maki/openairinterface5g
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

cmac0_prec128()/rotate_cpx_vector(): Add ARMv8.0-A fallback

Browse Source 
The current code, on ARM, uses ARMv8.1-A RDM instructions. Add a
fallback for processors that do not have this instruction set.
This commit is contained in:
Eric Mittag
2026-02-10 14:45:20 +01:00
committed by Robert Schmidt
parent 15ee75980b
commit d7ea319b59
2 changed files with 33 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
openair1/PHY/MODULATION/nr_modulation.c
+17 -1
View File
@@ -760,6 +760,8 @@ static inline __attribute__((always_inline)) int16x8_t cmac0_prec128(int16x8_t x
//
int16x8_t xr = vuzp1q_s16(x, x); // even lanes
int16x8_t xi = vuzp2q_s16(x, x); // odd lanes
#ifdef __ARM_FEATURE_QRDMX
// ARMv8.1-A: Use RDM instructions
// real = ar*br - ai*bi (Q15 scaling via high-half doubling muls)
int16x8_t real = vqdmulhq_s16(xr, wr); // ≈ round((2*xr*wr)/2^16)
real = vqrdmlshq_s16(real, xi, wi); // real -= round((2*xi*wi)/2^16)
@@ -767,7 +769,21 @@ static inline __attribute__((always_inline)) int16x8_t cmac0_prec128(int16x8_t x
// imag = ar*bi + ai*br
int16x8_t imag = vqdmulhq_s16(xr, wi);
imag = vqrdmlahq_s16(imag, xi, wr); // imag += round((2*xi*wr)/2^16)
//
#else
// ARMv8.0-A fallback: Use standard 32-bit multiply
int32x4_t real_lo = vmull_s16(vget_low_s16(xr), vget_low_s16(wr));
int32x4_t real_hi = vmull_s16(vget_high_s16(xr), vget_high_s16(wr));
real_lo = vmlsl_s16(real_lo, vget_low_s16(xi), vget_low_s16(wi));
real_hi = vmlsl_s16(real_hi, vget_high_s16(xi), vget_high_s16(wi));
int32x4_t imag_lo = vmull_s16(vget_low_s16(xr), vget_low_s16(wi));
int32x4_t imag_hi = vmull_s16(vget_high_s16(xr), vget_high_s16(wi));
imag_lo = vmlal_s16(imag_lo, vget_low_s16(xi), vget_low_s16(wr));
imag_hi = vmlal_s16(imag_hi, vget_high_s16(xi), vget_high_s16(wr));
int16x8_t real = vcombine_s16(vqrshrn_n_s32(real_lo, 15), vqrshrn_n_s32(real_hi, 15));
int16x8_t imag = vcombine_s16(vqrshrn_n_s32(imag_lo, 15), vqrshrn_n_s32(imag_hi, 15));
#endif
// Re-interleave [real, imag]
int16x8x2_t produ = vzipq_s16(real, imag);
return produ.val[0];