refactor(cam): remove read noise from noise architecture (Phase 2)

- Make cam_read_noise a pass-through module, removing all noise injection logic - Switch write noise to use noise_mask_bernoulli instead of noise_mask_grouped - Add state machine to cam_write_noise for mask generation timing - Remove noise_mask_grouped.sv (no longer needed) - Remove read noise parameters from cam_noisy and cam_top - Update simulation and benchmark code to reflect read noise removal - Sync documentation to reflect Phase 2 architecture
2026-07-12 20:15:31 +08:00 · 2026-05-26 23:02:22 +08:00
parent e5d13917b2
commit 8b4d4c1b57
29 changed files with 277 additions and 863 deletions
--- a/hw/rtl/cam_noisy.sv
+++ b/hw/rtl/cam_noisy.sv
@@ -5,13 +5,7 @@ module cam_noisy #(
    parameter bit          WRITE_NOISE_EN       = 1'b1,
    parameter int          WRITE_NOISE_RATE_NUM = 1,
    parameter int          WRITE_NOISE_RATE_DEN = 100,
-    parameter int          WRITE_NOISE_BITS     = 8,
-    parameter logic [63:0] WRITE_NOISE_SEED     = 64'hB504_F32D_B504_F32D,
-    parameter bit          READ_NOISE_EN        = 1'b1,
-    parameter int          READ_NOISE_RATE_NUM  = 1,
-    parameter int          READ_NOISE_RATE_DEN  = 100,
-    parameter int          READ_NOISE_BITS      = 8,
-    parameter logic [63:0] READ_NOISE_SEED      = 64'h6A09_E667_F3BC_C909
+    parameter logic [63:0] WRITE_NOISE_SEED     = 64'hB504_F32D_B504_F32D
 ) (
    input  logic                              clk,
    input  logic                              rst_n,
@@ -42,7 +36,6 @@ module cam_noisy #(
        .WRITE_NOISE_EN       (WRITE_NOISE_EN),
        .WRITE_NOISE_RATE_NUM (WRITE_NOISE_RATE_NUM),
        .WRITE_NOISE_RATE_DEN (WRITE_NOISE_RATE_DEN),
-        .WRITE_NOISE_BITS     (WRITE_NOISE_BITS),
        .WRITE_NOISE_SEED     (WRITE_NOISE_SEED)
    ) u_write_noise (
        .clk           (clk),
@@ -73,13 +66,7 @@ module cam_noisy #(
    );

    // ── Read noise pipeline ──
-    cam_read_noise #(
-        .READ_NOISE_EN       (READ_NOISE_EN),
-        .READ_NOISE_RATE_NUM (READ_NOISE_RATE_NUM),
-        .READ_NOISE_RATE_DEN (READ_NOISE_RATE_DEN),
-        .READ_NOISE_BITS     (READ_NOISE_BITS),
-        .READ_NOISE_SEED     (READ_NOISE_SEED)
-    ) u_read_noise (
+    cam_read_noise u_read_noise (
        .clk             (clk),
        .rst_n           (rst_n),
        .valid_i         (core_rd_valid),
--- a/hw/rtl/cam_top.sv
+++ b/hw/rtl/cam_top.sv
@@ -5,13 +5,7 @@ module cam_top #(
    parameter bit           WRITE_NOISE_EN       = 1'b1,
    parameter int           WRITE_NOISE_RATE_NUM = 1,
    parameter int           WRITE_NOISE_RATE_DEN = 100,
-    parameter int           WRITE_NOISE_BITS     = 8,
-    parameter logic [63:0]  WRITE_NOISE_SEED     = 64'hB504_F32D_B504_F32D,
-    parameter bit           READ_NOISE_EN        = 1'b1,
-    parameter int           READ_NOISE_RATE_NUM  = 1,
-    parameter int           READ_NOISE_RATE_DEN  = 100,
-    parameter int           READ_NOISE_BITS      = 8,
-    parameter logic [63:0]  READ_NOISE_SEED      = 64'h6A09_E667_F3BC_C909
+    parameter logic [63:0]  WRITE_NOISE_SEED     = 64'hB504_F32D_B504_F32D
 ) (
    input  logic clk,
    input  logic rst_n,
@@ -77,13 +71,7 @@ module cam_top #(
        .WRITE_NOISE_EN       (WRITE_NOISE_EN),
        .WRITE_NOISE_RATE_NUM (WRITE_NOISE_RATE_NUM),
        .WRITE_NOISE_RATE_DEN (WRITE_NOISE_RATE_DEN),
-        .WRITE_NOISE_BITS     (WRITE_NOISE_BITS),
-        .WRITE_NOISE_SEED     (WRITE_NOISE_SEED),
-        .READ_NOISE_EN        (READ_NOISE_EN),
-        .READ_NOISE_RATE_NUM  (READ_NOISE_RATE_NUM),
-        .READ_NOISE_RATE_DEN  (READ_NOISE_RATE_DEN),
-        .READ_NOISE_BITS      (READ_NOISE_BITS),
-        .READ_NOISE_SEED      (READ_NOISE_SEED)
+        .WRITE_NOISE_SEED     (WRITE_NOISE_SEED)
    ) u_noisy (
        .clk              (clk),
        .rst_n            (rst_n),
--- a/hw/rtl/noise/cam_read_noise.sv
+++ b/hw/rtl/noise/cam_read_noise.sv
@@ -1,13 +1,7 @@
 `timescale 1ns / 1ps
 `include "cam_params.svh"

-module cam_read_noise #(
-    parameter bit          READ_NOISE_EN       = 1'b1,
-    parameter int          READ_NOISE_RATE_NUM = 1,
-    parameter int          READ_NOISE_RATE_DEN = 100,
-    parameter int          READ_NOISE_BITS     = 8,
-    parameter logic [63:0] READ_NOISE_SEED     = 64'h6A09_E667_F3BC_C909
-) (
+module cam_read_noise (
    input  logic                              clk,
    input  logic                              rst_n,
    input  logic                              valid_i,
@@ -19,69 +13,19 @@ module cam_read_noise #(
    output logic [(`LANES)*(`HASH_BITS)-1:0]  hashes_noisy_o,
    output logic [(`LANES)-1:0]               lane_valid_o
 );
-    logic                             valid_q;
-    logic [(`LANES)*(`ROW_BITS)-1:0]  row_ids_q;
-    logic [(`LANES)*(`HASH_BITS)-1:0] hashes_q;
-    logic [(`LANES)-1:0]              lane_valid_q;
-
-    logic [127:0] random_num [0:`LANES-1];
-    logic [(`HASH_BITS)-1:0] mask [0:`LANES-1];
-
-    `ifndef SYNTHESIS
-    initial begin
-        if (READ_NOISE_SEED == 64'd0) $fatal(1, "READ_NOISE_SEED must be nonzero");
-    end
-    `endif
-
-    generate
-        for (genvar lane = 0; lane < `LANES; lane++) begin : gen_lane_noise
-            localparam logic [63:0] LANE_SALT = 64'(lane + 1) * 64'h9E37_79B9_7F4A_7C15;
-            random128 u_random_read (
-                .clk   (clk),
-                .rst_n (rst_n),
-                .enable(valid_i && lane_valid_i[lane] && READ_NOISE_EN && (READ_NOISE_RATE_NUM > 0)),
-                .seed  ({(READ_NOISE_SEED ^ LANE_SALT), (READ_NOISE_SEED ^ LANE_SALT)}),
-                .out   (random_num[lane])
-            );
-
-            noise_mask_grouped #(
-                .HASH_BITS      (`HASH_BITS),
-                .NOISE_BITS     (READ_NOISE_BITS),
-                .NOISE_RATE_NUM (READ_NOISE_RATE_NUM),
-                .NOISE_RATE_DEN (READ_NOISE_RATE_DEN)
-            ) u_mask (
-                .random_i(random_num[lane]),
-                .mask_o  (mask[lane])
-            );
-        end
-    endgenerate

