• <ins id="pjuwb"></ins>
    <blockquote id="pjuwb"><pre id="pjuwb"></pre></blockquote>
    <noscript id="pjuwb"></noscript>
          <sup id="pjuwb"><pre id="pjuwb"></pre></sup>
            <dd id="pjuwb"></dd>
            <abbr id="pjuwb"></abbr>
            posts - 43,  comments - 64,  trackbacks - 0

            注:本文的代碼圖片資料選自NVIDIA CUDAProgramming Guide,原作者保留所有著作權(quán)。

              NVIDIA近日終于發(fā)布了CUDA,有可能作為下一代SDK10的一部分奉送給樂于發(fā)掘GPU計算能力的專業(yè)人員。感興趣的朋友可以去這里一探究竟,下載嘗鮮,提供了大量的范例。
              我們都知道,GPU的并行運算性能是極為強悍的,如此豐富的計算資源如果浪費著不用,就用來跑跑游戲是遠遠不行的。而傳統(tǒng)的圖形API又單單的只提供了圖形操作的功能,沒有提供類似于CPU那樣通用計算的接口,所以說以往的方法都是很麻煩而且需要相當?shù)慕?jīng)驗的 —— 比如用HDR圖片作Cube Map的時候,如果使用的是Paul提供的那種類似于經(jīng)緯圖的紋理,就需要大量的反三角函數(shù)的計算,而用Vertex Shader作反三角又太浪費時間,于是人們用1D紋理作線性插值查找表進行快速計算(訪問數(shù)據(jù)紋理)。這個例子也可以看作是最基本的GPU計算。


            CUDA的誕生
              使用傳統(tǒng)API進行計算是個不可挽回的錯誤,CUDA的出現(xiàn)將改變這一狀況。CUDA主要在驅(qū)動程序方面和函數(shù)庫方面進行了擴充。在CUDA庫中提供了標準的FFT與BLAS庫,一個為NVDIA GPU設(shè)計的C編譯器。CUDA的特色如下,引自NVIDIA的官方說明:
                1、為并行計算設(shè)計的統(tǒng)一硬件軟件架構(gòu)。有可能在G80系列上得到發(fā)揮
                2、在GPU內(nèi)部實現(xiàn)數(shù)據(jù)緩存和多線程管理。這個強,思路有些類似于XB360 PS3上的CPU編程
                3、在GPU上可以使用標準C語言進行編寫。
                4、標準離散FFT庫和BLAS基本線性代數(shù)計算庫。
                5、一套CUDA計算驅(qū)動。
                6、提供從CPU到GPU的加速數(shù)據(jù)上傳性能。瓶頸就在于此
                7、CUDA驅(qū)動可以和OpenGL DirectX驅(qū)動交互操作。這強,估計也可以直接操作渲染管線
                8、與SLI配合實現(xiàn)多硬件核心并行計算。
                9、同時支持Linux和Windows。這個就是噱頭了
              看過了宣傳,您可以看一下CUDA提供的Programming Guide和其他的文檔。NVIDIA我覺得有些類似圖形界的Microsoft,提供精良的裝備諸如SDK和開發(fā)文檔等等,比ATi好多了。

            CUDA本質(zhì)
              CUDA的本質(zhì)是,NVIDIA為自家的GPU編寫了一套編譯器NVCC極其相關(guān)的庫文件。CUDA的應用程序擴展名可以選擇是.cu,而不是.cpp等。NVCC是一個預處理器和編譯器的混合體。當遇到CUDA代碼的時候,自動編譯為GPU執(zhí)行的代碼,也就是生成調(diào)用CUDA Driver的代碼。如果碰到Host C++代碼,則調(diào)用平臺自己的C++編譯器進行編譯,比如Visual Studio C++自己的Microsoft C++ Compiler。然后調(diào)用Linker把編譯好的模塊組合在一起,和CUDA庫與標準C\C++庫鏈接成為最終的CUDA Application。由此可見,NVCC模仿了類似于GCC一樣的通用編譯器的工作原理(GCC編譯C\C++代碼本質(zhì)上就是調(diào)用cc和g++)。NVCC有著復雜的選項,詳情參閱CUDA SDK中的NVCC相關(guān)文檔。

