b,*rxdataF128,*rxdataF_comp128;
unsigned char aarx;//,symbol_mod;
// symbol_mod = (symbol>=(7-frame_parms->Ncp)) ? symbol-(7-frame_parms->Ncp) : symbol;
#ifndef __SSE3__
zeroU = _mm_xor_si128(zeroU,zeroU);
#endif
// printf("comp: symbol %d\n",symbol);
if (Qm == 4)
QAM_amp128U = _mm_set1_epi16(QAM16_n1);
else if (Qm == 6) {
QAM_amp128U = _mm_set1_epi16(QAM64_n1);
QAM_amp128bU = _mm_set1_epi16(QAM64_n2);
}
for (aarx=0;aarx<frame_parms->nb_antennas_rx;aarx++) {
ul_ch128 = (__m128i *)&ul_ch_estimates_ext[aarx][symbol*frame_parms->N_RB_DL*12];
ul_ch_mag128 = (__m128i *)&ul_ch_mag[aarx][symbol*frame_parms->N_RB_DL*12];
ul_ch_mag128b = (__m128i *)&ul_ch_magb[aarx][symbol*frame_parms->N_RB_DL*12];
rxdataF128 = (__m128i *)&rxdataF_ext[aarx][symbol*frame_parms->N_RB_DL*12];
rxdataF_comp128 = (__m128i *)&rxdataF_comp[aarx][symbol*frame_parms->N_RB_DL*12];
for (rb=0;rb<nb_rb;rb++) {
// printf("comp: symbol %d rb %d\n",symbol,rb);
#ifdef OFDMA_ULSCH
if (Qm>2) {
// get channel amplitude if not QPSK
mmtmpU0 = _mm_madd_epi16(ul_ch128[0],ul_ch128[0]);
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift);
mmtmpU1 = _mm_madd_epi16(ul_ch128[1],ul_ch128[1]);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift);
mmtmpU0 = _mm_packs_epi32(mmtmpU0,mmtmpU1);
ul_ch_mag128[0] = _mm_unpacklo_epi16(mmtmpU0,mmtmpU0);
ul_ch_mag128b[0] = ul_ch_mag128[0];
ul_ch_mag128[0] = _mm_mulhi_epi16(ul_ch_mag128[0],QAM_amp128U);
ul_ch_mag128[0] = _mm_slli_epi16(ul_ch_mag128[0],2); // 2 to compensate the scale channel estimate
ul_ch_mag128[1] = _mm_unpackhi_epi16(mmtmpU0,mmtmpU0);
ul_ch_mag128b[1] = ul_ch_mag128[1];
ul_ch_mag128[1] = _mm_mulhi_epi16(ul_ch_mag128[1],QAM_amp128U);
ul_ch_mag128[1] = _mm_slli_epi16(ul_ch_mag128[1],2); // 2 to compensate the scale channel estimate
mmtmpU0 = _mm_madd_epi16(ul_ch128[2],ul_ch128[2]);
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift);
mmtmpU1 = _mm_packs_epi32(mmtmpU0,mmtmpU0);
ul_ch_mag128[2] = _mm_unpacklo_epi16(mmtmpU1,mmtmpU1);
ul_ch_mag128b[2] = ul_ch_mag128[2];
ul_ch_mag128[2] = _mm_mulhi_epi16(ul_ch_mag128[2],QAM_amp128U);
ul_ch_mag128[2] = _mm_slli_epi16(ul_ch_mag128[2],2); // 2 to compensate the scale channel estimate
ul_ch_mag128b[0] = _mm_mulhi_epi16(ul_ch_mag128b[0],QAM_amp128bU);
ul_ch_mag128b[0] = _mm_slli_epi16(ul_ch_mag128b[0],2); // 2 to compensate the scale channel estimate
ul_ch_mag128b[1] = _mm_mulhi_epi16(ul_ch_mag128b[1],QAM_amp128bU);
ul_ch_mag128b[1] = _mm_slli_epi16(ul_ch_mag128b[1],2); // 2 to compensate the scale channel estimate
ul_ch_mag128b[2] = _mm_mulhi_epi16(ul_ch_mag128b[2],QAM_amp128bU);
ul_ch_mag128b[2] = _mm_slli_epi16(ul_ch_mag128b[2],2);// 2 to compensate the scale channel estimate
}
#else
mmtmpU0 = _mm_madd_epi16(ul_ch128[0],ul_ch128[0]);
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift-1);
mmtmpU1 = _mm_madd_epi16(ul_ch128[1],ul_ch128[1]);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift-1);
mmtmpU0 = _mm_packs_epi32(mmtmpU0,mmtmpU1);
ul_ch_mag128[0] = _mm_unpacklo_epi16(mmtmpU0,mmtmpU0);
