还没吃饭 发表于 2024-2-27 12:08

FPGA硬件加速Sora文生视频源代码【星嵌电子XQ138F-EVM开发板体验】(原创)

<div class='showpostmsg'><p>嗨,亲爱的工程师、学生和爱好者们,我来啦!欢迎来到神秘的星嵌世界!如果你是一位FPGA工程师或者对嵌入式异构技术感兴趣,那么你来到的地方绝对没错!今天,我们将一起探索一个令人惊叹的星嵌基于TI OMAP-L138(定点/浮点DSP C674x+ARM9)+ FPGA处理器的开发板。</p>

<div style="text-align: center;"></div>

<p>&nbsp;</p>

<p>要用FPGA加速文本到视频生成模型Sora,首先需要了解该模型的复杂性和计算需求。文本到视频生成模型涉及深度学习算法,如循环神经网络(RNN)或Transformer,用于文本处理,以及卷积神经网络(CNN)或生成对抗网络(GAN)用于视频生成。通常涉及对模型中的计算密集型部分进行硬件加速。文本到视频生成模型Sora包含多个深度学习层,如卷积层、循环层、注意力机制等。由于FPGA的并行处理能力,特别适合加速这些计算密集型任务。</p>

<div style="text-align: center;"></div>

<p>&nbsp;</p>

<p><strong>FPGA加速通常通过并行处理、流水线设计和优化内存访问模式等技术实现。我将用Verilog HDL编写一些用于文本到视频生成模型Sora的FPGA加速器模块。请注意,由于时间关系 我现在用verilog语言只写几个简单的模型加速代码,更加复杂有机会再写:</strong></p>

<pre>
<code>module TextToVideoAccelerator(

&nbsp; &nbsp; <strong>input</strong> wire clk,

&nbsp; &nbsp; <strong>input</strong> wire reset,

&nbsp; &nbsp;

&nbsp; &nbsp; // 文本输入接口<strong>input</strong> wire text_input,

&nbsp; &nbsp; <strong>input</strong> wire text_valid,

&nbsp; &nbsp; output wire text_ready,

&nbsp; &nbsp;

&nbsp; &nbsp; // 视频输出接口

&nbsp; &nbsp; output wire video_output,

&nbsp; &nbsp; output wire video_valid,

&nbsp; &nbsp; <strong>input</strong> wire video_ready

);



&nbsp; &nbsp; // 文本到视频生成模型Sora参数

&nbsp; &nbsp; parameter TEXT_LENGTH = 1024; // 文本输入长度

&nbsp; &nbsp; parameter VIDEO_WIDTH = 640;&nbsp;&nbsp;// 视频宽度

&nbsp; &nbsp; parameter VIDEO_HEIGHT = 480; // 视频高度

&nbsp; &nbsp; parameter PIXEL_DEPTH = 8;&nbsp; &nbsp; // 像素深度// 内部状态和控制信号<strong>reg</strong> internal_text_buffer ; // 文本缓冲区<strong>reg</strong> internal_video_frame ; // 视频帧缓冲区<strong>reg</strong> text_processing; // 文本处理标志<strong>reg</strong> video_generation; // 视频生成标志<strong>reg</strong> text_index; // 当前处理的文本索引<strong>reg</strong> video_index; // 当前生成的视频像素索引<strong>reg</strong> text_buffer_full; // 文本缓冲区是否已满<strong>reg</strong> video_frame_ready; // 视频帧是否已准备好// FPGA内部处理函数

&nbsp; &nbsp; always @(posedge clk or posedge reset) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;<strong>if</strong> (reset) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 重置内部状态

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;text_processing &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_generation &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;text_index &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_index &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;text_buffer_full &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_frame_ready &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end <strong>else</strong> begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 文本输入处理<strong>if</strong> (text_valid &amp;&amp; !text_ready &amp;&amp; !text_processing &amp;&amp; !text_buffer_full) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 存储文本输入到缓冲区

