[GIS] GDAL usage of GPU with Windows

gdalgdal-translategdalwarp

I use GDAL 2.2.1, released 2017/06/23 on Windows 10. It was install with pre-compiled packages. I usually use gdalwarp, gdal_translate or gdal_grid as commands for single use, or with Python bindings for larger programs.

I haven't found a way to use Windows GPU, even with a few options I try such as -wo "USE_OPENCL=TRUE". The graph on my GPU stays still and flat (0.0%). Here is an example of code to resample an image:

gdalwarp -r cubicspline -tr 4.5e-5 4.5e-5 -ot Int16 -te  3 41 8 45 -multi -wo "USE_OPENCL=TRUE" DEM_75m.tif DEM_5m.tif

During the processing, RAM is used at 200Mo, 20% of CPU and of course 100% of hard-drive usage. 0% GPU.

Should I re-install GDAL with new build parameters, or is there a way of using it as it is?

Best Answer

Sounds like you need to recompile gdal with opencl / cuda. You will need the cuda toolkit installed. Once thats done, your call to gdalwarp looks correct. Even once it's working you'll likely see that there is a lot of time spent transferring the data to the GPU and back again.

Related Solutions

[GIS] gdal_array.SaveArray() leaves dataset open in Python

Edit based on the comments below:

Assigning the gdal_array.SaveArray(a, "test.tif") call to a variable returns an osgeo.gdal.Dataset object that can be managed as a per the below gotchas. Using the above example this should work:

    a = np.arange(300).reshape((3, 10, 10))
    ds = gdal_array.SaveArray(a, "test.tif")
    ds = None
    os.rename("test.tif", "test2.tif")

Checkout the Python gotchas documentation: https://trac.osgeo.org/gdal/wiki/PythonGotchas

Specifically:

Saving and closing datasets/datasources

To save and close GDAL raster datasets or OGR vector datasources, the object needs to be dereferenced, such as setting it to None, a different value, or deleting the object. If there are more than one copies of the dataset or datasource object, then each copy needs to be dereferenced.

For example, creating and saving a raster dataset:
>>> from osgeo import gdal
>>> driver = gdal.GetDriverByName('GTiff')
>>> dst_ds = driver.Create('new.tif', 10, 15)
>>> band = dst_ds.GetRasterBand(1)
>>> arr = band.ReadAsArray()  # raster values are all zero
>>> arr[2, 4:] = 50  # modify some data
>>> band.WriteArray(arr)  # raster file still unmodified
>>> band = None  # dereference band to avoid gotcha described previously
>>> dst_ds = None  # save, close
The last dereference to the raster dataset writes the data modifications and closes the raster file. WriteArray(arr) does not write the array to disk, unless the GDAL block cache is full (typically 40 MB).

With some drivers, raster datasets can be intermittently saved without closing using FlushCache(). Similarly, vector datasets can be saved using SyncToDisk(). However, neither of these methods guarantee that the data are written to disk, so the preferred method is to deallocate as shown above.

GDAL – Efficient Downsampling with GDAL

Try gdalwarp, it can utilize multithreading and might be more efficiant than gdal_translate for this kind of tasks:

gdalwarp -r average -ts 1295 464 \ 
    -wm 2048 -multi -wo NUM_THREADS=ALL_CPUS \
    -co BLOCKXSIZE=256 -co BLOCKYSIZE=256 -co TILED=YES \
    -co NUM_THREADS=ALL_CPUS -co COMPRESS=LZW \
    world.tif downsized_world.tif

Although average will be slower, nearest Interpolation is not suitable to be used with DEM data.

Change the used memory to fit your system, it's this option: -wm 2048.

Best Answer

Related Solutions

[GIS] gdal_array.SaveArray() leaves dataset open in Python

GDAL – Efficient Downsampling with GDAL

Related Question