% vim: set foldmethod=marker foldmarker=<<<,>>>:

\section{Instruction level optimization}
% https://www.youtube.com/watch?v=BP6NxVxDQIs

 %<<< How code executes on a computer
\begingroup
\setbeamertemplate{background canvas}{%
\begin{tikzpicture}[remember picture,overlay]
\only<3>{
\draw[line width=20pt,red!60!black] 
  (11,-2) -- (15,-8);
\draw[line width=20pt,red!60!black] 
  (15,-2) -- (11,-8);
}
\end{tikzpicture}}
\begin{frame}[fragile] \frametitle{How code executes on a computer}{}
  \begin{columns}
    \column{0.4\textwidth}
    \begin{overprint}
      \onslide<1->%<<<
      \begin{minted}[
          frame=lines,
          fontsize=\footnotesize,
          linenos,
          gobble=8,
          mathescape
        ]{C++}
        void laplace(double* u, double* x,
                     double* y, double* f,
                          long Ns, long Nt) {
          for (long t = 0; t < Nt; t++) {
            for (long s = 0; s < Ns; s++) {
              double rx, ry, rz;
              rx = x[s*3]-y[t*3];
              ry = x[s*3+1]-y[t*3+1];
              rz = x[s*3+2]-y[t*3+2];

              double r2 = rx*rx+ry*ry+rz*rz;
              if (r2 > 0) {
                double rinv = 1/sqrt(r2);
                u[t] += f[s] * rinv;
              }
            }
          }
        }
      \end{minted}
      %>>>
    \end{overprint}
    \column{0.25\textwidth}
      \center
      \resizebox{0.99\textwidth}{!}{\begin{tikzpicture} %<<<
        \draw[draw=black,ultra thick] (0,0) rectangle (4,4.2);
        \node at (2,3.8) {\Large CPU};

        \draw[draw=black,ultra thick] (0.25,0.125) rectangle (3.75,1.125);
        \node at (2,0.625) {\Large Cache};

        \draw[draw=black,ultra thick] (0.25,1.25) rectangle (3.75,2.25);
        \node at (2,1.75) {\Large Control Unit};

        \draw[draw=black,ultra thick] (0.25,2.375) rectangle (3.75,3.375);
        \node at (2,2.875) {\Large ALU};

        \draw[latex-latex, ultra thick] (1,0) -- (1,-1);
        \draw[latex-latex, ultra thick] (2,0) -- (2,-1);
        \draw[latex-latex, ultra thick] (3,0) -- (3,-1);

        \draw[draw=black,ultra thick] (0,-2.2) rectangle (4,-1);
        \node at (2,-1.6) {\Large RAM};
      \end{tikzpicture}} %>>>
    \column{0.31\textwidth}

    \begin{itemize}
      \setlength\itemsep{0.75em}
      \item code executes line-by-line
      \item one scalar operation at a time
      \item one operation per clock cycle
      \item sequentially and in order
    \end{itemize}
    \only<2>{}

  \end{columns}

  % Programming language and hardware abstraction go hand-in-hand
  % most languages don't make it easy to specify when it is safe to vectorize (aliasing)

  % lies! forget that!
  % you have been lied to!
  % that is not how code executes on a computer at all
  % instructions can execute in any order -- but you are guaranteed that the net effect is the same as sequential
  % execution
\end{frame}
\endgroup
%>>>

\begin{frame} \frametitle{Core microarchitecture}{} %<<<

  \begin{columns}[t]
    \column{0.55\textwidth}

      \resizebox{0.99\textwidth}{!}{\begin{tikzpicture}
      \only<1>{
        %\write18{wget -O figs/skylake-arch.svg https://en.wikichip.org/w/images/e/ee/skylake_server_block_diagram.svg}
        %\write18{convert figs/skylake-arch.svg figs/skylake-arch.png}
        \node at (0,0) {\includegraphics[width=0.9\textwidth]{figs/skylake-arch}};
      }
      \only<2>{
        \node[opacity=0] at (0,0) {\includegraphics[width=0.9\textwidth]{figs/skylake-arch}};
        \node at (0,0) {\includegraphics[width=0.99\textwidth]{figs/skylake_scheduler}};
        \node at (0,-3) {\small Skylake micro-architecture (wikichip.org)};
      }
      \end{tikzpicture}}