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; internal_text_buffer &lt;= text_input;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; text_index &lt;= text_index + 1;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; <strong>if</strong> (text_index == TEXT_LENGTH - 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 文本缓冲区已满

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;text_buffer_full &lt;= 1;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;text_processing &lt;= 1; // 开始处理文本

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; text_ready &lt;= 1; // 通知外部文本已接收

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end <strong>else</strong><strong>if</strong> (text_ready) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 清除接收准备信号

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; text_ready &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end



&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 文本处理与视频生成<strong>if</strong> (text_processing &amp;&amp; text_buffer_full) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 在这里实现文本到视频生成的算法逻辑// 例如,使用深度学习模型将文本转换为视频帧//我在下面会写<strong>for</strong> (int y = 0; y &lt; VIDEO_HEIGHT; y = y + 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; <strong>for</strong> (int x = 0; x &lt; VIDEO_WIDTH; x = x + 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 文本输入的第一个字节决定视频帧的颜色

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;internal_video_frame &lt;= internal_text_buffer;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 算法处理完成,准备输出视频帧// 调用硬件加速模块进行文本到视频帧的转换// 将文本缓冲区的内容&ldquo;传递&rdquo;给硬件加速器// TextToVideoHardwareAccelerator 是一个硬件模块// 它接收文本输入并输出视频帧数据// 实际的硬件加速模块设计

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;TextToVideoHardwareAccelerator #(

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; .TEXT_LENGTH(TEXT_LENGTH),

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; .VIDEO_WIDTH(VIDEO_WIDTH),

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; .VIDEO_HEIGHT(VIDEO_HEIGHT)

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;) accelerator (

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; .clk(clk),

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; .reset(reset),

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; .text_in(internal_text_buffer),

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; .video_frame_out(internal_video_frame)

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;);



&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 完成转换后,设置视频帧准备就绪信号

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_frame_ready &lt;= 1;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end



&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 视频输出处理<strong>if</strong> (video_ready &amp;&amp; video_frame_ready) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 输出视频帧的一个像素

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; video_output &lt;= internal_video_frame;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; video_valid &lt;= 1;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; video_index &lt;= video_index + 1;



&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 检查是否所有像素都已输出<strong>if</strong> (video_index == (VIDEO_WIDTH * VIDEO_HEIGHT) - 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 重置视频生成标志和索引

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;video_generation &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;video_index &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;video_frame_ready &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end <strong>else</strong><strong>if</strong> (video_valid) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 清除视频有效信号

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; video_valid &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end

&nbsp; &nbsp; end

endmodule

</code></pre>

<p>&nbsp;</p>

<p><strong>这个模块只是一个非常简化的模型加速代码,它展示了文本输入、处理、视频生成和输出的基本流程。由于文本到视频生成模型Sora的复杂性(例如深度学习中的Transformer模型),FPGA加速实现代码会相当长且高度专业化,并涉及多个层次的设计。下面我将再写一个简化的、概念性的Verilog文本到视频转换程序,展示如何在FPGA上搭建一个模块以加速部分处理过程:</strong></p>

<pre>
<code>// 我有一个简化版的基于FPGA的文本编码和帧预测模块

module TextToVideoAccelerator(

&nbsp; &nbsp; <strong>input</strong> wire clk, // 主时钟信号<strong>input</strong> wire rst_n, // 异步复位信号,低电平有效<strong>input</strong> wire text_in, // 输入的文本数据流

&nbsp; &nbsp; output <strong>reg</strong> video_out_rbg , // 输出的RGB格式视频帧缓冲区// 其他必要的输入/输出接口,如权重存储器接口、隐藏状态存储器接口等

);



