> CUDA architects [...] happily exposed every ugly details of underlying hardware to the programmer if that allows a bit more performance.
After spending more than a decade dancing around all the underlying x86 hidden stuff for low-level optimization, I appreciate CUDA a lot. Everything is there under your total control. No more one size fits all. Higher barrier of entry but no surprises and less time spent debugging to figure out what landmine your code stepped into.
(a) You mention that the NVidia docs push people to use libraries, etc. to really get high performance CUDA kernels, rather than writing them themselves. My argument would be that SIMD is exactly the same - they're something really that are perfect if you're writing a BLAS implementation but are too low level for most developers thinking about a problem to make use of.
(b) You show a problem where autovectorisation fails because of branching, and jump straight to intrinsics as the solution which you basically say are ugly. Looking at the intrinsic code, you're using a mask to deal with the branching. But there's a middle ground - almost always you would want to try restructuring the problem, e.g. splitting up loops and adding masks where there's conditions - i.e. lean into the SIMD paradigm. This would also be the same advice in CUDA.
(c) As you've found, GCC actually performs quite poorly for x86-64 optimisations compared to Intel. It's not always clear cut though, the Intel Compiler for e.g. sacrifices IEEE 764 float precision and go down to ~14 digits of precision in it's defaults, because it sets the flag `-fp-model=fast -fma`. This is true of both the legacy and new Intel compiler. If you switch to `-fp-model=strict` then you may find that the results are closer.
(d) AVX512 is quite hardware specific. Some processors execute these instructions much better than others. It's really a collection of extensions, and you get frequency downclocking that's better/worse on different processors as these instructions are executed.
Regarding (b), I would never rely on auto vectorization because I have no insight into it. The only way to check if my code is written such that auto-vectorization can do its thing is to compile my program with every combination of compiler, optimization setting and platform I intend to support, disassemble all the resulting binaries, and analyze the disassembly to try to figure out if it did autovectorization in the way I expect. That's a terrible developer experience; writing intrinsics by hand is much easier, and more robust. I'd need to re-check every piece of autovectorized code after every change and after every compiler upgrade.
I just treat autovectorization like I treat every other fancy optimization that's not just constant folding and inlining: nice when it happens to work out, it probably happens to make my code a couple percent faster on average, but I absolutely can't rely on it in any place where I depend on the performance.
> every combination of compiler, optimization setting and platform I intend to support, disassemble all the resulting binaries, and analyze the disassembly to try to figure out if it did autovectorization in the way I expect
I just used to fire up VTune and inspect the hot loops... typically if you care about this you're only really working on hardware targeting the latest instruction sets anyway in my experience. It's only if you're working on low level libraries I would bother doing intrinsics all over the place.
For most consumer software you want to be able to fall back to some lowest-common-denominator hardware anyway otherwise people using it run into issues - same reason that Debian, Conda, etc. only go up to really old instructions sets.
I work on games sometimes, where the goal is: "run as fast as possible on everyone's computer, whether it's a 15 year old netbook with an Intel Atom or a freshly built beast of a gaming desktop". As a result, the best approach is to discover supported instructions at runtime and dispatch to a function that's using those instructions (maybe populating a global vector function table at launch?). Next best is to assume some base level vector support (maybe the original AVX for x86, Neon for ARM) and unconditionally use those. Targeting only the latest instruction sets is a complete non-starter.
> (d) AVX512 is quite hardware specific. Some processors execute these instructions much better than others. It's really a collection of extensions, and you get frequency downclocking that's better/worse on different processors as these instructions are executed.
To re-iterate, this is our observation as well. The first AVX512 processors would execute such code quite fast for a short time, then overheat and throttle, leading to a worse wall-time performance than the corresponding AVX256 code.
I am not sure if there is a better way to find the fastest code path besides "measure on the target system", which of course comes with its own challenges.
The processors with severe throttling from AVX-512 were server CPUs, i.e. Skylake Server and its derivatives, like Cascade Lake and Cooper Lake.
Only few of those CPUs have been used in workstations, i.e. high-end desktop computers.
The vast majority of the CPUs with AVX-512 that can be encountered at the general population are either AMD Zen 4 and Zen 5 CPUs or some old Intel CPUs from the Tiger Lake, Ice Lake and Rocket Lake families. All these do not have AVX-512 throttling problems.
The owners of server computers are more likely to be knowledgeable about them and choose programs compiled with an appropriate target CPU model.
