Closed dchigarev closed 2 days ago
This can be implemented using xegpu.load_nd %x {transpose = {1, 0}} for B tiles.
xegpu.load_nd %x {transpose = {1, 0}}
We should support both patterns:
linalg.matmul_transpose_b
matmul
%b_tr = linalg.transpose %b ... %res = linalg.matmul %a, %b, ...
This functionality is required for OV integration (#207).
This can be implemented using
xegpu.load_nd %x {transpose = {1, 0}}
for B tiles.We should support both patterns:
linalg.matmul_transpose_b
.matmul
(this is how MLIR from OV will look like):This functionality is required for OV integration (#207).