SIGGRAPH, Vancouver 2014 Copyright Khronos Group Page 1

Transcription

1 Neil Trevett Vice President Mobile Ecosystem at NVIDIA President of Khronos and Chair of the OpenCL Working Group SIGGRAPH, Vancouver 2014 Copyright Khronos Group Page 1

2 Copyright Khronos Group Page 2 Speakers Neil Trevett OpenCL Chair, VP NVIDIA NVIDIA Introduction to Khronos and OpenCL Ecosystem Ralph Potter Research Engineer Codeplay SPIR Luke Iwanski Games Technology Programmer Codeplay SYCL Laszlo Kishonti CEO Kishonti Compute Benchmarking Neil Trevett OpenCL Chair, VP NVIDIA NVIDIA Wrap-up and Questions

3 Copyright Khronos Group Page 3 OpenCL Portable Heterogeneous Computing Portable Heterogeneous programming of diverse compute resources - Targeting supercomputers -> embedded systems -> mobile devices One code tree can be executed on CPUs, GPUs, DSPs and hardware - Dynamically interrogate system load and balance work across available processors OpenCL = Two APIs and C-based Kernel language - Platform Layer API to query, select and initialize compute devices - Kernel language - Subset of ISO C99 + language extensions - C Runtime API to build and execute kernels OpenCL across multiple devices Kernel OpenCL Code Kernel OpenCL Code Kernel OpenCL Code Kernel Code GPU DSP HW CPU CPU

4 Copyright Khronos Group Page 4 OpenCL Roadmap What markets has OpenCL been aimed at? What problems is OpenCL solving? How will OpenCL need to adapt in the future? Discussion Focus for New Capabilities HPC Desktop Mobile HPC Desktop Mobile Web HPC Desktop Mobile Web FPGA HPC Desktop Mobile Web FPGA Embedded Safety Critical Dec08 OpenCL 1.0 Specification 3-component vectors Additional image formats Multiple hosts and devices Buffer region operations Enhanced event-driven execution Additional OpenCL C built-ins Improved OpenGL data/event interop Jun10 OpenCL 1.1 Specification Device partitioning Separate compilation and linking Enhanced image support Built-in kernels / custom devices Enhanced DX and OpenGL Interop Nov11 18 months 18 months 24 months OpenCL 1.2 Specification Shared Virtual Memory On-device dispatch Generic Address Space Enhanced Image Support C11 Atomics Pipes Android ICD Nov13 OpenCL 2.0 Specification Roadmap Discussions Binning/Triaging SW and HW features Will use Provisional Specs Some common requests: - C++ Programming - SPIR in Core - Refine and evolve Memory and Execution Models - Better debug and profiling - Trans-API Interop

5 OpenCL Implementations Desktop 1.0 May Jul Jun Aug May Aug Feb May Mar Dec Jul May Jun Aug12 Mobile 1.0 Feb Sep Nov Apr Jan May Apr Aug Jul13 FPGA Dec08 OpenCL 1.0 Specification Jun10 OpenCL 1.1 Specification Nov11 OpenCL 1.2 Specification Nov13 OpenCL 2.0 Specification Copyright Khronos Group Page 5

6 Copyright Khronos Group Page 6 OpenCL Desktop Usage Broad commercial uptake of OpenCL - Mainly imaging, video and vision processing - Adobe, Apple, Corel, ArcSoft Etc. Etc. OpenCL on Sourceforge, Github, Google Code, Bitbucket finds over 2,000 projects - OpenCL implementations - Beignet, pocl - VLC, X264, FFMPEG, Handbrake - GIMP, ImageMagick, IrfanView - Hadoop, Memcached - WinZip, Crypto++ Etc. Etc. Desktop benchmarks use OpenCL - PCMark 8 video chat and edit - Basemark CL, CompuBench Desktop Basemark CL

7 Copyright Khronos Group Page 7 Teaching OpenCL International textbooks - US, Japan, Europe, China and India Research Paper momentum - Over 4000 papers in 2013 Commercial OpenCL training courses - Almost 100 University Courses with OpenCL OpenCL Research Papers on Google Scholar

