接这篇文章 OpenGL深入探索——像素缓冲区对象 （PBO）（附完整工程代码地址）
原理示意图：
首选检查显卡是否支持 PBO :
#if defined(_WIN32) // check PBO is supported by your video card // 检查显卡是否支持 PBO if (glInfo.isExtensionSupported("GL_ARB_pixel_buffer_object")) { // get pointers to GL functions glGenBuffersARB = (PFNGLGENBUFFERSARBPROC)wglGetProcAddress("glGenBuffersARB"); glBindBufferARB = (PFNGLBINDBUFFERARBPROC)wglGetProcAddress("glBindBufferARB"); glBufferDataARB = (PFNGLBUFFERDATAARBPROC)wglGetProcAddress("glBufferDataARB"); glBufferSubDataARB = (PFNGLBUFFERSUBDATAARBPROC)wglGetProcAddress("glBufferSubDataARB"); glDeleteBuffersARB = (PFNGLDELETEBUFFERSARBPROC)wglGetProcAddress("glDeleteBuffersARB"); glGetBufferParameterivARB = (PFNGLGETBUFFERPARAMETERIVARBPROC)wglGetProcAddress("glGetBufferParameterivARB"); glMapBufferARB = (PFNGLMAPBUFFERARBPROC)wglGetProcAddress("glMapBufferARB"); glUnmapBufferARB = (PFNGLUNMAPBUFFERARBPROC)wglGetProcAddress("glUnmapBufferARB"); // check once again PBO extension if (glGenBuffersARB && glBindBufferARB && glBufferDataARB && glBufferSubDataARB && glMapBufferARB && glUnmapBufferARB && glDeleteBuffersARB && glGetBufferParameterivARB) { pboSupported = true; pboMode = 1; // using 1 PBO cout << "Video card supports GL_ARB_pixel_buffer_object." << endl; } else { pboSupported = false; pboMode = 0; // without PBO cout << "Video card does NOT support GL_ARB_pixel_buffer_object." << endl; } } // Query the system memory page size and update the default value SYSTEM_INFO si; GetSystemInfo(&si); if (si.dwPageSize > 0) { systemPageSize = si.dwPageSize; } #elif defined (__gnu_linux__) // for linux, do not need to get function pointers, it is up-to-date if (glInfo.isExtensionSupported("GL_ARB_pixel_buffer_object")) { pboSupported = true; cout << "Video card supports GL_ARB_pixel_buffer_object" << endl; } else { cout << "Video card does NOT support GL_ARB_pixel_buffer_object" << endl; } if (glInfo.isExtensionSupported("GL_AMD_pinned_memory")) { amdSupported = true; cout << "Video card supports GL_AMD_pinned_memory" << endl; } else { cout << "Video card does NOT support GL_AMD_pinned_memory" << endl; } // Query the system memory page size and update the default value if (sysconf(_SC_PAGE_SIZE) > 0) { systemPageSize = sysconf(_SC_PAGE_SIZE); } #endif
创建PBO / 调整 PBO个数 的代码:
// 创建 count 个 PBO void setPboCount(int count) { if (!pboSupported) return; // 如果 count 大于 当前的 PBO 数 if (count > pboCount) { if (pboMethod != AMD) { // Generate each Pixel Buffer object and allocate memory for it（生成 PBO，并为其分配内存） // Hopefully, PBOs will get allocated in VRAM glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); // Unbind any buffer object previously bound（释放之前所绑定的 PBO） for (int i = pboCount; i < count; ++i) { GLuint pboId; glGenBuffers(1, &pboId); // Generate new Buffer Object ID glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pboId); // Create a zero-sized memory Pixel Buffer Object and bind it glBufferData(GL_PIXEL_UNPACK_BUFFER, DATA_SIZE, NULL, GL_STREAM_DRAW); // Reserve the memory space for the PBO glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); // Release the PBO binding pboIds.push_back(pboId); // Update our list of PBO IDs pboFences.push_back(NULL); cout << "Created PBO buffer #" << i << " of size: " << DATA_SIZE << endl; } pboCount = pboIds.size(); assert(GL_NO_ERROR == glGetError()); } // 特殊的 DMA 模式，需要自己手动分配对齐的内存，并提供内存指针 else { // Generate each Pixel Buffer object and allocate memory for it // PBOs will get allocated in System RAM, and GPU will access it through DMA glBindBuffer(GL_EXTERNAL_VIRTUAL_MEMORY_BUFFER_AMD, 0); // Unbind