KeynoteTHE HETEROGENEOUS SYSTEM ARCHITECTURE ITS (NOT) ALL ABOUT THE GPU
4 likes4,553 views
The document discusses the Heterogeneous System Architecture (HSA) which aims to make it easier to program heterogeneous systems containing CPUs and other accelerators like GPUs. The key goals of HSA are to provide a unified memory model, low-latency dispatch between devices, and support for a wide range of programming languages. The HSA Foundation is a consortium working to define specifications to achieve these goals and promote HSA adoption.
KeynoteTHE HETEROGENEOUS SYSTEM ARCHITECTURE ITS (NOT) ALL ABOUT THE GPU
- 1. CLICK TO EDIT MASTER TITLE STYLE Click to edit Master text styles ‒ Second level ‒ Third level ‒ Fourth level ‒ Fifth level THE HETEROGENEOUS SYSTEM ARCHITECTURE ITS (NOT) ALL ABOUT THE GPU PAUL BLINZER, FELLOW, HSA SYSTEM SOFTWARE, AMD SYSTEM ARCHITECTURE WORKGROUP CHAIR, HSA FOUNDATION 1 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 2. CLICK TO EDIT MASTER TITLE STYLE THREE ERAS OF PROCESSOR PERFORMANCE Single-Core Era Multi-Core Era Click to edit Master text styles Enabled by: ‒ Second level Moore’s Observation Voltage Scaling Micro-Architecture Constrained by: ‒ Third level ‒ Fourth level Enabled by: Moore’s Observation Desire for Throughput 20 years of SMP arch Constrained by: ‒ Fifth level Single-thread Performance ? o we are here Power Complexity Time Throughput Performance Power Parallel SW availability Scalability Time o we are here (# of Processors) 2 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014 Heterogeneous Systems Era Enabled by: Constrained by: Targeted Application Moore’s Observation Abundant data parallelism Power efficient data parallel Performance processing (GPUs) Programming models Communication overheads we are here Time (Data-parallel exploitation) o
- 3. CLICK TO EDIT MASTER TITLE STYLE THE PROGRESS OF FLOATING POINT PERFORMANCE Click to edit Master text styles ‒ Second level ‒ Third level ‒ Fourth level ‒ Fifth level 3 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 4. CLICK TO EDIT MASTER TITLE STYLE WHY THE DIFFERENCE? ITS ABOUT THE PURPOSE CPUs: built for general purpose, serial instruction threads, often high data locality, lot’s of conditional execution and dealing with data interdependency Click to edit Master text styles ‒ CPU cache hierarchy is focused on general purpose data access from/to execution units, feeding back ‒ Second level previously computed data to the execution units with very low latency ‒ Third level ‒ Comparatively few registers (vs GPUs), but large caches keep often used “arbitrary” data close to the execution ‒ units Fourth level ‒ Fifth level ‒ Execution model is synchronous GPUs: purpose-built for highly data parallel, latency tolerant workloads ‒ Apply the same sequence of instructions over and over on data with little variation but high throughput (“streaming data”), passing the data from one processing stage to another (latency tolerance) ‒ Compute units have a relatively large register file store ‒ Using a lot of “specialty caches” (constant cache, Texture Cache, etc), data caches optimized for SW data prefetch ‒ separate memory spaces (e.g. GPU local memory) ‒ Execution model is asynchronous 4 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 5. CLICK TO EDIT MASTER TITLE STYLE THE HSA (= HETEROGENEOUS SYSTEM ARCHITECTURE) GOALS To bring accelerators forward as a first class processor within the system Click to edit Master text styles ‒ Unified process address space access across all processors ‒ Second level ‒ Operates in pageable system memory ‒ Third level ‒ Full memory coherency between the CPU and HSA components ‒ Fourth level ‒ Well-defined relaxed consistency memory model suited for many high level languages ‒ Fifth level ‒ User mode dispatch/scheduling (eliminates “drivers” from the dispatch path) ‒ QoS through pre-emption and context switching* It is not dictating a specific platform or component microarchitecture! ‒ It defines the “what needs to be supported”, not the “how it needs to be supported” ‒ Focus is on providing a robust data parallel application execution infrastructure for “regular languages” 5 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 6. CLICK TO EDIT MASTER TITLE STYLE THE GOALS OF HSA, PART 2 To attract mainstream programmers Click to edit Master text styles ‒ By making it easier writing data parallel code ‒ Second level ‒ Support broader set of languages beyond traditional GPGPU Languages ‒ Third level ‒ Support for Task Parallel & Nested Data Parallel Runtimes ‒ Fourth