    always_ff @(posedge clk or negedge rst_n) begin
        if (!rst_n) begin
-            valid_q        <= 1'b0;
-            row_ids_q      <= '0;
-            hashes_q       <= '0;
-            lane_valid_q   <= '0;
            valid_o        <= 1'b0;
            row_ids_o      <= '0;
            hashes_noisy_o <= '0;
            lane_valid_o   <= '0;
        end else begin
-            valid_q      <= valid_i;
-            row_ids_q    <= row_ids_i;
-            hashes_q     <= hashes_i;
-            lane_valid_q <= lane_valid_i;
-
-            valid_o      <= valid_q;
-            row_ids_o    <= row_ids_q;
-            lane_valid_o <= lane_valid_q;
-            for (int lane = 0; lane < `LANES; lane++) begin
-                if (READ_NOISE_EN && (READ_NOISE_RATE_NUM > 0) && lane_valid_q[lane]) begin
-                    hashes_noisy_o[lane*`HASH_BITS +: `HASH_BITS] <= hashes_q[lane*`HASH_BITS +: `HASH_BITS] ^ mask[lane];
-                end else begin
-                    hashes_noisy_o[lane*`HASH_BITS +: `HASH_BITS] <= hashes_q[lane*`HASH_BITS +: `HASH_BITS];
-                end
-            end
+            valid_o        <= valid_i;
+            row_ids_o      <= row_ids_i;
+            hashes_noisy_o <= hashes_i;
+            lane_valid_o   <= lane_valid_i;
        end
    end
+
 endmodule
--- a/hw/rtl/noise/cam_write_noise.sv
+++ b/hw/rtl/noise/cam_write_noise.sv
@@ -5,7 +5,6 @@ module cam_write_noise #(
    parameter bit          WRITE_NOISE_EN       = 1'b1,
    parameter int          WRITE_NOISE_RATE_NUM = 1,
    parameter int          WRITE_NOISE_RATE_DEN = 100,
-    parameter int          WRITE_NOISE_BITS     = 8,
    parameter logic [63:0] WRITE_NOISE_SEED     = 64'hB504_F32D_B504_F32D
 ) (
    input  logic                    clk,
@@ -18,30 +17,46 @@ module cam_write_noise #(
    output logic [(`ROW_BITS)-1:0]  core_wr_row,
    output logic [(`HASH_BITS)-1:0] core_wr_hash
 );
-    logic pending_q;
-    logic [(`ROW_BITS)-1:0]  row_q;
-    logic [(`HASH_BITS)-1:0] hash_q;
-    logic [127:0] random_num;
+
+    localparam int PROB_BITS = 8;
+    localparam int SAMPLE_RANGE = 1 << PROB_BITS;
+    localparam int WRITE_NOISE_THRESHOLD_RAW =
+        (WRITE_NOISE_RATE_NUM * SAMPLE_RANGE) / WRITE_NOISE_RATE_DEN;
+    localparam int WRITE_NOISE_THRESHOLD =
+        (WRITE_NOISE_THRESHOLD_RAW > (SAMPLE_RANGE - 1)) ?
+            (SAMPLE_RANGE - 1) : WRITE_NOISE_THRESHOLD_RAW;
+
+    wire noise_active = WRITE_NOISE_EN && (WRITE_NOISE_RATE_NUM > 0) && (WRITE_NOISE_THRESHOLD > 0);
+
+    typedef enum logic [1:0] {
+        STATE_IDLE,
+        STATE_WAIT_MASK
+    } state_t;
+
+    state_t state_q;
+    logic mask_start_q;
+    logic mask_busy;
+    logic mask_done;
    logic [(`HASH_BITS)-1:0] flip_mask;
+    logic [(`ROW_BITS)-1:0] row_q;
+    logic [(`HASH_BITS)-1:0] hash_q;

-    assign wr_ready = !pending_q;
+    assign wr_ready = (state_q == STATE_IDLE);

-    random128 u_random_write (
-        .clk   (clk),
-        .rst_n (rst_n),
-        .enable(wr_valid && wr_ready && WRITE_NOISE_EN && (WRITE_NOISE_RATE_NUM > 0)),
-        .seed  ({WRITE_NOISE_SEED, WRITE_NOISE_SEED}),
-        .out   (random_num)
-    );
-
-    noise_mask_grouped #(
+    noise_mask_bernoulli #(
        .HASH_BITS      (`HASH_BITS),
-        .NOISE_BITS     (WRITE_NOISE_BITS),
-        .NOISE_RATE_NUM (WRITE_NOISE_RATE_NUM),
-        .NOISE_RATE_DEN (WRITE_NOISE_RATE_DEN)
-    ) u_mask (
-        .random_i(random_num),
-        .mask_o  (flip_mask)
+        .PROB_BITS      (PROB_BITS),
+        .PRNG_WORDS     (2),
+        .BITS_PER_CYCLE (32),
+        .SEED           (WRITE_NOISE_SEED)
+    ) u_bernoulli_mask (
+        .clk         (clk),
+        .rst_n       (rst_n),
+        .start_i     (mask_start_q),
+        .threshold_i (PROB_BITS'(WRITE_NOISE_THRESHOLD)),
+        .busy_o      (mask_busy),
+        .done_o      (mask_done),
+        .mask_o      (flip_mask)
    );

    `ifndef SYNTHESIS
@@ -52,26 +67,47 @@ module cam_write_noise #(

    always_ff @(posedge clk or negedge rst_n) begin
        if (!rst_n) begin
-            pending_q     <= 1'b0;
-            row_q         <= '0;
-            hash_q        <= '0;
+            state_q <= STATE_IDLE;
+            mask_start_q <= 1'b0;
+            row_q <= '0;
+            hash_q <= '0;
            core_wr_valid <= 1'b0;
-            core_wr_row   <= '0;
-            core_wr_hash  <= '0;
+            core_wr_row <= '0;
+            core_wr_hash <= '0;
        end else begin
-            core_wr_valid <= pending_q;
-            core_wr_row   <= row_q;
-            if (WRITE_NOISE_EN && (WRITE_NOISE_RATE_NUM > 0)) begin
-                core_wr_hash <= hash_q ^ flip_mask;
-            end else begin
-                core_wr_hash <= hash_q;
-            end
+            core_wr_valid <= 1'b0;
+            mask_start_q <= 1'b0;

-            pending_q <= wr_valid && wr_ready;
-            if (wr_valid && wr_ready) begin
-                row_q  <= wr_row;
-                hash_q <= wr_hash;
-            end
+            unique case (state_q)
+                STATE_IDLE: begin
+                    if (wr_valid && wr_ready) begin
+                        row_q <= wr_row;
+                        hash_q <= wr_hash;
+                        if (noise_active) begin
+                            mask_start_q <= 1'b1;
+                            state_q <= STATE_WAIT_MASK;
+                        end else begin
+                            // Noise inactive: pass through immediately (one-cycle)
+                            core_wr_valid <= 1'b1;
+                            core_wr_row <= wr_row;
+                            core_wr_hash <= wr_hash;
+                        end
+                    end
+                end
+
+                STATE_WAIT_MASK: begin
+                    if (mask_done) begin
+                        core_wr_valid <= 1'b1;
+                        core_wr_row <= row_q;
+                        core_wr_hash <= hash_q ^ flip_mask;
+                        state_q <= STATE_IDLE;
+                    end
+                end
+
+                default: begin
+                    state_q <= STATE_IDLE;
+                end
+            endcase
        end
    end
 endmodule
--- a/hw/rtl/noise/noise_mask_grouped.sv
+++ b/hw/rtl/noise/noise_mask_grouped.sv
@@ -1,39 +0,0 @@
-`timescale 1ns / 1ps
-`include "cam_params.svh"
-
-module noise_mask_grouped #(
-    parameter int HASH_BITS      = `HASH_BITS,
-    parameter int NOISE_BITS     = 8,
-    parameter int NOISE_RATE_NUM = 1,
-    parameter int NOISE_RATE_DEN = 100
-) (
-    input  logic [127:0]          random_i,
-    output logic [HASH_BITS-1:0]  mask_o
-);
-    localparam int GROUP_BITS       = HASH_BITS / NOISE_BITS;
-    localparam int BIT_INDEX_BITS   = 6;
-    localparam int SAMPLE_BITS      = 8;
-    localparam int GROUP_RAND_BITS  = BIT_INDEX_BITS + SAMPLE_BITS;
-    localparam int SAMPLE_RANGE     = 1 << SAMPLE_BITS;
-    localparam int THRESHOLD        = (NOISE_RATE_NUM * SAMPLE_RANGE) / NOISE_RATE_DEN;
-
-    `ifndef SYNTHESIS
-    initial begin
-        if (NOISE_BITS <= 0) $fatal(1, "NOISE_BITS must be > 0");
-        if (HASH_BITS % NOISE_BITS != 0) $fatal(1, "HASH_BITS must be divisible by NOISE_BITS");
-        if (GROUP_BITS != 64) $fatal(1, "GROUP_BITS must be 64 for 6-bit grouped noise");
-        if (NOISE_BITS * GROUP_RAND_BITS > 128) $fatal(1, "NOISE_BITS consumes more than 128 random bits");
-        if (NOISE_RATE_DEN <= 0) $fatal(1, "NOISE_RATE_DEN must be > 0");
-        if (NOISE_RATE_NUM < 0 || NOISE_RATE_NUM > NOISE_RATE_DEN) $fatal(1, "NOISE_RATE_NUM out of range");
-    end
-    `endif
-
-    always_comb begin
-        mask_o = '0;
-        for (int i = 0; i < NOISE_BITS; i++) begin
-            if (random_i[i * GROUP_RAND_BITS + BIT_INDEX_BITS +: SAMPLE_BITS] < THRESHOLD) begin
-                mask_o[i * GROUP_BITS + random_i[i * GROUP_RAND_BITS +: BIT_INDEX_BITS]] = 1'b1;
-            end
-        end
-    end
-endmodule
--- a/hw/sim/benchmarks/retrieval/Makefile
+++ b/hw/sim/benchmarks/retrieval/Makefile
@@ -9,7 +9,6 @@ VERILOG_SOURCES := $(RTL_CAM_TOP)
 TOPK_K ?= 5
 NUM_ROWS ?= 4096
 WRITE_NOISE_EN ?= 0
-READ_NOISE_EN ?= 0

