I'm new to the CUDA paradigm. My question is in determining the number of threads per block, and blocks per grid. Does a bit of art and trial play into this? What I've found is that many examples have seemingly arbitrary number chosen for these things.
I'm considering a problem where I would be able to pass matrices - of any size - to a method for multiplication. So that, each element of C (as in C = A * B) would be calculated by a single thread. How would you determine the threads/block, blocks/grid in this case?
In general you want to size your blocks/grid to match your data and simultaneously maximize occupancy, that is, how many threads are active at one time. The major factors influencing occupancy are shared memory usage, register usage, and thread block size.
A CUDA enabled GPU has its processing capability split up into SMs (streaming multiprocessors), and the number of SMs depends on the actual card, but here we'll focus on a single SM for simplicity (they all behave the same). Each SM has a finite number of 32 bit registers, shared memory, a maximum number of active blocks, AND a maximum number of active threads. These numbers depend on the CC (compute capability) of your GPU and can be found in the middle of the Wikipedia article http://en.wikipedia.org/wiki/CUDA.
First of all, your thread block size should always be a multiple of 32, because kernels issue instructions in warps (32 threads). For example, if you have a block size of 50 threads, the GPU will still issue commands to 64 threads and you'd just be wasting them.
Second, before worrying about shared memory and registers, try to size your blocks based on the maximum numbers of threads and blocks that correspond to the compute capability of your card. Sometimes there are multiple ways to do this... for example, a CC 3.0 card each SM can have 16 active blocks and 2048 active threads. This means if you have 128 threads per block, you could fit 16 blocks in your SM before hitting the 2048 thread limit. If you use 256 threads, you can only fit 8, but you're still using all of the available threads and will still have full occupancy. However using 64 threads per block will only use 1024 threads when the 16 block limit is hit, so only 50% occupancy. If shared memory and register usage is not a bottleneck, this should be your main concern (other than your data dimensions).
On the topic of your grid... the blocks in your grid are spread out over the SMs to start, and then the remaining blocks are placed into a pipeline. Blocks are moved into the SMs for processing as soon as there are enough resources in that SM to take the block. In other words, as blocks complete in an SM, new ones are moved in. You could make the argument that having smaller blocks (128 instead of 256 in the previous example) may complete faster since a particularly slow block will hog fewer resources, but this is very much dependent on the code.
Regarding registers and shared memory, look at that next, as it may be limiting your occupancy. Shared memory is finite for a whole SM, so try to use it in an amount that allows as many blocks as possible to still fit on an SM. The same goes for register use. Again, these numbers depend on compute capability and can be found tabulated on the wikipedia page.
https://docs.nvidia.com/cuda/cuda-occupancy-calculator/index.html
The CUDA Occupancy Calculator allows you to compute the multiprocessor occupancy of a GPU by a given CUDA kernel. The multiprocessor occupancy is the ratio of active warps to the maximum number of warps supported on a multiprocessor of the GPU. Each multiprocessor on the device has a set of N registers available for use by CUDA program threads. These registers are a shared resource that are allocated among the thread blocks executing on a multiprocessor. The CUDA compiler attempts to minimize register usage to maximize the number of thread blocks that can be active in the machine simultaneously. If a program tries to launch a kernel for which the registers used per thread times the thread block size is greater than N, the launch will fail...
With rare exceptions, you should use a constant number of threads per block. The number of blocks per grid is then determined by the problem size, such as the matrix dimensions in the case of matrix multiplication.
Choosing the number of threads per block is very complicated. Most CUDA algorithms admit a large range of possibilities, and the choice is based on what makes the kernel run most efficiently. It is almost always a multiple of 32, and at least 64, because of how the thread scheduling hardware works. A good choice for a first attempt is 128 or 256.
You also need to consider shared memory because threads in the same block can access the same shared memory. If you're designing something that requires a lot of shared memory, then more threads-per-block might be advantageous.
For example, in terms of context switching, any multiple of 32 works just the same. So for the 1D case, launching 1 block with 64 threads or 2 blocks with 32 threads each makes no difference for global memory accesses. However, if the problem at hand naturally decomposes into 1 length-64 vector, then the first option will be better (less memory overhead, every thread can access the same shared memory) than the second.