// 内部信号声明和存储器实例化<strong>reg</strong> word_embedding ; // 文本词嵌入向量缓存<strong>reg</strong> hidden_state ; // 隐藏层状态缓存<strong>reg</strong> frame_prediction; // 当前帧预测结果// 初始化逻辑

initial begin

&nbsp; &nbsp; // 初始化内部缓存和状态<strong>for</strong> (int i = 0; i &lt; 2048; i = i + 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;word_embedding &lt;= 64&#39;b0; // 初始化词嵌入向量为0

&nbsp; &nbsp; end

&nbsp; &nbsp; <strong>for</strong> (int j = 0; j &lt; 4096; j = j + 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;hidden_state &lt;= 512&#39;b0; // 初始化隐藏状态为0

&nbsp; &nbsp; end

end



// 文本编码及帧预测流水线

always @(posedge clk or negedge rst_n) begin

&nbsp; &nbsp; <strong>if</strong> (~rst_n) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 复位操作

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;frame_prediction &lt;= 512&#39;b0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 清空视频输出缓冲区<strong>else</strong> begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 此处简化处理步骤:// 1. 对文本进行词嵌入编码// word_embedding &lt;= Embedding(text_in);// 2. 将词嵌入向量送入模型核心进行计算// hidden_state &lt;= ProcessWithTransformer(word_embedding);// 3. 使用隐藏层状态预测下一视频帧// frame_prediction &lt;= PredictFrame(hidden_state);// 将预测的帧颜色值写入视频输出缓冲区// (假设已经完成了量化和颜色空间转换)<strong>for</strong> (int y = 0; y &lt; 1080; y = y + 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;<strong>for</strong> (int x = 0; x &lt; 1920; x = x + 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // video_out_rbg &lt;= FrameToPixel(frame_prediction);// 这里仅为示意,在此可以加上复杂的加速代码

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end

&nbsp; &nbsp; end

end



endmodule

</code></pre>

<p>&nbsp;</p>

<p><strong>因为这样的项目通常涉及复杂的深度学习模型、大量的硬件逻辑描述和高度定制化的IP核设计。在实际应用中,TextEncoder会将文本序列编码为适合于神经网络模型的格式,然后通过VideoAccelerator硬件模块将编码后的文本转化为视频帧。VideoAccelerator包含许多并行计算单元、内存控制器以及针对特定文本到视频转换模型优化的数据路径结构。<br />
EEWORLDIMGTK4</strong></p>

<p><strong>根据上面的代码进行再次优化,我将提供一个修改后的代码版本。这个版本移除了TextToVideoHardwareAccelerator的实例化,并添加了对文本到视频帧转换逻辑的占位符。同时,为了保证时序正确性,在实际设计中加入了硬件加速器处理完成的信号反馈:</strong></p>

<pre>
<code>module TextToVideoAccelerator(

&nbsp; &nbsp; <strong>input</strong> wire clk,

&nbsp; &nbsp; <strong>input</strong> wire reset,

&nbsp; &nbsp;

&nbsp; &nbsp; // 文本输入接口<strong>input</strong> wire text_input,

&nbsp; &nbsp; <strong>input</strong> wire text_valid,

&nbsp; &nbsp; output wire text_ready,

&nbsp; &nbsp;

&nbsp; &nbsp; // 视频输出接口

&nbsp; &nbsp; output wire video_output,

&nbsp; &nbsp; output wire video_valid,

&nbsp; &nbsp; <strong>input</strong> wire video_ready

);