 CAM_RETRIEVAL_DATASET ?=
 export CAM_RETRIEVAL_DATASET
--- a/hw/sim/benchmarks/retrieval/test_retrieval_benchmark.py
+++ b/hw/sim/benchmarks/retrieval/test_retrieval_benchmark.py
@@ -12,10 +12,8 @@ import numpy as np
 from cocotb.clock import Clock

 from model.ref_model import (
-    lane_seed_128,
    match_topk,
    match_topk_from_scores,
-    score_rows_with_read_noise,
 )
 from tests.top.utils import (
    dut_hash_bits,
@@ -199,13 +197,9 @@ def compute_metrics(topk_indices: list[int], row_labels: list[int], query_label:
    return precision, recall, f1


-def mode_from_params(write_noise_en: int, read_noise_en: int) -> str:
-    if write_noise_en and read_noise_en:
-        return "write_read_noise"
+def mode_from_params(write_noise_en: int) -> str:
    if write_noise_en:
        return "write_noise"
-    if read_noise_en:
-        return "read_noise"
    return "no_noise"


@@ -228,8 +222,8 @@ def write_outputs(out_dir: Path, result: dict) -> None:

    fieldnames = [
        "run_id", "mode", "num_rows", "hash_bits", "lanes", "topk_k",
-        "write_noise_en", "read_noise_en", "write_noise_rate_num",
-        "write_noise_rate_den", "read_noise_rate_num", "read_noise_rate_den",
+        "write_noise_en", "write_noise_rate_num",
+        "write_noise_rate_den",
        "num_queries", "k", "macro_precision", "retrieval_recall", "macro_f1",
        "recall@k", "exact_match_rate", "status",
    ]
@@ -245,11 +239,8 @@ def write_outputs(out_dir: Path, result: dict) -> None:
                "lanes": result["params"]["lanes"],
                "topk_k": result["params"]["topk_k"],
                "write_noise_en": result["params"]["write_noise_en"],
-                "read_noise_en": result["params"]["read_noise_en"],
                "write_noise_rate_num": result["params"]["write_noise_rate_num"],
                "write_noise_rate_den": result["params"]["write_noise_rate_den"],
-                "read_noise_rate_num": result["params"]["read_noise_rate_num"],
-                "read_noise_rate_den": result["params"]["read_noise_rate_den"],
                "num_queries": result["dataset"]["num_queries"],
                "k": int(k),
                "macro_precision": metrics["macro_precision"],
@@ -293,13 +284,9 @@ async def cam_retrieval_benchmark(dut):
    hash_bits = dut_hash_bits(dut)
    lanes = dut_lanes(dut)
    write_noise_en = int(get_param(dut, "WRITE_NOISE_EN", 0) or 0)
-    read_noise_en = int(get_param(dut, "READ_NOISE_EN", 0) or 0)
    write_noise_rate_num = int(get_param(dut, "WRITE_NOISE_RATE_NUM", 0) or 0)
    write_noise_rate_den = int(get_param(dut, "WRITE_NOISE_RATE_DEN", 100) or 100)
-    read_noise_rate_num = int(get_param(dut, "READ_NOISE_RATE_NUM", 0) or 0)
-    read_noise_rate_den = int(get_param(dut, "READ_NOISE_RATE_DEN", 100) or 100)
-    read_noise_bits = int(get_param(dut, "READ_NOISE_BITS", 8) or 8)
-    mode = mode_from_params(write_noise_en, read_noise_en)
+    mode = mode_from_params(write_noise_en)

    if write_noise_en:
        raise AssertionError("First retrieval benchmark version only supports WRITE_NOISE_EN=0")
@@ -315,7 +302,6 @@ async def cam_retrieval_benchmark(dut):
    await write_rows(dut, dataset.rows)

    accumulators = {k: MetricAccumulator() for k in BENCHMARK_KS}
-    read_lane_states = [lane_seed_128(0x6A09_E667_F3BC_C909, lane) for lane in range(lanes)]

    for query, query_label in zip(dataset.queries, dataset.query_labels):
        beats, _, _, _ = await query_topk_once(dut, query)
@@ -324,15 +310,7 @@ async def cam_retrieval_benchmark(dut):

        dut_topk = [int(beat[1]) for beat in beats[: max(BENCHMARK_KS)]]

-        if read_noise_en:
-            scores, read_lane_states = score_rows_with_read_noise(
-                query, dataset.rows, lane_states=read_lane_states,
-                width=hash_bits, lanes=lanes, noise_bits=read_noise_bits,
-                rate_num=read_noise_rate_num, rate_den=read_noise_rate_den,
-            )
-            golden_topk = match_topk_from_scores(scores, max(BENCHMARK_KS))
-        else:
-            golden_topk, _ = match_topk(query, dataset.rows, width=hash_bits, k=max(BENCHMARK_KS))
+        golden_topk, _ = match_topk(query, dataset.rows, width=hash_bits, k=max(BENCHMARK_KS))

        for k in BENCHMARK_KS:
            precision, recall, f1 = compute_metrics(dut_topk, dataset.row_labels, query_label, k)
@@ -352,11 +330,8 @@ async def cam_retrieval_benchmark(dut):
            "lanes": lanes,
            "topk_k": max(BENCHMARK_KS),
            "write_noise_en": write_noise_en,
-            "read_noise_en": read_noise_en,
            "write_noise_rate_num": write_noise_rate_num,
            "write_noise_rate_den": write_noise_rate_den,
-            "read_noise_rate_num": read_noise_rate_num,
-            "read_noise_rate_den": read_noise_rate_den,
        },
        "dataset": {
            "num_classes": dataset.num_classes,
--- a/hw/sim/mk/cocotb-common.mk
+++ b/hw/sim/mk/cocotb-common.mk
@@ -37,11 +37,6 @@ COMPILE_ARGS += $(EXTRA_DEFINES)
 WRITE_NOISE_EN          ?= $(NOISE_EN)
 WRITE_NOISE_RATE_NUM    ?= $(NOISE_RATE_NUM)
 WRITE_NOISE_RATE_DEN    ?= $(NOISE_RATE_DEN)
-WRITE_NOISE_BITS        ?= $(NOISE_BITS)
-READ_NOISE_EN           ?= $(NOISE_EN)
-READ_NOISE_RATE_NUM     ?= $(NOISE_RATE_NUM)
-READ_NOISE_RATE_DEN     ?= $(NOISE_RATE_DEN)
-READ_NOISE_BITS         ?= $(NOISE_BITS)

 ifneq ($(strip $(WRITE_NOISE_EN)),)
 COMPILE_ARGS += -GWRITE_NOISE_EN=$(WRITE_NOISE_EN)
@@ -52,21 +47,6 @@ endif
 ifneq ($(strip $(WRITE_NOISE_RATE_DEN)),)
 COMPILE_ARGS += -GWRITE_NOISE_RATE_DEN=$(WRITE_NOISE_RATE_DEN)
 endif
-ifneq ($(strip $(WRITE_NOISE_BITS)),)
-COMPILE_ARGS += -GWRITE_NOISE_BITS=$(WRITE_NOISE_BITS)
-endif
-ifneq ($(strip $(READ_NOISE_EN)),)
-COMPILE_ARGS += -GREAD_NOISE_EN=$(READ_NOISE_EN)
-endif
-ifneq ($(strip $(READ_NOISE_RATE_NUM)),)
-COMPILE_ARGS += -GREAD_NOISE_RATE_NUM=$(READ_NOISE_RATE_NUM)
-endif
-ifneq ($(strip $(READ_NOISE_RATE_DEN)),)
-COMPILE_ARGS += -GREAD_NOISE_RATE_DEN=$(READ_NOISE_RATE_DEN)
-endif
-ifneq ($(strip $(READ_NOISE_BITS)),)
-COMPILE_ARGS += -GREAD_NOISE_BITS=$(READ_NOISE_BITS)
-endif

 export PYTHONPATH := $(SIM_ROOT):$(PYTHONPATH)
 export QUIET ?= 1
--- a/hw/sim/mk/rtl-sources.mk
+++ b/hw/sim/mk/rtl-sources.mk
@@ -4,10 +4,9 @@ endif

 RTL_RANDOM := $(RTL_ROOT)/random/random128.sv

-RTL_NOISE_MASK := $(RTL_ROOT)/noise/noise_mask_grouped.sv
 RTL_BERNOULLI_NOISE_MASK := $(RTL_ROOT)/noise/noise_mask_bernoulli.sv $(RTL_RANDOM)
-RTL_WRITE_NOISE := $(RTL_NOISE_MASK) $(RTL_RANDOM) $(RTL_ROOT)/noise/cam_write_noise.sv
-RTL_READ_NOISE := $(RTL_NOISE_MASK) $(RTL_RANDOM) $(RTL_ROOT)/noise/cam_read_noise.sv
+RTL_WRITE_NOISE := $(RTL_BERNOULLI_NOISE_MASK) $(RTL_ROOT)/noise/cam_write_noise.sv
+RTL_READ_NOISE := $(RTL_ROOT)/noise/cam_read_noise.sv

 RTL_CAM_CORE_BANKED := $(RTL_ROOT)/core/cam_core_banked.sv
 RTL_MATCH_ENGINE := \
--- a/hw/sim/model/ref_model.py
+++ b/hw/sim/model/ref_model.py
@@ -94,204 +94,6 @@ def xorshift128(state: int) -> int:
    return ((next_x << 96) | (next_y << 64) | (next_z << 32) | next_w) & mask128


