Venturing into OptiX – cudaMemcpy breaks everything

I’ve been looking into using Nvidia Optix for my Fractal lookdev tool, as I need realtime feedback while traversing the landscape.

I really like the API so far, it even comes with a Julia example demo in the SDK! I’ve modified it to render the Mandelbulb instead, with little trouble (and with refraction).

12318072_10208255826582355_1853094794_o

However, I needed to read the buffer data to write .exr’s to disk (this is a vital feature, better to make sure it’s viable at the beginning of the project). That shouldn’t be a problem, as I can easily grab the buffer’s device pointer and then cudaMemcpy it over.

All I need to do is include and , no big deal right?

#include <cuda.h>
#include <cuda_runtime.h>

#include <optixu/optixpp_namespace.h>
#include <optixu/optixu_math_namespace.h>
Linking CXX executable ../bin/julia
CMakeFiles/julia.dir/julia.cpp.o: In function `AnimCamera::apply(optix::Handle<optix::ContextObj>)':
/home/tom/src/optix/SDK/julia/julia.cpp:151: undefined reference to `PinholeCamera::PinholeCamera(float3, float3, float3, float, float, PinholeCamera::AspectRatioMode)'
/home/tom/src/optix/SDK/julia/julia.cpp:153: undefined reference to `PinholeCamera::getEyeUVW(float3&, float3&, float3&, float3&)'
collect2: error: ld returned 1 exit status

Oh.

Well maybe it’s because I included them before OptiX.

#include <optixu/optixpp_namespace.h>
#include <optixu/optixu_math_namespace.h>

#include <cuda.h>
#include <cuda_runtime.h>

using namespace optix;

/usr/local/cuda-7.0/include/cuda_runtime_api.h:257:17: error: ‘cudaError_t’ does not name a type
 extern __host__ cudaError_t CUDARTAPI cudaDeviceReset(void);
                 ^
/usr/local/cuda-7.0/include/cuda_runtime_api.h:274:36: error: ‘cudaError_t’ does not name a type
 extern __host__ __cudart_builtin__ cudaError_t CUDARTAPI cudaDeviceSynchronize(void);
                                    ^
/usr/local/cuda-7.0/include/cuda_runtime_api.h:349:17: error: ‘cudaError_t’ does not name a type
 extern __host__ cudaError_t CUDARTAPI cudaDeviceSetLimit(enum cudaLimit limit, size_t value);
                 ^

...

/home/tom/src/optix/SDK/julia/julia.cpp:299:95: error: ‘cudaMemcpy’ was not declared in this scope
   cudaMemcpy( (void*)h_ptr, (void*)d_ptr, sizeof(float) * totalPixels, cudaMemcpyDeviceToHost );

Oh.

This is dire, could it be related to the optix namespace somehow?

#include <optixu/optixpp_namespace.h>
#include <optixu/optixu_math_namespace.h>
using namespace optix;

#include <cuda.h>
#include <cuda_runtime.h>
Linking CXX executable ../bin/julia
[100%] Built target tutorial

Success! Looks blindly using the optix namespace causes clashes with Cuda, setting it up like I did above fixes that in a janky way. Maybe this is mentioned in the documentation somewhere but I never found any reference to it.

So if you’re getting undefined references when you need cudaMemcpy in an OptiX project, check out your include order and namespace setup (and don’t forget to add ${CUDA_LIBRARIES} to target_link_libraries in the cmake configuration if it’s not there already).

The best solution would be to only bring in the stuff you actually need from the optix namespace, so there’s no namespace clashing at all.

So far OptiX seems pretty solid though, I hope I’ll have fun with it!

Update

Turns out in this particular instance I was better off using the map() and unmap() functions within the Optix Buffer class instead of cudaMemcpy(), they remove the dependency of including the CUDA headers. However, I would no doubt have run into the same problems using something like the NVRTC headers.

Playing with Fractals in WebGL

Playing with Fractals in WebGL

As an excuse to play with WebGL I decided to finally got around to playing with the Mandelbrot and Julia set. There’s a live demo at the end of the post, the zoom control/mouse movement needs some work but the basic idea works great.

Links

A couple of useful places I used as reference :
Ozone3D – The Mandelbrot Set: Colors of Infinity
Fractals: Useful Beauty
dat.GUI – A great JavaScript GUI

The Shader Code

Vertex Shader

Really simple, all we need is a fullscreen quad with smooth UV’s.

attribute vec3 vtx_pos;
attribute vec2 vtx_uv;

varying vec2 texcoord;

void main(void)
{
    texcoord = vtx_uv;
    gl_Position = vec4(vtx_pos, 1.0);
}
Mandelbrot Set Fragment Shader

This one is very similar to the code in Ozone3D’s tutorial, I basically copied it for the most part (there’s only so many ways to skin a cat anyway).

