fallensigh/ciciec2026
Archived
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

perf(mlkem): register butterfly inputs to cut the 100 MHz NTT critical path

Browse Source 
At mlkem_clk=100 MHz the worst path (WNS -2.14 ns) was coeff_bram read ->
butterfly diff (b-a) -> barrett first multiply stage, all in one cycle. The
butterfly took its operands combinationally from the BRAM read (unlike
poly_mul's basecase, which already registers them). Add an input register
stage (a/b/zeta/mode/valid) in butterfly_unit_pipe so the BRAM read and the
first multiply land in separate cycles; feed the a/b/mode alignment delay
lines from the registered inputs. Butterfly latency 8->9; ntt_core writeback
address delay lines deepened 9->10 to match.

Verified byte-exact vs NIST KAT: NTT standalone TB 19/19 (fwd+inv), Encaps
9/9, Decaps 9/9, hello_world e2e — all PASS.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
FallenSigh
2026-07-11 07:05:12 +08:00
parent b2c8d6eaf3
commit 4538ec496f
2 changed files with 30 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
rtl/ip/mlkem/ntt/butterfly_unit_pipe.v
View File

@@ -18,21 +18,32 @@ module butterfly_unit_pipe (
localparam [11:0] Q12 = 12'd3329;
localparam [12:0] Q13 = 13'd3329;
wire [11:0] diff = (b >= a) ? (b - a) : (b - a + Q12);
wire [11:0] mul_data = mode ? diff : b;
// Plan A: register the inputs. Operands arrive combinationally from the
// coeff_bram read; without this stage the path coeff_bram-read -> diff ->
// first multiply stage sits in one cycle (the mlkem_clk critical path at
// 100 MHz). Registering here splits it into (BRAM read -> reg) and
// (reg -> diff -> multiply). Adds one pipeline cycle (absorbed by the NTT
// writeback delay line).
reg [11:0] a_r, b_r, zeta_r;
reg mode_ri, vld_r;
wire [11:0] diff = (b_r >= a_r) ? (b_r - a_r) : (b_r - a_r + Q12);
wire [11:0] mul_data = mode_ri ? diff : b_r;
wire [11:0] mul_result;
wire mul_valid;
barrett_mul_pipe u_barrett (
.clk(clk),
.rst_n(rst_n),
.valid_i(valid_i),
.a(zeta),
.valid_i(vld_r),
.a(zeta_r),
.b(mul_data),
.product(mul_result),
.valid_o(mul_valid)
);
// a/b/mode delay lines (fed from the REGISTERED inputs a_r/b_r/mode_ri so
// a_d6 aligns with mul_valid = valid_i delayed 8).
reg [11:0] a_d0, a_d1, a_d2, a_d3, a_d4, a_d5, a_d6;
reg [11:0] b_d0, b_d1, b_d2, b_d3, b_d4, b_d5, b_d6;
reg mode_d0, mode_d1, mode_d2, mode_d3, mode_d4, mode_d5, mode_d6;
@@ -51,6 +62,7 @@ module butterfly_unit_pipe (
always @(posedge clk or negedge rst_n) begin
if (!rst_n) begin
a_r <= 12'd0; b_r <= 12'd0; zeta_r <= 12'd0; mode_ri <= 1'b0; vld_r <= 1'b0;
a_d0 <= 12'd0; a_d1 <= 12'd0; a_d2 <= 12'd0; a_d3 <= 12'd0; a_d4 <= 12'd0; a_d5 <= 12'd0; a_d6 <= 12'd0;
b_d0 <= 12'd0; b_d1 <= 12'd0; b_d2 <= 12'd0; b_d3 <= 12'd0; b_d4 <= 12'd0; b_d5 <= 12'd0; b_d6 <= 12'd0;
mode_d0 <= 1'b0; mode_d1 <= 1'b0; mode_d2 <= 1'b0; mode_d3 <= 1'b0; mode_d4 <= 1'b0; mode_d5 <= 1'b0; mode_d6 <= 1'b0;
@@ -58,9 +70,12 @@ module butterfly_unit_pipe (
b_out_r <= 12'd0;
valid_r <= 1'b0;
end else begin
a_d0 <= a; a_d1 <= a_d0; a_d2 <= a_d1; a_d3 <= a_d2; a_d4 <= a_d3; a_d5 <= a_d4; a_d6 <= a_d5;
b_d0 <= b; b_d1 <= b_d0; b_d2 <= b_d1; b_d3 <= b_d2; b_d4 <= b_d3; b_d5 <= b_d4; b_d6 <= b_d5;
mode_d0 <= mode; mode_d1 <= mode_d0; mode_d2 <= mode_d1;
// input register stage
a_r <= a; b_r <= b; zeta_r <= zeta; mode_ri <= mode; vld_r <= valid_i;
a_d0 <= a_r; a_d1 <= a_d0; a_d2 <= a_d1; a_d3 <= a_d2; a_d4 <= a_d3; a_d5 <= a_d4; a_d6 <= a_d5;
b_d0 <= b_r; b_d1 <= b_d0; b_d2 <= b_d1; b_d3 <= b_d2; b_d4 <= b_d3; b_d5 <= b_d4; b_d6 <= b_d5;
mode_d0 <= mode_ri; mode_d1 <= mode_d0; mode_d2 <= mode_d1;
mode_d3 <= mode_d2; mode_d4 <= mode_d3; mode_d5 <= mode_d4; mode_d6 <= mode_d5;
valid_r <= mul_valid;

