Xeon Phi MICROSOFT* WINDOWS* 1.4 1

Transcription

1 Xeon Phi MICROSOFT* WINDOWS* 1.4 1

2 Xeon Phi MICROSOFT* WINDOWS* :... 4 : Xeon Phi... 5 ( MIC) Xeon Phi Xeon Phi Xeon Phi uos Xeon Phi / : : Xeon Phi Xeon Phi

3 Xeon Phi MICROSOFT* WINDOWS* Xeon Phi Cilk Plus Xeon Phi Cilk Plus Xeon Phi TBB MKL SGEMM MKL Xeon Phi Xeon Phi A: Linux*

4 Xeon Phi MICROSOFT* WINDOWS* ( MIC) Xeon Phi Microsoft* Windows* C/C++ Fortran : 1. MPSS 2. MIC 3. Xeon Phi Parallel Studio XE 2015 Composer Edition for Windows* 4. ( MKL) 5. Xeon Phi 6. (BKM) : 1. ( ) 2. PCIe* XeonPhi Xeon (OS) : Windows* 7 Enterprise SP1 (64 )Windows* 8/8.1 Enterprise (64 )Windows Server* 2008 R2 SP1 (64 )Windows Server* 2012 (64 )Windows Server* 2012 R2 (64 ) Xeon Phi uos Linux* Xeon Phi ISA / I/O 1 VPU SIMD (Single Instruction Multiple Data) CPU NAcc Xeon Phi MKL C/C for Windows* Visual Fortran 15.0 for Windows* Xeon Phi 1 Intel acronyms dictionary, 8/6/2009, 4

5 Xeon Phi MICROSOFT* WINDOWS* SDP Xeon Phi KNC Xeon Phi Xeon Phi ( : Knights Corner) MPSS Xeon Phi SCIF Xeon Phi Xeon SCIF API PCIe ( Xeon Phi ) 2 Xeon PCIe* x Xeon Phi GPU Xeon Phi Xeon Phi 1: 5

6 Xeon Phi MICROSOFT* WINDOWS* Xeon Phi 1 : Xeon Phi Windows* : Xeon Phi : / PCIe* : <MPSS-installdir> \bin\micinfo.exe <MPSS-install-dir>\bin\MicFlash.exe <MPSS-install-dir>\bin\micctrl.exe <MPSS-install-dir>\bin\micsmc.exe <MPSS-install-dir>\service RAS MicRas MicRas Windows* <MPSS-install-dir> "c:\program Files\Intel\MPSS" OS (uos): Xeon Phi Linux* : uos Linux* SCIF (COIMYO) 6

7 Xeon Phi MICROSOFT* WINDOWS* ( MIC) Xeon Phi 50 MIC 1GHz ( 1.3GHz) MIC x86 ISA SIMD 4 2: MIC (VPU) VPU SIMD ISA VPU MIC Xeon Phi VPU MIC SIMD ISA ( MMX SSE AVX ) 32KB L1 32KB L1 512KB L2 L2 32MB LLC Xeon Phi () 7

8 Xeon Phi MICROSOFT* WINDOWS* Parallel Studio XE 2015 Composer Edition for Windows* ( ) TOOLS & DOWNLOADSSoftware Drivers:Intel Manycore Platform Software Stack(Intel M Downloads Microsoft* Windows* Readme (readme-windows.pdf) (releasenotes-windows.txt) MPSS (MPSS_Users_Guide-windows.pdf) 2. Microsoft* Windows* 7 Enterprise SP1 (64 ) Microsoft* Windows* 8 Enterprise (64 ) Microsoft* Windows Server* 2008 R2 SP1 (64 ) Microsoft* Windows Server* 2012 (64 ) Microsoft* Windows Server* 2012 R2 (64 ) NET Framework 4.0 ( uos PuTTY* PuTTYgen* () 5. Readme Preliminary Steps Windows* (mpss-3.*-windows.zip) 8. zip Windows* (Intel(R) Xeon Phi(TM) coprocessor.msi Intel(R) Xeon Phi(TM) coprocessor essentials.msi) 9. Readme Windows* Intel(R) Xeon Phi(TM) coprocessor.msi Xeon Phi Windows* [ ] MPSS c:\program Files\Intel\MPSS [Intel ] Xeon Phi (Intel(R) Xeon Phi(TM) coprocessor essentials.msi) 10. [ ] > [ ] > [ ] MPSS 3 Intel(R) Xeon Phi(TM) coprocessor Intel(R) Xeon Phi(TM) coprocessor essentials 8

