$Y[i]=0$\;
\ForAll {$ind$ $\in$ $c\_index[i]$} {
$Y[i] = Y[i] \oplus X[ind]$\;
- \CommentSty{$\oplus:$ bitwise xor operation}
+ \CommentSty{$\oplus:$ bitwise XOR operation}
}
}
% \end{algorithmic}
*, *, *, 10, *, *}
\end{lstlisting}
- \subsubsection*{Hybrid (HYB)\index{compressed storage format!HYB}} The HYB format heuristically computes a value $K$ and stores $K$ nonzeros per rows in the ELL format. When a row has more than $K$ non-zeros, the trailing nonzeros are stored in COO. This design decreases the storage overhead due to ELL padding elements and thus improves the overall performance.
+ \subsubsection*{Hybrid (HYB)\index{compressed storage format!HYB}} The HYB format heuristically computes a value $K$ and stores $K$ nonzeros per rows in the ELL format. When a row has more than $K$ non zeros, the trailing nonzeros are stored in COO. This design decreases the storage overhead due to ELL padding elements and thus improves the overall performance.
+\pagebreak
\begin{lstlisting}[caption={}]
hyb.nnz_per_row = 2
hyb.ell.col_index = {2, 1, 1, 0, 2, 0, *, 4, 3, 2, *, 5}
Avg row weight & 95.08 & 93.6 & 143 & 144\\
\hline
\end{tabular}
- \caption{Properties of some NFS matrices}
+ \caption{Properties of some NFS matrices.}
\label{table:rsa_matrix}
\end{table}
\subsection{Dense format}
The dense format is used for the dense part of the matrix. This format uses 1 bit per matrix entry. Within a column, 32 matrix entries are stored as a 32-bit integer. Thus, 32 rows are stored as $N$ consecutive integers.
- Each CUDA thread works on a column. Each thread fetches one element from the input vector in coalesced fashion. Then, each thread checks the 32 matrix entries one by one. When the matrix entry is a nonzero, the thread performs an xor operation between the element from the input vector and the partial result for the row. This means each thread accesses the input vector only once to do work on up to 32 nonzeros. The partial result from each thread needs to be stored and combined to get the final result for the 32 rows. These operations are performed in CUDA shared memory.
+ Each CUDA thread works on a column. Each thread fetches one element from the input vector in coalesced fashion. Then, each thread checks the 32 matrix entries one by one. When the matrix entry is a nonzero, the thread performs an XOR operation between the element from the input vector and the partial result for the row. This means each thread accesses the input vector only once to do work on up to 32 nonzeros. The partial result from each thread needs to be stored and combined to get the final result for the 32 rows. These operations are performed in CUDA shared memory.
- The 32 threads in a warp share 32 shared memory entries to store the partial results from the 32 rows. Since all threads in a warp execute a common instruction at the same time, access to these 32 entries can be made exclusive. The result from each warp in a thread block is combined using p-reduction on shared memory. The result from each thread block is combined using an atomic xor operation on global memory.
+ The 32 threads in a warp share 32 shared memory entries to store the partial results from the 32 rows. Since all threads in a warp execute a common instruction at the same time, access to these 32 entries can be made exclusive. The result from each warp in a thread block is combined using p-reduction on shared memory. The result from each thread block is combined using an atomic XOR operation on global memory.
When the blocking factor is larger than 64, access to the shared memory needs to be reorganized to avoid bank conflicts. Each thread can read/write up to 64-bit data at a time to the shared memory. If a thread is accessing 128-bit data for example, two read/write operations need to be performed. Thus, there will be bank conflicts if we store 128-bit data on contiguous addresses. We can avoid bank conflicts by having the threads in a warp first access consecutive 64-bit elements representing the first halves of the 128-bit elements. Then, the threads again access consecutive 64-bit elements representing the second halves of the 128-bit elements. The same modification can be applied to other formats as well.
\begin{figure}[t]
\centering
\includegraphics[height=5cm]{Chapters/chapter19/fig/scoo.pdf}
- \caption{Example of the memory access pattern for a $6 \times 6$ matrix stored in Sliced COO format (Slice Size = 3 rows).}
+ \caption{Example of the memory access pattern for a $6 \times 6$ matrix stored in sliced COO format (slice size = 3 rows).}
\label{fig:scoo-1}
\end{figure}
One thread block works on a slice, one group at a time. Neighboring threads work on neighboring nonzeros in the group in parallel. Each thread works on more than one row of the matrix. Thus, each thread needs some storage to store the partial results and combine them with the results from the other threads to generate the final output. Each entry costs equally as the element of the vector, i.e., the blocking factor in bytes. Shared memory is used as intermediate storage as it has low latency however it is limited to 48 KB per SM in Fermi. The global memory is accessed only once at the end to write the final output. The CUDA kernel for SpMV with Sliced COO format is shown in Listing \ref{ch19:lst:spmv}.
- Since neighboring nonzeros may or may not come from the same row (recall that we sorted them by column index), many threads may access the same entry in the shared memory if it is used to store result of the same rows. Thus, shared memory entries are either assigned exclusively to a single thread or shared by using atomic xor operations. Based on the way we allocate the shared memory, we further divide the Sliced COO format into three different subformats: small, medium, and large.