Therefore I believe that nowadays, when the percentage of computers with good AVX-512 support is increasing, and even Intel is expected to introduce by the end of the year Nova Lake with AVX-512 support, an application should be compiled such that whenever it detects AVX-512 support it should dispatch to the corresponding branch.
On the computers with AVX-512 support, using it can provide a significant increase in performance, while the computers where this could be harmful are more and more unlikely to be encountered outside datacenters that have failed to update their servers.
Skylake Server was introduced 9 years ago and Ice Lake Server, which corrected the behavior, was introduced 6 years ago. Therefore, wherever performance matters, the Skylake Server derivatives would have been replaced by now, as a single Epyc server can replace a cluster of servers with Skylake Server CPUs, at a much lower power consumption and with a higher performance.
> I am not sure if there is a better way to find the fastest code path besides "measure on the target system", which of course comes with its own challenges.
Yeah, and it's incredibly frustrating because there is almost zero theory on how to write performant code. Will caching things in memory be faster than re-requesting them over network? Who knows! Sometimes it won't! But you can't predict what those times will be beforehand which turns this whole field into pure black magic instead of anything remotely similar to engineering or science, since theoretical knowledge has no relation to reality.
At my last job we had one of the weirdest "memory access is slo-o-o-ow" scenarios I've ever seen (and it would reproduce pretty reliably... after about 20 hours of the service's continuous execution): somehow, due to peculiarities of the GC and Linux physical memory manager, almost all of the working set of our application would end up in a single physical DDR stick, as opposed to being evenly spread across 4 stick the server actually has. Since a single memory stick literally can't cope with such high data throughput, the performance tanked. And it took quite some time us to figure out what the hell was going on because nothing came up on the perf graphs or metrics or whatever: it's just that almost everything in the the application's userspace became slower. No, the CPU is definitely not throttled, it's actually 20–30% idle. No, there is almost zero disk activity, and the network is fine. Huh?!
If you wish to see some speedups using AVX512, without limiting yourself to C or C++, you might want to try ISPC (https://ispc.github.io/index.html).
You'll get sane aliasing rules from the perspective of performance, multi-target binaries with dynamic dispatching and a lot more control over the code generated.
Yes, works well with AMD. You can compile multi-target so that you'll have e.g. SSE4.2, AVX2, AVX512 support built to your binaries and the best (widest) version is picked by the runtime automatically.
> The answer, if it’s not obvious from my tone already:), is 8%.
Not if the data is small and in cache.
> The performant route with AVX-512 would probably include the instruction vpconflictd, but I couldn’t really find any elegant way to use it.
I think the best way to do this is duplicate sum_r and count 16 times, so each pane has a seperate accumulation bucket and there can't be any conflicts. After the loop, you quickly do a sum reduction for each of the 16 buckets.
Isn't it another way of saying what the author says in the previous paragraph, namely that "ideal SIMD speedup can only come from problems that are compute bound"?
If the cost of getting the input data into the cache is already large compared to processing it with the non-vectorized code, then SIMD cannot achieve meaningful speedup. The opposite of this condition (processing is expensive compared to the cost of data into the cache) is basically the definition "compute bound".
Yeah N is big enough that entire data isn't in the cache, but the memory access pattern here is the next best thing: totally linear, predictable access. I remember seeing around 94%+ L1d cache hit rate.
> In CPU world there is a desire to shield programmers from those low-level details, but I think there are two interesting forces at play now-a-days that’ll change it soon. On one hand, Dennard Scaling (aka free lunch) is long gone, hardware landscape is getting increasingly fragmented and specialized out of necessity, software abstractions are getting leakier, forcing developers to be aware of the lowest levels of abstraction, hardware, for good performance.
The problem is that not all programming languages expose SIMD, and even if they do it is only a portable subset, additionally the kind of skills that are required to be able to use SIMD properly isn't something everyone is confortable doing.
I certainly am not, still managed to get around with MMX and early SSE, can manage shading languages, and that is about it.
The good news is that the portable subset of SIMD is all you really need anyway. If you go beyond the portable subset, you need per-architecture code writing and testing, and you're mostly talking about pretty small gains relative to the cost.
Initial example takes array pointers without the __restrict__ keyword/extension so compiler might assume they could be aliased to same address space and will code defensively.
Would be interesting to see if auto vec performs better with that addition.