-def generate_write_flip_mask(
-    prng_state: int,
-    hash_bits: int,
-    noise_bits: int,
-    rate_num: int,
-    rate_den: int,
-) -> tuple[int, int]:
-    """Generate one write-noise flip mask using one xorshift128 step."""
-    assert hash_bits % noise_bits == 0
-    group_bits = hash_bits // noise_bits
-    bit_index_bits = 6
-    sample_bits = 8
-    group_random_bits = bit_index_bits + sample_bits
-    assert group_bits == 64
-    assert noise_bits * group_random_bits <= 128
-
-    sample_range = 1 << sample_bits
-    threshold = (rate_num * sample_range) // rate_den
-
-    state = xorshift128(prng_state)
-    mask = 0
-    for group_idx in range(noise_bits):
-        group_rand = (state >> (group_idx * group_random_bits)) & ((1 << group_random_bits) - 1)
-        bit_idx = group_rand & ((1 << bit_index_bits) - 1)
-        sample = (group_rand >> bit_index_bits) & (sample_range - 1)
-        if sample < threshold:
-            mask |= 1 << (group_idx * group_bits + bit_idx)
-
-    return mask, state
-
-
-def generate_grouped_flip_mask(
-    *,
-    random_value: int,
-    hash_bits: int,
-    noise_bits: int,
-    rate_num: int,
-    rate_den: int,
-) -> int:
-    """Generate a grouped flip mask from one 128-bit value.
-
-    This is the shared write/read noise model: 8 default 64-bit groups, one
-    candidate flip per group, 6-bit bit index and 8-bit threshold sample.
-    It is not independent Bernoulli sampling over all 512 bits.
-    """
-    assert noise_bits > 0
-    assert hash_bits % noise_bits == 0
-    group_bits = hash_bits // noise_bits
-    bit_index_bits = 6
-    sample_bits = 8
-    group_random_bits = bit_index_bits + sample_bits
-    assert group_bits == 64
-    assert noise_bits * group_random_bits <= 128
-    assert rate_den > 0
-    assert 0 <= rate_num <= rate_den
-
-    sample_range = 1 << sample_bits
-    threshold = (rate_num * sample_range) // rate_den
-    mask = 0
-
-    for group_idx in range(noise_bits):
-        group_rand = (random_value >> (group_idx * group_random_bits)) & ((1 << group_random_bits) - 1)
-        bit_idx = group_rand & ((1 << bit_index_bits) - 1)
-        sample = (group_rand >> bit_index_bits) & (sample_range - 1)
-        if sample < threshold:
-            mask |= 1 << (group_idx * group_bits + bit_idx)
-
-    return mask
-
-
-def lane_seed_128(seed: int, lane: int) -> int:
-    """Derive a nonzero 128-bit lane seed matching the RTL salt convention."""
-    mask128 = (1 << 128) - 1
-    salt = ((lane + 1) * 0x9E37_79B9_7F4A_7C15) & ((1 << 64) - 1)
-    mixed64 = (int(seed) ^ salt) & ((1 << 64) - 1)
-    state = ((mixed64 << 64) | mixed64) & mask128
-    assert state != 0
-    return state
-
-
-def generate_read_lane_masks(
-    lane_states: list[int],
-    *,
-    hash_bits: int,
-    noise_bits: int,
-    rate_num: int,
-    rate_den: int,
-    lane_valid: list[bool],
-) -> tuple[list[int], list[int]]:
-    """Advance valid lane PRNG states once and return one mask per lane."""
-    next_states: list[int] = []
-    masks: list[int] = []
-
-    for lane, state in enumerate(lane_states):
-        if lane_valid[lane]:
-            next_state = xorshift128(state)
-            mask = generate_grouped_flip_mask(
-                random_value=next_state,
-                hash_bits=hash_bits,
-                noise_bits=noise_bits,
-                rate_num=rate_num,
-                rate_den=rate_den,
-            )
-        else:
-            next_state = state
-            mask = 0
-        next_states.append(next_state)
-        masks.append(mask)
-
-    return masks, next_states
-
-
-def score_rows_with_read_noise(
-    query: int,
-    rows: Sequence[int],
-    *,
-    lane_states: Sequence[int],
-    width: int = 512,
-    lanes: int = 8,
-    noise_bits: int = 8,
-    rate_num: int = 1,
-    rate_den: int = 100,
-) -> tuple[np.ndarray, list[int]]:
-    """Score one query with read noise and return updated lane PRNG states.
-
-    Unlike match_top1_with_read_noise(), this helper is stateful across calls:
-    callers pass current lane states in and receive the next states back.
-    This matches a DUT that is reset once, then serves multiple queries.
-    """
-    assert lanes > 0
-    assert len(rows) % lanes == 0
-    assert len(lane_states) == lanes
-
-    scores = np.zeros(len(rows), dtype=np.int32)
-    next_lane_states = [int(state) for state in lane_states]
-
-    for base in range(0, len(rows), lanes):
-        lane_valid = [True] * lanes
-        masks, next_lane_states = generate_read_lane_masks(
-            next_lane_states,
-            hash_bits=width,
-            noise_bits=noise_bits,
-            rate_num=rate_num,
-            rate_den=rate_den,
-            lane_valid=lane_valid,
-        )
-
-        for lane in range(lanes):
-            row_idx = base + lane
-            noisy_row = int(rows[row_idx]) ^ int(masks[lane])
-            scores[row_idx] = xnor_popcount_score(int(query), noisy_row, width)
-
-    return scores, next_lane_states
-
-
-def match_top1_with_read_noise(
-    query: int,
-    rows: Sequence[int],
-    *,
-    width: int = 512,
-    lanes: int = 8,
-    noise_bits: int = 8,
-    rate_num: int = 1,
-    rate_den: int = 100,
-    seed: int = 0x6A09_E667_F3BC_C909,
-) -> MatchResult:
-    """Top-1 matching with dynamic read noise, one query in flight."""
-    assert lanes > 0
-    assert len(rows) % lanes == 0
-
-    scores = np.zeros(len(rows), dtype=np.int32)
-    best_index = 0
-    best_score = -1
-    lane_states = [lane_seed_128(seed, lane) for lane in range(lanes)]
-
-    for base in range(0, len(rows), lanes):
-        lane_valid = [True] * lanes
-        masks, lane_states = generate_read_lane_masks(
-            lane_states,
-            hash_bits=width,
-            noise_bits=noise_bits,
-            rate_num=rate_num,
-            rate_den=rate_den,
-            lane_valid=lane_valid,
-        )
-
-        for lane in range(lanes):
-            row_idx = base + lane
-            noisy_row = int(rows[row_idx]) ^ masks[lane]
-            score = xnor_popcount_score(int(query), noisy_row, width)
-            scores[row_idx] = score
-            if score > best_score:
-                best_score = score
-                best_index = row_idx
-
-    return MatchResult(top1_index=int(best_index), top1_score=int(best_score), scores=scores)
-
-
 def random_hashes(
    rng: np.random.Generator,
    n: int,
--- a/hw/sim/scripts/sweep_noise.py
+++ b/hw/sim/scripts/sweep_noise.py
@@ -16,7 +16,6 @@ if str(SIM_ROOT) not in sys.path:

 import numpy as np
 from model.ref_model import (
-    generate_write_flip_mask,
    match_top1,
    random_hashes,
 )
@@ -31,20 +30,12 @@ def apply_write_noise(
    noise_bits: int = 8,
    seed: int = 0,
 ) -> list[int]:
-    """Apply write-noise flip masks to every row, returning noisy copies.
+    """No-op: write-noise flip masks are now generated by Bernoulli RTL only.

-    *seed* is a 64-bit value (RTL NOISE_SEED). It is duplicated to form
-    the 128-bit xorshift initial state: {seed, seed}.
+    The sweep now measures top-1 stability of pure matching over queries,
+    since noise is applied at RTL write time, not in the Python model.
    """
-    noisy: list[int] = []
-    state = (seed << 64) | seed
-    mask_w = (1 << width) - 1
-    for row in rows:
-        flip, state = generate_write_flip_mask(
-            state, width, noise_bits, rate_num, rate_den
-        )
-        noisy.append((row ^ flip) & mask_w)
-    return noisy
+    return list(rows)


 def main() -> None:
--- a/hw/sim/tests/model/test_ref_model_noise.py
+++ b/hw/sim/tests/model/test_ref_model_noise.py
@@ -1,94 +1,31 @@
 # -*- coding: utf-8 -*-
 """
-参考模型（ref_model）的纯 Python 单元测试。
+参考模型（ref_model）的纯 Python 单元测试 — Phase 2 cleaned.

