Rule. Do not disable 8-bit SIMD optimizations or unconditionally fall back to C reference implementations using compile-time #if CONFIG_AV1_HIGHBITDEPTH inside SIMD functions. Instead, keep 8-bit SIMD routines enabled and dispatch between 8-bit and high-bit-depth routines at caller sites based on runtime bit depth (e.g., cm->seq_params.bit_depth == AOM_BITS_8).
Why. In libaom, CONFIG_AV1_HIGHBITDEPTH is a build-time configuration macro indicating that the compiled binary supports both 8-bit and 10/12-bit video streams dynamically at runtime. Using #if CONFIG_AV1_HIGHBITDEPTH inside low-bitdepth SIMD routines disables SIMD acceleration for standard 8-bit streams in high-bitdepth-enabled builds, causing severe performance regressions.
Bad:
// Inside 8-bit SIMD implementation (e.g., resize_neon.c or resize_ssse3.c): #if CONFIG_AV1_HIGHBITDEPTH // Disables 8-bit SIMD optimization whenever high-bitdepth build is enabled av1_resize_and_extend_frame_c(src, dst, filter, num_planes, phase); #else av1_resize_and_extend_frame_ssse3(src, dst, filter, num_planes, phase); #endif
Good:
// In caller dispatch (e.g., av1_scale_if_required()): #if CONFIG_AV1_HIGHBITDEPTH if ((cm->width << 1) == unscaled->y_crop_width && (cm->height << 1) == unscaled->y_crop_height && cm->seq_params.bit_depth == AOM_BITS_8) { av1_resize_and_extend_frame(unscaled, scaled, filter, phase, num_planes); } else { av1_resize_and_extend_frame_nonnormative( unscaled, scaled, (int)cm->seq_params.bit_depth, num_planes); } #else av1_resize_and_extend_frame(unscaled, scaled, filter, phase, num_planes); #endif