&nbsp; &nbsp; // 假设的文本到视频生成模型参数

&nbsp; &nbsp; parameter TEXT_LENGTH = 1024; // 文本输入长度

&nbsp; &nbsp; parameter VIDEO_WIDTH = 640;&nbsp;&nbsp;// 视频宽度

&nbsp; &nbsp; parameter VIDEO_HEIGHT = 480; // 视频高度

&nbsp; &nbsp; parameter PIXEL_DEPTH = 8;&nbsp; &nbsp; // 像素深度// 内部状态和控制信号
&nbsp;   <strong>reg</strong> internal_text_buffer ; // 文本缓冲区
&nbsp;   <strong>reg</strong> internal_video_frame ; // 视频帧缓冲区
&nbsp;   <strong>reg</strong> text_processing; // 文本处理标志
&nbsp;   <strong>reg</strong> video_generation; // 视频生成标志
&nbsp;   <strong>reg</strong> text_index; // 当前处理的文本索引
&nbsp;   <strong>reg</strong> video_index; // 当前生成的视频像素索引
&nbsp;   <strong>reg</strong> text_buffer_full; // 文本缓冲区是否已满
&nbsp;   <strong>reg</strong> video_frame_ready; // 视频帧是否已准备好
&nbsp;   <strong>reg</strong> hardware_accelerator_done; // 硬件加速器完成信号
&nbsp;   
&nbsp;   // FPGA内部处理函数

&nbsp; &nbsp; always @(posedge clk or posedge reset) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;<strong>if</strong> (reset) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 重置内部状态

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;text_processing &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_generation &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;text_index &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_index &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;text_buffer_full &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_frame_ready &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;hardware_accelerator_done &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end <strong>else</strong> begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 文本输入处理<strong>if</strong> (text_valid &amp;&amp; !text_ready &amp;&amp; !text_processing &amp;&amp; !text_buffer_full) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 存储文本输入到缓冲区

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; internal_text_buffer &lt;= text_input;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; text_index &lt;= text_index + 1;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; <strong>if</strong> (text_index == TEXT_LENGTH - 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 文本缓冲区已满

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;text_buffer_full &lt;= 1;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;text_processing &lt;= 1; // 开始处理文本

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; text_ready &lt;= 1; // 通知外部文本已接收

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end <strong>else</strong><strong>if</strong> (text_ready) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 清除接收准备信号

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; text_ready &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end



&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 文本处理与视频生成(占位符,调用硬件加速器或实现相应算法)<strong>if</strong> (text_processing &amp;&amp; text_buffer_full) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 触发硬件加速器并等待其完成// 硬件加速器在一个时钟周期内完成(这在不同算法情况下可能需要多个时钟周期)

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; hardware_accelerator_done &lt;= 1;



&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 模拟硬件加速器完成处理后,设置视频帧准备就绪信号<strong>if</strong> (hardware_accelerator_done) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;video_frame_ready &lt;= 1;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;hardware_accelerator_done &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end



&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 视频输出处理<strong>if</strong> (video_ready &amp;&amp; video_frame_ready) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 输出视频帧的一个像素

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; video_output &lt;= internal_video_frame;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; video_valid &lt;= 1;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; video_index &lt;= video_index + 1;



&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 检查是否所有像素都已输出<strong>if</strong> (video_index == (VIDEO_WIDTH * VIDEO_HEIGHT) - 1) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 重置视频生成标志和索引

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;video_generation &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;video_index &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;video_frame_ready &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end <strong>else</strong><strong>if</strong> (video_valid) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 清除视频有效信号

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; video_valid &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end

&nbsp; &nbsp; end

endmodule



// 在模块外部实例化硬件加速器模块(它是一个现成的IP核)// 注意:此部分未给出具体实现,因为`TextToVideoHardwareAccelerator`的具体行为取决于实际硬件或IP核// 应当根据该硬件加速器提供的接口和功能进行适配// 硬件加速器提供以下接口:// .start(text_in, video_frame_out)// .done()// 并且可以在内部处理时钟域中工作

TextToVideoHardwareAccelerator #(

&nbsp; &nbsp; .TEXT_LENGTH(TEXT_LENGTH),

&nbsp; &nbsp; .VIDEO_WIDTH(VIDEO_WIDTH),

&nbsp; &nbsp; .VIDEO_HEIGHT(VIDEO_HEIGHT)

