GPU Computing with Ruby

GPU Computing with Ruby

• 1.
  GPU Computing withRuby SpeedGo Computing Chung Shin Yee shinyee@speedgocomputing.com
• 2.
  CPU vs GPUArchitecture 6 Core vs 1024 Core 6 GB/s vs 300 GB/s Memory Bandwidth By CUDA C Programming Guide
• 3.
  CUDA Programming Model . . . . By CUDA C Programming Guide
• 4.
  Existing Programming Tools ● Cg ● BrookGPU ● GLSL (OpenGL Shading Language) ● Nvidia CUDA C/C++ ● OpenCL ● PyCUDA Where is the Red Ruby ?
• 5.
  Bridging Ruby &CUDA C/C++ ● Ruby C extension – Hard to manipulate Ruby objects in C. – Compilation problems. ● Ruby FFI – Bridging purely in Ruby. – Support multiple Ruby implementations.
• 6.
  Ruby Bridge Sample
• 7.
  Developing SGC RubyCUDA ● Object-oriented API. ● Start with crucial operations. – Memory allocation. – Memory transfer. – Kernel launch. – Wrapper for structures. ● Documented with YARD.
• 8.
  Driver vs RuntimeAPI ● CUDA Driver API – For system developers. – Supported by PyCUDA. ● CUDA Runtime API – For computation centric developers. We going to support both API !
• 9.
  Using SGC RubyCUDA ● Kernel program in CUDA C.
• 10.
  Using SGC RubyCUDA ● Compiling kernel into PTX. – nvcc --ptx vadd.cu
• 11.
  Using SGC RubyCUDA ● Setup require 'rubycu' include SGC::CU CUInit.init d = CUDevice.get(0) c = CUContext.create(d) m = CUModule.new.load(“vadd.ptx”) f = m.function(“vadd”)
• 12.
  Using SGC RubyCUDA ● Memory allocations da = CUDevice.malloc(10*4) db = CUDevice.malloc(10*4) dc = CUDevice.malloc(10*4) ha = Buffer.new(:int, 10) hb = Buffer.new(:int, 10) hc = Buffer.new(:int, 10)
• 13.
  Using SGC RubyCUDA ● Initialization (0...10).each { |i| ha[i] = i hb[i] = 1 hc[i] = ha[i] + hb[i] hd[i] = 0 }
• 14.
  Using SGC RubyCUDA ● Transfer inputs to the GPU CUMemory.memcpy_htod(da, ha, 4*10) CUMemory.memcpy_htod(db, hb, 4*10) CUMemory.memcpy_htod(dc, hc, 4*10)
• 15.
  Using SGC RubyCUDA ● Launch kernel on GPU # Launch with 1x1x1 grid, # 10x1x1 blocks, params = [da, db, dc, 10] f.launch_kernel(1, 1, 1, 10, 1, 1, 0, 0, params) By CUDA C Programming Guide By CUDA C Programming Guide
• 16.
  Using SGC RubyCUDA ● Transfer results back to system memory CUMemory.memcpy_dtoh(hd, dc, 4*10) ● Verify results (0...10).each { |i| assert_equal(hc[i], hd[i]) }
• 17.
  Problematic CUDA RuntimeAPI ● For use in a CUDA C/C++ program. ● Workaround – CUDA C/C++ effectively uses C/C++ bindings. – Create dynamic library for the kernel programs. – Load the library at runtime.
• 18.
  Current Limitations ● Support limited data types. – Fixnum → int – ?? → long – Float → float – ?? → double ● No supports for CUDA C++ templates. ● No Ruby in a kernel program.
• 19.
  To Support ● Texture memory. ● New features in CUDA 4.0 – Multi-GPU. – Unified Virtual Memory. ● More C data types. ● Mac platform.
• 20.
  Try It Now!Thank You ~ git clone git://github.com/xman/sgc-ruby-cuda.git cd sgc-ruby-cuda gem install ffi yard rake test rake yard