9 Xeon Phi MICROSOFT* WINDOWS* 3: [ ] MPSS 11. Readme-windows.pdf Xeon Phi ([ ] > [ ] > ] [ ] ) 9

10 Xeon Phi MICROSOFT* WINDOWS* 4: (R) Xeon Phi(TM) 14. Xeon Phi ( ) MPSS prompt> micctrl --start 15. micinfo prompt> micinfo.exe 10

11 Xeon Phi MICROSOFT* WINDOWS* 5: micinfo.exe Driver Version 3.4.* MPSS Version 3.4.* Flash Version 2.1.*.* 11

12 Xeon Phi MICROSOFT* WINDOWS* Parallel Studio XE 2015 Composer Edition for C++ Windows* Parallel Studio XE 2015 Composer Edition for Fortran Windows* ( ( MKL ) Parallel Studio XE 2015 Composer Edition for C++ Windows* Parallel Studio XE 2015 Composer Edition for Fortran Windows* [ ] (.EXE) Xeon Phi "set OFFLOAD_REPORT=3" (<install-dir>\samples\ja_jp\c++\mic_samples <installdir>\samples\ja_jp\fortran\mic_samples) ( "MIC:" ) () 3. SEP Vtune Amplifier XE 2015 SEP a) <VTune-installdir>\bin64\k1om\ prompt>.\sep_micboot_install.cmd b) MIC () ( ) prompt> micctrl --start prompt> micctrl -w micctrl w micx: online c) d) prompt> micctrl --stop prompt>.\sep_micboot_uninstall.cmd prompt> micctrl --start 12

13 Xeon Phi MICROSOFT* WINDOWS* prompt> micctrl w Xeon Phi Xeon Phi Xeon Phi micinfo [ ] > [ ] > ] [ ] [ ] [ ] prompt> micctrl --start prompt> micctrl -w prompt> micinfo Xeon Phi Xeon Phi PuTTY* Linux* PuTTY* 2 prompt> micctrl -s <micx> MPSS prompt> micctrl r <micx> prompt> micctrl -w prompt> micctrl b <micx> prompt> micctrl -w prompt> micinfo MPSS prompt> micctrl --stop prompt> micctrl --start prompt> micctrl -w prompt> micinfo 13

14 Xeon Phi MICROSOFT* WINDOWS* Xeon Phi uos IP <coprocessor>.100 IP <coprocessor>.99 mic<coprocessor> mic0 IP IP mic Linux* PuTTY* root root Linux* WinSCP* PuTTY* download.html ( ) "<MPSS-installdir>\bin" SSH PuTTYgen* <MPSS-install-dir>\bin 6: PuTTY* PuTTgen* <MPSS-install-dir>\bin PuTTYgen* [Generate] SSH ( 7) 14

15 Xeon Phi MICROSOFT* WINDOWS* 7: PuTTYgen* authorized_keys (.txt ) <MPSSinstall-dir>\bin [Public key pasting into OpenSSH authorized_keys file] authorized_keys [Save private key] id_rsa.ppk id_rsa.ppk <MPSS-install-dir>\bin ([ ] ) <MPSS-install-dir>\bin prompt> micctrl --addssh root f "<MPSS-install-dir>\bin\authorized_keys" "micctrl --stop" "micctrl --start" prompt> micctrl --stop prompt> micctrl --start PuTTY* PuTTY* [Host Name (or IP address)] root@ (mic0 ) (mic1 root@ ) 15

16 Xeon Phi MICROSOFT* WINDOWS* 8: PuTTY* [Connection] > [SSH] [Auth] [Private key file for authentication] [Browse] id_rsa.ppk 9: [Open] ( ) 16

17 Xeon Phi MICROSOFT* WINDOWS* 10: mic0 WinSCP* () WinSCP* [File protocol] SCP [Host name] [User name] root [Private key file] (application.mic) mic0 17

18 Xeon Phi MICROSOFT* WINDOWS* 11: WinSCP* [Login] GUI mic0 "ls" Xeon Phi Xeon Phi (NFS) MPSS 12.2 [ ] ] [ ] [ ] 18

19 Xeon Phi MICROSOFT* WINDOWS* 12: MPSS <MPSS-install-dir>\bin root micinfo - micflash - / micctrl micras Windows* Xeon Phi RAS RAS micsmc (GUI) 2 19

