bt | BitTorrent library and client with DHT, magnet links, encryption and more | Stream Processing library

Hi all I'm going to answer my own question here might be useful to others. The answer for me is that it was because I was using a generic OpenBLAS, not an Intel processor-specific version of BLAS, and running in debug mode.

With optimization at compile time and using an Intel processor-specific version of BLAS:

1. Bt = B.t() and then A = Bt * C is definitely slower than A = B.t() * C, as we would expect due to the storing of the intermediate step.
2. A = B.t() * C is faster than A = B * C, if B is square (I know this isn't the same number), but the difference is small, maybe 0-20% for the numbers I am using.
3. Along a similar line, A = B.rows(0, 499) * C is slower than A = B.cols(0, 499).t() * C.

The explanation is I believe that column access is faster than row access. B.t() * C uses columns of both B and C, whereas B * C uses rows of B and columns of C.

All of this is much faster than loops. So use BLAS over C++ manual loops -- this is way more important than worrying about rows vs columns.

One anomaly: A = B.rows(0, 499) is still faster than A = B.cols(0, 499). Any ideas on why would be appreciated!

P.S. also tips on handing tensors higher than 2D in C++ would be appreciated. I hate arma::Cubes although I do use them.

Here is a suggestion where TimeAndCode.Code is an enum rather than a string.

It will always order entries with indentical timespans in a repeated CO, GO, BT pattern; meaning that if e.g. five entries with identical timespans has the following Code selection: 2 x CO, 1 x BT, 2 x GO, it will always order them as CO, GO, BT, CO, GO (as opposed to CO, GO, CO, GO, BT).

I achieve this by generating an OrderBy property based on the timespan, an index (generated inside a nested group) and the numerical Code value for each entry.

Using the following types:

found some answers to my own question, sharing with community

6-sec auto-switch mode is a new feature in Android 12, which works only if (mode == AudioSystem.MODE_IN_COMMUNICATION) (check out flow related to MSG_CHECK_MODE_FOR_UID flag). This should help for MODE_IN_COMMUNICATION set to AudioManager and left after app exit, this was messing with global/system-level audio routing. There is also a brand new AudioManager.OnModeChangedListener called when mode is (auto-)changing

and setSpeakerphoneOn turns out to be deprecated, even if this isn't marked in doc... we have new method setCommunicationDevice(AudioDeviceInfo) and in its description we have info about startBluetoothSco(), stopBluetoothSco() and setSpeakerphoneOn(boolean) deprecation. I'm using all three methods and now on Android 12 I'm iterating through getAvailableCommunicationDevices(), comparing type of every item and if desired type found I'm calling setCommunicationDevice(targetAudioDeviceInfo). I'm NOT switching audio mode at all now, staying on MODE_NORMAL. All my streams are AudioManager.STREAM_VOICE_CALL type (where applicable)

for built-in earpiece audio playback aka. "ear-friendly mode" we were using

According to man7.org getifaddrs, any of the socket operations could be a cause for EBADF

ERRORS

getifaddrs() may fail and set errno for any of the errors specified for socket(2), bind(2), getsockname(2), recvmsg(2), sendto(2), malloc(3), or realloc(3).

Unrelated, but do you do freeifaddrs() somewhere?

You don't null terminate the buffer. Therefore strcpy wil just copy the contents of buf until a null character is encountered which is undefind behaviour.

Add this right after the forloop:

It looks like this might be the root cause:

https://c-ares.org/mail/c-ares-archive-2017-08/0002.shtml

The buffer size of pipes isn't unlimited. You're creating a deadlock, where the child process is hanging because it's trying to write to a buffer that's full, and your parent process doesn't try to read anything from the buffer until the child process has completed. To fix the problem, you need to use another thread to read from the buffer while the child process is still running. The simplest way to do this is to use readProcess or a similar function in place of createProcess. If this doesn't give you enough flexibility to do what you want, then you'll need to create and manage the other thread yourself, which you can see how to do by looking at how readProcess is implemented.

execve is special; it's the only one that has special interaction with PTRACE_TRACEME. The way strace works, other system calls do show the 32-bit call number. (And modern strace needs special help to know whether that's a 32-bit call number for int 0x80 / sysenter, or a 64-bit call number, since 64-bit processes can still invoke int 0x80, although they normally shouldn't. This support was only added in 2019, with PTRACE_GET_SYSCALL_INFO)

You're right, when the kernel is actually invoked, EAX holds 11, __NR_execve from unistd_32.h. It's set by mov $0xb,%eax before glibc's execve wrapper jumps to the VDSO page to enter the kernel via whatever efficient method is supported on this hardware (normally sysenter.)

But execution doesn't actually stop until it reaches some code in the main execve implementation that checks for PTRACE_TRACEME and raises SIGTRAP.

Apparently sometime before that happens, it calls void set_personality_64bit(void) in arch/x86/kernel/process_64.c, which includes