Instancing with OpenGL 3.3 seems very slow - c++

Question

I wrote a minimal code-sample in C++, which is rendering 10000 colored
quads on the screen. I am using "instancing" and so updating only
the model-matrix for each quad each frame. The data of the 6 vertices
are stored in an indivdual VBO und will be reused all the time.
The projection-matrix (orthographic) is injected once at program-start
via uniform. The model-matrix is calculated on the CPU with the library GLM.
I measured the rendering-time and I got only an average FPS of 52.
I think this is MUCH to less, but I cannot find the mistake/bottleneck in my little sample program.
After some analysis it seems, that the 3 calculations done with GLM
are very slow. Am I doing something wrong here? For example, If
I remove the rotating-calculation, I get an FPS-boost of 10 FPS!
Maybe you can help me to find out, what I can do better here and how
can I optimize my sample. It is important for me, that each quad is individual configurable during runtime, so I decided to use instancing.
Moving the matrix-calculations to the GPU seems another option, but I am really confused, why the CPU has so much problems calculating the 10000
model-matrices! Ok, my CPU is very bad (Athlon 2 Core-Duo M300, GPU is ATI Mobility Radeon 4100), but It should do this task in no measurable time, or?
Here is minimal, fully working, compilable example (If u have GLFW and GLM).
Maybe someone have some time and can help me out here :)
#define GLEW_STATIC
#define GLM_FORCE_INLINE
#define GLM_FORCE_SSE2
#include "glew.h"
#include "glfw3.h"
#include "glm.hpp"
#include "glm/gtc/matrix_transform.hpp"
#include <conio.h>
#include <cstdlib>
#include <iostream>
#include <ctime>
GLuint buildShader()
{
std::string strVSCode =
"#version 330 core\n"
"in vec3 vertexPosition;\n"
"in mat4 modelMatrix;\n"
"uniform mat4 projectionMatrix;\n"
"out vec4 m_color;\n"
"void main() {\n"
" vec4 vecVertex = vec4(vertexPosition, 1);\n"
" gl_Position = projectionMatrix * modelMatrix * vecVertex;\n"
" m_color = gl_Position;\n"
"}\n";
std::string strFSCode = "#version 330 core\n"
"out vec4 frag_colour;\n"
"in vec4 m_color;\n"
"void main() {\n"
" frag_colour = vec4(m_color.x, m_color.y, m_color.z, 0.5f);\n"
"}\n";
GLuint gluiVertexShaderId = glCreateShader(GL_VERTEX_SHADER);
char const * VertexSourcePointer = strVSCode.c_str();
glShaderSource(gluiVertexShaderId, 1, &VertexSourcePointer, NULL);
glCompileShader(gluiVertexShaderId);
GLuint gluiFragmentShaderId = glCreateShader(GL_FRAGMENT_SHADER);
char const * FragmentSourcePointer = strFSCode.c_str();
glShaderSource(gluiFragmentShaderId, 1, &FragmentSourcePointer, NULL);
glCompileShader(gluiFragmentShaderId);
GLuint gluiProgramId = glCreateProgram();
glAttachShader(gluiProgramId, gluiVertexShaderId);
glAttachShader(gluiProgramId, gluiFragmentShaderId);
glLinkProgram(gluiProgramId);
glDeleteShader(gluiVertexShaderId);
glDeleteShader(gluiFragmentShaderId);
return gluiProgramId;
}
struct Sprite
{
glm::vec3 position, dimension;
float speed, rotation, rx, ry;
};
struct Vertex
{
float x, y, z;
Vertex(){};
Vertex(float x, float y, float z) : x(x), y(y), z(z) {}
};
int main(int arc, char **argv)
{
// GLFW init
int displayResWith = 1366; //modify this here
int displayResHeight = 768; //modify this here
glfwInit();
glfwWindowHint(GLFW_CONTEXT_VERSION_MAJOR, 3);
glfwWindowHint(GLFW_CONTEXT_VERSION_MINOR, 3);
glfwWindowHint(GLFW_OPENGL_FORWARD_COMPAT, 1);
glfwWindowHint(GLFW_OPENGL_PROFILE, GLFW_OPENGL_CORE_PROFILE);
glfwWindowHint(GLFW_RED_BITS, 8);
glfwWindowHint(GLFW_GREEN_BITS, 8);
glfwWindowHint(GLFW_BLUE_BITS, 8);
glfwWindowHint(GLFW_ALPHA_BITS, 8);
glfwWindowHint(GLFW_DEPTH_BITS, 32);
glfwWindowHint(GLFW_STENCIL_BITS, 32);