            CUDA編程概念
            Device
              CUDA API提供接口枚舉出系統(tǒng)中可以作為計算設(shè)備使用的硬件為計算進行初始化等操作。類似于DX編程中的初始化COM接口。

            Texture
              線性內(nèi)存中的數(shù)據(jù)和數(shù)組中的數(shù)據(jù)都可以作為紋理使用。不過數(shù)組在緩存層面上更適合優(yōu)化。紋理的概念類似于傳統(tǒng)的圖像紋理,可以以8 16 32位的整數(shù)儲存,也可以用fp16格式進行儲存。而且當把數(shù)組轉(zhuǎn)換為紋理的時候,還有一些有點,比如整數(shù)與fp16數(shù)字可以選擇統(tǒng)一的轉(zhuǎn)換到32bit浮點數(shù),還可以使用數(shù)組邊界這個特性,還可以進行過濾操作。

            OpenGL/DirectX Interoperability
              OpenGL的幀緩沖與DirectX9的頂點緩沖可以被映射到CUDA可操作的地址空間中,讓CUDA讀寫幀緩沖里面的數(shù)據(jù)。不過CUDA Context一次只能操作一個Direct3D設(shè)備。當前CUDA還不支持對DX10進行類似的操作,除了DX9頂點緩沖也不允許進行映射,而且一次只能映射一次。(這個地方NVIDIA沒有說清楚,我估計是指只有一個Mapping Slot)

            Thread Block
            A thread block is a batch of threads that can cooperate together by efficiently sharing data through some fast shared memory and synchronizing their execution to coordinate memory accesses.
              ThreadBlock由一系列線程組成,這些線程可以快速共享內(nèi)存,同步內(nèi)存訪問。每個線程都有個ID,這個ID好像平面坐標一般。線程組成Grid。示意圖如下:

            CudaThread.PNG


            Memory Model
            A thread that executes on the device has only access to the device’s DRAM and on-chip memory through the following memory spaces:
            ? Read-write per-thread registers,
            ? Read-write per-thread local memory,
            ? Read-write per-block shared memory,
            ? Read-write per-grid global memory,
            ? Read-only per-grid constant memory,
            ? Read-only per-grid texture memory.
            The global, constant, and texture memory spaces can be read from or written to by the host and are persistent across kernel calls by the same application.
              在CUDA中我們要接觸到的內(nèi)存主要有:寄存器,Local內(nèi)存,Shared內(nèi)存,Global內(nèi)存,Constant內(nèi)存,Texture內(nèi)存。 有些類似于C內(nèi)存的分配類型了。而且內(nèi)存可以分配為數(shù)組或者是普通線性內(nèi)存,CUDA提供API可以正確的進行內(nèi)存拷貝等操作。

            cudamemory.PNG


              后面我們將談到如何優(yōu)化GPU內(nèi)存。從上面的資料我們可以看出,這里的Grid概念類似于Process,也就是為線程執(zhí)行分配資源的單元,而只有線程是真正計算的部分。Local Memory類似線程的棧。Texture Memory類似于堆內(nèi)存區(qū)。

            具體操作
              我以CUDA附帶的simpleCUBLAS作為例子。
            #include?<stdio.h>
            #include?
            <stdlib.h>
            #include?
            <string.h>

            /*?Includes,?cuda?*/
            #include?
            "cublas.h"

            /*?Matrix?size?*/
            #define?N??(275)

