NVidia CUDA: cache L2 and multiple kernel invocations - caching

I'm wondering whether L2 cache is freed between multiple kernel invocations. For example I have a kernel that does some preprocessing on data and the second one that uses it. Is it possible to achieve greater performance if data size is less than 768 KB? I see no reason for NVidia guys to implement it otherwise but maybe I'm wrong. Does anybody have an experience with that?

Assuming you are talking about L2 data cache in Fermi.
I think the caches are flushed after each kernel invocation. In my experience, running two consecutive launches of the same kernel with a lots of memory accesses (and #L2 cache misses) doesn't make any substantial changes to the L1/L2 cache statistics.
In your problem, I think, depending on the data dependency, it is possible to put two stages into one kernel (with some sync) so the second part of the kernel can reuse the data processed by the first part.
Here is another trick: You know the gpu has, for example N SMs, you can perform the first part using the first N * M1 blocks. The next N * M2 blocks for the second part. Make sure all the blocks in the first part finish at the same time (or almost) using sync. In my experience, the block scheduling order is really deterministic.
Hope it helps.

Related

Do software prefetching hints disable hardware prefetcher?

The motivation of this quesion is to understand how software memory prefetching affects my program.
I'm building a multi-threaded data partitioner. Each thread sequencially read over a local source array and randomly write to another local destination array. As the content of the source array won't be used in near future, I'd like to use prefetchtnta instruction to avoid them growing inside caches. On the other hand, each thread has a local write combiner that combines writes and commits to the local destination array using _mm_stream_si64. The intuition and goal is to make sure each thread has a fixed size of data cache to work with and never being occupied by unused bits.
Is this design reasonable? I'm not familiar of how CPU works and cannot be sure if this strategy actually disables hardware prefetchers that presumably invalidate this approach. If this is just me being naive, what's the right way to achieve this goal?

Is it possible to use CPU cache in Golang?

Consider some memory and CPU intensive task:
e.g.: Task Block: read 16 bytes from memory then do CPU job. Then write back to memory.
And this Task Block can be parallelizable meaning each core can ran one Task Block.
e.g.: 8 CPU needs 8*16 byte cache but concurrently.
Yes, and just like all other code running on your machine, they all use CPU cache.
It's much too broad of a question to tell you how to code your app to make it the most efficient use of cache. I highly recommend setting up Go Benchmarks and then refactor your code and compare times. (Note, do not benchmark within a VM - VMs, and kind on any platform, do not have accurate enough clocks for Go's benchmarking. Run all Benchmarks native to your OS instead, no VM).
It all comes down to your ability to code the application to make efficient use of that CPU cache. This is a much broader topic for how you use your variables, how often they get updated, what stays on the heap or gets GC on the stack and how often, etc.
One tiny example to point you in the right direction to read more about efficient L1 and L2 cache development...
L1 cache uses 64 bit rows. If you want to store 4x 16bit Int16s, typically they will be allocated on the stack and most likely all stored on the same row of cache.
Say you want to update one of the Int16s? Well, CPU cache cannot update part of the row: It will have to invalidate the entire row, and allocate a whole new row of cache with the previous 3 Int16s and your new updates value.
Very inefficient.
One solution to that problem is use Int64s, which the CPU cache will only invalidate 1 row but yet keep the other 3 in cache for quick reads. Are you doing more push or pops? etc.
Again, it highly depends on your use case: this may even slow things down if you are using a lot of context switching of those 4 ints (e.g. mutex locks). In which case that's a whole different problem to optimize.
I recommend reading up on high frequency scaling and memory allocations on the stack and heaps.

Techniques available to control data/instructions in/out of the cache?

