I am trying to understand below lines from here
How quickly can data be sent to the GPU or read back from it?
How fast can the GPU kernel read and write data?
How quickly can data be sent to the GPU? Which peripheral device sent data to GPU? Where GPU kernel read and write data? to the data bus?
Is this implementation showing us how everything [GPU and other peripheral devices] contribute towards computation performance?
Related
I am new to rocket chip. I want to design a coprocessor to accelerate data processing. I have a question about how to do a large chunk of data exchange between core and accelerator for one custom instruction. I am wondering if dcache can be used for this data exchange. I went to check LazyRoCC.scala and found io.mem.resp.bits.data only xLen bits long. Does it mean the length of data exchange between dcache and RoCC is limited to xLen? Is there any other way to do this data exchange? Thank you in advance!
I am learning about the STM32 F4 microcontroller. I'm trying to find out about limitations for using DMA.
Per my understanding and research, I know that if the data size is small (that is, the device uses DMA to generate or consume a small amount of data), the overhead is increased because DMA transfer requires the DMA controller to perform operations, thereby unnecessarily increasing system cost.
I did some reaserch and found the following:
Limitation of DMA
CPU puts all its lines at high impedance state so that the DMA controller can then transfer data directly between device and memory without CPU intervention. Clearly, it is more suitable for device with high data transfer rates like a disk.
Over a serial interface, data is transferred one bit at a time which makes it slow to use DMA.
Is that correct? What else do I need to know?
DMA -CPU puts all its lines at high impedance state
I do not know where did you take it from - but you should not use this source any more.
Frequency of the DMA transfers do not matter unless you reach the the BUS throughput. you can transfer one byte per week, month, year, decade ..... and it is absolutely OK.
In the STM32 microcontrollers it is a very important feature as we can transfer data from/to external devices even if the uC is in low power mode with the core (CPU) sleeping. DMA controller can even wake up the core when some conditions are met.
As #Vinci and #0___________ (f.k.a. #P__J__) already pointed out,
A DMA controller works autonomously and doesn't create overhead on the CPU it supplements (at least not by itself). But:
The CPU/software must perform some instructions to configure the DMA and to trigger it or have it triggered by some peripheral. For this, it needs CPU time and program memory space (usually ROM). Besides, it usually needs some additional RAM in variables to manage the software around the DMA.
Hence, you are right, using a DMA comes with some kinds of overhead.
And furthermore,
The DMA transfers make use of the memory bus(es) that connect the involved memories/registers/peripherals to the DMA controller. That is, while the DMA controller does its own work, it may cause the CPU which it tries to offload to stall in the meantime, at least for short moments when the data words are transferred (which in turn sum up for longer transfers...).
On the other hand, a DMA doesn't only help you to reduce the CPU load (regarding total CPU time to implement some feature). If used "in a smart way", it helps you to reduce software latencies to implement different functions because one part of the implementation can be "hidden" behind the DMA-driven data transfer of another part (unless, both rely on the same bus resources - see above...).
The information is right in that using a DMA requires some development work and some runtime to manage the DMA transfer itself (see also
a related question
here), which may not be worth the benefits of using DMA. That is, for small portions of data one doesn't gain as much performance (or latency) as during big transfers. On embedded systems, DMA controllers (and their channels) are limited resources so it is important to consider which part of the function benefits from such a resource most. Therefore, one would usually prefer using DMA for the data transfers to/from disks (if it is about "payload data" such as large files or video streams) over slow serial connections.
The information is wrong, however, in that DMA is not worth using on serial interfaces as those only transfer a single bit at a time. Please note that microcontrollers (as your
STM32F4)
have built-in peripheral components that convert the serial bit-by-bit stream into a byte-by-byte or word-by-word stream, which can easily be tranferred by DMA in a helpful way - especially if the size of the packets is known in advance and software doesn't have to analyse a non-formatted stream. Furthermore, not every serial connection is "slow" at all. If the project uses, e. g., an SPI flash chip, then the SPI serial connection is the one used for data transfer.
I need to take video frames from pi and send them further to zmq socket. These should be filtered to those only where some movement takes place on the scene.
Another requrement is to completely avoid disk writes. My idea is to use Motion and redirect all file output to /dev/shm. My process will read from there
and send further. Motion will be setup to save frames at the rate allowing to handle and remove them before memory is exhausted.
It would be great to have another solution which could give me frames one-by-one through the same process address space.
I am pretty new to tensorflow. I used to use theano for deep learning development. I notice a difference between these two, that is where input data can be stored.
