Why is OpenGL immediate mode faster than core? - c++

Question

I am using the following library to render text in OpenGL: fontstash. I have another header file which adds support for OpenGL 3.0+. The question is why is the render implementation with core profile much slower than the immediate mode?
Here is the render code with immediate mode:
static void glfons__renderDraw(void* userPtr, const float* verts, const float* tcoords, const unsigned int* colors, int nverts)
{
GLFONScontext* gl = (GLFONScontext*)userPtr;
if (gl->tex == 0) return;
glBindTexture(GL_TEXTURE_2D, gl->tex);
glEnable(GL_TEXTURE_2D);
glEnableClientState(GL_VERTEX_ARRAY);
glEnableClientState(GL_TEXTURE_COORD_ARRAY);
glEnableClientState(GL_COLOR_ARRAY);
glVertexPointer(2, GL_FLOAT, sizeof(float)*2, verts);
glTexCoordPointer(2, GL_FLOAT, sizeof(float)*2, tcoords);
glColorPointer(4, GL_UNSIGNED_BYTE, sizeof(unsigned int), colors);
glDrawArrays(GL_TRIANGLES, 0, nverts);
glDisable(GL_TEXTURE_2D);
glDisableClientState(GL_VERTEX_ARRAY);
glDisableClientState(GL_TEXTURE_COORD_ARRAY);
glDisableClientState(GL_COLOR_ARRAY);
}
Here is the core profile render code:
static void gl3fons__renderDraw(void* userPtr, const float* verts, const float* tcoords, const unsigned int* colors, int nverts)
{
GLFONScontext* gl = (GLFONScontext*)userPtr;
if (gl->tex == 0) return;
if (gl->shader == 0) return;
if (gl->vao == 0) return;
if (gl->vbo == 0) return;
// init shader
glUseProgram(gl->shader);
// init texture
glActiveTexture(GL_TEXTURE0);
glBindTexture(GL_TEXTURE_2D, gl->tex);
glUniform1i(gl->texture_uniform, 0);
// init our projection matrix
glUniformMatrix4fv(gl->projMat_uniform, 1, false, gl->projMat);
// bind our vao
glBindVertexArray(gl->vao);
// setup our buffer
glBindBuffer(GL_ARRAY_BUFFER, gl->vbo);
glBufferData(GL_ARRAY_BUFFER, (2 * sizeof(float) * 2 * nverts) + (sizeof(int) * nverts), NULL, GL_DYNAMIC_DRAW);
glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(float) * 2 * nverts, verts);
glBufferSubData(GL_ARRAY_BUFFER, sizeof(float) * 2 * nverts, sizeof(float) * 2 * nverts, tcoords);
glBufferSubData(GL_ARRAY_BUFFER, 2 * sizeof(float) * 2 * nverts, sizeof(int) * nverts, colors);
// setup our attributes
glEnableVertexAttribArray(0);
glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 0, 0);
glEnableVertexAttribArray(1);
glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(float) * 2, (void *) (sizeof(float) * 2 * nverts));
glEnableVertexAttribArray(2);
glVertexAttribPointer(2, 4, GL_UNSIGNED_BYTE, GL_TRUE, sizeof(int), (void *) (2 * sizeof(float) * 2 * nverts));
glDrawArrays(GL_TRIANGLES, 0, nverts);
glBindBuffer(GL_ARRAY_BUFFER, 0);
glBindVertexArray(0);
glUseProgram(0);
}
I made a small test for each implementation and the results show that immediate mode is significantly faster than core.
Both tests fill the screen with AAA... and I log the time it took to do this for each frame. This is the loop:
// Create GL stash for 512x512 texture, our coordinate system has zero at top-left.
struct FONScontext* fs = glfonsCreate(512, 512, FONS_ZERO_TOPLEFT);
// Add font to stash.
int fontNormal = fonsAddFont(fs, "sans", "fontstash/example/DroidSerif-Regular.ttf");
// Render some text
float dx = 10, dy = 10;
unsigned int white = glfonsRGBA(255,255,255,255);
std::chrono::high_resolution_clock::time_point t1 = std::chrono::high_resolution_clock::now();
fonsSetFont(fs, fontNormal);
fonsSetSize(fs, 20.0f);
fonsSetColor(fs, white);
for(int i = 0; i < 90; i++){
for( int j = 0; j < 190; j++){
dx += 10;
fonsDrawText(fs, dx, dy, "A", NULL);
}
dy += 10;
dx = 10;
}
std::chrono::high_resolution_clock::time_point t2 = std::chrono::high_resolution_clock::now();
std::chrono::duration<double, std::milli> time_span = t2 - t1;
std::cout<<"Time to render: "<<time_span.count()<<"ms"<<std::endl;
And the results show more than 400ms difference between the two:
Core profile (left) vs Immediate mode (right)
What should be changed in order to speed up performance?

Answer 1

I don't know exactly what gl is in this program, but it's pretty clear that, every time you want to render a piece of text, you perform the following operations:
Allocate storage for a buffer, reallocating whatever storage had been created from the last time this was called.
Perform three separate uploads of data to that buffer.
These are not good ways to stream vertex data to the GPU. There are specific techniques for doing this well, but this is not one of them. In particular, the fact that you are constantly reallocating the same buffer is going to kill performance.
The most effective way to deal with this is to have a single buffer with a fixed amount of storage. It gets allocated exactly once and never again. Ideally, whatever API you're getting vertex data from would provide it in an interleaved format, so that you would only need to perform one upload rather than three. But apparently, Fontstash is apparently not so generous.
In any case, the main idea is to avoid reallocation and synchronization. The latter means never trying to write over data that has been written to recently. So your buffer needs to be sufficiently large to hold twice the number of font vertices you ever expect to render. Essentially, you double-buffer the vertex data: writing to one set of data while the other set is being read from.
So at the beginning of the frame, you figure out what the byte offset to where you want to render will be. This will either be the start of the buffer or half-way through it. Then, for each blob of text, you write vertex data to this offset and increment the offset accordingly.
And to avoid having to change VAO state, you should interleave the vertex data manually. Instead of uploading three arrays, you should interleave the vertices so that you're effectively making one gigantic array of vertices. So you never need to call glVertexAttribPointer in the middle of this function; you just use the parameters to glDraw* to draw the part of the array you want.
This also means you only need one glBufferSubData call. But if you have access to persistent mapped buffers, you don't even need that, since you can just write to the memory directly while using the other portion of it. Though if you use persistent mapping, you will need to use a fence sync object when you switch buffer regions to make sure that you're not writing to vertex data that is still being read by the GPU.