accelerating matlab image processing toolbox functions on gpus

Accelerating MATLAB Image Processing Toolbox Functions on GPUs

Jingfei Kong, Martin Dimitrov, Yi Yang, Janaka Liyanage, Lin Cao, Jacob Staples, Mike Mantor,

Huiyang Zhou

University of Central Florida 2

Motivation

• With high memory bandwidth and teraflops computing capability, Graphics Processor Units (GPUs) become quite attractive for accelerating general purpose applications

• Developing high-performance GPU programs, however, requires deep understanding of both application algorithms and GPU hardware architecture

• A systematic way of dealing with a generic class of applications is missing

University of Central Florida 3

Our Contributions

• Compare performance-critical hardware features in different GPUs

• Develop high-quality open-source library code for some representative functions in MATLAB™ Image Processing Toolbox (IPT)– https://sites.google.com/site/iptatiproject/ [15]

• Reveal insights on efficiently accelerating a wide range of image processing algorithms

University of Central Florida 4

Presentation Outline

• Motivation• Our Contributions• Implication of GPU hardware on GPGPU programming• A GPGPU library for IPT functions

– categorization and optimization strategies• Case Studies

– 2D convolution– dither

• Conclusions

University of Central Florida 5

Implication of GPU hardware on GPGPU programmingPerformance-Critical Hardware Features

Implication on GPGPU Programs AMD/ATI HD5870 NVIDIA GTX280

Memory Access Bandwidth

Vector-type (float2 or float4) data access

Scalar-type (float) or vector-type (float2) data access

Register File A high number of registers (1k float4 registers or 16kB) per core implies more computational work in each core

A relatively small number of registers (2k float registers or 8kB) per core implies less computational work in each core.

Shared Memory/Local Data Share

Higher amount of shared memory (32kB per SM) implies large tile sizes (one tile is the workload of one thread block or work group)

Smaller amount of shared memory (16kB per SM) implies small tile sizes (one tile is the workload of one thread block or work group)

Ratio of (Peak Computation Throughput / Peak Memory Bandwidth)

(2.72 TFLOPS)/(154 GB/s) means more computation needs to be performed for each loaded data item

(0.62 TFLOPS)/(141 GB/s) means a relatively small amount of computation needs to be performed for each loaded data item

Stream processor pipeline

5-way VLIW ILP needed to make ALUs busy

Scalar pipeline less ILP needed

University of Central Florida 6

Implication of GPU hardware on GPGPU programmingPerformance-Critical Hardware Features

Implication on GPGPU Programs

AMD/ATI HD5870 (RV870)

NVIDIA GTX280

Memory Access Bandwidth

Vector-type (float2 or float4) data access

Scalar-type (float) or vector-type (float2) data access

Register File A high number of registers (1k float4 registers or 16kB) per core implies more computational work in each core

A relatively small number of registers (2k float registers or 8kB) per core implies less computational work in each core.

Shared Memory/Local Data Share

Higher amount of shared memory (32kB per SM) implies large tile sizes (one tile is the workload of one thread block or work group)

Smaller amount of shared memory (16kB per SM) implies small tile sizes (one tile is the workload of one thread block or work group)

Ratio of (Peak Computation Throughput / Peak Memory Bandwidth)

(2.72 TFLOPS)/(154 GB/s) means more computation needs to be performed for each loaded data item

(0.62 TFLOPS)/(141 GB/s) means a relatively small amount of computation needs to be performed for each loaded data item

Stream processor pipeline

5-way VLIW ILP needed to make ALUs busy

Scalar pipeline less ILP needed

Memory Access Bandwidth

Vector-type (float2 or float4) data access

Scalar-type (float) or vector-type (float2) data access

Our experiments show:

Using float instead would reduce bandwidth by at least 10%

Using float4 instead would reduce bandwidth by at least 16%

University of Central Florida 7

Implication of GPU hardware on GPGPU programmingPerformance-Critical Hardware Features

Implication on GPGPU Programs

AMD/ATI HD5870 (RV870)

NVIDIA GTX280

Memory Access Bandwidth

Vector-type (float2 or float4) data access

Scalar-type (float) or vector-type (float2) data access

Register File A high number of registers (1k float4 registers or 16kB) per core implies more computational work in each core

A relatively small number of registers (2k float registers or 8kB) per core implies less computational work in each core.

