Rule. In composite DSP and SIMD routines (such as compound average upsampled prediction or distance-weighted compound prediction), invoke the corresponding standalone DSP primitive functions (such as aom_comp_avg_pred_*, aom_highbd_comp_avg_pred_*, or aom_dist_wtd_comp_avg_pred_*) instead of duplicating or open-coding vector intrinsic loops. When chaining multi-stage operations, pass intermediate buffers between the specialized building blocks and maintain consistent delegation across low-bitdepth and high-bitdepth implementations.
Why. Inlining arithmetic or blending loops across multiple composite functions duplicates complex SIMD logic, increases maintenance burden, and creates redundant code across architecture variants. Delegating sub-operations to dedicated standalone primitives eliminates duplicate vector code, ensures consistent math, rounding, and bitdepth handling across call sites, reduces binary size, and allows composite routines to automatically benefit from optimizations or bug fixes in the underlying building blocks.
Bad:
// Anti-pattern: Open-coding inline SIMD loops inside composite prediction functions void aom_highbd_comp_avg_upsampled_pred_neon(...) { aom_highbd_upsampled_pred_neon(xd, cm, mi_row, mi_col, mv, comp_pred8, ...); for (int i = 0; i < width * height; i += 8) { uint16x8_t cp = vqrshlq_u16(vaddq_u16(vld1q_u16(&pred[i]), vld1q_u16(&comp_pred[i])), shift); vst1q_u16(&comp_pred[i], cp); } } void aom_highbd_dist_wtd_comp_avg_upsampled_pred_neon(...) { aom_highbd_upsampled_pred_neon(xd, cm, mi_row, mi_col, mv, comp_pred8, ...); for (int i = 0; i < width * height; i += 8) { // Inlined distance-weighted blending calculation loop uint32x4_t cp0 = vmlal_u16(vmull_u16(vget_low_u16(p), bck_offset_u16), vget_low_u16(cp), fwd_offset_u16); ... } }
Good:
// Delegate composite sub-operations to dedicated standalone SIMD primitives void aom_highbd_comp_avg_upsampled_pred_neon(...) { aom_highbd_upsampled_pred_neon(xd, cm, mi_row, mi_col, mv, comp_pred8, width, height, subpel_x_q3, subpel_y_q3, ref8, ref_stride, bd, subpel_search); aom_highbd_comp_avg_pred_neon(comp_pred8, pred8, width, height, comp_pred8, width); } void aom_highbd_dist_wtd_comp_avg_upsampled_pred_neon(...) { aom_highbd_upsampled_pred_neon(xd, cm, mi_row, mi_col, mv, comp_pred8, width, height, subpel_x_q3, subpel_y_q3, ref8, ref_stride, bd, subpel_search); aom_highbd_dist_wtd_comp_avg_pred_neon(comp_pred8, pred8, width, height, comp_pred8, width, jcp_param); }