1

Structured Grids and Sparse Matrix Vector Multiplication on the Cell Processor

Sam WilliamsLawrence Berkeley National LabUniversity of California Berkeley

SWWilliams@lbl.gov

November 1, 2006

2

• Cell Architecture• Programming Cell• Benchmarks & Performance

– Stencils on Structured Grids– Sparse Matrix-Vector Multiplication

• Summary

Outline

3

Cell Architecture

4

Cell Architecture

• 3.2GHz, 9 Core SMP– One core is a conventional cache based PPC

– The other 8 are local memory based SIMD processors (SPEs)

• 25.6GB/s memory bandwidth (128b @ 1.6GHz) to XDR• 2 chip (16 SPE) SMP blades

DRAM_0

25.6 GB/sI/OMEM

SPU0

PPU0

SPU1

SPU2

SPU3

SPU4

SPU5

SPU6

SPU7

MEMI/O

SPU14

SPU15

SPU13

SPU12

SPU11

SPU10

SPU9

SPU8

PPU1

DRAM_1

25.6 GB/s

5

Memory Architectures

Conventional

Register File

MainMemory

ld reg, addr

Cache

ld reg, Local_Addr

DMA LocalAddr, GlobalAddr

Register File

LocalMemory

MainMemory

Three Level

6

SPE notes

• SIMD

• Limited dual issue– 1 float/ALU + 1 load/store/permute/etc… per cycle

• 4 FMA single precision datapaths, 6 cycle latency

• 1 FMA double precision datapath, 13 cycle latency

• Double precision instructions are not dual issued, and will stall all subsequent instruction issues by 7 cycles.

• 128b aligned loads/stores (local store to/from register file)– Must rotate to access unaligned data– Must permute to operate on scalar granularities

7

Cell Programming

• Modified SPMD (Single Program Multiple Data)– Dual Program Multiple Data (control + computation)

– Data access similar to MPI, but data is shared like pthreads

• Power core is used to:– Load/initialize data structures

– Spawn SPE threads

– Parallelize data structures

– Pass pointers to SPEs

– Synchronize SPE threads

– Communicate with other processors

– Perform other I/O operations

8

Processors Evaluated

Cell SPE X1E SSP Power5 Opteron Itanium2Architecture SIMD Vector Super Scalar Super Scalar VLIW

Frequency 3.2 GHz 1.13 GHz 1.9 GHz 2.2 GHz 1.4 GHz

GFLOP/s (double) 1.83 4.52 7.6 4.4 5.6

Cores used 8 4 (MSP) 1 1 1

Aggregate:L2 Cache - 2MB 1.9MB 1MB 256KB

L3 Cache - - 36MB - 3MB

DRAM Bandwidth 25.6 GB/s 34 GB/s 10+5 GB/s 6.4 GB/s 6.4 GB/s

GFLOP/s (double) 14.6 18 7.6 4.4 5.6

9

Stencil Operations on Structured Grids

10

Stencil Operations

• Simple Example - The Heat Equation– dT/dt = k2T

– Parabolic PDE on 3D discretized scalar domain

• Jacobi Style (read from current grid, write to next grid)– 8 FLOPs per point, typically double precision

– Next[x,y,z] = Alpha*Current[x,y,z] +

Beta*( Current[x-1,y,z] + Current[x+1,y,z] +

Current[x,y-1,z] + Current[x,y+1,z] +

Current[x,y,z-1] + Current[x,y,z+1] )– Doesn’t exploit the FMA pipeline well

– Basically 6 streams presented to the memory subsystem

• Explicit ghost zones bound grid

[x,y,z] [x+1,y,z]

[x,y-1,z]

[x,y,z-1][x,y+1,z]

[x,y,z+1]

[x-1,y,z]

11

Optimization - Planes• Naïve approach (cacheless vector machine) is to load 5

streams and store one.• This is 8 flops per 48 bytes

– memory limits performance to 4.2 GFLOP/s

• A better approach is to make each DMA the size of a plane– cache the 3 most recent planes (z-1, z, z+1)

– there are only two streams (one load, one store)

– memory now limits performance to 12.8 GFLOP/s

• Still must compute on each plane after it is loaded– e.g. forall Current_local[x,y] update Next_local[x,y]

– Note: computation can severely limit performance

12

Optimization - Double Buffering

• Add a input stream buffer and and output stream buffer (keep 6 planes in local store)

• Two phases (transfer & compute) are now overlapped• Thus it is possible to hide the faster of DMA transfer time

and computation time

13

Optimization - Cache Blocking

• Domains can be quite large (~1GB)• A single plane, let alone 6, might not fit in the local store• Partition domain into cache blocked slabs so that 6 cache

blocked planes can fit in the local store• Partitioning in the Y dimension maintains good spatial and

temporal locality• Has the added benefit that cache blocks are independent

and thus can be parallelized across multiple SPEs• Memory efficiency can be diminished as an intra grid ghost

zone is implicitly created.

