From b98e1901f5d7035cd121b7c71b31a78afa89afc3 Mon Sep 17 00:00:00 2001
From: Francis Russell <francis@unchartedbackwaters.co.uk>
Date: Wed, 18 Apr 2012 17:20:26 +0100
Subject: [PATCH] Presented version.

---
 presentation.tex | 63 ++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 61 insertions(+), 2 deletions(-)

diff --git a/presentation.tex b/presentation.tex
index 0c4e249..a5ce7ba 100644
--- a/presentation.tex
+++ b/presentation.tex
@@ -11,10 +11,10 @@
 %\input{pygments.sty}
 
 \title[ONETEP - DSLs and GPUs]{ONETEP - Code generation from a Quantum Chemistry
-DSL / GPU Parallelisation}
+DSL}
 
 \subtitle{PSL Presentation}
-\author[K. Wilkinson \& F. Russell]{Karl Wilkinson \& Francis Russell \\ Joint work with Chris-Kriton Skylaris \& Paul Kelly}
+\author[F. Russell]{Francis Russell \\ Joint work with Chris-Kriton Skylaris \& Karl Wilkinson}
 \date{18/04/2012}
 \institute[ICL \& Soton]{Imperial College London \& Southampton University}
 
@@ -167,6 +167,33 @@ output is FortranFunction("integrals_kinetic", ["kinet",
 
 \frame{
 
+\frametitle{Mapping indices to real code}
+
+\begin{itemize}
+
+\item We don't expose position indices in our DSL since these are
+related to the choice of discretisation. However we use them internally
+for ONETEP code generation.
+
+\item When we use DSL indices to index multiple operands,
+we are forcing the compiler to generate code that can iterate over all
+the operands simultaneously.
+
+\item The mapping between DSL indices and loop-indices in generated code
+is complex. Loops in the generated code may correspond to zero or more DSL
+indices.
+
+\item The code we want to generate is primarily related to control flow and data
+movement. The most computationally intensive part of the kinetic energy
+calculation is the Fourier transform, and we intend to call a black-box function
+for that anyway.
+
+\end{itemize}
+
+}
+
+\frame{
+
 \frametitle{Implementation Challenges: Interfacing with ONETEP}
 
 We do not have a well-defined abstract interface to interact with ONETEP. We
@@ -259,10 +286,42 @@ manipulation of spatial indices, support reduction operations etc.
 
 \frame{
 
+\frametitle{Current Status}
+
+\begin{itemize}
+
+\item Our compiler can parse expressions in our DSL.
+
+\item Using a relatively abstract description of the ONETEP data
+structures, it constructs representations of the iterations and
+predicates required to generate the points required from a particular
+operand.
+
+\item Inferring an efficient loop structure from this when indices are
+bound together is a work in progress.
+
+\item Thie approach captures data dependency well, but not anything
+involving mutable state. This is needed for allocating and populating
+the array passed to the black-box Fourier transform implementation.
+
+It will also be necessary for correctly reasoning about the
+communication required for MPI code generation.
+
+\end{itemize}
+
+}
+
+\frame{
+
 \frametitle{The Future}
 
 \begin{itemize}
 
+\item{\bf Inferring data movement.} ONETEP already has MPI
+parallelisation. We would like to be able to automatically reason about
+data transfers, exploring the possibility of improving scalability.
+Similarly, this is required if we wish to generate code targetting GPUs.
+
 \item{\bf A mechanism for systematically exploring different code generation
 variants.} If we can generate a search space of valid implementations, we
 can work on defining an objective function that lets us choose the best
-- 
2.47.3