minutes of requirements meeting on single vs. multiple CX state contexts

meetings/2024-08-13-minutes.md

@@ -0,0 +1,52 @@
+# Meeting Minutes 2024-08-13
+
+- CX TG Requirements Ad-Hoc Meeting: Single or Multiple Contexts
+
+- Attendees
+
+ Darius Rad (Bluespec)
+ Jan Gray (Individual)
+ Guy Lemieux (Individual)
+ Victor Lu (Individual)
+
+- Discussion
+
+Jan proposed meeting rules, objectives (better mutual understanding of
+the proposals, but not a decision).
+
+Guy asked for clarification about the specific topic of the discussion.
+Darius stated that the topic is multiple potential requirements related
+to the general concept of one or multiple state contexts per hart,
+and not a simple binary question.
+
+Jan presented his position that multiple state contexts per hart is a more
+novel, more complex, but more flexible, more performant, affordable, and
+preferable approach to achieve the goals of the task group. His position
+was presented from the perspective of the user programming model and
+suggested other perspectives be clear about which level of abstraction
+they are referring to. He described a number of use cases in support
+of this position.
+
+Guy questioned whether objection to multiple state contexts is due to
+memory table mechanism in basis specification, for which he is designing
+an alternative.
+
+Guy proposed the following requirements:
+* Must efficiently support hardware implementation for single context per hart
+* Should efficiently support hardware implementation for multiple contexts per hart
+* Must convey a single programming model in either case
+* Must support >1 context per thread without context switching
+
+Darius presented his position that a single state context per hart is more
+consistent with existing RISC-V extensions and that existing mechanisms
+for context management in the operating system and in user level software
+may be utilized to achieve the goals of the task group. He indicated
+that the behavior of one state context per hart is more consistent with
+the expected programming model at the ISA level in all privilege modes.
+
+There was discussion regarding whether a single state per hart or multiple
+states per hart need to be mutually exclusive requirements.
+
+There was discussion regarding whether extension state should be
+considered hart state, or not, or if both behaviors could be or should
+be supported.