> let the compilers know that the float* are aligned
Reminded me of way back before OpenGL 2.0, and I was trying to get Vertex Buffer Objects working in my Delphi program using my NVIDIA graphics card. However it kept crashing occasionally, and I just couldn't figure out why.
I've forgotten a lot of the details, but either the exception message didn't make sense or I didn't understand it.
Anyway, after bashing my head for a while I had an epiphany of sorts. NVIDIA liked speed, vertices had to be manipulated before uploading to the GPU, maybe the driver used aligned SIMD instructions and relied on the default alignment of the C memory allocator?
In Delphi the default memory allocator at the time only did 4 byte aligned allocations, and so I searched and found that Microsoft's malloc indeed was default aligned to 16 bytes. However the OpenGL standard and VBO extension didn't say anything about alignment...
Manually aligned the buffers and voila, the crashes stopped. Good times.
which honestly, shouldn't be neccessary today with avx512. There's essentially no reason to prefer the aligned load/store commands over the unaligned ones - if the actual pointer is unaligned it will function correctly at half the throughput, while if it_is_ aligned you will get the same performance as the aligned-only load.
No reason for the compiler to balk at vectorizing unaligned data these days.
> There's essentially no reason to prefer the aligned load/store commands over the unaligned ones - if the actual pointer is unaligned it will function correctly at half the throughput
Getting a fault instead of half the performance is actually a really good reason to prefer aligned load/store. To be fair, you're talking about a compiler here, but I never understood why people use the unaligned intrinsics...
There are many situations where your data is essentially _majority_ unaligned. Considerable effort by the hardware guys has gone into making that situation work well.
A great example would be a convolution-kernel style code - with AVX512 you are using 64 bytes at a time (a whole cacheline), and sampling a +- N element neighborhood around a pixel. By definition most of those reads will be unaligned!
A lot of other great use cases for SIMD don't let you dictate the buffer alignment. If the code is constrained by bandwidth over compute, I have found it to be worth doing a head/body/tail situation where you do one misaligned iteration before doing the bulk of the work in alignment, but honestly for that to be worth it you have to be working almost completely out of L1 cache which is rare... otherwise you're going to be slowed down to L2 or memory speed anyways, at which point the half rate penalty doesn't really matter.
The early SSE-style instructions often favored making two aligned reads and then extracting your sliding window from that, but there's just no point doing that on modern hardware - it will be slower.
What I get in these article is that the original intent on C language stands true.
Use C as a common platform denominator without crazy optimizations (like tcc). If you need performance, specialize, C gives you the tools to call assembly (or use compiler some intrinsic or even inline assembly).
Complex compiler doing crazy optimizations, in my opinion, is not worth it.
I remember being quite surprised that my implementation which uses manual stack updates is much slower than what compiler had with recursion.
Turns out, I was pushing and popping from stack on every conceptual "recursive call", but compiler figured out it can keep 2-3 recursive levels in registers and pop/push 30% of the time, had more stuff in memory than my version as well.
Even when I reduced memory read/writes to ~50% of the recursive program, kept most of the state in registers, the recursive program was faster anyway due to just using more registers than me.
I realized then that I cannot reason about the microoptimizations at all if I'm coding in a high-level language like C or C++.
Hard to predict the CPU pipeline, sometimes profile guided optimization gets me there faster than my own silliness of assuming I can reason about it.
> Complex compiler doing crazy optimizations, in my opinion, is not worth it.
For these optimisations that are in the back-end, they are used for other languages that can be higher-level or that cannot drop to assembler as easily. C is just one of the front-ends of modern compiler suites.
Assembly isnt how they work under the hood, but its the highest fidelity interface we have. C as "portable assembler" is targeting the surface that chip designers give us and the one that they try to make fast via all their microcode tricks.
>On one hand, Dennard Scaling (aka free lunch) is long gone, hardware landscape is getting increasingly fragmented and specialized out of necessity, software abstractions are getting leakier, forcing developers to be aware of the lowest levels of abstraction, hardware, for good performance.
There are lots of people using Javascript frameworks to build slow desktop and mobile software.
I wonder if the excess CO2 emitted by devices around the world using bloated software that has no need to be so (hullo MS Teams) could be calculated in terms of # of cross atlantic voyages of jets.
After spending more than a decade dancing around all the underlying x86 hidden stuff for low-level optimization, I appreciate CUDA a lot. Everything is there under your total control. No more one size fits all. Higher barrier of entry but no surprises and less time spent debugging to figure out what landmine your code stepped into.
