opengl - glMultiDrawElementsIndirect 很慢

Question

我正在重写我的旧渲染管道。我根据自己的喜好创建了一个非常精简的原型(prototype)，令我震惊的是，我原来相当复杂且优化不佳的管道与 super 简单的原型(prototype)具有完全相同的性能。

任务是渲染 1024 个任意大小的网格(总共 1400 万个三角形)，每个网格使用一组不同的制服。

我现在做的是使用统一缓冲区 + glMultiDrawElementsIndirect 并使用 gl_DrawIDARB 索引到统一缓冲区。这是渲染循环:

function renderloop(window, N, frame_times, program, commandbuff)
    glUseProgram(program)
    glEnable(GL_DEPTH_TEST)
    glClearColor(1, 1, 1, 1)
    GLAbstraction.bind(commandbuff)
    n = 0
    while isopen(window) && n <= N
        tic()
        glFinish() # make sure we time the right thing
        GLWindow.poll_glfw()
        #glBindVertexArray(vbo.id) doesn't change timing much
        glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT)
        glMultiDrawElementsIndirect(
            GL_TRIANGLES,
            GL_UNSIGNED_INT,
            C_NULL, length(commandbuff), 0
        )
        #glBindVertexArray(0)
        GLWindow.swapbuffers(window)
        push!(frame_times, toq())
        n += 1
    end
    frame_times
end

我的其他管道太复杂了，无法在这里写下来，但简而言之，它是未优化的 Julia 代码、带制服的 GLSL 3.0 绘图代码 + 光线拾取 + fxaa + 几个渲染目标等等。除了具有统一 block 等的现代化之外，着色器几乎相同。

新的(几乎)完整的代码可以在这里看到:

vert = """
#version 450
#extension GL_ARB_shader_draw_parameters : enable

struct VertexArgument{
    vec4 color;
    mat4 model;
};

layout (location = 0) in vec3 position;
layout (location = 1) in vec3 normal;

layout (std140) uniform Scene{
    vec4 lightposition;
    mat4 proj;
    mat4 view;
    mat4 projview;
    vec2 resolution;
} scene;

layout (std140) uniform VertexArguments{
    VertexArgument[1024] args;
} vertex_arguments;


out VertexOut{
    vec3 vertex;
    vec3 normal;
    vec3 lightdir;
    vec4 color;
} vertex_out;

void main(){
    VertexArgument arg = vertex_arguments.args[gl_DrawIDARB];
    vec4 position_camspace = scene.view * arg.model * vec4(position,  1.0);
    gl_Position = scene.proj * position_camspace;
    vertex_out.lightdir = normalize(vec3(-10) - position.xyz);
    vertex_out.vertex = -position_camspace.xyz;
    vertex_out.normal = normal;
    vertex_out.color = arg.color;
}
"""

frag = """
#version 450

vec3 blinnphong(vec3 V, vec3 N, vec3 L, vec3 color){

    float diff_coeff = max(dot(L,N), 0.0);

    // specular coefficient
    vec3 H = normalize(L+V);

    float spec_coeff = pow(max(dot(H,N), 0.0), 8.0);
    if (diff_coeff <= 0.0)
        spec_coeff = 0.0;

    // final lighting model
    return vec3(
        vec3(0.1) * vec3(0.3)  +
        vec3(0.9) * color * diff_coeff +
        vec3(0.3) * spec_coeff
    );
}

in VertexOut{
    vec3 vertex;
    vec3 normal;
    vec3 lightdir;
    vec4 color;
} vertex_in;

layout (location = 0) out vec4 frag_color;

void main(){
    vec3 L = normalize(vertex_in.lightdir);
    vec3 N = normalize(vertex_in.normal);
    vec3 light1 = blinnphong(vertex_in.vertex, N, L, vertex_in.color.rgb);
    vec3 light2 = blinnphong(vertex_in.vertex, N, -L, vertex_in.color.rgb);
    frag_color = vec4(light1 + light2, 1.0);
}
"""

window = create_glcontext(
    major = 4, minor = 5, debugging = false,
    windowhints = [
        (GLFW.SAMPLES,      0),
        (GLFW.DEPTH_BITS,   32),