level ‒ Rich debugging and performance analysis support ‒ Fifth level Create a platform architecture accessible for all accelerator types, not only GPU ‒ Focused on the APU/SOC and compute, but not preventing other accelerator types to participate But one can’t do it alone… 6 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 7. CLICK TO EDIT MASTER TITLE STYLE WHAT IS THE HSA FOUNDATION? This is the short version… Click to edit Master text styles “The HSA Foundation is a not-for-profit consortium of SOC and SOC IP vendors, OEMs, academia, OSVs and ISVs defining a consistent ‒ Second level heterogeneous platform architecture to make it dramatically easier to ‒ Third level program heterogeneous parallel devices” ‒ Fourth level ‒ Fifth level 7 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014 Feb 2014 It spans multiple host platform architectures and programmable data parallel components collaborating within the same HSA system architecture - CPU: x86, ARM, MIPS, … - Accelerator types: GPUs, DSPs, … Specifications define HW & SW platform requirements for applications to target the feature set from high level languages and APIs The System Architecture specification defines the required component and platform features for HSA compliant components Other HSA specifications leverage these properties for the SW definition
- 8. CLICK TO EDIT MASTER TITLE STYLE THE KEY DELIVERABLES OF THE HSA FOUNDATION IP, Specifications and API’s The primary specifications are Click to edit Master text styles ‒ The HSA Platform System Architecture Specification ‒ Second level ‒ Focus on hardware requirements and low level system software Support ‒ Third level ‒ “Small Model” (32bit) and “Large Model” ( 64bit) ‒ The HSA Programmer Reference Manual ‒ Fourth level ‒ Definition of an HSAIL Virtual ISA ‒ Fifth level ‒ Binary format (BRIG) ‒ Compiler writers guide and Libraries developer guide ‒ The HSA System Runtime Specification ‒ HSA Tools ‒ LLVM to HSAIL Compiler ‒ HSAIL Assembler HSA Conformance Suite Open source community and “reference” implementation of SW stack (Linux) by AMD ‒ To “jump start” the ecosystem ‒ Allow a single shared implementation where appropriate ‒ Enable university research in all areas 8 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014 Platform (Software) System Architecture Specification Programmer’s Reference Manual System Runtime Specification Tools Conformance
- 9. OPENCL™ WITH HSA, NOT OPENCL™ VS HSA! CLICK TO EDIT MASTER TITLE STYLE HSA is an optimized platform architecture, which will run OpenCL™ very well Click to edit Master text styles ‒ It is a complementary standard, not a competitor to OpenCL™ ‒ It is focused on the hardware and system platform runtime definition more than an API itself ‒ Second level ‒ It is ready to support many more languages than C/C++, including managed code languages ‒ Third level ‒ Fourth level OpenCL™ on HSA benefits from a rich and consistent platform infrastructure ‒ Fifth level ‒ Pointers shared between CPU and GPU (Shared Virtual Memory), Avoidance of wasteful copies ‒ Low latency dispatch ‒ Improved and consistent memory model ‒ Virtual function calls ‒ Flexible control flow ‒ Exception generation and handling ‒ Device and platform atomics 9 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 10. CLICK TO EDIT MASTER TITLE STYLE THE SYSTEM ARCHITECTURE WORKGROUP OF THE HSA FOUNDATION The membership spans a wide variety of IP and platform architecture owners Click to edit Master text styles ‒ Several host platform architectures (x86, ARM, MIPS, …) ‒ Second level ‒ Third level The specifications define a common set of platform properties ‒ Fourth level ‒ provide a dependable hardware and system foundation for application software, libraries and ‒ runtimes Fifth level ‒ Eliminate “weak points” in the system software- and hardware architecture of traditional platforms, causing overhead processing data parallel workloads ‒ Focusing on a few key concepts (queues, signals, memory) optimized and used throughout the architecture 10 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 11. CLICK TO EDIT MASTER TITLE STYLE THE SYSTEM ARCHITECTURE WORKGROUP OF THE HSA FOUNDATION The main deliverables are: ‒ Well-defined, consistent and dependable memory model all HSA agents operate in Click to edit Master text styles ‒ Share access to process virtual memory between HSA agents (“ptr-is-ptr”) ‒ Second level ‒ Low-latency workload dispatch, contained in user-mode queues ‒ Third level ‒ Scalability across a wide range of platforms, from smartphones to clients