I have encountered some Intel compiler intrinsic functions which I believe allow developers to bypass the cache?
http://software.intel.com/sites/products/documentation/doclib/stdxe/2013/composerxe/compiler/fortran-mac/GUID-AF42A867-B796-4D29-8FED-C20193FD87E0.htm
I have also come across the GCC compiler prefetch keyword, although I cannot admit to fully appreciating what this does.
With the above in mind I wondered if any members could either elaborate on the above (which I badly described) or provide other techniques which allow the developer to have close control over which data (or instructions) is/isn't loaded in the CPU cache?
This page contains a lot of information about all intrinsics:
Intel Intrinsics Guide
The series of instructions that will write data to memory, avoiding cache evictions are generally named _mm_stream_.... As the name implies, these are ideal for applications that write a large stream of data that is basically contiguous in memory and unlikely to be accessed again in the near future. So, for example, if you are mixing audio buffers and producing a single waveform output this would work well.
One of the keys to using these instructions effectively is taking advantage of write combining. If your write locations are scattered throughout memory, these instructions will stall as badly, or possibly worse than any other kind of memory storage instruction you attempt. Since these writes do not wind up in cache, if you're not filling an entire write buffer then essentially your operation becomes a write-through operation, requiring a stall until the write is completed. If you are writing contiguous memory locations then write combining will apply, and make your data writes much more efficient.
The flip side of that coin is prefetching. Prefetching tells the system to start pulling a memory address into the desired level of cache so that by the time the memory read is complete, you are ready to use the data. This is much harder to use, and requires an appropriate data "stride" which takes into account the cache sizes, cache line size, and the number of instructions which can execute before the memory read completes. Using the hinting parameter, you can "suggest" that the data goes into the L1, L2, or L3 cache, or that it is "non-temporal", meaning that you're just going to use it once and it should be evicted first before any other cache evictions. The hardware has its own prefetching heuristics that work well for most problems without explicit prefetching instructions, but the classic counter-example is a matrix transpose:
Prefetching examples
Prefetching is generally very difficult to use effectively except in some very specific cases like this. Without a more specific problem statement from you, this is about all I can provide.

CUDA: When to use shared memory and when to rely on L1 caching?

After Compute Capability 2.0 (Fermi) was released, I've wondered if there are any use cases left for shared memory. That is, when is it better to use shared memory than just let L1 perform its magic in the background?
Is shared memory simply there to let algorithms designed for CC < 2.0 run efficiently without modifications?
To collaborate via shared memory, threads in a block write to shared memory and synchronize with __syncthreads(). Why not simply write to global memory (through L1), and synchronize with __threadfence_block()? The latter option should be easier to implement since it doesn't have to relate to two different locations of values, and it should be faster because there is no explicit copying from global to shared memory. Since the data gets cached in L1, threads don't have to wait for data to actually make it all the way out to global memory.
With shared memory, one is guaranteed that a value that was put there remains there throughout the duration of the block. This is as opposed to values in L1, which get evicted if they are not used often enough. Are there any cases where it's better too cache such rarely used data in shared memory than to let the L1 manage them based on the usage pattern that the algorithm actually has?
2 big reasons why automatic caching is less efficient than manual scratch pad memory (applies to CPUs as well)