8 Copyright Khronos Group Page 8 Khronos Foundational APIs Developer Innovation Market Momentum Applications, libraries and frameworks that find OpenCL acceleration can deliver a better end-user experience A successful standard enables and encourages innovation in implementation and usage Deliver the lowest level abstraction possible API that still provides portability this is functionality needed on every platform Implementer Innovation Market Momentum.. Many devices competing on performance and power to tap into the value of OpenCL content

9 Copyright Khronos Group Page 9 OpenCL as Parallel Language Backend JavaScript binding for initiation of OpenCL C kernels Language for image processing and computational photography MulticoreWare open source project on Bitbucket Embedded array language for Haskell Java language extensions for parallelism River Trail Language extensions to JavaScript Compiler directives for Fortran, C and C++ PyOpenCL Python wrapper around OpenCL Harlan High level language for GPU programming OpenCL provides vendor optimized, cross-platform, cross-vendor access to heterogeneous compute resources

10 Copyright Khronos Group Page 10 Libraries and Languages using OpenCL Library Name Overview Website Accelerate accelerate: An embedded language for accelerated array processing amgcl Simple and generic algebraic multigrid framework Aparapi API for data parallel Java. Allows suitable code to be executed on GPU via OpenCL. ArrayFire Array-based function library Bolt Bolt C++ Template Library Boost.Compute Boost.Compute is a GPU/parallel-computing library for C++ based on OpenCL. Bullet Physics Bullet Physic OpenCL accelerated Rigid Body Pipeline C++ AMP CLANG/LLVM based C++AMP 1.2 standard and transforms it into OpenCL-C clblas cl BLAS implementation clfft OpenCL FFT Libarary clmagma clmagma 1.1 is an OpenCL port of MAGMA clpp OpenCL Data Parallel Primitives Library clspmv Sparse Matrix Solver Clyther Python just-in-time specialization engine for OpenCL Codeplay Math Lib OpenCL 1.2 Math library Concord C++ Hetrogenous Programing Framework ( Support OpenCL 1.2 ) TBB like COPRTHR CO-PRocessing THReads (COPRTHR) SDK DL- Data Layout DL Enables Optimized Data Layout Across Heterogeneous Processors ForOpenCL Fortran to OpenCL tool fortrancl FortranCL is an OpenCL interface for Fortran FSCL.Compiler FSharp to OpenCL Compiler GATLAS GPU Automatically Tuned Linear Algebra Software ( Project looks stalled) GMAC Global Memory for Accelerators GPULib Iterative sparse solvers gpumatrix A matrix and array library on GPU with interface compatible with Eigen. GPUVerify GPUVerify is a tool for formal analysis of GPU kernels written in OpenCL Halide Halide Programming language for high-performance image processing Harlan Harlan: A Scheme-Based GPU Programming Language HOpenCL Haskell OpenCL Wrapper API libcl C++ Generic parallel algorithms library Libra SDK Cross Platform Acceleration API M³ Platform Parallel Framework and Primitive Libraries MUMPS Direct Sparse solver Octave Octave acceleration via OpenCL Courtesy: AMD

