VBOs slower than obsolete method of drawing primitives - why?
Pio*_*cki 7 c++ opengl graphics performance vbo
I'm working on a Tile-based OpenGL, C++ application. I'm adding sample screen from application, so that it will be more clear:
I have Tile class which contains an array of Objects. Each tile can store up to 15 objects - the example of that is Tile with green and yellow square on it (two objects), so it's 10x10x15 = 1500 Objects to draw (in the worst case, because I'm not handling 'empty' ones). Usually it's less, in my testings I use around 600 of them. Object has it's own graphic, that can be drawn. Each Object belongs to one Tile at a time, but it can be moved (as for example red squares in the picture).
Objects backgrounds are going to have a border and they need to be nicely scalable, so I'm using 9-patch pattern to draw them (they are made of 9 quads).
Without drawing Tiles (their Objects to be precise), my application has around 600 fps.
At first, I've been using obsolete method to draw those Tiles - using glBegin(GL_QUADS)/glEnd() and glDisplayLists. I had a big drop of performance due to that drawing - from 600 to 320 fps. This is how I've been drawing them:
bool Background::draw(const TPoint& pos, int width, int height)
{
if(width <= 0 || height <= 0)
return false;
//glFrontFace(GL_CW);
glPushMatrix();
glTranslatef((GLfloat)pos.x, (GLfloat)pos.y, 0.0f); // Move background to right direction
if((width != m_savedWidth) || (height != m_savedHeight)) // If size to draw is different than the one saved in display list,
// then recalculate everything and save in display list
{
// That size will be now saved in display list
m_savedWidth = width;
m_savedHeight = height;
// If this background doesn't have unique display list id specified yet,
// then let OpenGL generate one
if(m_displayListId == NO_DISPLAY_LIST_ID)
{
GLuint displayList;
displayList = glGenLists(1);
m_displayListId = displayList;
}
glNewList(m_displayListId, GL_COMPILE);
GLfloat texelCentersOffsetX = (GLfloat)1/(2*m_width);
// Instead of coordinates range 0..1 we need to specify new ones
GLfloat maxTexCoordWidth = m_bTiling ? (GLfloat)width/m_width : 1.0;
GLfloat maxTexCoordHeight = m_bTiling ? (GLfloat)height/m_height : 1.0;
GLfloat maxTexCoordBorderX = (GLfloat)m_borderWidth/m_width;
GLfloat maxTexCoordBorderY = (GLfloat)m_borderWidth/m_height;
/* 9-cell-pattern
-------------------
| 1 | 2 | 3 |
-------------------
| | | |
| 4 | 9 | 5 |
| | | |
-------------------
| 6 | 7 | 8 |
-------------------
*/
glBindTexture(GL_TEXTURE_2D, m_texture); // Select Our Texture
// Top left quad [1]
glBegin(GL_QUADS);
// Bottom left
glTexCoord2f(0.0, maxTexCoordBorderY);
glVertex2i(0, 0 + m_borderWidth);
// Top left
glTexCoord2f(0.0, 0.0);
glVertex2i(0, 0);
// Top right
glTexCoord2f(maxTexCoordBorderX, 0.0);
glVertex2i(0 + m_borderWidth, 0);
// Bottom right
glTexCoord2f(maxTexCoordBorderX, maxTexCoordBorderY);
glVertex2i(0 + m_borderWidth, 0 + m_borderWidth);
glEnd();
// Top middle quad [2]
glBegin(GL_QUADS);
// Bottom left
glTexCoord2f(maxTexCoordBorderX + texelCentersOffsetX, maxTexCoordBorderY);
glVertex2i(0 + m_borderWidth, 0 + m_borderWidth);
// Top left
glTexCoord2f(maxTexCoordBorderX + texelCentersOffsetX, 0.0);
glVertex2i(0 + m_borderWidth, 0);
// Top right
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX - texelCentersOffsetX, 0.0);
glVertex2i(0 + width - m_borderWidth, 0);
// Bottom right
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX - texelCentersOffsetX, maxTexCoordBorderY);
glVertex2i(0 + width - m_borderWidth, 0 + m_borderWidth);
glEnd();
// Top right quad [3]
glBegin(GL_QUADS);
// Bottom left
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX, maxTexCoordBorderY);
glVertex2i(0 + width - m_borderWidth, 0 + m_borderWidth);
// Top left
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX, 0.0);
glVertex2i(0 + width - m_borderWidth, 0);
// Top right
glTexCoord2f(1.0, 0.0);
glVertex2i(0 + width, 0);
// Bottom right
glTexCoord2f(1.0, maxTexCoordBorderY);
glVertex2i(0 + width, 0 + m_borderWidth);
glEnd();
// Middle left quad [4]
glBegin(GL_QUADS);
// Bottom left
