415 #define _pM128(a) _mm_castsi128_ps(_pM128i(a))  macro
2823     res = _mm_add_ps(_pM128(a),_pM128(b)); //SSE, use only low 64 bits  in vadd_f32()
3546     tmp =  _mm_mul_ps(_pM128(a),_pM128(b));  in vmul_f32()
3848     res = _mm_mul_ps (_pM128(c), _pM128(b));  in vmla_f32()
3849     res = _mm_add_ps (_pM128(a), res);  in vmla_f32()
4046     res = _mm_mul_ps (_pM128(c), _pM128(b));  in vmls_f32()
4047     res = _mm_sub_ps (_pM128(a), res);  in vmls_f32()
5048     res = _mm_cmpeq_ps(_pM128(a), _pM128(b) );  in vceq_f32()
5143     res = _mm_cmpge_ps(_pM128(a),_pM128(b)); //use only 2 first entries  in vcge_f32()
5288     res = _mm_cmple_ps(_pM128(a),_pM128(b));  in vcle_f32()
5412     res = _mm_cmpgt_ps(_pM128(a),_pM128(b)); //use only 2 first entries  in vcgt_f32()
5544     a0 = _mm_and_ps (_pM128(a), *(__m128*)&c7fffffff);  in vcage_f32()
5545     b0 = _mm_and_ps (_pM128(b), *(__m128*)&c7fffffff);  in vcage_f32()
5571     a0 = _mm_and_ps (_pM128(a), *(__m128*)&c7fffffff);  in vcale_f32()
5572     b0 = _mm_and_ps (_pM128(b), *(__m128*)&c7fffffff);  in vcale_f32()
5598     a0 = _mm_and_ps (_pM128(a), *(__m128*)&c7fffffff);  in vcagt_f32()
5599     b0 = _mm_and_ps (_pM128(b), *(__m128*)&c7fffffff);  in vcagt_f32()
5625     a0 = _mm_and_ps (_pM128(a), *(__m128*)&c7fffffff);  in vcalt_f32()
5626     b0 = _mm_and_ps (_pM128(b), *(__m128*)&c7fffffff);  in vcalt_f32()
5784     res = vabdq_f32(_pM128(a), _pM128(b));  in vabd_f32()
6346     hadd128 = _mm_hadd_ps (_pM128(a), _pM128(b));  in vpadd_f32()
6779     res = _mm_rcp_ps(_pM128(a));  in vrecpe_f32()
6847     res = _mm_rsqrt_ps(_pM128(a));  in vrsqrte_f32()
6922     res = vrecpsq_f32(_pM128(a), _pM128(b));  in vrecps_f32()
12695     res =  _mm_cvtps_epi32(_pM128(a)); //use low 64 bits of result only  in vcvt_s32_f32()
12704     res = vcvtq_u32_f32(_pM128(a));  in vcvt_u32_f32()
12748     return64(vcvtq_n_s32_f32(_pM128(a),b));
14609     res = _mm_and_ps (_pM128(a), *(__m128*)c7fffffff); //use 64 low bits only  in vabs_f32()
14743     res = _mm_xor_ps (_pM128(a), *(__m128*) c80000000); //use low 64 bits  in vneg_f32()
15592     sel1 = _mm_and_ps   (_pM128(a), _pM128(b));  in vbsl_f32()
15593     sel2 = _mm_andnot_ps (_pM128(a), _pM128(c));  in vbsl_f32()