11 Copyright Khronos Group Page 11 Libraries and Languages using OpenCL #2 Open Fortran Parser ANTLR-based parsing tools that support the Fortran 2008 standard OpenACC to OpenCL Compiler Rose based OpenACC to OpenCL Compiler. OpenCL.jl Julia OpenCL 1.2 bindings OpenCLIPP OpenCL Integrated Performance Primitives - A library of optimized OpenCL image processing functions OpenCLLink Mathematica to use the OpenCL parallel computing language OpenClooVision Computer vision framework based on OpenCL and C# OpenCV-CL OpenCL accelerated OpenCV OpenHMPP Directive-based OpenACC and OpenHMPP Source to OpenCL compiler Paralution C++ sparse iterative solvers and preconditioners library with OpenCL support Pardiso Direct Sparse solver Pencil PENCIL to be a suitable target language for the compilation of domain-specific languages (DSLs). PETSc Portable, Extensible Toolkit for Scientific Computation PyOpenCL OpenCL parallel computation API from Python QT with OpenCL Using OpenCL with QT RaijinCL library for matrix operations for OpenCL Rivertrail JavaScript which supports Data Parallelism via OpenCL RNG Random number generation for parallelcomputations ROpenCL Parallel Computing for R Using OpenCL Rose Compiler Rose Compiler with OpenCL Support Rust-OpenCl OpenCL bindings for Rust. ScalaCL Scala support of OpenCL SkelCL SkelCL is a library providing high-level abstractions for alleviated programming of modern parallel heterogeneohttps://github.com/skelcl/skelcl SnuCL SnuCL naturally extends the original OpenCL semantics to the heterogeneous cluster SpeedIT 2.4 OpenCl based OpenFoam acceleration library streamscan StreamScan: Fast Scan Algorithms for GPUs without Global Barrier Synchronization- SuperLU Direct Sparse solver TM-Task Management Heterogeneous Task Scheduling and Management Trilinos Building blocks for the development of scientific applications; constructing and using sparse and dense matriceshttp://trilinos.sandia.gov/ VexCL VexCL is a C++ vector expression template library for OpenCL/CUDA ViennaCL open-source linear algebra library for computations on many-core architectures (GPUs, MIC) and multi-core CPUhttp://viennacl.sourceforge.net/ VirtualCL VirtualCL (VCL) cluster platform is a wrapper for OpenCL VOBLA Vehicle for Optimized Basic Linear Algebra - Optimized Basic Linear Algebra DSL VOCL Virtualized OpenCL enviornment VSI/Pro VSIPL implementation in OpenCL WAMS Algebraic Multigrid Solver using state-of-the-art wavelet preconditioners- solver for sparse linear equations Courtesy: AMD

12 Widening OpenCL Ecosystem OpenCL C Kernel Source Alternative Alternative Alternative Language for Language for Language Kernels for Kernels Kernels High-level High-level Apps Frameworks and Frameworks Frameworks SPIR Generator (e.g. patched Clang) SPIR is easier compiler target than C SPIR Standard Portable Intermediate Representation CLOSE COOPERATION WITH LLVM COMMUNITY SPIR 2.0 at SIGGRAPH 2014 (uses LLVM 3.4) OpenCL run-time can consume SPIR OpenCL C Runtime SYCL Programming abstraction that combines portability and efficiency of OpenCL with ease of use and flexibility of C++ SYCL 1.0 Provisional Released March 2014 Device X Device Y Device Z Copyright Khronos Group Page 12

13 The Future is Mobile Mobile SOCs now beginning to need more than just GPU Compute - Multi-core CPUs, GPUs, DSPs, ISPs, specialized hardware blocks OpenCL can provide a single programming framework for all processors on a SOC - OpenCL 1.2 Built-in Kernels for custom HW Image Courtesy Qualcomm Copyright Khronos Group Page 13

14 Copyright Khronos Group Page 14 APIs for Mobile Compute GPU Compute Shaders (OpenGL 4.4 and OpenGL ES 3.1) Pervasively available on almost any mobile device or OS Easy integration into graphics apps no API interop needed Program in GLSL not C Limited to acceleration on a single GPU C/C++ Language Integrated GPU Compute Easy programmability and low level access to GPU: Unified Memory, Virtual Addressing, Mature and optimized tools and performance Extensive compute and imaging libraries available (NPP, cufft, cublas, cuda-gdb, nvprof etc.) NVIDIA only, GPU only General Purpose Heterogeneous Programming Framework Flexible, low-level access to any devices with OpenCL compiler Open standard for any device or OS being used as backend by many languages and frameworks Single programming and run-time framework for CPUs, GPUs, DSPs, hardware Needs full compiler stack and IEEE precision RS Easy, High-level Compute Offload from Java C99 based kernel language for simple offload from Java apps to CPU and GPU JIT Compilation provide host and device portability Android only Limited control over acceleration configuration

