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
}
void test_vmuls16 (void)
{
int16x4_t out_int16x4_t;
int16x4_t arg0_int16x4_t;
int16x4_t arg1_int16x4_t;
out_int16x4_t = vmul_s16 (arg0_int16x4_t, arg1_int16x4_t);
}
static void inline ff_dct_unquantize_h263_neon(int qscale, int qadd, int nCoeffs,
int16_t *block)
{
int16x8_t q0s16, q2s16, q3s16, q8s16, q10s16, q11s16, q13s16;
int16x8_t q14s16, q15s16, qzs16;
int16x4_t d0s16, d2s16, d3s16, dzs16;
uint16x8_t q1u16, q9u16;
uint16x4_t d1u16;
dzs16 = vdup_n_s16(0);
qzs16 = vdupq_n_s16(0);
q15s16 = vdupq_n_s16(qscale << 1);
q14s16 = vdupq_n_s16(qadd);
q13s16 = vnegq_s16(q14s16);
if (nCoeffs > 4) {
for (; nCoeffs > 8; nCoeffs -= 16, block += 16) {
q0s16 = vld1q_s16(block);
q3s16 = vreinterpretq_s16_u16(vcltq_s16(q0s16, qzs16));
q8s16 = vld1q_s16(block + 8);
q1u16 = vceqq_s16(q0s16, qzs16);
q2s16 = vmulq_s16(q0s16, q15s16);
q11s16 = vreinterpretq_s16_u16(vcltq_s16(q8s16, qzs16));
q10s16 = vmulq_s16(q8s16, q15s16);
q3s16 = vbslq_s16(vreinterpretq_u16_s16(q3s16), q13s16, q14s16);
q11s16 = vbslq_s16(vreinterpretq_u16_s16(q11s16), q13s16, q14s16);
q2s16 = vaddq_s16(q2s16, q3s16);
q9u16 = vceqq_s16(q8s16, qzs16);
q10s16 = vaddq_s16(q10s16, q11s16);
q0s16 = vbslq_s16(q1u16, q0s16, q2s16);
q8s16 = vbslq_s16(q9u16, q8s16, q10s16);
vst1q_s16(block, q0s16);
vst1q_s16(block + 8, q8s16);
}
}
if (nCoeffs <= 0)
return;
d0s16 = vld1_s16(block);
d3s16 = vreinterpret_s16_u16(vclt_s16(d0s16, dzs16));
d1u16 = vceq_s16(d0s16, dzs16);
d2s16 = vmul_s16(d0s16, vget_high_s16(q15s16));
d3s16 = vbsl_s16(vreinterpret_u16_s16(d3s16),
vget_high_s16(q13s16), vget_high_s16(q14s16));
d2s16 = vadd_s16(d2s16, d3s16);
d0s16 = vbsl_s16(d1u16, d0s16, d2s16);
vst1_s16(block, d0s16);
}
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);
}