 本文件不涉及任何 RTL / Verilator 仿真，仅验证 Python 参考模型的正确性。
-所有 RTL-vs-模型 的对比测试（如顶层 test_cam_basic.py）都依赖此参考模型，
-因此这里是整个测试体系的「基石」——参考模型如果有 bug，所有对比测试都将失效。
+Phase 2 后只保留 pure matching 函数；所有 grouped/read-noise helpers 已删除。

 测试覆盖：
-  1. 分组翻转掩码 — 完全速率 (rate=1/1) 的正确位翻转模式
-  2. 分组翻转掩码 — 零速率 (rate=0/100) 不应产生任何翻转
-  3. 评分函数语义 — 确认是「匹配位数」而非「汉明距离」
-  4. 读取噪声模型 — 相同输入 + 相同种子 = 可复现结果
+  1. XNOR 评分语义 — 确认是「匹配位数」而非「汉明距离」
+  2. Top-1 matching — 纯匹配，正确选出最高分索引
+  3. Top-K matching — 返回排序后的行索引列表
+  4. Top-K 排序规则 — 分数降序、平局行号升序
+  5. Top-K k 超过行数时 clamp
 """
+
 from __future__ import annotations

 from model.ref_model import (
-    generate_grouped_flip_mask,
-    match_top1_with_read_noise,
+    match_top1,
+    match_topk,
+    match_topk_from_scores,
    xnor_popcount_score,
 )
+import numpy as np


 # ==============================================================================
-# 测试 1：完全速率下的分组翻转掩码生成
-# ==============================================================================
-
-
-def test_grouped_flip_mask_full_rate_one_bit_per_64_bit_group():
-    """
-    验证 generate_grouped_flip_mask 在 rate_num=1, rate_den=1 时的行为。
-
-    背景：
-      - CAM 的 write noise 模块将 512-bit 哈希按 64-bit 分组，每组最多翻转 1 位。
-      - random_value 的位域含义（每 group 14 bits）：
-        bits [5:0]   → sample（未使用）
-        bits [13:6]  → bit_idx（选择该组内翻转哪一位）
-
-    本测试：
-      - 构造一个 random_value，使每个 group 的 bit_idx = group+1
-      - 断言生成的 mask 恰好有 8 个位被置位（每 group 一个）
-      - 断言每个被翻转的位位置与预期一致
-    """
-    random_value = 0
-    for group in range(8):
-        bit_idx = group + 1
-        sample = 0
-        random_value |= bit_idx << (group * 14)
-        random_value |= sample << (group * 14 + 6)
-
-    mask = generate_grouped_flip_mask(
-        random_value=random_value,
-        hash_bits=512,      # 8 组 × 64 bits/组
-        noise_bits=8,       # 每组的 bit_idx 位宽
-        rate_num=1,         # 分子 = 1
-        rate_den=1,         # 分母 = 1 → 100% 概率，每组都翻转
-    )
-
-    # 预期：每组的 bit_idx 位被翻转
-    expected = 0
-    for group in range(8):
-        expected |= 1 << (group * 64 + group + 1)
-
-    assert mask == expected
-    assert mask.bit_count() == 8   # 恰好 8 位被翻转（每组一位）
-
-
-# ==============================================================================
-# 测试 2：零速率下不应产生任何翻转
-# ==============================================================================
-
-
-def test_grouped_flip_mask_zero_rate_no_flips():
-    """
-    验证 rate_num=0 时，无论 random_value 为何值，mask 都应为 0。
-
-    这是写入噪声的「零噪声」配置边界测试——
-    确保 RTL 参数 WRITE_NOISE_RATE_NUM=0 能真正关闭噪声注入。
-    """
-    mask = generate_grouped_flip_mask(
-        random_value=(1 << 128) - 1,  # 全 1 的 random_value
-        hash_bits=512,
-        noise_bits=8,
-        rate_num=0,                   # 分子 = 0 → 翻转概率为 0
-        rate_den=100,
-    )
-    assert mask == 0                  # mask 必须全 0，一个位都不翻
-
-
-# ==============================================================================
-# 测试 3：评分函数语义 — 确认是「XNOR 匹配位数」而非「汉明距离」
+# 测试 1：评分函数语义 — 确认是「XNOR 匹配位数」而非「汉明距离」
 # ==============================================================================


@@ -114,64 +51,27 @@ def test_score_is_bit_match_popcount_not_hamming_distance():


 # ==============================================================================
-# 测试 4：读取噪声模型的可复现性（确定性种子）
+# 测试 2：Top-1 matching
 # ==============================================================================


-def test_read_noise_model_is_reproducible_after_reset_seed():
-    """
-    验证 match_top1_with_read_noise 在相同参数下产生相同结果。
-
-    为什么这个测试至关重要：
-      - RTL 中的 read noise PRNG 使用固定种子 (0x6A09E667F3BCC909)
-      - 参考模型必须使用相同的种子来复现 RTL 的噪声行为
-      - 如果两次调用结果不同，说明模型存在非确定性 bug
-        （如未重置 PRNG 状态、或使用了非确定性随机源）
-
-    测试数据：
-      - 8 行不同模式的 512-bit 哈希（全0、全1、稀疏值）
-      - 噪声配置：rate=1%, lanes=8, noise_bits=8
-    """
-    rows = [0, (1 << 512) - 1, 0x1234, 0x5678, 0x9ABC, 0xDEF0, 0x1357, 0x2468]
-    query = rows[2]
-    kwargs = dict(
-        query=query,
-        rows=rows,
-        width=512,
-        lanes=8,
-        noise_bits=8,
-        rate_num=1,
-        rate_den=100,
-        seed=0x6A09_E667_F3BC_C909,
-    )
-
-    first = match_top1_with_read_noise(**kwargs)
-    second = match_top1_with_read_noise(**kwargs)
-
-    # 两次调用的 Top-1 结果和分数数组必须完全一致
-    assert first.top1_index == second.top1_index
-    assert first.top1_score == second.top1_score
-    assert first.scores.tolist() == second.scores.tolist()
+def test_match_top1_selects_highest_xnor_score_with_row_index_tiebreak():
+    """Top-1 应选出 XNOR 分最高的行；平局时选最小行号。"""
+    rows = [0b0000, 0b1111, 0b0011, 0b0101]
+    query = 0b0000
+    result = match_top1(query, rows, width=4)
+    assert result.top1_index == 0
+    assert result.top1_score == 4
+    assert result.scores.tolist() == [4, 0, 2, 2]


 # ==============================================================================
-# 测试 5：Top-K 排序 — 分数降序、平局行号升序
+# 测试 3：Top-K matching
 # ==============================================================================


-def test_match_topk_from_scores_uses_score_desc_then_row_asc():
-    """Top-K 排序规则：分数越大越优先；分数相同时行号越小越优先。"""
-    from model.ref_model import match_topk_from_scores
-    import numpy as np
-
-    scores = np.array([7, 9, 9, 2, 7], dtype=np.int32)
-    assert match_topk_from_scores(scores, 4) == [1, 2, 0, 4]
-
-
 def test_match_topk_scores_rows_by_xnor_popcount():
    """match_topk 通过 xnor_popcount 计算分数，返回排序后的行索引和分数数组。"""
-    from model.ref_model import match_topk
-
    rows = [0b0000, 0b1111, 0b0011, 0b0101]
    query = 0b0000
    indices, scores = match_topk(query, rows, width=4, k=3)
@@ -179,43 +79,24 @@ def test_match_topk_scores_rows_by_xnor_popcount():
    assert indices == [0, 2, 3]


+# ==============================================================================
+# 测试 4：Top-K 排序 — 分数降序、平局行号升序
+# ==============================================================================
+
+
+def test_match_topk_from_scores_uses_score_desc_then_row_asc():
+    """Top-K 排序规则：分数越大越优先；分数相同时行号越小越优先。"""
+    scores = np.array([7, 9, 9, 2, 7], dtype=np.int32)
+    assert match_topk_from_scores(scores, 4) == [1, 2, 0, 4]
+
+
+# ==============================================================================
+# 测试 5：Top-K k 超过行数时 clamp
+# ==============================================================================
+
+
 def test_match_topk_clamps_k_to_row_count():
    """当 k 超过实际行数时，返回所有行（按排序）。"""
-    from model.ref_model import match_topk
-
    indices, scores = match_topk(0, [0, 1], width=1, k=5)
    assert scores.tolist() == [1, 0]
    assert indices == [0, 1]
-
-
-# ==============================================================================
-# 测试 6：读取噪声 stateful 评分助手的跨查询状态推进
-# ==============================================================================
-
-
-def test_score_rows_with_read_noise_stateful_across_queries():
-    """score_rows_with_read_noise 在多次调用间正确推进 lane PRNG 状态。
-
-    两次调用使用相同的 rows/query 和零噪声率：
-      - 分数应一致（无噪声翻转）
-      - 但 lane states 应该变化（PRNG 已推进）
-    """
-    from model.ref_model import score_rows_with_read_noise
-
-    rows = [0, 0, 0, 0]
-    query = 0
-    lane_states = [1, 2]
-
-    scores_1, next_states_1 = score_rows_with_read_noise(
-        query, rows, lane_states=lane_states, width=128, lanes=2,
-        noise_bits=2, rate_num=0, rate_den=100,
-    )
-    scores_2, next_states_2 = score_rows_with_read_noise(
-        query, rows, lane_states=next_states_1, width=128, lanes=2,
-        noise_bits=2, rate_num=0, rate_den=100,
-    )
-
-    assert scores_1.tolist() == [128, 128, 128, 128]
-    assert scores_2.tolist() == [128, 128, 128, 128]
-    assert next_states_1 != lane_states
-    assert next_states_2 != next_states_1
--- a/hw/sim/tests/modules/cam_read_noise/Makefile
+++ b/hw/sim/tests/modules/cam_read_noise/Makefile
@@ -7,9 +7,5 @@ COCOTB_TEST_MODULES := tests.modules.cam_read_noise.test_cam_read_noise
 VERILOG_SOURCES := $(RTL_READ_NOISE)