15 Copyright Khronos Group Page 15 RenderScript and OpenCL RenderScript and OpenCL do not directly compete - RS addressing very different needs to OpenCL at a different level in the stack RenderScript designed for 99% of Android developers - using Java - Code critical sections as native C - automatic offload to CPU/GPU - Programmer Simplicity and Portability across 1,000 s Android handsets - Future - Dynamic load balancing through integration with Android instrumentation and power management systems BUT - other types of developer need OpenCL-class control in native code - Middleware engines: Unity, Epic Unreal, metaio AR, Bullet Physics - Leading edge apps: real-time video/vision/camera - OEM functionality: e.g. camera pipeline - These are the developers/apps/engines that hardware vendors want for differentiation Java Compute RS Graphics Java Binding to OpenGL ES (similar to JSR239) OpenCL on Android can enable specialized access to native acceleration and be an effective backend for RenderScript innovation Native

16 Mixamo - Avatar Videoconferencing Real time facial animation capture on mobile ported directly from PC Animate an avatar while conferencing Full GPU acceleration of vision processing using OpenCL NVIDIA Tegra K1 Development Board Copyright Khronos Group Page 16

17 Copyright Khronos Group Page 17 CompuBench Preview OpenGL ES Compute Shaders vs. OpenCL - After each compute iteration the current level-set is visualized with OpenGL Medical data of a human brain - Processed by level-set segmentation, measuring execution time Implemented API features: - 3D image writes, OpenCL-OpenGL interop, geometry shaders

18 SPIR 2.0 Provisional SIGGRAPH, Vancouver August 2014 Copyright Khronos Group Page 1

19 Standard Portable Intermediate Representation Goals 1. Portable interchange format for partially compiled OpenCL C 2. Target format for other languages Enables compiler ecosystem for portable parallel programs Copyright Khronos Group Page 2

20 OpenCL as Parallel Language Backend JavaScript binding for initiation of OpenCL C kernels Language for image processing and computational photography MulticoreWare open source project on Bitbucket Embedded array language for Haskell Java language extensions for parallelism River Trail Language extensions to JavaScript Compiler directives for Fortran, C and C++ PyOpenCL Python wrapper around OpenCL Harlan High level language for GPU programming OpenCL provides vendor optimized, cross-platform, cross-vendor access to heterogeneous compute resources Copyright Khronos Group Page 3

21 Builds on LLVM and OpenCL Optimizing compiler toolkit Portable, flexible, well understood Open source platform for innovation Proven platform for heterogeneous parallel programming Multi-vendor: CPU, GPU, FPGA etc. Copyright Khronos Group Page 4

22 Why use SPIR? Without SPIR: Vendors shipping source - Risk IP leakage Vendors shipping multiple binaries - Complexity - Miss optimizations in new compilers - Forward compatibility issues With SPIR: Ship a single binary per platform - E.g. SPIR file can support Intel & AMD Many vendors support SPIR consumption Shipped application can retarget new devices and new vendors Opportunity to unleash innovation: Domain Specific Languages, C++ Compilers, Halide,. Copyright Khronos Group Page 5

23 What s new in SPIR 2.0? Full support of OpenCL 2.0 C kernel language - Generic address space - Device side kernel enqueue - C++11 atomics - Pipes - More LLVM 3.4 with restrictions and conventions If you can do it in OpenCL C You can do it in SPIR Copyright Khronos Group Page 6

24 SPIR ecosystem is IR definition - Portable non-source encoding for OpenCL 1.2 or 2.0 device programs - SPIR 1.2 is based on LLVM SPIR 2.0 is based on LLVM 3.4 Consumption API for target hardware - cl_khr_spir extension to OpenCL runtime API Example generator - Open source patch to Clang translates OpenCL C to SPIR IR - Available in github: Ease of use tools - SPIR Verifier, SPIR built-ins name mangler - Available in github: Copyright Khronos Group Page 7

25 Longevity and Versioning SPIR to track both LLVM and OpenCL versions - SPIR 1.2 LLVM OpenCL SPIR 2.0 LLVM OpenCL 2.0 SPIR consumer tells you what versions can be loaded Khronos members contributing to mainline LLVM+Clang - Backward compatibility fixes and tests - Full SPIR support in Clang - Ease of use tools Copyright Khronos Group Page 8