The most notable aspect is MAX_ITER, which I replace with the actual iteration value before I compile the shader code. This was to get around the limitation of most webgl implementations which disallow you from using non-constant loop conditions, as they will be unrolled. This isn’t very convenient for real time display of any possible iteration size, so I simply recompile the shader when the iteration size changes.

Note that the “uniform vec2 c” uniform isn’t used here, it’s just there to keep the uniform layout identical for both shaders (I’m lazy and it’s just a demo).

Finally, I use a 1D texture (passed in through a 2D sampler, but the actual texture is only 1 pixel tall) to lookup the expected colour based on the iteration level.

#define iterations MAX_ITER

precision highp float;

varying vec2 texcoord;

uniform sampler2D gradientSampler;
uniform vec2 center;
uniform float zoom;
uniform vec2 c;

void main(void)
{
    // Mandelbrot
    vec2 C = vec2(1.0, 0.75) * (texcoord - vec2(0.5, 0.5)) * vec2(zoom, zoom) - center;
    vec2 z = C;

    gl_FragColor = vec4(0,0,0,1);

    for(float i = 0.0; i < float(iterations); i += 1.0)
    {
        z = vec2( z.x*z.x - z.y*z.y, 2.0 * z.x * z.y) + C;

        if(dot(z,z) > 4.0)
        {
            gl_FragColor = texture2D(gradientSampler, vec2(i / float(iterations), 0.5));
            break;
        }
    }
}
Julia Set Fragment Shader

The base code is identical to the Mandelbrot shader, except now we can pass in the desired complex parameter through c.

I can’t explain the maths as well as the pages I referenced earlier; but the main change is now the smooth UVs of the fullscreen quad are now assigned to Z instead of C, as the complex parameter is now a user defined constant and Z varies.

#define iterations MAX_ITER

precision highp float;

varying vec2 texcoord;

uniform sampler2D gradientSampler;
uniform vec2 center;
uniform float zoom;
uniform vec2 c;

void main(void)
{
    // Julia
    vec2 z = vec2(1.0, 0.75) * (texcoord - vec2(0.5, 0.5)) * vec2(zoom, zoom) - center;

    gl_FragColor = vec4(0,0,0,1);

    for(float i = 0.0; i < float(iterations); i += 1.0)
    {
        z = vec2( z.x*z.x - z.y*z.y, 2.0 * z.x * z.y) + c;
        if(dot(z,z) > 4.0)
        {
            gl_FragColor = texture2D(gradientSampler, vec2(i / float(iterations), 0.5));
            break;
        }
    }
}

The Live Demo (full page demo)

  • Toggle between the Mandelbrot/Julia set.
  • Use the zoom control to zoom in (this isn’t ideal).
  • Expect slow down on high iterations.
  • If you want to tinker around with the code then I apologise for how messy it is, it’s my very first adventure into JavaScript and I kept bolting more stuff onto it until I ended up with this.

ToDo

The Ozone3D page mentions that “a better approach is to use a multipass algorithm.”, which is next in line for this project.

It would also be cool to choose a complex parameter for the Julia set by clicking on the Mandelbrot itself, as both sets are connected.

Unreal Engine 4 on Linux – Problems with Optimus/Bumblebee “[VGL] ERROR: Could not open display :8”

The problem

Unreal Engine 4 works pretty well on the Linux desktop now, but I ran into unexpected problems when trying to run the editor natively on my NVIDIA Optimus equipped laptop. I tried to enable my dedicated gpu by using optirun but the “[VGL] ERROR: Could not open display :8”  error message was logged when I tried to load a project from the project browser.

This seems to be because optirun doesn’t like Unreal closing and forking a new editor process when the project is chosen. This is not a problem when opening a .uproject file directly, as nothing is closed and no new editor processes are forked. The problem (and potential solutions) are briefly described here, Bumblebee Documentation.

For most people the primus solution would probably work fine, but I never had any luck in getting primus to work so I had to figure out a tidy workaround for the alternative solution (opening a bash session with optirun and running the editor within that). 

A neat way of asking a new bash instance to run a bunch of commands is described here : http://stackoverflow.com/questions/7120426/invoke-bash-run-commands-inside-new-shell-then-give-control-back-to-user

My Solution

In the end I ended up using something like this to open a new optirun powered bash shell, start unreal and then wait until it exits before finally exiting the shell again.