            /*?Host?implementation?of?a?simple?version?of?sgemm?*///使用CPU進行Matrix乘法計算的算式
            static?void?simple_sgemm(int?n,?float?alpha,?const?float?*A,?const?float?*B,
            ?????????????????????????
            float?beta,?float?*C)
            {
            ????
            int?i;
            ????
            int?j;
            ????
            int?k;
            ????
            for?(i?=?0;?i?<?n;?++i)?{
            ????????
            for?(j?=?0;?j?<?n;?++j)?{
            ????????????
            float?prod?=?0;
            ????????????
            for?(k?=?0;?k?<?n;?++k)?{
            ????????????????prod?
            +=?A[k?*?n?+?i]?*?B[j?*?n?+?k];
            ????????????}

            ????????????C[j?
            *?n?+?i]?=?alpha?*?prod?+?beta?*?C[j?*?n?+?i];
            ????????}

            ????}

            }


            /*?Main?*/
            int?main(int?argc,?char**?argv)
            {????
            ????cublasStatus?status;
            ????
            float*?h_A;
            ????
            float*?h_B;
            ????
            float*?h_C;
            ????
            float*?h_C_ref;
            ????
            float*?d_A?=?0;
            ????
            float*?d_B?=?0;
            ????
            float*?d_C?=?0;
            ????
            float?alpha?=?1.0f;
            ????
            float?beta?=?0.0f;
            ????
            int?n2?=?N?*?N;
            ????
            int?i;
            ????
            float?error_norm;
            ????
            float?ref_norm;
            ????
            float?diff;

            ????
            /*?Initialize?CUBLAS?*///初始化CUBLAS庫
            ????status?
            =?cublasInit();
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?CUBLAS?initialization?error\n");
            ????????
            return?EXIT_FAILURE;
            ????}


            ????
            /*?Allocate?host?memory?for?the?matrices?*///分配內(nèi)存,這3個是257*257的大矩陣
            ????h_A?
            =?(float*)malloc(n2?*?sizeof(h_A[0]));
            ????
            if?(h_A?==?0)?{
            ????????fprintf?(stderr,?
            "!!!!?host?memory?allocation?error?(A)\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????h_B?
            =?(float*)malloc(n2?*?sizeof(h_B[0]));
            ????
            if?(h_B?==?0)?{
            ????????fprintf?(stderr,?
            "!!!!?host?memory?allocation?error?(B)\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????h_C?
            =?(float*)malloc(n2?*?sizeof(h_C[0]));
            ????
            if?(h_C?==?0)?{
            ????????fprintf?(stderr,?
            "!!!!?host?memory?allocation?error?(C)\n");
            ????????
            return?EXIT_FAILURE;
            ????}


            ????
            /*?Fill?the?matrices?with?test?data?*/
            ????
            for?(i?=?0;?i?<?n2;?i++)?{
            ????????h_A[i]?
            =?rand()?/?(float)RAND_MAX;
            ????????h_B[i]?
            =?rand()?/?(float)RAND_MAX;
            ????????h_C[i]?
            =?rand()?/?(float)RAND_MAX;
            ????}


            ????
            /*?Allocate?device?memory?for?the?matrices?*/ //在GPU設(shè)備上分配內(nèi)存
            ????status?
            =?cublasAlloc(n2,?sizeof(d_A[0]),?(void**)&d_A);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?device?memory?allocation?error?(A)\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????status?
            =?cublasAlloc(n2,?sizeof(d_B[0]),?(void**)&d_B);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?device?memory?allocation?error?(B)\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????status?
            =?cublasAlloc(n2,?sizeof(d_C[0]),?(void**)&d_C);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?device?memory?allocation?error?(C)\n");
            ????????
            return?EXIT_FAILURE;
            ????}


            ????
            /*?Initialize?the?device?matrices?with?the?host?matrices?*/ //把HOST內(nèi)的矩陣上傳到GPU去
            ????status?
            =?cublasSetVector(n2,?sizeof(h_A[0]),?h_A,?1,?d_A,?1);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?device?access?error?(write?A)\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????status?
            =?cublasSetVector(n2,?sizeof(h_B[0]),?h_B,?1,?d_B,?1);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?device?access?error?(write?B)\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????status?
            =?cublasSetVector(n2,?sizeof(h_C[0]),?h_C,?1,?d_C,?1);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?device?access?error?(write?C)\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????
            ????
            /*?Performs?operation?using?plain?C?code?*/ //使用CPU進行矩陣乘法計算
            ????simple_sgemm(N,?alpha,?h_A,?h_B,?beta,?h_C);
            ????h_C_ref?
            =?h_C;

