How does Nvidia's Fermi GPU issue threadblocks to streaming multiprocessor - cuda

Assume I have 8 threadblocks and my GPU has 8 SMs. Then how does GPU issue this threadblocks to the SMs?
I found some programs or articles suggest a breadth-first manner, that is , each SM runs a threadblock in this example.
However, according to a few documents, increasing occupancy may be a good idea if GPU kernels are latency-limited. It might be inferred that 8 threadblocks will run on 4 or less SMs if it can.
I wonder which one is the reality.
Thanks in advance.

It's hard to tell what the GPU is doing exactly. If you have a specific kernel you're interested in, you could try reading and storing the %smid register for each block.
An example of how to do this is given here.

You ask the wrong question: you shouldn't worry about how hardware allocates thread-blocks to SMs. That's GPU's responsibility. In fact, since their programming model makes no assumptions as for which blocks will run on which SMs, you get scalability across a pool of computing devices/future generations.
Instead, you should try to feed GPU with the optimal number of thread-blocks. That's non-trivial, since it's subject to many restrictions

Related

Maximum number of GPU Threads on Hardware and used memory

I have already read several threads about the capacity about the GPU and understood that the concept of blocks and threads has to be seperated from the physical Hardware. Although the maximum amount of threads per block is 1024, there is no limit on the number of blocks one can use. However, as the number of streaming processors is finite, there has to be a physical limit. After I wrote a GPU program, I would be interested in evaluating the used capacity of my GPU. To do this, I have to know how many threads I could start theoretically at one time on hardware. My graphics card is a Nvidia Geforce 1080Ti, so I have 3584 CUDA-Cores. As far as I understood, each Cuda core executes one Thread, so in theory, I would be able to execute 3584 threads per cycle. Is this correct?
Another question is the one about memory. I installed and used nvprof to get some insight into the used kernels. What is displayed there is for example the number of used registers. I transfer my arrays to the GPU using cuda.to_device (in Python Numba) and as far as I understood, the arrays then reside in global memory. How do I find out how big this global memory is? Is it equivalent to the DRAM size?
Thanks in advance
I'll focus on the first part of the question. The second should really be its own separate question.
CUDA cores do not map 1-to-1 to threads. They are more like ports in a multiscalar CPU. Multiple threads can issue instructions to the same CUDA core in different clock cycles. Sort of like hyperthreading in a CPU.
You can see the relation and numbers in the documentation in chapter K Compute Capabilities compared to the table Table 3. Throughput of Native Arithmetic Instructions. Depending on your architecture, you may have for example for your card (compute capability 6.1) 2048 threads per SM and 128 32 bit floating point operations per clock cycle. That means you have 128 CUDA cores shared by a maximum of 2048 threads.
Within one GPU generation, the absolute number of threads and CUDA cores only scales with the number of multiprocessors (SMs). TechPowerup's excellent GPU database documents 28 SMs for your card which should give you 28 * 2048 threads unless I did something wrong.

threadblocks and stream multiprocessors in C

I was told that in order to have a fully efficient CUDA C program, the number of threadblocks should be at least 3 or 4 times the number of stream multiprocessors.
My questions is: is the statement true? If yes/no why? what should the ratio ideally be?
It's generally a good idea to have multiple threadblocks that can be launched on each SM, when a kernel launches.
The GPU is a latency-hiding machine. In order to hide latency, it needs as much potential work as possible. Potential work can be translated into "warps that are ready to execute". This scenario can be maximized by having more than one threadblock per SM.
At some point, the GPU (SMs) run out of resources to host additional threadblocks. This running-out point might occur at around 3-4 threadblocks per SM, depending on the specifics of resource usage (registers, threads, shared memory, etc.) of the threadblocks, and the GPU type. Therefore, launching more than the amount that can be actually scheduled on the SMs won't help with concurrency, latency-hiding, occupancy, or other figures of merit for a parallel program. Those threadblocks will just wait until scheduling slots open up on the SMs.
There is no fixed ratio, but an analysis of typical threadblocks with 256 or 512 threads per block suggest that you will want at least 3-8 threadblocks per SM, to maximize occupancy (this varies based on GPU architecture as well). With 1024 threads per block, it might only require 2 threadblocks per SM.
GPU programs typically don't dramatically slow down if work is partitioned into more threadblocks, so the numbers are not a hard-and-fast rule, and the actual behavior will depend on other factors like shared memory usage (if any). It's just a general guideline.

