I was curious to see the performance of texture uploads with my configuration using OpenGL and noticed something I think is odd. I create a 4K texture using glTexStorage2D with one MIP level and a format of GL_RGBA8. Then, every frame I use glTexSubImage2D to re-upload a static image buffer to the texture. Based off the frame rate I get about 5.19GB/s. Next, I changed the format of the texture to GL_SRGB8_ALPHA8 and re-try the experiment. This time I am getting 2.81GB/s, a significant decrease. This seems odd because as far as I know there shouldn't be anything different about uploading sRGB data verses uploading RGB data, as there is no conversion that should be taking place (sRGB conversion takes place in the shader, during sampling).

Some additional information. All I'm rendering is a fullscreen quad with a pixel shader that simply outputs vec4(1), I'm not even sampling from the texture or doing anything else each frame other than calling glTexSubImage2D. For the first test I use GL_RGBA and GL_UNSIGNED_INT_8_8_8_8_REV in the call to glTexSubImage2D, as this is what the driver tells me is ideal. For the second test I use GL_UNSIGNED_INT_8_8_8_8, as per the drivers suggestion. A bit of testing confirms that these are the fastest formats to use respectively. This is using an Nvidia GPU.

Are both source and destination SRGB?

Edit - just to clarify - by one mip level do you mean you're only using mip0, rather than mip0 & mip1.

Are both source and destination SRGB?

Edit - just to clarify - by one mip level do you mean you're only using mip0, rather than mip0 & mip1.

#include <GL/glew.h>
#include <GLFW/glfw3.h>
#include <vector>
#include <cstdlib>
#include <cstdio>

#define SCREEN_SIZE_X 1024
#define SCREEN_SIZE_Y 1024

#define GLSL(src) "#version 440 core\n" #src

const char* vertex_shader = GLSL(
    const vec2 data[4] = vec2[]
    (
        vec2(-1.0,  1.0),
        vec2(-1.0, -1.0),
        vec2( 1.0,  1.0),
        vec2( 1.0, -1.0)
    );

    void main()
    {
        gl_Position = vec4(data[gl_VertexID], 0.0, 1.0);
    }
);

const char* fragment_shader = GLSL(
    layout(location = 0) uniform sampler2D texture0;
    layout(location = 1) uniform vec2 screenSize;
    out vec4 frag_color;

    void main()
    {
        frag_color = texture(texture0, gl_FragCoord.xy / screenSize);
    }
);

int main(int argc, char *argv[])
{
    if(!glfwInit())
        exit(EXIT_FAILURE);

    glfwWindowHint(GLFW_RESIZABLE, GL_FALSE);
    glfwWindowHint(GLFW_CONTEXT_VERSION_MAJOR, 4);
    glfwWindowHint(GLFW_CONTEXT_VERSION_MINOR, 4);
    glfwWindowHint(GLFW_OPENGL_PROFILE, GLFW_OPENGL_CORE_PROFILE);
    glfwWindowHint(GLFW_OPENGL_FORWARD_COMPAT, GL_TRUE);

    GLFWwindow* window = glfwCreateWindow(SCREEN_SIZE_X, SCREEN_SIZE_Y, "OpenGL Texture Upload", nullptr, nullptr);

    if(!window)
    {
        glfwTerminate();
        exit(EXIT_FAILURE);
    }

    glfwMakeContextCurrent(window);
    glfwSwapInterval(0);

    glewExperimental = GL_TRUE;

    if(glewInit() != GLEW_OK)
    {
        glfwTerminate();
        exit(EXIT_FAILURE);
    }
    
    GLuint vao = 0;
    glGenVertexArrays(1, &vao);
    glBindVertexArray(vao);

    GLuint vs = glCreateShader(GL_VERTEX_SHADER);
    glShaderSource(vs, 1, &vertex_shader, nullptr);
    glCompileShader(vs);

    GLuint fs = glCreateShader(GL_FRAGMENT_SHADER);
    glShaderSource(fs, 1, &fragment_shader, nullptr);
    glCompileShader(fs);

    GLuint shader_program = glCreateProgram();
    glAttachShader(shader_program, fs);
    glAttachShader(shader_program, vs);
    glLinkProgram(shader_program);
    glUseProgram(shader_program);

    glProgramUniform2f(shader_program, 1, SCREEN_SIZE_X, SCREEN_SIZE_Y);