            ????
            /*?Clear?last?error?*/
            ????cublasGetError();

            ????
            /*?Performs?operation?using?cublas?*/ //Wow !使用GPU計算
            ????cublasSgemm(
            'n',?'n',?N,?N,?N,?alpha,?d_A,?N,?d_B,?N,?beta,?d_C,?N);
            ????status?
            =?cublasGetError();
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?kernel?execution?error.\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????
            ????
            /*?Allocate?host?memory?for?reading?back?the?result?from?device?memory?*/ //分配HOST內(nèi)存準備存放結(jié)果
            ????h_C?
            =?(float*)malloc(n2?*?sizeof(h_C[0]));
            ????
            if?(h_C?==?0)?{
            ????????fprintf?(stderr,?
            "!!!!?host?memory?allocation?error?(C)\n");
            ????????
            return?EXIT_FAILURE;
            ????}


            ????
            /*?Read?the?result?back?*/ //回讀
            ????status?
            =?cublasGetVector(n2,?sizeof(h_C[0]),?d_C,?1,?h_C,?1);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?device?access?error?(read?C)\n");
            ????????
            return?EXIT_FAILURE;
            ????}


            ????
            /*?Check?result?against?reference?*/
            ????error_norm?
            =?0;
            ????ref_norm?
            =?0;
            ????
            for?(i?=?0;?i?<?n2;?++i)?{
            ????????diff?
            =?h_C_ref[i]?-?h_C[i];
            ????????error_norm?
            +=?diff?*?diff;
            ????????ref_norm?
            +=?h_C_ref[i]?*?h_C_ref[i];
            ????}

            ????error_norm?
            =?(float)sqrt((double)error_norm);
            ????ref_norm?
            =?(float)sqrt((double)ref_norm);
            ????
            if?(fabs(ref_norm)?<?1e-7)?{
            ????????fprintf?(stderr,?
            "!!!!?reference?norm?is?0\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????printf(?
            "Test?%s\n",?(error_norm?/?ref_norm?<?1e-6f)???"PASSED"?:?"FAILED");

            ????
            /*?Memory?clean?up?*/
            ????free(h_A);
            ????free(h_B);
            ????free(h_C);
            ????free(h_C_ref);
            ????status?
            =?cublasFree(d_A);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?memory?free?error?(A)\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????status?
            =?cublasFree(d_B);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?memory?free?error?(B)\n");
            ????????
            return?EXIT_FAILURE;
            ????}

            ????status?
            =?cublasFree(d_C);
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?memory?free?error?(C)\n");
            ????????
            return?EXIT_FAILURE;
            ????}


            ????
            /*?Shutdown?*///關(guān)閉CUBLAS卸載資源
            ????status?
            =?cublasShutdown();
            ????
            if?(status?!=?CUBLAS_STATUS_SUCCESS)?{
            ????????fprintf?(stderr,?
            "!!!!?shutdown?error?(A)\n");
            ????????
            return?EXIT_FAILURE;
            ????}


            ????
            if?(argc?<=?1?||?strcmp(argv[1],?"-noprompt"))?{
            ????????printf(
            "\nPress?ENTER?to?exit\n");
            ????????getchar();
            ????}