Shared Memory/Local Data Share

Higher amount of shared memory (32kB per SM) implies large tile sizes (one tile is the workload of one thread block or work group)

Smaller amount of shared memory (16kB per SM) implies small tile sizes (one tile is the workload of one thread block or work group)

Ratio of (Peak Computation Throughput / Peak Memory Bandwidth)

(2.72 TFLOPS)/(154 GB/s) means more computation needs to be performed for each loaded data item

(0.62 TFLOPS)/(141 GB/s) means a relatively small amount of computation needs to be performed for each loaded data item

Stream processor pipeline

5-way VLIW ILP needed to make ALUs busy

Scalar pipeline less ILP needed

Register File

256KB per SIMD engine * 20 SIMD engines = 5MB in total

64KB per SM * 30 SMs = 1.875MB in total

University of Central Florida 8

Implication of GPU hardware on GPGPU programmingPerformance-Critical Hardware Features

Implication on GPGPU Programs

AMD/ATI HD5870 (RV870)

NVIDIA GTX280

Memory Access Bandwidth

Vector-type (float2 or float4) data access

Scalar-type (float) or vector-type (float2) data access

Register File A high number of registers (1k float4 registers or 16kB) per core implies more computational work in each core

A relatively small number of registers (2k float registers or 8kB) per core implies less computational work in each core.

Shared Memory/Local Data Share

Higher amount of shared memory (32kB per SM) implies large tile sizes (one tile is the workload of one thread block or work group)

Smaller amount of shared memory (16kB per SM) implies small tile sizes (one tile is the workload of one thread block or work group)

Ratio of (Peak Computation Throughput / Peak Memory Bandwidth)

(2.72 TFLOPS)/(154 GB/s) means more computation needs to be performed for each loaded data item

(0.62 TFLOPS)/(141 GB/s) means a relatively small amount of computation needs to be performed for each loaded data item

Stream processor pipeline

5-way VLIW ILP needed to make ALUs busy

Scalar pipeline less ILP needed

Shared memory/ local data share

32KB per SIMD engine * 20 SIMD engines = 640KB in total

16KB per SM * 30 SMs = 480KB in total

University of Central Florida 9

Implication of GPU hardware on GPGPU programmingPerformance-Critical Hardware Features

Implication on GPGPU Programs

AMD/ATI HD5870 (RV870)

NVIDIA GTX280

Memory Access Bandwidth

Vector-type (float2 or float4) data access

Scalar-type (float) or vector-type (float2) data access

Register File A high number of registers (1k float4 registers or 16kB) per core implies more computational work in each core

A relatively small number of registers (2k float registers or 8kB) per core implies less computational work in each core.

Shared Memory/Local Data Share

Higher amount of shared memory (32kB per SM) implies large tile sizes (one tile is the workload of one thread block or work group)

Smaller amount of shared memory (16kB per SM) implies small tile sizes (one tile is the workload of one thread block or work group)

Ratio of (Peak Computation Throughput / Peak Memory Bandwidth)

(2.72 TFLOPS)/(154 GB/s) means more computation needs to be performed for each loaded data item

(0.62 TFLOPS)/(141 GB/s) means a relatively small amount of computation needs to be performed for each loaded data item

Stream processor pipeline

5-way VLIW ILP needed to make ALUs busy

Scalar pipeline less ILP needed

Ratio of Peak Computation Throughput / Peak Memory Bandwidth

(2720 GFLOPS)/(154 GB/s) = 17.7 flop/B= 70.6 flop/word

(624 GFLOPS)/(141 GB/s) = 4.4 flop/B= 17.7 flop/word

University of Central Florida 10

Implication of GPU hardware on GPGPU programmingPerformance-Critical Hardware Features

Implication on GPGPU Programs

AMD/ATI HD5870 (RV870)

NVIDIA GTX280

Memory Access Bandwidth

Vector-type (float2 or float4) data access

Scalar-type (float) or vector-type (float2) data access

Register File A high number of registers (1k float4 registers or 16kB) per core implies more computational work in each core

A relatively small number of registers (2k float registers or 8kB) per core implies less computational work in each core.

