Accelerated Library Framework for Cell B'E'

1
Accelerated Library Framework for Cell B.E.

• Jung-Yin, Chien

2
What is ALF?

• The Accelerated Library Framework (ALF) provides
  a programming environment for
• Data and Task parallel applications and libraries
• The ALF provides you with
• a set of interfaces to simplify library
  development on heterogeneous multi-core systems
• ALF supports the multiple-program-multiple-data
  (MPMD) programming model where multiple programs
  can be scheduled to run on multiple accelerator
  elements at the same time

3
What is ALF?

• The ALF functionality includes
• Data transfer management
• Parallel task management
• Double buffering
• Dynamic load balancing for data parallel tasks
• you can also create descriptions for multiple
  compute tasks and define their execution orders
  by defining task dependency
• The ALF runtime provides an optimal parallel
  scheduling scheme for the tasks based on given
  dependencies

4

• Using a layered approach to develop applications
  in ALF
• Application
• You develop programs only at the host level
• To use the provided accelerated libraries
• Alf_fft(data)
• Accelerated library
• You use the ALF APIs to implement the internal
  logic of libraries
• You just needs knowledge on what data to pass to
  the kernel and how to partition this data up
• Computation kernel
• You write optimized accelerator code at the
  accelerator level

5
Overview of ALF external components

• ALF work block task
• a computational kernel is defined as an
  accelerator routine that takes a given set of
  input data, and computes the resulting output
  data based on the given input
• ALF lightweight task
• a lightweight task is scheduled to run, ALF
  runtime gives the control to the task instances
  running on the accelerators

6
Basic structure of an ALF application

• ALF application structure and process flow

7
Concepts

• Computation kernel
• A computational kernel is a user-defined
  accelerator routine that does the computation for
  a corresponding task.
• Task descriptor
• For work block tasks
• The name of the accelerator computational kernel
  function
• Optionally, the name of the accelerator input
  data transfer list prepare function
• Optionally, the name of the accelerator output
  data transfer list prepare function
• Optionally, the name of the accelerator task
  context setup function
• Optionally, the name of the accelerator task
  context merge function
• For lightweight tasks
• The name of the accelerator entry point function

8
Concepts

• Task
• A task is a program that is executed on a number
  of accelerators. For each accelerator that runs a
  task, a task instance is spawned
• Task instance
• A task can be scheduled to run on multiple
  accelerators. Each task running on an accelerator
  is known as a task instance
• Task dependency
• D (AB)CT
• Dependency (MPMD)
• Task 1
• AB
• Task2
• CT
• Task3
• Multiply
• Task3 executes after Task1 and Task2 complete

9
Concepts

• Work Block

10
Concepts

• Work block
• A work block represents an invocation of a
  computational kernel of a work block task with a
  specific set of related input data, output data,
  and parameters
• The input and output data are described by
  corresponding data transfer lists
• Data transfer list

11
Work Block scheduling(1)

• Default work block scheduling policy

12
Work Block scheduling(2)

• Cyclic work block distribution policy

13
Work Block scheduling(3)

• Bundled work block distribution policy

14
Platform-specific constraints for the ALF
implementation

• SPE accelerator memory constraints

15

• And so on
• Input Buffer
• Output Buffer
• Common buffer