) accelerator (

&nbsp; &nbsp; .clk(clk),&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 时钟输入(通常硬件加速器有自己的内部时钟网络,此处仅为示例)

&nbsp; &nbsp; .reset(reset),&nbsp; &nbsp;&nbsp; &nbsp;// 复位信号

&nbsp; &nbsp; .text_in(internal_text_buffer),

&nbsp; &nbsp; .video_frame_out(internal_video_frame),

&nbsp; &nbsp; // 在此可以添加其他控制或状态信号

);

</code></pre>

<p>EEWORLDIMGTK5</p>

<p>请注意,上述代码可以使用一个握手信号来启动硬件加速器,并在加速器完成处理后通过done信号触发video_frame_ready信号。<br />
EEWORLDIMGTK6</p>

<p>现在用我的语言解释一下代码:</p>

<p>模块名:TextToVideoAccelerator,就是个超级翻译机,把文字瞬间变视频!不过它不是哈利波特魔法世界的产品,而是通过FPGA技术实现的。</p>

<pre>
<code><strong>input</strong> wire clk, // 这是我们的宇宙脉搏,每一跳都代表着时间滴答一声向前。<strong>input</strong> wire reset, // 哎呀喂,紧急刹车信号,一旦拉响,整个系统得从头再来!// 文本输入接口,就像你给外星人写信的传送门<strong>input</strong> wire text_input, // 一个大大的信息包,里面装着32位的神秘文字信息<strong>input</strong> wire text_valid, // &ldquo;嘿,我这里有货真价实的文字哦!&rdquo;&mdash;&mdash;来自上一级的通知信号

output wire text_ready, // &ldquo;OK,我已经准备好了,快把你的文本扔过来吧!&rdquo;&mdash;&mdash;对上级的回答// 视频输出接口,就像电影院的大银幕投放口

output wire video_output, // 八位彩色像素,组成视频的最小单位,每一帧画面就靠它们拼接而成

output wire video_valid, // &ldquo;看这里,现在我给出的是有效像素数据,别眨眼!&rdquo;<strong>input</strong> wire video_ready; // &ldquo;小加速器,我这边准备好接收下一像素了,放马过来吧!&rdquo;// 高级设定(其实就是程序员偷懒用的魔法数字)

parameter TEXT_LENGTH = 1024; // 文本缓冲区能装1024个字节,够编一部微型小说啦

parameter VIDEO_WIDTH = 640;&nbsp;&nbsp;// 每帧视频宽度640像素,小屏也清晰

parameter VIDEO_HEIGHT = 480; // 高度480像素,经典分辨率带你回味复古风

parameter PIXEL_DEPTH = 8;&nbsp; &nbsp; // 每个像素有8位深度,足够表现各种颜色斑斓// 内部状态和控制信号,相当于我们设备的&ldquo;五脏六腑&rdquo;<strong>reg</strong> internal_text_buffer [...]; // 文本暂存仓库,满满的都是智慧语言<strong>reg</strong> internal_video_frame [...]; // 视频帧临时存放所,一行行像素在等待组装成画面<strong>reg</strong> text_processing; // 文字正在炼丹炉里进行神奇转换?!<strong>reg</strong> video_generation; // 视频生成大法已启动?!<strong>reg</strong> text_index; // 当前处理的文字位置指示牌<strong>reg</strong> video_index; // 正在搬运的像素坐标指南针<strong>reg</strong> text_buffer_full; // 文字仓库是否塞满的警示灯<strong>reg</strong> video_frame_ready; // 视频帧出炉通知铃铛<strong>reg</strong> hardware_accelerator_done; // 硬件加速器工作完成的小旗子(幻想中)// FPGA内核操作法则

always @(posedge clk or posedge reset) begin // 当宇宙脉搏跳动一次或有人喊&ldquo;重来&rdquo;时...<strong>if</strong> (reset) begin // 如果重启按钮被按下...// 把所有状态归零,重新开始冒险旅程

