Re: [Openexr-devel] Writing Deep tiled images: help !

[peterh]

Yes!

On 2014-04-12 17:55, Larry Gritz wrote:

OK, my own code is making more sense now, although it embodies some minor misunderstandings.

 

Let me regurgitate one more time for clarity:

 

Scanline files can be either increasing or decreasing Y order and MUST be presented in the order specified. No random writes for scanlines.

 

Tiled files can be increasing Y, decreasing Y, or random Y, and for the first two cases, out-of-order tiles will be compressed and buffered internal to libIlmImf.

 

Yes?

 

On Apr 11, 2014, at 10:45 PM, peterh <address@hidden> wrote:

It appears that support for buffering tiles to handle out-of-order calls to writeTiles was present in the first version of OpenEXR that supported tiles (OpenEXR-1.1.0)

I realise my last reply about scanlines was incorrect due to some poor editing skills on my part:

Scanline images have no buffering between calls to writePixels; there's no way of delivering scanlines in any order since you don't pass writePixels a y offset, merely a number of scanlines to write. Calling writePixels(1) will always write currentScanline()

The buffering that happens is with tiled images in INCREASING_Y or DECREASING_Y, and then only when you specify tiles out of order.

Also worth pointing out that OpenEXR buffers the compressed data, not the raw pixel data. If you are creating tiles out of order and want INCREASING_Y or DECREASING_Y tile order, it will be more memory efficient to pass the tiles immediately to the library and have it manage buffering, rather than buffering them yourself until you have them ready in the correct order. It may also be appropriate to write a temporary file in RANDOM_Y order, and then read it back in and write it out in INCREASING_Y order

If the image is small enough to fit in memory and you have multithreading enabled it is best to write the entire image in one call to writeTiles or writePixels.

 

 

On 2014-04-12 06:35, Piotr Stanczyk wrote:

I am not sure about that ... let me check

 

Piotr

On 11 April 2014 10:19, Larry Gritz <address@hidden> wrote:

Was it always like this, in 1.x as well? Some of my code seems to bend over backwards to not allow random order writes unless a specific request was made for random Y order. I don't remember now why I wrote it that way, but the code is structured as if at some time in the past, the library would not have done the buffering without RANDOM_Y lineOrder. Was I just always wrong about this?

        -- lg

On Apr 11, 2014, at 10:13 AM, Larry Gritz <address@hidden> wrote:

