Jump to content

  • Log In with Google      Sign In   
  • Create Account

We're offering banner ads on our site from just $5!

1. Details HERE. 2. GDNet+ Subscriptions HERE. 3. Ad upload HERE.


TexSubImage2D performance


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
8 replies to this topic

#1 Chris_F   Members   -  Reputation: 2460

Like
1Likes
Like

Posted 19 February 2014 - 06:31 PM

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.



Sponsor:

#2 mark ds   Members   -  Reputation: 1482

Like
0Likes
Like

Posted 20 February 2014 - 07:47 AM

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.


Edited by mark ds, 20 February 2014 - 07:49 AM.


#3 Chris_F   Members   -  Reputation: 2460

Like
0Likes
Like

Posted 20 February 2014 - 02:28 PM

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);
}


#4 mark ds   Members   -  Reputation: 1482

Like
0Likes
Like

Posted 20 February 2014 - 03:39 PM

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...



#5 Chris_F   Members   -  Reputation: 2460

Like
0Likes
Like

Posted 20 February 2014 - 05:02 PM

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.


Edited by Chris_F, 20 February 2014 - 05:03 PM.


#6 Ashaman73   Crossbones+   -  Reputation: 7987

Like
0Likes
Like

Posted 21 February 2014 - 07:07 AM


with one MIP level

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 *):

 

 24) How should mipmap generation work for sRGB textures?

        RESOLVED:  The best way to perform mipmap generation for sRGB
        textures is by downsampling the sRGB image in a linear color
        space.

*        This involves converting the RGB components of sRGB texels
*        in a given texture image level to linear RGB space, filtering
*        appropriately in that linear RGB space, and then converting the
*        linear RGB values to sRGB for storage in the downsampled texture
*        level image.

        (Remember alpha, when present, is linear even in sRGB texture
        formats.)

        The OpenGL specification says "No particular filter algorithm
        is required, though a box filter is recommended as the default
        filter" meaning there is no requirement for how even non-sRGB
        mipmaps should be generated.  So while the resolution to this
        issue is technically a recommendation, it is however a strongly
        advised recommendation.

        The rationale for why sRGB textures should be converted to
        linear space prior to filtering and converted back to sRGB after
        filtering is clear.  If an implementation naively simply performed
        linear filtering on (non-linear) sRGB components as if they were
        in a linear space, the result tends to be a subtle darkening of
        the texture images as mipmap generation continues recursively.
        This darkening is an inappropriate basis that the resolved
        "best way" above would avoid.

Edited by Ashaman73, 21 February 2014 - 07:09 AM.


#7 samoth   Crossbones+   -  Reputation: 5032

Like
0Likes
Like

Posted 21 February 2014 - 11:51 AM

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?



#8 Chris_F   Members   -  Reputation: 2460

Like
0Likes
Like

Posted 21 February 2014 - 04:18 PM


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.



#9 Chris_F   Members   -  Reputation: 2460

Like
0Likes
Like

Posted 26 March 2014 - 04:18 PM

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.


Edited by Chris_F, 26 March 2014 - 04:19 PM.





Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS