From 6412bd76d95866a86f9db265a0d45e39e51e4f76 Mon Sep 17 00:00:00 2001
From: Francis Russell <francis@unchartedbackwaters.co.uk>
Date: Tue, 11 Jun 2013 10:36:09 +0100
Subject: [PATCH] Add details about IMKL and running on CX1.

---
 presentation.tex | 40 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 40 insertions(+)

diff --git a/presentation.tex b/presentation.tex
index ee1e939..5e21281 100644
--- a/presentation.tex
+++ b/presentation.tex
@@ -269,7 +269,47 @@ possible to pass in data in data in the split format.
 
 }
 
+\frame{
+
+\frametitle{Running on CX1}
+
+\begin{itemize}
+
+\item We benchmarked on CX1 cluster at Imperial using the Intel Math
+  Kernel Libraries instead of FFTW.
+
+\item Performance was terrible, significantly worse than vanilla ONETEP
+  ($\approx$ 20\%
+  slowdowns).
 
+\item It looks like MKL really dislikes working with split layout data.
+
+\item If so, this means that ONETEP's data marshalling that we worked to remove
+actually helps with IMKL.
+
+\item I updated the interpolate library to support staging data so that the FFT
+  routines could work on entirely \emph{contiguous}, \emph{interleaved} data.
+  
+\item The other implementation would gather and scatter directly from and to the input
+  and output arrays using the FFT routines.
+
+\item When planning the interpolate, we benchmark and select the best technique.
+  When using IMKL, we almost always do the scatter/gather ourselves.
+
+\end{itemize}
+
+}
+
+\frame{
+
+\frametitle{Running on CX1}
+
+\large
+\centering
+
+We don't have performance results for this new implementation\ldots yet.
+
+}
 
 \end{document}
 
-- 
2.47.3