Is it possible to set a limit on the number of cores to be used in Cuda programming for a given code?

Assume I have Nvidia K40, and for some reason, I want my code only uses portion of the Cuda cores(i.e instead of using all 2880 only use 400 cores for examples), is it possible?is it logical to do this either?
In addition, is there any way to see how many cores are being using by GPU when I run my code? In other words, can we check during execution, how many cores are being used by the code, report likes "task manger" in Windows or top in Linux?
It is possible, but the concept in a way goes against fundamental best practices for cuda. Not to say it couldn't be useful for something. For example if you want to run multiple kernels on the same GPU and for some reason want to allocate some number of Streaming Multiprocessors to each kernel. Maybe this could be beneficial for L1 caching of a kernel that does not have perfect memory access patterns (I still think for 99% of cases manual shared memory methods would be better).
How you could do this, would be to access the ptx identifiers %nsmid and %smid and put a conditional on the original launching of the kernels. You would have to only have 1 block per Streaming Multiprocessor (SM) and then return each kernel based on which kernel you want on which SM's.
I would warn that this method should be reserved for very experienced cuda programmers, and only done as a last resort for performance. Also, as mentioned in my comment, I remember reading that a threadblock could migrate from one SM to another, so behavior would have to be measured before implementation and could be hardware and cuda version dependent. However, since you asked and since I do believe it is possible (though not recommended), here are some resources to accomplish what you mention.
PTS register for SM index and number of SMs...
http://docs.nvidia.com/cuda/parallel-thread-execution/#identifiers
and how to use it in a cuda kernel without writing ptx directly...
https://gist.github.com/allanmac/4751080
Not sure, whether it works with the K40, but for newer Ampere GPUs there is the MIG Multi-Instance-GPU feature to partition GPUs.
https://docs.nvidia.com/datacenter/tesla/mig-user-guide/
I don't know such methods, but would like to get to know.
As to question 2, I suppose sometimes this can be useful. When you have complicated execution graphs, many kernels, some of which can be executed in parallel, you want to load GPU fully, most effectively. But it seems on its own GPU can occupy all SMs with single blocks of one kernel. I.e. if you have a kernel with 30-blocks grid and 30 SMs, this kernel can occupy entire GPU. I believe I saw such effect. Really this kernel will be faster (maybe 1.5x against 4 256-threads blocks per SM), but this will not be effective when you have another work.
GPU can't know whether we are going to run another kernel after this one with 30 blocks or not - whether it will be more effective to spread it onto all SMs or not. So some manual way to say this should exist
As to question 3, I suppose GPU profiling tools should show this, Visual Profiler and newer Parallel Nsight and Nsight Compute. But I didn't try. This will not be Task manager, but a statistics for kernels that were executed by your program instead.
As to possibility to move thread blocks between SMs when necessary,
#ChristianSarofeen, I can't find mentions that this is possible. Quite the countrary,
Each CUDA block is executed by one streaming multiprocessor (SM) and
cannot be migrated to other SMs in GPU (except during preemption,
debugging, or CUDA dynamic parallelism).
https://developer.nvidia.com/blog/cuda-refresher-cuda-programming-model/
Although starting from some architecture there is such thing as preemption. As I remember NVidia advertised it in the following way. Let's say you made a game that run some heavy kernels (say for graphics rendering). And then something unusual happened. You need to execute some not so heavy kernel as fast as possible. With preemption you can unload somehow running kernels and execute this high priority one. This increases execution time (of this high pr. kernel) a lot.
I also found such thing:
CUDA Graphs present a new model for work submission in CUDA. A graph
is a series of operations, such as kernel launches, connected by
dependencies, which is defined separately from its execution. This
allows a graph to be defined once and then launched repeatedly.
Separating out the definition of a graph from its execution enables a
number of optimizations: first, CPU launch costs are reduced compared
to streams, because much of the setup is done in advance; second,
presenting the whole workflow to CUDA enables optimizations which
might not be possible with the piecewise work submission mechanism of
streams.
https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#cuda-graphs
I do not believe kernels invocation take a lot of time (of course in case of a stream of kernels and if you don't await for results in between). If you call several kernels, it seems possible to send all necessary data for all kernels while the first kernel is executing on GPU. So I believe NVidia means that it runs several kernels in parallel and perform some smart load-balancing between SMs.

