Recursion problem

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, I have a recursive function that I am trying to parallelize. I have a C2050 and cuda toolkit 3.1. I’m compiling using

nvcc -arch=sm_20 -o out in.cu

If I use cuda-gdb (and compile with -g -G), whenever I step into the line where the recursive function calls itself, I get the following warnings:

approximateMatchRecursive (approxRes=warning: Variable stored in multiple registers. Not supported in GDB yet. Returning garbage value.
0xfffd1000000004, approxResCount=warning: Variable stored in multiple registers. Not supported in GDB yet. Returning garbage value.
0xfffd1000000004,
maxMatches=50, query=warning: Variable stored in multiple registers. Not supported in GDB yet. Returning garbage value.
0xfffd1000000004 <Address 0xfffd1000000004 out of bounds>,
errsLeft=0, loc=3, lo=0, hi=22, rankArr=warning: Variable stored in multiple registers. Not supported in GDB yet. Returning garbage value.
0xfffd1000000004, bwArr=warning: Variable stored in multiple registers. Not supported in GDB yet. Returning garbage value.

Running the program with cuda-memcheck doesn’t report any problems. Any fixes/suggestions?

2

Looks like recursion isnt supported by cuda-gdb, yet. It has just been introduced in CUDA 3.1 for Fermi-cards.

3

Looks like recursion isnt supported by cuda-gdb, yet. It has just been introduced in CUDA 3.1 for Fermi-cards.

4

So the problem must be a bug in the code? Too bad; I thought there might be a quick fix External Media

5

So the problem must be a bug in the code? Too bad; I thought there might be a quick fix External Media

6

Nope. Gdb clearly says that sth isnt supported by gdb, yet. As recursion is supported in CUDA 3.1 with Fermi cards, and you r running a Fermi card with CUDA 3.1, your code might be ok.

7

Nope. Gdb clearly says that sth isnt supported by gdb, yet. As recursion is supported in CUDA 3.1 with Fermi cards, and you r running a Fermi card with CUDA 3.1, your code might be ok.