SSE2NEON Performance Tier Analysis (AArch64)

December 31, 2025 · View on GitHub

This document classifies SSE intrinsics by their NEON implementation complexity on AArch64 (ARMv8-A 64-bit). Analysis uses unifdef to preprocess conditional compilation paths, providing accurate instruction counts for the AArch64 target.

Note: ARMv7 (32-bit ARM) implementations may differ significantly due to missing instructions like vrndnq_f32 (directed rounding) and require fallback paths. Use --no-unifdef flag to analyze the raw file including all code paths.

To regenerate this file:

python3 scripts/gen-perf-report.py > perf-tier.md

For detailed analysis:

python3 scripts/analyze-tiers.py              # AArch64 analysis (default)
python3 scripts/analyze-tiers.py --no-unifdef # Raw file (all architectures)
python3 scripts/analyze-tiers.py --json       # JSON output
python3 scripts/analyze-tiers.py --tier 4     # Filter by tier
python3 scripts/analyze-tiers.py --weighted   # Instruction-weighted counts

SSE2NEON Performance Classification (AArch64)

Based on static analysis of NEON instruction counts per SSE intrinsic. Cycle estimates based on ARM Cortex-A72 (ARMv8-A) Software Optimization Guide.

Summary

Metric	Value
Total SSE Intrinsics	468
Direct Mappings (T1)	367 (78.4%)
Moderate Emulation (T2-T3)	87 (18.6%)
Complex Emulation (T4)	14 (3.0%)
Avg NEON Ops/Intrinsic	1.89

Performance Tiers

Tier	NEON Ops	Estimated Cycles	Description
T1	1-2	1-3	Direct NEON mapping, near-native performance
T2	3-5	4-8	Few NEON operations, slight overhead
T3	6-10	8-15	Moderate emulation, noticeable overhead
T4	10+ or special	15-50+	Complex/algorithmic, significant overhead

Note: T4 classification uses multi-factor analysis beyond raw NEON counts:

Table lookups (e.g., _mm_shuffle_epi8) are T4 due to algorithmic complexity

Pure scalar fallbacks (e.g., _mm_crc32_u16 without HW CRC) are T4

Loop-based implementations are T4 regardless of instruction count

Mixed scalar+NEON with high effective cost may be promoted to T4

Performance-Critical Intrinsics (T4)

These intrinsics have significant emulation overhead. Consider alternative algorithms when porting performance-critical code.

Intrinsic	NEON Ops	Notes
`_mm_mpsadbw_epu8`	22	SAD computation, very expensive
`_mm_sqrt_ps`	15	Newton-Raphson refinement
`_mm_cvttpd_pi32`	14
`_mm_rsqrt_ps`	13	Newton-Raphson refinement
`_mm_aesdec_si128`	12	Use HW crypto when available
`_mm_dp_ps`	9
`_mm_minpos_epu16`	9	Horizontal minimum search
`_mm_aesenc_si128`	9	Use HW crypto when available
`_mm_aesimc_si128`	9	Use HW crypto when available
`_mm_dp_pd`	4
`_mm_shuffle_epi8`	3
`_mm_shuffle_pi8`	3
`_mm_aesenclast_si128`	3	Use HW crypto when available
`_mm_aesdeclast_si128`	3	Use HW crypto when available

Efficient Intrinsics (T1 - Single NEON Instruction)

These intrinsics map directly to single NEON instructions:

Arithmetic: _mm_abs_epi16, _mm_abs_epi32, _mm_abs_epi8, _mm_abs_pi16, _mm_abs_pi32, _mm_abs_pi8, _mm_add_epi16, _mm_add_epi32, ... (+29 more)
Comparison: _mm_cmpeq_epi16, _mm_cmpeq_epi32, _mm_cmpeq_epi64, _mm_cmpeq_epi8, _mm_cmpeq_pd, _mm_cmpeq_ps, _mm_cmpgt_epi16, _mm_cmpgt_epi32, ... (+9 more)
Logical: _mm_and_pd, _mm_and_ps, _mm_and_si128, _mm_andnot_pd, _mm_andnot_ps, _mm_andnot_si128, _mm_floor_pd, _mm_floor_ps, ... (+18 more)
Load/Store: _mm_load1_pd, _mm_load1_ps, _mm_load_pd, _mm_load_ps, _mm_load_si128, _mm_loadu_ps, _mm_loadu_si128, _mm_set1_epi16, ... (+24 more)
Conversion: _mm_cvt_ps2pi, _mm_cvt_si2ss, _mm_cvt_ss2si, _mm_cvtepi32_ps, _mm_cvtps_epi32, _mm_cvtps_pi16, _mm_cvtsd_f64, _mm_cvtsd_si32, ... (+13 more)
Math: _mm_ceil_pd, _mm_ceil_ps, _mm_floor_pd, _mm_floor_ps, _mm_max_epi16, _mm_max_epi32, _mm_max_epi8, _mm_max_epu16, ... (+13 more)

