Skip to content

PERFORMANCE.md

Latest commit

 

History

History
67 lines (42 loc) · 4.09 KB

PERFORMANCE.md

File metadata and controls

67 lines (42 loc) · 4.09 KB

Performance notes — DimenCache & Scaling Engine

Fast bypass (getOrPut)

When aspect ratio is off (cache key bit 63 clear, i.e. Long key ≥ 0 in the signed interpretation used in the fast check) and CalcType is one of:

  • PERCENT (ordinal 7)
  • SCALED (ordinal 11)
  • DENSITY (ordinal 14)
  • DIAGONAL (ordinal 1)
  • INTERPOLATED (ordinal 5)
  • PERIMETER (ordinal 8)

DimenCache.getOrPut returns compute() directly and does not store the result in the shard table. These paths reduce to baseValue * preComputedFactor — a single multiplication using a value computed once per configuration change.

All other strategy ordinals (AUTO, FLUID, POWER, LOGARITHMIC, FIT, FILL, …) go through the normal cache path when the cache is enabled.

Why

For the six bypassed types, measured cost of a single multiply is lower than a full cache-slot lookup; memoization is still provided by Compose remember (and by call-site batching where used). When aspect ratio is on, the computation is heavier and the cache path is used.

Pre-computed strategy scale factors (ScreenFactors)

ScreenFactors.updateFactors() runs only on configuration changes and pre-computes:

Field Formula
scale sw / 300
arMultiplier 1 + (sw - 300) * (ADJUSTMENT_SCALE + SENSITIVITY_DEFAULT * logNormalizedAr)
diagonalScale sqrt(sm² + lg²) / BASE_DIAGONAL_DP
powerScale (sw / BASE_WIDTH_DP) ^ 0.75
logScale 1 ± 0.4 * ln(sw * INV_BASE_RATIO)
interpolatedScale 1 + (sw * INV_BASE_RATIO - 1) * 0.5
perimeterScale (sm + lg) / BASE_PERIMETER_DP
aspectRatioMul 1 + SENSITIVITY_DEFAULT * logNormalizedAr

Each calculate*Dp function reads the pre-computed factor from ScreenFactors for the default path (qualifier = SMALL_WIDTH, inverter = DEFAULT, customSensitivityK = null). Non-default paths still compute inline but avoid Double conversions where possible.

Cached UiModeType

UiModeType.fromConfiguration(context, null) — which accesses SensorManager, hinge sensor lookup, and WindowMetricsCalculator — is now cached per configuration hash in DimenCache.getCachedUiModeType(context). The cache is invalidated automatically when the configuration hash changes. All *Mode / *Screen facilitators across 48 extension files read from this cache.

Eliminated Float→Double→Float conversions

  • Diagonal: sqrt((sm² + lg²).toDouble()).toFloat() eliminated — uses pre-computed diagonalScale.
  • Power: ratio.toDouble().pow(0.75).toFloat() eliminated on default path — uses pre-computed powerScale. Non-default paths use Math.pow.
  • Logarithmic: raw kotlin.math.ln() eliminated on default path — uses pre-computed logScale.

buildResizeStepsPx — zero-boxing

ResizeMath.buildResizeStepsPx writes directly to a pre-allocated FloatArray, avoiding ArrayList<Float> boxing/unboxing overhead.

Int / Float overloads

toDynamicScaledPx, toDynamicScaledDp, sdp, hdp, wdp (and their a/i/ia variants) have specialized Int and Float receiver overloads that avoid Number.toFloat() boxing.

Consumer R8/ProGuard rules

consumer-rules.pro keeps the public API surface (code, compose, common, and the listed core types). The rest of core is not blanket-kept, so R8 may still shrink, obfuscate, and inline internal helpers where no other rule forbids it. (ProGuard’s -allowoptimization is not valid for R8 and was removed.)

Persistence

DimenCache writes to a Preferences DataStore with namespace com.appdimens.dynamic.cache. The write flow uses sample(500) (not debounce) so that a first-startup burst of cache misses flushes within 500 ms of the first miss, instead of waiting until the burst quiets. For testing, call DimenCache.shutdown() to cancel the internal CoroutineScope and avoid leaked writes during teardown.

Benchmarks

Do not use SCALED / PERCENT / DENSITY / DIAGONAL / INTERPOLATED / PERIMETER without AR to measure cache throughput — those calls intentionally bypass shard storage.
⚡