LAPACK on the NVIDIA G80 Processor

1
LAPACK on the NVIDIA G80 Processor

• Robert Liao
• Tracy Wang
• CS252 Spring 2007

2
Overview

• Traditional GPU Architecture
• The NVIDIA G80 Processor
• CUDA (Compute Unified Device Architecture)
• LAPACK
• Performance and Issues

3
A Quick Note on Naming

• G80 is the codename for the GPU found in the
  following graphics cards.
• NVIDIA GeForce 8 Series Graphics Cards
• NVIDIA Quadro FX 4600
• NVIDIA Quadro FX 5600

4
Traditional GPUs
From Intel Corporation
5
Traditional GPUs

• GPUs talk Polygons

Vertex Processor
Pixel Fragmenting Creation
Merge Output
Process Fragments
From CPU
Display
6
Traditional GPUs

• OpenGL and DirectX abstract this away.

Vertex Processor
Pixel Fragmenting Creation
Merge Output
Process Fragments
From CPU
Display
7
The NVIDIA G80 Architecture

• Reconfigurable Processor Pipeline

From NVIDIA
8
G80 History and Specifications

• Project Started in Summer of 2002.
• 128 Compute Cores
• 1.35 GHz in the GeForce 8800
• Floating Point Ops
• Stream Processor Architecture
• One Computing Unit Streams into another
  Computing Unit

9
The CUDA Interface to the G80

• Compute Unified Device Architecture
• C Interface for Performing Operations on the
  NVIDIA Processor
• Contains traditional C memory semantics with the
  context of a GPU

10
Working with CUDA

• Custom compiler provided to compile C code that
  the GPU can understand.
• The API functions provide a whole host of ways to
  interface with the GPU.
• CUDA Libraries are provided for common tasks.
• CUDA Runtime helps management of memory
• No DirectX or OpenGL knowledge needed!

11
Working with CUDA

• Running C on the CPU

• Running C on the GPU

• malloc
• free
• CPU Code

• cudaMalloc
• cudaFree
• GPU Code

Pointers on one side stay on one side. This will
create issues for existing applications
12
LAPACK

• Linear Algebra PACKage
• Implemented in Fortran 77
• Interfaces with BLAS (Basic Linear Algebra
  Subprograms)
• Professor James Demmel involved in Project

13
CLAPACK

• An F2Ced version of LAPACK.
• Very ugly!
• s_rsle(io___8)
• do_lio(c__3, c__1, (char )nm,
  (ftnlen)sizeof(integer))
• e_rsle()
• if (nm lt 1)
• s_wsfe(io___10)
• do_fio(c__1, " NM ", (ftnlen)4)
• do_fio(c__1, (char )nm, (ftnlen)sizeof(integer
  ))
• do_fio(c__1, (char )c__1, (ftnlen)sizeof(integ
  er))
• e_wsfe()
• nm 0
• fatal TRUE_
• else if (nm gt 12)
• s_wsfe(io___11)
• do_fio(c__1, " NM ", (ftnlen)4)
• do_fio(c__1, (char )nm, (ftnlen)sizeof(integer
  ))
• do_fio(c__1, (char )c__12, (ftnlen)sizeof(inte
  ger))
• e_wsfe()

14
CUBLAS

• NVIDIAs CUDA Based Implementation of BLAS
• Many functions are similar, but argument
  signatures are slightly different
• Adds some other functions as well
• cublasAlloc
• cublasFree
• CUBLAS lives in the GPU world

15
CLAPACK and CUBLAS

• Putting them together is not as easy as just
  linking CLAPACK to CUBLAS.
• Matrices and data structures must be moved into
  GPU memory space.
• CLAPACK executes on the CPU.
• CUBLAS executes on the GPU.

CLAPACK Function
CUBLAS
Memory copy CPU-gtGPU
Memory copy GPU-gtCPU
16
CLAPACK Concentration

• General Solve
• sgesv
• Computes solution to linear system of equationsA
  X B
• To Solve, A is factored into three matrices, P,
  L, and U.
• P Permutation Matrix
• L Lower Triangular
• U Upper Triangular
• Currently, our results cover the triangular
  factoring step

17
Performance Results
18
Performance Results
19
Performance Issues

• Much copying must be done from the CPU to GPU and
  GPU to CPU to communicate results.
• Why not convert all pointers into GPU pointers?
• Requires CLAPACK to run in GPU memory.
• Could be someones research paper

20
Other Issues

• Floating Point Behaves Differently
• Section 5.2 of the CUDA Programming Guide
  Discusses Deviations from IEEE-754
• No support for denormalized numbers
• Underflowed numbers are flushed to zero
• We noticed some results appearing as 0.0001
  instead of 0, for example

21
Current State

• Investigating some interesting memory issues on
  the GPU side.
• Allocations Mysteriously Fail.

22
Conclusions To Date

• Small data sets are better left off on the CPU.
• GPU calculations may not be appropriate for
  scientific computing depending on needs.

23
Future Directions

• Moving all of LAPACK into GPU
• Resolving the copying issue
• Perhaps resolved by unifying the CPU and GPU?
• Want to give it a try?
• Cant find Quadro FX 5600 on Market (MSRP 2,999)
• GeForce 8 Series have the G80 Processor
• GeForce 8500GT (99.99)
• GeForce 8800GTX (939.99)

24
Questions