Whether ImageCache is a good idea depends on what the access pattern is. The case that ImageCache is most helpful is for when your access pattern is visiting and revisiting the same image (out of many many images), and each time you need to visit the image, you only need to access a few of the pixels, more or less coherently, and then the next time you need pixels from that image, it's either likely to be in a short time and needing nearby pixels, or not for a long time until you need a far-away part of that image. Also, this whole scheme works best when the images are *tiled*, which BMPs are not. When the usage patterns align, ImageCache can be used, for example, for an offline renderer to access 2TB of texture data per frame, spread across 20,000 texture files, all without using more than a few GB in memory at any one time (and with maybe only 20% redundant disk reads).
If you are reading one image at a time, need all the pixels in the image before moving on to the next image, and then are done with that first image, then you should not be using ImageCache at all. You can still use an ImageBuf if you want, bypassing the ImageCache, by making sure that when you call ImageBuf::read(), you pass the optional parameter force=true. But when you free one ImageBuf and make the next one, it will free and then re-allocate the memory. I assure you, although this constant alloc/free cycle seems wasteful, it will be a tiny fraction of the cost of the actual disk I/O and the compute necessary to decompress the pixel data and you will barely notice it. Another route you could take -- again, assuming that all these images are the same size and you only need one in memory at a time -- is to not use ImageBuf either. Just allocate your own buffer of the right size, use ImageInput::read_image to get the pixels into it, process it, and then read the next image into the same memory buffer without ever freeing it. ImageInput, ImageBuf, and ImageCache are different levels of abstraction, intended for different use cases. > On Apr 5, 2021, at 4:55 PM, Phil Miller <philip.mil...@sri.com> wrote: > > Well, our application is a little newer than horse-drawn wagons, and we do > use bmp files from time to time. This particular usage is for an instrument > that is outputting sensor data at scan locations using bmp images. Bmp images > tend to be simple, easy to use, well supported on both windows and linux, and > (still 😊) popular for certain applications. For my immediate use case, I am > not in control of the format and am just given the images that we need to > process. I have been able to use opencv to load the bmp images, but the > opencv reader is doing a memory allocation for each image that I read. > Consequently, I am churning through way too many memory allocations, as > opencv is doing an alloc, read, and free for each image that I read. What I > am trying to do is have an image cache that resuses images and avoids > unnecessary allocations. So I was looking to see if oiio would be better. > > Since oiio considers bmp files to be an obsolete format (and I concur that > there are many better image formats than bmp), let me ask a different > question: > If I choose to store my images (and their pyramids/mipmaps) in a tif file, > can I use oiio and its ImageCache to re-use a pre-existing memory block to > read the pixel values into? In particular, suppose I am willing to use a 10 > GB image cache. After the cache has filled up and it needs to free an image > before reading another one, will the cache reuse the memory block or free the > existing one and allocate a new one? Does it matter, that is will it help, if > all of my images are the same size? If I can get oiio to reduce the memory > allocations, then I can write a pre-processing application to use opencv to > convert the single channel bmp files to mipmapped tif files and then write > the real image processing app using oiio. > > Thanks for your time! > > From: Larry Gritz <l...@larrygritz.com> > Sent: Monday, April 5, 2021 6:53 PM > To: OpenImageIO dev list <oiio-dev@lists.openimageio.org> > Cc: Phil Miller <philip.mil...@sri.com> > Subject: Re: [Oiio-dev] [EXTERNAL] Re: Reading single channel bmp > > Yeah, sorry about that. I fixated so quickly on "is the texture 1->3 channel > conversion somehow going awry" that I failed to notice that you said BMP, > which is an unusual enough file format that there could be something > particular to its reader. > > And indeed, that is the case. > > There's not really such a thing as a BMP "grayscale" image. > > As I understand it, BMP files are either RGB[A] full color per pixel (24 or > 32 bits), or 16 bits per pixel (where those 16 bits encode R, G, and B), or > else are "palette" images. For palette images, the number of bits per pixel > indicates the number of palette entries, e.g., 8 bits means 256 entries in > the color palette, 4 bits means only 16 colors in the palette. But either > way, the palette itself consists of entries of what they call an "RGBQUAD", > so these palette images are reported by OIIO to be RGB images (OIIO hides the > palette itself from you so that apps don't need to worry about it and can > just ask for pixel colors). > > You could make an 8-bit per pixel BMP appear to encode a grayscale image by > making it a palette image in which all the