14

• Instead of computing on pencils, compute on ribbons (4x2)• Hides functional unit & local store latency • Minimizes local store memory traffic• Minimizes loop overhead

• May not be beneficial / noticeable for cache based machines

Optimization - Register Blocking

15

Double Precision Results

• ~2563 problem• Performance model matches well with hardware• 5-9x performance of Opteron/Itanium/Power5• X1E ran a slight variant (beta hard coded

to be 1.0, not cache blocked)

7.35

3.91

1.100.80

1.40

0.00

1.00

2.00

3.00

4.00

5.00

6.00

7.00

8.00

Cache Blocked

GFLOPs

3.2GHz Cell X1E MSP* 1.9GHz Power5

2.2GHz Opteron 1.4GHz Itanium2

16

Optimization - Temporal Blocking

• If the application allows it, perform block the outer (temporal) loop

• Only appropriate on memory bound implementations– Improves computational intensity– Cell SP or Cell with faster DP

• Simple approach– Overlapping trapezoids in time-space plot – Can be inefficient due to duplicated work– If performing n steps, the local store must hold

3(n+1) planes

• Time Skewing is algorithmically superior, but harder to parallelize.

• Cache Oblivious is similar but implemented recursively.

Time tTime t+1

Time t+2

17

Temporal Blocking Results

• Cell is computationally bound in double precision (no benefit in temporal blocking, so only cache blocked shown)

• Cache machines show the average of 4 steps

of time skewing

7.35

1.10 1.10

1.70

0.00

1.00

2.00

3.00

4.00

5.00

6.00

7.00

8.00

Time Skewed

GFLOPs

3.2GHz Cell 1.9GHz Power5 2.2GHz Opteron 1.4GHz Itanium2

18

Temporal Blocking Results (2)

• Temporal blocking on Cell was implemented in

single precision (four step average)

• Others still use double precision

65

1.10 1.10 1.70

0.00

10.00

20.00

30.00

40.00

50.00

60.00

70.00

Time Skewed

GFLOPs

3.2GHz Cell (Single Precision) 1.9GHz Power5

2.2GHz Opteron 1.4GHz Itanium2

19

Sparse Matrix-Vector Multiplication

20

Sparse Matrix Vector Multiplication

• Most of the matrix entries are zeros, thus the non zero entries are sparsely distributed

• Dense methods compute on all the data, sparse

methods only compute the nonzeros (as only

they compute to the result) • Can be used for unstructured grid problems• Issues

– Like DGEMM, can exploit a FMA well– Very low computational intensity

(1 FMA for every 12+ bytes)– Non FP instructions can dominate– Can be very irregular– Row lengths can be unique and very short

21

Compressed Sparse Row

• Compressed Sparse Row (CSR) is the standard format– Array of nonzero values

– Array of corresponding column for each nonzero value

– Array of row starts containing index (in the above arrays) of first nonzero in the row

22

Optimization - Double Buffer Nonzeros

• Computation and Communication are approximately equally expensive

• While operating on the current set of nonzeros, load the next (~1K nonzero buffers)

• Need complex (thought) code to stop and restart a row between buffers

• Can nearly double performance

23

Optimization - SIMDization

• Row Padding– Pad rows to the nearest multiple of 128b

– Might requires O(N) explicit zeros

– Loop overhead still present

– Generally works better in double precision

• BCSR– Nonzeros are grouped into dense r x c blocks(sub matrices)

– O(nnz) explicit zeros are added

– Choose r&c so that it meshes well with 128b registers

– Performance can hurt especially in DP as computing on zeros is very wasteful

– Can hide latency and amortize loop overhead

24

Optimization - Cache Blocked Vectors• Doubleword DMA gathers from DRAM can be expensive• Cache block source and destination vectors• Finite LS, so what’s the best aspect ratio?• DMA large blocks into local store• Gather operations into local store

– ISA vs. memory subsystem inefficiencies– Exploits temporal and spatial locality within

the SPE

• In effect, the sparse matrix is explicitly blocked

into submatrices, and we can skip, or otherwise

optimize empty submatrices

• Indices are now relative to the cache block– half words– reduces memory traffic by 16%

25

Optimization - Load Balancing

• Potentially irregular problem• Load imbalance can severely hurt performance• Partitioning by rows is clearly insufficient• Partitioning by nonzeros is inefficient when

the matrix has few but highly variable nonzeros

per row.

