[QGIS-Developer] How to speed up rendering

afernandez afernandez at odyhpc.com
Tue May 30 05:37:34 PDT 2023


Hello Hannes,
Sorry to say that I'm at a loss here. I had not previously heard of 'overviews' and am having a hard time finding much information. I couldn't find anything in the PyQGIS (v3.28) developer cookbook, which is my main reference in this QGIS journey. The found information is at https://docs.qgis.org/3.28/en/docs/user_manual/processing_algs/gdal/rastermiscellaneous.html but didn't exactly clarify how to implement overviews with a script or why this would sidestep reading the file several times. I don't quite understand this last point either, as the file is large but the size of the band (after everything is prepped and before rendering) is relatively small, and should fit amply into the available memory.
Thanks,
Arturo
Johannes Kröger (WhereGroup) via QGIS-Developer wrote:
Hi Arturo,
make sure you create overviews on your big rasters or QGIS will re-read the whole file all the time to render an appropriate representation in the map.
Cheers, Hannes
Am 30.05.23 um 04:35 schrieb afernandez via QGIS-Developer:
Hello Martin,
I'm trying to understand what you're saying but my background is in numerical methods where tasks are split into threads for acceleration purposes. If a single layer is rendered by a single core (as you write and I was observing), how would a 1,000 x 1,000 (or a 5,000 x 5,000) layer be handled? Wouldn't it take too long? (However, you want to define 'too long'). 
As far as more details about the process, the tensor 'var' (n x m x index where I have tried different combinations but let's say n x m is 100 x 200) of floating point numbers is manipulated according to the following pseudo-code:
# Initial manipulations
dims = var.dimensions
shape = var.shape
driver_name = 'GTIFF'
driver = gdal.GetDriverByName(driver_name) 
np_dtype = var.dtype
type_code = gdal_array.NumericTypeCodeToGDALTypeCode(np_dtype)
gdal_ds = driver.Create(_my_path_, cols, rows, 1, gdal.GDT_UInt16) 
gdal_ds.SetProjection(_my_projection_)
gdal_ds.SetGeoTransform(_my_transformation_)
# Creation of the bands and scaled matrix 
band = gdal_ds.GetRasterBand(1) 
data = var[_chosen_index_]
data = ma.getdata(data)
data_scaled = np.interp(data, (data.min(), data.max()), (0, 255))
data_scaled2 = data_scaled.astype(int) # This is to rescale into integers so that it can color the layer
# *** Lines to set up the color palette ***
# Write the array to band once everything has been rescaled
band.WriteArray(data_scaled2)
gdal_ds.FlushCache() 
This procedure seems to work well and takes a reasonable amount of time. Then, the time is consumed when doing the rendering itself. I'm not entirely sure what you mean by 'sub-optimal raster format' or 'thre are overviews missing' so I cannot address them specifically.
Thanks,
Arturo
Martin Dobias wrote:
Hi
On Fri, May 26, 2023 at 11:30 PM afernandez via QGIS-Developer <qgis-developer at lists.osgeo.org <mailto:qgis-developer at lists.osgeo.org> > wrote:
Hello,
One of my project would greatly benefit from a much faster rendering. After some testing, the bulk (>99%) of the time is spent at rendering the layer with QgsRasterLayer(...,...). What I have also noticed is that rendering time does not change with the number of cores, which suggests that the renderer might be using a single thread. I checked the API documentation but couldn't find anything about multithreading or some other procedure to accelerate rendering. I was wondering if anyone has any suggestion or has faced a similar difficulty. 
Multi-threaded rendering is done at the granularity of whole map layers, so a single layer is rendered by a single CPU core. Therefore if you have just one layer that is slow to render, the multi-threading is not going to help. It is best to do some testing how you can improve the layer's rendering - for example, it could be that you are using some sub-optimal raster format, or there are overviews missing, or something else... one would need more details.
Regards
Martin
_______________________________________________ QGIS-Developer mailing list QGIS-Developer at lists.osgeo.org <mailto:QGIS-Developer at lists.osgeo.org> List info: https://lists.osgeo.org/mailman/listinfo/qgis-developer <https://lists.osgeo.org/mailman/listinfo/qgis-developer> Unsubscribe: https://lists.osgeo.org/mailman/listinfo/qgis-developer <https://lists.osgeo.org/mailman/listinfo/qgis-developer> -- Johannes Kröger / GIS-Entwickler/-Berater --------------------------------------------- Aufwind durch Wissen! Web-Seminare und Online-Schulungen bei der www.foss-academy.com <http://www.foss-academy.com> --------------------------------------------- WhereGroup GmbH c/o KK03 GmbH Lange Reihe 29 20099 Hamburg Germany Tel: +49 (0)228 / 90 90 38 - 36 Fax: +49 (0)228 / 90 90 38 - 11 johannes.kroeger at wheregroup.com <mailto:johannes.kroeger at wheregroup.com> www.wheregroup.com <http://www.wheregroup.com> Geschäftsführer: Olaf Knopp, Peter Stamm Amtsgericht Bonn, HRB 9885 ------------------------------- 
Hi Arturo,