any buffer object previously bound for (int i = pboCount; i < count; ++i) { GLuint pboId; glGenBuffers(1, &pboId); // Generate new Buffer Object ID glBindBuffer(GL_EXTERNAL_VIRTUAL_MEMORY_BUFFER_AMD, pboId); // Create a zero-sized memory Pixel Buffer Object and bind it assert(GL_NO_ERROR == glGetError()); // Memory alignment functions are compiler-specific GLubyte *ptAlignedBuffer = (GLubyte *)alignedMalloc(systemPageSize, DATA_SIZE); if (NULL == ptAlignedBuffer) { cout << "ERROR [setPboCount] (alignedMalloc) size: " << DATA_SIZE << " alignment: " << systemPageSize << endl; break; } cout << "Created memory buffer #" << i << " of size: " << DATA_SIZE << " alignment: " << systemPageSize << endl; glBufferData(GL_EXTERNAL_VIRTUAL_MEMORY_BUFFER_AMD, DATA_SIZE, ptAlignedBuffer, GL_STREAM_DRAW); // Take control of the memory space for the PBO GLenum error = glGetError(); if (GL_NO_ERROR != error) { cout << "ERROR [setPboCount] (glBufferData): " << (char *)gluErrorString(error) << endl; alignedFree(ptAlignedBuffer); cout << "Freed memory buffer #" << i << endl; break; } glBindBuffer(GL_EXTERNAL_VIRTUAL_MEMORY_BUFFER_AMD, 0); // Release the PBO binding assert(GL_NO_ERROR == glGetError()); pboIds.push_back(pboId); // Update our list of PBO IDs pboFences.push_back(NULL); alignedBuffers.push_back((GLubyte *)ptAlignedBuffer); cout << "Created PBO buffer #" << i << endl; } pboCount = pboIds.size(); assert(GL_NO_ERROR == glGetError()); } } // 如果 count 小于当前的 PBO 数 else if (count < pboCount) { if (pboMethod != AMD) { glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); // Unbind any buffer object previously bound for (int i = pboCount - 1; i >= count; --i) { glDeleteSync(pboFences.back()); pboFences.pop_back(); GLuint pboId = pboIds.back(); glDeleteBuffers(1, &pboId); pboIds.pop_back(); // Update our list of PBO IDs cout << "Deleted PBO buffer #" << i << endl; } pboCount = pboIds.size(); assert(GL_NO_ERROR == glGetError()); } else { glBindBuffer(GL_EXTERNAL_VIRTUAL_MEMORY_BUFFER_AMD, 0); // Unbind any buffer object previously bound for (int i = pboCount - 1; i >= count; --i) { glDeleteSync(pboFences.back()); pboFences.pop_back(); GLuint pboId = pboIds.back(); glDeleteBuffers(1, &pboId); pboIds.pop_back(); // Update our list of PBO IDs cout << "Deleted PBO buffer #" << i << endl; // 手动释放自己分配的内存 alignedFree(alignedBuffers.back()); alignedBuffers.pop_back(); cout << "Freed memory buffer #" << i << endl; } pboCount = pboIds.size(); assert(GL_NO_ERROR == glGetError()); } } cout << "PBO Count: " << pboCount << endl; }
最关键的显示回调方法：
void displayCB() { if (pboMethod == NONE) { /* * Update data in System Memory. */ t1.start(); updatePixels(imageData, DATA_SIZE); // 更新 imageData 的像素 t1.stop(); updateTime = t1.getElapsedTimeInMilliSec(); /* * Copy data from System Memory to texture object. （将 imageData 的内容从内存拷贝到纹理当中） */ t1.start(); glBindTexture(GL_TEXTURE_2D, textureId); glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, IMAGE_WIDTH, IMAGE_HEIGHT, PIXEL_FORMAT, GL_UNSIGNED_BYTE, (GLvoid *)imageData); t1.stop(); copyTime = t1.getElapsedTimeInMilliSec(); } else { /* * Update buffer indices used in data upload & copy. * * "uploadIdx": index used to upload pixels to a Pixel Buffer Object.（上传像素至 uploadIdx 指定的 PBO） * "copyIdx": index used to copy pixels from a Pixel Buffer Object to a GPU texture.