and servers ‒ Fourth level ‒ These properties are ‒ Fifth level leveraged in the “HSA Programmer’s Reference”, HSAIL and HSA Runtime specifications 11 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 12. DESIGN CRITERIA FOR THE HSA INFRASTRUCTURE CLICK TO EDIT MASTER TITLE STYLE HSA defines platform & HW requirements Software can depend on ‒ Comprehensive Memory model (well-defined visibility and ordering rules for transactions) Click to edit Master text styles ‒ Shared Virtual Memory (same page table mapping, HW access enforcement) ‒ Second level ‒ Page fault capability for every HSA agent ‒ Third level ‒ Cache Coherency Domains ‒ Fourth level ‒ Architected, memory-based signaling and synchronization ‒ Fifth level ‒ User Mode Queues, Architected Queuing Language (AQL), Service Queues for runtime callbacks ‒ Preemption / Quality of Service* ‒ Error Reporting ‒ Hardware Debug and profiling support requirements ‒ Architected Topology Discovery Thin System SW Layer enables HW features for use by an application runtime ‒ Mainly responsible for hw init, access enforcement and resource management ‒ Provides a consistent, dependable feature set for application layer through SW primitives 12 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 13. CLICK TO EDIT MASTER TITLE STYLE HSA SECURITY AND EXECUTION MODEL HSA components operate in the same security infrastructure as the host CPU Click to edit Master text styles ‒ User and privileged memory distinction ‒ Second level ‒ Hardware enforced process space isolation ‒ Third level ‒ Page attributes (Read, write, execute) protections enforced by HW and apply as defined by system software ‒ Fourth level ‒ Fifth level Internally, the platform partitions functionality by privilege level ‒ User mode queues can only run AQL packets within the defined process context HSA defines Quality of Service requirements ‒ Requires support for mechanisms to schedule both HSA and non-HSA workloads for devices that support both task types with appropriate priority, latency, throughput and scheduling constraints. ‒ Context Switch ‒ Preempt ‒ Terminate and Context Reset 13 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 14. CLICK TO EDIT MASTER TITLE STYLE HSA COMMAND AND DISPATCH FLOW Click to edit Master text styles ‒ Second level ‒ Third level Application ‒ Fourth level ‒ Fifth level C Application B Application A Optional Dispatch Buffer C C C C C Hardware Queue B B B Hardware Queue A A A 14 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014 Accelerator Hardware (GPU) Hardware Queue SW view: User-mode dispatches to HW No Kernel Driver overhead Low dispatch times CPU & GPU dispatch APIs HW view: HW / microcode controlled HW scheduling Architected Queuing Language (AQL) HW-managed protection
- 15. WHAT DEFINES HSA PLATFORMS AND COMPONENTS? CLICK TO EDIT MASTER TITLE STYLE An HSA compatible platform consists of “HSA components” and “HSA agents” Click to edit Master text styles ‒ Both adhere to the same various system architecture requirements (memory model, etc) ‒ Second level ‒ HSA agents can submit AQL packets for execution, may be but are not required to be an HSA component (e.g. host CPU cores can be an HSA agent, but “step out” of the role) ‒ Third level ‒ HSA components can be a “target” of AQL, by implication they are also HSA agents (= generate AQL packets) ‒ Fourth level ‒ Fifth level Defined in the “System Architecture” specification ‒ The HSAIL and “Programmer’s Reference Manual” specifications define the software execution model ‒ Architected mechanisms to enqueue and dispatch workloads from one HSA agent queue to another eliminate the need to use the host CPU for these purposes for a lot of scenarios ‒ Architected infrastructure allows exchanging data with non-HSA compliant components in a platform ‒ Fundamental data types are naturally aligned 15 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 16. CLICK TO EDIT MASTER TITLE STYLE WHAT DEFINES HSA PLATFORMS AND COMPONENTS? There are two different machine models (“small” and “large”) ‒ target different functionality levels ‒ It takes into account different feature requirements for different platform environments ‒ In all cases, the same HSA application programming model is used to target HSA agents and provides the same power–efficient and low-latency dispatch mechanisms, synchronization primitives and SW programming model Click to edit Master text styles ‒ Second level ‒ Third level ‒ Fourth level Applications written to target HSA small model machines will generally work on large ‒ Fifth level model machines ‒ If the large model platform and host Operating System provides a 32bit process environment Properties Small Machine Model Large Machine Model Platform targets embedded or personal device space (controllers, smartphones, etc.) 16 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014 PC, workstation, cloud Server, etc running more demanding workloads Native pointer size 32bit 64bit (+ 32bit ptr if 32bit processes are supported) Floating point size Half (FP16*), Single (FP32) precision Half (FP16*), Single (FP32), Double (FP64) precision Atomic ops size 32bit 32bit, 64bit *min. Load and store on memory