            ????
            return?EXIT_SUCCESS;
            }


              除了那些個容錯的代碼,我們可以看出使用CUBLAS庫進行計算還是非常簡潔直觀的。詳細的資料請看CUDA SDK自帶的范例。

            我的展望
              ATi(AMD)坐不住的,應該會積極開發(fā)CPU與GPU融合的相關(guān)組建。
              瓶頸在CPU - GPU帶寬上,NV很有可能推出優(yōu)化過的nForce芯片組提供高帶寬。
              用ICE配上CUDA組成分布式的GPU計算平臺怎么樣?!大伙不妨暢想暢想。
              下一代BOINC計算平臺內(nèi)的項目能夠提供基于GPU的計算客戶端。
            posted on 2007-02-24 14:42 周波 閱讀(4531) 評論(6)  編輯 收藏 引用 所屬分類: Cg藝術(shù)無庸技術(shù)奇思妙想

            FeedBack:
            # re: Pure GPU Computing Platform : NVIDIA CUDA Tutorial
            2007-02-24 22:58 | lai3d
            CUDA stands for Compute Unified Device Architecture and is a new hardware and software architecture for issuing and managing computations on the GPU as a data-parallel computing device without the need of mapping them to a graphics API. It is available for the GeForce 8800 Series and beyond.

            8800及以上才支持啊,看來暫時不用考慮在游戲引擎里應用  回復  更多評論
              
            # re: Pure GPU Computing Platform : NVIDIA CUDA Tutorial
            2007-02-24 23:54 | Jedimaster
            不過我的6200A NV44A也可以運行調(diào)試版本的DEMO...而且我沒有安裝推薦的驅(qū)動程序。估計有向下兼容的考慮。畢竟架構(gòu)沒有太大的變化。  回復  更多評論
              
            # re: Pure GPU Computing Platform : NVIDIA CUDA Tutorial
            2007-02-26 11:25 | 空明流轉(zhuǎn)
            我用的是A卡。。。
            觀望中,準備將目前手里的項目用CUDA和普通的分布式程序都實現(xiàn)一下。
            實際上CUDA主要是為Workstation提供更加強勁的計算能力,至于游戲嘛,按照目前的顯卡速度,對于新的游戲,主要還是用于應付渲染了,想有足夠多余的資源參與常規(guī)運算,還不太現(xiàn)實。  回復  更多評論
              
            # re: Pure GPU Computing Platform : NVIDIA CUDA Tutorial
            2007-02-26 19:46 | Jedimaster
            @空明流轉(zhuǎn)
            前幾天去了下德國BOINC論壇,把這個想法和德國人交流了一下,引用了一篇回復。

            Zitat:
            Zitat von Jedimaster Beitrag anzeigen
            NVIDIA has released the CUDA, a parrallel library use NVIDIA GPU.
            I also got interested in it but unfortunately it seems to work solely with GeForce 8 series.

            Zitat:
            if we can supply client program use GPU & CPU, maybe we can highly improve our speed.
            Sure we could but when some projects released their applications as open source some people started to recompile them optimized (making use of MMX, SSE, all kinds of technologies the original binaries still lack). This had mainly 2 effects:

            1. People started using "optimized" core-clients (manipulated to demand a multiple of credits since they are calculated by CPU time which has been decreased by optimizations). From my point of view these people just did not understand the credit system although demanding more credits for completing 2 WUs in the time of one unoptimized may seem reasonable.

            2. and more dramatic: Some projects noticed a large discrepancy in the returned results. I think it was Einstein@Home that first asked their users not to use optimized clients. That caused problems at validation when erroneous results should have been sorted out. I don't know what the accuracy of GPU-based calculations is.

            Zitat:
            Sorry for my poor German, entschuldigung.
            Shouldn't matter too much for most users here.


            Kurz nochmal auf Deutsch: Jedimaster schlägt vor, daß man die kürzlich von NVIDIA freigegebene Bibliothek CUDA für GPU-basierte Berechnungen von z.B. Fouriertransformationen etc. benutzen könnte um die Anwendungen drastisch zu optimieren. Ich habe daraufhin geantwortet, daß CUDA nur mit der GeForce 8-Reihe zusammenarbeitet und derartige Berechnungen wie einst die MMX/SSE-Optimierungen Einbrüche bei der Genauigkeit zur Folge haben könnten, die zu denselben Problemen wie schon bei Einstein im Validator-Prozess führen könnten. Darüberhinaus würden wieder mehr "optimierte" Core-Clients zum Einsatz kommen.