Shared Memory/Local Data Share

Higher amount of shared memory (32kB per SM) implies large tile sizes (one tile is the workload of one thread block or work group)

Smaller amount of shared memory (16kB per SM) implies small tile sizes (one tile is the workload of one thread block or work group)

Ratio of (Peak Computation Throughput / Peak Memory Bandwidth)

(2.72 TFLOPS)/(154 GB/s) means more computation needs to be performed for each loaded data item

(0.62 TFLOPS)/(141 GB/s) means a relatively small amount of computation needs to be performed for each loaded data item

Stream processor pipeline

5-way VLIW ILP needed to make ALUs busy

Scalar pipeline less ILP needed

Stream processor pipeline

5-way VLIW Scalar pipeline

University of Central Florida 11

Summary of the LibraryMATLAB Image Processing Toolbox (IPT) Function Classification

Function Category Function Name Function Description

(A) Data independent

intlut Convert integer values using lookup tableimadjust Adjust image intensity valuesimlincomb Linear combination of images

(B) Data sharing

edge Find edges in grayscale imageimregionalmax Regional maxima of an imageordfilt2 2-D order-statistic filtering conv2 2D convolution of an imagemean2 Average of matrix elementsimdilate/imerode Dilate/erode a grayscale image

(C) Algorithm dependent

bwdist Euclidean distance transform of a binary imageradon Radon transform

(D) Data dependent

dither Represent grayscale images in binary format

University of Central Florida 12

MATLAB IPT Function Classification and Optimization Strategies

Function Category Function Name Function Description

(A) Data independent

intlut Convert integer values using lookup tableimadjust Adjust image intensity valuesimlincomb Linear combination of images

(B) Data sharing

edge Find edges in grayscale imageimregionalmax Regional maxima of an imageordfilt2 2-D order-statistic filtering conv2 2D convolution of an imagemean2 Average of matrix elementsimdilate/imerode Dilate/erode a grayscale image

(C) Algorithm dependent

bwdist Euclidean distance transform of a binary imageradon Radon transform

(D) Data dependent

dither Represent grayscale images in binary format

Data Independent

Strategies:effectively utilize bandwidth by packing multiple pixels,

perform multiple such light-weight tasks if possible to amortize the CPU-GPU data transfer overhead

Characteristics: straightforward one on one mapping, abundant

parallelism

University of Central Florida 13

MATLAB IPT Function Classification and Optimization Strategies

Function Category Function Name Function Description

(A) Data independent

intlut Convert integer values using lookup tableimadjust Adjust image intensity valuesimlincomb Linear combination of images

(B) Data sharing

edge Find edges in grayscale imageimregionalmax Regional maxima of an imageordfilt2 2-D order-statistic filtering conv2 2D convolution of an imagemean2 Average of matrix elementsimdilate/imerode Dilate/erode a grayscale image

(C) Algorithm dependent

bwdist Euclidean distance transform of a binary imageradon Radon transform

(D) Data dependent

dither Represent grayscale images in binary format

Data sharing

Strategies:data reuse, computation reuse

Characteristics: still one on one mapping, but there is an overlapping

over input pixels for computing adjacent output pixel

University of Central Florida 14

MATLAB IPT Function Classification and Optimization Strategies

Function Category Function Name Function Description

(A) Data independent

intlut Convert integer values using lookup tableimadjust Adjust image intensity valuesimlincomb Linear combination of images