- 17. CLICK TO EDIT MASTER TITLE STYLE THE HSA MEMORY MODEL REQUIREMENTS A memory model defines how writes by one work item or agent becomes Click to edit Master text styles visible other work items and agents ‒ Second level ‒ Rules that need to be adhered to by compilers and application threads ‒ Third level ‒ It defines visibility and ordering rules of write and read events across work items, HSA agents and interactions with non-HSA components in the system ‒ Fourth level ‒ Important to define ‒ Fifth level scope for performance optimizations in the compiler and allow reordering in the Finalizer HSA is defined to be compatible with C++11, Java and .NET Memory Models ‒ Relaxed consistency memory model for parallel compute performance ‒ Inherently maps to many CPU and device architectures ‒ Efficient sequential consistency mechanisms supported to fit high-level language programming models ‒ Scope visibility controlled by Load.Acquire / Store.Release & Barrier semantics 17 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 18. CLICK TO EDIT MASTER TITLE STYLE THE HSA MEMORY MODEL REQUIREMENTS A consistent, full set of atomic operations is available Click to edit Master text styles ‒ Naturally aligned on size, small machine model supports 32bit, large machine model supports 32bit and 64bit ‒ Second level Cache Coherency between HSA agents (& host CPU) is maintained by default ‒ Third level ‒ Key feature of the HSA system & platform environment ‒ Fourth level ‒ Defined transition points to non-HSA “data domains” (e.g. graphics, audio, …) ‒ Fifth level 18 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 19. THE HSA SIGNALING INFRASTRUCTURE CLICK TO EDIT MASTER TITLE STYLE HSA supports hardware-assisted signaling and synchronization primitives Click to edit Master text styles ‒ ‒ Defines Second consistent, level memory based semantics to synchronize with work items processed by HSA agents ‒ Third level ‒ e.g. 32bit or 64bit value, content update, wait on value by HSA agents and AQL packets ‒ Fourth level ‒ Typically hardware-‒ Fifth level assisted, power-efficient & low-latency way to synchronize execution of work items between threads on HSA agents Allows one-to-one and one-to-many signaling ‒ The signaling semantics follow general atomicity requirements defined in the memory model ‒ System Software, runtime & application SW can use this infrastructure to efficiently build higher-level synchronization primitives like mutexes, semaphores, … 19 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014 Source: wikimedia commons
- 20. CLICK TO EDIT MASTER TITLE STYLE WHAT IS HSAIL? HSAIL is the intermediate language for parallel compute in HSA Click to edit Master text styles ‒ HSAIL is a low level instruction set designed for parallel compute in a shared virtual memory environment. ‒ Second level ‒ Generated by a high level compiler (LLVM, gcc, Java VM, etc) ‒ Third level ‒ Compiled down to GPU ISA or other parallel processor ISA by an IHV Finalizer ‒ Fourth level ‒ Finalizer may execute at run time, install time or build time, depending on platform type ‒ Fifth level HSAIL is SIMT in form and does not dictate hardware microarchitecture ‒ HSAIL is designed for fast compile time, moving most optimizations to HL compiler ‒ Limited register set avoids full register allocation in Finalizer ‒ HSAIL is at the same level as PTX: an intermediate assembly or Virtual Machine Target ‒ Represented as bit-code in in a Brig file format with support late binding of libraries. 