1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
|
\begin{table}[htb] %[htbp]
\begin{threeparttable}
\caption{\label{tbl:algs} Suggested algorithms}
\centering
\noindent
%\resizebox{\columnwidth}{!}{\begin{tabular}{} ... \end{\tabular}}
\begin{tabularx}{\columnwidth}{ | X c | c l l | c l l | }
\hline
& & \multicolumn{3}{c|}{Linear} & \multicolumn{3}{c|}{Nearest Neighbor} \\
\mhd{|c}{GPU} & \mhd{c|}{Mode} & \mhd{c}{S} & \mhd{c}{Precise} & \mhd{c|}{Appr.} & \mhd{c}{S} & \mhd{c}{Precise} & \mhd{c|}{Appr.} \\
\hline
% Slices & Linear/Prec. & Linear/Overs. Slices & NN/Prec. & NN/half.
\multirow{1}{*}{GT200}
& Single & 1 & TEX & TEX & 1 & \multicolumn{2}{c|}{TEX} \\
\hline
\multirow{1}{*}{Fermi}
& * & 4 & ALU & Overs. & 4 & \multicolumn{2}{c|}{ALU} \\
\hline
\multirow{2}{*}{Kepler}
& Single & 1 & TEX & Overs. & 1 & \multicolumn{2}{c|}{ALU} \\
& Multi & 2 & TEX & TEX & 4 & \multicolumn{2}{c|}{ALU} \\
\hline
\multirow{2}{*}{Mxwl+}
& Single & 1 & Hybrid & Overs. & 1 & \multicolumn{2}{c|}{ALU} \\
& Multi & 2 & Hybrid & Overs. & 4 & ALU & TEX/Half \\
\hline
\multirow{1}{*}{VLIW}
& Single & 4 & ALU & Overs. & 1 & \multicolumn{2}{c|}{ALU} \\
& Multi & 4 & ALU & Overs. & 4 & \multicolumn{2}{c|}{TEX} \\
\hline
\multirow{1}{*}{GCN}
& * & 4 & ALU & Overs. & 4 & \multicolumn{2}{c|}{ALU} \\
\hline
%\multirow{3}{*}{HD7970}
%\hline
%\multirow{3}{*}{R9-290}
%\hline
\end{tabularx}
\begin{tablenotes}
\item The table specifies the fastest algorithms to implement back-projection kernel with linear or nearest-neighbor interpolation at each platform. Individual recommendations are given for the single-slice and multi-slice reconstruction modes. The recommended number of slices is given in column \textbf{S}. The options for precise and approximate reconstructions are proposed. In precise mode, the obtained reconstruction is exactly the same as one produced by the standard reconstruction method. In approximate mode, either a half-float data representation is used to accelerate nearest-neighbor interpolation or the oversampling approach is combined with nearest neighbor interpolation to substitute linear interpolation. The performance and optimal configuration for the texture-based algorithm is listed in \tablename~\ref{tbl:newtex}. The ALU-based algorithm is described in \tablename~\ref{tbl:alurec} and its oversampling modification is given in \tablename~\ref{tbl:overs}. The hybrid approach is defined in \tablename~\ref{tbl:hybrid}.
\end{tablenotes}
\end{threeparttable}
\end{table}
|