Complete Tier Classification

Click to expand full list

Tier 1 (367 intrinsics)

_mm_abs_epi16, _mm_abs_epi32, _mm_abs_epi8, _mm_abs_pi16 _mm_abs_pi32, _mm_abs_pi8, _mm_add_epi16, _mm_add_epi32 _mm_add_epi64, _mm_add_epi8, _mm_add_pd, _mm_add_ps, _mm_add_sd _mm_add_si64, _mm_adds_epi16, _mm_adds_epi8, _mm_adds_epu16 _mm_adds_epu8, _mm_addsub_pd, _mm_addsub_ps, _mm_and_pd, _mm_and_ps _mm_and_si128, _mm_andnot_pd, _mm_andnot_ps, _mm_andnot_si128 _mm_avg_epu16, _mm_avg_epu8, _mm_avg_pu16, _mm_avg_pu8 _mm_blendv_epi8, _mm_blendv_pd, _mm_blendv_ps, _mm_castpd_ps _mm_castpd_si128, _mm_castps_pd, _mm_castps_si128, _mm_castsi128_pd _mm_castsi128_ps, _mm_ceil_pd, _mm_ceil_ps, _mm_ceil_sd _mm_ceil_ss, _mm_clflush, _mm_clmulepi64_si128, _mm_cmpeq_epi16 _mm_cmpeq_epi32, _mm_cmpeq_epi64, _mm_cmpeq_epi8, _mm_cmpeq_pd _mm_cmpeq_ps, _mm_cmpeq_sd, _mm_cmpeq_ss, _mm_cmpestra _mm_cmpestrc, _mm_cmpestri, _mm_cmpestrm, _mm_cmpestro _mm_cmpestrs, _mm_cmpestrz, _mm_cmpge_pd, _mm_cmpge_ps _mm_cmpge_sd, _mm_cmpge_ss, _mm_cmpgt_epi16, _mm_cmpgt_epi32 _mm_cmpgt_epi64, _mm_cmpgt_epi8, _mm_cmpgt_pd, _mm_cmpgt_ps _mm_cmpgt_sd, _mm_cmpgt_ss, _mm_cmpistra, _mm_cmpistrc _mm_cmpistri, _mm_cmpistrm, _mm_cmpistro, _mm_cmpistrs _mm_cmpistrz, _mm_cmple_pd, _mm_cmple_ps, _mm_cmple_sd _mm_cmple_ss, _mm_cmplt_epi16, _mm_cmplt_epi32, _mm_cmplt_epi8 _mm_cmplt_pd, _mm_cmplt_ps, _mm_cmplt_sd, _mm_cmplt_ss _mm_cmpneq_pd, _mm_cmpneq_ps, _mm_cmpneq_sd, _mm_cmpneq_ss _mm_cmpnge_ps, _mm_cmpnge_sd, _mm_cmpnge_ss, _mm_cmpngt_ps _mm_cmpngt_sd, _mm_cmpngt_ss, _mm_cmpnle_ps, _mm_cmpnle_sd _mm_cmpnle_ss, _mm_cmpnlt_ps, _mm_cmpnlt_sd, _mm_cmpnlt_ss _mm_cmpord_sd, _mm_cmpord_ss, _mm_cmpunord_sd, _mm_cmpunord_ss _mm_comieq_sd, _mm_comieq_ss, _mm_comige_sd, _mm_comige_ss _mm_comigt_sd, _mm_comigt_ss, _mm_comile_sd, _mm_comile_ss _mm_comilt_sd, _mm_comilt_ss, _mm_comineq_sd, _mm_comineq_ss _mm_cvt_pi2ps, _mm_cvt_ps2pi, _mm_cvt_si2ss, _mm_cvt_ss2si _mm_cvtepi16_epi32, _mm_cvtepi16_epi64, _mm_cvtepi32_epi64 _mm_cvtepi32_pd, _mm_cvtepi32_ps, _mm_cvtepi8_epi16 _mm_cvtepi8_epi32, _mm_cvtepu16_epi32, _mm_cvtepu16_epi64 _mm_cvtepu32_epi64, _mm_cvtepu8_epi16, _mm_cvtepu8_epi32 _mm_cvtpd_epi32, _mm_cvtpd_ps, _mm_cvtpi16_ps, _mm_cvtpi32_pd _mm_cvtpi32_ps, _mm_cvtpi32x2_ps, _mm_cvtps_epi32, _mm_cvtps_pd _mm_cvtps_pi16, _mm_cvtps_pi8, _mm_cvtpu16_ps, _mm_cvtsd_f64 _mm_cvtsd_si32, _mm_cvtsd_si64, _mm_cvtsi128_si32, _mm_cvtsi128_si64 _mm_cvtsi32_sd, _mm_cvtsi32_si128, _mm_cvtsi64_sd, _mm_cvtsi64_si128 _mm_cvtsi64_ss, _mm_cvtss_f32, _mm_cvtss_sd, _mm_cvtss_si64 _mm_cvtt_ps2pi, _mm_cvtt_ss2si, _mm_cvttpd_epi32, _mm_cvttps_epi32 _mm_cvttsd_si32, _mm_cvttsd_si64, _mm_cvttss_si64, _mm_div_pd _mm_div_ps, _mm_div_ss, _mm_empty, _mm_floor_pd, _mm_floor_ps _mm_floor_sd, _mm_floor_ss, _mm_free, _mm_lfence, _mm_load1_pd _mm_load1_ps, _mm_load_pd, _mm_load_ps, _mm_load_sd, _mm_load_si128 _mm_load_ss, _mm_loadh_pd, _mm_loadh_pi, _mm_loadl_epi64 _mm_loadl_pd, _mm_loadl_pi, _mm_loadr_pd, _mm_loadu_pd _mm_loadu_ps, _mm_loadu_si128, _mm_loadu_si16, _mm_loadu_si32 _mm_loadu_si64, _mm_max_epi16, _mm_max_epi32, _mm_max_epi8 _mm_max_epu16, _mm_max_epu32, _mm_max_epu8, _mm_max_pi16 _mm_max_pu8, _mm_max_sd, _mm_max_ss, _mm_mfence, _mm_min_epi16 _mm_min_epi32, _mm_min_epi8, _mm_min_epu16, _mm_min_epu32 _mm_min_epu8, _mm_min_pi16, _mm_min_pu8, _mm_min_sd, _mm_min_ss _mm_monitor, _mm_move_epi64, _mm_move_sd, _mm_move_ss _mm_movedup_pd, _mm_movehdup_ps, _mm_movehl_ps, _mm_moveldup_ps _mm_movelh_ps, _mm_movepi64_pi64, _mm_movpi64_epi64, _mm_mul_epi32 _mm_mul_epu32, _mm_mul_pd, _mm_mul_ps, _mm_mul_sd, _mm_mul_ss _mm_mul_su32, _mm_mulhi_epu16, _mm_mulhi_pu16, _mm_mulhrs_epi16 _mm_mulhrs_pi16, _mm_mullo_epi16, _mm_mullo_epi32, _mm_mwait _mm_or_pd, _mm_or_ps, _mm_or_si128, _mm_packs_epi16 _mm_packs_epi32, _mm_packus_epi16, _mm_packus_epi32, _mm_pause _mm_popcnt_u32, _mm_popcnt_u64, _mm_prefetch, _mm_rcp_ss _mm_round_pd, _mm_round_ps, _mm_round_sd, _mm_round_ss _mm_rsqrt_ss, _mm_sad_epu8, _mm_set1_epi16, _mm_set1_epi32 _mm_set1_epi64, _mm_set1_epi64x, _mm_set1_epi8, _mm_set1_pd _mm_set1_ps, _mm_set_epi64, _mm_set_epi64x, _mm_set_ps1, _mm_set_sd _mm_set_ss, _mm_setcsr, _mm_setr_epi64, _mm_setr_pd, _mm_setzero_pd _mm_setzero_ps, _mm_setzero_si128, _mm_sfence, _mm_shuffle_epi_0101 _mm_shuffle_epi_0122, _mm_shuffle_epi_0321, _mm_shuffle_epi_1001 _mm_shuffle_epi_1010, _mm_shuffle_epi_1032, _mm_shuffle_epi_2103 _mm_shuffle_epi_2211, _mm_shuffle_epi_2301, _mm_shuffle_epi_3332 _mm_shuffle_ps_0011, _mm_shuffle_ps_0022, _mm_shuffle_ps_0101 _mm_shuffle_ps_0321, _mm_shuffle_ps_1001, _mm_shuffle_ps_1010 _mm_shuffle_ps_1032, _mm_shuffle_ps_1133, _mm_shuffle_ps_2103 _mm_shuffle_ps_2200, _mm_shuffle_ps_2301, _mm_shuffle_ps_3202 _mm_shuffle_ps_3210, _mm_slli_epi16, _mm_slli_epi32, _mm_slli_epi64 _mm_sqrt_pd, _mm_sqrt_sd, _mm_sqrt_ss, _mm_store_pd, _mm_store_pd1 _mm_store_ps, _mm_store_sd, _mm_store_si128, _mm_store_ss _mm_storeh_pd, _mm_storeh_pi, _mm_storel_epi64, _mm_storel_pd _mm_storel_pi, _mm_storer_pd, _mm_storeu_pd, _mm_storeu_ps _mm_storeu_si128, _mm_storeu_si16, _mm_storeu_si32, _mm_storeu_si64 _mm_stream_load_si128, _mm_stream_pd, _mm_stream_pi, _mm_stream_ps _mm_stream_si128, _mm_stream_si32, _mm_stream_si64, _mm_sub_epi16 _mm_sub_epi32, _mm_sub_epi64, _mm_sub_epi8, _mm_sub_pd, _mm_sub_ps _mm_sub_sd, _mm_sub_si64, _mm_sub_ss, _mm_subs_epi16, _mm_subs_epi8 _mm_subs_epu16, _mm_subs_epu8, _mm_test_all_ones, _mm_undefined_pd _mm_undefined_ps, _mm_undefined_si128, _mm_unpackhi_epi16 _mm_unpackhi_epi32, _mm_unpackhi_epi64, _mm_unpackhi_epi8 _mm_unpackhi_pd, _mm_unpackhi_ps, _mm_unpacklo_epi16 _mm_unpacklo_epi32, _mm_unpacklo_epi64, _mm_unpacklo_epi8 _mm_unpacklo_pd, _mm_unpacklo_ps, _mm_xor_pd, _mm_xor_ps _mm_xor_si128