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;(此处列出一堆状态变量清零语句)



&nbsp; &nbsp; end <strong>else</strong> begin // 否则,我们继续正常工作流程...// 文本输入处理部分:像是在做快递分拣<strong>if</strong> (text_valid &amp;&amp; !text_ready &amp;&amp; !text_processing &amp;&amp; !text_buffer_full) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 收到有效文本且尚未处理完毕,赶紧放入仓库

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;(将text_input存入internal_text_buffer,并更新text_index)

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 仓库满了?那咱们就开始施展魔法吧

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;(判断并设置text_processing标志和text_buffer_full标志)

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;text_ready &lt;= 1; // 对外界说:&ldquo;亲,你的文本我收到了哦!&rdquo;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end <strong>else</strong><strong>if</strong> (text_ready) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 已经确认收到文本,撤下&ldquo;待命&rdquo;信号

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;text_ready &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 文本转视频的核心区域,此处脑洞大开,实际操作交给硬件加速器<strong>if</strong> (text_processing &amp;&amp; text_buffer_full) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 幻想一下,我们摇响硬件加速器的小铃铛:&ldquo;嘿,开工啦!&rdquo;// (实际上,应使用真实接口触发硬件加速器,并等待其完成)

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;hardware_accelerator_done &lt;= 1; // 想象加速器瞬间完成任务// 加速器完成后,仿佛魔术师挥舞魔杖,&ldquo;叮&rdquo;一声,视频帧闪亮登场<strong>if</strong> (hardware_accelerator_done) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; video_frame_ready &lt;= 1; // 设置视频帧已完成的信号

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; hardware_accelerator_done &lt;= 0; // 重置完成标志,准备下一轮挑战

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end



&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;// 视频输出环节:像放映员逐帧播放胶片<strong>if</strong> (video_ready &amp;&amp; video_frame_ready) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 把仓库里的一个像素搬上荧幕

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_output &lt;= internal_video_frame[计算出的坐标];

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_valid &lt;= 1; // &ldquo;注意啦,这一像素很靠谱,可以展示出来!&rdquo;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_index &lt;= video_index + 1; // 移动到下一个像素的位置// 检查是否播完一整帧<strong>if</strong> (video_index已经指向最后一像素) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; // 结束本帧播放,收拾好心情,迎接下一帧

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp; (重置相关标志和索引)

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;end

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end <strong>else</strong><strong>if</strong> (video_valid) begin

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 如果当前像素不再有效,赶紧关掉绿灯

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;video_valid &lt;= 0;

&nbsp; &nbsp;&nbsp; &nbsp;&nbsp;&nbsp;end

&nbsp; &nbsp; end

end



// 实例化那个神乎其技的硬件加速器// 就像请来一位神秘的魔法师助手(假设他/她叫&quot;accelerator&quot;)// (此处提供了一些参数和接口连接,但具体怎么调用还得看他/她的说明书)

TextToVideoHardwareAccelerator #(

&nbsp; &nbsp; .TEXT_LENGTH(TEXT_LENGTH),

&nbsp; &nbsp; .VIDEO_WIDTH(VIDEO_WIDTH),

&nbsp; &nbsp; .VIDEO_HEIGHT(VIDEO_HEIGHT)

) accelerator (

&nbsp; &nbsp; .clk(clk),&nbsp; &nbsp;&nbsp; &nbsp;&nbsp; &nbsp;// 给魔法师同步心跳

&nbsp; &nbsp; .reset(reset),&nbsp; &nbsp;&nbsp; &nbsp;// 发生意外情况,魔法师也需要知道&ldquo;重新施法&rdquo;的命令

&nbsp; &nbsp; .text_in(internal_text_buffer),

&nbsp; &nbsp; .video_frame_out(internal_video_frame),

&nbsp; &nbsp; // 在此可以添加其他的控制线和反馈信号,可以续炼上乘秘籍

);

</code></pre>

<p><strong>这部分需要基于具体的文本到视频生成模型Sora(如Transformer等)进行设计,并将其优化成适合FPGA并行计算的形式,其中涉及大量的矩阵运算、注意力机制、卷积操作等硬件加速模块的设计与整合。FPGA加速文本到视频生成模型的代码涉及到大量的硬件设计细节和深度学习算法实现:</strong></p>

<p>&nbsp;</p>

<p>&nbsp;</p>

<p>// 一个基于Transformer的文本编码器和解码器</p>

<p>module TextToVideoAccelerator(</p>

<p>&nbsp; &nbsp; input wire clk,</p>

<p>&nbsp; &nbsp; input wire rst_n,</p>

<p>&nbsp; &nbsp; input wire text_in, // 输入的文字数据流</p>

<p>&nbsp; &nbsp; output reg video_out , // 输出的视频像素矩阵</p>

<p>);</p>

<p>// 定义内部需要的存储和计算资源</p>

<p>reg encoder_hidden_state ;</p>

<p>wire decoder_hidden_state ;</p>

<p>reg attention_weights ;</p>

<p>// 文本编码模块</p>

<p>encoder_text_to_hidden_encoder encoder_instance (</p>

<p>&nbsp; &nbsp; .clk(clk),</p>

<p>&nbsp; &nbsp; .rst_n(rst_n),</p>

<p>&nbsp; &nbsp; .text_input(text_in),</p>

<p>&nbsp; &nbsp; .hidden_state(encoder_hidden_state)</p>

<p>);</p>

<p>// 注意力机制模块(跨多个头)</p>

<p>multi_head_attention attention_instance (</p>

<p>&nbsp; &nbsp; .clk(clk),</p>

<p>&nbsp; &nbsp; .rst_n(rst_n),</p>

<p>&nbsp; &nbsp; .encoder_states(encoder_hidden_state),</p>

<p>&nbsp; &nbsp; .decoder_states(decoder_hidden_state),</p>

<p>&nbsp; &nbsp; .attention_weights(attention_weights)</p>

<p>);</p>

<p>// 解码器模块</p>

<p>decoder_hidden_to_video_decoder decoder_instance (</p>

<p>&nbsp; &nbsp; .clk(clk),</p>

<p>&nbsp; &nbsp; .rst_n(rst_n),</p>

<p>&nbsp; &nbsp; .encoder_states(encoder_hidden_state),</p>

<p>&nbsp; &nbsp; .attention_weights(attention_weights),</p>

<p>&nbsp; &nbsp; .decoder_hidden_state(decoder_hidden_state),</p>

<p>&nbsp; &nbsp; .video_output(video_out)</p>

<p>);</p>

<p>// 以上各模块的具体实现将非常复杂,涉及大量并行化处理、流水线操作以及自定义IP核设计</p>

<p>endmodule</p>

<p>&nbsp;</p>

<p>&nbsp;</p>

<div style="text-align: center;"></div>

<p>&nbsp;</p>

<p>在上面的模型中:</p>

<ul type="1">
        <li>encoder_text_to_hidden_encoder 模块负责将输入的文本序列转换为连续的隐藏状态向量。</li>
        <li>multi_head_attention 模块根据编码器产生的隐藏状态执行注意力运算,这通常在Transformer模型中非常重要。</li>
        <li>decoder_hidden_to_video_decoder 模块则利用编码器隐藏状态和注意力权重来生成视频帧。<br />
        &nbsp;</li>
</ul>

<p>每个模块都需要详细设计,包括逻辑综合、优化布局布线以适应FPGA的结构,并且必须考虑大量的并行计算单元、内存层次结构以及数据搬移策略,以便有效地加速文本到视频的生成过程。此外,还需要对接如卷积神经网络(CNN)等其他组件以生成图像帧,这些通常不在Transformer框架内直接完成。</p>