 HASH_BITS ?= 512
-READ_NOISE_EN ?= 0
-READ_NOISE_RATE_NUM ?= 0
-READ_NOISE_RATE_DEN ?= 100
-READ_NOISE_BITS ?= $(shell echo $$(( $(HASH_BITS) / 64 )))

 include $(SIM_ROOT)/mk/cocotb-common.mk
--- a/hw/sim/tests/modules/cam_read_noise/test_cam_read_noise.py
+++ b/hw/sim/tests/modules/cam_read_noise/test_cam_read_noise.py
@@ -2,7 +2,7 @@ from __future__ import annotations

 import cocotb
 from cocotb.clock import Clock
-from cocotb.triggers import RisingEdge
+from cocotb.triggers import RisingEdge, Timer


 async def reset_read_noise(dut):
@@ -38,11 +38,45 @@ async def read_noise_disabled_forwards_hashes_after_one_stage(dut):
    dut.row_ids_i.value = rows
    dut.lane_valid_i.value = all_lanes_valid
    dut.valid_i.value = 1
-    await RisingEdge(dut.clk)
+    await Timer(1, unit="step")
+    await RisingEdge(dut.clk)       # valid_o ← valid_i=1 internally
+    await Timer(1, unit="step")
    dut.valid_i.value = 0
-    await RisingEdge(dut.clk)
-    await RisingEdge(dut.clk)

+    # One-stage pass-through: valid_o holds the latched value for this cycle
+    assert int(dut.valid_o.value) == 1
+    assert int(dut.hashes_noisy_o.value) == hashes
+    assert int(dut.row_ids_o.value) == rows
+    assert int(dut.lane_valid_o.value) == all_lanes_valid
+
+
+@cocotb.test()
+async def read_noise_enabled_still_forwards_hashes_unmodified(dut):
+    """With READ_NOISE_EN=1, the pass-through still forwards hashes unmodified."""
+    cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
+    await reset_read_noise(dut)
+
+    LANES = len(dut.lane_valid_i)
+    ROW_BITS = len(dut.row_ids_i) // LANES
+    HASH_BITS_PER_LANE = len(dut.hashes_i) // LANES
+    all_lanes_valid = (1 << LANES) - 1
+
+    hashes = 0
+    rows = 0
+    for lane in range(LANES):
+        hashes |= (lane + 0x55) << (lane * HASH_BITS_PER_LANE)
+        rows |= lane << (lane * ROW_BITS)
+
+    dut.hashes_i.value = hashes
+    dut.row_ids_i.value = rows
+    dut.lane_valid_i.value = all_lanes_valid
+    dut.valid_i.value = 1
+    await Timer(1, unit="step")
+    await RisingEdge(dut.clk)       # valid_o ← valid_i=1 internally
+    await Timer(1, unit="step")
+    dut.valid_i.value = 0
+
+    # One-stage pass-through: valid_o holds latched value from previous cycle
    assert int(dut.valid_o.value) == 1
    assert int(dut.hashes_noisy_o.value) == hashes
    assert int(dut.row_ids_o.value) == rows
--- a/hw/sim/tests/modules/cam_write_noise/Makefile
+++ b/hw/sim/tests/modules/cam_write_noise/Makefile
@@ -10,6 +10,4 @@ HASH_BITS ?= 512
 WRITE_NOISE_EN ?= 1
 WRITE_NOISE_RATE_NUM ?= 1
 WRITE_NOISE_RATE_DEN ?= 100
-WRITE_NOISE_BITS ?= $(shell echo $$(( $(HASH_BITS) / 64 )))
-
 include $(SIM_ROOT)/mk/cocotb-common.mk
--- a/hw/sim/tests/modules/cam_write_noise/test_cam_write_noise.py
+++ b/hw/sim/tests/modules/cam_write_noise/test_cam_write_noise.py
@@ -2,8 +2,32 @@ from __future__ import annotations

 import cocotb
 from cocotb.clock import Clock
-from cocotb.triggers import RisingEdge
-from model.ref_model import generate_write_flip_mask
+from cocotb.triggers import RisingEdge, Timer
+
+# Bernoulli: 1 PRIME + 16 RUN = 17 cycles internal
+# + 1 cycle for mask_start propagation + 1 cycle for core_wr_valid output = 19
+DEFAULT_WRITE_NOISE_LATENCY = 19
+
+
+async def pulse_write(dut, row: int, value: int):
+    dut.wr_row.value = row
+    dut.wr_hash.value = value
+    dut.wr_valid.value = 1
+    await Timer(1, unit="step")
+    assert int(dut.wr_ready.value) == 1
+    await RisingEdge(dut.clk)
+    await Timer(1, unit="step")
+    dut.wr_valid.value = 0
+
+
+async def wait_core_write(dut, max_cycles: int = 128) -> int:
+    cycles = 0
+    while int(dut.core_wr_valid.value) == 0:
+        assert cycles < max_cycles, "timed out waiting for core_wr_valid"
+        await RisingEdge(dut.clk)
+        await Timer(1, unit="step")
+        cycles += 1
+    return cycles


 async def reset_write_noise(dut):
@@ -19,23 +43,52 @@ async def reset_write_noise(dut):


@cocotb.test()
-async def write_noise_outputs_grouped_noisy_hash(dut):
+async def write_noise_enabled_applies_bernoulli_mask_after_generation(dut):
+    """Noise active: FSM enters WAIT_MASK, core_wr_hash deterministic across reset."""
    cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
    await reset_write_noise(dut)

-    value = 0x123456789ABCDEF
-    dut.wr_row.value = 3
-    dut.wr_hash.value = value
-    dut.wr_valid.value = 1
+    value = (1 << 512) - 1  # all-ones: even low-rate Bernoulli may flip some bits
+    await pulse_write(dut, row=3, value=value)
+    await Timer(1, unit="step")
+    assert int(dut.wr_ready.value) == 0
+
+    cycles = await wait_core_write(dut)
+    assert cycles == DEFAULT_WRITE_NOISE_LATENCY
+
+    assert int(dut.core_wr_row.value) == 3
+    hash_after_first = int(dut.core_wr_hash.value)
+
    await RisingEdge(dut.clk)
+    await Timer(1, unit="step")
+    assert int(dut.core_wr_valid.value) == 0
+    assert int(dut.wr_ready.value) == 1
+
+    # Deterministic across reset: same seed → same mask → same noisy hash
+    await reset_write_noise(dut)
+    await pulse_write(dut, row=3, value=value)
+    await wait_core_write(dut)
+    assert int(dut.core_wr_hash.value) == hash_after_first
+
+
+@cocotb.test()
+async def write_noise_backpressures_second_write_until_done(dut):
+    cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
+    await reset_write_noise(dut)
+
+    await pulse_write(dut, row=1, value=0xAA55)
+
+    dut.wr_row.value = 2
+    dut.wr_hash.value = 0x55AA
+    dut.wr_valid.value = 1
+    await Timer(1, unit="step")
+    for _ in range(4):
+        assert int(dut.wr_ready.value) == 0
+        assert int(dut.core_wr_valid.value) == 0
+        await RisingEdge(dut.clk)
+        await Timer(1, unit="step")
    dut.wr_valid.value = 0

-    while int(dut.core_wr_valid.value) == 0:
-        await RisingEdge(dut.clk)
-
-    seed = 0xB504_F32D_B504_F32D
-    hash_bits = len(dut.wr_hash)
-    noise_bits = hash_bits // 64
-    flip, _ = generate_write_flip_mask((seed << 64) | seed, hash_bits, noise_bits, 1, 100)
-    assert int(dut.core_wr_row.value) == 3
-    assert int(dut.core_wr_hash.value) == (value ^ flip)
+    cycles = await wait_core_write(dut)
+    assert cycles == DEFAULT_WRITE_NOISE_LATENCY - 4  # 19 - 4 = 15
+    assert int(dut.core_wr_row.value) == 1
--- a/hw/sim/tests/perf/Makefile
+++ b/hw/sim/tests/perf/Makefile
@@ -8,8 +8,5 @@ VERILOG_SOURCES := $(RTL_CAM_TOP)

 HASH_BITS ?= 512
 WRITE_NOISE_EN := 0
-READ_NOISE_EN := 0
-WRITE_NOISE_BITS := $(shell echo $$(( $(HASH_BITS) / 64 )))
-READ_NOISE_BITS := $(shell echo $$(( $(HASH_BITS) / 64 )))

 include $(SIM_ROOT)/mk/cocotb-common.mk
--- a/hw/sim/tests/perf/test_cam_perf.py
+++ b/hw/sim/tests/perf/test_cam_perf.py
@@ -243,7 +243,6 @@ async def cam_perf_benchmark(dut):
    hash_bits = dut_hash_bits(dut)
    lanes = dut_lanes(dut)
    write_noise_en = _get_param(dut, "WRITE_NOISE_EN", 1)
-    read_noise_en = _get_param(dut, "READ_NOISE_EN", 0)

    # ── Deterministic query ─────────────────────────────────────────────
    query_hash = 0xAA55_AA55_AA55_AA55_AA55_AA55_AA55_AA55
@@ -271,7 +270,7 @@ async def cam_perf_benchmark(dut):
    )