(a) You mention that the NVidia docs push people to use libraries, etc. to really get high performance CUDA kernels, rather than writing them themselves. My argument would be that SIMD is exactly the same - they're something really that are perfect if you're writing a BLAS implementation but are too low level for most developers thinking about a problem to make use of.
(b) You show a problem where autovectorisation fails because of branching, and jump straight to intrinsics as the solution which you basically say are ugly. Looking at the intrinsic code, you're using a mask to deal with the branching. But there's a middle ground - almost always you would want to try restructuring the problem, e.g. splitting up loops and adding masks where there's conditions - i.e. lean into the SIMD paradigm. This would also be the same advice in CUDA.
(c) As you've found, GCC actually performs quite poorly for x86-64 optimisations compared to Intel. It's not always clear cut though, the Intel Compiler for e.g. sacrifices IEEE 764 float precision and go down to ~14 digits of precision in it's defaults, because it sets the flag `-fp-model=fast -fma`. This is true of both the legacy and new Intel compiler. If you switch to `-fp-model=strict` then you may find that the results are closer.
(d) AVX512 is quite hardware specific. Some processors execute these instructions much better than others. It's really a collection of extensions, and you get frequency downclocking that's better/worse on different processors as these instructions are executed.
I just treat autovectorization like I treat every other fancy optimization that's not just constant folding and inlining: nice when it happens to work out, it probably happens to make my code a couple percent faster on average, but I absolutely can't rely on it in any place where I depend on the performance.
I just used to fire up VTune and inspect the hot loops... typically if you care about this you're only really working on hardware targeting the latest instruction sets anyway in my experience. It's only if you're working on low level libraries I would bother doing intrinsics all over the place.
For most consumer software you want to be able to fall back to some lowest-common-denominator hardware anyway otherwise people using it run into issues - same reason that Debian, Conda, etc. only go up to really old instructions sets.
To re-iterate, this is our observation as well. The first AVX512 processors would execute such code quite fast for a short time, then overheat and throttle, leading to a worse wall-time performance than the corresponding AVX256 code.
I am not sure if there is a better way to find the fastest code path besides "measure on the target system", which of course comes with its own challenges.
Only few of those CPUs have been used in workstations, i.e. high-end desktop computers.
The vast majority of the CPUs with AVX-512 that can be encountered at the general population are either AMD Zen 4 and Zen 5 CPUs or some old Intel CPUs from the Tiger Lake, Ice Lake and Rocket Lake families. All these do not have AVX-512 throttling problems.
The owners of server computers are more likely to be knowledgeable about them and choose programs compiled with an appropriate target CPU model.
Therefore I believe that nowadays, when the percentage of computers with good AVX-512 support is increasing, and even Intel is expected to introduce by the end of the year Nova Lake with AVX-512 support, an application should be compiled such that whenever it detects AVX-512 support it should dispatch to the corresponding branch.
On the computers with AVX-512 support, using it can provide a significant increase in performance, while the computers where this could be harmful are more and more unlikely to be encountered outside datacenters that have failed to update their servers.
Skylake Server was introduced 9 years ago and Ice Lake Server, which corrected the behavior, was introduced 6 years ago. Therefore, wherever performance matters, the Skylake Server derivatives would have been replaced by now, as a single Epyc server can replace a cluster of servers with Skylake Server CPUs, at a much lower power consumption and with a higher performance.
Yeah, and it's incredibly frustrating because there is almost zero theory on how to write performant code. Will caching things in memory be faster than re-requesting them over network? Who knows! Sometimes it won't! But you can't predict what those times will be beforehand which turns this whole field into pure black magic instead of anything remotely similar to engineering or science, since theoretical knowledge has no relation to reality.
At my last job we had one of the weirdest "memory access is slo-o-o-ow" scenarios I've ever seen (and it would reproduce pretty reliably... after about 20 hours of the service's continuous execution): somehow, due to peculiarities of the GC and Linux physical memory manager, almost all of the working set of our application would end up in a single physical DDR stick, as opposed to being evenly spread across 4 stick the server actually has. Since a single memory stick literally can't cope with such high data throughput, the performance tanked. And it took quite some time us to figure out what the hell was going on because nothing came up on the perf graphs or metrics or whatever: it's just that almost everything in the the application's userspace became slower. No, the CPU is definitely not throttled, it's actually 20–30% idle. No, there is almost zero disk activity, and the network is fine. Huh?!