Role of Warps in NVIDIA GPU architecture with OpenCL

I'm studying OpenCL concepts as well as the CUDA architecture for a small project, and there is one thing that is unclear to me: the necessity for Warps.
I know a lot of questions have been asked on this subject, however after having read some articles i still don't get the "meaning" of warps.
As far as I understand (speaking for my GPU card which is a Tesla, but i guess this easily translates to other boards):
A work-item is linked to a CUDA thread, which several of them can be executed by a Streaming Processor (SP). BTW, does a SP treats those WI in parallel?
Work-items are grouped into Work-groups. Work-groups operate on a Stream Multiprocessor and can not migrate. However, work-items in a work-group can collaborate via shared memory (a.k.a local memory). One or more work-groups may be executed by a Stream MultiProcessor. BTW, does a SM treats those WG in parallel?
Work-item are executed in parallel inside a work-group. However, synchronization is NOT guaranteed, that's why you need concurrent programming primitives, such as barriers.
As far as I understand, all of this is rather a logical view than a 'physical', hardware perspective.
If all of the above is correct, can you help me on the following. Is that true to say that:
1 - Warps execute 32 threads or work-items simultaneously. Thus, they will 'consume' parts of a work-group. And that's why in the end you need stuff like memory fences to synchronize work-items in work groups.
2 - The Warp scheduler allocates the registers for the 32 threads of warp when it becomes active.
3 - Also, are executed thread in a warp synchronized at all?
Thanks for any input on Warps, and especially why they are necessary in the CUDA architecture.
My best analogon is that a Warp is the vector that be processed in parallel, not unlike an AVX or SSE vector with an Intel CPU. This makes an SM a 32-length vector processor.
Then, to your questions:
Yes, all 32 elements will be run in parallel. Note that also a GPU puts hyperthreading to the extreme: a workgroup will consist of multiple Warps, which all are run more-or-less in parallel. You will need memory fences to sychronise that all.
Yes, typically all 32 work elements (CUDA: thread) in a Warp will work in parallel. Note that you typically will have multiple regsters per work element.
Not guaranteed, AFAIK.

Is it fair to compare SSE/AVX units to GPU cores?