There is no silver bullet. The best number of threads per block depends a lot on the characteristics of the specific application being parallelized. CUDA's design guide recommends using a small amount of threads per block when a function offloaded to the GPU has several barriers, however, there are experiments showing that for some applications a small number of threads per block increases the overhead of synchronizations, imposing a larger overhead. In contrast, a larger number of threads per block may decrease the amount of synchronizations and improve the overall performance.
For an in-depth discussion (too lengthy for StackOverflow) about the impact of the number of threads per block on CUDA kernels, check this journal article, it shows tests of different configurations of the number of threads per block in the NPB (NAS Parallel Benchmarks) suite, a set of CFD (Computational Fluid Dynamics) applications.
Related
I am trying to understand the basic architecture of a GPU. I have gone through a lot of material including this very good SO answer. But I am still confused not able to get a good picture of it.
My Understanding:
A GPU contains two or more Streaming Multiprocessors (SM) depending upon the compute capablity value.
Each SM consists of Streaming Processors (SP) which are actually responisible for the execution of instructions.
Each block is processed by SP in form of warps (32 threads).
Each block has access to a shared memory. A different block cannot access the data of some other block's shared memory.
Confusion:
In the following image, I am not able to understand which one is the Streaming Multiprocessor (SM) and which one is SP. I think that Multiprocessor-1 respresent a single SM and Processor-1 (upto M) respresent a single SP. But I am not sure about this because I can see that each Processor (in blue color) has been provided a Register but as far as I know, a register is provided to a thread unit.
It would be very helpful to me if you could provide some basic overview w.r.t this image or any other image.
First, some comments on the "My understanding" portion of the question:
The number of SMs depends on GPU model - there are low-end models with just one SM, and high-end ones with as many as 30! Compute capability defines what those SMs are capable of, but not how many SMs there are in a GPU.
Each thread block is assigned to an SM, not SP. There can be multiple thread blocks running on a given SM, subject to its resource limitations.
On to the diagram:
Orange boxes are indeed SMs, just as they are labeled. Each SM has shared memory pool, divided between all thread blocks running on this SM.
Blue boxes are SPs. Since SP is a scalar lane, it runs one thread, and each thread is provided with its own set of registers, again, just like the diagram shows.
Addressing the follow-up question:
Each SM can have multiple resident thread blocks. The maximum number of thread blocks resident on SM is determined by compute capability. Achieved number can be lower than maximum when it is limited by the number of registers or the amount of shared memory consumed by each thread block.
SM will then schedule instruction from all warps resident on it, picking among warps that have instructions ready for execution - and those warps may come from any thread block resident on this SM. You generally want to have many warps resident, so that at any given moment of time SPs can be kept busy running instructions from whatever warps are ready.
Number of cores per SM is not a very useful metric, and you need not think too much about it at this point.
In CUDA, can a SM run multiple blocks simultaneously if each block won't cost too much resource.
On Fermi, we know that a SM consists of 32kb register space for use. suppose a thread use 32 register, so this SM can lanuch one block which contains 256 ((32*1024)/(32*4)) threads. If SM can run multiple blocks simultaneously, we can also configure 32 theards for a block, and 8 block for the SM. Is there any difference?
As #talonmies commented, your math is not entirely correct. But the key point is that an SM contains a balance of many different types of resources. The better your kernel and kernel launch parameters fit with this balance, the better your performance.
I haven't checked the numbers for Kepler (compute capability 3.x) but for Fermi (2.x), an SM can keep track of 48 concurrent warps (1,536 threads) and 8 concurrent blocks. This means that if you chose a low thread count for your blocks, the 8 concurrent blocks becomes the limiting factor to occupancy in your kernel. For instance, if you chose 32 threads per block, you get up to 256 (8 * 32) concurrent threads running on the SM while the SM can run up to 1,536 threads (48 * 32).
In the occupancy calculator, you can see what the different hardware limits are and it will tell you which of them becomes the limiting factor with your specific kernel. You can experiment with variations in launch parameters, shared memory usage and register usage to see how they affect your occupancy.
Occupancy is not everything when it comes to performance. Increased occupancy translates to increased ability to hide the latency of memory transfers. When the memory bandwidth is saturated, increasing occupancy further does not help. There is another effect in play as well. Increasing the size of a block may decrease occupancy but at the same time increase the amount of instruction level parallelism (ILP) available in your kernel. In this case, decreasing occupancy can increase performance.
