feat(poly_mul): implement synchronous PolyMul with base-case multiply

Phase 2.2: NTT-domain polynomial pointwise multiplication.
- basecase_mul.v: degree-1 base-case multiply (c0,c1) with Barrett
- poly_mul_zeta_rom.v: 128-entry zeta ROM for PolyMul
- poly_mul_sync.v: FSM (IDLE→LOAD 256 cycles→COMPUTE 256 cycles→DONE)

Verified: 5/5 vectors bit-exact vs Python PolyMul reference
This commit is contained in:
2026-06-24 23:10:18 +08:00
parent c4cd10c2c1
commit 39dd36994b
6 changed files with 604 additions and 0 deletions

View File

@@ -0,0 +1,166 @@
// poly_mul_sync.v - Synchronous NTT-domain polynomial multiplier
//
// Computes pointwise (Karatsuba-like base-case) multiplication of two
// 256-coefficient NTT-domain polynomials.
//
// Operation flow:
// IDLE LOAD (256× A+B pairs) COMP_CALC (read+compute, 1 cycle)
// COMP_C0 (output c0) COMP_C1 (output c1) DONE IDLE
//
// The LOAD phase accepts both A and B coefficients simultaneously
// (one pair per cycle) on coeff_a_in/coeff_b_in.
//
// The COMPUTE phase outputs the 256 result coefficients one per cycle
// via valid/ready handshake on coeff_out/valid_o/ready_i.
//
// Memory: 256×12-bit register arrays for A and B coefficients.
//
// Interface:
// clk, rst_n - Clock, active-low reset
// coeff_a_in[11:0]- Polynomial A coefficient input
// coeff_b_in[11:0]- Polynomial B coefficient input
// valid_i - Input valid
// ready_o - Ready to accept input (high in IDLE/LOAD)
// coeff_out[11:0] - Result coefficient output
// valid_o - Output valid (high in COMP_C0/COMP_C1)
// ready_i - Output consumer ready
module poly_mul_sync (
input clk, rst_n,
input [11:0] coeff_a_in,
input [11:0] coeff_b_in,
input valid_i,
output ready_o,
output [11:0] coeff_out,
output valid_o,
input ready_i
);
// State definitions
localparam S_IDLE = 3'd0;
localparam S_LOAD = 3'd1;
localparam S_COMP_CALC = 3'd2;
localparam S_COMP_C0 = 3'd3;
localparam S_COMP_C1 = 3'd4;
localparam S_DONE = 3'd5;
reg [2:0] state, next_state;
// Coefficient storage (register arrays)
reg [11:0] mem_A [0:255];
reg [11:0] mem_B [0:255];
// Counters
reg [7:0] load_cnt; // 0..256 for loading 256 pairs
reg [6:0] comp_k; // 0..127, current base-case index
// Registered basecase_mul results
reg [11:0] c0_reg, c1_reg;
// Combinational read signals for COMP_CALC
wire [7:0] addr_even = {comp_k, 1'b0}; // comp_k * 2 (7+1 = 8 bits)
wire [7:0] addr_odd = {comp_k, 1'b1}; // comp_k * 2 + 1
wire [11:0] mem_a0 = mem_A[addr_even];
wire [11:0] mem_a1 = mem_A[addr_odd];
wire [11:0] mem_b0 = mem_B[addr_even];
wire [11:0] mem_b1 = mem_B[addr_odd];
// Zeta ROM
wire [11:0] zeta;
poly_mul_zeta_rom u_zeta (
.addr (comp_k),
.zeta (zeta)
);
// Basecase multiply
wire [11:0] bc_c0, bc_c1;
basecase_mul u_bc (
.a0 (mem_a0),
.a1 (mem_a1),
.b0 (mem_b0),
.b1 (mem_b1),
.zeta(zeta),
.c0 (bc_c0),
.c1 (bc_c1)
);
// Output interface
assign ready_o = (state == S_IDLE) || (state == S_LOAD);
assign valid_o = (state == S_COMP_C0) || (state == S_COMP_C1);
assign coeff_out = (state == S_COMP_C0) ? c0_reg : c1_reg;
// State transition logic (combinational)
always @* begin
next_state = state;
case (state)
S_IDLE: if (valid_i && ready_o) next_state = S_LOAD;
S_LOAD: if (load_cnt >= 255 && valid_i && ready_o)
next_state = S_COMP_CALC;
S_COMP_CALC: next_state = S_COMP_C0;
S_COMP_C0: if (valid_o && ready_i) next_state = S_COMP_C1;
S_COMP_C1: if (valid_o && ready_i) begin
if (comp_k >= 127) next_state = S_DONE;
else next_state = S_COMP_CALC;
end
S_DONE: next_state = S_IDLE;
default: next_state = S_IDLE;
endcase
end
// Sequential logic
integer i;
always @(posedge clk or negedge rst_n) begin
if (!rst_n) begin
state <= S_IDLE;
load_cnt <= 8'd0;
comp_k <= 7'd0;
c0_reg <= 12'd0;
c1_reg <= 12'd0;
for (i = 0; i < 256; i = i + 1) begin
mem_A[i] <= 12'd0;
mem_B[i] <= 12'd0;
end
end else begin
state <= next_state;
// ---- LOAD phase ----
// First coefficient captured on IDLE LOAD transition
if (state == S_IDLE && valid_i && ready_o) begin
mem_A[0] <= coeff_a_in;
mem_B[0] <= coeff_b_in;
load_cnt <= 8'd1;
end
// Subsequent coefficients in LOAD state
if (state == S_LOAD && valid_i && ready_o) begin
mem_A[load_cnt] <= coeff_a_in;
mem_B[load_cnt] <= coeff_b_in;
load_cnt <= load_cnt + 8'd1;
end
// ---- COMPUTE phase ----
// COMP_CALC: capture basecase_mul results
if (state == S_COMP_CALC) begin
c0_reg <= bc_c0;
c1_reg <= bc_c1;
end
// COMP_C0 COMP_C1: c0 was consumed, increment comp_k
if (state == S_COMP_C0 && valid_o && ready_i) begin
// comp_k stays same, c1 still to output
end
// COMP_C1 COMP_CALC: c1 was consumed, advance to next pair
if (state == S_COMP_C1 && valid_o && ready_i) begin
comp_k <= comp_k + 7'd1;
end
// ---- DONE IDLE: reset counters ----
if (state == S_DONE) begin
load_cnt <= 8'd0;
comp_k <= 7'd0;
end
end
end
endmodule