I have a presentation to make to people who have (almost) no clue of how a GPU works. I think saying that a GPU has a thousand cores where a CPU only has four to eight of them is a non-sense. But I want to give my audience an element of comparison.
After a few months working with NVidia's Kepler and AMD's GCN architectures, I'm tempted to compare a GPU "core" to a CPU's SIMD ALU (I don't know if they have a name for that at Intel). Is it fair ? After all, when looking at an assembly level, those programming models have much in common (at least with GCN, take a look at p2-6 of the ISA manual).
This article states that an Haswell processor can do 32 single-precision operations per cycle, but I suppose there is pipelining or other things happening to achieve that rate. In NVidia parlance, how many Cuda-cores does this processor have ? I would say 8 per CPU-core for 32 bits operations, but this is just a guess based on the SIMD width.
Of course there is many other things to take into account when comparing CPU and GPU hardware, but this is not what I'm trying to do. I just have to explain how the thing is working.
PS: All pointers to CPU hardware documentations or CPU/GPU presentations are greatly appreciated !
EDIT:
Thanks for your answers, sadly I had to chose only one of them. I marked Igor's answer because it sticks the most to my initial question and gave me enough informations to justify why this comparison shouldn't be taken too far, but CaptainObvious provided very good articles.
I'd be very caution on making this kind of comparison. After all even in the GPU world the term "core" depending on the context has really different capability: the new AMD GCN is quite different from the old VLIW4 one which itself is quite different from the CUDA core one.
Besides that, you will bring more puzzlement than understanding to your audience if you make just one small comparison with CPU and that's it. If I were you I'd still go for a more detailed (can still be quick) comparison. For instance someone used to CPU and with little knowledge of GPU, might wonder how come a GPU can have so many registers though it's so expensive (in the CPU world). An explanation to that question is given at the end of this post as well as some more comparison GPU vs CPU.
This other article gives a nice comparison between these two kind of processing units by explaining how GPUs work but also how they evolved and showing the differences with CPUs. It addresses topics like data flow, memory hierarchy but also for what kind of applications a GPU is useful. After all the power a GPU can developed is accessible (efficiently) only for some types of problems.
And personally, If I had to make a presentation about GPU and had the possibility to make only one reference to CPU it would be this: presenting the problems a GPU can solve efficiently vs those a CPU can handle better.
As a bonus even though it's not related directly to your presentation here is an article that put GPGPU in perspective, showing that some speedup claimed by some people are overrated (this is linked to my last point btw :))
Very loosely speaking, it is not entirely unreasonable to say that a Haswell core has about 16 CUDA cores, but you definitely don't want to take that comparison too far. You may want to be cautious about making that statement directly in a presentation, but I've found it to be useful to think of a CUDA core as being somewhat related to a scalar FP unit.
It may help if I explain why Haswell can perform 32 single-precision operations per cycle.
8 single-precision operations execute in each AVX/AVX2 instruction. When writing code that will run on a Haswell CPU, you can use AVX and AVX2 instructions which operate on 256-bit vectors. These 256-bit vectors can represent 8 single-precision FP numbers, 8 integers (32-bit) or 4 double-precision FP numbers.
2 AVX/AVX2 instructions can execute in each core per cycle, although there are some restrictions on which instructions can be paired up.
A fused multiply add (FMA) instruction technically performs 2 single-precision operations. FMA instructions perform "fused" operations such as A = A * B + C, so there are arguably two operations per scalar operand: a multiplication and an addition.
This article explains the above points in more detail: http://www.realworldtech.com/haswell-cpu/4/
In the total accounting, a Haswell core can perform 8 * 2 * 2 single-precision operations per cycle. Since CUDA cores support FMA operations as well, you cannot count that factor of 2 when comparing CUDA cores to Haswell cores.
A Kepler CUDA core has one single-precision floating-point unit, so it can perform one floating-point operation per cycle: http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf, http://www.realworldtech.com/kepler-brief/
If I was putting together slides on this, I would have one section explaining how many FP operations Haswell can do per cycle: the three points above, plus you have multiple cores and possibly multiple processors. And, I'd have another section explaining how many FP operations a Kepler GPU can do per cycle: 192 per SMX, and you have multiple SMX units on the GPU.
PS.: I may be stating the obvious, but just to avoid confusion: the Haswell architecture also includes an integrated GPU, which has an altogether different architecture from the Haswell CPU.
I completely agree with CaptainObvious, especially that presenting the problems a GPU can solve efficiently vs those a CPU can handle better would be a good idea.
One way I like to compare CPUs and GPUs is by the number of operation/sec that they can perorm. But of course don't compare one cpu core to a multi-core gpu.
A SandyBridge core can perform 2 AVX op/cycles, that is crunch 8 double precision numbers/cycle. Hence, a computer with 16 Sandy-Bridge cores clocked at 2.6 GHz has a peak power of 333 Gflops.
A K20 computing module GK110 has a peak of 1170 Gflops, that is 3.5 times more. This is a fair comparaison in my opinion, and it should be emphasized that the peak performance is much easier to reach on CPU (some applications reach 80%-90% of peak) than on GPU (best cases I know are less than 50% of peak).
So to summerize, I would not go into architecture details, but rather state some shear numbers with the perspective that the peak is often far from reach on GPUs.
It's more fair to compare GPU to vectorized CPU units however if your audience has zero idea of how GPUs work, it seems fair to assume that they have a similar knowledge of vectorized SSE instructions.
For audiences such as these it's important to point out the high level differences, like how blocks of "cores" on the gpu share a scheduler and register file.
I would refer to the GTC Kepler architecture overview for a better idea of what the Kepler architecture looks like.
This is also a reasonably graspable comparison between the two if you want to stick to the "gpu core" idea.