20 Xeon Phi MICROSOFT* WINDOWS* Xeon Phi MPSS Xeon Phi MPSS 6 Xeon Phi MIC SIMD (C/C++ FORTRAN ) ( ) SIMD / Xeon Phi MIC API Xeon Phi MIC : o Parallel Studio XE 2015 Composer Edition for C++ Windows* 64 MIC o Parallel Studio XE 2015 Composer Edition for Fortran Windows* 64 MIC (): o ( MKL) MIC MKL Xeon Phi Parallel Studio XE 2015 Composer Edition for C++ Windows* Parallel Studio XE 2015 Composer Edition for Fortran Windows* o ( TBB) o ( IPP) : 20

21 Xeon Phi MICROSOFT* WINDOWS* Debugger Extension o 64 MIC o VTune Amplifier XE 2015 for Windows*Windows* Xeon Phi Microsoft* Visual Studio* 2008 o Parallel Studio XE 2015 Composer Edition for C++ Windows* Parallel Studio XE 2015 Composer Edition for Fortran Windows*: [ ] > [ ] > [Intel Parallel Studio XE 2015 ( (R) Parallel Studio XE 2015)] > [Command Prompt ( )] > [Parallel Studio XE 2015 Composer Edition ( (R) Parallel Studio XE 2015 Composer Edition)] > [Intel 64 Visual Studio XXXX mode ( (R) 64 Visual Studio XXXX )] <install-dir>\documentation\ja_jp\ o o o compiler_c\cl\compiler_ug_c compiler_f\cl\compiler_ug_f - C for Windows* Visual Fortran 15.0 for Windows* ([ ] > [ ] > [Intel Parallel Studio XE 2015 ( (R) Parallel Studio XE 2015)] > [Documentation ( )] ) MIC > MIC > MIC MIC > > MIC Release_Notes-C-2015_W_JA.pdf Release_Notes-F-2015_W_ JA.pdf - MIC MIC 4 : Parallel Studio XE 2015 Composer Edition for C++ Windows* Parallel Studio XE 2015 Composer Edition for Fortran Windows* (Release_Notes-*-2015_ W_EN.pdf) <install-dir>\documentation\ja_jp\debugger\migdb_quickstart_win.pdf () Xeon Phi Visual Studio* Windows* Visual Studio* GDB Xeon Phi 21

22 Xeon Phi MICROSOFT* WINDOWS* Xeon Phi o o MKL <install-dir>\documentation\ja_jp\mkl mkl_documentation.htm ( [ ] > [ ] > [Intel Parallel Studio XE 2015 ( (R) Parallel Studio XE 2015)] > [Documentation ( )] ) Xeon Phi MKL MKL VTune Amplifier XE 2015 for Windows* Xeon Phi <VTune-install-dir>\documentation \en\tutorials\find_hotspots\c++\index.htm () (Linux* ) Web : o Xeon Phi o RESOURCES (including downloads) Xeon Phi uos GDB MIC Xeon Phi () ABI System V Application Binary Interface K1OM Architecture Processor Supplement () Xeon Phi : o o o C++: <install-dir>\samples\ja_jp\c++\mic_samples Fortran: <install-dir>\samples\ja_jp\fortran\mic_samples\ MKL: o <install-dir>\mkl\examples\examples_mic\mic_offload : o C: <install-dir>\samples\ja_jp\c++\mic_sample\leo_tutorial Xeon Phi.dll Xeon Phi release-notes-*-2015-w-ja.pdf Xeon Phi > > /Qoffload-option /Qoffload-attribute-target /Qoffload-attribute-target ( )/Qoffload - 22

23 Xeon Phi MICROSOFT* WINDOWS* /Qoffload:none) ( ) OFFLOAD_REPORT prompt> set OFFLOAD_REPORT=3 1 2 prompt> set OFFLOAD_REPORT=<1 2> > CPU 2 CPU Xeon Phi (C/C++) (Fortran) ( MKL) CPU Xeon Phi : Xeon Phi Xeon Phi Xeon Phi Xeon Phi ans = a[0] + a[1] + + a[n-1] : C

24 Xeon Phi MICROSOFT* WINDOWS* float reduction_serial(float *data, int size) float ret = 0.f; for (int i=0; i<size; ++i) ret += data[i]; return ret; 1: (C/C++) ( ) #pragma offload target(mic) Xeon Phi ( ) ( ) in out inout ( ) () ret 1 MIC 1 float reduction_offload(float *data, int size) float ret = 0.f; #pragma offload target(mic) in(data:length(size)) for (int i=0; i<size; ++i) ret += data[i]; return ret; 2: Xeon Phi VPU Cilk Plus MIC