> OK, so let me restate it to make sure I have it right:
>
> For scanline files, storage in the file is either increasing or decreasing (determined by lineOrder); libIlmImf will accept scanlines in any order, but if they are sent other than with the line order you specified, the library will buffer internally.
>
> For tiled files, libIlmImf will accept tiles in any order, buffering if they are not in canonical order, UNLESS you set lineOrder to RANDOM_Y, in which case it really will store it in random order rather than buffering (but at the cost of reading that file in canonical order having more seeks than if it had been written in canonical tile order).
>
> Do I have that all correct?
>
>
>
> On Apr 10, 2014, at 5:10 PM, Peter Hillman <address@hidden> wrote:
>
>> RANDOM_Y is not supported for scanline images - you either start at datawindow.min.y and work to datawindow.max.y or vice versa. You can deliver scanlines out of order: they'll be buffered before writing.
>>
>> With tiled images, you can write tiles in any order, but you must set RANDOM_Y to allow it write without buffering.
>>
>>
>>
>> On 11/04/14 12:02, Larry Gritz wrote:
>>> Uh oh.
>>>
>>> I knew about RANDOM_Y for scanline images. Are you saying that for tiled images, if we don't set lineorder to RANDOM_Y, then there's internal buffering if we send the tiles in anything but canonical order?
>>>
>>>
>>>
>>> On Apr 10, 2014, at 4:54 PM, Peter Hillman <address@hidden> wrote:
>>>
>>>> In general with OpenEXR, you cannot write to a file from different threads unless they have a shared framebuffer. Calling "setFrameBuffer" from one thread while another is writing may cause a crash.
>>>>
>>>> Even if you do have a shared framebuffer, the threads lock to serialise the writes: only one "writePixels" or "writeTile" request goes through at once.
>>>> If you don't use the IlmThread library to run your threads it may be wise to implement your own lock mechanism around writePixel and writeTile calls.
>>>>
>>>>
>>>> There are two ways of writing tiled data: if you use xTileCoords/yTileCoords, then pixel(0,0) is the bottom left corner of each tile, implying the image buffer pointed to by the slice is the same size as the tile.
>>>> if you don't do that, then pixel(0,0) is the bottom left corner of the displayWindow, implying the image buffer pointed to be the slice is the size of the dataWindow, accounting for offset data windows appropriately.
>>>>
>>>> In your case, it would seem like you want the second option: preallocate your dataWindow sized array of pointers-to-samples, and call setFrameBuffer before any rendering begins.
>>>> Then each thread sets the pointers to the samples for the tile it is going to write and fills in the corresponding part of the samplecount array before calling writeTile.
>>>>
>>>> The pointers need not be valid when setFrameBuffer is called. When you call writeTile, only the pointers and sampleCount data for the pixels being written need to be valid, and need not remain valid afterwards. So yes, you can delete data for the actual sample data of a tile after calling writeTile, though you'll need to keep the sampleCount array and the pointers-to-samples array for each channel.
>>>>
>>>> Note that if you are writing tiles in anything other than RANDOM_Y order, then whenever your threads deliver tiles out of order OpenEXR will have to buffer data so it can write the tiles sequentially.
>>>> I suspect this may well be the case from multithreaded write applications. With deep data this could take a significant amount of memory.
>>>> You may wish to enforce RANDOM_Y order to allow the library to write tiles to disk as soon as you call writeTile. If your read pattern is generally also random, rather than in 'scanline order', there'll be little performance hit.
>>>>
>>>> You might also consider having one worker thread managing the writing of tile data and marshalling into the format required for the DeepSlices, and have other threads compute data and signal the worker thread when they have tiles ready for processing.
>>>>
>>>> Yes, you can use an Array2D<float*> for your data points (dataZ and dataA in the example). Each value is a pointer to the first sample of the channel data for each pixel.
>>>> Subsequent samples are found using the offset provided. That means you can store pixel data in separate or interleaved channels.
>>>>
>>>>
>>>>
>>>> On 11/04/14 02:16, Michel Lerenard wrote:
>>>>> Hi,
>>>>>
>>>>> I'm trying to write deep data into an EXR file and am having trouble understanding how the API works.
>>>>> I'd like to be able to write tiles from different threads. The application I'm working on (Isotropix Clarisse) runs several threads, each one render a part of an image.
>>>>> I'd like to write deep data when the current tile is completely computed.
>>>>>
>>>>> I'm using the "Reading And Writing Image Files" documentation, and the section about Deep Tile File writing on page 26 is confusing to me.
>>>>>
>>>>> From what I understand, I need a DeepFrameBuffer to write an image. It seems it should be the same for all threads, as a DeepTiledOutputFile can only use one. Is it thread safe ? Can I use the same FrameBuffer from all threads ?
>>>>> I was puzzled by the type of dataZ and data0, reading the next example (reading a deep tiled file) I understand that the arrays are only storing pointers to deep data for a particular pixel, and that they could as well be Array2D<float*>. Is that correct ?
>>>>>
>>>>> If it is, this means that each one of my threads should set pointers in the framebuffer DeepSlice for each pixel of the tile it's computing, then call writeTile. After that can I delete the data and reset the pointers in the DeepSlice ?
>>>>>
>>>>>
>>>>> Thanks,
>>>>>
>>>>> Michel
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> Openexr-devel mailing list
>>>>> address@hidden
>>>>> https://lists.nongnu.org/mailman/listinfo/openexr-devel
>>>>
>>>> _______________________________________________
>>>> Openexr-devel mailing list
>>>> address@hidden
>>>> https://lists.nongnu.org/mailman/listinfo/openexr-devel
>>> --
>>> Larry Gritz
>>> address@hidden
>>>
>>>
>>>
>>
>
> --
> Larry Gritz
> address@hidden
>
>
>
>
> _______________________________________________
> Openexr-devel mailing list
> address@hidden
> https://lists.nongnu.org/mailman/listinfo/openexr-devel

--
Larry Gritz
address@hidden




_______________________________________________
Openexr-devel mailing list
address@hidden
https://lists.nongnu.org/mailman/listinfo/openexr-devel

 

--
Larry Gritz
address@hidden