“Classic” GPGPU Prof. Aaron Lanterman School of Electrical and Computer Engineering Georgia Institute of Technology2 “Classic” vs. “Modern” GPGPU • “Classic GPGPU” – User must map their algorithm to a graphics framework not designed with general computation in mind – GPGPU implementations on the Playstation 3s RSX and the Xbox 360’s Xenos will need to be programmed in this mindset – Many PC users will have older graphics cards • “Modern GPGPU” – Runs on hardware (e.g., NVIDIA G80) specifically designed to be friendly to GPGPU computations – NVIDIA’s Compute Unified Device Architecture (CUDA) framework hides low-level details • See Hyesoon Kimʼs Spring 2008 class, CS8803SC: Software and Hardware Cooperative Computing – ATI’s CTM (Close-to-Metal) provides powerful low-level access to GPGPU-friendly hardware, but requires users to “roll their own”3 Typical classic GPGPU setup • Load inputs into textures – GPU textures <-> CPU arrays – Texture coordinates <-> computational domain • Draw a quadrangle • Do computation in pixel shaders – Older GPUs typically have more pixel shader units than vertex shader units – Older GPUs can’t do texture fetch in vertex shaders – GPU pixel shaders <-> CPU inner loops – Texture sample (tex2D) <-> CPU memory read • Render output as pixels in the quadrangle – Render-to-texture <-> CPU memory write – Vertex coordinates <-> computational range4 Reference • A lot of this discussion is inspired by various chapters in GPU Gems 2 – Excellent reference for “classic GPGPU” – GPU Gems 3 focuses more on “modern GPGPU” solutions5 Scattering vs. gathering on a typical GPU • Gather: – Indirect read a = x[i] • Texture fetch tex2D(x,...) • Scatter: – Indirect write a[i] = x • No arbitrary texture writes on the GPU! – Can do “ordered writes” with “render to texture”6 Mapping • Apply some function to each element in parallel • Store input values as textels in a texture • Draw quadrangle with as many pixels as textels in input • Pixel shader just computes the function • Output values in pixels of rendered quadrangle – Use “render-to-texture” to use mapped values in another stage7 Reduction • Associative operations – Ex: sums, products, min, max • Do multiple passes, rendering to smaller textures with each pass 72 15 15 32 64 102 8 301 134 47 218 52 26 19 38 98 83 8 19 38 8 Min example • Could do more than four per pixel – Depending on how many “texture fetches” are allowed – Less passes, but increased time per pass8 Flow control is tricky • Most older GPUs don’t “really” support branching • Loops typically “unrolled” by the compiler • Predicated branching: – Compute both parts of the branch – Use results from only one branch • On some GPUs, vertex shaders support branching but pixel shaders don’t • Even newer GPUs that directly support branching may give you a significant performance hit – Parallel execution units may be restricted to executing only one branch at a time (locality is important)9 Static branch resolution • Instead of branching in the pixel shader, execute different pixel shaders on different output quadrangles • Ex: boundary conditions in PDE Info & examples from p. 549-550 of GPU Gems 210 Skipping unnecesary work with Z-cull • Setup – Write 0 to z-buffer of pixels where you want to skip computation – Write 1 to z-buffer of pixels where you do • Feed pixel shader doing the computation z-buffer value of 0.5 • Ex: landlocked cells in fluid simulation • Warning: GPU may do z-culling at coarser resolution than the pixels – Will skip shading only if all pixels in a region fail the depth test • Can do similar tricks with alpha stenciling Info & examples from p. 550-552 of GPU Gems 211 Applications tailor-made for GPGPU • Applications with high compute-to-communication ratios • Partial differential equations • Cellular automata – Sim City! • Finance – Options pricing • Linear algebra – Chapter 44 of GPU Gems 2 – There are clever ways to handle banded matrices, sparce matrices, etc.12 GPGPU applications in games • Collision detection • Physics – Rigid bodies – Fluids, clouds, smoke, cloth • Particle systems • Line of sight calculations for AI? • Aside: – AGEIA marketed Physics Processing Unit (PPU) to accelerate their “PhysX” SDK • But now owned by NVIDIA – Havok FX can exploit NVIDIA and ATI GPUs13 General advice • Quite often better to store long 1D arrays as “wrapped” 2D arrays – 1D arrays limited in length – GPUs seem to be faster at handling 2D textures than 1D • Pre-compute constants on the CPU • Pre-compute low-dimensional functions and store them as textures – GPU will naturally do an interpolated lookup14 Limitations of GPGPU • Usually 32-bit floats (used to be worse!) – Some newer high-end cards support 64-bit floats • May not have integers – Can be a problem with precise texture accesses • May not have bitwise operations • Not-so-great with dynamic data structures (queues, stacks, trees, etc.) – But lots of clever people have come up with tricks • Think outside the box: strange algorithms that might be silly on a CPU might map well to a GPU15 Some successful GPGPU applications • Computed Tomography (CT) • MRI (application of FFTs) • Phase unwrapping for Synthetic Aperture Radar (SAR) – 35x speedup obtained by Peter Karasev, Dan Campbell, and Mark Richards • Data mining • Raytracing • GPU-accelerated Photoshop!16 Quake 3: Raytraced Images from graphics.cs.uni-sb.de/~sidapohl/egoshooter
View Full Document
Unlocking...