something wrong with cuda visual profiler

dear all:

I use cuda 2.3 in winxp 64, but parameters, gld_uncoalesced, gld_coalesced, gst_uncoalesced and gst_coalesced in cuda visual profiler are disabled, see figure 1

figure 1,

How can I use these parameters?

with cards < compute 1.2 these are visible, and the ones below are grayed out.


dear cbuchner1: thanks for your comment, I notice your comment on another thread,…rt=#entry603122

you said “My compute 1.2 card shows me the 32/64/128 bit load/store counters (which represent coalesced loads of various bit widths)”.

if “32/64/128 bit load/store” is coalesced, how to count number of non-coalesced under knowledge of

  1. gld request : Number of global memory load requests (available only for compute capability 1.2 or higher)

  2. gld_32/64/128b : Number of 32 byte, 64 byte and 128 byte global memory load transactions (available only for compute capability 1.2 or higher)

for example:

gld request = 280

gld 64b = 1648

no gld 128, and gld 32

I believe the gld request represents uncoalesced loads only. I think I’ve seen this counter as 0, whereas I had plenty of gld_64 in my matrix multiplication code.

But please refer to the reference documentation for reliable information.