        (GLFW.ALPHA_BITS,   8),
        (GLFW.RED_BITS,     8),
        (GLFW.GREEN_BITS,   8),
        (GLFW.BLUE_BITS,    8),

        (GLFW.STENCIL_BITS, 0),
        (GLFW.AUX_BUFFERS,  0)
    ]
)

events = WindowEvents(Window => window)

cam = PerspectiveCamera(
    TranslationSpeed => 1f0,
    LookAt => Vec3f0(0),
    EyePosition => Vec3f0(6, 6, 8),
    Rotation => Vec3f0(0),
    Area => events[Area],
    RotationSpeed => 0.1f0
)

vertshader = compile_shader(Vector{UInt8}(vert), GL_VERTEX_SHADER, :vertexshader)
fragshader = compile_shader(Vector{UInt8}(frag), GL_FRAGMENT_SHADER, :fragshader)

program = compile_program(vertshader, fragshader)

scene = (
    Vec4f0(10),
    cam[Projection],
    cam[View],
    cam[ProjectionView],
    Vec2f0(widths(cam[Area]))
)

scene_buff = UniformBuffer(scene) # create UniformBuffer GL_STATIC_DRAW

FieldTraits.on(cam, ProjectionView) do projview
    # write new values to scene buffer.. if not doing this, timings stay the same
    scene_buff[1] = (
        Vec4f0(10),
        cam[Projection],
        cam[View],
        projview,
        Vec2f0(widths(cam[Area]))
    )
end

vals = (Vec4f0(1, 0, 0, 1), eye(Mat4f0))
uniform_array = UniformBuffer(typeof(vals))

function loadmeshes(folder)
  # load 1024 meshes
    meshpaths = filter(x-> endswith(x, ".ifs"), readdir(folder))[1:1024]
    faces = GLTriangle[]
    vertices = Tuple{Point3f0, Normal{3, Float32}}[]
    fidx = 0; vidx = 0;
    drawcommands = Vector{Command}(length(meshpaths))
    for (i, meshpath) in enumerate(meshpaths)
        mesh = read_ifs(joinpath(folder, meshpath))
        fs, vs = mesh.indexes[1], mesh.parent
        append!(faces, fs)
        ns = normals(vs, fs)
        append!(vertices, zip(vs, ns))
        mini, maxi = extrema(mesh.parent)
        x, y = ind2sub((32, 32), i)
        trans = translationmatrix(Vec3f0(x, y, 0f0))
        s = maximum(maxi .- mini)
        scale = scalematrix(Vec3f0(1f0 ./ s))
    # add uniform attributes to buffer
        push!(uniform_array, (
            Vec4f0(rand(Vec3f0)..., 1f0),
            trans * scale * translationmatrix(-Vec3f0(mini))
        ))
        drawcommands[i] = Command(length(fs) * 3, 1, fidx, vidx, 0)
        fidx += length(fs) * 3; vidx += length(vs)
    end

    vbo = VertexArray(view(vertices, faces)) # vertexarray
    ibuff = GLBuffer(drawcommands, buffertype = GL_DRAW_INDIRECT_BUFFER)
    vbo, ibuff
end

vbo, commandbuff = loadmeshes(homedir() * "/3dstuff/models")
sceneidx = glGetUniformBlockIndex(program, "Scene")
vertex_arts_idx = glGetUniformBlockIndex(program, "VertexArguments")

glUniformBlockBinding(program, sceneidx, 0)
glUniformBlockBinding(program, vertex_arts_idx, 1)
glBindBufferBase(GL_UNIFORM_BUFFER, 0, scene_buff.buffer.id)
glBindBufferBase(GL_UNIFORM_BUFFER, 1, uniform_array.buffer.id)


function renderloop(window, N, frame_times, commandbuff)
    glUseProgram(program)
    glEnable(GL_DEPTH_TEST)
    glClearColor(1, 1, 1, 1)
    GLAbstraction.bind(commandbuff)
    n = 0
    while isopen(window) && n <= N
        tic()
        glFinish() # make sure we time the real thing
        GLWindow.poll_glfw()
        #glBindVertexArray(vbo.id) doesn't change timing much
        glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT)
        glMultiDrawElementsIndirect(
            GL_TRIANGLES,
            GL_UNSIGNED_INT,
            C_NULL, length(commandbuff), 0
        )
        #glBindVertexArray(0)
        GLWindow.swapbuffers(window)
        push!(frame_times, toq())
        n += 1
    end
    frame_times
end
times = Float64[]
renderloop(window, 2000, times, commandbuff)
mean(times) * 1000 # ~ 14 ms

GPU 是 FirePro 9100。

旧管道的计时:每帧约 13 毫秒。新原型(prototype):~15 毫秒和 0.2 毫秒，没有调用 glMultiDrawElementsIndirect。

我还尝试打开和关闭 vsync，并稍微移动了代码，在时间上没有任何差异。新原型(prototype)手感也不太顺畅，看来这不仅仅是测量问题。

Answer 1

glMultiDrawElementsIndirect( GL_TRIANGLES, GL_UNSIGNED_INT, C_NULL，长度(命令缓冲区)，0 )

这个参数应该是你要绘制多少个元素。将 1024 放在这里以查看它是否解决了性能问题。