Document NVIDIA OpenCL broken on Volta and Turing

Refs #3125

Change-Id: Id200c1092b279d45a97b98c6fe1d03c2ffb14fb0

diff --git a/docs/dev-manual/known-issues.rst b/docs/dev-manual/known-issues.rst
index 50c468210e854dbf3f59c3b212adbe5e0eab7458..1d97c22e85643081990e1beb50afdc3a6cffe922 100644 (file)
--- a/docs/dev-manual/known-issues.rst
+++ b/docs/dev-manual/known-issues.rst
@@ -21,3 +21,11 @@ GPU emulation does not work
 The non-bonded GPU emulation mode does not work, at least for builds
 with GPU support; then a GPU setup call is called.
 Also dynamic pruning needs to be implemented for GPU emulation.
+
+OpenCL on NVIDIA Volta and later broken
+---------------------------------------
+
+The OpenCL code produces incorrect results on Volta and Turing GPU architectures
+from NVIDIA (CC 7.0 and 7.5). This is an issue that affects certain flavors of 
+the nonboded kernels, most likely a result of miscompilation, and there is no
+known workaround.

diff --git a/docs/release-notes/2020/2020.1.rst b/docs/release-notes/2020/2020.1.rst
index 3d28b5c48c65c5c0dba09e4f0408f818901154a8..fe2d306396b205dabf17e3effa0f8ce0a0484ac2 100644 (file)
--- a/docs/release-notes/2020/2020.1.rst
+++ b/docs/release-notes/2020/2020.1.rst
@@ -143,6 +143,10 @@ Add support for ICC NextGen
 Add support for Intel Compiler based on LLVM technology.
 To compile GROMACS with this compiler use ``CXX=icpc CXXFLAGS=-qnextgen cmake``.
 
+Document known issues with OpenCL on Volta and Turing
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+:issue:`3125`
+
 Miscellaneous
 ^^^^^^^^^^^^^
-

diff --git a/docs/user-guide/mdrun-performance.rst b/docs/user-guide/mdrun-performance.rst
index dbe5dcfcbd0d83578efb8410bd47c730e87d40ce..322eb94be50d2733830145ecde729bb939070bfa 100644 (file)
--- a/docs/user-guide/mdrun-performance.rst
+++ b/docs/user-guide/mdrun-performance.rst
@@ -1322,6 +1322,9 @@ Limitations in the current OpenCL support of interest to |Gromacs| users:
 - On NVIDIA GPUs the OpenCL kernels achieve much lower performance
   than the equivalent CUDA kernels due to limitations of the NVIDIA OpenCL
   compiler.
+- On the NVIDIA Volta an Turing architectures the OpenCL code is known to produce
+  incorrect results with driver version up to 440.x (most likely due to compiler issues).
+  Runs typically fail on these architectures.
 
 Limitations of interest to |Gromacs| developers: