[clfft] 32/109: Workaround for 2D FFT failures on NVIDIA GPUs

Jérôme Kieffer kieffer-guest at moszumanska.debian.org
Wed May 20 07:29:23 UTC 2015


This is an automated email from the git hooks/post-receive script.

kieffer-guest pushed a commit to branch debian
in repository clfft.

commit 3c94e569e92826a6fdd23e489c2210f80ed1fed8
Author: Pavan Yalamanchili <contact at pavanky.com>
Date:   Thu Oct 31 17:06:10 2013 -0400

    Workaround for 2D FFT failures on NVIDIA GPUs
    
    - Issue: https://github.com/clMathLibraries/clFFT/issues/25
    - Inplace transpose were being used for power of 2 dimensions
    - If device from NVIDIA, then alternative path is taken
---
 src/library/plan.cpp | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/src/library/plan.cpp b/src/library/plan.cpp
index 61ae195..b988ecd 100644
--- a/src/library/plan.cpp
+++ b/src/library/plan.cpp
@@ -1512,6 +1512,20 @@ clfftStatus	clfftBakePlan( clfftPlanHandle plHandle, cl_uint numQueues, cl_comma
 					return	CLFFT_SUCCESS;
 				}
 
+                // TODO : Check for a better way to do this.
+                bool isnvidia = false;
+                for (size_t Idx = 0; !isnvidia && Idx < numQueues; Idx++)
+                {
+                    cl_command_queue QIdx = commQueueFFT[Idx];
+                    cl_device_id Device;
+                    clGetCommandQueueInfo(QIdx, CL_QUEUE_DEVICE, sizeof(Device), &Device, NULL);
+                    char Vendor[256];
+                    clGetDeviceInfo(Device, CL_DEVICE_VENDOR, sizeof(Vendor), &Vendor, NULL);
+                    isnvidia |= (strncmp(Vendor, "NVIDIA", 6) == 0);
+                }
+                // nvidia gpus are failing when doing transpose for 2D FFTs
+                if (isnvidia) break;
+
 				if (fftPlan->length.size() != 2) break;
 				if (!(IsPo2(fftPlan->length[0])) || !(IsPo2(fftPlan->length[1])))
 					break;

-- 
Alioth's /usr/local/bin/git-commit-notice on /srv/git.debian.org/git/debian-science/packages/clfft.git



More information about the debian-science-commits mailing list