26 Call to Action Seeking feedback on SPIR 2.0 provisional - A Provisional specification Innovate on the Front end - New languages, abstractions - Target production quality backends Innovate on the Back end - New target platforms: Multi core, Vector, VLIW - Reuse production quality frontends Innovate on Tooling - Program analysis, optimization Copyright Khronos Group Page 9

27 Getting Started IR Specification - Khronos SPIR registry - Front end - Khronos-patched Clang from Github Verifier - LLVM pass checks SPIR validity - Khronos Github Backend - Check your favorite OpenCL implementation for cl_khr_spir Same open source license as mainline LLVM and Clang Copyright Khronos Group Page 10

28 More About Flows Copyright Khronos Group Page 11

29 OpenCL: Source Compilation Flow ISV ships their kernel source - Exposes their IP Supports only OpenCL C User application OpenCL C Kernel Source Vendor specific OpenCL Host Library Copyright Khronos Group Page 12

30 OpenCL: Binary compilation flow OpenCL C Kernel Source Vendor specific OpenCL Host Library Vendor specific binary Platform specific container Vendor specific binary Vendor specific ISV ships vendor-specific binary - Proliferation: devices, driver revisions, vendors - Market-lagging: target shipped products OpenCL Host Library Copyright Khronos Group Page 13

31 OpenCL: SPIR flow ISV ships kernels in SPIR form User runs application on platform of their choice OpenCL C Kernel Source OpenCL Host Library Vendor specific Standard Portable Intermediate Platform specific container Standard Portable Intermediate OpenCL Host Library Vendor specific Copyright Khronos Group Page 14

32 SPIR Reference Flow Generation Device program source SPIR Generator Standard Portable Intermediate Consumption Platform specific container Standard Portable Intermediate Vendor specific OpenCL Runtime cl_khr_spir Copyright Khronos Group Page 15

33 SPIR Today Generation OpenCL C Device program source Patched Clang SPIR Generator Standard Portable Intermediate Consumption Platform specific container Standard Portable Intermediate Vendor specific OpenCL Runtime cl_khr_spir Copyright Khronos Group Page 16

34 Sample SPIR Consumption Flow Standard Portable Intermediate clcreateprogramwithbinary Device specific binary clbuildprogram( -x spir -spir-std=2.0.) Copyright Khronos Group Page 17

35 Sample SPIR Flow: Room for Optimizations Standard Portable Intermediate cl_program Device executable clbuildprogram SPIR Verifier Standard LLVM optimizations Custom optimizations E.g. vectorize Materialization (Convert to device specific IR) ABI fixup, custom optimizations JIT LLVM IR Target IR Copyright Khronos Group Page 18

36 Resources IR Specification - Khronos SPIR registry - Feedback Forum Thread - Khronos-patch Clang and Tools Backend - Check your favorite OpenCL implementation for cl_khr_spir Copyright Khronos Group Page 19

37 Questions? Copyright Khronos Group Page 20

38 Copyright Khronos Group 2014 SYCL for OpenCL in a Nutshell Luke Iwanski, Games Technology Codeplay SIGGRAPH Vancouver

39 2 Copyright Khronos Group 2014

40 SYCL for OpenCL in a nutshell Copyright Khronos Group 2014 Why? Where in the OpenCL ecosystem? Motivation Features overview Example time Roadmap 3

41 Copyright Khronos Group 2014 Why SYCL? Modern C++ programming model for OpenCL (compiler, runtime) Ease to use High performance Single source Allows multi-compiler implementation. SYCL device compiler + Host compiler of your choice Portability across platforms and compilers Providing the full OpenCL feature set and seamless integration with existing OpenCL code Enabling the creation of higher level programming models and C++ templated libraries based on OpenCL 4