    \column{0.45\textwidth}
      \begin{itemize}
        \setlength\itemsep{0.85em}
        \item {Speculative execution and branch predction}

        \item {Out-of-order execution}

        \only<2>{
        \item {Superscalar execution:} \\
          \quad 2-FP, 2-reads, 1-write

        \item {Vector instructions}

        \item {Pipelining:} \\
          \quad latency and throughput
        }

        %Instruction pipelining where the execution of multiple instructions can be partially overlapped.

        %Superscalar execution, VLIW, and the closely related explicitly parallel instruction computing concepts, in which
        %multiple execution units are used to execute multiple instructions in parallel.

        %Out-of-order execution where instructions execute in any order that does not violate data dependencies. Note that
        %this technique is independent of both pipelining and superscalar execution. Current implementations of out-of-order
        %execution dynamically (i.e., while the program is executing and without any help from the compiler) extract ILP from
        %ordinary programs. An alternative is to extract this parallelism at compile time and somehow convey this information
        %to the hardware. Due to the complexity of scaling the out-of-order execution technique, the industry has re-examined
        %instruction sets which explicitly encode multiple independent operations per instruction.

        %Register renaming which refers to a technique used to avoid unnecessary serialization of program operations imposed
        %by the reuse of registers by those operations, used to enable out-of-order execution.

        %Speculative execution which allows the execution of complete instructions or parts of instructions before being
        %certain whether this execution should take place. A commonly used form of speculative execution is control flow
        %speculation where instructions past a control flow instruction (e.g., a branch) are executed before the target of
        %the control flow instruction is determined. Several other forms of speculative execution have been proposed and are
        %in use including speculative execution driven by value prediction, memory dependence prediction and cache latency
        %prediction.

        %Branch prediction which is used to avoid stalling for control dependencies to be resolved. Branch prediction is used
        %with speculative execution.

      \end{itemize}
  \end{columns}

  % CPU core complexity: https://www.youtube.com/watch?v=eICYHA-eyXM&t=555s
  % out-of-order, vector, branch-prediction
\end{frame}
%>>>

\begin{frame} \frametitle{Instruction level parallelism}{} %<<<

  \center
  \includegraphics[width=0.8\textwidth]{figs/intel-core-gflops}

  {\footnotesize John McCalpin - Memory bandwidth and system balance in HPC systems, 2016}

\end{frame}
%>>>

\begin{frame}[fragile] \frametitle{Instruction latency and throughput}{} %<<<

  \begin{columns}[t]
    \column{0.45\textwidth}
    \footnotesize
    \begin{overprint}

      \onslide<1>%<<<
      \begin{minted}[
          frame=lines,
          fontsize=\footnotesize,
          linenos,
          gobble=8,
          mathescape
        ]{C++}
        #include <iostream>
        #include <omp.h>

        int main(int argc, char** argv) {
          double x = 3.141, one = 1.0;

          double T = -omp_get_wtime();
          for (long i = 0; i < 1000000000L; i++) {
            x = one + x;
          }
          T += omp_get_wtime();
          std::cout<<"T = "<< T <<'\n';
          std::cout<<"cycles/iter = "<< 3.3*T <<'\n';

          return 0;
        }
      \end{minted}
      %>>>

      \onslide<2-3>%<<<
      \begin{minted}[
          frame=lines,
          fontsize=\footnotesize,
          linenos,
          gobble=8,
          mathescape
        ]{C++}
        #include <iostream>
        #include <omp.h>

        int main(int argc, char** argv) {
          double x = 3.141, one = 1.0;

          double T = -omp_get_wtime();
          for (long i = 0; i < 1000000000L; i++) {
            x = one + x;
          }
          T += omp_get_wtime();
          std::cout<<"T = "<< T <<'\n';
          std::cout<<"cycles/iter = "<< 3.3*T <<'\n';