(B) Data sharing

edge Find edges in grayscale imageimregionalmax Regional maxima of an imageordfilt2 2-D order-statistic filtering conv2 2D convolution of an imagemean2 Average of matrix elementsimdilate/imerode Dilate/erode a grayscale image

(C) Algorithm dependent

bwdist Euclidean distance transform of a binary imageradon Radon transform

(D) Data dependent

dither Represent grayscale images in binary format

Algorithm dependent

Strategies:re-think algorithms, explore inherent

parallelism

Characteristics: lack of explicit parallelism

University of Central Florida 15

MATLAB IPT Function Classification and Optimization Strategies

Function Category Function Name Function Description

(A) Data independent

intlut Convert integer values using lookup tableimadjust Adjust image intensity valuesimlincomb Linear combination of images

(B) Data sharing

edge Find edges in grayscale imageimregionalmax Regional maxima of an imageordfilt2 2-D order-statistic filtering conv2 2D convolution of an imagemean2 Average of matrix elementsimdilate/imerode Dilate/erode a grayscale image

(C) Algorithm dependent

bwdist Euclidean distance transform of a binary imageradon Radon transform

(D) Data dependent

dither Represent grayscale images in binary format

DataDependent

Strategies:give it a shot and you might have some surprise

Characteristics: lack of explicit parallelism, sequential nature with

data dependency and fine-grain communication requirements

University of Central Florida 16

Summary of the LibraryPerformance Comparison against MATLAB CPU (single-threaded)

Function Category

Function Name Kernel Speedup on GTX 280 Kernel Speedup on HD5870

CUDA OpenCL OpenCL

(A) Data independent

intlut 17.7 17.5 12.7imadjust 21.4 15.7 11.9imlincomb 944.6 593.7 1385.4

(B) Data sharing

edge 3385.9 1175.2 4955.1imregionalmax 2117.8 798.4 3694.0ordfilt2 1199.6 171.6 1727.1conv2 345.5 156.9 649.8mean2 50.5 25.2 34.7imdilate/imerode 951.5 523.3 1579.8

(C) Algorithm dependentbwdist 134.8 126.2 104.3radon 84.3 67.4 61.2

(D) Data dependent dither 10.2 6.5 7.6

University of Central Florida 17

Summary of the LibraryPerformance Comparison against MATLAB CPU (single-threaded)

Function Category

Function Name Kernel Speedup on GTX 280 Kernel Speedup on HD5870

CUDA OpenCL OpenCL

(A) Data independent

intlut 17.7 17.5 12.7imadjust 21.4 15.7 11.9imlincomb 944.6 593.7 1385.4

(B) Data sharing

edge 3385.9 1175.2 4955.1imregionalmax 2117.8 798.4 3694.0ordfilt2 1199.6 171.6 1727.1conv2 345.5 156.9 649.8mean2 50.5 25.2 34.7imdilate/imerode 951.5 523.3 1579.8

(C) Algorithm dependentbwdist 134.8 126.2 104.3radon 84.3 67.4 61.2

(D) Data dependent dither 10.2 6.5 7.6

Geometric mean

206x 110x 218x

Kernel speedup on GTX 280

Kernel speedup onHD 5870

CUDA OpenCL OpenCL

University of Central Florida 18

Summary of the LibraryPerformance Comparison against MATLAB CPU (single-threaded)

Function Category

Function Name Kernel Speedup on GTX 280 Kernel Speedup on HD5870

CUDA OpenCL OpenCL

(A) Data independent

intlut 17.7 17.5 12.7imadjust 21.4 15.7 11.9imlincomb 944.6 593.7 1385.4

(B) Data sharing

edge 3385.9 1175.2 4955.1imregionalmax 2117.8 798.4 3694.0ordfilt2 1199.6 171.6 1727.1conv2 345.5 156.9 649.8mean2 50.5 25.2 34.7imdilate/imerode 951.5 523.3 1579.8

