fixed_p_rounding - Set up fixed-point arithmetic rounding information

Overview

The fixed_p_rounding module implements fixed-point rounding logic for Ara’s vector execution stage. It operates on 64-bit wide elements and generates rounding result bits for each element stream based on the rounding mode and vector element width (VEW).

---

Ports

Inputs

Name	Type	Description
operand_a_i	elen_t	Encodes the shift amount j for each sub-element.
operand_b_i	elen_t	Source operand to extract rounding bits from.
valid_i	logic	Enables rounding logic when high.
op_i	ara_op_e	Operation type; rounding occurs only on selected operations.
vew_i	vew_e	Specifies the vector element width: EW8, EW16, EW32, or EW64.
vxrm_i	vxrm_t	RISC-V vector rounding mode: 00, 01, 10, or 11.

Output

Name	Type	Description
r_o	strb_t	Output rounding bits for each element.

---

Internal Structures

rounding_args

A union packed type that allows simultaneous interpretation of a 64-bit value as:

• 8 x 8-bit (w8)

• 4 x 16-bit (w16)

• 2 x 32-bit (w32)

• 1 x 64-bit (w64)

---

Functional Behavior

Rounding logic is triggered only when valid_i is high and op_i matches one of:

• VSSRA

• VSSRL

• VNCLIP

• VNCLIPU

Depending on vew_i and vxrm_i, different rounding strategies are selected.

Rounding Modes (vxrm_i)

• 00 (RNU): Round to Nearest, Up.

• 01 (RNE): Round to Nearest, ties to Even.

• 10 (RTZ): Round Towards Zero (no rounding).

• 11 (ROD): Round towards -∞ for negative and +∞ for positive.

Element Width Handling

Each vew_i type (EW8, EW16, etc.) leads to different slicing of input operands.

Example for EW8, vxrm = 00:

for (int i = 0; i < 8; i++) begin
  j      = opa.w8[i];
  r_o[i] = opb.w8[i][j-1];
end

Example for EW32, vxrm = 01:

for (int i = 0; i < 2; i++) begin
  j      = opa.w32[i];
  r_o[i] = opb.w32[i][j-1] & opb.w32[i][j];
end

For vxrm = 10, no rounding is performed:

r_o = '0;

For vxrm = 11 (ROD), more logic is applied:

• Negates the jth bit

• ORs the masked remaining bits using a lookup from bit_select

---

Lookup Table: bit_select

This is a constant table used to mask and check lower bits for rounding:

bit_select = {
  64'h0000000000000000,
  64'h0000000000000001,
  64'h0000000000000003,
  ...
  64'h7FFFFFFFFFFFFFFF
};

Each index j selects bit_select[j] to help evaluate whether remaining bits are non-zero.