glTexCoord2f(0.0, (GLfloat)1.0 - maxTexCoordBorderY );
glVertex2i(0, 0 + height - m_borderWidth);
// Top left
glTexCoord2f(0.0, maxTexCoordBorderY );
glVertex2i(0, 0 + m_borderWidth);
// Top right
glTexCoord2f(maxTexCoordBorderX, maxTexCoordBorderY );
glVertex2i(0 + m_borderWidth, 0 + m_borderWidth);
// Bottom right
glTexCoord2f(maxTexCoordBorderX, (GLfloat)1.0 - maxTexCoordBorderY );
glVertex2i(0 + m_borderWidth, 0 + height - m_borderWidth);
glEnd();
// Middle right quad [5]
glBegin(GL_QUADS);
// Bottom left
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX, (GLfloat)1.0 - maxTexCoordBorderY);
glVertex2i(0 + width - m_borderWidth, 0 + height - m_borderWidth);
// Top left
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX, maxTexCoordBorderY);
glVertex2i(0 + width - m_borderWidth, 0 + m_borderWidth);
// Top right
glTexCoord2f(1.0, maxTexCoordBorderY);
glVertex2i(0 + width, 0 + m_borderWidth);
// Bottom right
glTexCoord2f(1.0, (GLfloat)1.0 - maxTexCoordBorderY);
glVertex2i(0 + width, 0 + height - m_borderWidth);
glEnd();
// Bottom left quad [6]
glBegin(GL_QUADS);
// Bottom left
glTexCoord2f(0.0f, 1.0);
glVertex2i(0, 0 + height);
// Top left
glTexCoord2f(0.0f, (GLfloat)1.0 - maxTexCoordBorderY);
glVertex2i(0, 0 + height - m_borderWidth);
// Top right
glTexCoord2f(maxTexCoordBorderX, (GLfloat)1.0 - maxTexCoordBorderY);
glVertex2i(0 + m_borderWidth, 0 + height - m_borderWidth);
// Bottom right
glTexCoord2f(maxTexCoordBorderX, 1.0);
glVertex2i(0 + m_borderWidth, 0 + height);
glEnd();
// Bottom middle quad [7]
glBegin(GL_QUADS);
// Bottom left
glTexCoord2f(maxTexCoordBorderX + texelCentersOffsetX, 1.0);
glVertex2i(0 + m_borderWidth, 0 + height);
// Top left
glTexCoord2f(maxTexCoordBorderX + texelCentersOffsetX, (GLfloat)1.0 - maxTexCoordBorderY);
glVertex2i(0 + m_borderWidth, 0 + height - m_borderWidth);
// Top right
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX - texelCentersOffsetX, (GLfloat)1.0 - maxTexCoordBorderY);
glVertex2i(0 + width - m_borderWidth, 0 + height - m_borderWidth);
// Bottom right
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX - texelCentersOffsetX, 1.0);
glVertex2i(0 + width - m_borderWidth, 0 + height);
glEnd();
// Bottom right quad [8]
glBegin(GL_QUADS);
// Bottom left
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX, 1.0);
glVertex2i(0 + width - m_borderWidth, 0 + height);
// Top left
glTexCoord2f((GLfloat)1.0 - maxTexCoordBorderX, (GLfloat)1.0 - maxTexCoordBorderY);
glVertex2i(0 + width - m_borderWidth, 0 + height - m_borderWidth);
// Top right
glTexCoord2f(1.0, (GLfloat)1.0 - maxTexCoordBorderY);
glVertex2i(0 + width, 0 + height - m_borderWidth);
// Bottom right
glTexCoord2f(1.0, 1.0);
glVertex2i(0 + width, 0 + height);
glEnd();
GLfloat xTexOffset;
GLfloat yTexOffset;
if(m_borderWidth > 0)
{
glBindTexture(GL_TEXTURE_2D, m_centerTexture); // If there's a border, we have to use
// second texture now for middle quad
xTexOffset = 0.0; // We are using another texture, so middle middle quad
yTexOffset = 0.0; // has to be texture with a whole texture
}
else
{
// Don't bind any texture here - we're still using the same one
xTexOffset = maxTexCoordBorderX; // But it implies using offset which equals
yTexOffset = maxTexCoordBorderY; // maximum texture coordinates
}
// Middle middle quad [9]
glBegin(GL_QUADS);
// Bottom left
glTexCoord2f(xTexOffset, maxTexCoordHeight - yTexOffset);
glVertex2i(0 + m_borderWidth, 0 + height - m_borderWidth);
// Top left
glTexCoord2f(xTexOffset, yTexOffset);
glVertex2i(0 + m_borderWidth, 0 + m_borderWidth);
// Top right
glTexCoord2f(maxTexCoordWidth - xTexOffset, yTexOffset);
glVertex2i(0 + width - m_borderWidth, 0 + m_borderWidth);
// Bottom right
glTexCoord2f(maxTexCoordWidth - xTexOffset, maxTexCoordHeight - yTexOffset);
glVertex2i(0 + width - m_borderWidth, 0 + height - m_borderWidth);
glEnd();
glEndList();
}
glCallList(m_displayListId); // Now we can call earlier or now created display list
glPopMatrix();
return true;
}
There is probably too much of code there, but I wanted to show everything. The main thing about this version is use of display lists and glVertex2i which are deprecated.