parallel accesses to random addresses are more efficient. Example: histogramming. Let's say you want to increment N bins, and each are > 256 bytes apart. Then due to coalescing rules, that will result in N serial reads/writes since global and cache memory is organized in large ~256byte blocks. Shared memory doesn't have that problem.
Also to access global memory, you have to do virtual to physical address translation. Having a TLB that can do lots of translations in || will be quite expensive. I haven't seen any SIMD architecture that actually does vector loads/stores in || and I believe this is the reason why.
avoids writing back dead values to memory, which wastes bandwidth & power. Example: in an image processing pipeline, you don't want your intermediate images to get flushed to memory.
Also, according to an NVIDIA employee, current L1 caches are write-through (immediately writes to L2 cache), which will slow down your program.
So basically, the caches get in the way if you really want performance.
As far as i know, L1 cache in a GPU behaves much like the cache in a CPU. So your comment that "This is as opposed to values in L1, which get evicted if they are not used often enough" doesn't make much sense to me
Data on L1 cache isn't evicted when it isn't used often enough. Usually it is evicted when a request is made for a memory region that wasn't previously in cache, and whose address resolves to one that is already in use. I don't know the exact caching algorithm employed by NVidia, but assuming a regular n-way associative, then each memory entry can only be cached in a small subset of the entire cache, based on it's address
I suppose this may also answer your question. With shared memory, you get full control as to what gets stored where, while with cache, everything is done automatically. Even though the compiler and the GPU can still be very clever in optimizing memory accesses, you can sometimes still find a better way, since you're the one who knows what input will be given, and what threads will do what (to a certain extent of course)
Caching data through several memory layers always needs to follow a cache-coherency protocol. There are several such protocols and the decision on which one is the most suitable is always a trade off.
You can have a look at some examples:
Related to GPUs
Generally for computing units
I don't want to get in many details, because it is a huge domain and I am not an expert. What I want to point out is that in a shared-memory system (here the term shared does not refer to the so called shared memory of GPUs) where many compute-units (CUs) need data concurrently there is a memory protocol that attempts to keep the data close to the units so that can fetch them as fast as possible. In the example of a GPU when many threads in the same SM (symmetric multiprocessor) access the same data there should be a coherency in the sense that if thread 1 reads a chunk of bytes from the global memory and in the next cycle thread 2 is going to access these data, then an efficient implementation would be such that thread 2 is aware that data are found already in L1 cache and can access it fast. This is what the cache coherency protocol attempts to achieve, to let all compute units be up to date with what data exist in caches L1, L2 and so on.
However, keeping threads up to date, or else, keeping threads in coherent states, comes at some cost which is essentially missing cycles.
In CUDA by defining the memory as shared rather than L1-cache you free it from that coherency protocol. So access to that memory (which is physically the same piece of whatever material it is) is direct and does not implicitly call the functionality of coherency protocol.
I don't know how fast should this be, I didn't perform any such benchmark but the idea is that since you don't pay anymore for this protocol the access should be faster!
Of course, the shared memory on NVIDIA GPUs is split in banks and if someone wants to use it for performance improvement should have a look at this before. The reason is bank conflicts that occur when two threads access the same bank and this causes serialization of the access..., but that's another thing link