          std::cout<<x<<'\n';
          return 0;
        }
      \end{minted}
      %>>>

      \onslide<4-5>%<<<
      \begin{minted}[
          frame=lines,
          fontsize=\footnotesize,
          linenos,
          gobble=10,
          mathescape
        ]{C++}
          double x[32], one = 1;
          // ... initialize x

          double T = -omp_get_wtime();
          for (long i = 0; i < 1000000000L; i++) {
            x[0] = one + x[0];
            x[1] = one + x[1];
            x[2] = one + x[2];
            x[3] = one + x[3];
            ...
            x[31] = one + x[31];
          }
          T += omp_get_wtime();
          std::cout<<"T = "<< T <<'\n';
          std::cout<<"cycles/iter = "<< 3.3*T <<'\n';
      \end{minted}
      %>>>

      \onslide<6-7>%<<<
      \begin{minted}[
          frame=lines,
          fontsize=\footnotesize,
          linenos,
          gobble=10,
          mathescape
        ]{C++}
          sctl::Vec<double,8> x[8], one = 1;
          // ... initialize x

          double T = -omp_get_wtime();
          for (long i = 0; i < 1000000000L; i++) {
            x[0] = one + x[0];
            x[1] = one + x[1];
            x[2] = one + x[2];
            x[3] = one + x[3];
            ...
            x[8] = one + x[8];
          }
          T += omp_get_wtime();
          std::cout<<"T = "<< T <<'\n';
          std::cout<<"cycles/iter = "<< 3.3*T <<'\n';
      \end{minted}
      %>>>

      \onslide<8->%<<<
      \begin{minted}[
          frame=lines,
          fontsize=\footnotesize,
          linenos,
          gobble=10,
          mathescape
        ]{C++}
          sctl::Vec<double,8> x[8], one = 1;
          // ... initialize x

          double T = -omp_get_wtime();
          for (long i = 0; i < 1000000000L; i++) {
            x[0] = one / x[0];
            x[1] = one / x[1];
            x[2] = one / x[2];
            x[3] = one / x[3];
            ...
            x[8] = one / x[8];
          }
          T += omp_get_wtime();
          std::cout<<"T = "<< T <<'\n';
          std::cout<<"cycles/iter = "<< 3.3*T <<'\n';
      \end{minted}
      %>>>

    \end{overprint}

    \column{0.1\textwidth}

    \column{0.45\textwidth}

    \begin{overprint}
      \onslide<1-2>%<<<
      \begin{minted}[gobble=8,fontsize=\footnotesize]{text}
        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 0
        cycles/iter = 0
      \end{minted}
      %>>>

      \onslide<3-4>%<<<
      \begin{minted}[gobble=8,fontsize=\footnotesize]{text}
        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 0
        cycles/iter = 0


        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 1.22387
        cycles/iter = 4.03876
      \end{minted}
      %>>>

      \onslide<5-5>%<<<
      \begin{minted}[gobble=8,fontsize=\footnotesize]{text}
        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 0
        cycles/iter = 0


        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 1.22387
        cycles/iter = 4.03876


        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 1.22366
        cycles/iter = 4.03809
      \end{minted}

      \textcolor{red}{\qquad 8 adds/cycle!}
      %>>>

      \onslide<7-8>%<<<
      \begin{minted}[gobble=8,fontsize=\footnotesize]{text}
        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 1.22806
        cycles/iter = 4.05259
      \end{minted}

      \textcolor{red}{\qquad 16 adds/cycle!}
      %>>>

      \onslide<9-9>%<<<
      \begin{minted}[gobble=8,fontsize=\footnotesize]{text}
        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 1.22806
        cycles/iter = 4.05259
      \end{minted}

      \textcolor{red}{\qquad 16 adds/cycle!}

      \vspace{1em}
      \begin{minted}[gobble=8,fontsize=\footnotesize]{text}
        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 39.1521
        cycles/iter = 129.202
      \end{minted}

      \textcolor{red}{\qquad \sim 32$\times$ slower!}
      %>>>

      \onslide<10->%<<<
      \begin{minted}[gobble=8,fontsize=\footnotesize]{text}
        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 1.22806
        cycles/iter = 4.05259
      \end{minted}

