A Low-Power Scalable Stream Compute Accelerator for General Matrix Multiply (GEMM)

Algorithm 1

divide 𝑋 and 𝑌 into blocks 𝑋′ and 𝑌′ of size (PE × cache depth);

For  each block of 𝑋′and 𝑌′  do

  prefetch 𝑌′ into cache via stream C;

  preload any W, 𝛼, 𝛽 or 𝑍′ to stream 𝐵;

   for  each row of 𝑋′  do

    stream new elements of the row of 𝑋′ via S;

    multiply-accumulate elements of 𝑋′ and 𝑌^'across PEs;

     if  𝑍′contains final elements  of 𝑋𝑌  then

       shift new partial results of 𝑍′ from PEs  via 𝑃;

       perform scalar operations using 𝛼, 𝛽 and

       𝑊 at output of 𝑃 and 𝐵 via ASE (§ 3.3);

     else

      shift new elements of 𝑍′ from PEs via 𝑃

      to memory or cache;

     end

   end

end