(C) Algorithm dependentbwdist 134.8 126.2 104.3radon 84.3 67.4 61.2

(D) Data dependent dither 10.2 6.5 7.6

Function name

CUDA on GTX 280

OpenCL on HD 5870

imlincomb 944.6 1385.4edge 3385.9 4955.1imregionalmax

2117.8 3694.0

ordfilt2 1199.6 1727.1conv2 345.5 649.8imdilate 951.5 1579.8

University of Central Florida 19

Summary of the LibraryPerformance Comparison against MATLAB CPU (single-threaded)

Function Category

Function Name Kernel Speedup on GTX 280 Kernel Speedup on HD5870

CUDA OpenCL OpenCL

(A) Data independent

intlut 17.7 17.5 12.7imadjust 21.4 15.7 11.9imlincomb 944.6 593.7 1385.4

(B) Data sharing

edge 3385.9 1175.2 4955.1imregionalmax 2117.8 798.4 3694.0ordfilt2 1199.6 171.6 1727.1conv2 345.5 156.9 649.8mean2 50.5 25.2 34.7imdilate/imerode 951.5 523.3 1579.8

(C) Algorithm dependentbwdist 134.8 126.2 104.3radon 84.3 67.4 61.2

(D) Data dependent dither 10.2 6.5 7.6

Geometric mean

206x 110x

Kernel speedup on GTX 280

CUDA OpenCL

University of Central Florida 20

2D Convolution Overview

1 2

4 5

9

6

3

7 8

1 1

1 1

1

1

2

2 1

3 x 3 filter

55

input pixels output pixels

University of Central Florida 21

2D Convolution Overview

• Drag the filter over the each pixel of the source image and multiply and accumulate the overlapped input elements to generate an output pixel.

Input Image

filter

pixel

University of Central Florida 22

2D Convolution: Intra-Thread Data Reuse

• Each thread computes multiple pixels along the column

• Intra-Thread reuse: – For a 7x7 filter we reuse each input pixel up to 7 times

22

Input Image

Thread iThread iThread i

University of Central Florida 23

2D Convolution: Inter-Thread Data Reuse

• Threads in the same warp/wavefront access the same row. • Inter-thread reuse

– The row is fetched into texture cache/shared memory and reused by different threads on subsequent accesses.

Input Image

0 1

Reused row in texture cache/shared memory

2 3threads

University of Central Florida 24

2D Convolution Performance

A 4096 x 4096 image with a 7 x 7 filter

• Jacket ‘s: – around 20 GFLOPS on GTX 280– Jacket 1.2.2 trial version (released on 1/4/2010) from

Accelereyes®

• Ours:– around 350 GFLOPS on GTX 280– around 733 GFLOPS on HD 5870

University of Central Florida 25

Data Dependent Case Study: Dither

University of Central Florida 26

Dither

0/1?

input pixels

230 < 128?

1230error

Error = 230 – 128 = 102output pixels

University of Central Florida 27

Dither – Data Dependency

j

i i+j

pixel at (i, j)

University of Central Florida 28

Dither – Parallel Processing Schedule

...1 2 4 53 6 7

3 4 6 75 9 10

5 6 97 10 11 12

7 10 119 12 13 14

9 10 12 1311 14 15 16

11 12 14 1513 16 17 18

13 14 16 1715 18 19 20

15 16 18 1917 20 21 22

8

8

8

8

From P. Metaxas [8]

...

University of Central Florida 29

Dither – Our GPU Implementation

12 3 4 5

45 6 7

8

78 9 10

11

1011 12 13

14

12 3

4

4

5

5

A relatively small amount of thread blocks/threads are active at any given time

• low resource utilization• synchronization overhead (among thread blocks/threads)

We still get up to 10.3x kernel speedup and 3.5x overall speedup!

University of Central Florida 30

Conclusions

• We identify performance-critical hardware features for GPGPU programs

• We present our experience and optimization strategies in developing high performance GPU code for functions from MATLAB Image Processing Toolbox

University of Central Florida 31

Our Open-source Library Project Website

https://sites.google.com/site/iptatiproject/ [15]You are more than welcome to contribute!

Thank you and Questions?