bench.md

SoC: A53 (Odroid C2)

Android: Tensor G3 (Pixel 8a)

Transform functions - speedup vs C

N (imdct)	SoC tx	SoC NE10	Android tx	Android NE10	M1 tx	M1 NE10
1920 (960)	2.55×	1.60×	2.26×	1.37×	2.21×	1.60×
960 (480)	2.37×	1.62×	2.07×	1.42×	2.17×	1.58×
480 (240)	2.12×	1.48×	1.89×	1.31×	1.97×	1.41×
240 (120)	1.88×	1.41×	1.74×	1.33×	1.75×	1.32×

Post-filter function (comb_filter_const) - speedup vs C

Android = performance core (cpu7, 2.37 GHz); SoC = A53; M1 = perf core.

N	SoC	Android	M1
240	2.52×	1.36×	4.00×
480	2.53×	1.06×	5.71×
960	2.54×	0.99×	5.81×

De-emphasis functions - speedup vs C

Android = performance core (cpu7, 2.37 GHz); SoC = A53; M1 = perf core.

kernel / N	SoC	Android	M1
mono 240	3.18×	2.63×	2.41×
mono 480	3.23×	2.64×	2.43×
mono 960	3.26×	2.67×	2.43×
stereo 240	2.75×	1.97×	2.46×
stereo 480	2.78×	2.05×	2.43×
stereo 960	2.79×	2.02×	2.41×

Power of Two transforms

N/2	SoC C (ns)	SoC tx (ns)	SoC speedup	M1 C (ns)	M1 tx (ns)	M1 speedup
64	1878	1119	1.68×	138	71	1.94×
128	3895	2317	1.68×	263	146	1.80×
256	7361	3856	1.91×	518	262	1.97×
512	14612	7231	2.02×	1095	512	2.14×
1024	30876	14478	2.13×	2298	1066	2.16×

Total decoder speedup across 8 random tracks:

SoC: 25-32% Android: 20-25% M1: 15-20%