Ok I know that related questions have been asked over and over again and I read pretty much everything I found about this, but things are still unclear. Probably also because I found and read things contradicting each other (maybe because, being from different times, they referred to devices with different compute capability, between which there seems to be quite a gap). I am looking to be more efficient, to reduce my execution time and thus I need to know exactly how many threads/warps/blocks can run at once in parallel. Also I was thinking of generalizing this and calculating an optimal number of threads and blocks to pass to my kernel based only on the number of operations I know I have to do (for simpler programs) and the system specs.
I have a GTX 550Ti, btw with compute capability 2.1.
4 SMs x 48 cores = 192 CUDA cores.
Ok so what's unclear to me is:
Can more than 1 block run AT ONCE (in parallel) on a multiprocessor (SM)? I read that up to 8 blocks can be assigned to a SM, but nothing as to how they're ran. From the fact that my max number of threads per SM (1536) is barely larger than my max number of threads per block (1024) I would think that blocks aren't ran in parallel (maybe 1 and a half?). Or at least not if I have a max number of threads on them. Also if I set the number of blocks to, let's say 4 (my number of SMs), will they be sent to a different SM each?
Or I can't really control how all this is distributed on the hardware and then this is a moot point, my execution time will vary based on the whims of my device ...
Secondly, I know that a block will divide it's threads into groups of 32 threads that run in parallel, called warps. Now these warps (presuming they have no relation to each other) can be ran in parallel aswell? Because in the Fermi architecture it states that 2 warps are executed concurrently, sending one instruction from each warp to a group of 16 (?) cores, while somewhere else i read that each core handles a warp, which would explain the 1536 max threads (32*48) but seems a bit much. Can 1 CUDA core handle 32 threads concurrently?
On a simpler note, what I'm asking is: (for ex) if I want to sum 2 vectors in a third one, what length should I give them (nr of operations) and how should I split them in blocks and threads for my device to work concurrently (in parallel) at full capacity (without having idle cores or SMs).
I'm sorry if this was asked before and I didn't get it or didn't see it. Hope you can help me. Thank you!
The distribution and parallel execution of work are determined by the launch configuration and the device. The launch configuration states the grid dimensions, block dimensions, registers per thread, and shared memory per block. Based upon this information and the device you can determine the number of blocks and warps that can execute on the device concurrently. When developing a kernel you usually look at the ratio of warps that can be active on the SM to the maximum number of warps per SM for the device. This is called the theoretical occupancy. The CUDA Occupancy Calculator can be used to investigate different launch configurations.
When a grid is launched the compute work distributor will rasterize the grid and distribute thread blocks to SMs and SM resources will be allocated for the thread block. Multiple thread blocks can execute simultaneously on the SM if the SM has sufficient resources.
In order to launch a warp, the SM assigns the warp to a warp scheduler and allocates registers for the warp. At this point the warp is considered an active warp.
Each warp scheduler manages a set of warps (24 on Fermi, 16 on Kepler). Warps that are not stalled are called eligible warps. On each cycle the warp scheduler picks an eligible warp and issue instruction(s) for the warp to execution units such as int/fp units, double precision floating point units, special function units, branch resolution units, and load store units. The execution units are pipelined allowing many warps to have 1 or more instructions in flight each cycle. Warps can be stalled on instruction fetch, data dependencies, execution dependencies, barriers, etc.
Each kernel has a different optimal launch configuration. Tools such as Nsight Visual Studio Edition and the NVIDIA Visual Profiler can help you tune your launch configuration. I recommend that you try to write your code in a flexible manner so you can try multiple launch configurations. I would start by using a configuration that gives you at least 50% occupancy then try increasing and decreasing the occupancy.
Answers to each Question
Q: Can more than 1 block run AT ONCE (in parallel) on a multiprocessor (SM)?
Yes, the maximum number is based upon the compute capability of the device. See Tabe 10. Technical Specifications per Compute Capability : Maximum number of residents blocks per multiprocessor to determine the value. In general the launch configuration limits the run time value. See the occupancy calculator or one of the NVIDIA analysis tools for more details.
Q:From the fact that my max number of threads per SM (1536) is barely larger than my max number of threads per block (1024) I would think that blocks aren't ran in parallel (maybe 1 and a half?).
The launch configuration determines the number of blocks per SM. The ratio of maximum threads per block to maximum threads per SM is set to allow developer more flexibility in how they partition work.
