1. Martin Kruliš 17. 12. 2015 by Martin Kruliš (v1.1)1

2.  History ◦ Original idea – there is a lot of silicon on a CPU die  Use many x86 cores to create a GPU ◦ 2006 Project Larabee  The GPU could not keep up with AMD and NVIDIA ◦ 2007 Teraflops Research Chip (96-bit VLIW arch) ◦ 2009 Single Chip Cloud Computer (48 cores) ◦ 2010 Knights Ferry (32 cores) ◦ 2011 Knights Corner (60 cores based on Pentium 1) ◦ 2013 Knights Landing (72 cores based on Atom) ◦ Knights Hill announced 17. 12. 2015 by Martin Kruliš (v1.1)2

3.  The Xeon Phi Device ◦ Many simpler (Pentium/Atom) cores ◦ Each equipped with powerful 512bit vector engine 17. 12. 2015 by Martin Kruliš (v1.1)3

4.  Software Architecture ◦ Xeon Phi is basically an independent Linux machine 17. 12. 2015 by Martin Kruliš (v1.1)4 Example

5.  Modes of Xeon Phi Usage ◦ OpenCL device ◦ Standalone computational device  Connected over TCP/IP (SSH, …)  Using low-level symmetric communications interface ◦ MPI device  Communicating over TCP/IP or OFED  May be used in both ways (ranked from host or device) ◦ Offload device  Explicit mode offloading  Implicit mode offloading 17. 12. 2015 by Martin Kruliš (v1.1)5

6.  Xeon Phi as OpenCL Accelerator Card ◦ Requires Intel OpenCL platform ◦ HW to OpenCL mapping  Xeon Phi card = OCL compute device  Virtual core = OCL compute unit ◦ Work items in work group are processed in a loop  Unrolled 16x and vectorized when possible ◦ OCL manages thread pool  One thread per virtual core  Work groups are assigned to threads as tasks ◦ Better to run more WGs and less WIs then on GPU 17. 12. 2015 by Martin Kruliš (v1.1)6 Example

7.  Using Xeon Phi as Standalone Device ◦ The device is autonomous linux machine ◦ The code is cross-compiled on host and deployed on Xeon Phi (including libraries) ◦ Complete freedom in parallelization techniques  OpenMP, Intel TBB, Intel CILK, pthreads, … ◦ Communication has to be performed manually by  Symmetric communication interface (SCIF)  TCP/IP stack, which uses SCIF as data link layer ◦ Useful for extending master-worker applications 17. 12. 2015 by Martin Kruliš (v1.1)7 Example

8.  Symmetric Communications Interface ◦ Socket-like interface that encapsulates PCI-Express data transfers  Message passing and RMA transfers ◦ Memory mapping techniques  Device memory may be mapped to host address space  Host memory and memory of other devices can be mapped to address space of a device  Upper 512G (32x 16G pages) ◦ Supports direct assignment virtualization model ◦ All other communication methods are built on SCIF 17. 12. 2015 by Martin Kruliš (v1.1)8

9.  Offload Execution Model ◦ Both host and device code is written together ◦ Offload parts for the device are explicitly marked  The compiler performs the dual compilation, inserts the stubs, handles the data transfers, … ◦ Explicit offload model (a.k.a. Pragma offload)  Everything is controlled by programmer  Only binary-safe data structures can be transferred ◦ Implicit offload model (a.k.a. Shared VM model)  Data transfers are handled automatically  Complex data structures and pointers may be transferred 17. 12. 2015 by Martin Kruliš (v1.1)9

10.  Code Compilation 17. 12. 2015 by Martin Kruliš (v1.1)10 Source Host MIC #pragma offload _Cilk_offload Compilation Stub

11.  Pragma Offload ◦ Functions and variables are declared with __attribute__((target(mic))) ◦ Offloaded code is invoked as #pragma offload  A clause may select target card target(mic[:id])  Or list data structures which are used for the offload  in(varlist) – copied to the device before the offload  out(varlist) – copied back to host after the offload  inout, nocopy, length, align  Allocation control alloc_if(), free_if() 17. 12. 2015 by Martin Kruliš (v1.1)11

12.  Example __attribute__((target(mic))) void preprocess(…) { … }... __attribute__((target(mic))) static float *X; // of N items #pragma offload target(mic:0) \ in(X:length(N) alloc_if(1) free_if(0)) preprocess(X); #pragma offload target(mic:0) \ nocopy(X:length(N) alloc_if(0) free_if(0)) process(X); #pragma offload target(mic:0) \ out(X:length(N) alloc_if(0) free_if(1)) finalize(X); 17. 12. 2015 by Martin Kruliš (v1.1)12

13.  Asynchronous Operations ◦ Execution char sigVar; #pragma offload target(mic) signal(&sigVar) long_lasting_func();... concurrent CPU work... ◦ Data transfers #pragma offload_transfer clauses signal(&sigVar) ◦ Waiting, polling #pragma offload_wait wait(&sigVar) if (_Offload_signaled(micID, &sigVar))... 17. 12. 2015 by Martin Kruliš (v1.1)13

14.  Shared Virtual Memory Mode ◦ Code and variables that are shared have attribute _Cilk_shared (e.g., float _Cilk_shared *X ) ◦ Shared memory allocation must be performed via specialized functions  _Offload_shared_malloc, _Offload_shared_free ◦ Offloaded code gets executed by _Cilk_offload _Cilk_offload_to(micId) ◦ The data are transferred as compiler see fit 17. 12. 2015 by Martin Kruliš (v1.1)14 Example

15.  A Few More Things ◦ Compile time macros for MIC detection __INTEL_OFFLOAD, __MIC__ ◦ Runtime MIC detection  _Offload_number_of_devices()  _Offload_get_device_number()  Querying device capabilities is more complicated  You can read /proc/cpuinfo on the device  Or use special MicAccessAPI ◦ Asynchronous data transfers and execution  Using signal variables for synchronization 17. 12. 2015 by Martin Kruliš (v1.1)15

16. 17. 12. 2015 by Martin Kruliš (v1.1)16