Once an image has been rendered onto the GPU, what are the steps of getting that image from GPU memory onto CPU RAM?
vkCmdCopyImage or vkCmdCopyImageToBuffer. Record it into command buffer and submit it to queue.
Related
I recently upgraded to colab pro. I am trying to use GPU resources from colab pro to train my Mask RCNN model. I was allocated around 15 GB of memory when I tried to run the model right after I signed up for Pro. However, for some reason, I was allocated just 1 GB of memory from the next morning. Since then, I haven't been allocated more than 1 GB. I was wondering if I am missing something or I perturbed the VM inherent packages. I understand that the allocation varies from day to day, but it's been like this for almost 3 days now. Following attempts have already made to improve, but none seems to work.
I have made sure that GPU and "High-RAM" option is selected.
I have tried restarting runtimes several times
I have tried running other scripts (just to make sure that problem was not with mask rcnn script)
I would appreciate any suggestions on this issue.
GPU info
The high memory setting in the screen controls the system RAM rather than GPU memory.
The command !nvidia-smi will show GPU memory. For example:
The highlighted output shows the GPU memory utilization: 0 of 16 GB.
I am working on APTOS Blindness detection challenge datasets from Kaggle. Post uploading the files; when I try to unzip the train images folder ; I get an error of file size limit saying the limited space available on RAM and Disk. Could any one please suggest an alternative to work with large size of image data.
If you get that error while unzipping the archive, it is a disk space problem. Colab gives you about 80 gb by default, try switching runtime to GPU acceleration, aside from better performance during certain tasks as using tensorflow, you will get about 350 gb of available space.
From Colab go to Runtime -> Change runtime type, and in the hardware acceleration menu select GPU.
If you need more disk space, Colab now offers a Pro version of the service with double disk space available in the free version.
If the (discrete) GPU has its own video RAM, I have to copy my data from RAM to VRAM to be able to use them. But if the GPU is integrated with the CPU (e.g. AMD Ryzen) and shares the memory, do I still have to make copies, or can they both alternatively access the same memory block?
It is possible to avoid copying in case of integrated graphics, but this feature is platform specific, and it may work differently for different vendors.
How to Increase Performance by Minimizing Buffer Copies on Intel® Processor Graphics article describes how to achieve this for Intel hardware:
To create zero copy buffers, do one of the following:
Use CL_MEM_ALLOC_HOST_PTR and let the runtime handle creating a zero copy allocation buffer for you
If you already have the data and want to load the data into an OpenCL buffer object, then use CL_MEM_USE_HOST_PTR with a buffer allocated at a 4096 byte boundary (aligned to a page and cache line boundary) and a total size that is a multiple of 64 bytes (cache line size).
When reading or writing data to these buffers from the host, use clEnqueueMapBuffer(), operate on the buffer, then call clEnqueueUnmapMemObject().
GPU and CPU memory sharing ?
GPU have multiple cores without control unit but the CPU controls the GPU through control unit. dedicated GPU have its own DRAM=VRAM=GRAM faster then integrated RAM. when we say integrated GPU its mean that GPU placed on same chip with CPU, and CPU & GPU used same RAM memory (shared memory ).
References to other similar Q&As:
GPU - System memory mapping
Data sharing between CPU and GPU on modern x86 hardware with OpenCL or other GPGPU framework
I tried to do Action Recognition using this Kinetics labels in colab. I refer this link
When I gave the input video below 2 MB this model was working fine. But if I give the input video more than 2 MB I got ResourceExhausted Error after few mins I got GPU memory usage is close to the limit.
Even I terminate the notebook and start the new one I got the same error.
As the error says, the physical limitations of your hardware has been reached. It requires more GPU memory than there is available.
You could prevent this by reducing the models' batch size, or resize the resolution of your input video sequence.
Alternatively, you could try to use Google's cloud training to gain additional hardware resources, however it is not free.
https://cloud.google.com/tpu/
I found some info in the internet that Core-Image process images on CPU if any of it's size bigger than 2048 (width or height or both). And it looks to be true because applying CIFilter even on 3200x2000 image is very slow. If I do the same on 2000x2000 image it is much faster. Is it possible to tell Core-Image to process all images on GPU always? Or maybe information I found was incorrect?
Processing on the GPU is not always faster, because your image data first has to be loaded to the GPU memory, processed, and then transferred back.
You can use kCIContextUseSoftwareRenderer to force software rendering (on the CPU) but there is no constant to force rendering on a GPU, I'm afraid. Also, software rendering does not work in the Simulator.
The maximum size depends on the device you're working on. For iPhone 3GS/4 & iPad 1, it's 2048*2048. For later iPhone/iPad, it's 4096*4096. On OSX, it would depend on your graphic card and/or OS version (2, 4, 8, or 16K²).
One possible way around the limit is to tile your image into pieces below the limit, and process each tile separately. Then you'd have to put the pieces back together after.