ul_ch_mag128[1] = _mm_unpackhi_epi16(mmtmpU0,mmtmpU0);
mmtmpU0 = _mm_madd_epi16(ul_ch128[2],ul_ch128[2]);
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift-1);
mmtmpU1 = _mm_packs_epi32(mmtmpU0,mmtmpU0);
ul_ch_mag128[2] = _mm_unpacklo_epi16(mmtmpU1,mmtmpU1);
// printf("comp: symbol %d rb %d => %d,%d,%d\n",symbol,rb,*((short*)&ul_ch_mag128[0]),*((short*)&ul_ch_mag128[1]),*((short*)&ul_ch_mag128[2]));
#endif
// multiply by conjugated channel
mmtmpU0 = _mm_madd_epi16(ul_ch128[0],rxdataF128[0]);
// print_ints("re",&mmtmpU0);
// mmtmpU0 contains real part of 4 consecutive outputs (32-bit)
mmtmpU1 = _mm_shufflelo_epi16(ul_ch128[0],_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_shufflehi_epi16(mmtmpU1,_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_sign_epi16(mmtmpU1,*(__m128i*)&conjugate[0]);
// print_ints("im",&mmtmpU1);
mmtmpU1 = _mm_madd_epi16(mmtmpU1,rxdataF128[0]);
// mmtmpU1 contains imag part of 4 consecutive outputs (32-bit)
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift);
// print_ints("re(shift)",&mmtmpU0);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift);
// print_ints("im(shift)",&mmtmpU1);
mmtmpU2 = _mm_unpacklo_epi32(mmtmpU0,mmtmpU1);
mmtmpU3 = _mm_unpackhi_epi32(mmtmpU0,mmtmpU1);
// print_ints("c0",&mmtmpU2);
// print_ints("c1",&mmtmpU3);
rxdataF_comp128[0] = _mm_packs_epi32(mmtmpU2,mmtmpU3);
// print_shorts("rx:",rxdataF128[0]);
// print_shorts("ch:",ul_ch128[0]);
// print_shorts("pack:",rxdataF_comp128[0]);
// multiply by conjugated channel
mmtmpU0 = _mm_madd_epi16(ul_ch128[1],rxdataF128[1]);
// mmtmpU0 contains real part of 4 consecutive outputs (32-bit)
mmtmpU1 = _mm_shufflelo_epi16(ul_ch128[1],_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_shufflehi_epi16(mmtmpU1,_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_sign_epi16(mmtmpU1,*(__m128i*)conjugate);
mmtmpU1 = _mm_madd_epi16(mmtmpU1,rxdataF128[1]);
// mmtmpU1 contains imag part of 4 consecutive outputs (32-bit)
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift);
mmtmpU2 = _mm_unpacklo_epi32(mmtmpU0,mmtmpU1);
mmtmpU3 = _mm_unpackhi_epi32(mmtmpU0,mmtmpU1);
rxdataF_comp128[1] = _mm_packs_epi32(mmtmpU2,mmtmpU3);
// print_shorts("rx:",rxdataF128[1]);
// print_shorts("ch:",ul_ch128[1]);
// print_shorts("pack:",rxdataF_comp128[1]);
// multiply by conjugated channel
mmtmpU0 = _mm_madd_epi16(ul_ch128[2],rxdataF128[2]);
// mmtmpU0 contains real part of 4 consecutive outputs (32-bit)
mmtmpU1 = _mm_shufflelo_epi16(ul_ch128[2],_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_shufflehi_epi16(mmtmpU1,_MM_SHUFFLE(2,3,0,1));
mmtmpU1 = _mm_sign_epi16(mmtmpU1,*(__m128i*)conjugate);
mmtmpU1 = _mm_madd_epi16(mmtmpU1,rxdataF128[2]);
// mmtmpU1 contains imag part of 4 consecutive outputs (32-bit)
mmtmpU0 = _mm_srai_epi32(mmtmpU0,output_shift);
mmtmpU1 = _mm_srai_epi32(mmtmpU1,output_shift);
mmtmpU2 = _mm_unpacklo_epi32(mmtmpU0,mmtmpU1);
mmtmpU3 = _mm_unpackhi_epi32(mmtmpU0,mmtmpU1);
rxdataF_comp128[2] = _mm_packs_epi32(mmtmpU2,mmtmpU3);
// print_shorts("rx:",rxdataF128[2]);
