NvMon segfault on stop_counters #34
Labels
bug
Something isn't working
Comments
|
FWIW, gdb backtrace |
|
Perhaps related: When I try to use the GPU Marker API instead (see code below) I get the following error. All output (with Command: Julia File: # perfctr_gpu.jl
using LIKWID
using LinearAlgebra
using CUDA
@assert CUDA.functional()
const N = 10_000
const a = 3.141
# Note: CUDA defaults to Float32
const x = CUDA.rand(N)
const y = CUDA.rand(N)
const z = CUDA.zeros(N)
GPUMarker.init()
GPUMarker.startregion("saxpy")
for _ in 1:100
z .= a .* x .* y
end
GPUMarker.stopregion("saxpy")
GPUMarker.close() |
|
I get the same segfault (see first post) on an entirely different machine (our DGX-A100). To me, that seems to suggest that this is not system / installation related but actually a LIKWID bug. @TomTheBear what do you think? |
|
I also don't understand the |
|
@TomTheBear With your PR the segfault is gone but, I believe, things still aren't working as they should.
|
|
Oh, regarding point 2, it shows up under "FLOPS_DP", but why?! Update: Because julia> a = 3.141f0;
julia> events = @nvmon "FLOPS_SP" saxpy!(z, a, x, y);
Group: FLOPS_SP
┌────────────────────────────────────────────────────┬─────────┐
│ Event │ GPU 1 │
├────────────────────────────────────────────────────┼─────────┤
│ SMSP_SASS_THREAD_INST_EXECUTED_OP_FADD_PRED_ON_SUM │ 0.0 │
│ SMSP_SASS_THREAD_INST_EXECUTED_OP_FMUL_PRED_ON_SUM │ 0.0 │
│ SMSP_SASS_THREAD_INST_EXECUTED_OP_FFMA_PRED_ON_SUM │ 10000.0 │
└────────────────────────────────────────────────────┴─────────┘ |
|
With the latest fixes on the likwid PR this works again. Hope we'll get a new likwid release with this soon 😀 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
To reproduce / for copy-paste:
Not sure yet whether this is an issue with likwid, LIKWID.jl, or the likwid installation on Noctua 2.
(cc @TomTheBear)
The text was updated successfully, but these errors were encountered: