模塊介紹

卷積神經(jīng)網(wǎng)絡(luò)（CNN)可以分為卷積層、池化層、激活層、全鏈接層結(jié)構(gòu)，本篇要實(shí)現(xiàn)的，就是CNN的卷積層中的window窗。

在卷積過程中，最復(fù)雜的就是卷積運(yùn)算，也就是Filter和圖片（輸入）相乘然后在相加的這一步驟。

我此處的構(gòu)想就是將其卷積這個(gè)步驟進(jìn)行拆分：加窗、載入權(quán)重、卷積運(yùn)算。因而對(duì)應(yīng)3個(gè)模塊，而此處實(shí)現(xiàn)的就是加窗這個(gè)模塊。而他主要負(fù)責(zé)的功能就是：提取輸入圖片中的數(shù)據(jù)，生成對(duì)應(yīng)的窗口。 如上圖所示，對(duì)x[:,:,0]圖片進(jìn)行窗口提起，提取的第一個(gè)窗口(左上角第一個(gè))就是

[ 0 0 0 0 0 1 0 0 1 ] egin{bmatrix}0&0&0\0&0&1\0&0&1end{bmatrix}000000011

代碼

可配置參數(shù)、輸入和輸出定義

STRIDE為窗口滑動(dòng)的步長，KERNEL_SIZE對(duì)應(yīng)輸入卷積核的大小，PADDING 為補(bǔ)充的長度

pixel_in 為輸出的圖片數(shù)據(jù)，frame_start 為圖片開始輸入的標(biāo)志，pixel_valid為輸入有效標(biāo)志

window_out是圖片展成一維的窗口數(shù)據(jù)

module window#(  parameterDATA_WIDTH = 16,       // Widthofeachpixeldata  parameterIMG_WIDTH = 32,       // Widthofinputimage  parameterIMG_HEIGHT = 32,      // Heightofinputimage  parameterKERNEL_SIZE = 3,      // Sizeofconvolutionwindow (square)  parameterSTRIDE = 1,         // Strideofconvolution  parameterPADDING = (KERNEL_SIZE - 1) / 2 // PaddingsizecalculatedforSAMEmode)(   input wire clk,            // Clock signal   input wire rst_n,           // Active low reset   input wire [DATA_WIDTH-1:0] pixel_in, // Input pixel data   input wire pixel_valid,        // Input pixel valid signal   input wire frame_start,        // Start of new frame signal     output reg [KERNEL_SIZE*KERNEL_SIZE*DATA_WIDTH-1:0] window_out, // Flattened window output   output reg window_valid       // Window data valid );

內(nèi)部信號(hào)定義

輸入的圖片數(shù)據(jù)是一個(gè)一個(gè)輸入的，用x_pos和y_pos 來記錄當(dāng)前pixel位于圖片中的位置

窗口在圖片上滑動(dòng)，用x_window，y_window用來判斷窗口目前的位置

line_Buffer緩存輸入的數(shù)據(jù)，同時(shí)進(jìn)行padding操作, 形成數(shù)據(jù)窗口，而window_buffer 在line_buffer上進(jìn)行滑動(dòng)，形成窗口

狀態(tài)機(jī)，分為三個(gè)狀態(tài) IDLE, LOAD,PROCESS, 分別對(duì)應(yīng)空閑，載入（開始載入數(shù)據(jù)），處理（形成window）

// Internal signals

reg [5:0] x_pos, y_pos; // Current input pixel position

reg [5:0] x_window, y_window; // Window center position

reg [DATA_WIDTH-1:0] line_buffer [0:KERNEL_SIZE][0:IMG_WIDTH+2*PADDING-1]; // Line buffer

reg [DATA_WIDTH-1:0] window_buffer [0:KERNEL_SIZE-1][0:KERNEL_SIZE-1]; // Window buffer

reg signed [6:0] src_y, src_x; // Temporary variables for coordinate calculation

// State machine

reg [1:0] current_state, next_state;

localparam IDLE = 2'b00, LOAD = 2'b01, PROCESS = 2'b10;

// Loop variables

integer i, j, k;

狀態(tài)的賦值以及跳轉(zhuǎn)

當(dāng)接收到frame_start信號(hào)（圖片開始輸入），狀態(tài)從空閑進(jìn)入到LOAD狀態(tài)；

當(dāng)目前的圖片數(shù)據(jù)可以已經(jīng)足夠，可以用來生成穩(wěn)定的輸出窗口時(shí)，進(jìn)入到PROCESS狀態(tài)

當(dāng)目前滑窗口提取完對(duì)應(yīng)數(shù)據(jù)窗口后，回到IDLE狀態(tài)

注：y_pos從0到KERNEL_SIZE-1時(shí)，已經(jīng)有了KERNEL_SIZE行數(shù)據(jù)了，可以進(jìn)入窗口數(shù)據(jù)提取階段；實(shí)際上可以更早進(jìn)入，因?yàn)榇嬖赑adding。當(dāng)y_pos=KERNEL_SIZE-Padding-1的時(shí)候，就可以進(jìn)入了

// FSM state transitions

always @(posedge clk or negedge rst_n) begin

if (!rst_n)

current_state <= IDLE;

else

current_state <= next_state;

end

always @(*) begin

case (current_state)

IDLE: next_state = frame_start ? LOAD : IDLE;

LOAD: next_state = (y_pos >= KERNEL_SIZE-1) ? PROCESS : LOAD;

PROCESS: next_state = (y_window >= IMG_HEIGHT && x_window == 0) ? IDLE : PROCESS;

default: next_state = IDLE;

endcase

end

狀態(tài)執(zhí)行

推薦使用拆分的方法，把一個(gè)狀態(tài)執(zhí)行的大always塊，分成很多子always塊。

a. 輸入圖片數(shù)據(jù)位置捕獲

當(dāng)前狀態(tài)為IDLE，圖片即將開始輸入時(shí)，將定位信號(hào)復(fù)原

當(dāng)前狀態(tài)不為IDLE, 同時(shí)輸入有效，那么坐標(biāo)根據(jù)情況自增

// Input pixel position tracking

always @(posedge clk or negedge rst_n) begin

if (!rst_n) begin

x_pos <= 0;

y_pos <= 0;

end else if (current_state == IDLE && frame_start) begin

x_pos <= 0;

y_pos <= 0;

end else if (pixel_valid && current_state != IDLE) begin

if (x_pos == IMG_WIDTH-1) begin

x_pos <= 0;

y_pos <= y_pos + 1;

end else begin

x_pos <= x_pos + 1;

end

end

end

b. Line_Buffer 的緩沖

每次開啟新的一行的數(shù)據(jù)，對(duì)Line_Buffer 全部復(fù)位

然后對(duì)對(duì)應(yīng)的位置進(jìn)行實(shí)際數(shù)據(jù)的填充

// Line buffer management

always @(posedge clk or negedge rst_n) begin

if (!rst_n) begin

for (i = 0; i <= KERNEL_SIZE; i = i + 1)

for (j = 0; j < IMG_WIDTH + 2*PADDING; j = j + 1)

line_buffer[i][j] <= 0;

end else if (pixel_valid && current_state != IDLE) begin

if (x_pos == 0) begin

// Clear the line buffer row at the start of each new line

for (k = 0; k < IMG_WIDTH + 2*PADDING; k = k + 1)

line_buffer[y_pos % (KERNEL_SIZE + 1)][k] <= 0;

end

line_buffer[y_pos % (KERNEL_SIZE + 1)][x_pos + PADDING] <= pixel_in;

end

end

c .Window position tracking

復(fù)位、一幀圖片的開始或即將進(jìn)入PROCESS狀態(tài)，對(duì)window記位進(jìn)行復(fù)位

當(dāng)前狀態(tài)位PROCESS狀態(tài)，同時(shí)沒有超過當(dāng)前圖片的高度時(shí)，對(duì)window的位置進(jìn)行對(duì)應(yīng)的變化

// Window position tracking

always @(posedge clk or negedge rst_n) begin

if (!rst_n || frame_start || (current_state == LOAD && next_state == PROCESS)) begin

x_window <= 0;

y_window <= 0;

end else if (current_state == PROCESS && y_window < IMG_HEIGHT) begin

if (x_window + STRIDE >= IMG_WIDTH) begin

x_window <= 0;

y_window <= y_window + STRIDE;

end else begin

x_window <= x_window + STRIDE;

end

end

end

d. window_buffer的處理

// Window generation and output

always @(posedge clk or negedge rst_n) begin

if (!rst_n) begin

window_valid <= 0;

for (i = 0; i < KERNEL_SIZE; i = i + 1)

for (j = 0; j < KERNEL_SIZE; j = j + 1)

window_buffer[i][j] <= 0;

end else begin

window_valid <= 0; // Default

if (current_state == PROCESS &&

x_window < IMG_WIDTH &&?

y_window < IMG_HEIGHT &&?