• Define a cost function of number of row starts

and number of nonzeros.• Determine the parameters via static timing analysis or

tuning.

26

Other Approaches

• BeBop / OSKI on the Itanium2 & Opteron– uses BCSR – auto tunes for optimal r x c blocking– can also exploit symmetry– Cell implementation is base case

• Cray’s routines on the X1E– Report best of CSRP, Segmented Scan &

Jagged Diagonal

27

Benchmark Matrices

#15 - PDE

N=40K, NNZ=1.6M

#17 - FEM

N=22K, NNZ=1M

#18 - Circuit

N=17K, NNZ=125K

#36 - CFD

N=75K, NNZ=325K

#06 - FEM Crystal

N=14K, NNZ=490K

#09 - 3D Tube

N=45K, NNZ=1.6M

#25 - Financial

N=74K, NNZ=335K

#27 - NASA

N=36K, NNZ=180K

#28 - Vibroacoustic

N=12K, NNZ=177K

#40 - Linear Programming

N=31K, NNZ=1M

28

Double Precision SpMV PerformanceDiagonally Dominant

Notes:

• few nonzeros per row severely limited performance on CFD

• BCSR & symmetry were clearly exploited on first 2

• 3-8x faster than Opteron/Itanium

4.23

3.26

3.92

1.61

0.93 0.86

0.420.28

1.14 1.16

0.49

0.21

1.55 1.61

0.00

0.50

1.00

1.50

2.00

2.50

3.00

3.50

4.00

4.50

FEM Crystal 3D Tube FEM CFD

Matrix

GFLOP/s

3.2GHz Cell(CSR) 2.2GHz Opteron(OSKI)1.4GHz Itanium2(OSKI) X1E MSP(best of JAG,CSRP,ELL,…)

29

Double Precision SpMV Performanceless structured/partitionable

Notes:• few nonzeros per row (hurts a lot)• Not structured as well as previous set• 4-6x faster than Opteron/Itanium

1.77

1.46

1.61

0.300.37

0.42

0.27 0.240.32

0.57

0.00

0.20

0.40

0.60

0.80

1.00

1.20

1.40

1.60

1.80

2.00

Circuit Financial NASA

Matrix

GFLOP/s

3.2GHz Cell(CSR) 2.2GHz Opteron(OSKI)1.4GHz Itanium2(OSKI) X1E MSP(best of JAG,CSRP,ELL,…)

30

Double Precision SpMV Performancemore random

Notes:• many nonzeros per row• Load imbalance hurt cell’s performance by as much as 20%• 5-7x faster than Opteron/Itanium

3.35 3.42 3.48

0.440.57

0.470.460.56 0.63

0.84

0.00

0.50

1.00

1.50

2.00

2.50

3.00

3.50

4.00

PDE Vibroacoustic Linear Programming

Matrix

GFLOP/s

3.2GHz Cell(CSR) 2.2GHz Opteron(OSKI)1.4GHz Itanium2(OSKI) X1E MSP(best of JAG,CSRP,ELL,…)

31

Conclusions

32

Conclusions

• Even in double precision, Cell obtains much better performance on a surprising variety of codes.

• Cell can eliminate unneeded memory traffic, hide memory latency, and thus achieves a much higher percentage of memory bandwidth.

• Instruction set can be very inefficient for poorly SIMDizable or misaligned codes.

• Loop overheads can heavily dominate performance.

• Programming model could be streamlined

33

Acknowledgments

• This work is a collaboration with the following LBNL/FTG members:– John Shalf, Lenny Oliker, Shoaib Kamil, Parry

Husbands, Kaushik Datta and Kathy Yelick

• Additional thanks to– Joe Gebis and David Patterson

• X1E FFT numbers provided by:– Bracy Elton, and Adrian Tate

• Cell access provided by IBM

34

Questions?

35

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted, provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers, or to redistribute to lists, requires prior specific permission. GSPx 2006. October 30-November 2, 2006. Santa Clara, CA. Copyright 2006 Global Technology Conferences, Inc.