    GLuint texture = 0;
    glGenTextures(1, &texture);
#ifdef USE_SRGB
    glTextureStorage2DEXT(texture, GL_TEXTURE_2D, 1, GL_SRGB8_ALPHA8, 4096, 4096);
#else
    glTextureStorage2DEXT(texture, GL_TEXTURE_2D, 1, GL_RGBA8, 4096, 4096);
#endif
    glTextureParameteriEXT(texture, GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
    glTextureParameteriEXT(texture, GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
    glTextureParameteriEXT(texture, GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
    glTextureParameteriEXT(texture, GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
    glBindMultiTextureEXT(GL_TEXTURE0, GL_TEXTURE_2D, texture);
    glProgramUniform1i(shader_program, 0, 0);

    std::vector<unsigned int> image_buffer(4096*4096, 0xFF0000FFul);

    double lastTime = glfwGetTime();
    double nbFrames = 0;

    while(!glfwWindowShouldClose(window))
    {
        double currentTime = glfwGetTime();
        nbFrames++;
        if (currentTime - lastTime >= 1.0)
        {
            char cbuffer[50];
            snprintf(cbuffer, sizeof(cbuffer), "OpenGL Texture Upload [%.1f fps, %.3f ms]", nbFrames, 1000.0 / nbFrames);
            glfwSetWindowTitle(window, cbuffer);
            nbFrames = 0;
            lastTime++;
        }
#ifdef USE_SRGB
        glTextureSubImage2DEXT(texture, GL_TEXTURE_2D, 0, 0, 0, 4096, 4096, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8, image_buffer.data());
#else
        glTextureSubImage2DEXT(texture, GL_TEXTURE_2D, 0, 0, 0, 4096, 4096, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8_REV, image_buffer.data());
#endif
        glClear(GL_COLOR_BUFFER_BIT);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
        glfwSwapBuffers(window);
        glfwPollEvents();
    }

    glfwDestroyWindow(window);
    glfwTerminate();
    exit(EXIT_SUCCESS);
}

Seems OK - I was guessing some internal conversion between SRGB & RGB.

However, GL_UNSIGNED_INT_8_8_8_8 vs GL_UNSIGNED_INT_8_8_8_8_REV stands out as a likely candidate...

Seems OK - I was guessing some internal conversion between SRGB & RGB.

However, GL_UNSIGNED_INT_8_8_8_8 vs GL_UNSIGNED_INT_8_8_8_8_REV stands out as a likely candidate...

I used glGetInternalformativ to find the optimum pixel format for each internal format. As I said, experimental testing confirms that no other pixel format is faster for these respective formats, so no matter what, uploading sRGB is always slower than uploading RGB. As for internal conversion, as I said, that should not be the case. If it were, that would be horrendous.

here is no conversion that should be taking place (sRGB conversion takes place in the shader, during sampling).

Some additional information. All I'm rendering is a fullscreen quad with a pixel shader that simply outputs vec4(1), I'm not even sampling from the texture or doing anything else

This, however, looks like you are:

frag_color = texture(texture0, gl_FragCoord.xy / screenSize);

Maybe it is the actual sampling that slows you down, not the texture upload.

Can you, instead of measuring frame time, use a timer query (and maybe a fence sync) so you really measure glTexImage2D?

This could be the problem. Refering to its specification , it is possible, that your texture (-parts) will be converted into linear RGB first, then downsampled and converted back to sRGB afterwards. Take a look at the specification (marked it with *):

That should not be the case. I create the texture through the glTexStorage API, asking for only 1 MIP level, which means that I get an immutable texture that only has room for one MIP level. The driver wouldn't (shouldn't) be calculating MIPs on a texture that can't store them. Experimentation with textureLod confirms that there are no other MIPs present.

Maybe it is the actual sampling that slows you down, not the texture upload.



Can you, instead of measuring frame time, use a timer query (and maybe a fence sync) so you really measure glTexImage2D?

I suppose I could give it a try (I have not used those features before), however my intuition tells me that the cost of doing sRGB conversion for a single texture on a single quad at 1024x1024 screen resolution is going to be negligible. Also, I modified it so that I only uploaded the texture once and then simply drew the quad each frame, and in that scenario there is no difference in performance for sRGB or RGB.

OK, can anyone actually help me measure this correctly. No matter what, the result of a timer query seems to be the same, about 0.001056ms.

For example, if I do this:

GLuint query;
glGenQueries(1, &query);
glBeginQuery(GL_TIME_ELAPSED, query);

for(int i = 0; i < 100; ++i) {
    glTextureSubImage2DEXT(texture, GL_TEXTURE_2D, 0, 0, 0, 8192, 8192, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8_REV, image_buffer.data());
}

glEndQuery(GL_TIME_ELAPSED);
GLuint elapsed_time;
glGetQueryObjectuiv(query, GL_QUERY_RESULT, &elapsed_time);
glDeleteQueries(1, &query);
float time_ms = elapsed_time / 1000000.0f;
printf("time: %f ms\n", time_ms);

It will take many seconds to complete and yet still 0.001056ms is printed.