Library for fast mapping of Java records to native memory

Great stuff!!!

Yup. Totally agree. Java does needs an array of structs. Hopefully value classes will help out through flattened array. But in future, one can use value records with this library with probable zero cost allocation. But the library doesn't use any reflection calls for get and set hence high performance as a result, and using records helps a lot with escape analysis. Planning to do some serious benchmarks soon. Some preliminary tests shows it's similar to c code (example code in test package). Performance suffers if record fields are arrays due to heap allocation of arrays.

I find it weird that the people steering Java have been seemingly willing to sit out the use case high performance computation while it has so dominated the computing landscape. They are just patiently incrementally iterating on all these JEP's that would support dramatically improved capabilities and make Java a very attractive platform for ML - but they keep fretting over minor interface adjustments, cycle after cycle. I get there is a philosophy of keeping the language stable and well designed, but this is really taking it to an extreme in the face of missing an entire segment of computing.

Interesting approach. I think Project Babylon did the same thing https://github.com/openjdk/babylon/blob/code-reflection/hat/...

I had tested it and it's quite fast. Actually, you don't need to generate any bytecode on the fly. Problem is when you deal with array as fields, implementation becomes difficult. You can revisit if interested to come back to such an implementation one day.

I use c-struct layout. I should be more explicit in the readme. I use classfile api to generate bytecode during initialisation of the Mem<T> and bytecode stored in cache in case if initialised again somewhere based on the same record type (I don't cache for records that are declared locally in a method). The class created from implementing Mem is a hidden class. So, basically, given a record, one can be able to analyse the layout based on record state description, and then for that Mem implementation (hidden class) we generate static final field varhandles + layout, segment is an instance field, and then generate bytecode the get and set to avoid reflection (actually, this is where most headache is in implementation). Go to the test package and see simple code for some adhoc rudimentary java (and native) files for benchmarks. Planning to test JMH benchmarks soon.

I'm actually planning to resurrect a dead raytracer project (https://github.com/mambastudio/MambaTracer) that has a GPU backend. And possible develop a future 2D API (a lot of work I know). There is an interesting world in multipurpose programming of GPUs, and it took me a while to realise prefix sum is the “hello world” in that area. As a matter of fact, most GPU code without efficient GPU prefix sum, don’t perform well for specific algorithms that might require some scan data, and that’s where CUDA dominates, which has an amazing implementation, and not in other areas such as OpenCL. The raytracer I was making, I encountered a bug that took me I think a month to deduce after some frustration, based on a discovery of my rudimentary struct implementation (through classes and annotation – which I totally agree with you they are a bit annoying) had an misalignment layout which made me rethink, something is wrong with my model approach. So, I needed to solve data and layout representation between java and native code, to be simple, concise, with little headache. If you notice the TypedMemory, the idea is to ensure a user can implement their own libraries with data models without depending on it, and thus becomes a simple plug. Unfortunately, @size will make your code have it as a dependency. Hopefully one day we have multifields as stack allocated arrays (https://web.archive.org/web/20251101203905/https://cr.openjd...) , hence maybe have value `record(char c, float[3] array) {}`

Hahaha... inspired by it actually.

I have not used SBE but looking at it, my understanding is that it starts from an explicit schema, typically XML, and generates encoder/decoder flyweights over a binary buffer. That gives much more control to the user in terms of field order (very important), sizes. Here, TypedMemory takes a different starting point in which the Java record shape is the schema, and the library derives the FFM MemoryLayout/accessors from that. I think the difference is schema/codegen/protocol orientation vs Java-type/FFM/in memorylayout orientation.

Thanks. Main goal. Unions is where I decided to pause - no simple and ergonomic way to do it at the moment.