.macro vmul_iter reg0 reg1 reg2 idx | |
.irp op, vmul.i16 vmul.f16 vmul.i32 vmul.f32 | |
\op d\reg0, d\reg1, d\reg2[\idx] | |
\op q\reg0, q\reg1, d\reg2[\idx] | |
.endr | |
.endm | |
.macro vmul_acc_iter reg0 reg1 reg2 idx | |
.irp op, vmla.i16 vmla.i32 vmla.f16 vmla.f32 vmls.i16 vmls.i32 vmls.f16 vmls.f32 | |
\op d\reg0, d\reg1, d\reg2[\idx] | |
\op q\reg0, q\reg1, d\reg2[\idx] | |
.endr | |
.endm | |
# There are two restriction on the scalar operand: | |
# * The scalar operand is restricted to D0-D7 if size is 16bit wide, | |
# or D0 - D15 otherwise. | |
# * The scalar index should within range, 0-3 if size is 16bit wide, | |
# 0-1 if size is 32bit wide. | |
vmul_iter 0 1 3 0 | |
vmul_iter 3 12 7 2 | |
vmul_iter 4 9 8 1 | |
vmul_iter 13 6 15 3 | |
vmul_acc_iter 2 7 1 0 | |
vmul_acc_iter 5 4 6 2 | |
vmul_acc_iter 4 13 10 1 | |
vmul_acc_iter 12 6 13 3 |