How does one write code that best utilizes the CPU cache to improve performance?

This could sound like a subjective question, but what I am looking for are specific instances, which you could have encountered related to this.
How to make code, cache effective/cache friendly (more cache hits, as few cache misses as possible)? From both perspectives, data cache & program cache (instruction cache),
i.e. what things in one's code, related to data structures and code constructs, should one take care of to make it cache effective.
Are there any particular data structures one must use/avoid, or is there a particular way of accessing the members of that structure etc... to make code cache effective.
Are there any program constructs (if, for, switch, break, goto,...), code-flow (for inside an if, if inside a for, etc ...) one should follow/avoid in this matter?
I am looking forward to hearing individual experiences related to making cache efficient code in general. It can be any programming language (C, C++, Assembly, ...), any hardware target (ARM, Intel, PowerPC, ...), any OS (Windows, Linux,S ymbian, ...), etc..
The variety will help to better to understand it deeply.
The cache is there to reduce the number of times the CPU would stall waiting for a memory request to be fulfilled (avoiding the memory latency), and as a second effect, possibly to reduce the overall amount of data that needs to be transfered (preserving memory bandwidth).
Techniques for avoiding suffering from memory fetch latency is typically the first thing to consider, and sometimes helps a long way. The limited memory bandwidth is also a limiting factor, particularly for multicores and multithreaded applications where many threads wants to use the memory bus. A different set of techniques help addressing the latter issue.
Improving spatial locality means that you ensure that each cache line is used in full once it has been mapped to a cache. When we have looked at various standard benchmarks, we have seen that a surprising large fraction of those fail to use 100% of the fetched cache lines before the cache lines are evicted.
Improving cache line utilization helps in three respects:
It tends to fit more useful data in the cache, essentially increasing the effective cache size.
It tends to fit more useful data in the same cache line, increasing the likelyhood that requested data can be found in the cache.
It reduces the memory bandwidth requirements, as there will be fewer fetches.
Common techniques are:
Use smaller data types
Organize your data to avoid alignment holes (sorting your struct members by decreasing size is one way)
Beware of the standard dynamic memory allocator, which may introduce holes and spread your data around in memory as it warms up.
Make sure all adjacent data is actually used in the hot loops. Otherwise, consider breaking up data structures into hot and cold components, so that the hot loops use hot data.
avoid algorithms and datastructures that exhibit irregular access patterns, and favor linear datastructures.
We should also note that there are other ways to hide memory latency than using caches.
Modern CPU:s often have one or more hardware prefetchers. They train on the misses in a cache and try to spot regularities. For instance, after a few misses to subsequent cache lines, the hw prefetcher will start fetching cache lines into the cache, anticipating the application's needs. If you have a regular access pattern, the hardware prefetcher is usually doing a very good job. And if your program doesn't display regular access patterns, you may improve things by adding prefetch instructions yourself.
Regrouping instructions in such a way that those that always miss in the cache occur close to each other, the CPU can sometimes overlap these fetches so that the application only sustain one latency hit (Memory level parallelism).
To reduce the overall memory bus pressure, you have to start addressing what is called temporal locality. This means that you have to reuse data while it still hasn't been evicted from the cache.
Merging loops that touch the same data (loop fusion), and employing rewriting techniques known as tiling or blocking all strive to avoid those extra memory fetches.
While there are some rules of thumb for this rewrite exercise, you typically have to carefully consider loop carried data dependencies, to ensure that you don't affect the semantics of the program.
These things are what really pays off in the multicore world, where you typically wont see much of throughput improvements after adding the second thread.
I can't believe there aren't more answers to this. Anyway, one classic example is to iterate a multidimensional array "inside out":
pseudocode
for (i = 0 to size)
for (j = 0 to size)
do something with ary[j][i]
The reason this is cache inefficient is because modern CPUs will load the cache line with "near" memory addresses from main memory when you access a single memory address. We are iterating through the "j" (outer) rows in the array in the inner loop, so for each trip through the inner loop, the cache line will cause to be flushed and loaded with a line of addresses that are near to the [j][i] entry. If this is changed to the equivalent:
for (i = 0 to size)
for (j = 0 to size)
do something with ary[i][j]
It will run much faster.
The basic rules are actually fairly simple. Where it gets tricky is in how they apply to your code.
The cache works on two principles: Temporal locality and spatial locality.
The former is the idea that if you recently used a certain chunk of data, you'll probably need it again soon. The latter means that if you recently used the data at address X, you'll probably soon need address X+1.
