[INF] Linear Algebra Implementation Plan

proposals/infra/INF-XXXX-Linear-Algebra-Implementation-Plan.md

@@ -0,0 +1,111 @@
+---
+title: "INF-XXXX - Linear Algebra Implementation Plan"
+params:
+  authors:
+    - V-FEXrt: Ashley Coleman
+  sponsors:
+    - V-FEXrt: Ashley Coleman
+  status: Under Consideration
+---
+
+
+* Impacted Projects: DXC, Clang implementation not considered here
+
+## Introduction
+
+This implementation plan covers a high level description of the work necessary
+to implement the
+[linalg spec](https://github.com/microsoft/hlsl-specs/blob/main/proposals/0035-linalg-matrix.md)
+for SM6.10 including back-of-the-napkin time
+estimates. This document is not intended to be a comprehensive implementation
+timeline, instead the estimates are purely for strawmanning discussion and
+should be expected to vary in at least an order of magnetude in either direction.
+
+## Time Estimate Windows
+
+In order to facilitate useful discussion, the table below defines terms for
+various low resolution time windows. Each term intentionally overlaps with
+nearby terms.
+
+| Term    | Window     |
+|---------|------------|
+| Days    | 1-10 days  |
+| Weeks   | 1-2 weeks  |
+| Sprints | 2-4 weeks  |
+| Months  | 1-3 months |
+
+Anything smalller than days is too small resolution to discuss and should be
+bundled with other tasks while anything larger than months is too large to discuss
+and should be split up.
+
+## Top Level Work Items
+
+The table below represents identified work items with top level estimates.
+See [Commentary](#Commentary) for discussion on specific work items.
+
+| Item Class | Time Window | Item Count  |
+|----------------------|---------|---|
+| Publish HLSL Headers | Days    | 1   |
+| Header Distribution  | Weeks   | 1   |
+| Groupshared Args     | Weeks   | 1   |
+| HLSL __builtins api  | Weeks   | 19* |
+| DXIL api             | Days    | 19* |
+| DXIL Validator       | Sprints | 1   |
+| HLK Testing          | Months  | 1   |
+| PSV0/RDAT Updates    | ???     | 1   |
+| Shader Reflection    | ???     | 1   |
+
+Loosely adding everything up gives ~7 "Months" of serialized work with a wide error margin.
+
+*: At the time of writing there are 19 identified matrix operations. The table assumes 1 DXIL Op and 1 HLSL __builtin per operation which is likely incorrect.
+
+## Commentary
+
+### Publish HLSL Headers
+Very trivial task, just need to make a PR with header code that already exists.
+
+### Header Distribution
+There is an open discussion on how to ship HLSL headers in DXC. Shipping HLSL
+source with the compiler is a very new thing for DXC and consideration is needed
+to make sure we have the right solution/process for users.
+
+## Groupshared Args
+See [the groupshared args proposal](https://github.com/microsoft/hlsl-specs/blob/main/proposals/0043-groupshared-arguments.md)
+
+## HLSL __builtins api
+Each function/operation in the HLSL linalg spec needs an underlying builtin
+function to call to actually do the underlying work.
+
+The work for each operation is as follows:
+ - Perform sema checks
+ - Lower the call to DXCs "High level ops"
+ - Write unit/integration level tests
+
+## DXIL api
+The DXIL api is the final api that IHVs operate over.
+
+The High Level Ops emitted by the __builtins api need to be lowered into DXIL.
+Each lowering could use any number of new DXIL ops but the average tends towards
+one so that is assumed here.
+
+The work for each HLOp give then above context is roughly:
+ - Lower the op to DXIL
+ - Create one new DXIL op
+ - Write unit/integration level tests
+
+## DXIL Validator
+I don't have strong context here, but members of the team implied that there
+is a large chunck of work here, so marked as "Sprints" kind of arbirarily.
+
+This should be revisted
+
+## HLK Testing
+Based on challenges and timelines for Long Vector HLK testing, I choose "Months" here.
+
+This should be revisted
+
+## PSV0/RDAT Updates
+I don't have enough context to even posit a guess.
+
+## Shader Reflection
+I don't have enough context to even posit a guess.