Slides-part2_scalarization.tex

\talksection{Scalarization Stage}

\begin{frame}[fragile]{Scalarization Overview}

\begin{itemize}
    \item Eliminates vector operations from the source function
    \item Vector types used likely to be narrower than the native SIMD width
    \item To be combined with packetization
    \begin{itemize}
        \item Generate vector instructions with the native SIMD width
        %\item Implicitely performs 'Structure-of-Arrays to Array-of-Structures' conversion
    \end{itemize}
    \item On its own, does not change the the behaviour of the code
    %\item Alternative is instantiating vector instructions (and users) $N$ times
\end{itemize}

%\vspace{1.5ex}
\vfill
\hspace{1em}\includegraphics[scale=0.55]{images/stages-scalar.pdf}

\end{frame}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\begin{frame}[fragile]{Scalarization Example}

\begin{itemize}
    \item Example: Extract audio samples from left and right channels, scale by 2
    \item Scalarizing $n$-element loads and stores introduces a stride of $n$
    \begin{itemize}
        \item Results in interleaved loads and stores after packetization
    \end{itemize}
\end{itemize}

\begin{minipage}[t]{0.45\linewidth}

\vspace{0.1ex}
Before Scalarization (fragment):
\begin{codebox}[commandchars=\\\[\]]
int2 \uniform[*src], int \uniform[*left], int \uniform[*right];
int \varying[tid] = \varying[get_global_id(0)];


int2 \varying[sample] = \uniform[src]\idx[\varying[tid]];


\uniform[left]\idx[\varying[tid]] = (\varying[sample.x] << \uniform[1]);
\uniform[right]\idx[\varying[tid]] = (\varying[sample.y] << \uniform[1]);
\end{codebox}

\end{minipage}
\hspace{1em}
\begin{minipage}[t]{0.49\linewidth}

\vspace{0.1ex}
After Scalarization (reconstructed):
\begin{codebox}[commandchars=\\\[\]]
int2 \uniform[*src], int \uniform[*left], int \uniform[*right];
int \varying[tid] = \varying[get_global_id(0)];

int \uniform[*srcScalar] = ((int *)\uniform[src]);
int \varying[sampleLeft] = \uniform[srcScalar]\idx[(\varying[tid] * \uniform[2]) + \uniform[0]];
int \varying[sampleRight] = \uniform[srcScalar]\idx[(\varying[tid] * \uniform[2]) + \uniform[1]];

\uniform[left]\idx[\varying[tid]] = (\varying[sampleLeft] << \uniform[1]);
\uniform[right]\idx[\varying[tid]] = (\varying[sampleRight] << \uniform[1]);
\end{codebox}
%int \varying[sampleLeft] = ((int *)&\uniform[src]\idx[\varying[tid]])\idx[\uniform[0]];

\end{minipage}

\end{frame}