| Line 24: | Line 24: | ||
=== Optimization flags (mail discussion regarding NEON, VFPv3 D16/D32) === | === Optimization flags (mail discussion regarding NEON, VFPv3 D16/D32) === | ||
| + | Short Background: | ||
| + | * VFPv3 = VFPv3-D32 (32 registers). Tegra and Dove only have -D16 | ||
| + | * Pointer: http://wiki.debian.org/ArmHardFloatPort , https://wiki.linaro.org/Linaro-arm-hardfloat | ||
| + | * Least common denominator would be vfpv3-d16, and no neon (tegra2) - currently. | ||
| + | * Currently we use VFPv3(-d32) which would mean we don't run on these. | ||
| + | |||
| + | Proposal: | ||
| + | * softfloat port could do vfpv3-d16 and hardfp port can do vfpv3(-d32) or even neon. Thus we get board board support (including tegra2/dove) and maximum speed for hardfp | ||
| + | |||
| + | Decision: | ||
| + | * Two modes: VFPv3-D16 for hardfp (-mfpu=vfpv3-d16) and optional set NEON enabled (-mfpu=neon) -> two architectures | ||
| + | |||
| + | === AOB === | ||
| + | * Do we need to import some patches for binutils for ARM hardfp? | ||
Contents |
Monday 2010-11-08 @8.30 - 9.30 UTC
ARM:
x86:
Short Background:
Proposal:
Decision: