How to build sysutils/edk2 at rpi4 (for example) that matches https://github.com/pftf/RPi4/releases/tag/v1.18 (for example)

[Mark Millard]

On 2020-Aug-11, at 00:46, Mark Murray <markm at freebsd.org> wrote:

>> On 11 Aug 2020, at 07:44, Mark Millard via freebsd-arm <freebsd-arm at freebsd.org> wrote:
>> 
>> Unfortunately, I'm not aware of anything for macchiatobin
>> that is analogous to https://github.com/pftf/RPi3 and
>> https://github.com/pftf/RPi4 off which to derive what
>> source would reproduce some known release that might
>> have been put to use with FreeBSD (a tested combination).
> 
> I have a MacchiatoBin Double Shot doing dual duty as an ARM
> build box and firewall. It has a dual-port PCI NIC in it.
> 
> I may have the same edk2 build as you, and it works really well.
> 
> In an ideal world, I'd be using the native ethernet ports and I'd
> have a cheap video card in the PCI slot, but the serial console
> works well enough for now.

Nothing the this one's PCI slot.

Realtek USB 10/100/1000 LAN via using
https://reviews.freebsd.org/D25809 .

Serial console.

SATA SSD.

FYI:

Its -j4 buildworld buildkernel results are somewhat slower than
for an OverDrive 1000. (Similarly for poudriere bulk runs. Single
threaded activity is a different story.) Builds are CPU/memory
bound, not I/O bound, for both systems.


ENVIRONMENT: -mcpu=cortex-a57 OverDrive 1000 @1.7G Hz, 8 GiByte RAM total
             (2 DIMMs slots, both populated):

World built in 15508 seconds, ncpu: 4, make -j4
Kernel(s)  GENERIC-NODBG built in 1117 seconds, ncpu: 4, make -j4


ENVIRONMENT: -mcpu=cortex-a72 MACCHIATObin Double Shot @ 2G Hz, 16 GiByte RAM
             (1 DIMM slot):

World built in 18789 seconds, ncpu: 4, make -j4
Kernel(s)  GENERIC-NODBG built in 1296 seconds, ncpu: 4, make -j4


The only differences in my equivalents of src.conf are the 57's
vs. 72's in -mcpu= usage (same arch different tuning). head
-r363590 doing self-hosted, from-scratch rebuilds.

What contributes? L0-L2 cache size differences and dual DIMM
(dual channel? interleaving?) vs. single DIMM. Cache performance
for 4 active threads keeps the two systems similarly performant
for problems that fit in both sizes of RAM cache. But the
OverDrive looks to have 4 times the RAM cache. That matters for
-j4 buildworld buildkernel .

[The .png (that will not make it to the list) is a graph
of benchmark data from my C++17 variant of some old HINT
benchmarks. Data types: d=double, ull=unsigned long long,
ul=unsigned long, ui=unsigned int. In this context,
variations in ull vs. ul results for a given system give
a clue as to the degree of other sources of variability:
both are 64-bit unsigned here. Y-axis: larger is faster.
While gcc10 was used, the system libc++ and such were
used instead of gcc's libraries.]


===
Mark Millard
marklmi at yahoo.com
( dsl-only.net went
away in early 2018-Mar)