<p>同时,根据具体模型的特性,还需要增加更多的内部状态变量以及控制信号来协调不同阶段的数据流和流水线操作,确保在保持数据正确性的同时最大化FPGA的并行处理能力。</p>

<p>FPGA加速通常会涉及将模型的不同部分映射到FPGA上,并使用高级综合工具(如Xilinx Vitis或Intel OpenVINO),然后在FPGA上进行编译和部署。</p>

<p>由于文本到视频生成模型Sora的复杂性,FPGA加速的具体实现将取决于模型的细节、FPGA的硬件资源以及可用的高级综合工具。通常,这需要一个专业的团队,包括硬件工程师、深度学习专家和FPGA软件工程师,来共同设计和实现这样的解决方案。</p>

<p>为了在FPGA上高效实现文本到视频的生成模型,需要:</p>

<ul>
        <li><strong><strong>设计并实现模型的核心算法</strong></strong>:包括词嵌入、自注意力机制、解码等,这些通常需要通过高层次综合工具从已训练好的神经网络模型映射至硬件结构。</li>
        <li><strong><strong>优化内存访问与数据流</strong></strong>:合理利用FPGA的BRAM资源存储权重和中间结果,并实现高效的读写操作。</li>
        <li><strong><strong>采用流水线技术</strong></strong>:最大化利用FPGA并行处理能力,对计算流程进行流水线划分。</li>
        <li><strong><strong>量化与定制化IP核</strong></strong>:对模型进行量化以适应FPGA资源,并可能需要设计定制化IP核来执行特定运算。<br />
        &nbsp;</li>
</ul>

<p>&nbsp;</p>

<div style="text-align: center;"></div>

<p>&nbsp;</p>

<p>在实际应用中,文本到视频生成模型Sora涉及更复杂的网络结构,如Transformer模型,并且需要处理大量数据和计算。编写FPGA加速文本到视频生成模型Sora的代码是一个复杂的过程,因为文本到视频生成模型(如 Sora、DALL-E 2等)有大量的计算需求。FPGA加速的关键在于利用并行处理能力来优化计算密集型任务。通常,这涉及到将深度学习模型的某些层映射到FPGA的逻辑资源上,并优化数据传输和计算过程。</p>

<p>今天先写到这里...</p>

<p>总之,上面的内容就是描绘了一个奇幻世界的场景,其中有一位勤劳的扫地僧电子工匠 &ldquo;FPGA&rdquo;变身为神秘的硬件加速器,将传来的文本信息一步步转化为生动活泼的视频流的武功心法。</p>

<p>&nbsp;</p>

<p>希望上面的经验能对您有所帮助!</p>

<p>谢谢!</p>

<p>本人网名:还没吃饭中<br />
2024年2月18日</p>
</div><script>                                        var loginstr = '<div class="locked">查看本帖全部内容,请<a href="javascript:;"   style="color:#e60000" class="loginf">登录</a>或者<a href="https://bbs.eeworld.com.cn/member.php?mod=register_eeworld.php&action=wechat" style="color:#e60000" target="_blank">注册</a></div>';
                                       
                                        if(parseInt(discuz_uid)==0){
                                                                                                (function($){
                                                        var postHeight = getTextHeight(400);
                                                        $(".showpostmsg").html($(".showpostmsg").html());
                                                        $(".showpostmsg").after(loginstr);
                                                        $(".showpostmsg").css({height:postHeight,overflow:"hidden"});
                                                })(jQuery);
                                        }                </script><script type="text/javascript">(function(d,c){var a=d.createElement("script"),m=d.getElementsByTagName("script"),eewurl="//counter.eeworld.com.cn/pv/count/";a.src=eewurl+c;m.parentNode.insertBefore(a,m)})(document,523)</script>
页: [1]
查看完整版本: FPGA硬件加速Sora文生视频源代码【星嵌电子XQ138F-EVM开发板体验】(原创)