GLFWwindow* window = glfwCreateWindow(displayResWith, displayResHeight,"Instancing", glfwGetPrimaryMonitor(),NULL);
int width, height;
glfwMakeContextCurrent(window);
glfwSwapInterval(0);
glfwGetFramebufferSize(window, &width, &height);
//GLEW init
glewExperimental = GL_TRUE;
glewInit();
const GLubyte* renderer = glGetString(GL_RENDERER);
const GLubyte* version = glGetString(GL_VERSION);
std::cout << "Renderer: " << renderer << std::endl;
std::cout << "OpenGL supported version: " << version << std::endl;
//OpenGL init
glEnable(GL_CULL_FACE);
glCullFace(GL_BACK);
glEnable(GL_DEPTH_TEST);
glDepthFunc(GL_LESS);
glEnable(GL_BLEND);
glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
glClearColor(255.0f, 255.0f, 255.0f, 255.0f);
//Shader
GLuint programID = buildShader();
//VBO vertexBuffer
GLuint vertexBuffer;
glGenBuffers(1, &vertexBuffer);
glBindBuffer(GL_ARRAY_BUFFER, vertexBuffer);
Vertex VertexBufferData[6];
VertexBufferData[0] = Vertex(-0.5f, 0.5f, 0.0f); //Links oben
VertexBufferData[1] = Vertex(-0.5f, -0.5f, 0.0f); //Links unten
VertexBufferData[2] = Vertex(0.5f, -0.5f, 0.0f); //Rechts unten
VertexBufferData[3] = VertexBufferData[2]; //Rechts unten
VertexBufferData[4] = Vertex(0.5f, 0.5f, 0.0f); //Rechts oben
VertexBufferData[5] = VertexBufferData[0]; //Links oben
glBufferData(GL_ARRAY_BUFFER, sizeof(Vertex)*6, VertexBufferData, GL_STATIC_DRAW);
//VBO instanceBuffer
GLuint instanceBuffer;
glGenBuffers(1, &instanceBuffer);
glBindBuffer(GL_ARRAY_BUFFER, instanceBuffer);
int iMaxInstanceCount = 30000;
glm::mat4 *ptrInstanceBufferData = new glm::mat4[iMaxInstanceCount];
glBufferData(GL_ARRAY_BUFFER, iMaxInstanceCount * sizeof(glm::mat4), NULL, GL_STREAM_DRAW);
//VAO - Start
GLuint vertexArrayObject;
glGenVertexArrays(1, &vertexArrayObject);
glBindVertexArray(vertexArrayObject);
//For VBO vertexbuffer
glEnableVertexAttribArray(glGetAttribLocation(programID, "vertexPosition"));
glBindBuffer(GL_ARRAY_BUFFER, vertexBuffer);
glVertexAttribPointer(
glGetAttribLocation(programID, "vertexPosition"),
3,
GL_FLOAT,
GL_FALSE,
sizeof(Vertex),
(void*)0
);
glVertexAttribDivisor(0, 0);
//For VBO instanceBuffer
int pos = glGetAttribLocation(programID, "modelMatrix");
int pos1 = pos + 0;
int pos2 = pos + 1;
int pos3 = pos + 2;
int pos4 = pos + 3;
glEnableVertexAttribArray(pos1);
glEnableVertexAttribArray(pos2);
glEnableVertexAttribArray(pos3);
glEnableVertexAttribArray(pos4);
glBindBuffer(GL_ARRAY_BUFFER, instanceBuffer);
glVertexAttribPointer(pos1, 4, GL_FLOAT, GL_FALSE, sizeof(GLfloat) * 4 * 4, (void*)(0));
glVertexAttribPointer(pos2, 4, GL_FLOAT, GL_FALSE, sizeof(GLfloat) * 4 * 4, (void*)(sizeof(float) * 4));
glVertexAttribPointer(pos3, 4, GL_FLOAT, GL_FALSE, sizeof(GLfloat) * 4 * 4, (void*)(sizeof(float) * 8));
glVertexAttribPointer(pos4, 4, GL_FLOAT, GL_FALSE, sizeof(GLfloat) * 4 * 4, (void*)(sizeof(float) * 12));
glVertexAttribDivisor(pos1, 1);
glVertexAttribDivisor(pos2, 1);
glVertexAttribDivisor(pos3, 1);
glVertexAttribDivisor(pos4, 1);
glBindVertexArray(0); //VAO - End
//Matrix vars
glm::mat4 Projection, Rotating, Scaling, Translation, Identity;
glm::vec3 ZRotateVec(0.0f, 0.0f, 1.0f);
//Calc projection-matrix and put shader (uniform)
Projection = glm::ortho(0.0f, (float)width, 0.0f, (float)height, 0.0f, 1.0f);
glUseProgram(programID);
glUniformMatrix4fv(glGetUniformLocation(programID, "projectionMatrix"), 1, GL_FALSE, &Projection[0][0]);
//Creating sprites
std::srand(static_cast<unsigned int>(std::time(0)));
int iActInstanceCount = 10000;
Sprite *ptrSprites = new Sprite[iActInstanceCount];
for (int i = 0; i < iActInstanceCount; ++i)
{
ptrSprites[i].dimension = glm::vec3(16, 16, 1.0f);
ptrSprites[i].position = glm::vec3(std::rand()%(width-32),std::rand()%(height-32),-1.0f *((std::rand()%256)/256.0f));
