Slide 1

Warp-Aware Trace Scheduling for GPUS James Jablin (Brown) Thomas Jablin (UIUC) Onur Mutlu (CMU) Maurice Herlihy (Brown)

Slide 2

Historical Trends in GFLOPS: CPUs vs. GPUs Reproduced from NVIDIA C Programming Guide (Version 5.0)

Slide 3

Performance Pitfalls Control flow can negatively affect performance.

Slide 4

Performance Pitfalls Pipeline Stall execution delay in an instruction pipeline to resolve a dependency

Slide 5

Hardware: CPU versus GPU Reproduced from NVIDIA C Programming Guide (Version 5.0)

Slide 6

Slide 7

Slide 8

Slide 9

Slide 10

Slide 11

Slide 12

Slide 13

Slide 14

Slide 15

Slide 16

Slide 17

Slide 18

Performance Pitfalls Pipeline Stall execution delay in an instruction pipeline to resolve a dependency

Slide 19

Performance Pitfalls Pipeline Stall execution delay in an instruction pipeline to resolve a dependency Warp Divergence threads within a warp take different paths and the different execution paths are serialized

Slide 20

Warp Divergence Example

Slide 21

Slide 22

Slide 23

Slide 24

Slide 25

Slide 26

Slide 27

Slide 28

Warp-Aware Trace Scheduling Schedule instructions across basic block boundaries to expose additional ILP

Slide 29

Warp-Aware Trace Scheduling Schedule instructions across basic block boundaries to expose additional ILP while managing and optimizing warp divergence.

Slide 30

Origins: Microcode Trace Scheduling generalizing local and disparate vertical-to- horizontal microcode compaction StepDescription

Slide 31

Origins: Microcode Trace Scheduling generalizing local and disparate vertical-to- horizontal microcode compaction StepDescription 1. Trace Selection

Slide 32

Origins: Microcode Trace Scheduling generalizing local and disparate vertical-to- horizontal microcode compaction StepDescription 1. Trace Selection 2. Trace Formation

Slide 33

Origins: Microcode Trace Scheduling generalizing local and disparate vertical-to- horizontal microcode compaction StepDescription 1. Trace Selection 2. Trace Formation 3. Local Scheduling

Slide 34

Origins: Microcode Trace Scheduling generalizing local and disparate vertical-to- horizontal microcode compaction StepDescription 1. Trace SelectionPartition basic blocks into regions 2. Trace Formation Facilitate local scheduling, potentially adding nodes and edges 3. Local SchedulingSchedule instructions within each region

Slide 35

Slide 36

Slide 37

Slide 38

Slide 39

Slide 40

Slide 41

Slide 42

Slide 43

Slide 44

Slide 45

Slide 46

Slide 47

Slide 48

Slide 49

Slide 50

Slide 51

Backup Slides

Slide 52

Slide 53

GPU Programming Model

Slide 54

Slide 55

Slide 56

Characterizing the Grid

Slide 57

Characterizing the Grid, Block

Slide 58

Slide 59

Characterizing the Grid, Block, and Thread

Slide 60

Slide 61

Warp Divergence Examples

Slide 62

Slide 63

Slide 64

Slide 65