palette entries coincidentally > have their RGB triplets all have equal values. But that is just trickery -- > it's still a color image, just a very, very gray-looking one because of the > special choice of which colors are in the palette. > > I could imagine some surgery of the BMP reader that would make it operate as > follows: The palette is examined for whether every palette entry has the > property that R==G==B, and if so, it could report the file as being a true > 1-channel image and the palette conversion for pixels can just be result = > palette[pixelval][0]. If any of the palette entries have R,G,B not all equal, > then it reports back as if it were a RGB 3-channel image as before. This is > kind of a hack -- it's taking that the BMP space says is a RGB image and > lying to report it back as at 1-channel grayscale, just because "all the RGB > colors in that image look gray enough to me". Still, it's a neat idea, and I > would welcome a patch from anybody who would make this change, but I don't > have time at the moment to do it myself. > > I always thought that BMP was primarily for low-res old Windows images such > as file icons and such, and other stuff from back when images were assembled > by horse-drawn wagons. I'm a little surprised to hear that anybody is dealing > with a MILLION bmp images at 4k resolution. Can I ask what is the application > of this and whether it might be better to use a different file format? Maybe > one that not only has a proper grayscale, but also has better compression > methods than are supported by BMP? > > -- lg > > > > On Apr 5, 2021, at 3:11 PM, Larry Gritz <l...@larrygritz.com > <mailto:l...@larrygritz.com>> wrote: > > Oh, bmp specifically, I get it. Hang on, let me look at that case. > > > > On April 5, 2021 2:53:56 p.m. PDT, Phil Miller <philip.mil...@sri.com > <mailto:philip.mil...@sri.com>> wrote: > No, when I was debugging, the spec comes out with nchannels=3 instead of 1. > Let me extract some sample reader code that I can share and I will get back > to you. > > From: Oiio-dev <oiio-dev-boun...@lists.openimageio.org > <mailto:oiio-dev-boun...@lists.openimageio.org>> On Behalf Of Larry Gritz > Sent: Monday, April 5, 2021 2:58 PM > To: OpenImageIO dev list <oiio-dev@lists.openimageio.org > <mailto:oiio-dev@lists.openimageio.org>> > Subject: [EXTERNAL] Re: [Oiio-dev] Reading single channel bmp > > There is logic in the TextureSystem so that individual texture lookups will > expand a 1-channel greyscale image to all 3 channels of a 3-channel lookup > (versus using the "fill" color for the second two channels). > > Even when using that TextureSystem option, it's only about what data value it > copies into a user result for that individual texture lookup. It doesn't > actually duplicate any image data in memory in the cache. > > But there is nothing in ImageCache or ImageBuf, as far as I can tell, that > does anything like that. An ImageBuf that references a 1-channel file should > just report as 1 channel. > > Are you doing something that indicates that it's actually replicating the > data? Or are you just worried, based on this option of the texture system, > that something weird might end up happening for an ordinary ImageBuf? > > > > > On Apr 5, 2021, at 11:30 AM, Phil Miller <philip.mil...@sri.com > <mailto:philip.mil...@sri.com>> wrote: > > I am having trouble using an ImageBuf with an ImageCache to read a single > channel (grayscale/monochrome) bmp image as a single channel image. The spec > in the internal oiio logic seems to always expand to 3 channels. I can use > the force flag when I read it, but then that bypasses the cache. The images > are fairly large (4k x 3k) and I need to read a million of them, so the > difference between one and three channels is significant. > > I appreciate any help and guidance. > > Sincerely, > Phil > _______________________________________________ > Oiio-dev mailing list > Oiio-dev@lists.openimageio.org <mailto:Oiio-dev@lists.openimageio.org> > http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org > <https://urldefense.us/v3/__http:/lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org__;!!Nv3xtKNH_4uope0!xOYMMZQ_G451K0HfPRe1lqeW_Z4EG0Oep39UNfj0cKUGt0-jxrGFh-9CA29gUrlK$> > > -- > Larry Gritz > l...@larrygritz.com <mailto:l...@larrygritz.com> > > > > > > -- > Larry Gritz > l...@larrygritz.com <mailto:l...@larrygritz.com> > _______________________________________________ > Oiio-dev mailing list > Oiio-dev@lists.openimageio.org <mailto:Oiio-dev@lists.openimageio.org> > http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org > <https://urldefense.us/v3/__http:/lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org__;!!Nv3xtKNH_4uope0!0-H9hVMU1slRpUi7NepKlIBTy-jWTsIhbzrnS5Bn8mOuMUAKBqCUcUgc2CfVaPes$> > > -- > Larry Gritz > l...@larrygritz.com <mailto:l...@larrygritz.com> > > > > > _______________________________________________ > Oiio-dev mailing list > Oiio-dev@lists.openimageio.org <mailto:Oiio-dev@lists.openimageio.org> > http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org > <http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org> -- Larry Gritz l...@larrygritz.com
_______________________________________________ Oiio-dev mailing list Oiio-dev@lists.openimageio.org http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org