            德國人認為把CUDA等等技術(shù)用于分布式計算,在數(shù)值的有效性上還欠妥,甚至連專門為CPU優(yōu)化的程序都不推薦實用。結(jié)合目前G8系列還沒有上市,而且相關(guān)的中間件還不夠成熟,可能開發(fā)相應的程序還不是很現(xiàn)實。

            為了應付游戲還是有些浪費,這邊物理卡剛出來,Havok就又來湊熱鬧了,SM2人們用的正順手,SM3早就到來了。還是把優(yōu)化工作做好先。  回復  更多評論
              
            # re: Pure GPU Computing Platform : NVIDIA CUDA Tutorial
            2007-08-15 20:02 | sdfsdfsdf
            would be nice to read your site in english..  回復  更多評論
              
            # re: Pure GPU Computing Platform : NVIDIA CUDA Tutorial
            2007-08-23 15:18 | Dimitris
            @Jedimaster
            As long as accuracy is concerned, cuda complies with the error margin specifications (except for doubles i believe). There is are alot of SDK examples on their site with reference execution comparison, only once it failed to pass but the accuracy threshold was pretty tight! I'm using it now on computational physics thesis.

            For starters i just ported a simple random number generator to the gpu, the results were identical to the digits i cared about (6th crucial). Mind though that i had to convert it to float!
            Double should be working on CUDA 1.0 i think, but they didnt work for me, it wouldnt assign the value at all when double.
            I didnt even bother parallelizing it, just a 1x1 block and It was WAY OVER a magnitude faster. I'm impressed. All that with a 8600GTS.  回復  更多評論
              
            <2006年10月>
            24252627282930
            1234567
            891011121314
            15161718192021
            22232425262728
            2930311234

            周波 87年出生 南京林業(yè)大學05421班242信箱 專業(yè)木材科學與工程工業(yè)裝備與過程自動化 遷移到 jedimaster(dot)cnblogs(dot)com

            常用鏈接

            留言簿(4)

            隨筆分類

            隨筆檔案

            新聞檔案

            同學們Blog

            搜索

            •  

            積分與排名

            • 積分 - 54187
            • 排名 - 421

            最新評論

            閱讀排行榜

            亚洲中文字幕久久精品无码APP | 久久精品国产清高在天天线| 国产成人无码精品久久久久免费 | 亚洲国产精品久久久久| 亚洲av成人无码久久精品| 中文字幕久久亚洲一区| 久久久久久毛片免费看| 精品欧美一区二区三区久久久| 久久综合香蕉国产蜜臀AV| 亚洲国产精品无码久久一线| 人人妻久久人人澡人人爽人人精品 | 国产婷婷成人久久Av免费高清| 亚洲精品蜜桃久久久久久| 亚洲国产另类久久久精品黑人| 久久综合给合久久国产免费| 精品熟女少妇av免费久久| 国产成人久久激情91| 欧美精品一区二区精品久久| 青青热久久综合网伊人| 久久99久久无码毛片一区二区| 久久本道综合久久伊人| 日韩影院久久| 亚洲欧洲日产国码无码久久99| www久久久天天com| 国产综合免费精品久久久| 色婷婷久久综合中文久久一本 | 久久精品国产亚洲AV香蕉| 色狠狠久久AV五月综合| 亚洲国产精品久久久久久| 亚洲日本va午夜中文字幕久久 | 国产精品99久久久久久人| 99久久精品免费看国产| 狠狠色丁香婷婷久久综合| 精品免费久久久久久久| 国产精品免费看久久久香蕉| 99久久做夜夜爱天天做精品| 91视频国产91久久久| 亚洲Av无码国产情品久久| 狠狠色丁香久久婷婷综合五月| 国产精品无码久久综合网| 三上悠亚久久精品|