SPMD Is Not Intel’s Cup of Tea
joshbarczak.com
joshbarczak.com
Now the author is again saying his SPMD hand-coded assembly language can outperform higher level languages by a factor of two for some test cases.
I don't see the economics working out any better for the assembly language solution this time than last time. It won't port across different hardware generations, is hell to write and maintain, and requires super-specialist knowledge.
What this might point to is room for a new high-level language that maps better to the reality of current hardware.
What this might point to is room for a new high-level
language that maps better to the reality of current
hardware.
For me it's something with value types & memory coeherancy sematics. That's what you need to get the most out of your hardware.
Good candidates for that today are Rust, C, C++ and maybe C#(don't have enough experience with their value types to make a judgement yet).
See also maxas versus CUDA in Nervana's Neon library on github. In this case, naively, it's used to make SGEMM operations that are 20-30% faster, but across the entire universe of SGEMM inputs, it can be 3-6x faster. Whether this is because of the use of assembler or because Scott Gray is a better dense matrix programmer than the current cuBLAS team is an open question because cuBLAS is closed source.
Finally, you've ignored the 800 lb gorilla in the room: Intel is making processors that look more and more like GPUs in terms of cores, vector units, and SIMD width whilst investing in FPGAs through the purchase of Altera. But what they're not making is a consistent paradigm for programming them. They could embrace an open standard like OpenCL. They could give NVIDIA a flaming middle finger and build a CUDA compiler like AMD did (Rhythm). Or they could make one of their proprietary languages like Intel TBB (https://www.threadingbuildingblocks.org/) or ISPC (https://ispc.github.io/) a first-class citizen and throw their massive support engine behind it. But instead, they seem to throw these APIs over the fence and to continue to rely on "recompile and run" in the hopes that optimizing compilers can magically parallelize otherwise serial code.
Which is why I continue to stick with CUDA/NVIDIA...
For games, it might be about your game running at 30 fps or 60 fps.
If your future server farm is projected to cost $20M, wouldn't you want to try something like this to save a few millions?
> I don't see the economics working out any better for the assembly language solution this time than last time. It won't port across different hardware generations, is hell to write and maintain, and requires super-specialist knowledge.
Sometimes that's the price you have to pay. Such is life.
It would be better if hardware manufacturers designed their CPUs around high-level languages, not the other way around. Burroughs did with their mainframes, which ran Algol 60 natively and also had instructions for efficient COBOL execution. Later, Lisp machine manufacturers made machines specifically designed to run Lisp.
When hardware is designed without reference to the software which later runs on it, never mind the problems the software is designed to solve, it often leads to solutions tailored better for the problem, but which don't fit the hardware particularly well, not being attempted.
Many if not most problems involve processing different data, usually non-numeric, in different ways. MIMD hardware would be better for that.
Unless you have explicit value types and memory coherancy guarantees then you'll always see this tension.
It's not like hardware vendors set out to build slow chips or don't know about high-level languages, it's just that hardware doesn't work that way.
(1) Hardware designers need to do much more to speed up general memory accesses, so that a CPU cache is no longer needed.
(2) Programmers need to accept that if they use pointers (either explicitly like in C, or implicitly like in Lisp or Java), they will have to tolerate the performance hit.
(3) Programming is made much more difficult by requiring programmers to take into account memory layout, which is almost like a return to the bad old days when you had (by today's standards) tiny memories.
(4) Memory is distributed over multiple CPUs, each working on a different part of the problem.
There aren't many problems where users (who today, are almost everybody) care about raw speed. What they do care about are response times.
Split application into logic(branches, special-case, and all the gnarly stuff that CPUs don't like) and data processing(nice structured data that fits well into memory). This is essentially what Sony did on the PS3, and made for a very fast system. They just did a poor job of messaging it ahead to devs who were familiar with n-core systems.
That way you can speed up the things that need to be fast and the slow stuff doesn't interfere with it. We aren't going to see any fundamental shifts in memory tech so I think this is the best mitigating solution.
Relational Databases already kinda nail the structured part to a degree, I think where we fall over is mixing logic with the data we want to process.
There's also a continuum of prototyping -> high performance that gets missed in some of these discussions. If I want to see if something works I can knock out a quick prototype in JS but if I want to ship a performant, solid product I'm going to use Rust/C++/etc.
