Skip to content

Latest commit

 

History

History
17 lines (13 loc) · 1.34 KB

param.adoc

File metadata and controls

17 lines (13 loc) · 1.34 KB

Implementation-defined Constant Parameters

Each hart supporting a matrix extension defines four parameters:

  1. The maximum size in bits of a matrix element that any operation can produce or consume, \$\rm{ELEN} \ge 8\$, which must be a power of 2.

  2. The number of bits in a single matrix tile register, MLEN, which must be a power of 2, and must be no greater than 232.

  3. The number of bits in a row of a single matrix tile register, RLEN, which must be a power of 2, and must be no greater than 216.

  4. The multiple of length for matrix accumulation registers, AMUL, where the number of bits in a row of a single matrix accumulation register is \$\rm{RLEN} \times \rm{AMUL}\$, and the number of bits in a single matrix accumulation register is \$\rm{MLEN} \times \rm{AMUL}\$.

Some constraints on these parameters are defined as following.

  1. \$\rm{ELEN} \le \rm{RLEN} \le \rm{MLEN}\$, this supports matrix tile size from \$1 \times 1\$ to \$2^{16} \times 2^{16}\$.

  2. For implmentations without widening accumulation space, \$\rm{AMUL} = 1\$.

  3. For implmentations with double-widening accumulation space, \$\rm{AMUL} = 2\$.

  4. For implmentations with quadruple-widening accumulation space, \$\rm{AMUL} = 4\$.

  5. For implmentations with octuple-widening accumulation space, \$\rm{AMUL} = 8\$.

  6. AMUL with any other value is not allowed.