naive.md

benchmark_name	cpp_code	accera_code
Accera_Naive	src/batched_softmax/accera_naive.cpp	src/batched_softmax/naive.py

Naive Accera

Note

The following shows the implementation of the {{benchmark_name}}. The full source code listing of the Accera code generator can be found in {{accera_code}} :fas fa-code: and the benchmark runner is found in {{cpp_code}} :fas fa-code: .

The pseudocode of the naive implementation is:

\begin{algorithm} 
\begin{algorithmic} 
\PROCEDURE{BatchedSoftmax}{$Input$}
    \STATE maxVal[\texttt{BATCH\_SIZE}] = \{-$\infty$, $\ldots$, -$\infty$\}
    \STATE denom[\texttt{BATCH\_SIZE}] = \{$0$, $\ldots$, $0$\}
    \FOR{$bm$ = 0 \TO \texttt{BATCH\_SIZE}} 
        \FOR{$m$ = 0 \TO \texttt{N}} 
            \STATE maxVal[bm] = $max$(maxVal[bm], Input[bm, m])
        \ENDFOR 
    \ENDFOR 
    \FOR{$bi$ = 0 \TO \texttt{BATCH\_SIZE}} 
        \FOR{$i$ = 0 \TO \texttt{N}} 
            \STATE Output[bi, i] = $e^{\text{Input[bi, i]} - \text{maxVal[bi]}}$
        \ENDFOR 
    \ENDFOR 
    \FOR{$ba$ = 0 \TO \texttt{BATCH\_SIZE}} 
        \FOR{$a$ = 0 \TO \texttt{N}} 
            \STATE demon[ba] = denom[ba] + Output[ba, a]
        \ENDFOR 
    \ENDFOR 
    \FOR{$bj$ = 0 \TO \texttt{BATCH\_SIZE}} 
        \FOR{$j$ = 0 \TO \texttt{N}} 
            \STATE Output[bj,j] = $\frac{\text{Output[bj,j]}}{\text{denom[bj]}}$ 
        \ENDFOR 
    \ENDFOR  
    \RETURN Output
\ENDPROCEDURE
\end{algorithmic}
\end{algorithm}

A batched Softmax implementation in Accera follows from the above pseudocode. First, we need to import the required packages.

naive.py

We then define the problem size, input and output arrays, and the auxillary temporary arrays.

naive.py

We initialize the values in the temporary arrays.

naive.py

We compute the maximum value of the inputs within each batch.

naive.py

We compute the exponential value and store the results in the Output array.

naive.py

We accumulate the values of Output within each batch.

naive.py

We normalize the Output using the accumulated values.

naive.py

We fuse all the schedules into a single schedule and construct an action plan from that.

naive.py

We then export the plan as a naive named function in naive.o object file.

naive.py

Usage

The package can then be used within our C code base. To do so, we first need to import the HAT package created:

accera_naive.cpp

We then declare our inputs and outputs:

accera_naive.cpp

We then can use the exported function within our C code:

accera_naive.cpp