If you wish to see some speedups using AVX512, without limiting yourself to C or C++, you might want to try ISPC (https://ispc.github.io/index.html).
You'll get sane aliasing rules from the perspective of performance, multi-target binaries with dynamic dispatching and a lot more control over the code generated.
See for instance https://github.com/ispc/ispc/pull/2160
Not if the data is small and in cache.
> The performant route with AVX-512 would probably include the instruction vpconflictd, but I couldn’t really find any elegant way to use it.
I think the best way to do this is duplicate sum_r and count 16 times, so each pane has a seperate accumulation bucket and there can't be any conflicts. After the loop, you quickly do a sum reduction for each of the 16 buckets.
Isn't it another way of saying what the author says in the previous paragraph, namely that "ideal SIMD speedup can only come from problems that are compute bound"?
If the cost of getting the input data into the cache is already large compared to processing it with the non-vectorized code, then SIMD cannot achieve meaningful speedup. The opposite of this condition (processing is expensive compared to the cost of data into the cache) is basically the definition "compute bound".
The problem is that not all programming languages expose SIMD, and even if they do it is only a portable subset, additionally the kind of skills that are required to be able to use SIMD properly isn't something everyone is confortable doing.
I certainly am not, still managed to get around with MMX and early SSE, can manage shading languages, and that is about it.
See also https://www.numberworld.org/blogs/2024_8_7_zen5_avx512_teard...
Would be interesting to see if auto vec performs better with that addition.
auto aligned_p = std::assume_aligned<16>(p)
Reminded me of way back before OpenGL 2.0, and I was trying to get Vertex Buffer Objects working in my Delphi program using my NVIDIA graphics card. However it kept crashing occasionally, and I just couldn't figure out why.
I've forgotten a lot of the details, but either the exception message didn't make sense or I didn't understand it.
Anyway, after bashing my head for a while I had an epiphany of sorts. NVIDIA liked speed, vertices had to be manipulated before uploading to the GPU, maybe the driver used aligned SIMD instructions and relied on the default alignment of the C memory allocator?
In Delphi the default memory allocator at the time only did 4 byte aligned allocations, and so I searched and found that Microsoft's malloc indeed was default aligned to 16 bytes. However the OpenGL standard and VBO extension didn't say anything about alignment...
Manually aligned the buffers and voila, the crashes stopped. Good times.
No reason for the compiler to balk at vectorizing unaligned data these days.
Getting a fault instead of half the performance is actually a really good reason to prefer aligned load/store. To be fair, you're talking about a compiler here, but I never understood why people use the unaligned intrinsics...
A great example would be a convolution-kernel style code - with AVX512 you are using 64 bytes at a time (a whole cacheline), and sampling a +- N element neighborhood around a pixel. By definition most of those reads will be unaligned!
A lot of other great use cases for SIMD don't let you dictate the buffer alignment. If the code is constrained by bandwidth over compute, I have found it to be worth doing a head/body/tail situation where you do one misaligned iteration before doing the bulk of the work in alignment, but honestly for that to be worth it you have to be working almost completely out of L1 cache which is rare... otherwise you're going to be slowed down to L2 or memory speed anyways, at which point the half rate penalty doesn't really matter.
The early SSE-style instructions often favored making two aligned reads and then extracting your sliding window from that, but there's just no point doing that on modern hardware - it will be slower.
Use C as a common platform denominator without crazy optimizations (like tcc). If you need performance, specialize, C gives you the tools to call assembly (or use compiler some intrinsic or even inline assembly).
Complex compiler doing crazy optimizations, in my opinion, is not worth it.
Turns out, I was pushing and popping from stack on every conceptual "recursive call", but compiler figured out it can keep 2-3 recursive levels in registers and pop/push 30% of the time, had more stuff in memory than my version as well.
Even when I reduced memory read/writes to ~50% of the recursive program, kept most of the state in registers, the recursive program was faster anyway due to just using more registers than me.
I realized then that I cannot reason about the microoptimizations at all if I'm coding in a high-level language like C or C++.
Hard to predict the CPU pipeline, sometimes profile guided optimization gets me there faster than my own silliness of assuming I can reason about it.
For these optimisations that are in the back-end, they are used for other languages that can be higher-level or that cannot drop to assembler as easily. C is just one of the front-ends of modern compiler suites.
There are lots of people using Javascript frameworks to build slow desktop and mobile software.