Sure with CPU's you have variety but less so and certainly more dynamic changes over time with SPMD type area's with GPU architectures still changing more generation to generation compared to CPU architectures.
Given that perhaps this highlights that area as an opportunity.
But I suspect the whole area of such a dynamic area with GPU's from 5 years ago at a hardware level being much more different than CPU's with changes that any investment in low-level optimization will be cost wise higher in many cases than just having portability and the ability to easily run on the next generation and gain accordingly.
With this in mind the number of architectures and associated hardware changes more often than the core CPU's and it is perhaps that alone that curtails such optimization process's being able to mature and come into play. Though once things settle down more.
Though I'm mindful this is in many ways the old Betamax VS VHS and from that we learned that it is not always the best technology that progresses with the masses, but the ease of use/cost of attaining and running it. In short perfection rarely wins out in the face of just good enough for a lot less effort in so many cases.
You will always be more efficient if you can tailor your code to a particular device, a standard abstraction will never let you get to that level.
What you need to ask is: 1) Do you really need the extra 90% effort for the 10% extra performance? 2) are the tools good enough? are the abstractions wrong?
This article seems to be saying that the abstractions are wrong... how do these techniques fare on other devices?
From the article:
"SPMD has a serious advantage in that it’s able to target AMD’s massive 64-wide SIMDs just as easily as Intel’s itty-bitty 8-wide ones, at least for simple usage patterns like do this operation on 1 million pixels. For simple stream processing applications, SPMD makes perfect sense, and it’s relatively easy to write one source code that will run close to optimal on all the architectures."> These are not rock-solid results. I admit I’ve cut a lot of corners and left some room for doubt, but it looks like significant speedups are possible by using explicit thread-level programming. There are optimizations that can be done at this level which are simply not possible using a conventional SPMD model.
Yeah. This is the issue with abstracting using OpenCL, CUDA, etc.
If you need absolute performance, sometimes explicit control, even if it means using SIMD intrinsics, is better than a compiler that can't understand what you are really trying to do. Of course the other side of the coins is it's more work and less portable to use explicit width intrinsics.
For example, compilers don't understand memory access patterns and it's hard to share work between execution lanes (work units).
Think of an algorithm that processes images in rectangular blocks. Each block needs also rightmost column of pixels from the block on the left. I know swizzling performed by the cache/memory controller would make this example less worse, but humor me. I'm trying to make a generic point, not just about spatial image processing.
If the order of execution is suitable, the code that computes first block could set the registers ready to transfer rightmost column of pixels for the next block on right.
So you knowing and controlling the processing order could get this single column of pixels from left of block from registers (set by previous processing of left side block), otherwise just issue one memory read for current block.
Instead of just reading both blocks again, which is what a compiler would do. Sure, it would likely come from cache. But it means more cache pressure, 95% of those pixels might not be needed anymore -- just that right edge of the left side block. In other words, in the worst case you need almost twice as much data in cache than otherwise.
> Example 1: Block Min/Max
> My first test case is taking a scalar image and computing the min and max over 4×4 tiles. This is the first half of a BC4 compressor. ...
I'm surprised first example running time dropped just from 1.7M to 1.15M. Maybe the code is cache/memory access (or some other resource, not familiar with Intel's GPUs) limited.
This might be a good place to try random evolution (genetic algorithms, monte carlo, etc.) in instruction and memory access scheduling, variations, etc. I suspect hardware level details about Intel's GPUs are not public. Some instruction stream orderings might just perform faster.
We have better compilers than that already. If you have a static data flow, there's no reason your compiler can not take it into account when optimizing your code.
Optimizing a dynamic data flow is the hard one (still possible, but not done in practice), but people suck at this too.
That's right, but this is not about static data flow.
This is about choosing a particular strategy for execution and caching a small amount of data in registers to maximize usefulness of cache by having less "hot memory". This reduces expensive memory traffic, which is very often the bottleneck.
More generally, you can often do significant part of the work in other lane, but can't communicate this efficiently to somewhere else, because it might be executing simultaneously. So you have to for example access memory unnecessarily or do extra computation for little benefit. When you can control execution order, this is not an issue anymore.
I know how amazing compilers are nowadays, but I have never seen any of them optimizing this type of case.