Liquid%20SIMD:%20Abstracting%20SIMD%20Hardware%20Using%20Lightweight%20Dynamic%20Mapping

1
Liquid SIMD Abstracting SIMD Hardware Using
Lightweight Dynamic Mapping

• Nathan Clark, Amir Hormati, Scott Mahlke,
• Sami Yehia, Krisztián Flautner
• University of Michigan ARM
  Ltd.

2
Computational Efficiency

• Low power envelope
• More useful work/transistors
• Hardware accelerators
• Niagara II encryption engine

Source AMD Analyst Day 12/14/06
3
How Are Accelerators Used?

• Control statically placed in binary

4
Problem With Static Control
CPU

• Not forward/backward compatible

5
Solution Virtualization

• Statically identify accelerated computation
• Abstract accelerator features
• Dynamically retarget binary

6
Liquid SIMD

• Virtualize SIMD accelerators
• Why virtualize SIMD?
• Intel MMX to SSE2
• ARM v6 to Neon
• Wide vectors useful Lin 06

7
SIMD Accelerator Assumptions
SIMD Exec
Fetch
Decode
Retire
Scalar Exec

• Same instruction stream
• Separate pipeline memory interface

8
How to Virtualize

• Use scalar ISA to represent SIMD operations
• Compatibility, low overhead
• Key easy to translate

Program
Branch
9
Virtualization Architecture
10
1. Data Parallel Operations
for(i 0 i lt 8 i) r1 Ai r2
Bi r3 r1 r2 r4 r3 constant
Ci r4
C
11
1a. What If Theres No Scalar Equivalent?
for(i 0 i lt 8 i) r1 Ai r2
Bi r3 r1 r2 cmp r3, FF r3
movgt FF ...
Idioms can always be constructed
12
2. Scalarizing Permutations
for(i 0 i lt 8 i) r1 r2 r3
tmpi r1 for(i 0 i lt 8 i) r1
offseti r2 tmpr1 i r3 r2
const
offset 4, 4, 4, 4, -4, -4, -4, -4
offset 4, 4, 4, 4, -4, -4, -4, -4
offset 4, 4, 4, 4, -4, -4, -4, -4
13
3. Scalarizing Reductions
for(i 0 i lt 8 i) r1 Ai r2
r2 r1
14
Applied to ARM Neon

• All instructions supported except
• VTBL indirect indexing
• v1 vtbl v2, v3
• Interleaved memory accesses
• Not needed in evaluated benchmarks

15
Translation to SIMD
for(i 0 i lt 8 i) r1 Ai r2
Bi r3 r1 r2 r4 offseti Ci
r4 r3
for(i 0 i lt 8 i 4) v1 Ai v2
Bi v3 v1 v2 v4 v3 constant
for(i 0 i lt 8 i 4) v1 Ai v2
Bi v3 v1 v2 v4
i 4
for(i 0 i lt 8 i 4) v1 Ai v2
Bi v3 v1 v2 v4 offseti
for(i 0 i lt 8 i 4) v1 Ai v2
Bi v3 v1 v2 v3 shuffle v3
Ci v3

• Update induction variable
• Use inverse of defined translation rules

16
Translator Design

• Translator efficiency, speed, flexibility

17
Evaluation

• Trimaran ARM
• Hand SIMDized loops
• SimpleScalar model ARM926 w/ Neon SIMD
• VHDL translator, 130nm std. cell

18
Liquid SIMD Issues

• Code bloat
• lt1 overhead beyond baseline
• Register pressure
• Not a problem
• Translator cost
• 0.2 mm2 2KB cache
• Translation overhead

19
Translation Overhead
MediaBench
Kernels
SPECfp
20
Summary

• Accelerators are more common and evolving
• Costly binary migration
• SIMD virtualization using scalar ISA
• One binary forward/backward compatibility
• Negligible overhead

21
Questions
?
?
?
?

• ?

?
?
?
?
?
?
?