      \textcolor{red}{\qquad 16 adds/cycle!}

      \vspace{1em}
      \begin{minted}[gobble=8,fontsize=\footnotesize]{text}
        $ g++ -O3 -march=native -fopenmp test.cpp
        $ ./a.out
        T = 39.1521
        cycles/iter = 129.202
      \end{minted}

      \textcolor{red}{\qquad \sim 32$\times$ slower!}

      \footnotesize
      \vspace{1em}
      \quad {\normalsize Fast}: bitwise ops, int \& fp ops ($+,-,*$)

      \vspace{0.5em}
      \quad {\normalsize Slow}: branches, $/, {\sqrt \cdot}, \sin, \cos, \cdots$
      %>>>

    \end{overprint}

  \end{columns}

  % coding example
\end{frame}
%>>>

\begin{frame}[fragile] \frametitle{Pipelining: evaluating polynomials} %<<<
  \begin{columns}[T]
    \column{0.15\textwidth}
      {\bf Input:} \\
      x,~a,~b,~c,~d,~e,~f,~g,~h \\

      \vspace{1em}
      {\bf Compute:} \\
      ((((((ax+b)x+c)x+d)x\\
      ~~~~+e)x+f)x+g)x+h

    \column{0.6\textwidth}
      \resizebox{0.88\textwidth}{!}{\begin{tikzpicture}[nodes={draw, ellipse}, latex-]
        \node{$\times, +$}
          child { node {$\times, +$}
            child { node {$\times, +$}
              child { node {$\times, +$}
                child { node {$\times, +$}
                  child { node {$\times, +$}
                    child { node {$\times, +$}
                      child { node {a} }
                      child { node {x} }
                      child { node {b} }
                    }
                    child { node {x} }
                    child { node {c} }
                  }
                  child { node {x} }
                  child { node {d} }
                }
                child { node {x} }
                child { node {e} }
              }
              child { node {x} }
              child { node {f} }
            }
            child { node {x} }
            child { node {g} }
          }
          child { node {x} }
          child { node {h} };
      \end{tikzpicture}}%

    \column{0.25\textwidth}
      \textcolor{c1}{u = a * x + b}\only<1-4>{ $\leftarrow$} \\
      \textcolor{c2}{v = u * x + c}\only<5-8>{ $\leftarrow$} \\
      \textcolor{c3}{w = v * x + d}\only<9-12>{ $\leftarrow$} \\
      \textcolor{c4}{p = w * x + e}\only<13-16>{ $\leftarrow$} \\
      \textcolor{c5}{q = p * x + f}\only<17-20>{ $\leftarrow$} \\
      \textcolor{c6}{r = q * x + g}\only<21-24>{ $\leftarrow$} \\
      \textcolor{c7}{s = r * x + h}\only<25-28>{ $\leftarrow$} \\

      \vspace{1em}
      {\bf Pipeline:}

      \vspace{0.5em}
      \resizebox{0.99\textwidth}{!}{\begin{tikzpicture}
        \draw[draw=none] (0,0) rectangle (4,1);
        \only<1-28>{
        \draw[fill=white] (0,0) rectangle (1,0.5);
        \draw[fill=white] (1,0) rectangle (2,0.5);
        \draw[fill=white] (2,0) rectangle (3,0.5);
        \draw[fill=white] (3,0) rectangle (4,0.5);

        \draw[fill=white] (0,0.6) rectangle (1,1.1);
        \draw[fill=white] (1,0.6) rectangle (2,1.1);
        \draw[fill=white] (2,0.6) rectangle (3,1.1);
        \draw[fill=white] (3,0.6) rectangle (4,1.1);
        }

        \only<1>{\draw[fill=c1] (0,0) rectangle (1,0.5);}
        \only<2>{\draw[fill=c1] (1,0) rectangle (2,0.5);}
        \only<3>{\draw[fill=c1] (2,0) rectangle (3,0.5);}
        \only<4>{\draw[fill=c1] (3,0) rectangle (4,0.5);}

        \only<5>{\draw[fill=c2] (0,0) rectangle (1,0.5);}
        \only<6>{\draw[fill=c2] (1,0) rectangle (2,0.5);}
        \only<7>{\draw[fill=c2] (2,0) rectangle (3,0.5);}
        \only<8>{\draw[fill=c2] (3,0) rectangle (4,0.5);}