Q: If I set the number of blocks to, let's say 4 (my number of SMs), will they be sent to a different SM each? Or I can't really control how all this is distributed on the hardware and then this is a moot point, my execution time will vary based on the whims of my device ...
You have limited control of work distribution. You can artificially control this by limiting occupancy by allocating more shared memory but this is an advanced optimization.
Q: Secondly, I know that a block will divide it's threads into groups of 32 threads that run in parallel, called warps. Now these warps (presuming they have no relation to each other) can be ran in parallel as well?
Yes, warps can run in parallel.
Q: Because in the Fermi architecture it states that 2 warps are executed concurrently
Each Fermi SM has 2 warps schedulers. Each warp scheduler can dispatch instruction(s) for 1 warp each cycle. Instruction execution is pipelined so many warps can have 1 or more instructions in flight every cycle.
Q: Sending one instruction from each warp to a group of 16 (?) cores, while somewhere else i read that each core handles a warp, which would explain the 1536 max threads (32x48) but seems a bit much. Can 1 CUDA core handle 32 threads concurrently?
Yes. CUDA cores is the number of integer and floating point execution units. The SM has other types of execution units which I listed above. The GTX550 is a CC 2.1 device. On each cycle a SM has the potential to dispatch at most 4 instructions (128 threads) per cycle. Depending on the definition of execution the total threads in flight per cycle can range from many hundreds to many thousands.
I am looking to be more efficient, to reduce my execution time and thus I need to know exactly how many threads/warps/blocks can run at once in parallel.
In short, the number of threads/warps/blocks that can run concurrently depends on several factors. The CUDA C Best Practices Guide has a writeup on Execution Configuration Optimizations that explains these factors and provides some tips for reasoning about how to shape your application.
One of the concepts that took a whle to sink in, for me, is the efficiency of the hardware support for context-switching on the CUDA chip.
Consequently, a context-switch occurs on every memory access, allowing calculations to proceed for many contexts alternately while the others wait on theri memory accesses. ne of the ways that GPGPU architectures achieve performance is the ability to parallelize this way, in addition to parallelizing on the multiples cores.
Best performance is achieved when no core is ever waiting on a memory access, and is achieved by having just enough contexts to ensure this happens.
There has been much discussion about how to choose the #blocks & blockSize, but I still missing something. Many of my concerns address this question: How CUDA Blocks/Warps/Threads map onto CUDA Cores? (To simplify the discussion, there is enough perThread & perBlock memory. Memory limits are not an issue here.)
kernelA<<<nBlocks, nThreads>>>(varA,constB, nThreadsTotal);
1) To keep the SM as busy as possible, I should set nThreads to a multiple of warpSize. True?
2) An SM can only execute one kernel at a time. That is all HWcores of that SM are executing only kernelA. (Not some HWcores running kernelA, while others run kernelB.) So if I have only one thread to run, I'm "wasting" the other HWcores. True?
3)If the warp-scheduler issues work in units of warpSize (32 threads), and each SM has 32 HWcores, then the SM would be full utilized. What happens when the SM has 48 HWcores? How can I keep all 48 cores full utilized when the scheduler is issuing work in chunks of 32? (If the previous paragraph is true, wouldn't it be better if the scheduler issued work in units of HWcore size?)
4) It looks like the warp-scheduler queues up 2 tasks at a time. So that when the currently-executing kernel stalls or blocks, the 2nd kernel is swapped in. (It is not clear, but I'll guess the queue here is more than 2 kernels deep.) Is this correct?
5) If my HW has an upper limit of 512 threads-per-block (nThreadsMax), that doesn't mean the kernel with 512 threads will run fastest on one block. (Again, mem not an issue.) There is a good chance I'll get better performance if I spread the 512-thread kernel across many blocks, not just one. The block is executed on one or many SM's. True?
5a) I'm thinking the smaller the better, but does it matter how small I make nBlocks? The question is, how to choose the value of nBlocks that is decent? (Not necessarily optimal.) Is there a mathematical approach to choosing nBlocks, or is it simply trial-n-err.
1) Yes.
2) CC 2.0 - 3.0 devices can execute up to 16 grids concurrently. Each SM is limited to 8 blocks so in order to reach full concurrency the device has to have at least 2 SMs.