20 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 21. CLICK TO EDIT MASTER TITLE STYLE LOOKING BEYOND THE TRADITIONAL GPU LANGUAGES Dynamic Languages are one of the most interesting areas for exploration of Click to edit Master text styles heterogeneous parallel runtimes and have many applications ‒ Second level Dynamic Languages have different compilation foundations targeting HSA / HSAIL ‒ Third level ‒ Java/Scala, JavaScript, Dart, etc ‒ Fourth level ‒ Fifth level See Project Sumatra - http://openjdk.java.net/projects/sumatra/ ‒ Formal Project for GPU Acceleration for Java in OpenJDK HSA allows efficient support for other standards based language environments ‒ like OpenMP, Fortran, GO, Haskell ‒ And domain specific languages like Halide, Julia, and many others 21 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 22. CLICK TO EDIT MASTER TITLE STYLE LOOKING BEYOND THE DATA PARALLEL COMPUTE APPLICATION The initial release of the HSA specifications focuses on data parallel compute language and application scenarios Click to edit Master text styles ‒ Focus is on integrating GPUs into the general software infrastructure ‒ Second level ‒ But the next generations of the specifications may apply to other domains ‒ Third level ‒ With their domain-specific HW processor language focus ‒ Fourth level ‒ Fifth level By design the HSA infrastructure is quite easy to extend ‒ Initial focus is on data parallel compute tasks ‒ But other areas of support are under consideration for the future ‒ Architected Topology infrastructure allows to reliably identify and address domain specific accelerator capabilities By design the HSA infrastructure is easy to virtualize ‒ Programming model does leverage few, simple hardware & platform paradigms (queues, signals, memory) for its operation ‒ Future spec work may put additional requirements to cover such environments 22 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014 CPU GPU Audio Processor Image Processor Shared Memory and Coherency Fabric Video Decode Encode DSP Security Processor Fixed Function Accelerator
- 23. CLICK TO EDIT MASTER TITLE STYLE IN SUMMARY… HSA is not about a specific API, feature or runtime Click to edit Master text styles ‒ It is about a paradigm to efficiently access the various heterogeneous components in a system by software ‒ Second level ‒ It allows application programmers to use the languages of their choice to efficiently implement their code ‒ Third level HSA is not about a specific hardware or vendor or Operating System ‒ Fourth level ‒ It defines a few fundamental requirements and concepts as building blocks software at all levels can depend on ‒ Fifth level ‒ HW vendors can efficiently expose their compute acceleration features to software in an architected way ‒ OS, runtimes and application frameworks can build efficient data and task parallel runtimes leveraging these ‒ Application software can more easily use the right tool for the job through high level language support HSA is an open and flexible concept ‒ Collaborative participation through the HSA Foundation is encouraged for companies and academia ‒ The first set of standards by the HSA Foundation is either released or imminent ‒ This is a good time to engage 23 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 24. WHERE TO FIND FURTHER INFORMATION ON HSA? CLICK TO EDIT MASTER TITLE STYLE HSA Click Foundation to edit Master Website: text http://styles www.hsafoundation.com ‒ The ‒ main Second location level for specs, developer info, tools, publications and many things more ‒ Also the location to apply for membership ‒ Third level ‒ Fourth level Tools are now available at GitHUB – HSA Foundation ‒ Fifth level ‒ libHSA Assembler and Disassembler ‒ https://github.com/HSAFoundation/HSAIL-Tools ‒ HSAIL Instruction Set Simulator ‒ https://github.com/HSAFoundation/HSAIL-Instruction-Set-Simulator ‒ HSA ISS Loader Library for Java and C++ for creation and dspatch HSAIL Kernels ‒ https://github.com/HSAFoundation/Okra-Interface-to-HSAIL-Simulator More to come soon … 24 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 25. CLICK TO EDIT MASTER TITLE STYLE Click to edit Master text styles With thanks to Phil Rogers, Greg Stoner Sasa Marinkovic and others for some materials and feedback ‒ Second level ‒ Third level ‒ Fourth level ‒ Fifth level Trademark Attribution HSA Foundation, the HSA Foundation logo and combinations thereof are trademarks of HSA Foundation, Inc. in the United States and/or other jurisdictions. Other names used in this presentation are for identification purposes only and may be trademarks of their respective owners. ©2014 HSA Foundation, Inc. All rights reserved. 25 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014
- 26. ANY QUESTIONS? CLICK TO EDIT MASTER TITLE STYLE Of course there are, so go ahead Click to edit Master text styles ‒ Second level ‒ Third level ‒ Fourth level ‒ Fifth level 26 | THE HETEROGENEOUS SYSTEM ARCHITECTURE – IT’S (NOT) ALL ABOUT THE GPU | MARCH 1, 2014