I thought the problem of such slow down was use of this obsolete method which I read is quite slow, so I decided to go for VBO. I've used this tutorial and according to it I changed my method like this:
bool Background::draw(const TPoint& pos, int width, int height)
{
if(width <= 0 || height <= 0)
return false;
glPushMatrix();
glTranslatef((GLfloat)pos.x, (GLfloat)pos.y, 0.0f); // Move background to right direction
if((width != m_savedWidth) || (height != m_savedHeight)) // If size to draw is different than the one saved in display list,
// then recalculate everything and save in display list
{
// That size will be now saved in display list
m_savedWidth = width;
m_savedHeight = height;
GLfloat texelCentersOffsetX = (GLfloat)1/(2*m_width);
// Instead of coordinates range 0..1 we need to specify new ones
GLfloat maxTexCoordWidth = m_bTiling ? (GLfloat)width/m_width : 1.0;
GLfloat maxTexCoordHeight = m_bTiling ? (GLfloat)height/m_height : 1.0;
GLfloat maxTexCoordBorderX = (GLfloat)m_borderWidth/m_width;
GLfloat maxTexCoordBorderY = (GLfloat)m_borderWidth/m_height;
/* 9-cell-pattern, each number represents one quad
-------------------
| 1 | 2 | 3 |
-------------------
| | | |
| 4 | 9 | 5 |
| | | |
-------------------
| 6 | 7 | 8 |
-------------------
*/
/* How vertices are distributed on one quad made of two triangles
v1 ------ v0
| / |
| / |
| / |
v2 ------ v3
*/
GLfloat vertices[] = {
// Top left quad [1]
m_borderWidth, 0, 0, // v0
0, 0, 0, // v1
0, m_borderWidth, 0, // v2
0, m_borderWidth, 0, // v2
m_borderWidth, m_borderWidth, 0, // v3
m_borderWidth, 0, 0, // v0
// Top middle quad [2]
width-m_borderWidth, 0, 0, // v0
m_borderWidth, 0, 0, // v1
m_borderWidth, m_borderWidth, 0, // v2
m_borderWidth, m_borderWidth, 0, // v2
width-m_borderWidth, m_borderWidth, 0, // v3
width-m_borderWidth, 0, 0, // v0
// Top right quad [3]
width, 0, 0, // v0
width-m_borderWidth, 0, 0, // v1
width-m_borderWidth, m_borderWidth, 0, // v2
width-m_borderWidth, m_borderWidth, 0, // v2
width, m_borderWidth, 0, // v3
width, 0, 0, // v0
// Middle left quad [4]
m_borderWidth, m_borderWidth, 0, // v0
0, m_borderWidth, 0, // v1
0, height-m_borderWidth, 0, // v2
0, height-m_borderWidth, 0, // v2
m_borderWidth, height-m_borderWidth, 0, // v3
m_borderWidth, m_borderWidth, 0, // v0
// Middle right quad [5]
width, m_borderWidth, 0, // v0
width-m_borderWidth, m_borderWidth, 0, // v1
width-m_borderWidth, height-m_borderWidth, 0, // v2
width-m_borderWidth, height-m_borderWidth, 0, // v2
width, height-m_borderWidth, 0, // v3
width, m_borderWidth, 0, // v0
// Bottom left quad [6]
m_borderWidth, height-m_borderWidth, 0, // v0
0, height-m_borderWidth, 0, // v1
0, height, 0, // v2
0, height, 0, // v2
m_borderWidth, height, 0, // v3
m_borderWidth, height-m_borderWidth, 0, // v0
// Bottom middle quad [7]
width-m_borderWidth, height-m_borderWidth, 0, // v0
m_borderWidth, height-m_borderWidth, 0, // v1