+ Since neighboring nonzeros may or may not come from the same row (recall that we sorted them by column index), many threads may access the same entry in the shared memory if it is used to store result of the same rows. Thus, shared memory entries are either assigned exclusively to a single thread or shared by using atomic XOR operations. Based on the way we allocate the shared memory, we further divide the Sliced COO format into three different subformats: small, medium, and large.
\begin{table}
\centering
Sliced COO subformats & Small & Medium & Large \\
\hline
Memory sharing & No sharing & Among warp & Among block\\
- Access method & Direct & Atomic xor & Atomic xor\\
+ Access method & Direct & Atomic XOR & Atomic XOR\\
Bank conflict & No & No & Yes\\
\# Rows per Slice & 12 & 192 & 6144 \\
\hline
The maximum number of rows per slice is calculated as \emph{size of shared memory per SM in bits} / \emph{(number of threads per block * blocking factor)}. We use 512 threads per block for 64-bit blocking factor which gives 12 rows, and 256 threads per block for 128 and 256-bit blocking factor, which gives 12 and 6 rows, respectively. Hence, one byte per row index is sufficient for this subformat.
\subsubsection*{Medium sliced (MS) COO}
- In this subformat, each thread in a warp gets an entry in the shared memory to store the partial result for each row. However, this entry is shared with the threads in other warps. Access to the shared memory uses an atomic xor operation. Each thread in a warp accesses only one bank, avoiding bank conflicts. A p-reduction operation on shared memory is required to combine the 32 partial results.
+ In this subformat, each thread in a warp gets an entry in the shared memory to store the partial result for each row. However, this entry is shared with the threads in other warps. Access to the shared memory uses an atomic XOR operation. Each thread in a warp accesses only one bank, avoiding bank conflicts. A p-reduction operation on shared memory is required to combine the 32 partial results.
The maximum number of rows per slice is calculated as \emph{size of shared memory per SM in bits} / \emph{(32 * blocking factor)} where 32 is the number of threads in a warp. This translates to 192, 96, and 48 rows per slice for blocking factors of 64, 128, and 256-bits, respectively. Hence, one byte per row index is sufficient for this format.
\subsubsection*{Large sliced (LS) COO}
- In this subformat, the result for each row gets one entry in shared memory, which is shared among all threads in the thread block. Access to shared memory uses an atomic xor operation. Thus, there will be bank conflicts. However, this drawback can be compensated for by a higher texture cache hit rate. There is no p-reduction on shared memory required.
+ In this subformat, the result for each row gets one entry in shared memory, which is shared among all threads in the thread block. Access to shared memory uses an atomic XOR operation. Thus, there will be bank conflicts. However, this drawback can be compensated for by a higher texture cache hit rate. There is no p-reduction on shared memory required.
The maximum number of rows per slice is calculated as \emph{size of shared memory per SM in bits} / \emph{blocking factor}. This translates to 6144, 3072, and 1536 rows per slice for blocking factors of 64, 128, and 256-bits, respectively. We need two bytes for the row index.
\noindent Table \ref{table:scoo} summarizes the differences between the three subformats.
\begin{itemize}
\item[a)] Each thread in a block is allocated one memory entry for the result of a row and works on consecutive 12 rows. A p-reduction to Thread\#1 is needed to get the final result of one row.
- \item[b)] Each warp in a block is allocated one memory entry per row and works on 192 consecutive rows, threads in each warp use atomic xor to update the value in a given entry. A p-reduction for each row to memory in Warp\#1 is needed to obtain the final result of that row.
- \item[c)] A thread block works on 6144 consecutive rows and shares the memory entry of each row. Any update to the memory has to use atomic xor.
+ \item[b)] Each warp in a block is allocated one memory entry per row and works on 192 consecutive rows, threads in each warp use atomic XOR to update the value in a given entry. A p-reduction for each row to memory in Warp\#1 is needed to obtain the final result of that row.
+ \item[c)] A thread block works on 6144 consecutive rows and shares the memory entry of each row. Any update to the memory has to use atomic XOR.
\end{itemize}
\subsection{Determining the cut-off point of each format}
\hline
\end{tabular}
\end{center}
-\caption{Overview of hardware used in the experiments}
+\caption{Overview of hardware used in the experiments.}
\label{table:hardware}
\end{table}
\hline
\end{tabular}
- \caption{Performance of SpMV on RSA-170 matrix}
+ \caption{Performance of SpMV on RSA-170 matrix.}
\label{table:rsa170}
\end{table}
Table \ref{table:rsa170} shows the result. The dual-GPU implementation achieves speedups between 1.93 and 1.96 compared to the single-GPU performance.
- Table \ref{table:individual} shows the individual performance of each subformat (in \emph{gnnz/s}), the percentage of nonzeros are included in parentheses. The results show that the performance decreases when the matrix gets sparser. The MS-COO and LS-COO performance degrade when the blocking factor is increased from 64 to 128 and 256-bit. This is caused by the increased number of bank conflicts and serialization of atomic xor operations on larger blocking factors. Thus, the SS-COO format gets a higher percentage of nonzeros with 128 and 256-bit blocking factor.
+ Table \ref{table:individual} shows the individual performance of each subformat (in \emph{gnnz/s}), the percentage of nonzeros are included in parentheses. The results show that the performance decreases when the matrix gets sparser. The MS-COO and LS-COO performance degrade when the blocking factor is increased from 64 to 128 and 256-bit. This is caused by the increased number of bank conflicts and serialization of atomic XOR operations on larger blocking factors. Thus, the SS-COO format gets a higher percentage of nonzeros with 128 and 256-bit blocking factor.
More detail results of our full block Wiedemann CUDA implementation as well as a multi-GPU implementation can be found in \cite{ch19:spmv-ccpe}
\begin{table}
\begin{center}
+\begin{small}
\begin{tabular}{|l|n{8}{0}|n{9}{0}|n{3}{2}|l|}
\hline
{Name} & row & column & nz/row & Description \\
uk-2002 &18520486&18520486&16.10&directed graph\\
\hline
\end{tabular}
+\end{small}
\end{center}
\caption{Overview of sparse matrices used for performance evaluation.}
\label{table:matrices}