3) Yes the warp schedulers select and issue warps at at time. Forget the concept of CUDA cores they are irrelevant. In order to hide latency you need to have high instruction level parallelism or a high occupancy. It is recommended to have >25% for CC 1.x and >50% for CC >= 2.0. In general CC 3.0 requires higher occupancy than 2.0 devices due to the doubling of schedulers but only a 33% increase in warps per SM. The Nsight VSE Issue Efficiency experiment is the best way to determine if you had sufficient warps to hide instruction and memory latency. Unfortunately, the Visual Profiler does not have this metric.
4) The warp scheduler algorithm is not documented; however, it does not consider which grid the thread block originated. For CC 2.x and 3.0 devices the CUDA work distributor will distribute all blocks from a grid before distributing blocks from the next grid; however, this is not guaranteed by the programming model.
5) In order to keep the SM busy you have to have sufficient blocks to fill the device. After that you want to make sure you have sufficient warps to reach a reasonable occupancy. There are both pros and cons to using large thread blocks. Large thread blocks in general use less instruction cache and have smaller footprints on cache; however, large thread blocks stall at syncthreads (SM can become less efficient as there are less warps to choose from) and tend to keep instructions executing on similar execution units. I recommend trying 128 or 256 threads per thread block to start. There are good reasons for both larger and smaller thread blocks.
5a) Use the occupancy calculator. Picking too large of a thread block size will often cause you to be limited by registers. Picking too small of a thread block size can find you limited by shared memory or the 8 blocks per SM limit.
Let me try to answer your questions one by one.
That is correct.
What do you mean exactly by "HWcores"? The first part of your statement is correct.
According to the NVIDIA Fermi Compute Architecture Whitepaper: "The SM schedules threads in groups of 32 parallel threads called warps. Each SM features two warp schedulers and two instruction dispatch units, allowing two warps to be issued and executed concurrently. Fermi’s dual warp scheduler selects two warps, and issues one instruction from each warp to a group of sixteen cores, sixteen load/store units, or four SFUs. Because warps execute independently, Fermi’s scheduler does not need to check for dependencies from within the instruction stream".
Furthermore, the NVIDIA Keppler Architecture Whitepaper states: "Kepler’s quad warp scheduler selects four warps, and two independent instructions per warp can be dispatched each cycle."
The "excess" cores are therefore used by scheduling more than one warp at a time.
The warp scheduler schedules warps of the same kernel, not different kernels.
Not quite true: Each block is locked in to a single SM, since that's where its shared memory resides.
That's a tricky issue and depends on how your kernel is implemented. You may want to have a look at the nVidia Webinar Better Performance at Lower Occupancy by Vasily Volkov which explains some of the more important issues. Primarily, though, I would suggest you choose your thread count to improve occupancy, using the CUDA Occupancy Calculator.
I have CUDA 2.1 installed on my machine and it has a graphic card with 64 cuda cores.
I have written a program in which I initialize simultaneously 30000 blocks (and 1 thread per block). But am not getting satisfying results from the gpu (It performs slowly than the cpu)
Is it that the number of blocks must be smaller than or equal to the number of cores for good performance? Or is it that the performance has nothing to do with number of blocks
CUDA cores are not exactly what you might call a core on a classical CPU. Indeed, they have to be viewed as nothing more than ALUs (Arithmetic and Logic Units), which are just able to compute ready operations.
You might know that threads are handled per warps (groups of 32 threads) inside the blocks you've defined. When your blocks are dispatched on the different SMs (Streaming Multiprocessors, they are the actual cores of the GPU), each SM schedules warps within a block to optimize the computation time in regard of the memory access time needed to get threads' input data.
The problem is threads are always handled through their belonging warp, so if you have only one thread per block, the SM it is running on won't be able to schedule through warps and you won't take advantage of the multiple CUDA cores available. Your CUDA cores will be waiting for data to process, since CUDA cores compute far quicker than data are retrieved through memory.
Having lots of blocks with few threads is not what the GPU is awaiting. In this case, you face the block per SM limitation (this number depends on your device), which force your GPU to spend a lot of time to put blocks on SM and then remove them to treat the next ones. You should rather increase the number of threads in your blocks instead of the number of blocks in your application.
The warp size in all current CUDA hardware is 32. Using less than 32 threads per block (or not using a round multiple of 32 threads per block) just wastes cycles. As it stands, using 1 thread per block is leaving something like 95% of the ALU cycles of your GPU idle. That is the underlying reason for the poor performance.