make sure you create overviews on your big rasters or QGIS will re-read the whole file all the time to render an appropriate representation in the map.
Cheers, Hannes
Am 30.05.23 um 04:35 schrieb afernandez via QGIS-Developer:
Hello Martin,
I'm trying to understand what you're saying but my background is in numerical methods where tasks are split into threads for acceleration purposes. If a single layer is rendered by a single core (as you write and I was observing), how would a 1,000 x 1,000 (or a 5,000 x 5,000) layer be handled? Wouldn't it take too long? (However, you want to define 'too long'). 
As far as more details about the process, the tensor 'var' (n x m x index where I have tried different combinations but let's say n x m is 100 x 200) of floating point numbers is manipulated according to the following pseudo-code:
# Initial manipulations
dims = var.dimensions
shape = var.shape
driver_name = 'GTIFF'
driver = gdal.GetDriverByName(driver_name) 
np_dtype = var.dtype
type_code = gdal_array.NumericTypeCodeToGDALTypeCode(np_dtype)
gdal_ds = driver.Create(_my_path_, cols, rows, 1, gdal.GDT_UInt16) 
gdal_ds.SetProjection(_my_projection_)
gdal_ds.SetGeoTransform(_my_transformation_)
# Creation of the bands and scaled matrix 
band = gdal_ds.GetRasterBand(1) 
data = var[_chosen_index_]
data = ma.getdata(data)
data_scaled = np.interp(data, (data.min(), data.max()), (0, 255))
data_scaled2 = data_scaled.astype(int) # This is to rescale into integers so that it can color the layer
# *** Lines to set up the color palette ***
# Write the array to band once everything has been rescaled
band.WriteArray(data_scaled2)
gdal_ds.FlushCache() 
This procedure seems to work well and takes a reasonable amount of time. Then, the time is consumed when doing the rendering itself. I'm not entirely sure what you mean by 'sub-optimal raster format' or 'thre are overviews missing' so I cannot address them specifically.
Thanks,
Arturo
Martin Dobias wrote:
Hi
On Fri, May 26, 2023 at 11:30 PM afernandez via QGIS-Developer <qgis-developer at lists.osgeo.org <mailto:qgis-developer at lists.osgeo.org> > wrote:
Hello,
One of my project would greatly benefit from a much faster rendering. After some testing, the bulk (>99%) of the time is spent at rendering the layer with QgsRasterLayer(...,...). What I have also noticed is that rendering time does not change with the number of cores, which suggests that the renderer might be using a single thread. I checked the API documentation but couldn't find anything about multithreading or some other procedure to accelerate rendering. I was wondering if anyone has any suggestion or has faced a similar difficulty. 
Multi-threaded rendering is done at the granularity of whole map layers, so a single layer is rendered by a single CPU core. Therefore if you have just one layer that is slow to render, the multi-threading is not going to help. It is best to do some testing how you can improve the layer's rendering - for example, it could be that you are using some sub-optimal raster format, or there are overviews missing, or something else... one would need more details.
Regards
Martin
_______________________________________________ QGIS-Developer mailing list QGIS-Developer at lists.osgeo.org <mailto:QGIS-Developer at lists.osgeo.org> List info: https://lists.osgeo.org/mailman/listinfo/qgis-developer <https://lists.osgeo.org/mailman/listinfo/qgis-developer> Unsubscribe: https://lists.osgeo.org/mailman/listinfo/qgis-developer <https://lists.osgeo.org/mailman/listinfo/qgis-developer> -- Johannes Kröger / GIS-Entwickler/-Berater --------------------------------------------- Aufwind durch Wissen! Web-Seminare und Online-Schulungen bei der www.foss-academy.com <http://www.foss-academy.com> --------------------------------------------- WhereGroup GmbH c/o KK03 GmbH Lange Reihe 29 20099 Hamburg Germany Tel: +49 (0)228 / 90 90 38 - 36 Fax: +49 (0)228 / 90 90 38 - 11 johannes.kroeger at wheregroup.com <mailto:johannes.kroeger at wheregroup.com> www.wheregroup.com <http://www.wheregroup.com> Geschäftsführer: Olaf Knopp, Peter Stamm Amtsgericht Bonn, HRB 9885 ------------------------------- 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/qgis-developer/attachments/20230530/48c434ed/attachment-0001.htm>


More information about the QGIS-Developer mailing list