Skip to content

Latest commit

 

History

History
128 lines (109 loc) · 5.33 KB

README.md

File metadata and controls

128 lines (109 loc) · 5.33 KB
Raw

Vaspview 2

This is an experiment to modernize the old vaspview application for viewing VASP charge density files.

Installation

Dependencies

Vaspview 2 requires the following tools and libraries for building:

  • A C++ compiler.
  • cmake.
  • GLEW.
  • GLUT.
  • OpenGL libraries (GL and GLU).

On Ubuntu 20.04 you require

build-essential cmake libglew-dev freeglut3-dev libxi-dev libgl1-mesa-dev libglu1-mesa-dev

On Ubuntu 10.04 you can install everything required by installing the packages

build-essential cmake libglew1.5-dev freeglut3-dev libgl1-mesa-dev libglu1-mesa-dev libxi-dev libxmu-dev

On Redhat Enterprise Linux 5 + EPEL repo, you need (at least)

gcc-c++ cmake glew-devel freeglut-devel libXmu-devel

After that, make a directory for building, run cmake and make, e.g.

  • mkdir build
  • cd build
  • cmake -D CMAKE_BUILD_TYPE=Release ..
  • make

Assuming everything is successful, the binary will be found in BUILD_DIRECTORY/src/vaspview .

Development Projects

DONE

  • Previously vaspview used manual setting up of GL function pointers on Windows, with only lowest common denominator fallback on non-Windows platforms. This has been replaced with the GLEW library. This provides an easy way to set up all the entry points, as well as an easy way to query for available extensions at runtime rather than compile-time.
  • As a side-effect of the above, 3D textures are now enabled on all current hw, leading to MASSIVE speedup when moving the slice plane around.
  • Use GL_FLOAT instead of GL_DOUBLE for vertices and normals. It's accurate enough, and on everything but the latest generation hardware (e.g. NVIDIA Fermi and later) doubles are an order of magnitude slower.
  • Added pad fields so that the vertex structure is a multiple of 32 bytes in size (previously 24 with GL_FLOAT, 48 with GL_DOUBLE); this is supposedly better on Ati according to http://www.opengl.org/wiki/Vertex_Buffer_Object
  • Switch implementation language to C++98, allowing the replacement of CDynArray with std::string and std::vector, as appropriate, and replacing CHashTable with std::map (std::unordered_map in C++2011 would be even better, but it doesn't matter performance-wise). Also, switching object allocation and destruction to new/delete and classes with constructors and destructors simplifies the code a bit.
  • Optionally use Vertex Buffer Objects (VBO) for uploading isosurface data to the GPU. Time spent in ds3ViewIsoDrawNode dropped from 38% to 11%, nice!
  • Use CMAKE for building.
  • Enabled support for compressed 3D textures, but image quality was bad so it had to be removed.
  • Removed old EXT_paletted_textures code path, as no current hardware supports paletted textures: http://www.opengl.org/registry/specs/EXT/paletted_texture.txt . NVIDIA hardware as of the NV4x generation (2004) does not support this, ATI/AMD and Intel never really supported it. Managed to test this code path on an old Geforce 2MX, but it didn't work there either, maybe a bug in the codepath, or a bug in the driver; in any case it doesn't matter as the codepath is now removed.
  • Removed old 2D slice texture generation-on-CPU codepath, as all current hardware supports EXT_texture3D (part of OpenGL 1.2 core), and the fallback was unusably slow and buggy. Tested on GeForce 2MX and Radeon 9200, both ~10 year old HW in 2011, and both support 3D textures.
  • Fix a bunch of bugs wrt using uninitialized memory.

TODO

  • Triangle stripping? http://tomsdxfaq.blogspot.com/ claims stripping is an outdated optimization, as is http://home.comcast.net/~tom_forsyth/blog.wiki.html#Strippers (same guy, I think). OTOH rendering in vertex buffer ordering seems more important. http://home.comcast.net/~tom_forsyth/papers/fast_vert_cache_opt.html
  • CUDA experiment: Move the marching cubes algorithm to the GPU. CUDA SDK has an example implementing marching cubes, integrate that? Currently this is the biggest performance bottleneck, up to 46% spent in DS3IsoSurface::isoMake() when moving the isosurface value slider back and forth.
  • Replace vect.hh/vect.cc with Eigen?
  • Someone could test, and contribute, OS X and Windows support (wink). The code itself should all be relatively platform neutral (thanks to GLEW), so mainly some patches to the cmake build scripts would be needed. There is also an old src/win32/winmain.c from the original vasputil, this could perhaps be resurrected.
  • VBO rendering in batches to prevent exhausting GPU memory on older cards. Batch size should be about the size of the pre-T&L cache size on the GPU, which on slightly older cards (Geforce 6800) is apparently about ~64000 vertices.
  • Supposedly element rendering is faster if the indices are GL_UNSIGNED_SHORT rather than GL_UNSIGNED_INT. Since it's easy to have more than 2**16 vertices, this would imply batching and sorting the vertices etc.
  • http://www.sci.utah.edu/~bavoil/opengl/vbo/batching/
  • Replace builtin GLUT-based widget toolkit with Qt?
  • Better use of RAII, per 'good' C++ style. This is, unfortunately, tedious as currently object ownership and lifetimes are not very clear.