optimize NEON functions required for libjpeg-turbo

[nemequ]

@kleisauke is trying to get libjpeg-turbo working on WASM using SIMDe. Here is a list of functions which aren't implemented yet:

• vaddhn_s32
• vld1q_dup_s16
• vld1q_lane_s16
• vld2_u8
• vmlal_lane_s16
• vmlal_lane_u16
• vmlsl_lane_s16
• vmlsl_lane_u16
• vmull_lane_s16
• vmull_lane_u16
• vqdmulh_lane_s16
• vqdmulhq_lane_s16
• vqrdmulhq_lane_s16
• vqrshrn_n_s16
• vqshluq_n_s16
• vqshrn_n_s16
• vrshrn_n_s32
• vrshrn_n_u16
• vrshrn_n_u32
• vshll_n_s16
• vshrn_n_s32
• vshrn_n_u16
• vshrn_n_u32
• vsriq_n_u16
• vst2_lane_u16
• vst2q_u8
• vst3_lane_u8
• vst4_lane_u16
• vst4_lane_u8
• vsubhn_s32

[kleisauke]

Here's a list of completed functions with their corresponding commits:

• vaddhn_s32 (commit e9ee066)
• vld1q_dup_s16 (commit 650d531)
• vld1q_lane_s16 (commit 9051a51)
• vld2_u8 (commit 85d2ed2)
• vmlal_lane_s16 (commit 82e36ed)
• vmlal_lane_u16 (commit 82e36ed)
• vmlsl_lane_s16 (commit de78ae9)
• vmlsl_lane_u16 (commit de78ae9)
• vmull_lane_s16 (commit 4dd488d)
• vmull_lane_u16 (commit 4dd488d)
• vqdmulh_lane_s16 (commit 79dc1ee)
• vqdmulhq_lane_s16 (commit 79dc1ee)
• vqrdmulhq_lane_s16 (commit dc2ea75)
• vqrshrn_n_s16 (commit 2595b3e)
• vqshluq_n_s16 (commit 77af9f1)
• vqshrn_n_s16 (commit d9260dc)
• vrshrn_n_s32 (commit a703711)
• vrshrn_n_u16 (commit a703711)
• vrshrn_n_u32 (commit a703711)
• vshll_n_s16 (commit 98ac861)
• vshrn_n_s32 (commit 8810cdd)
• vshrn_n_u16 (commit 8810cdd)
• vshrn_n_u32 (commit 8810cdd)
• vsriq_n_u16 (commit aa832e1)
• vst2_lane_u16 (commit 8ee1eb4)
• vst2q_u8 (commit 1e38dcb)
• vst3_lane_u8 (commit ae308b2)
• vst4_lane_u16 (commit b231820)
• vst4_lane_u8 (commit b231820)
• vsubhn_s32 (commit ca62754)

[nemequ]

Thanks for the reminder! I added some more earlier today, and we'll try to get that last one done soon; I think @Glitch18 is planning to take care of it.

[Glitch18]

Yes. Will be pushing the commit soon!

[nemequ]

BTW, once this is done I'd be very interested in any performance data which could point us to something we might be able to optimize in SIMDe. See https://github.com/simd-everywhere/simde/wiki/Performance-Tuning#finding-performance-problems

[kleisauke]

Great, thanks! I'll re-run the benchmark in test/bench within Chrome/Firefox once this is done. For Node.js, this requires an update of V8 to 9.1 (nodejs/node#38273) to match the renumbered/finalized WASM SIMD opcodes.

[kleisauke]

vqshluq_n_s16 was implemented with commit 77af9f1, which makes it possible to compile libjpeg-turbo for WebAssembly with SIMD support (by reusing the Arm Neon intrinsics, see commit kleisauke/wasm-vips@acd4c81). 🎉

I'll re-run the benchmarks and post the results soon, feel free to close this issue.

[kleisauke]

First set of benchmarking/profiling results can be found here: test/bench/README-simde.md.

It seems that reusing the Arm Neon intrinsics for WASM made it ~3.5x slower than its C implementation (on this benchmark). The most number of ticks (>= 10) can be observed in these functions (ordered from high to low):

• simde_vshlq_u16
• simde_vld3_u8
• simde_vld4q_s16
• simde_vclzq_s16
• simde_vld1q_lane_s16
• simde_vtrn1q_s32
• simde_vtrn2q_s32
• simde_vtrn1q_s16

Note that libjpeg-turbo is considering a whole new SIMD implementation just for WASM, so please don't spend too much time on this.