// print_shorts("ch:",ul_ch128[2]);
// print_shorts("pack:",rxdataF_comp128[2]);
ul_ch128+=3;
ul_ch_mag128+=3;
ul_ch_mag128b+=3;
rxdataF128+=3;
rxdataF_comp128+=3;
}
}
_mm_empty();
_m_empty();
}
void lte_idft(LTE_DL_FRAME_PARMS *frame_parms,int *z, unsigned short Msc_PUSCH) {
int *idft_in0=(int*)idft_in128[0],*idft_out0=(int*)idft_out128[0];
int *idft_in1=(int*)idft_in128[1],*idft_out1=(int*)idft_out128[1];
int *idft_in2=(int*)idft_in128[2],*idft_out2=(int*)idft_out128[2];
int *z0,*z1,*z2,*z3,*z4,*z5,*z6,*z7,*z8,*z9,*z10,*z11;
int i,ip;
// printf("Doing lte_idft for Msc_PUSCH %d\n",Msc_PUSCH);
if (frame_parms->Ncp == 0) { // Normal prefix
z0 = z;
z1 = z0+(frame_parms->N_RB_DL*12);
z2 = z1+(frame_parms->N_RB_DL*12);
//pilot
z3 = z2+(2*frame_parms->N_RB_DL*12);
z4 = z3+(frame_parms->N_RB_DL*12);
z5 = z4+(frame_parms->N_RB_DL*12);
z6 = z5+(frame_parms->N_RB_DL*12);
z7 = z6+(frame_parms->N_RB_DL*12);
z8 = z7+(frame_parms->N_RB_DL*12);
//pilot
z9 = z8+(2*frame_parms->N_RB_DL*12);
z10 = z9+(frame_parms->N_RB_DL*12);
// srs
z11 = z10+(frame_parms->N_RB_DL*12);
}
else { // extended prefix
z0 = z;
z1 = z0+(frame_parms->N_RB_DL*12);
//pilot
z2 = z1+(2*frame_parms->N_RB_DL*12);
z3 = z2+(frame_parms->N_RB_DL*12);
z4 = z3+(frame_parms->N_RB_DL*12);
z5 = z4+(frame_parms->N_RB_DL*12);
z6 = z5+(frame_parms->N_RB_DL*12);
//pilot
z7 = z6+(2*frame_parms->N_RB_DL*12);
z8 = z7+(frame_parms->N_RB_DL*12);
// srs
z9 = z8+(frame_parms->N_RB_DL*12);
}
// conjugate input
for (i=0;i<(Msc_PUSCH>>2);i++) {
*&(((__m128i*)z0)[i])=_mm_sign_epi16(*&(((__m128i*)z0)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z1)[i])=_mm_sign_epi16(*&(((__m128i*)z1)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z2)[i])=_mm_sign_epi16(*&(((__m128i*)z2)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z3)[i])=_mm_sign_epi16(*&(((__m128i*)z3)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z4)[i])=_mm_sign_epi16(*&(((__m128i*)z4)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z5)[i])=_mm_sign_epi16(*&(((__m128i*)z5)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z6)[i])=_mm_sign_epi16(*&(((__m128i*)z6)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z7)[i])=_mm_sign_epi16(*&(((__m128i*)z7)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z8)[i])=_mm_sign_epi16(*&(((__m128i*)z8)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z9)[i])=_mm_sign_epi16(*&(((__m128i*)z9)[i]),*(__m128i*)&conjugate2[0]);
if (frame_parms->Ncp==0) {
*&(((__m128i*)z10)[i])=_mm_sign_epi16(*&(((__m128i*)z10)[i]),*(__m128i*)&conjugate2[0]);
*&(((__m128i*)z11)[i])=_mm_sign_epi16(*&(((__m128i*)z11)[i]),*(__m128i*)&conjugate2[0]);
}
}
for (i=0,ip=0;i<Msc_PUSCH;i++,ip+=4) {
idft_in0[ip+0] = z0[i];
idft_in0[ip+1] = z1[i];
idft_in0[ip+2] = z2[i];
idft_in0[ip+3] = z3[i];
idft_in1[ip+0] = z4[i];
idft_in1[ip+1] = z5[i];
idft_in1[ip+2] = z6[i];
idft_in1[ip+3] = z7[i];
idft_in2[ip+0] = z8[i];
idft_in2[ip+1] = z9[i];
if (frame_parms->Ncp==0) {
idft_in2[ip+2] = z10[i];
idft_in2[ip+3] = z11[i];
}
}
switch (Msc_PUSCH) {