        \only<9 >{\draw[fill=c3] (0,0) rectangle (1,0.5);}
        \only<10>{\draw[fill=c3] (1,0) rectangle (2,0.5);}
        \only<11>{\draw[fill=c3] (2,0) rectangle (3,0.5);}
        \only<12>{\draw[fill=c3] (3,0) rectangle (4,0.5);}

        \only<13>{\draw[fill=c4] (0,0) rectangle (1,0.5);}
        \only<14>{\draw[fill=c4] (1,0) rectangle (2,0.5);}
        \only<15>{\draw[fill=c4] (2,0) rectangle (3,0.5);}
        \only<16>{\draw[fill=c4] (3,0) rectangle (4,0.5);}

        \only<17>{\draw[fill=c5] (0,0) rectangle (1,0.5);}
        \only<18>{\draw[fill=c5] (1,0) rectangle (2,0.5);}
        \only<19>{\draw[fill=c5] (2,0) rectangle (3,0.5);}
        \only<20>{\draw[fill=c5] (3,0) rectangle (4,0.5);}

        \only<21>{\draw[fill=c6] (0,0) rectangle (1,0.5);}
        \only<22>{\draw[fill=c6] (1,0) rectangle (2,0.5);}
        \only<23>{\draw[fill=c6] (2,0) rectangle (3,0.5);}
        \only<24>{\draw[fill=c6] (3,0) rectangle (4,0.5);}

        \only<25>{\draw[fill=c7] (0,0) rectangle (1,0.5);}
        \only<26>{\draw[fill=c7] (1,0) rectangle (2,0.5);}
        \only<27>{\draw[fill=c7] (2,0) rectangle (3,0.5);}
        \only<28>{\draw[fill=c7] (3,0) rectangle (4,0.5);}

        \only<29>{\node at (2,0.75) {\Large 28 cycles};}
        \only<29>{\node at (2,0.25) {\Large 12.5\% utilization!};}

      \end{tikzpicture}}%

  \end{columns}


  % Helmholtz kernel code example
  % sample sort code
  % evaluating a polynomial

  % what we think happens
  % reality!
\end{frame}
%>>>

\begin{frame}[fragile] \frametitle{Pipelining: evaluating polynomials} %<<<

  \begin{columns}[T]
    \column{0.75\textwidth}
      {\bf Input:} \\
      x,~a,~b,~c,~d,~e,~f,~g,~h \\

      \vspace{1em}
      {\bf Compute:} \\
      ((ax+b)x\textsuperscript{2}+(cx+d))x\textsuperscript{4}+(ex+f)x\textsuperscript{2}+(gx+h)

      \resizebox{0.99\textwidth}{!}{\begin{tikzpicture}[
        baseline,
        level distance=15mm,
        %text depth=.5em,
        %text height=.8em,
        level 1/.style={sibling distance=10em},
        level 2/.style={sibling distance=5em},
        level 3/.style={sibling distance=2.5em},
        level 4/.style={sibling distance=1em},
        nodes={draw, ellipse}, latex-]

        \node{$\times,+$}
          child { node {$\times,+$}
            child { node {$\times,+$}
              child { node {a} }
              child { node {x} }
              child { node {b} }
            }
            child { node {$\times$}
              child { node {x} }
            }
            child { node {$\times,+$}
              child { node {c} }
              child { node {x} }
              child { node {d} }
            }
          }
          child { node {$\times$}
            child { node {$\times$}
              child { node {x} }
            }
          }
          child { node {$\times,+$}
            child { node {$\times,+$}
              child { node {e} }
              child { node {x} }
              child { node {f} }
            }
            child { node {$\times$}
              child { node {x} }
            }
            child { node {$\times,+$}
              child { node {g} }
              child { node {x} }
              child { node {h} }
            }
          };

