Rule. When implementing or modifying SIMD routines, use established inline helper functions (such as abs_diff16() or abs_diff()) rather than open-coding equivalent multi-instruction intrinsic sequences (such as _mm_or_si128(mm_subs_epu16(a, b), mm_subs_epu16(b, a))). When repetitive multi-step arithmetic sequences (such as multiply-accumulate followed by rounding shifts), mask calculations, or sub-register operations across low/high halves (such as vget_low*/vget_high, _mm_extract, or _mm256_extract*) recur across vector lanes, loop unrolls, or related kernel functions without an existing helper, factor the shared sequence into a static inline helper function.
Why. Using established inline SIMD helpers and factoring repetitive multi-instruction sequences into reusable static inline functions clarifies mathematical intent, reduces visual clutter, and eliminates copy-paste errors across sub-registers and unrolled lanes without incurring runtime call overhead.
Bad:
// Open-coded unsigned absolute difference instead of using helper const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p[1], p[0]), _mm_subs_epu16(p[0], p[1])); // Duplicated multi-instruction arithmetic across low and high sub-registers uint32x4_t m0_lo = vmull_u16(vget_low_u16(m0), vget_low_u16(s0)); m0_lo = vmlal_u16(m0_lo, vget_low_u16(m0_inv), vget_low_u16(s1)); uint16x4_t res_lo = vshrn_n_u32(m0_lo, AOM_BLEND_A64_ROUND_BITS); uint32x4_t m0_hi = vmull_u16(vget_high_u16(m0), vget_high_u16(s0)); m0_hi = vmlal_u16(m0_hi, vget_high_u16(m0_inv), vget_high_u16(s1)); uint16x4_t res_hi = vshrn_n_u32(m0_hi, AOM_BLEND_A64_ROUND_BITS);
Good:
// Use existing SIMD helper function const __m128i abs_p1p0 = abs_diff16(p[1], p[0]); // Factored repetitive lane/sub-register sequence into a static inline helper static inline uint16x4_t blend_u16x4(uint16x4_t m, uint16x4_t m_inv, uint16x4_t a, uint16x4_t b) { uint32x4_t acc = vmull_u16(m, a); acc = vmlal_u16(acc, m_inv, b); return vshrn_n_u32(acc, AOM_BLEND_A64_ROUND_BITS); } uint16x4_t res_lo = blend_u16x4(vget_low_u16(m0), vget_low_u16(m0_inv), vget_low_u16(s0), vget_low_u16(s1)); uint16x4_t res_hi = blend_u16x4(vget_high_u16(m0), vget_high_u16(m0_inv), vget_high_u16(s0), vget_high_u16(s1));