case 12:
// dft12f(dft_in0,dft_out0,1);
// dft12f(idft_in1,idft_out1,1);
// dft12f(idft_in2,idft_out2,1);
break;
case 24:
dft24(idft_in0,idft_out0,1);
dft24(idft_in1,idft_out1,1);
dft24(idft_in2,idft_out2,1);
break;
case 36:
dft36(idft_in0,idft_out0,1);
dft36(idft_in1,idft_out1,1);
dft36(idft_in2,idft_out2,1);
break;
case 48:
dft48(idft_in0,idft_out0,1);
dft48(idft_in1,idft_out1,1);
dft48(idft_in2,idft_out2,1);
break;
case 60:
dft60(idft_in0,idft_out0,1);
dft60(idft_in1,idft_out1,1);
dft60(idft_in2,idft_out2,1);
break;
case 72:
dft72(idft_in0,idft_out0,1);
dft72(idft_in1,idft_out1,1);
dft72(idft_in2,idft_out2,1);
break;
case 96:
dft96(idft_in0,idft_out0,1);
dft96(idft_in1,idft_out1,1);
dft96(idft_in2,idft_out2,1);
break;
case 108:
dft108(idft_in0,idft_out0,1);
dft108(idft_in1,idft_out1,1);
dft108(idft_in2,idft_out2,1);
break;
case 120:
dft120(idft_in0,idft_out0,1);
dft120(idft_in1,idft_out1,1);
dft120(idft_in2,idft_out2,1);
break;
case 144:
dft144(idft_in0,idft_out0,1);
dft144(idft_in1,idft_out1,1);
dft144(idft_in2,idft_out2,1);
break;
case 180:
dft180(idft_in0,idft_out0,1);
dft180(idft_in1,idft_out1,1);
dft180(idft_in2,idft_out2,1);
break;
case 192:
dft192(idft_in0,idft_out0,1);
dft192(idft_in1,idft_out1,1);
dft192(idft_in2,idft_out2,1);
break;
case 240:
dft240(idft_in0,idft_out0,1);
dft240(idft_in1,idft_out1,1);
dft240(idft_in2,idft_out2,1);
break;
case 288:
dft288(idft_in0,idft_out0,1);
dft288(idft_in1,idft_out1,1);
dft288(idft_in2,idft_out2,1);
break;
case 300:
dft300(idft_in0,idft_out0,1);
dft300(idft_in1,idft_out1,1);
dft300(idft_in2,idft_out2,1);
break;
}
for (i=0,ip=0;i<Msc_PUSCH;i++,ip+=4) {
z0[i] = idft_out0[ip];
/*
printf("out0 (%d,%d),(%d,%d),(%d,%d),(%d,%d)\n",
((short*)&idft_out0[ip])[0],((short*)&idft_out0[ip])[1],
((short*)&idft_out0[ip+1])[0],((short*)&idft_out0[ip+1])[1],
((short*)&idft_out0[ip+2])[0],((short*)&idft_out0[ip+2])[1],
((short*)&idft_out0[ip+3])[0],((short*)&idft_out0[ip+3])[1]);
*/
z1[i] = idft_out0[ip+1];
z2[i] = idft_out0[ip+2];
z3[i] = idft_out0[ip+3];
z4[i] = idft_out1[ip+0];
z5[i] = idft_out1[ip+1];
z6[i] = idft_out1[ip+2];
z7[i] = idft_out1[ip+3];
z8[i] = idft_out2[ip];
z9[i] = idft_out2[ip+1];
if (frame_parms->Ncp==0) {
z10[i] = idft_out2[ip+2];
z11[i] = idft_out2[ip+3];
}
}
// conjugate output
for (i=0;i<(Msc_PUSCH>>2);i++) {
((__m128i*)z0)[i]=_mm_sign_epi16(((__m128i*)z0)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z1)[i]=_mm_sign_epi16(((__m128i*)z1)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z2)[i]=_mm_sign_epi16(((__m128i*)z2)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z3)[i]=_mm_sign_epi16(((__m128i*)z3)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z4)[i]=_mm_sign_epi16(((__m128i*)z4)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z5)[i]=_mm_sign_epi16(((__m128i*)z5)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z6)[i]=_mm_sign_epi16(((__m128i*)z6)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z7)[i]=_mm_sign_epi16(((__m128i*)z7)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z8)[i]=_mm_sign_epi16(((__m128i*)z8)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z9)[i]=_mm_sign_epi16(((__m128i*)z9)[i],*(__m128i*)&conjugate2[0]);