    # ── Correctness assertions (conditional on noise state) ─────────────
-    if not write_noise_en and not read_noise_en:
+    if not write_noise_en:
        # Without noise: stored hash at row 0 == query_hash → exact match.
        assert top1_index == 0, (
            f"Noise disabled: expected top1_index=0 (exact match), got "
@@ -282,7 +281,7 @@ async def cam_perf_benchmark(dut):
            f"{top1_score}"
        )
    else:
-        # With noise: write/read flip masks may corrupt stored values, so
+        # With noise: write flip masks may corrupt stored values, so
        # we cannot reliably assert the exact match.  Instead, confirm a
        # valid non-zero score was produced (the match engine ran).
        assert top1_score > 0, (
@@ -290,10 +289,10 @@ async def cam_perf_benchmark(dut):
            "Match engine returned invalid result."
        )
        dut._log.info(
-            "Noise enabled (WRITE_NOISE_EN=%s, READ_NOISE_EN=%s) — "
+            "Noise enabled (WRITE_NOISE_EN=%s) — "
            "skipping exact top1_index/top1_score assertion. "
            "top1_index=%d top1_score=%d",
-            write_noise_en, read_noise_en, top1_index, top1_score,
+            write_noise_en, top1_index, top1_score,
        )

    # ── Machine-readable performance marker ─────────────────────────────
--- a/hw/sim/tests/top/no_noise/Makefile
+++ b/hw/sim/tests/top/no_noise/Makefile
@@ -8,6 +8,5 @@ VERILOG_SOURCES := $(RTL_CAM_TOP)

 # 禁用所有噪声模块
 WRITE_NOISE_EN := 0
-READ_NOISE_EN := 0

 include $(SIM_ROOT)/mk/cocotb-common.mk
--- a/hw/sim/tests/top/no_noise/test_no_noise.py
+++ b/hw/sim/tests/top/no_noise/test_no_noise.py
@@ -1,6 +1,6 @@
 # -*- coding: utf-8 -*-
 """
-CAM 顶层（cam_top）no_noise 配置集成测试（WRITE_NOISE_EN=0, READ_NOISE_EN=0）。
+CAM 顶层（cam_top）no_noise 配置集成测试（WRITE_NOISE_EN=0）。

 所有噪声模块禁用，验证 CAM 在无噪声下的标准行为。

@@ -17,8 +17,8 @@ CAM 顶层（cam_top）no_noise 配置集成测试（WRITE_NOISE_EN=0, READ_NOIS

 === 配置背景 ===

-  本目录固定使用 WRITE_NOISE_EN=0 和 READ_NOISE_EN=0 编译，
-  因此所有测试无需运行时参数门控——Makefile 保证配置正确。
+   本目录固定使用 WRITE_NOISE_EN=0 编译，
+   因此所有测试无需运行时参数门控——Makefile 保证配置正确。
 """

 from __future__ import annotations
@@ -62,7 +62,7 @@ async def compile_includes_grouped_noise_helper(dut):


 # ═══════════════════════════════════════════════════════════════════════════════
-# 测试 A：基线（WRITE_NOISE_EN=0, READ_NOISE_EN=0）
+# 测试 A：基线（WRITE_NOISE_EN=0）
 # ── 验证写+查在噪声关闭时与旧 CAM 行为完全一致
 # ═══════════════════════════════════════════════════════════════════════════════

--- a/hw/sim/tests/top/read_noise/Makefile
+++ b/hw/sim/tests/top/read_noise/Makefile
@@ -6,10 +6,7 @@ TOPLEVEL := cam_top
 COCOTB_TEST_MODULES := tests.top.read_noise.test_read_noise
 VERILOG_SOURCES := $(RTL_CAM_TOP)

-# 读取噪声开启，写入噪声默认关闭
-READ_NOISE_EN := 1
-READ_NOISE_RATE_NUM := 1
-READ_NOISE_RATE_DEN := 100
+# 读取噪声开启（Phase 2 后为 pass-through），写入噪声默认关闭
 WRITE_NOISE_EN := 0

 include $(SIM_ROOT)/mk/cocotb-common.mk
--- a/hw/sim/tests/top/read_noise/test_read_noise.py
+++ b/hw/sim/tests/top/read_noise/test_read_noise.py
@@ -1,27 +1,9 @@
 # -*- coding: utf-8 -*-
 """
-CAM 读取噪声（read_noise）集成测试。
+CAM 读取路径 pass-through 集成测试 — Phase 2 cleaned.

-本文件针对 READ_NOISE_EN=1 的编译配置，验证 RTL 的读取噪声行为
-与 Python 参考模型（ref_model）一致。
-
-=== 测试内容 ===
-
-  read_noise_model_match — 读取噪声模型匹配：
-    写入原始哈希，预测含写入噪声（如果 WRITE_NOISE_EN=1）的存储值，
-    再用 match_top1_with_read_noise 计算含读取噪声的期望结果，
-    与 RTL 实际 Top-1 进行比对。
-
-=== 架构背景 ===
-
-  CAM 硬件由以下流水线组成：
-    Write Noise → Banked Core Storage → Read Noise → Match Engine Pipeline
-                                                     ↓
-                                               Top-K Tracker → Result Serializer
-
-  本测试覆盖的是 Read Noise → Match Engine 段。
-  写入噪声（WRITE_NOISE_EN）通过 Makefile 的 test-with-write-noise 子目标
-  启用，测试代码内部已兼容两种配置。
+Read noise 已退休；cam_read_noise 是纯 pass-through。
+本测试验证查询返回的 scores 与 pure matching 一致。
 """

 from __future__ import annotations
@@ -31,83 +13,40 @@ import numpy as np
 from cocotb.clock import Clock
 from cocotb.triggers import RisingEdge
 from model.ref_model import (
-    generate_write_flip_mask,
-    match_top1_with_read_noise,
+    match_top1,
    random_hashes,
    unpack_score_debug_flat,
 )
 from tests.top.utils import (
+    collect_topk,
    dut_hash_bits,
    dut_lanes,
    dut_num_rows,
    get_param,
    query_once,
+    query_topk_once,
    reset_dut,
+    write_row,
    write_rows,
 )


-# ═══════════════════════════════════════════════════════════════════════════════
-# 测试：读取噪声模型匹配
-# ── READ_NOISE_EN=1 由 Makefile 保证，测试代码中不再重复门控
-# ═══════════════════════════════════════════════════════════════════════════════
-
-
@cocotb.test()
-async def read_noise_model_match(dut):
-    """读取噪声模型匹配：验证 RTL 的读取噪声行为与 Python 参考模型一致。
-
-    与写入噪声不同，读取噪声发生在查询阶段（每次查询向哈希值注入噪声），
-    因此：
-      - 如果先有写入噪声，存储行已经被翻转过一次
-      - 然后查询时还会再注入一层读取噪声
-      - 两层噪声使用不同的种子（写: 0xB504..., 读: 0x6A09...）
-
-    本测试：
-      1. 用 Python 模型预计算存储后的哈希（含写入噪声）
-      2. 用 match_top1_with_read_noise 预计算含读取噪声的期望结果
-      3. 写入原始值到 RTL，查询，比对结果
-    """
+async def read_path_pass_through_produces_pure_matching(dut):
+    """写 4 行，查询存过的行，验证 Top-1/Top-K 与 pure matching 一致。"""
    cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
    await reset_dut(dut)

    num_rows = dut_num_rows(dut)
    hash_bits = dut_hash_bits(dut)
-    lanes = dut_lanes(dut)
-    rng = np.random.default_rng(123)
+    rng = np.random.default_rng(42)
    rows = random_hashes(rng, num_rows, width=hash_bits)

-    # If write noise is enabled, apply write flip masks to predict stored rows
-    stored_rows = list(rows)
-    if get_param(dut, "WRITE_NOISE_EN", 0):
-        seed = 0xB504_F32D_B504_F32D
-        prng_state = (seed << 64) | seed
-        stored_rows = []
-        for row in rows:
-            flip, prng_state = generate_write_flip_mask(
-                prng_state,
-                hash_bits,
-                get_param(dut, "WRITE_NOISE_BITS", 8),
-                get_param(dut, "WRITE_NOISE_RATE_NUM", 1),
-                get_param(dut, "WRITE_NOISE_RATE_DEN", 100),
-            )
-            stored_rows.append(row ^ flip)
-
-    query = rows[min(5, num_rows - 1)]
-
    await write_rows(dut, rows)
-    top1_index, top1_score, score_debug = await query_once(dut, query)
+    query = rows[min(50, num_rows - 1)]

-    expected = match_top1_with_read_noise(
-        query,
-        stored_rows,
-        width=hash_bits,
-        lanes=lanes,
-        noise_bits=get_param(dut, "READ_NOISE_BITS", 8),
-        rate_num=get_param(dut, "READ_NOISE_RATE_NUM", 1),
-        rate_den=get_param(dut, "READ_NOISE_RATE_DEN", 100),
-        seed=0x6A09_E667_F3BC_C909,
-    )
+    top1_index, top1_score, score_debug = await query_once(dut, query)
+    expected = match_top1(query, rows, width=hash_bits)

    assert top1_index == expected.top1_index
    assert top1_score == expected.top1_score
--- a/hw/sim/tests/top/write_noise/Makefile
+++ b/hw/sim/tests/top/write_noise/Makefile
@@ -10,7 +10,6 @@ VERILOG_SOURCES := $(RTL_CAM_TOP)
 WRITE_NOISE_EN := 1
 WRITE_NOISE_RATE_NUM := 1
 WRITE_NOISE_RATE_DEN := 100
-READ_NOISE_EN := 0

 include $(SIM_ROOT)/mk/cocotb-common.mk

--- a/hw/sim/tests/top/write_noise/test_write_noise.py
+++ b/hw/sim/tests/top/write_noise/test_write_noise.py
@@ -2,7 +2,7 @@
 """
 CAM 写入噪声（Write Noise）集成测试 —— 专用配置。