      \end{tikzpicture}}%

    \column{0.25\textwidth}
      %<<<
      \textcolor{c1}{x\textsuperscript{2} = x * x}                                      \only<1-4>{ $\leftarrow$} \\ %
      \textcolor{c2}{x\textsuperscript{4} = x\textsuperscript{2} * x\textsuperscript{2}}\only<5-8>{ $\leftarrow$} \\ %
      \textcolor{c3}{u = a * x + b}                                                     \only<1-4>{ $\leftarrow$} \\
      \textcolor{c4}{v = c * x + d}                                                     \only<2-5>{ $\leftarrow$} \\ %
      \textcolor{c5}{w = e * x + f}                                                     \only<2-5>{ $\leftarrow$} \\
      \textcolor{c6}{p = g * x + h}                                                     \only<3-6>{ $\leftarrow$} \\ %
      \textcolor{c7}{q = u * x\textsuperscript{2} + v}                                  \only<6-9>{ $\leftarrow$} \\ %
      \textcolor{c8}{r = w * x\textsuperscript{2} + p}                                  \only<7-10>{ $\leftarrow$} \\ %
      \textcolor{c9}{s = q * x\textsuperscript{4} + r}                                  \only<11-14>{ $\leftarrow$} \\ %

      \vspace{0.5em}
      {\bf Pipeline:}

      \vspace{0.1em}
      \resizebox{0.99\textwidth}{!}{\begin{tikzpicture}
        \draw[draw=none] (0,0) rectangle (4,1);
        \only<1-14>{
        \draw[fill=white] (0,0) rectangle (1,0.5);
        \draw[fill=white] (1,0) rectangle (2,0.5);
        \draw[fill=white] (2,0) rectangle (3,0.5);
        \draw[fill=white] (3,0) rectangle (4,0.5);

        \draw[fill=white] (0,0.6) rectangle (1,1.1);
        \draw[fill=white] (1,0.6) rectangle (2,1.1);
        \draw[fill=white] (2,0.6) rectangle (3,1.1);
        \draw[fill=white] (3,0.6) rectangle (4,1.1);
        }

        \only<1>{\draw[fill=c1] (0,0) rectangle (1,0.5);}
        \only<2>{\draw[fill=c1] (1,0) rectangle (2,0.5);}
        \only<3>{\draw[fill=c1] (2,0) rectangle (3,0.5);}
        \only<4>{\draw[fill=c1] (3,0) rectangle (4,0.5);}

        \only<5>{\draw[fill=c2] (0,0) rectangle (1,0.5);}
        \only<6>{\draw[fill=c2] (1,0) rectangle (2,0.5);}
        \only<7>{\draw[fill=c2] (2,0) rectangle (3,0.5);}
        \only<8>{\draw[fill=c2] (3,0) rectangle (4,0.5);}

        \only<1>{\draw[fill=c3] (0,0.6) rectangle (1,1.1);}
        \only<2>{\draw[fill=c3] (1,0.6) rectangle (2,1.1);}
        \only<3>{\draw[fill=c3] (2,0.6) rectangle (3,1.1);}
        \only<4>{\draw[fill=c3] (3,0.6) rectangle (4,1.1);}

        \only<2>{\draw[fill=c4] (0,0) rectangle (1,0.5);}
        \only<3>{\draw[fill=c4] (1,0) rectangle (2,0.5);}
        \only<4>{\draw[fill=c4] (2,0) rectangle (3,0.5);}
        \only<5>{\draw[fill=c4] (3,0) rectangle (4,0.5);}

        \only<2>{\draw[fill=c5] (0,0.6) rectangle (1,1.1);}
        \only<3>{\draw[fill=c5] (1,0.6) rectangle (2,1.1);}
        \only<4>{\draw[fill=c5] (2,0.6) rectangle (3,1.1);}
        \only<5>{\draw[fill=c5] (3,0.6) rectangle (4,1.1);}

        \only<3>{\draw[fill=c6] (0,0) rectangle (1,0.5);}
        \only<4>{\draw[fill=c6] (1,0) rectangle (2,0.5);}
        \only<5>{\draw[fill=c6] (2,0) rectangle (3,0.5);}
        \only<6>{\draw[fill=c6] (3,0) rectangle (4,0.5);}