if (frame_parms->Ncp==0) {
((__m128i*)z10)[i]=_mm_sign_epi16(((__m128i*)z10)[i],*(__m128i*)&conjugate2[0]);
((__m128i*)z11)[i]=_mm_sign_epi16(((__m128i*)z11)[i],*(__m128i*)&conjugate2[0]);
}
}
}
__m128i test_mm_sign_epi16(__m128i a, __m128i b) {
// CHECK-LABEL: test_mm_sign_epi16
// CHECK: call <8 x i16> @llvm.x86.ssse3.psign.w.128(<8 x i16> %{{.*}}, <8 x i16> %{{.*}})
return _mm_sign_epi16(a, b);
}
/* ------------------------------------------------------------------------- */
pstatus_t ssse3_sign_16s(
const INT16 *pSrc,
INT16 *pDst,
INT32 len)
{
const INT16 *sptr = (const INT16 *) pSrc;
INT16 *dptr = (INT16 *) pDst;
size_t count;
if (len < 16)
{
return general_sign_16s(pSrc, pDst, len);
}
/* Check for 16-byte alignment (eventually). */
if ((ULONG_PTR) pDst & 0x01)
{
return general_sign_16s(pSrc, pDst, len);
}
/* Seek 16-byte alignment. */
while ((ULONG_PTR) dptr & 0x0f)
{
INT16 src = *sptr++;
*dptr++ = (src < 0) ? (-1) : ((src > 0) ? 1 : 0);
if (--len == 0) return PRIMITIVES_SUCCESS;
}
/* Do 32-short chunks using 8 XMM registers. */
count = len >> 5; /* / 32 */
len -= count << 5; /* * 32 */
if ((ULONG_PTR) sptr & 0x0f)
{
/* Unaligned */
while (count--)
{
__m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
xmm0 = _mm_set1_epi16(0x0001U);
xmm1 = _mm_set1_epi16(0x0001U);
xmm2 = _mm_set1_epi16(0x0001U);
xmm3 = _mm_set1_epi16(0x0001U);
xmm4 = _mm_lddqu_si128((__m128i *) sptr); sptr += 8;
xmm5 = _mm_lddqu_si128((__m128i *) sptr); sptr += 8;
xmm6 = _mm_lddqu_si128((__m128i *) sptr); sptr += 8;
xmm7 = _mm_lddqu_si128((__m128i *) sptr); sptr += 8;
xmm0 = _mm_sign_epi16(xmm0, xmm4);
xmm1 = _mm_sign_epi16(xmm1, xmm5);
xmm2 = _mm_sign_epi16(xmm2, xmm6);
xmm3 = _mm_sign_epi16(xmm3, xmm7);
_mm_store_si128((__m128i *) dptr, xmm0); dptr += 8;
_mm_store_si128((__m128i *) dptr, xmm1); dptr += 8;
_mm_store_si128((__m128i *) dptr, xmm2); dptr += 8;
_mm_store_si128((__m128i *) dptr, xmm3); dptr += 8;
}
}
else
{
/* Aligned */
while (count--)
{
__m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
xmm0 = _mm_set1_epi16(0x0001U);
xmm1 = _mm_set1_epi16(0x0001U);
xmm2 = _mm_set1_epi16(0x0001U);
xmm3 = _mm_set1_epi16(0x0001U);
xmm4 = _mm_load_si128((__m128i *) sptr); sptr += 8;
xmm5 = _mm_load_si128((__m128i *) sptr); sptr += 8;
xmm6 = _mm_load_si128((__m128i *) sptr); sptr += 8;
xmm7 = _mm_load_si128((__m128i *) sptr); sptr += 8;
xmm0 = _mm_sign_epi16(xmm0, xmm4);
xmm1 = _mm_sign_epi16(xmm1, xmm5);
xmm2 = _mm_sign_epi16(xmm2, xmm6);
xmm3 = _mm_sign_epi16(xmm3, xmm7);
_mm_store_si128((__m128i *) dptr, xmm0); dptr += 8;
_mm_store_si128((__m128i *) dptr, xmm1); dptr += 8;
_mm_store_si128((__m128i *) dptr, xmm2); dptr += 8;
_mm_store_si128((__m128i *) dptr, xmm3); dptr += 8;
}
}
/* Do 8-short chunks using two XMM registers. */
count = len >> 3;
len -= count << 3;
while (count--)
{