42 Copyright Khronos Group 2014 OpenCL ecosystem OpenCL C Kernel Source SPIR Generator (e.g. Khronos patched Clang open source on GitHUB) Alternative Alternative Alternative Language for Language for Language Kernels for Kernels Kernels High-level High-level Apps Frameworks and Frameworks Frameworks SPIR Standard Portable Intermediate Representation SPIR 1.2 Released January 2014 OpenCL Runtime SYCL A programming abstraction that combines the portability and efficiency of OpenCL with the ease of use and flexibility of C++ SYCL 1.2 Provisional Released March 2014 Device X Device Y Device Z 5

43 The layering of SYCL: Building an ecosystem Copyright Khronos Group 2014 User application code C++ template libraries SYCL for OpenCL OpenCL 6

44 Copyright Khronos Group 2014 Motivation We want to enable C++ for the OpenCL ecosystem Where more C++ developers can get the benefits of OpenCL With C++ libraries supported on OpenCL platforms C++ tools supported on OpenCL platforms Aim to achieve long-term support for OpenCL features with C++ Multiple Sources of implementations (multiple vendors) Reliability by providing host fall-back Enable future innovations 7

45 Copyright Khronos Group 2014 SYCL features: Overview 8

46 Copyright Khronos Group 2014 OpenCL/SYCL interoperability Seamless integration of OpenCL C applications with SYCL applications OpenCL C data types and built-in functions available SYCL / OpenGL interoperability Based on OpenCL/OpenGL interoperability extensions C++ exception handling Host fall-back mode - using SYCL without OpenCL Introduced in SYCL Hierarchical data parallelism 9

47 Copyright Khronos Group 2014 Hierarchical Data Parallelism Workgroup Task (nd-range) Work Work item Workgroup item Work Work item Workgroup item Work Work item Work itemworkgroup Work item Work item Work Work item item Work Work item item item Work item Work Work item Work item Work Work item item item item Work Work item item Work Work item item buffer<int> my_buffer(data, 10); auto in_access = my_buffer.get_access<cl::sycl::access:read>(); auto out_access = my_buffer.access<cl::sycl::access:write>(); command_group(my_queue, [&]() { " parallel_for_workgroup(nd_range(range(size), range(groupsize)), " " lambda<class hierarchical>([=](group group) " { " " parallel_for_workitem(group, [=](item tile) " " { " " " out_access[tile] = in_access[tile] * 2; " " }); " })); }); Advantages: 1. Easy to understand the concept of work-groups 2. Performance-portable between CPU and GPU 3. Barriers are automatically deduced 4. Easier to compose components and algorithms 10

48 Copyright Khronos Group 2014 Example time: Simple kernel 11

49 12 Copyright Khronos Group 2014

50 Copyright Khronos Group 2014 Simple kernel summary Simple kernel demo source is only 20 lines of actual C++/ SYCL code Equivalent of simple kernel demo in OpenCL takes over 100 lines of code This code can be easily templated by changing 17 lines of code Plain OpenCL C will take many, many,.. many more lines of code 13

51 Copyright Khronos Group 2014 Example time: Templated kernel 14

52 15 Copyright Khronos Group 2014

53 16 Copyright Khronos Group 2014

54 17 Copyright Khronos Group 2014

55 Copyright Khronos Group 2014 Templated kernel summary Only 52 lines of code to create a templated kernel for the subtract operation Templates on the device factor of 5 lines per new datatype (including initialisation and printing) SYCL is simple 18

56 Copyright Khronos Group 2014 Final notes about SYCL Keep in mind Advantages of modern C++ (lambdas, templates, struct arguments, static polymorphism) but, limitations of current OpenCL ( recursion, dynamic allocation, static variables) It will get better with the next OpenCL iterations 19

57 Copyright Khronos Group 2014 SYCL roadmap GDC, March 2014 Released a provisional specification to enable feedback Developers can provide input into the standardisation process Feedback via Khronos forums Next steps Full specification, based on feedback Khronos test suite for implementations Release of implementations 20

58 Copyright Khronos Group 2014 SYCL Useful Links SYCL spec and forums: trisycl github: Codeplay s blogs: Examples github 21

59 Copyright Khronos Group 2014 Thanks Luke 22