www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.ldc - dcompute issues?

reply Bruce Carneal <bcarneal gmail.com> writes:
I'm trying to get the dcompute saxpy test program to run and am 
having some issues.

The first issue was an ldc2 1.24.0 seg fault over this line:
 compute(CompileFor.deviceOnly) module ...

I commented out the  compute... annotation leaving the top line 
module declaration after which the test program compiled and 
linked but it errored out when attempting to queue up the saxpy 
kernel itself. Everything up to the queue attempt went well: 
Platform initialization, getDevices, Context, Buffer! 
allocations, data copies, ... so it appears that the 
libdcompute.a build is fine.

I examined the newly created kernels_cuda210_64.ptx file and did 
not find saxpy code.  I only found an 8-liner that, from the 
mangled name, looks like a GlobalIndex calculation.

If this problem rings a bell with anyone in the LDC community, 
please chime in.  I'd *really* prefer to work with dcompute 
rather than fall back to CUDA shimming or SycL exploration.

Finally, many thanks for maintaining/advancing LDC.  It has been 
great for all of my CPU/SIMD work.
Feb 09 2021
parent reply kinke <noone nowhere.com> writes:
On Wednesday, 10 February 2021 at 02:44:37 UTC, Bruce Carneal 
wrote:

 The first issue was an ldc2 1.24.0 seg fault over this line:
  compute(CompileFor.deviceOnly) module ...


Please try to reduce it and file an issue. You can also use a CI 
build with enabled assertions from 
https://github.com/ldc-developers/ldc/releases/tag/CI, which 
might give a usable hint.


 Finally, many thanks for maintaining/advancing LDC.  It has 
 been great for all of my CPU/SIMD work.


Thanks, you're welcome.
Feb 10 2021
next sibling parent Bruce Carneal <bcarneal gmail.com> writes:
On Wednesday, 10 February 2021 at 17:54:44 UTC, kinke wrote:

 On Wednesday, 10 February 2021 at 02:44:37 UTC, Bruce Carneal 
 wrote:

 The first issue was an ldc2 1.24.0 seg fault over this line:
  compute(CompileFor.deviceOnly) module ...


 Please try to reduce it and file an issue. You can also use a 
 CI build with enabled assertions from 
 https://github.com/ldc-developers/ldc/releases/tag/CI, which 
 might give a usable hint.


 Finally, many thanks for maintaining/advancing LDC.  It has 
 been great for all of my CPU/SIMD work.


 Thanks, you're welcome.


I've narrowed it to a noop routine with a single 
GlobalPointer!(float) argument (the no-argument noop routine 
compiled and ran).

I'll pull down the assert-enabled compiler and keep digging.  
Thanks for the pointer.
Feb 10 2021
prev sibling parent reply Bruce Carneal <bcarneal gmail.com> writes:
On Wednesday, 10 February 2021 at 17:54:44 UTC, kinke wrote:

 On Wednesday, 10 February 2021 at 02:44:37 UTC, Bruce Carneal 
 wrote:

 The first issue was an ldc2 1.24.0 seg fault over this line:
  compute(CompileFor.deviceOnly) module ...


 Please try to reduce it and file an issue. You can also use a 
 CI build with enabled assertions from 
 https://github.com/ldc-developers/ldc/releases/tag/CI, which 
 might give a usable hint.


 Finally, many thanks for maintaining/advancing LDC.  It has 
 been great for all of my CPU/SIMD work.


 Thanks, you're welcome.


The ldc2-4bee4e6f-linux-x86_64 CI compiler that you pointed me to 
works, no segv during compilation and the saxpy results coming 
back from the GPU check out.  For reference, I'm using a cuda-620 
target and running on a cuda-750 card.

You've got me up and running on dcompute.  Thanks!  I'm *really* 
happy that I might not have to go back to C++/CUDA.
Feb 10 2021
parent reply Imperatorn <johan_forsberg_86 hotmail.com> writes:
On Wednesday, 10 February 2021 at 19:59:45 UTC, Bruce Carneal 
wrote:

 On Wednesday, 10 February 2021 at 17:54:44 UTC, kinke wrote:

 On Wednesday, 10 February 2021 at 02:44:37 UTC, Bruce Carneal 
 wrote:

 [...]


 Please try to reduce it and file an issue. You can also use a 
 CI build with enabled assertions from 
 https://github.com/ldc-developers/ldc/releases/tag/CI, which 
 might give a usable hint.


 [...]


 Thanks, you're welcome.


 The ldc2-4bee4e6f-linux-x86_64 CI compiler that you pointed me 
 to works, no segv during compilation and the saxpy results 
 coming back from the GPU check out.  For reference, I'm using a 
 cuda-620 target and running on a cuda-750 card.

 You've got me up and running on dcompute.  Thanks!  I'm 
 *really* happy that I might not have to go back to C++/CUDA.


Nice! D compute looks interesting. Are/were there any patches for 
the problem(s) you encountered?
Feb 11 2021
parent reply Bruce Carneal <bcarneal gmail.com> writes:
On Thursday, 11 February 2021 at 19:11:28 UTC, Imperatorn wrote:

 On Wednesday, 10 February 2021 at 19:59:45 UTC, Bruce Carneal 
 wrote:

 On Wednesday, 10 February 2021 at 17:54:44 UTC, kinke wrote:

 On Wednesday, 10 February 2021 at 02:44:37 UTC, Bruce Carneal 
 wrote:

 [...]


 Please try to reduce it and file an issue. You can also use a 
 CI build with enabled assertions from 
 https://github.com/ldc-developers/ldc/releases/tag/CI, which 
 might give a usable hint.


 [...]


 Thanks, you're welcome.


 The ldc2-4bee4e6f-linux-x86_64 CI compiler that you pointed me 
 to works, no segv during compilation and the saxpy results 
 coming back from the GPU check out.  For reference, I'm using 
 a cuda-620 target and running on a cuda-750 card.

 You've got me up and running on dcompute.  Thanks!  I'm 
 *really* happy that I might not have to go back to C++/CUDA.


 Nice! D compute looks interesting. Are/were there any patches 
 for the problem(s) you encountered?


No patches.  Once I installed the CI compiler the saxpy demo/test 
"just worked".

There are some, documented, "TODOs" in the very readable dcompute 
source code but no show stoppers so far.  derelict CUDA/CL are 
also available if you need them.
Feb 11 2021
parent Imperatorn <johan_forsberg_86 hotmail.com> writes:
On Friday, 12 February 2021 at 06:39:05 UTC, Bruce Carneal wrote:

 On Thursday, 11 February 2021 at 19:11:28 UTC, Imperatorn wrote:

 On Wednesday, 10 February 2021 at 19:59:45 UTC, Bruce Carneal 
 wrote:

 [...]


 Nice! D compute looks interesting. Are/were there any patches 
 for the problem(s) you encountered?


 No patches.  Once I installed the CI compiler the saxpy 
 demo/test "just worked".

 There are some, documented, "TODOs" in the very readable 
 dcompute source code but no show stoppers so far.  derelict 
 CUDA/CL are also available if you need them.


Ok, good to hear 👍
Feb 11 2021