        \only<6>{\draw[fill=c7] (0,0) rectangle (1,0.5);}
        \only<7>{\draw[fill=c7] (1,0) rectangle (2,0.5);}
        \only<8>{\draw[fill=c7] (2,0) rectangle (3,0.5);}
        \only<9>{\draw[fill=c7] (3,0) rectangle (4,0.5);}

        \only<7>{\draw[fill=c8] (0,0) rectangle (1,0.5);}
        \only<8>{\draw[fill=c8] (1,0) rectangle (2,0.5);}
        \only<9>{\draw[fill=c8] (2,0) rectangle (3,0.5);}
        \only<10>{\draw[fill=c8] (3,0) rectangle (4,0.5);}

        \only<11>{\draw[fill=c9] (0,0) rectangle (1,0.5);}
        \only<12>{\draw[fill=c9] (1,0) rectangle (2,0.5);}
        \only<13>{\draw[fill=c9] (2,0) rectangle (3,0.5);}
        \only<14>{\draw[fill=c9] (3,0) rectangle (4,0.5);}

        \only<15>{\node at (2,0.75) {\Large 14 cycles};}
        \only<15>{\node at (2,0.25) {\Large 2\times speedup!};}

      \end{tikzpicture}}%
      %>>>
      %%<<<
      %\textcolor{c1}{x\textsuperscript{2} = x * x}                                      \only<1-4>{ $\leftarrow$} \\
      %\textcolor{c2}{x\textsuperscript{4} = x\textsuperscript{2} * x\textsuperscript{2}}\only<5-8>{ $\leftarrow$} \\
      %\textcolor{c3}{u = a * x + b}                                                     \only<2-5>{ $\leftarrow$} \\
      %\textcolor{c4}{v = c * x + d}                                                     \only<3-6>{ $\leftarrow$} \\
      %\textcolor{c5}{w = e * x + f}                                                     \only<4-7>{ $\leftarrow$} \\
      %\textcolor{c6}{p = g * x + h}                                                     \only<6-9>{ $\leftarrow$} \\
      %\textcolor{c7}{q = u * x\textsuperscript{2} + v}                                  \only<7-10>{ $\leftarrow$} \\
      %\textcolor{c8}{r = w * x\textsuperscript{2} + p}                                  \only<10-13>{ $\leftarrow$} \\
      %\textcolor{c9}{s = q * x\textsuperscript{4} + r}                                  \only<14-17>{ $\leftarrow$} \\

      %\vspace{0.5em}
      %{\bf Pipeline:}

      %\vspace{0.1em}
      %\resizebox{0.99\textwidth}{!}{\begin{tikzpicture}
      %  \draw[draw=none] (0,0) rectangle (4,1);
      %  \only<1-17>{
      %  \draw[fill=white] (0,0) rectangle (1,1);
      %  \draw[fill=white] (1,0) rectangle (2,1);
      %  \draw[fill=white] (2,0) rectangle (3,1);
      %  \draw[fill=white] (3,0) rectangle (4,1);
      %  }

      %  \only<1>{\draw[fill=c1] (0,0) rectangle (1,1);}
      %  \only<2>{\draw[fill=c1] (1,0) rectangle (2,1);}
      %  \only<3>{\draw[fill=c1] (2,0) rectangle (3,1);}
      %  \only<4>{\draw[fill=c1] (3,0) rectangle (4,1);}

      %  \only<5>{\draw[fill=c2] (0,0) rectangle (1,1);}
      %  \only<6>{\draw[fill=c2] (1,0) rectangle (2,1);}
      %  \only<7>{\draw[fill=c2] (2,0) rectangle (3,1);}
      %  \only<8>{\draw[fill=c2] (3,0) rectangle (4,1);}