In Theano, it supports shared variable to store input data on GPU memory to reduce the data transfer between CPU and GPU.
In tensorflow, we need to feed data into placeholder, and the data can come from CPU memory or files.
My question is: is it possible to store input data on GPU memory for tensorflow? or does it already do it in some magic way?
Thanks.
If your data fits on the GPU, you can load it into a constant on GPU from e.g. a numpy array:
with tf.device('/gpu:0'):
tensorflow_dataset = tf.constant(numpy_dataset)
One way to extract minibatches would be to slice that array at each step instead of feeding it using tf.slice:
batch = tf.slice(tensorflow_dataset, [index, 0], [batch_size, -1])
There are many possible variations around that theme, including using queues to prefetch the data to GPU dynamically.
It is possible, as has been indicated, but make sure that it is actually useful before devoting too much effort to it. At least at present, not every operation has GPU support, and the list of operations without such support includes some common batching and shuffling operations. There may be no advantage to putting your data on GPU if the first stage of processing is to move it to CPU.
Before trying to refactor code to use on-GPU storage, try at least one of the following:
1) Start your session with device placement logging to log which ops are executed on which devices:
config = tf.ConfigProto(log_device_placement=True)
sess = tf.Session(config=config)
2) Try to manually place your graph on GPU by putting its definition in a with tf.device('/gpu:0'): block. This will throw exceptions if ops are not GPU-supported.
Considering that a processor runs at 100 MHz and the data is coming to the processor from an external device/peripheral at the rate of 1000 Mbit/s (8 Bits/Clockcycle # 125 MHz), which is the best way to handle traffic that comes at a higher speed to the processor ?
First off, you can't do it in software. There would be no way to sample the digital lines at a sufficient rate, or to doing anything useful with it.
You need to use a hardware FIFO buffer or memory cell. When a data burst comes in, it can be buffered in the high speed FIFO and then read out as needed by the processor.
Drop in high speed FIFO chips are surprisingly expensive (though most are dual ported). To cut cost, you would be best off using an SRAM chip, and a hardware adder to increment the address lines on incoming data.
This is not an uncommon situation for software. semaj said the right word. This is a system engineering issue. Other folks have the right answer too. If you want to look at or process that data with the 100MHz processor, it is not going to happen, dont bother trying. You CAN look at snapshots of it or have the hardware filter out a specific percentage of it that you are looking for. At the end of the day though it is a systems issue, what does the hardware provide, where does it put this data, what is the softwares task for this data, does it see X buffers of data come in on the goesinta, and the notify the goesouta hardware that there are X buffers ready to go? Does the hardware examine and align the buffers so that you can look at a header, and then decide where to route the hardware? Once you do your system engineering you will know if you can use that processor or not, and if you can use it what its job is and how to do it.
Your direct question. What is the best way to handle it. The best way to handle it is to have hardware (fpga, asic, etc) move it into and out of some storage device (ram of some sort probably). Not necessarily the same ram the processor runs out of (DMA is a good thing to avoid). The hardware is something the software can talk to but you cannot examine all of that data so dont try. Without knowing what kind of data this is, what form, what the software looks at how much work you are willing to force the hardware to do, etc determines the rest of the answer. If you expect a certain (guaranteed) percentage to be bad or not belong to this processor, etc have the hardware filter that out and then what is left you can process.
Networking is a good example of this, PCs have gige ports but cannot process GigE line rate data. That is why we use switches now instead of hubs, hardware slices out a percentage of the data so the pc can handle it, the protocols take care of the data that cannot be processed by resending it later. And the switches processors dont look at all of the data, the hardware slices it up so the software can examine just the header. Or sometimes the software simply manages tables that drive the hardware and the hardware does all the work of processing the data.
Do your system engineering the answers will simply fall out.
You buffer it. Typically data from a device is written to a memory buffer (circular queue) using DMA (no cpu involved). The cpu reads from the memory buffer at a constant rate. Usually devices send data in bursts. This keeps the buffer from filling up. If there is too much data, buffer overflow.
DMA (direct memory access) is possibly the solution, however, it seems unlikely that the memory bus could run faster than the processor core, so the receiving peripheral would have to accept data into a larger register than 8 bit because 125MHz could not be sustained. For example a 16bit register would allow memory writes at 62.5MHz which may be achievable. Also the receiving device would have to be able to accept an external clock that is both faster and asynchronous to the core clock. Also of course the receiving peripheral must have support for DMA.
Unless you are more specific about your hardware and the communication protocol it is difficult to give anything other than a general answer.