no speedup from coalescing global reads?! Surprising profile results

Accelerated Computing CUDA CUDA Programming and Performance
1

I recently reordered some data in my CUDA program so that my read from global memory would be coalesced. I believe I have nearly all reads coalesced properly, as evidenced by the profile output below:

gputime: 43422.2
gld_incoherent: 746
gld_coherent: 304708
gst_incoherent: 0
gst_coherent: 215656
branch: 1451145
divergent_branch: 8397

And here is the profile output for the old version:

gputime: 43259.3
gld_incoherent: 398499
gld_coherent: 778
gst_incoherent: 55134
gst_coherent: 14728
branch: 1450768
divergent_branch: 8395

So my question is: Why hasn’t the gputime gone down? Was the latency simply hidden that well in the original version?

Thanks in advance.

2

it looks like you have many more global stores 215k vs 70k, that may be a reason
another possible reason is that your kernel is compute bound, not memory bound