fallensigh/mlkem-sync
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Reduce NTT butterfly dispatch latency

Browse Source
This commit is contained in:
FallenSigh
2026-07-08 00:57:05 +08:00
parent ce998bb49a
commit a0bd07bdf8
1 changed files with 3 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
sync_rtl/ntt/ntt_core.v
View File

@@ -2,7 +2,7 @@
// //
// Uses 256 individual 12-bit registers and a deeply pipelined butterfly path. // Uses 256 individual 12-bit registers and a deeply pipelined butterfly path.
// The arithmetic hot path is split into: // The arithmetic hot path is split into:
// address -> operand/zeta register -> pipelined Barrett butterfly -> writeback // address/operand/zeta register -> pipelined Barrett butterfly -> writeback
// In inverse mode, final x3303 output scaling also uses a pipelined Barrett // In inverse mode, final x3303 output scaling also uses a pipelined Barrett
// multiplier so the output path does not reintroduce a combinational reducer. // multiplier so the output path does not reintroduce a combinational reducer.
@@ -96,7 +96,7 @@ module ntt_core (
case (state) case (state)
S_IDLE: if (valid_i) next_state = S_LOAD; S_IDLE: if (valid_i) next_state = S_LOAD;
S_LOAD: if (load_cnt >= 8'd255 && valid_i) next_state = S_CMP_A; S_LOAD: if (load_cnt >= 8'd255 && valid_i) next_state = S_CMP_A;
S_CMP_A: next_state = bf_done ? S_OUT_PREP : S_CMP_B; S_CMP_A: next_state = bf_done ? S_OUT_PREP : S_CMP_ISSUE;
S_CMP_B: next_state = bf_done ? S_OUT_PREP : S_CMP_ISSUE; S_CMP_B: next_state = bf_done ? S_OUT_PREP : S_CMP_ISSUE;
S_CMP_ISSUE: next_state = S_CMP_WAIT; S_CMP_ISSUE: next_state = S_CMP_WAIT;
S_CMP_WAIT: if (bf_valid) next_state = S_CMP_WB; S_CMP_WAIT: if (bf_valid) next_state = S_CMP_WB;
@@ -169,12 +169,9 @@ module ntt_core (
load_cnt <= load_cnt + 8'd1; load_cnt <= load_cnt + 8'd1;
end end
if (state == S_CMP_A) begin if (state == S_CMP_A && !bf_done) begin
r_wa <= j; r_wa <= j;
r_wb <= j + layer_len; r_wb <= j + layer_len;
end
if (state == S_CMP_B) begin
r_a <= cr[j]; r_a <= cr[j];
r_b <= cr[j + layer_len]; r_b <= cr[j + layer_len];
r_zeta <= zeta; r_zeta <= zeta;