The cache tries to accomodate this by remembering the most recently used chunks of data. It operates with cache lines, typically sized 128 byte or so, so even if you only need a single byte, the entire cache line that contains it gets pulled into the cache. So if you need the following byte afterwards, it'll already be in the cache.
And this means that you'll always want your own code to exploit these two forms of locality as much as possible. Don't jump all over memory. Do as much work as you can on one small area, and then move on to the next, and do as much work there as you can.
A simple example is the 2D array traversal that 1800's answer showed. If you traverse it a row at a time, you're reading the memory sequentially. If you do it column-wise, you'll read one entry, then jump to a completely different location (the start of the next row), read one entry, and jump again. And when you finally get back to the first row, it will no longer be in the cache.
The same applies to code. Jumps or branches mean less efficient cache usage (because you're not reading the instructions sequentially, but jumping to a different address). Of course, small if-statements probably won't change anything (you're only skipping a few bytes, so you'll still end up inside the cached region), but function calls typically imply that you're jumping to a completely different address that may not be cached. Unless it was called recently.
Instruction cache usage is usually far less of an issue though. What you usually need to worry about is the data cache.
In a struct or class, all members are laid out contiguously, which is good. In an array, all entries are laid out contiguously as well. In linked lists, each node is allocated at a completely different location, which is bad. Pointers in general tend to point to unrelated addresses, which will probably result in a cache miss if you dereference it.
And if you want to exploit multiple cores, it can get really interesting, as usually, only one CPU may have any given address in its L1 cache at a time. So if both cores constantly access the same address, it will result in constant cache misses, as they're fighting over the address.
I recommend reading the 9-part article What every programmer should know about memory by Ulrich Drepper if you're interested in how memory and software interact. It's also available as a 104-page PDF.
Sections especially relevant to this question might be Part 2 (CPU caches) and Part 5 (What programmers can do - cache optimization).
Apart from data access patterns, a major factor in cache-friendly code is data size. Less data means more of it fits into the cache.
This is mainly a factor with memory-aligned data structures. "Conventional" wisdom says data structures must be aligned at word boundaries because the CPU can only access entire words, and if a word contains more than one value, you have to do extra work (read-modify-write instead of a simple write). But caches can completely invalidate this argument.
Similarly, a Java boolean array uses an entire byte for each value in order to allow operating on individual values directly. You can reduce the data size by a factor of 8 if you use actual bits, but then access to individual values becomes much more complex, requiring bit shift and mask operations (the BitSet class does this for you). However, due to cache effects, this can still be considerably faster than using a boolean[] when the array is large. IIRC I once achieved a speedup by a factor of 2 or 3 this way.
The most effective data structure for a cache is an array. Caches work best, if your data structure is laid out sequentially as CPUs read entire cache lines (usually 32 bytes or more) at once from main memory.
Any algorithm which accesses memory in random order trashes the caches because it always needs new cache lines to accomodate the randomly accessed memory. On the other hand an algorithm, which runs sequentially through an array is best because:
It gives the CPU a chance to read-ahead, e.g. speculatively put more memory into the cache, which will be accessed later. This read-ahead gives a huge performance boost.
Running a tight loop over a large array also allows the CPU to cache the code executing in the loop and in most cases allows you to execute an algorithm entirely from cache memory without having to block for external memory access.
One example I saw used in a game engine was to move data out of objects and into their own arrays. A game object that was subject to physics might have a lot of other data attached to it as well. But during the physics update loop all the engine cared about was data about position, speed, mass, bounding box, etc. So all of that was placed into its own arrays and optimized as much as possible for SSE.
So during the physics loop the physics data was processed in array order using vector math. The game objects used their object ID as the index into the various arrays. It was not a pointer because pointers could become invalidated if the arrays had to be relocated.
In many ways this violated object-oriented design patterns but it made the code a lot faster by placing data close together that needed to be operated on in the same loops.
This example is probably out of date because I expect most modern games use a prebuilt physics engine like Havok.
A remark to the "classic example" by user 1800 INFORMATION (too long for a comment)
I wanted to check the time differences for two iteration orders ( "outter" and "inner"), so I made a simple experiment with a large 2D array:
measure::start();
for ( int y = 0; y < N; ++y )
for ( int x = 0; x < N; ++x )
sum += A[ x + y*N ];
measure::stop();
and the second case with the for loops swapped.
The slower version ("x first") was 0.88sec and the faster one, was 0.06sec. That's the power of caching :)
I used gcc -O2 and still the loops were not optimized out. The comment by Ricardo that "most of the modern compilers can figure this out by itselves" does not hold