Tier 2 (73 intrinsics)

_mm_add_ss, _mm_cmpnge_pd, _mm_cmpngt_pd, _mm_cmpnle_pd _mm_cmpnlt_pd, _mm_cmpord_pd, _mm_cmpord_ps, _mm_cmpunord_pd _mm_cmpunord_ps, _mm_cvtepi8_epi64, _mm_cvtepu8_epi64, _mm_cvtpi8_ps _mm_cvtpu8_ps, _mm_cvtsd_ss, _mm_div_sd, _mm_hadd_epi16 _mm_hadd_epi32, _mm_hadd_pd, _mm_hadd_pi16, _mm_hadd_pi32 _mm_hadd_ps, _mm_hadds_epi16, _mm_hadds_pi16, _mm_hsub_epi16 _mm_hsub_epi32, _mm_hsub_pd, _mm_hsub_pi16, _mm_hsub_pi32 _mm_hsub_ps, _mm_hsubs_epi16, _mm_hsubs_pi16, _mm_loadr_ps _mm_madd_epi16, _mm_maskmove_si64, _mm_maskmoveu_si128, _mm_max_pd _mm_max_ps, _mm_min_pd, _mm_min_ps, _mm_movemask_pd _mm_mulhi_epi16, _mm_set_epi16, _mm_set_epi32, _mm_set_epi8 _mm_set_pd, _mm_set_ps, _mm_setr_epi16, _mm_setr_epi32 _mm_setr_epi8, _mm_setr_ps, _mm_shuffle_ps_2001, _mm_shuffle_ps_2010 _mm_shuffle_ps_2032, _mm_sign_epi16, _mm_sign_epi32, _mm_sign_epi8 _mm_sign_pi16, _mm_sign_pi32, _mm_sign_pi8, _mm_sll_epi16 _mm_sll_epi32, _mm_sll_epi64, _mm_sra_epi16, _mm_sra_epi32 _mm_srai_epi16, _mm_srl_epi16, _mm_srl_epi32, _mm_srl_epi64 _mm_store_ps1, _mm_storer_ps, _mm_test_all_zeros, _mm_testc_si128 _mm_testz_si128

Tier 3 (14 intrinsics)

_mm_aeskeygenassist_si128, _mm_crc32_u16, _mm_crc32_u32 _mm_crc32_u64, _mm_crc32_u8, _mm_cvtpd_pi32, _mm_maddubs_epi16 _mm_maddubs_pi16, _mm_movemask_epi8, _mm_movemask_pi8 _mm_movemask_ps, _mm_rcp_ps, _mm_sad_pu8, _mm_test_mix_ones_zeros

Tier 4 (14 intrinsics)

_mm_aesdec_si128, _mm_aesdeclast_si128, _mm_aesenc_si128 _mm_aesenclast_si128, _mm_aesimc_si128, _mm_cvttpd_pi32, _mm_dp_pd _mm_dp_ps, _mm_minpos_epu16, _mm_mpsadbw_epu8, _mm_rsqrt_ps _mm_shuffle_epi8, _mm_shuffle_pi8, _mm_sqrt_ps