Unkreativ

@OjasSharma276 My suggestion is adding the new compiler directly to the CI after a new version is released. In our projects (OpenSource) we support 4 or more releases per compiler gcc/clang/nvcc/msvc(depending on the project because main OS is linux), icpx.

In my company’s project, we use C++11. This was the first version of C++ considered modern. The current latest version is C++23. I asked my team members why we were still using such an old version and why we couldn’t migrate to newer ones. The answer I got was simple: migration isn’t easy. The existing codebase is heavily optimized around older standards, and moving to a newer version means re evaluating and testing everything again. At the same time, the team has to maintain and develop the product, so doing a full migration alongside active development is extremely difficult. Most tech companies still use older versions of Java, C++, or other languages for this exact reason. Migration is the real problem.

The baby is growing, run vector addition on CPU: godbolt.org/z/WvosWvMe8 CUDA: godbolt.org/z/aYc33eK9n alpaka3.readthedocs.io/en/latest/ #alpaka" target="_blank" rel="nofollow noopener">enccs.github.io/gpu-programmin…

@zuhaitz_dev Beside that fact that this benchmark is useless CPU, compiler and flags are missing, for C/C++ this is on of the most important information.

And this is what many benchmarks fail to understand. You cannot compare a O(n³) nested loop that anyone can do to highly optimized BLAS libraries. A terrible algorithm will lead to terrible results regardless of the language.

@jithu83 @_streetdogg They are equal. We added in our projects the rule that you should write the constbalways on the right side so beginners know that the left hand of the const is constant. I dislike this rule but for someone who is seldom writing C++ it is easier.

@_streetdogg Any difference between first two or do they mean same thing ?

I used to get confused all the time with the usage of “const” with pointer declaration. What becomes constant (read-only) depends on which side of *️⃣ you add the “const” keyword. On the left means Data (of the datatype) is constant and on the right would mean the pointer variable is constant (readonly). On both sides? - data and pointer both become constant! Let me know of any tricks you know of which makes life with pointers... easy! I recorded a course on pointers. More details here: pyjamabrah.com/pointers/

@madhav1 if you are interessed at some.point to learn whats required to get the matmul fast, see this very good blog post aleksagordic.com/blog/matmul

Day 2 of CUDA programming - naive SGEMM on h100 - couldn’t use ncu, so timed with CUDA events - throughput: 6.35 TFLOPs, which is <1% of cuBLAS

@madhav1 Here is also a good overview over native whys e.g CUDA, ROCm and portable abstraction as alpaka, kokkos and Sycl enccs.github.io/gpu-programmin…

@madhav1 If you do not want to run your code only on NVidia GPUs have a look into alpaka alpaka3.readthedocs.io/en/latest/ Here is an example for vector multiply, change std::multiply with std::plus and you have a vectorAdd running on GPUs and CPUs. Run it in godbolt godbolt.org/z/bE7dbrv5f

Day 1 of CUDA programming - vector addition on 3050 - profiled with ncu - practiced device mem alloc + pointers

@blueberry_55555 @basit_ayantunde Write code which is running in GPUs. If the stl function is jot constexpr you can not use it on GPU. The you start to write your own linkedblist implementation, chunked allocators, ...

@basit_ayantunde name me one good situation, in regular user code (so no “i need to write a malloc”), where you should manually write a linked list instead of using a standard library type

C++'s strength is that I don't need a PhD in type-theory to implement a simple and locally-safe intrusive linked-list.

@SebAaltonen Data access is not required to return the identity of data. An access can be a transformation e.g. compression, a derivation from other data fields of the object, ...

@SebAaltonen The problem is that many develops think that the oop objects are layout in memory must be equal to the object hierarchy. That's wrong, you should seperate the user domain oop layout from the memory layout. C++ ca do this with zero overhead: github.com/alpaka-group/l…

I think the biggest issue with OOP is that people think about concrete real life objects. Computer science is about data transforms, not about "physical" objects living in the CPU memory. Wrong abstractions lead to wrong data coupling. Which hurts performance and maintainability.

🚀 alpaka 1.2.0 is here! 🎉 Now you can easily target the wide variety of GPU accelerators and CPU hardware with our latest release. 🦙 github.com/alpaka-group/a…

Good morning Pirna!

@ronnybrendel Das freut jedem der auf dem Dorf lebt und 15km bis zur nächsten Apotheke fahren muß, besonders wenn man kein Auto mehr fahren kann/darf.

The White House Memory Safety Appeal is a Security Red Herring hackaday.com/2024/02/29/the…

[#cpp20][reflect] Minimal static reflection library * C++20 - single header (gcc, clang, msvc) * Minimal API (no dependencies) * Verifies itself upon include (via static_asserts) * Compiler changes agnostic (no ifdefs) github.com/boost-ext/refl… godbolt.org/z/qYjP41vP5

@lefticus Even if you use standard C++17 the traveler will be lost at the first initialization. Even with the specs it is not easy.

Would a time traveler from 1998 be able to read YOUR C++ code today?

There is no excuse why your software is not running on CPUs and GPUs! We released alpaka 1.1.0. Thanks to all contributors and users! github.com/alpaka-group/a…

@HenschelRobert Nice to see this topic popping up this was always an issue for us when using ISAAC to live visualize PIConGPU. Please consider time slicing equal to a CPU scheduler but on minute base instead of (sub)-milliseconds as a possible way to solve the issue.

Now starting in room 607. Come and contribute your thoughts! #SC23

@btcecho @Tesla @btcecho Wo hat denn Tesla den größten SuperComputer. Wenn ihr mit groß schnell meint dann bitte ein link zur TOP500 wo dieser auf Platz 1 steht! Auch Die BF16 Operationen könne nicht an Frontier vorbei ziehen.

Der Autobauer hält an seiner #Bitcoin-Position fest. Zudem wolle @Tesla stärker in #KI investieren und habe den "größten Supercomputer der Welt in Betrieb genommen." btc-echo.de/schlagzeilen/t…

@SebAaltonen IMO the problem is that we think in types and not in concepts. The type of an object is not important and only reflects one possible implementation. Unfortionally we can not express 'auto(ConceptX) foobar = foo(bar);'

I like strongly typed languages. Being able to read the types of variables primes my brain for understanding the code I read. Auto in C++ is horrible. Makes code reading process much slower. Have to hover with mouse. Not surprised it's not any better in shading language.