（拷贝 cpyIdx 指定的 PBO 的像素到 GPU 纹理） * * When (pboCount > 1), this will allow to perform（当 pboCount 数大于1时，就允许使用备用buffer来同时进行上传和拷贝） * simultaneous upload & copy, by using alternative buffers. * That is a good thing, unless the double buffering is being already * done somewhere else in the code. */ static int copyIdx = 0; copyIdx = (copyIdx + 1) % pboCount; int uploadIdx = (copyIdx + 1) % pboCount; /* * Upload new data to a Pixel Buffer Object. */ t1.start(); glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pboIds[uploadIdx]); // Access the Pixel Buffer Object and bind it // pboMethod 表示不同的纹理流机制（Texture Stream methods） if (pboMethod == ORPHAN) { // GL_STREAM_DRAW 表示每次渲染都会更新该 PBO的像素数据 // GL_DYNAMITC_DRAW 表示每帧都会更新该 Buffer // GL_STATIC_DRAW 表示几乎或从不更新该 Buffer glBufferDataARB(GL_PIXEL_UNPACK_BUFFER_ARB, DATA_SIZE, NULL, GL_STREAM_DRAW_ARB); // 获得 PBO 的映射 Buffer 指针，以待写入操作 GLubyte *ptr = (GLubyte *)glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_WRITE_ONLY_ARB); if (NULL == ptr) { cout << "ERROR [displayCB] (glMapBufferARB): " << (char *)gluErrorString(glGetError()) << endl; return; } else { // update data directly on the mapped buffer（在映射的 Buffer 上直接更新 PBO 的像素数据） updatePixels(ptr, DATA_SIZE); // release pointer to mapping buffer（释放映射 Buffer 的指针） if (!glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB)) { cout << "ERROR [displayCB] (glUnmapBufferARB): " << (char *)gluErrorString(glGetError()) << endl; } } } // 异步模式 else if (pboMethod == UNSYNCH_ORPHAN || pboMethod == UNSYNCH_FENCES) { if (pboMethod == UNSYNCH_FENCES) { // 检查索引的pboFences 是否是同步对象（Sync Object） if (glIsSync(pboFences[uploadIdx])) { // 阻塞并 wait 同步对象，等待被 signal GLenum result = glClientWaitSync(pboFences[uploadIdx], 0, GL_TIMEOUT_IGNORED); switch (result) { case GL_ALREADY_SIGNALED: // Transfer was already done when trying to use buffer // （说明此时传输已经完成） cout << "DEBUG (glClientWaitSync): ALREADY_SIGNALED (good timing!) uploadIdx: " << uploadIdx << endl; break; case GL_CONDITION_SATISFIED: // This means that we had to wait for the fence to synchronize us after using all the buffers, // which implies that the GPU command queue is full and that we are GPU-bound (DMA transfers aren't fast enough). // （说明我们不得不等待 fence 的同步，即 我们所绑定的 GPU 命令队列已满[DMA传输还不够快]） cout << "WARNING (glClientWaitSync): CONDITION_SATISFIED (had to wait for the sync) uploadIdx: " << uploadIdx << endl; break; case GL_TIMEOUT_EXPIRED: cout << "WARNING (glClientWaitSync): TIMEOUT_EXPIRED (DMA transfers are too slow!) uploadIdx: " << uploadIdx << endl; break; case GL_WAIT_FAILED: cout << "ERROR (glClientWaitSync): WAIT_FAILED: " << (char *)gluErrorString(glGetError()) << endl; break; } // 删除同步对象 glDeleteSync(pboFences[uploadIdx]); pboFences[uploadIdx] = NULL; } } // 注意和 ORPHAN 的区别 else if (pboMethod == UNSYNCH_ORPHAN) { // Buffer 需要重新指定 glBufferData(GL_PIXEL_UNPACK_BUFFER, DATA_SIZE, NULL, GL_STREAM_DRAW); // Buffer re-specification (orphaning) } // 获得 PBO 的映射 Buffer 指针，以待写入操作 GLubyte *ptr = (GLubyte *)glMapBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, DATA_SIZE, GL_MAP_WRITE_BIT | GL_MAP_UNSYNCHRONIZED_BIT); if (NULL == ptr) { cout << "ERROR [displayCB] (glMapBufferRange): " << (char *)gluErrorString(glGetError()) << endl; return; } else { updatePixels(ptr, DATA_SIZE); // Update data directly on the mapped buffer（直接更新映射 Buffer 的数据） if (!glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER)) { cout << "ERROR [displayCB] (glUnmapBuffer): " << (char *)gluErrorString(glGetError()) << endl; } } } else if (pboMethod == AMD) { // 同样要进行手动同步 if (glIsSync(pboFences[uploadIdx])) { GLenum result = glClientWaitSync(pboFences[uploadIdx], 0, GL_TIMEOUT_IGNORED); switch (result) { case GL_ALREADY_SIGNALED: // Transfer was already done when trying to use buffer //cout << "DEBUG (glClientWaitSync): ALREADY_SIGNALED (good timing!) uploadIdx: " << uploadIdx << endl; break; case GL_CONDITION_SATISFIED: // This means that we had to wait for the fence to synchronize us after using all the buffers, // which implies that the GPU command queue is full and that we are GPU-bound (DMA transfers aren't fast enough). //cout << "WARNING (glClientWaitSync): CONDITION_SATISFIED (had to wait for the sync) uploadIdx: " << uploadIdx << endl; break; case GL_TIMEOUT_EXPIRED: cout << "WARNING (glClientWaitSync): TIMEOUT_EXPIRED (DMA transfers are too slow!) uploadIdx: " << uploadIdx << endl; break; case GL_WAIT_FAILED: cout << "ERROR (glClientWaitSync): WAIT_FAILED: " << (char *)gluErrorString(glGetError()) << endl; break; } glDeleteSync(pboFences[uploadIdx]); pboFences[uploadIdx] = NULL; } // alignedBuffers 存储的是手动分配的一块对齐过的内存指针 updatePixels(alignedBuffers[uploadIdx], DATA_SIZE); // Update data directly on the mapped buffer } t1.stop(); updateTime = t1.getElapsedTimeInMilliSec(); /* * Protect each Pixel Buffer Object against being overwritten.（防止 PBO 被重复写入） * * Tipically the data upload will be slower than our main loop, so this * function will be called again before the previous frame was uploaded * and processed. The main bottleneck is the PCI bus transfer speed, * which limits how fast the DMA (System Memory --> VRAM) can work. * * 通常数据上传将会慢于主循环，意味着在先前的帧被上传处理之前，该方法将会被再次调用。 * 主要的性能瓶颈在于 PCI 总线的传输速度， 它限制了 DMA 的传输速度（内存到显存）。 * * OpenGL Sync Fences will block until the PBO is released.（GL 同步对象将会阻塞主线程，直到 PBO 被释放为止） */ if (pboMethod == UNSYNCH_FENCES || pboMethod == AMD) { // 创建一个同步对象，并将其加入 GL 的命令流中（❤ 具体请参看第八版红宝书的 P589 第11章 Memory ❤） pboFences[uploadIdx] = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0); } /* * Copy data from a Pixel Buffer Object to a GPU texture. * glTexSubImage2D() will copy pixels to the corresponding texture in the GPU. * 传输 PBO 中的数据到 GPU 纹理 */ t1.start(); glBindTexture(GL_TEXTURE_2D, textureId); // Bind the texture glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, pboIds[copyIdx]); // Access the Pixel Buffer Object and bind it // Use offset instead of pointer（由于使用了 PBO，所以传递的是偏移量） glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, IMAGE_WIDTH, IMAGE_HEIGHT, PIXEL_FORMAT, GL_UNSIGNED_BYTE, 0); t1.stop(); copyTime = t1.getElapsedTimeInMilliSec(); // it is good idea to release PBOs with ID 0 after use. // Once bound with 0, all pixel operations behave normal ways. glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0); } // clear buffer glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT); // save the initial ModelView matrix before modifying ModelView matrix glPushMatrix(); // tramsform camera glTranslatef(0, 0, -cameraDistance); glRotatef(cameraAngleX, 1, 0, 0); // pitch glRotatef(cameraAngleY, 0, 1, 0); // heading // draw a point with texture glBindTexture(GL_TEXTURE_2D, textureId); glColor4f(1, 1, 1, 1); glBegin(GL_QUADS); glNormal3f(0, 0, 1); glTexCoord2f(0.0f, 0.0f); glVertex3f(-1.0f, -1.0f, 0.0f); glTexCoord2f(1.0f, 0.0f); glVertex3f(1.0f, -1.0f, 0.0f); glTexCoord2f(1.0f, 1.0f); glVertex3f(1.0f, 1.0f, 0.0f); glTexCoord2f(0.0f, 1.0f); glVertex3f(-1.0f, 1.0f, 0.0f); glEnd(); // unbind texture glBindTexture(GL_TEXTURE_2D, 0); // draw info messages showInfo(); //showTransferRate(); printTransferRate(); glPopMatrix(); glutSwapBuffers(); }
运行结果对比：
【未使用 PBO 的情况】
【Orphaning模式，PBO 个数 = 1】
【异步 Orphaning模式， PBO个数 = 1】
【Fences 的 同步模式， PBO个数 = 1】
【Orphaning模式， PBO个数=3】
【异步的 Orphaning 模式， PBO个数=3】
可见使用了 PBO 的确比未使用 PBO 性能要略优，但一味增加 PBO 的数量，并不能显著提高性能。
本例中 不同的 PBO 模式性能差别不大，但是 Orphan 模式的写法最简单，不需要自己手动同步和创建额外对齐的内存。