m_borderWidth, height, 0, // v2
m_borderWidth, height, 0, // v2
width-m_borderWidth, height, 0, // v3
width-m_borderWidth, height-m_borderWidth, 0, // v0
// Bottom right quad [8]
width, height-m_borderWidth, 0, // v0
width-m_borderWidth, height-m_borderWidth, 0, // v1
width-m_borderWidth, height, 0, // v2
width-m_borderWidth, height, 0, // v2
width, height, 0, // v3
width, height-m_borderWidth, 0, // v0
// Middle middle quad [9]
width-m_borderWidth, m_borderWidth, 0, // v0
m_borderWidth, m_borderWidth, 0, // v1
m_borderWidth, height-m_borderWidth, 0, // v2
m_borderWidth, height-m_borderWidth, 0, // v2
width-m_borderWidth, height-m_borderWidth, 0, // v3
width-m_borderWidth, m_borderWidth, 0 // v0
};
copy(vertices, vertices + 162, m_vCoords); // 162, because we have 162 coordinates
int dataSize = 162 * sizeof(GLfloat);
m_vboId = createVBO(m_vCoords, dataSize);
}
// bind VBOs for vertex array
glBindBufferARB(GL_ARRAY_BUFFER_ARB, m_vboId); // for vertex coordinates
glEnableClientState(GL_VERTEX_ARRAY); // activate vertex coords array
glVertexPointer(3, GL_FLOAT, 0, 0);
glDrawArrays(GL_TRIANGLES, 0, 162);
glDisableClientState(GL_VERTEX_ARRAY); // deactivate vertex array
// bind with 0, so, switch back to normal pointer operation
glBindBufferARB(GL_ARRAY_BUFFER_ARB, NO_VBO_ID);
glPopMatrix();
return true;
}
It is quite similar to previous version, but instead of glDisplayList and glVertex2i() I used VBO which is being created from data stored in an array.
But results disappointed me, because I got performance drop instead of boost, I got barely
~260 fps and I must note that in this method version I haven't yet implemented use of textures, so there are only quads for now without any texture bound to it.
I've read a few articles to find what could be the reason of such slow down and found out that maybe it is due to big amount of small VBOs and I should probably have one VBO containing all backgrounds data instead of separate VBO for each background. But the problem is that Objects can move around and they have different textures (and texture atlas is not a good solution for me), so it would be difficult for me to update those changes for those Objects that changed their state. For now, when Objects is being changed, I just recreate it's VBO and the rest VBOs stay untouched.
所以我的问题是 - 我做错了什么？使用较大(~600)小VBOs的数量是否真的比过时的绘图方法慢glVertex2i？在我的情况下,可能是什么 - 也许不是最好但更好的解决方案？
从它的外观来看,你每帧都在重建VBO.如果你只是想改变数据的使用glBufferSubData,glBufferData那就要经历整个冗长的VBO初始化.
如果数据是静态的,只创建一次VBO,然后重复使用它.
- 请注意,如果您希望VBO的数据发生很大变化,则应使用GL_DYNAMIC_DRAW作为VBO的数据类型. (3认同)
仅仅因为固定功能的东西已经过时、不推荐使用并且通常不推荐，并不一定意味着它总是很慢。
缓冲区和着色器等奇特的“新”功能（它已经存在了一段时间）也并不一定意味着一切都会快如闪电。
当您将绘图包装在显示列表中时，您基本上将一系列操作传递给驱动程序。这实际上为驾驶员提供了相当大的空间来优化正在发生的事情。它可以很好地将您正在做的大部分事情打包到一个非常高效的预打包 GPU 操作块中。这很可能比将数据打包到缓冲区并将其发送出去时的效率稍高一些。
这并不是说我会建议坚持使用旧式界面，但对于它在某些情况下做得很好，我当然并不感到惊讶。
| 归档时间： |
|
| 查看次数： |
812 次 |
| 最近记录： |