#5772 closed defect (fixed)
OpenCL not working with Nvidia GRID K520
| Reported by: | mojodna | Owned by: | warmerdam |
|---|---|---|---|
| Priority: | normal | Milestone: | 1.11.2 |
| Component: | Algorithms | Version: | unspecified |
| Severity: | normal | Keywords: | opencl warp |
| Cc: |
Description
GDAL built per https://gist.github.com/mojodna/6f3d8eaa56fffa687a98 using https://aws.amazon.com/marketplace/pp/B00FYCDDTE on a g2.2xlarge.
This is the output from:
sudo LD_LIBRARY_PATH=/usr/local/lib /usr/local/bin/gdalwarp -q -wo "USE_OPENCL=TRUE" -t_srs "EPSG:3857" -tr 10 10 -tap -multi -co "TILED=YES" -co compress=lzw -r lanczos /vsizip/data/n43w125.zip/n43w125/floatn43w125_13.flt test.tiff
(n43w125.zip is available from http://ned.stamen.com.s3.amazonaws.com/13arcsec/n43w125.zip)
ERROR 1: Error: Failed to build program executable!
Build Log:
:55:20: error: cannot decrement value of type 'float attribute((address_space(1)))'
dstPtr[iDstOffset] --;
ERROR 1: Error at file gdalwarpkernel_opencl.c line 2325: CL_BUILD_PROGRAM_FAILURE
ERROR 1: OpenCL routines reported failure (-11) on line 3250.
ERROR 1: Error: Failed to build program executable!
Build Log:
:55:20: error: cannot decrement value of type 'float attribute((address_space(1)))'
dstPtr[iDstOffset] --;
ERROR 1: Error at file gdalwarpkernel_opencl.c line 2325: CL_BUILD_PROGRAM_FAILURE ERROR 1: OpenCL routines reported failure (-11) on line 3250.
Change History (4)
comment:1 by , 9 years ago
comment:2 by , 9 years ago
Success!
Non-benchmark-worthy timing:
# OpenCL-enabled $ time $(sudo LD_LIBRARY_PATH=/usr/local/lib /usr/local/bin/gdalwarp -q -wo "USE_OPENCL=TRUE" -t_srs "EPSG:3857" -r lanczos /vsizip/n43w125.zip/n43w125/floatn43w125_13.flt test-gpu.tiff) real 0m26.536s user 0m22.908s sys 0m3.664s # single-core CPU $ time $(sudo LD_LIBRARY_PATH=/usr/local/lib /usr/local/bin/gdalwarp -q -wo "USE_OPENCL=FALSE" -t_srs "EPSG:3857" -r lanczos /vsizip/n43w125.zip/n43w125/floatn43w125_13.flt test-cpu.tiff) real 0m56.892s user 0m55.180s sys 0m1.404s # multi-processor CPU $ nproc 8 $ time $(sudo LD_LIBRARY_PATH=/usr/local/lib /usr/local/bin/gdalwarp -q -wo "USE_OPENCL=FALSE" -t_srs "EPSG:3857" -multi -r lanczos /vsizip/n43w125.zip/n43w125/floatn43w125_13.flt test-cpu-multi.tiff) real 0m45.119s user 0m55.120s sys 0m1.720s # multi-processor, OpenCL-enabled $ time $(sudo LD_LIBRARY_PATH=/usr/local/lib /usr/local/bin/gdalwarp -q -t_srs "EPSG:3857" -multi -r lanczos /vsizip/n43w125.zip/n43w125/floatn43w125_13.flt test-cpu.tiff) real 0m13.018s user 0m14.896s sys 0m3.696s
I've been using watch nvidia-smi to ensure that the GPU is actually in use.
comment:3 by , 9 years ago
| Component: | default → Algorithms |
|---|---|
| Keywords: | opencl warp added |
| Milestone: | → 1.11.2 |
| Resolution: | → fixed |
| Status: | new → closed |
For CPU-only computation, you might also want to try -wo NUM_THREADS=ALL_CPUS (see http://www.gdal.org/structGDALWarpOptions.html#a0ed77f9917bb96c7a9aabd73d4d06e08) which will spread the computation, whereas -multi only spreads I/O and computation.
trunk r28300, branches/1.11 r28301 "OpenCL warper: fix OpenCL code compilation with NVIDIA OpenCL (#5772)"
Can you try the following change:
Index: alg/gdalwarpkernel_opencl.c =================================================================== --- alg/gdalwarpkernel_opencl.c (révision 28173) +++ alg/gdalwarpkernel_opencl.c (copie de travail) @@ -593,7 +593,7 @@ "if (dstPtr[iDstOffset] == dstMinVal)\n" "dstPtr[iDstOffset] = dstMinVal + 1;\n" "else\n" - "dstPtr[iDstOffset] --;\n" + "dstPtr[iDstOffset] = dstPtr[iDstOffset] - 1;\n" "}\n" "}\n" "#endif\n"