      %  \only<2>{\draw[fill=c3] (0,0) rectangle (1,1);}
      %  \only<3>{\draw[fill=c3] (1,0) rectangle (2,1);}
      %  \only<4>{\draw[fill=c3] (2,0) rectangle (3,1);}
      %  \only<5>{\draw[fill=c3] (3,0) rectangle (4,1);}
      %
      %  \only<3>{\draw[fill=c4] (0,0) rectangle (1,1);}
      %  \only<4>{\draw[fill=c4] (1,0) rectangle (2,1);}
      %  \only<5>{\draw[fill=c4] (2,0) rectangle (3,1);}
      %  \only<6>{\draw[fill=c4] (3,0) rectangle (4,1);}
      %
      %  \only<4>{\draw[fill=c5] (0,0) rectangle (1,1);}
      %  \only<5>{\draw[fill=c5] (1,0) rectangle (2,1);}
      %  \only<6>{\draw[fill=c5] (2,0) rectangle (3,1);}
      %  \only<7>{\draw[fill=c5] (3,0) rectangle (4,1);}
      %
      %  \only<6>{\draw[fill=c6] (0,0) rectangle (1,1);}
      %  \only<7>{\draw[fill=c6] (1,0) rectangle (2,1);}
      %  \only<8>{\draw[fill=c6] (2,0) rectangle (3,1);}
      %  \only<9>{\draw[fill=c6] (3,0) rectangle (4,1);}

      %  \only<7>{\draw[fill=c7] (0,0) rectangle (1,1);}
      %  \only<8>{\draw[fill=c7] (1,0) rectangle (2,1);}
      %  \only<9>{\draw[fill=c7] (2,0) rectangle (3,1);}
      %  \only<10>{\draw[fill=c7] (3,0) rectangle (4,1);}

      %  \only<10>{\draw[fill=c8] (0,0) rectangle (1,1);}
      %  \only<11>{\draw[fill=c8] (1,0) rectangle (2,1);}
      %  \only<12>{\draw[fill=c8] (2,0) rectangle (3,1);}
      %  \only<13>{\draw[fill=c8] (3,0) rectangle (4,1);}

      %  \only<14>{\draw[fill=c9] (0,0) rectangle (1,1);}
      %  \only<15>{\draw[fill=c9] (1,0) rectangle (2,1);}
      %  \only<16>{\draw[fill=c9] (2,0) rectangle (3,1);}
      %  \only<17>{\draw[fill=c9] (3,0) rectangle (4,1);}

      %  \only<18>{\node at (2,0.75) {\Large 17 cycles};}
      %  \only<18>{\node at (2,0.25) {\Large 60\% faster!};}

      %\end{tikzpicture}}%
      %%>>>

  \end{columns}


  % Helmholtz kernel code example
  % sample sort code
  % evaluating a polynomial

  % what we think happens
  % reality!
\end{frame}
%>>>

\begin{frame} \frametitle{Pipelining: actual performance} %<<<

  % perf - show stalled cycles

\end{frame}
%>>>

\begin{frame} \frametitle{Vectorization}{} %<<<

  % benefits from fixed-size blocking (compiler can unroll)
  % loops have conditionals, so unrolling is difficult

  % vector dot product: show data dependency stalls

  % data arrangement: AoS vs SoA

  % most languages don't make it easy to specify when it is safe to vectorize (aliasing)
  % MMX, SSE, AVX, AVX512

  % Use fast operations instead of slow
  % remove un-necessary operations (pre-allocate memory)
  % reduce number of operations (caching)
  % batch operations, loop unrolling/fixed-length loops, expose instruction level parallelism

  % unaligned memory accesses

\end{frame}
%>>>

\begin{frame} \frametitle{Vectorization - GEMM micro-kernel}{} %<<<
  % show different ways of vectorizing that don't work
  % most languages don't make it easy to specify when it is safe to vectorize (aliasing)

  % start with triple loop
  % compiler options
  % loop unrolling
  % __restrict__
  %
\end{frame}
%>>>

\begin{frame} \frametitle{Instruction-level parallelism -- summary}{} %<<<

  % Cast all computations in additions, multiplications, bitwise ops (eg. baobzi)
  % Avoid expensive ops (div), branches
  % show penalty from branches
  % out-of-order execution, pipelining, vectorization:
  % - refactor code to expose instruction level parallelism (sometimes even at the cost of extra work)

\end{frame}
%>>>

\begin{frame} \frametitle{Optimized libraries for function evaluationa and vectorization} %<<<
  % Fast function evaluation using polynomial evaluation
  % baobzi
  % sf_benchmarks :
\end{frame}
%>>>