-本文件测试 WRITE_NOISE_EN=1, READ_NOISE_EN=0 配置下，
+本文件测试 WRITE_NOISE_EN=1 配置下，
 写入噪声模块的正确性。默认噪声率约 1%（NUM=1, DEN=100）。

 === 测试列表 ===
@@ -14,7 +14,7 @@ CAM 写入噪声（Write Noise）集成测试 —— 专用配置。

 === 架构背景 ===

-  写入噪声流水线位置：Write Noise → Banked Core Storage → Read Noise → Match Engine
+  写入噪声流水线位置：Write Noise → Banked Core Storage → Match Engine
  本测试覆盖完整的 cam_top 链路，写入噪声为唯一活跃噪声源。

 === Makefile 子目标 ===
@@ -30,7 +30,6 @@ import numpy as np
 from cocotb.clock import Clock
 from cocotb.triggers import RisingEdge
 from model.ref_model import (
-    generate_write_flip_mask,
    match_top1,
    random_hashes,
 )
@@ -89,71 +88,7 @@ async def default_noise_reproducible(dut):


 # ═══════════════════════════════════════════════════════════════════════════════
-# 测试 2：精确 RTL-vs-模型 PRNG 掩码匹配
-# ── RTL 存储的哈希与 ref_model.py 生成的掩码逐位一致
-# ═══════════════════════════════════════════════════════════════════════════════
-
-
-@cocotb.test()
-async def exact_noise_model_match(dut):
-    """精确噪声模型匹配：RTL 的 PRNG 输出必须与 Python 参考模型逐位一致。
-
-    测试方法：
-      1. 用固定 RTL seed 和已知噪声参数，在 Python 中预计算每行的 flip 掩码
-      2. 预期存储值 = 原始值 XOR flip_mask
-      3. 写入原始值到 RTL，查询预期存储值
-      4. 断言每行 score = HASH_BITS（完全匹配）
-
-    这验证了 RTL 的 LFSR 实现与 Python 模型的 PRNG 使用相同的
-    多项式、相同的位宽、相同的种子初始化序列。
-    """
-    if not hasattr(dut, "score_debug_flat"):
-        dut._log.info("Skipping exact_noise_model_match: requires SIM_DEBUG.")
-        return
-
-    rtol = None
-    atol = None
-    cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
-    await reset_dut(dut)
-
-    hash_bits = dut_hash_bits(dut)
-    noise_bits = get_param(dut, "WRITE_NOISE_BITS", 8)
-    rate_num = get_param(dut, "WRITE_NOISE_RATE_NUM", 1)
-    rate_den = get_param(dut, "WRITE_NOISE_RATE_DEN", 100)
-
-    n_test_rows = 4
-    rng = np.random.default_rng(99)
-    rows = random_hashes(rng, n_test_rows, width=hash_bits)
-
-    RTL_SEED = 0xB504_F32D_B504_F32D
-    prng_state = (RTL_SEED << 64) | RTL_SEED
-    expected_stored = []
-    for row in rows:
-        flip, prng_state = generate_write_flip_mask(
-            prng_state,
-            hash_bits,
-            noise_bits,
-            rate_num,
-            rate_den,
-        )
-        expected_stored.append(row ^ flip)
-
-    for idx, val in enumerate(rows):
-        await write_row(dut, idx, val)
-
-    for idx, expected in enumerate(expected_stored):
-        top1_index, top1_score, score_debug = await query_once(dut, expected)
-        assert score_debug is not None, (
-            "score_debug required for mask match verification"
-        )
-        assert int(score_debug[idx]) == hash_bits, (
-            f"Row {idx}: expected stored hash to match model prediction, "
-            f"score={score_debug[idx]} != {hash_bits}"
-        )
-
-
-# ═══════════════════════════════════════════════════════════════════════════════
-# 测试 3：零噪声率（WRITE_NOISE_EN=1, RATE_NUM=0）
+# 测试 2：零噪声率（WRITE_NOISE_EN=1, RATE_NUM=0）
 # ── 噪声模块已连接但翻转概率为 0 → 行为应与无噪声一致
 # ═══════════════════════════════════════════════════════════════════════════════

@@ -195,80 +130,4 @@ async def zero_rate_noise(dut):
        assert np.array_equal(score_debug, expected.scores)


-# ═══════════════════════════════════════════════════════════════════════════════
-# 测试 4：100% 噪声率（RATE_NUM=1, RATE_DEN=1）
-# ── 每组都翻转 → 精确验证 PRNG 掩码生成
-# ═══════════════════════════════════════════════════════════════════════════════

-
-@cocotb.test()
-async def full_rate_noise(dut):
-    """完全速率噪声：每组 100% 翻转概率。
-
-    使用固定 RTL seed (0xB504F32DB504F32D)，用 Python 模型预计算
-    写入全 0 和全 1 行后应存储的哈希值，然后验证 RTL 实际存储的哈希
-    与模型预测完全一致。
-
-    这是最低容忍度的噪声测试——要求 score_debug_flat（SIM_DEBUG）
-    且每行的分数必须精确等于 HASH_BITS。
-    """
-    rate_num = get_param(dut, "WRITE_NOISE_RATE_NUM", 1)
-    rate_den = get_param(dut, "WRITE_NOISE_RATE_DEN", 100)
-    if rate_num != 1 or rate_den != 1:
-        dut._log.info(
-            "Skipping full_rate_noise: requires WRITE_NOISE_RATE_NUM=1, RATE_DEN=1."
-        )
-        return
-    if not hasattr(dut, "score_debug_flat"):
-        dut._log.info(
-            "Skipping full_rate_noise: requires SIM_DEBUG (score_debug_flat)."
-        )
-        return
-
-    cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
-    await reset_dut(dut)
-
-    hash_bits = dut_hash_bits(dut)
-    num_rows = dut_num_rows(dut)
-    noise_bits = get_param(dut, "WRITE_NOISE_BITS", 8)
-    all_zero = 0
-    all_one = (1 << hash_bits) - 1
-
-    RTL_SEED = 0xB504_F32D_B504_F32D
-    prng_state = (RTL_SEED << 64) | RTL_SEED
-
-    flip0, prng_state = generate_write_flip_mask(
-        prng_state,
-        hash_bits,
-        noise_bits,
-        rate_num,
-        rate_den,
-    )
-    expected_row0 = all_zero ^ flip0
-
-    flip1, prng_state = generate_write_flip_mask(
-        prng_state,
-        hash_bits,
-        noise_bits,
-        rate_num,
-        rate_den,
-    )
-    expected_row1 = all_one ^ flip1
-
-    rows = [0] * num_rows
-    rows[0] = all_zero
-    rows[1] = all_one
-
-    await write_rows(dut, rows)
-
-    top1_index, top1_score, score_debug = await query_once(dut, expected_row0)
-    assert score_debug is not None, "score_debug required for full_rate_noise"
-    assert int(score_debug[0]) == hash_bits, (
-        f"Row 0: expected exact match, score={score_debug[0]} != {hash_bits}"
-    )
-
-    top1_index, top1_score, score_debug = await query_once(dut, expected_row1)
-    assert score_debug is not None
-    assert int(score_debug[1]) == hash_bits, (
-        f"Row 1: expected exact match, score={score_debug[1]} != {hash_bits}"
-    )
--- a/hw/syn/synth_cam_top_flat.ys
+++ b/hw/syn/synth_cam_top_flat.ys
@@ -10,7 +10,7 @@ verilog_defaults -add -I../rtl/random

 # Read RTL sources in canonical order
 read_verilog -sv -D SYNTHESIS ../rtl/random/random128.sv
-read_verilog -sv -D SYNTHESIS ../rtl/noise/noise_mask_grouped.sv
+read_verilog -sv -D SYNTHESIS ../rtl/noise/noise_mask_bernoulli.sv
 read_verilog -sv -D SYNTHESIS ../rtl/noise/cam_write_noise.sv
 read_verilog -sv -D SYNTHESIS ../rtl/noise/cam_read_noise.sv
 read_verilog -sv -D SYNTHESIS ../rtl/core/cam_core_banked.sv
--- a/hw/syn/synth_cam_top_hier.ys
+++ b/hw/syn/synth_cam_top_hier.ys
@@ -10,7 +10,7 @@ verilog_defaults -add -I../rtl/random

 # Read RTL sources in canonical order
 read_verilog -sv -D SYNTHESIS ../rtl/random/random128.sv
-read_verilog -sv -D SYNTHESIS ../rtl/noise/noise_mask_grouped.sv
+read_verilog -sv -D SYNTHESIS ../rtl/noise/noise_mask_bernoulli.sv
 read_verilog -sv -D SYNTHESIS ../rtl/noise/cam_write_noise.sv
 read_verilog -sv -D SYNTHESIS ../rtl/noise/cam_read_noise.sv
 read_verilog -sv -D SYNTHESIS ../rtl/core/cam_core_banked.sv