Only one post touched on it, but a big issue comes up when sharing data between processes. You want to avoid having multiple processes attempting to modify the same cache line simultaneously. Something to look out for here is "false" sharing, where two adjacent data structures share a cache line and modifications to one invalidates the cache line for the other. This can cause cache lines to unnecessarily move back and forth between processor caches sharing the data on a multiprocessor system. A way to avoid it is to align and pad data structures to put them on different lines.
I can answer (2) by saying that in the C++ world, linked lists can easily kill the CPU cache. Arrays are a better solution where possible. No experience on whether the same applies to other languages, but it's easy to imagine the same issues would arise.
Cache is arranged in "cache lines" and (real) memory is read from and written to in chunks of this size.
Data structures that are contained within a single cache-line are therefore more efficient.
Similarly, algorithms which access contiguous memory blocks will be more efficient than algorithms which jump through memory in a random order.
Unfortunately the cache line size varies dramatically between processors, so there's no way to guarantee that a data structure that's optimal on one processor will be efficient on any other.
To ask how to make a code, cache effective-cache friendly and most of the other questions , is usually to ask how to Optimize a program, that's because the cache has such a huge impact on performances that any optimized program is one that is cache effective-cache friendly.
I suggest reading about Optimization, there are some good answers on this site.
In terms of books, I recommend on Computer Systems: A Programmer's Perspective which has some fine text about the proper usage of the cache.
(b.t.w - as bad as a cache-miss can be, there is worse - if a program is paging from the hard-drive...)
There has been a lot of answers on general advices like data structure selection, access pattern, etc. Here I would like to add another code design pattern called software pipeline that makes use of active cache management.
The idea is borrow from other pipelining techniques, e.g. CPU instruction pipelining.
This type of pattern best applies to procedures that
could be broken down to reasonable multiple sub-steps, S[1], S[2], S[3], ... whose execution time is roughly comparable with RAM access time (~60-70ns).
takes a batch of input and do aforementioned multiple steps on them to get result.
Let's take a simple case where there is only one sub-procedure.
Normally the code would like:
def proc(input):
return sub-step(input))
To have better performance, you might want to pass multiple inputs to the function in a batch so you amortize function call overhead and also increases code cache locality.
def batch_proc(inputs):
results = []
for i in inputs:
// avoids code cache miss, but still suffer data(inputs) miss
results.append(sub-step(i))
return res
However, as said earlier, if the execution of the step is roughly the same as RAM access time you can further improve the code to something like this:
def batch_pipelined_proc(inputs):
for i in range(0, len(inputs)-1):
prefetch(inputs[i+1])
# work on current item while [i+1] is flying back from RAM
results.append(sub-step(inputs[i-1]))
results.append(sub-step(inputs[-1]))
The execution flow would look like:
prefetch(1) ask CPU to prefetch input[1] into cache, where prefetch instruction takes P cycles itself and return, and in the background input[1] would arrive in cache after R cycles.
works_on(0) cold miss on 0 and works on it, which takes M
prefetch(2) issue another fetch
works_on(1) if P + R <= M, then inputs[1] should be in the cache already before this step, thus avoid a data cache miss
works_on(2) ...
There could be more steps involved, then you can design a multi-stage pipeline as long as the timing of the steps and memory access latency matches, you would suffer little code/data cache miss. However, this process needs to be tuned with many experiments to find out right grouping of steps and prefetch time. Due to its required effort, it sees more adoption in high performance data/packet stream processing. A good production code example could be found in DPDK QoS Enqueue pipeline design:
http://dpdk.org/doc/guides/prog_guide/qos_framework.html Chapter 21.2.4.3. Enqueue Pipeline.
More information could be found:
https://software.intel.com/en-us/articles/memory-management-for-optimal-performance-on-intel-xeon-phi-coprocessor-alignment-and
http://infolab.stanford.edu/~ullman/dragon/w06/lectures/cs243-lec13-wei.pdf
Besides aligning your structure and fields, if your structure if heap allocated you may want to use allocators that support aligned allocations; like _aligned_malloc(sizeof(DATA), SYSTEM_CACHE_LINE_SIZE); otherwise you may have random false sharing; remember that in Windows, the default heap has a 16 bytes alignment.
Write your program to take a minimal size. That is why it is not always a good idea to use -O3 optimisations for GCC. It takes up a larger size. Often, -Os is just as good as -O2. It all depends on the processor used though. YMMV.
Work with small chunks of data at a time. That is why a less efficient sorting algorithms can run faster than quicksort if the data set is large. Find ways to break up your larger data sets into smaller ones. Others have suggested this.
In order to help you better exploit instruction temporal/spatial locality, you may want to study how your code gets converted in to assembly. For example:
for(i = 0; i < MAX; ++i)
for(i = MAX; i > 0; --i)
The two loops produce different codes even though they are merely parsing through an array. In any case, your question is very architecture specific. So, your only way to tightly control cache use is by understanding how the hardware works and optimising your code for it.

Resources