16
rtl/ip/mlkem/ntt/ntt_core.v
View File

@@ -53,8 +53,8 @@ module ntt_core (
reg rd_fire_r; // issue_fire delayed 1 (BRAM read latency) -> feed
// 9-deep writeback address delay lines (BRAM read + 8-stage butterfly).
reg [7:0] a0d[0:8]; reg [7:0] b0d[0:8];
reg [7:0] a1d[0:8]; reg [7:0] b1d[0:8];
reg [7:0] a0d[0:9]; reg [7:0] b0d[0:9];
reg [7:0] a1d[0:9]; reg [7:0] b1d[0:9];
integer k;
reg [DW-1:0] coeff_out_r;
@@ -128,10 +128,10 @@ module ntt_core (
wire wr_en_1 = bf0_valid & ~loading_we;
wire wr_en_2 = bf0_valid & ~loading_we;
wire wr_en_3 = bf0_valid & ~loading_we;
wire [7:0] wr_addr_0 = loading_we ? load_addr : a0d[8];
wire [7:0] wr_addr_1 = b0d[8];
wire [7:0] wr_addr_2 = a1d[8];
wire [7:0] wr_addr_3 = b1d[8];
wire [7:0] wr_addr_0 = loading_we ? load_addr : a0d[9];
wire [7:0] wr_addr_1 = b0d[9];
wire [7:0] wr_addr_2 = a1d[9];
wire [7:0] wr_addr_3 = b1d[9];
wire [11:0] wr_data_0 = loading_we ? coeff_in : bf0_a_out;
wire [11:0] wr_data_1 = bf0_b_out;
wire [11:0] wr_data_2 = bf1_a_out;
@@ -185,7 +185,7 @@ module ntt_core (
layer <= 3'd0; mode_r <= 1'b0;
layer_issue_done <= 1'b0; bf_inflight <= 6'd0;
rd_fire_r <= 1'b0;
for (k = 0; k <= 8; k = k + 1) begin
for (k = 0; k <= 9; k = k + 1) begin
a0d[k] <= 8'd0; b0d[k] <= 8'd0; a1d[k] <= 8'd0; b1d[k] <= 8'd0;
end
coeff_out_r <= 12'd0; valid_o_r <= 1'b0;
@@ -200,7 +200,7 @@ module ntt_core (
// writeback address delay lines: sample j-derived addrs at read cycle
a0d[0] <= cmp_a0; b0d[0] <= cmp_b0; a1d[0] <= cmp_a1; b1d[0] <= cmp_b1;
for (k = 1; k <= 8; k = k + 1) begin
for (k = 1; k <= 9; k = k + 1) begin
a0d[k] <= a0d[k-1]; b0d[k] <= b0d[k-1];
a1d[k] <= a1d[k-1]; b1d[k] <= b1d[k-1];
end