# Where does unreal live
unreal_path=$HOME/src/UnrealEngine/Engine/Binaries/Linux

# Make a temporary file to store our commands in
tmp_file=$(mktemp)

# Open project window
echo "$unreal_path/UE4Editor &gt;/dev/null 2&gt;@1" &gt;&gt; $tmp_file

# Figure out the PID of the new editor instance that the project window spawned 
# (there are probably better ways of doing this, like pgrep)

# Tokenize the string
echo "editor_process=\$(ps -ejH | tail -n1)" &gt;&gt; $tmp_file
# We only want the first token from the string so we just grab the first item here
echo "editor_pid=( \$editor_process )" &gt;&gt; $tmp_file
    
# Wait until the editor closes
echo "while kill -0 \$editor_pid &gt;/dev/null 2&gt;@1; do sleep 1; done" &gt;&gt; $tmp_file

# Clean up and exit
echo "rm -f $tmp_file" &gt;&gt; $tmp_file
echo "exit" &gt;&gt; $tmp_file

# Launch a new optirun powered bash instance and ask it to execute the commands
# in the file
optirun bash --rcfile $tmp_file

My polished solution

Note that I use the fish shell, but the syntax should be simple enough to convert to bash if you feel inspired. Notable features include storing the log file for later reading if necessary, automatically figuring out if we should open a specific project file or the project window and fixing the project file path if it’s ambiguous.

function goUnreal
  set -l unreal_path $HOME/src/UnrealEngine/Engine/Binaries/Linux
  set -l tmp_file (mktemp)

  set -q UNREAL_LOG_FILE; and rm -f $UNREAL_LOG_FILE;
  set -U UNREAL_LOG_FILE (mktemp)

  echo "source ~/.bashrc" &gt; $tmp_file
  echo "PS1=\"[Unreal Shell] ~&gt; \"" &gt; $tmp_file

  if count $argv[1] &gt; /dev/null
    # Test if project file exists
    set -l project_file $argv[1]
    if test -e $project_file
      # Test if full path exists
      set project_file (pwd)/$project_file
      if test ! -e $project_file
        echo "Project file $project_file doesn't exist"
        return 1
      end
    end

    # Open specific uproject
    echo "$unreal_path/UE4Editor &gt;$UNREAL_LOG_FILE \"$project_file\" 2&gt;@1" &gt;&gt; $tmp_file
  else
    # Open project window
    echo "$unreal_path/UE4Editor &gt;$UNREAL_LOG_FILE 2&gt;@1" &gt;&gt; $tmp_file

    # Figure out the PID of the new editor instance that the project window spawned

    # Tokenize the string
    echo "editor_process=\$(ps -ejH | tail -n1)" &gt;&gt; $tmp_file
    # We only want the first token from the string so we just grab the first item here
    echo "editor_pid=( \$editor_process )" &gt;&gt; $tmp_file
    
    # Wait until the editor closes
    echo "while kill -0 \$editor_pid &gt;/dev/null 2&gt;@1; do sleep 1; done" &gt;&gt; $tmp_file
  end

  # Clean up and exit
  echo "rm -f $tmp_file" &gt;&gt; $tmp_file
  echo "exit" &gt;&gt; $tmp_file

  optirun bash --rcfile $tmp_file
end

Copy blendshape to similiar geometry script (Maya)

A quick script I wrote that helps you copy blendshapes from a target mesh to only the selected vertices of another mesh, which is useful if you have models with identical faces but different bodies (that are a single object).

For example, if you want to copy a face blendshape from a character onto a copy of the character with identical face geometry but different body topology, you would :

  1. Duplicate the character with different body topology.
  2. Select the original blendshape mesh that you want to transfer to the duplicate.
  3. Shift select the vertices on the duplicate character that you want to be affected (such as all the head vertices).
  4. Run my script.
  5. Blendshape >> Add the result to your skinned character.

I’m unsure if there is a better way of doing this, but this script worked for a simple cartoon character.

import maya.cmds as cmds
 
selection = cmds.ls(sl=True)
 
if len(selection) > 0:
    targetMesh = selection[-1]
    affectedVertices = selection[0:-1] # Assume the rest of the selection is made of vertices
    affectedMesh = cmds.listRelatives(cmds.listRelatives(affectedVertices[0], p=True), p=True)
    
    vtxSet = cmds.sets(affectedVertices, v=True, n="SelectionSet")
    print cmds.sets(vtxSet, q=True)
    
    # Move to origin
    cmds.move(32, 0, 0, affectedMesh, ws=True)
    cmds.move(32, 0, 0, targetMesh, ws=True)
    
    # Select target mesh and affected vertices
    cmds.select(cl=True)
    cmds.select(targetMesh)
    cmds.select(vtxSet, add=True)
    
    cmds.transferAttributes(transferPositions=1, transferNormals=1, transferUVs=2, transferColors=0, sourceUvSpace="map1", sampleSpace=3, searchMethod=0, flipUVs=0, colorBorders=1)
    
    cmds.hide(targetMesh)
    
    cmds.select(cl=True)
else:
    print "Select vertices to move and target mesh"