Simd | C++ image processing and machine | Computer Vision library

I got the idea of using pre-defined functions to do this: calculate "a + b", "c * 5", "d * 3" and then add the result.

But this way seems generate a lot of code. Is there any better methods to do this?

By the way, does Apache Arrow use SIMD by default(c++ version)? If not, how can I make it use SIMD?

I have trouble understanding something about simd_packed vectors in the simd module in Swift. I use the example of float4, I hope someone can help.

My understanding is that simd_float4 is a typealias of SIMD4< Float>, and MemoryLayout< Float>>.alignment = 16 (bytes), hence MemoryLayout.alignment = 16. Makes sense.

But the following I do not understand: simd_packed_float4 is also a typealias of SIMD4. And so MemoryLayout.alignment = 16.

What is the point of the "packed" in simd_packed_float4, then? Where is the "relaxed alignment" that the documentation talks about?

In the Metal Shader Language Specification (Version 2.4) ( https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf) in Table 2.4 (p.28), it says the alignment of packed_float4 is 4 (which is also the alignment of the scalar type, float), so this IS a "relaxed alignment" (as compared to the 16). That makes sense on its own, but how do I reconcile this to the above (simd_packed_float4 is typealias of SIMD4 and MemoryLayout = 16)?

I have a minimally reproducible sample which is as follows -

I am working on a C++ intrinsic wrapper for x64 and neon. I want my functions to be constexpr. My motivation is similar to Constexpr and SSE intrinsics, but #pragma omp simd and intrinsics may not be supported by the compiler (GCC) in a constexpr function. The following code is just a demonstration (auto-vectorization is good enough for addition).

this is yet another SSE is slower than normal code! Why? type of question.
I know that there are a bunch of similar questions but they don't seem to match my situation.

I am trying to implement Miller-Rabin primality test with Montgomery Modular Multiplication for fast modulo operations.
I tried to implement it in both scalar and SIMD way and it turns out that the SIMD version was around 10% slower.
that [esp+16] or [esp+12] is pointing to the modulo inverse of N if there's anyone wondering.

I am really puzzled over the fact that a supposedly 1 Latency 1c Throughput 1uops instruction psrldq takes more than 3 Latency 0.5c Throughput 1uops pmuludq.

Below is the code and the run time analysis on visual studio ran on Ryzen 5 3600.

Any idea on how to improve SIMD code and/or why is it slower than a scalar code is appreciated.

P.S. Seems like the run time analysis is off by one instruction for some reason

EDIT 1: the comment on the image was wrong, I attached a fixed version below:

I am confused about the data sharing scope of the variable acc in the flowing two cases. In the case 1 I get following compilation error: error: reduction variable ‘acc’ is private in outer context, whereas the case 2 compiles without any issues.

According to this article variables defined outside parallel region are shared.

Why is adding for-loop parallelism privatizing acc? How can I in this case accumulate the result calculated in the the for-loop and distribute a loop's iteration space across a thread team?

case 1

I wrote a function to add up all the elements of a double[] array using SIMD (System.Numerics.Vector) and the performance is worse than the naïve method.

On my computer Vector.Count is 4 which means I could create an accumulator of 4 values and run through the array adding up the elements by groups.

For example a 10 element array, with a 4 element accumulator and 2 remaining elements I would get

First, some relevant background info: I've got a CoreAudio-based low-latency audio processing application that does various mixing and special effects on audio that is coming from an input device on a purpose-dedicated Mac (running the latest version of MacOS) and delivers the results back to one of the Mac's local audio devices.

In order to obtain the best/most reliable low-latency performance, this app is designed to hook in to CoreAudio's low-level audio-rendering callback (via AudioDeviceCreateIOProcID(), AudioDeviceStart(), etc) and every time the callback-function is called (from the CoreAudio's realtime context), it reads the incoming audio frames (e.g. 128 frames, 64 samples per frame), does the necessary math, and writes out the outgoing samples.

This all works quite well, but from everything I've read, Apple's CoreAudio implementation has an unwritten de-facto requirement that all real-time audio operations happen in a single thread. There are good reasons for this which I acknowledge (mainly that outside of SIMD/SSE/AVX instructions, which I already use, almost all of the mechanisms you might employ to co-ordinate parallelized behavior are not real-time-safe and therefore trying to use them would result in intermittently glitchy audio).

However, my co-workers and I are greedy, and nevertheless we'd like to do many more math-operations per sample-buffer than even the fastest single core could reliably execute in the brief time-window that is necessary to avoid audio-underruns and glitching.

My co-worker (who is fairly experienced at real-time audio processing on embedded/purpose-built Linux hardware) tells me that under Linux it is possible for a program to requisition exclusive access for one or more CPU cores, such that the OS will never try to use them for anything else. Once he has done this, he can run "bare metal" style code on that CPU that simply busy-waits/polls on an atomic variable until the "real" audio thread updates it to let the dedicated core know it's time to do its thing; at that point the dedicated core will run its math routines on the input samples and generate its output in a (hopefully) finite amount of time, at which point the "real" audio thread can gather the results (more busy-waiting/polling here) and incorporate them back into the outgoing audio buffer.

My question is, is this approach worth attempting under MacOS/X? (i.e. can a MacOS/X program, even one with root access, convince MacOS to give it exclusive access to some cores, and if so, will big ugly busy-waiting/polling loops on those cores (including the polling-loops necessary to synchronize the CoreAudio callback-thread relative to their input/output requirements) yield results that are reliably real-time enough that you might someday want to use them in front of a paying audience?)

It seems like something that might be possible in principle, but before I spend too much time banging my head against whatever walls might exist there, I'd like some input about whether this is an avenue worth pursuing on this platform.

I can't make my MTKView clear its background. I've set the view's and its layer's isOpaque to false, background color to clear and tried multiple solutions found on google/stackoverflow (most in the code below like loadAction and clearColor of color attachment) but nothing works.

All the background color settings seem to be ignored. Setting loadAction and clearColor of MTLRenderPassColorAttachmentDescriptor does nothing.

I'd like to have my regular UIView's drawn under the MTKView. What am I missing?