y_window + (KERNEL_SIZE>>1) <= y_pos) begin

// Generate window

for (i = 0; i < KERNEL_SIZE; i = i + 1) begin

for (j = 0; j < KERNEL_SIZE; j = j + 1) begin

src_y = y_window + i - (KERNEL_SIZE>>1);

src_x = x_window + j - (KERNEL_SIZE>>1);

if (src_y >= 0 && src_y < IMG_HEIGHT &&?

src_x >= 0 && src_x < IMG_WIDTH) begin

window_buffer[i][j] <= line_buffer[src_y % (KERNEL_SIZE + 1)][src_x + PADDING];

end else begin

window_buffer[i][j] <= 0; // Padding

end

end

end

window_valid <= 1;

end

end

end

當(dāng)window坐標(biāo)沒有超過圖片大小，確?？梢陨纱翱跁r(shí)，獲取生成。對(duì)KERNEL_SIZE>>1，等價(jià)于KERNEL_SIZE/2,表示中心位置的偏移量

e.g.

這樣就可以將卷積索引轉(zhuǎn)換為相對(duì)于中心的坐標(biāo)，這樣可以用于判斷是否越界,從而進(jìn)行padding補(bǔ)充

以KERNEL_SIZE=3為例

卷積核位置	src坐標(biāo)計(jì)算	結(jié)果	取值
0,0	scr_y=0+0-1=-1	越界	padding
0,1	src_y=0+0-1=-1	越界	padding
0,2	src_y=0+0-1=-1	越界	padding
1,0	src_x=0+0-1=-1	越界	padding
1,1	src_y=0,src_x=0	有效	原圖[0,0]
1,2	src_y=0,src_x=1	有效	原圖[0,1]
2,0	src_x=0+0-1=-1	越界	padding
2,1	src_y=1,src_x=0	有效	原圖[1,0]
2,2	scr_y=1,src_x=1	有效	原圖[1,1]

e. 數(shù)據(jù)窗口的展平

// Flatten window buffer for output

always @(*) begin

for (i = 0; i < KERNEL_SIZE; i = i + 1) begin

for (j = 0; j < KERNEL_SIZE; j = j + 1) begin

window_out[(KERNEL_SIZE*KERNEL_SIZE-(i*KERNEL_SIZE+j))*DATA_WIDTH-1 -: DATA_WIDTH] = window_buffer[i][j];

end

end

end

endmodule

將原本二維的的數(shù)據(jù)(寬為KERNEL_SIZE, 高為KERNEL_SIZE, 位寬為DATA_WIDTH)的數(shù)據(jù)，按照從罪小位排在最高位的順序，壓縮成一維的數(shù)據(jù)

測(cè)試

`timescale 1ns / 1ps

module window_tb();

// 測(cè)試用參數(shù) - 使用小尺寸便于觀察

parameter DATA_WIDTH = 8;

parameter IMG_WIDTH = 32;

parameter IMG_HEIGHT = 32;

parameter KERNEL_SIZE = 3;

parameter STRIDE = 1;

parameter PADDING = (KERNEL_SIZE - 1) / 2;

// 測(cè)試信號(hào)

reg clk;

reg rst_n;

reg [DATA_WIDTH-1:0] pixel_in;

reg pixel_valid;

reg frame_start;

wire [KERNEL_SIZE*KERNEL_SIZE*DATA_WIDTH-1:0] window_out;

wire window_valid;

// 實(shí)例化被測(cè)模塊

window #(

.DATA_WIDTH(DATA_WIDTH),

.IMG_WIDTH(IMG_WIDTH),

.IMG_HEIGHT(IMG_HEIGHT),

.KERNEL_SIZE(KERNEL_SIZE),

.STRIDE(STRIDE),

.PADDING(PADDING)

) dut (

.clk(clk),

.rst_n(rst_n),

.pixel_in(pixel_in),

.pixel_valid(pixel_valid),

.frame_start(frame_start),

.window_out(window_out),

.window_valid(window_valid)

);

// 時(shí)鐘生成

initial begin

clk = 0;

forever#5clk = ~clk;

end

// 測(cè)試數(shù)據(jù) - 5x5圖像

reg [DATA_WIDTH-1:0] test_image [0:IMG_HEIGHT-1][0:IMG_WIDTH-1];

// 窗口計(jì)數(shù)器

integer window_count = 0;

