tvm_tirx.backend.cuda.op.TCGen05InstrDescriptorBlockScaledAttrs

tvm_tirx.backend.cuda.op.TCGen05InstrDescriptorBlockScaledAttrs#

class tvm_tirx.backend.cuda.op.TCGen05InstrDescriptorBlockScaledAttrs(d_dtype, a_dtype, b_dtype, sfa_dtype, sfb_dtype, M, N, K, trans_a, trans_b, n_cta_groups=<MISSING>, neg_a=<MISSING>, neg_b=<MISSING>, is_sparse=<MISSING>)#

Bases: Attrs

Static options carried by tcgen05_encode_instr_descriptor_block_scaled().

Parameters:

  • d_dtype (str) – Element type of the output matrix D.

  • a_dtype (str) – Element type of the input matrix A.

  • b_dtype (str) – Element type of the input matrix B.

  • sfa_dtype (str) – Element type of the scale factors for A.

  • sfb_dtype (str) – Element type of the scale factors for B.

  • M (int) – Number of rows in the output tile.

  • N (int) – Number of columns in the output tile.

  • K (int) – Reduction dimension of the input tiles.

  • trans_a (bool) – Whether to transpose A.

  • trans_b (bool) – Whether to transpose B.

  • n_cta_groups (int) – Number of CTA groups, 1 (default) or 2.

  • neg_a (bool) – Whether to negate A; defaults to False.

  • neg_b (bool) – Whether to negate B; defaults to False.

  • is_sparse (bool) – Whether to select sparse MMA; defaults to False.

property K#
property M#
property N#
property a_dtype#
property b_dtype#
property d_dtype#
property is_sparse#
property n_cta_groups#
property neg_a#
property neg_b#
property sfa_dtype#
property sfb_dtype#
property trans_a#
property trans_b#