ptrSprites[i].rotation = rand() % 360 + 0.0f;
ptrSprites[i].rx = static_cast<float>(std::rand() % 2);
ptrSprites[i].ry = static_cast<float>(std::rand() % 2);
ptrSprites[i].speed = (std::rand() % 100) + 1.0f;
if (ptrSprites[i].speed < 1.0f) ptrSprites[i].speed = 1.0f;
}
//FPS init
double fFramesRendered = 0.0f;
double fFrameMeasurementStart = 0.0f;
double fFPS = 0.0f;
double fCurrentTime = 0.0f;
glfwSetTime(0);
//Main-loop (also renderloop)
while (!glfwWindowShouldClose(window))
{
//application-logic
if (glfwGetKey(window, GLFW_KEY_ESCAPE)== GLFW_PRESS)
glfwSetWindowShouldClose(window, GL_TRUE);
const double fNewTime = glfwGetTime();
double fDeltaTime = fNewTime - fCurrentTime;
fCurrentTime = fNewTime;
for (int i = 0; i < iActInstanceCount; ++i)
{
float fSpeed = ptrSprites[i].speed * static_cast<float>(fDeltaTime);
ptrSprites[i].rotation += fSpeed;
if (ptrSprites[i].rotation >= 360.0f) ptrSprites[i].rotation = 0.0f;
if (ptrSprites[i].rx == 1) ptrSprites[i].position.x = ptrSprites[i].position.x + fSpeed;
if (ptrSprites[i].rx == 0) ptrSprites[i].position.x = ptrSprites[i].position.x - fSpeed;
if (ptrSprites[i].ry == 1) ptrSprites[i].position.y = ptrSprites[i].position.y + fSpeed;
if (ptrSprites[i].ry == 0) ptrSprites[i].position.y = ptrSprites[i].position.y - fSpeed;
if (ptrSprites[i].position.x <= 0) ptrSprites[i].rx = 1;
if (ptrSprites[i].position.x + ptrSprites[i].dimension.x >= width) ptrSprites[i].rx = 0;
if (ptrSprites[i].position.y <= 0) ptrSprites[i].ry = 1;
if (ptrSprites[i].position.y + ptrSprites[i].dimension.y >= height) ptrSprites[i].ry = 0;
//matrix-calculations (saved in local buffer)
Translation = glm::translate(Identity, ptrSprites[i].position + glm::vec3(ptrSprites[i].dimension.x / 2.0f, ptrSprites[i].dimension.y / 2.0f, 0.0f));
Scaling = glm::scale(Translation, ptrSprites[i].dimension);
ptrInstanceBufferData[i] = glm::rotate(Scaling, ptrSprites[i].rotation, ZRotateVec);
}
//render-call
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
glUseProgram(programID);
glBindVertexArray(vertexArrayObject);
glBindBuffer(GL_ARRAY_BUFFER, instanceBuffer);
glBufferData(GL_ARRAY_BUFFER, iMaxInstanceCount * sizeof(glm::mat4), NULL, GL_STREAM_DRAW); // Buffer orphaning
glBufferSubData(GL_ARRAY_BUFFER, 0, iActInstanceCount * sizeof(glm::mat4), ptrInstanceBufferData);
glDrawArraysInstanced(GL_TRIANGLES, 0, 6, iActInstanceCount);
glBindVertexArray(0);
glfwSwapBuffers(window);
glfwPollEvents();
//FPS-stuff
++fFramesRendered;
if ((fCurrentTime*1000.0f) >= (fFrameMeasurementStart*1000.0f) + 1000.0f)
{
fFPS = ((fCurrentTime*1000.0f) - (fFrameMeasurementStart*1000.0f)) / 1000.0f * fFramesRendered;
fFrameMeasurementStart = fCurrentTime;
fFramesRendered = 0;
std::cout << "FPS: " << fFPS << std::endl;
}
}
//Termination and cleanup
glDeleteBuffers(1, &vertexBuffer);
glDeleteBuffers(1, &instanceBuffer);
glDeleteVertexArrays(1, &vertexArrayObject);
glDeleteProgram(programID);
glfwDestroyWindow(window);
glfwTerminate();
return _getch();
}

Answer 1

Well, after testing it on my machine, it is definitely CPU limited, so nothing you do with OGL is going to make much difference. I get about ~300fps with GCC on at least -O1, but only ~80 with -O0. My CPU is very fast (i7 2600k, 4.7ghz), but my GPU is rather slow (GT 520). I'm also on Ubuntu.
Some quick ideas for things that might speed it up a little:
Put the vertex positions in an array in the vertex shader and use gl_VertexID to access them
Use GL_TRIANGLE_STRIP instead of GL_TRIANGLES
Use radians for angles, as otherwise GLM has to convert them
None of these are likely to make much of any impact, really. Just make sure your compiler is set up right, and there probably isn't much more to do.