// 初始化測(cè)試圖像

task reset_test_image;

integer i, j;

begin

for(i = 0; i < IMG_HEIGHT; i = i + 1) begin

for(j = 0; j < IMG_WIDTH; j = j + 1) begin

test_image[i][j] =0;

end

end

end

endtask

task init_test_image;

integer i, j;

begin

for(i = 0; i < IMG_HEIGHT; i = i + 1) begin

for(j = 0; j < IMG_WIDTH; j = j + 1) begin

test_image[i][j] = i * IMG_WIDTH + j + 1;

end

end

end

endtask

// 顯示測(cè)試圖像

task display_test_image;

integer i, j;

begin

$display(" === 4x4 Test Image ===");

for(i = 0; i < IMG_HEIGHT; i = i + 1) begin

$write("Row %0d: ", i);

for(j = 0; j < IMG_WIDTH; j = j + 1) begin

$write("%3d ", test_image[i][j]);

end

$display("");

end

$display("====================== ");

end

endtask

// 發(fā)送一幀圖像數(shù)據(jù)

task send_frame;

integer i, j;

begin

$display("Sending 4x4 frame...");

init_test_image();

display_test_image();

// 發(fā)送frame_start信號(hào)

@(posedge clk);

frame_start = 1;

@(posedge clk);

frame_start = 0;

// 逐像素發(fā)送數(shù)據(jù)

for(i = 0; i < IMG_HEIGHT; i = i + 1) begin

for(j = 0; j < IMG_WIDTH; j = j + 1) begin

@(posedge clk);

pixel_in = test_image[i][j];

pixel_valid = 1;

$display("Sending pixel[%0d][%0d] = %0d at time %0t", i, j, pixel_in, $time);

end

end

@(posedge clk);

pixel_valid = 0;

$display("All pixels sent at time %0t", $time);

end

endtask

// 主測(cè)試序列

initial begin

$display("========================================");

$display("Window Test - Focus on Last Window");

$display("IMG_SIZE: %0dx%0d, KERNEL: %0dx%0d", IMG_WIDTH, IMG_HEIGHT, KERNEL_SIZE, KERNEL_SIZE);

$display("Expected windows: %0d", IMG_WIDTH * IMG_HEIGHT);

$display("========================================");

// 初始化信號(hào)

rst_n = 0;

pixel_in = 0;

pixel_valid = 0;

frame_start = 0;

reset_test_image();

// 復(fù)位序列

repeat(5) @(posedge clk);

rst_n = 1;

repeat(3) @(posedge clk);

// 發(fā)送測(cè)試幀

send_frame();

// 等待所有窗口輸出

repeat(50) @(posedge clk);

$display(" ========================================");

$display("Test Summary:");

$display("Total Windows Generated: %0d", window_count);

$display("Expected Windows: %0d", IMG_WIDTH * IMG_HEIGHT);

if(window_count == IMG_WIDTH * IMG_HEIGHT) begin

$display("SUCCESS: All windows generated!");

end else begin

$display("FAILURE: Missing windows!");

end

$display("========================================");

$finish;

end

// 窗口監(jiān)控

always @(posedge clk) begin

if(window_valid) begin

window_count = window_count + 1;

$display("Window %0d: pos(%0d,%0d) at time %0t",

window_count, dut.x_window, dut.y_window, $time);

// 顯示窗口內(nèi)容

$write("Window content: ");

$write("[%0d %0d %0d] ",

window_out[71:64], window_out[63:56], window_out[55:48]);

$write("[%0d %0d %0d] ",

window_out[47:40], window_out[39:32], window_out[31:24]);

$write("[%0d %0d %0d]",

window_out[23:16], window_out[15:8], window_out[7:0]);

$display("");

end

end

// 狀態(tài)機(jī)監(jiān)控

reg [1:0] prev_state = 2'b00;

always @(posedge clk) begin

if(dut.current_state != prev_state) begin

case(dut.current_state)

2'b00: $display("Time %0t: State -> IDLE", $time);

2'b01: $display("Time %0t: State -> LOAD", $time);

2'b10: $display("Time %0t: State -> PROCESS", $time);

default: $display("Time %0t: State -> UNKNOWN(%0d)", $time, dut.current_state);

endcase

prev_state = dut.current_state;

end

end

// 波形轉(zhuǎn)儲(chǔ)

initial begin

$dumpfile("window_tb.vcd");

$dumpvars(0, window_tb);

// 限制仿真時(shí)間

#2000;

$display("ERROR: Simulation timeout!");

$finish;

end

endmodule

結(jié)果

輸入數(shù)據(jù)

Row 0: 1 2 3 4 5
Row 1: 6 7 8 9 10
Row 2: 11 12 13 14 15
Row 3: 16 17 18 19 20

Line_Buffer 緩沖數(shù)據(jù)

Window_Buffer輸出數(shù)據(jù)

valid為高，window_buffer開始提取line_buffer數(shù)據(jù)，同時(shí)輸出展平的window_out；

window_buffer提取完畢，valid拉低