__m128i xmm0 = _mm_set1_epi16(0x0001U);
__m128i xmm1 = LOAD_SI128(sptr); sptr += 8;
xmm0 = _mm_sign_epi16(xmm0, xmm1);
_mm_store_si128((__m128i *) dptr, xmm0); dptr += 8;
}
/* Do leftovers. */
while (len--)
{
INT16 src = *sptr++;
*dptr++ = (src < 0) ? -1 : ((src > 0) ? 1 : 0);
}
return PRIMITIVES_SUCCESS;
}
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);
}
int multadd_cpx_vector(int16_t *x1,
int16_t *x2,
int16_t *y,
uint8_t zero_flag,
uint32_t N,
int output_shift)
{
// 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)|
//
// zero_flag - Set output (y) to zero prior to disable accumulation
//
// N - the size f the vectors (this function does N cpx mpy. WARNING: N>=4;
//
// output_shift - shift to be applied to generate output
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_sign_epi16(*x1_128,*(__m128i*)&conjug2[0]);
tmp_re = _mm_madd_epi16(tmp_re,*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_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);
//print_ints("unpack lo:",&tmpy0[i]);
tmpy1 = _mm_unpackhi_epi32(tmp_re,tmp_im);
//print_ints("unpack hi:",&tmpy1[i]);
if (zero_flag == 1)
*y_128 = _mm_packs_epi32(tmpy0,tmpy1);
else
*y_128 = _mm_adds_epi16(*y_128,_mm_packs_epi32(tmpy0,tmpy1));
//print_shorts("*y_128:",&y_128[i]);
#elif defined(__arm__)
msg("mult_cpx_vector not implemented for __arm__");
#endif
x1_128++;
x2_128++;
y_128++;
}
_mm_empty();
_m_empty();
return(0);
}
int mult_cpx_vector(int16_t *x1, //Q15
int16_t *x2,//Q13
int16_t *y,
uint32_t N,
int output_shift)
{
// Multiply elementwise 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
uint32_t i; // loop counter
simd_q15_t *x1_128;
simd_q15_t *x2_128;
simd_q15_t *y_128;
simd_q15_t tmp_re,tmp_im;
simd_q15_t tmpy0,tmpy1;
x1_128 = (simd_q15_t *)&x1[0];
x2_128 = (simd_q15_t *)&x2[0];
y_128 = (simd_q15_t *)&y[0];
//print_shorts("x1_128:",&x1_128[0]);
// print_shorts("x2_128:",&x2_128[0]);
//right shift by 13 while p_a * x0 and 15 while
// we compute 4 cpx multiply for each loop
for(i=0; i<(N>>2); i++) {
tmp_re = _mm_sign_epi16(*x1_128,*(__m128i*)&conjug2[0]);// Q15
//print_shorts("tmp_re1:",&tmp_re[i]);
tmp_re = _mm_madd_epi16(tmp_re,*x2_128); //Q28
//print_ints("tmp_re2:",&tmp_re[i]);
tmp_im = _mm_shufflelo_epi16(*x1_128,_MM_SHUFFLE(2,3,0,1)); //Q15
//print_shorts("tmp_im1:",&tmp_im[i]);
tmp_im = _mm_shufflehi_epi16(tmp_im,_MM_SHUFFLE(2,3,0,1)); //Q15
//print_shorts("tmp_im2:",&tmp_im[i]);
tmp_im = _mm_madd_epi16(tmp_im, *x2_128); //Q28
//print_ints("tmp_im3:",&tmp_im[i]);
tmp_re = _mm_srai_epi32(tmp_re,output_shift);//Q(28-shift)
//print_ints("tmp_re shifted:",&tmp_re[i]);
tmp_im = _mm_srai_epi32(tmp_im,output_shift); //Q(28-shift)
//print_ints("tmp_im shifted:",&tmp_im[i]);
tmpy0 = _mm_unpacklo_epi32(tmp_re,tmp_im); //Q(28-shift)
//print_ints("unpack lo :",&tmpy0[i]);
tmpy1 = _mm_unpackhi_epi32(tmp_re,tmp_im); //Q(28-shift)
//print_ints("mrc rho0:",&tmpy1[i]);
*y_128 = _mm_packs_epi32(tmpy0,tmpy1); //must be Q15
//print_shorts("*y_128:",&y_128[i]);
x1_128++;
x2_128++;
y_128++;
}
_mm_empty();
_m_empty();
return(0);
}
