Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- Nice. Very clean api.by kosolam
- Thanks. Main goal. Unions is where I decided to pause - no simple and ergonomic way to do it at the moment.by joe_mwangi
- Wouldn’t apache arrow serve the same purpose?by wwarner
- Why not use graal?
- This is very similar to SBE encoder/decoder flyweights over raw memory? What are the differences?by c-fe
- 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.by joe_mwangi
- At first glance it reminds me of C#'s Span<T>.by jayd16
- Hahaha... inspired by it actually.by joe_mwangi
- Ha, I wrote something like this ages ago when I had to interact with a Java application and really wanted/needed Python utilities. I created my own concurrent, type-safe DDL to shepherd that data.by hansvm
- I don't usually like annotations in Java, but I'm rather liking your use of @size here.
Are you working on anything in particular that drove you to build this? Looks great!@size(3) int[] zby elric - 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) {}`by joe_mwangi
- What is the positioning for this and how does it work? A comparison to SBE might be nice.
I understand the issue about using Layout and MemorySegment being verbose but the reason I'm using those things it to develop high performance software that uses off-help memory and bypasses object allocation.
What does "map Java record types onto native memory" actually mean? Did you somehow turn a Java record into a flyweight or is `Point point = points.get(0);` just instantiating a record instance using data read from off-help memory? If it's a dynamic mapping library using reflection, that's cool but doesn't it kill the performance goals for most Java off heap usage?
Is this more of a off-heap to heap bridge for pulling data into the normal Java space when performance isn't critical?
by matt_heimer - 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.by joe_mwangi
- I did something similar a few years back, with a slightly different approach to declaration, using interfaces to denote the layout of the struct. Mutation was opt-in by exposing setters using the (of the time) standard JavaBeans layout and an annotation processor took care of the codegen of an implementing class, which could be used where you wanted an on-heap box of an off-heap structure.
One benefit of this approach was that by using the interface as the type you could fairly easily support a flyweight pattern, reducing GC pressure when working with large off-heap collections. The parallels between stateless interfaces and offheap structs was also quite pleasing.
I'd love to see a similar effort using more modern techniques than Unsafe et al.
by steve_barham - 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.
by joe_mwangi - This is interesting. Java desperately needs an array of struct for type safe sugar over high performance arenas, but the areas you’d turn to this would be in a zero allocation effort where the cost of the this library’s off-heap and the object allocation in the getters and setters etc largely negate the advantages for a lot of use cases.by wood_spirit
- 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.by zmmmmm
- 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.by joe_mwangi
- Inspired by this, I wrote a quick prototype using MethodHandle combinators: https://gist.github.com/DasBrain/19804df69c78cee257dd0294b00...
It doesn't have all the bells and whistles - so no support for annotations or arrays, but for ~2h of work, it shows what is possible without resorting to bytecode generation.
by J-Kuhn - Great stuff!!!by joe_mwangi