cuTENSOR: Destroy plan description and preference after construction.

[maleadt]

Should fix https://github.com/JuliaGPU/CUDA.jl/issues

@lkdevos Can you confirm?

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic master) to apply these changes.

Click here to view the suggested changes.

diff --git a/lib/cutensor/src/operations.jl b/lib/cutensor/src/operations.jl
index 0c4f3d276..6dea1325b 100644
--- a/lib/cutensor/src/operations.jl
+++ b/lib/cutensor/src/operations.jl
@@ -104,7 +104,7 @@ function plan_elementwise_trinary(
     plan_pref = Ref{cutensorPlanPreference_t}()
     cutensorCreatePlanPreference(handle(), plan_pref, algo, jit)
 
-    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref=workspace)
+    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref = workspace)
     cutensorDestroyOperationDescriptor(desc[])
     cutensorDestroyPlanPreference(plan_pref[])
     return plan
@@ -189,7 +189,7 @@ function plan_elementwise_binary(
     plan_pref = Ref{cutensorPlanPreference_t}()
     cutensorCreatePlanPreference(handle(), plan_pref, algo, jit)
 
-    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref=workspace)
+    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref = workspace)
     cutensorDestroyOperationDescriptor(desc[])
     cutensorDestroyPlanPreference(plan_pref[])
     return plan
@@ -259,7 +259,7 @@ function plan_permutation(
     plan_pref = Ref{cutensorPlanPreference_t}()
     cutensorCreatePlanPreference(handle(), plan_pref, algo, jit)
 
-    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref=workspace)
+    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref = workspace)
     cutensorDestroyOperationDescriptor(desc[])
     cutensorDestroyPlanPreference(plan_pref[])
     return plan
@@ -349,7 +349,7 @@ function plan_contraction(
     plan_pref = Ref{cutensorPlanPreference_t}()
     cutensorCreatePlanPreference(handle(), plan_pref, algo, jit)
 
-    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref=workspace)
+    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref = workspace)
     cutensorDestroyOperationDescriptor(desc[])
     cutensorDestroyPlanPreference(plan_pref[])
     return plan
@@ -427,7 +427,7 @@ function plan_reduction(
     plan_pref = Ref{cutensorPlanPreference_t}()
     cutensorCreatePlanPreference(handle(), plan_pref, algo, jit)
 
-    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref=workspace)
+    plan = CuTensorPlan(desc[], plan_pref[]; workspacePref = workspace)
     cutensorDestroyOperationDescriptor(desc[])
     cutensorDestroyPlanPreference(plan_pref[])
     return plan

[kshyatt]

This should fix things but I wonder if we should add an option to pass a plan in so that users can re-use a contraction plan?

[maleadt]

Probably, the API is currently pretty happy to re-create a bunch of state over and over.

For now this is the quick fix though. Feel free to add options; maybe @lkdvos can suggest some based on his usage in TensorOperations.jl.

[kshyatt]

ITensor uses these a lot as well, cc @mtfishman

[mtfishman]

I don't think we're using that right now.

[lkdvos]

Thanks for investigating and fixing this. I ran the script again and the memory differences now look more like reasonable fluctuations because of other processes on my machine. (note difference in magnitude as well)

---

At least for TensorOperations, plans is not something I ever made use of either.
I didn't see an obvious way of making use of this through @tensor, so I was just creating a plan for every single (binary) contraction/permutation (hooked into the functions without plan argument).
I guess this is also what made the memory loss show up :).

I would argue that if you really want to have the control of reusing the plans etc, you could technically do that already, possibly having to re-wrap implementations a little.
Perhaps if you care about that level of customization/performance this is a fair tradeoff though?

Anyways, looks resolved to me, thank you!