25 Xeon Phi MICROSOFT* WINDOWS* sec_reduce_add() float reduction_vectorreduction(float *data, int size) float ret = 0; #pragma offload target(mic) in(data:length(size)) ret = sec_reduce_add(data[0:size]); // Cilk Plus return ret; 3: (C/C++) Xeon Phi C++ > MIC > MIC <install-dir>\samples\ja_jp\c++\mic_samples\ intro_samplec\samplec13.c C/C++ C/C++ > MIC > MIC > > Parallel Studio XE 2015 Composer Edition C/C++ 2 (_Cilk_shared _Cilk_offload) ( Fortran ) (_Cilk_shared ) _Cilk_offload / API: void *_Offload_shared_malloc(size_t size); _Offload_shared_free(void *p); API: void *_Offload_shared_aligned_malloc(size_t size, size_t alignment); _Offload_shared_aligned_free(void *p); 25

26 Xeon Phi MICROSOFT* WINDOWS* 2 1 (_Cilk_offload ) _Cilk_shared _Cilk_offload float * _Cilk_shared data; // _Cilk_shared float MIC_OMPReduction(int size) float Result; #ifdef MIC int nthreads = 60; omp_set_num_threads(nthreads); #pragma omp parallel for reduction(+:result) for (int i=0; i<size; ++i) Result += data[i]; return Result; #else printf("intel(r) Xeon Phi(TM) Coprocessor not available\n"); #endif return 0.0f; int main() size_t size = 1*1e6; int n_bytes = size*sizeof(float); float Result; data = (_Cilk_shared float *)_Offload_shared_malloc (n_bytes); for (int i=0; i<size; ++i) data[i] = i%10; Result = _Cilk_offload MIC_OMPReduction(size); Printf( Cilk Offload Result=%.0f\n,Result); _Offload_shared_free(data); return 0; 4: _Cilk_shared _Cilk_offload (C/C++) C: <install-dir>\samples\ja_jp\c++\mic_samples shrd_samplec LEO_tutorial C++: <install-dir>\samples\ja_jp\c++\mic_samples\shrd_samplecpp 26

27 Xeon Phi MICROSOFT* WINDOWS* C++ Fortran C++ > MIC > MIC > > Xeon Phi Xeon Phi ( NFS ) : 1. <install-dir>\samples\ja_jp\c++\openmp_samples.zip openmp_sample.c 2. /Qmic icl /Qmic -openmp openmp_sample.c 3. WinSCP* a.out mic0 ( /tmp ) 4. c:\program Files (x86)\common Files\Intel\Shared Libraries\compiler\lib \mic\lib\libiomp5.so OpenMP* /tmp 5. PuTTY* ( OpenMP* ) export LD_LIBRARY_PATH=/tmp 6. ulimit s unlimited 7. /tmp a.out cd /tmp./a.out 27

28 Xeon Phi MICROSOFT* WINDOWS* Xeon Phi 1. ( TBB) 2. OpenMP* 3. Cilk Plus 4. Pthreads* Xeon Phi Xeon Phi : OpenMP* CPU OpenMP* Xeon Phi OpenMP* / OpenMP* 1 Xeon Phi OpenMP* Xeon Phi OpenMP* CPU Xeon Phi CPU Xeon Phi omp parallel Xeon Phi 4 1 uos 4 ( ) OpenMP* 1 CPU Xeon Phi float OMP_reduction_OMP(float *data, int size) float ret = 0; #pragma offload target(mic) in(size) in(data:length(size)) #pragma omp parallel for reduction(+:ret) for (int i=0; i<size; ++i) ret += data[i]; return ret; 5: OpenMP* (C/C++) OpenMP* Fortran <install-dir>\samples \ja_jp\fortran\mic_samples\leo_fortran_intro 28

29 Xeon Phi MICROSOFT* WINDOWS* real function FTNReductionOMP(data, size) implicit none integer :: size, i real, dimension(size) :: data real :: ret = 0.0!dir$ omp offload target(mic) in(size) in(data:length(size))!$omp parallel do reduction(+:ret) do i=1,size ret = ret + data(i) enddo!$omp end parallel do FTNReductionOMP = ret return end function FTNReductionOMP 6: OpenMP* (Fortran) Xeon Phi : OpenMP* + Cilk Plus OpenMP* Cilk Plus Cilk Plus sec_reduce_add() MIC float OMPnthreads_CilkPlusEAN_reduction(float *data, int size) float ret=0; #pragma offload target(mic) in(data:length(size)) int nthreads = omp_get_max_threads(); int ElementsPerThread = size/nthreads; #pragma omp parallel for reduction(+:ret) for(int i=0;i<nthreads;i++) ret = sec_reduce_add( data[i*elementsperthread:elementsperthread]); // for(int i=nthreads*elementsperthread; i<size; i++) ret+=data[i]; return ret; 7: Open MP* Cilk Plus (C/C++) 29

30 Xeon Phi MICROSOFT* WINDOWS* Xeon Phi : Cilk Plus Cilk Plus Cilk Plus MIC #pragma offload_attribute(push,target(mic)) #pragma offload_attribute(pop) #pragma offload_attribute(push,target(mic)) #include <cilk/cilk.h> #include <cilk/reducer_opadd.h> #pragma offload_attribute(pop) 8: (C/C++) cilk_for float ReduceCilk(float*data, int size) float ret = 0; #pragma offload target(mic) in(data:length(size)) cilk::reducer_opadd<int> total; cilk_for (int i=0; i<size; ++i) total += data[i]; ret = total.get_value(); return ret; 9: cilk_for Xeon Phi : TBB Cilk Plus TBB Cilk Plus MIC #pragma offload_attribute (push,target(mic)) #include "tbb/task_scheduler_init.h" #include "tbb/blocked_range.h" #include "tbb/parallel_reduce.h" #include "tbb/task.h" #pragma offload_attribute (pop) using namespace tbb; 10: TBB (C/C++) 30

31 Xeon Phi MICROSOFT* WINDOWS* Xeon Phi declspec(target(mic)) parallel_reduce 1 () join 1. MIC declspec(target(mic)) #ifdef MIC class declspec(target(mic)) ReduceTBB private: float *my_data; public: float sum; void operator()( const blocked_range<size_t>& r ) float *data = my_data; for( size_t i=r.begin(); i!=r.end(); ++i) sum += data[i]; ReduceTBB( ReduceTBB& x, split) : my_data(x.my_data), sum(0) void join( const ReduceTBB& y) sum += y.sum; ReduceTBB( float data[] ) : my_data(data), sum(0) ; #endif 11: MIC TBB (C/C++) 2. Xeon Phi declspec(target(mic)) 31

32 Xeon Phi MICROSOFT* WINDOWS* declspec(target(mic)) float MICReductionTBB(float *data, int size) ReduceTBB redc(data); // task_scheduler_init init; parallel_reduce(blocked_range<size_t>(0, size), redc); return redc.sum; 12: MIC TBB (C/C++) 3. #pragma offload target(mic) TBB float MICReductionTBB(float *data, int size) float ret(0.f); #pragma offload target(mic) in(size) in(data:length(size)) out(ret) ret = _MICReductionTBB(data, size); return ret; 13: TBB (C/C++) : TBB /Qtbb 32

33 Xeon Phi MICROSOFT* WINDOWS* MKL MKL (NAcc) NAcc Xeon Phi NAcc BLAS LAPACK FFT VML VSL ( MKL MKL NAcc MIC MKL Xeon Phi SGEMM 13: MKL BLAS SGEMM sgemm 1: 2: #pragma offload Xeon Phi free_if(0) Xeon Phi #define PHI_DEV 0 #pragma offload_transfer target(mic:phi_dev) \ in(a:length(matrix_elements) free_if(0)) \ in(b:length(matrix_elements) free_if(0)) \ in(c:length(matrix_elements) free_if(0)) 14: Xeon Phi 3: sgemm Xeon Phi MKL NAcc nocopy() 2 33

34 Xeon Phi MICROSOFT* WINDOWS* #pragma offload target(mic:phi_dev) \ in(transa, transb, N, alpha, beta) \ nocopy(a: alloc_if(0) free_if(0)) nocopy(b: alloc_if(0) free_if(0)) \ out(c:length(matrix_elements) alloc_if(0) free_if(0)) // output data sgemm(&transa, &transb, &N, &N, &N, &alpha, A, &N, B, &N, &beta, C, &N); 15: sgemm 4: 2 alloc_if(0) free_if(1) #pragma offload_transfer target(mic:phi_dev) \ nocopy(a:length(matrix_elements) alloc_if(0) free_if(1)) \ nocopy(b:length(matrix_elements) alloc_if(0) free_if(1)) \ nocopy(c:length(matrix_elements) alloc_if(0) free_if(1)) 16: MKL OpenMP* #pragma offload target(mic:phi_dev) \ in(transa, transb, N, alpha, beta) \ nocopy(a: alloc_if(0) free_if(0)) nocopy(b: alloc_if(0) free_if(0)) out(c:length(matrix_elements) alloc_if(0) free_if(0)) // output data omp_set_num_threads(64); // set num threads in openmp sgemm(&transa, &transb, &N, &N, &N, &alpha, A, &N, B, &N, &beta, C, &N); 17: omp_set_num_threads() Xeon Phi MKL MKL mkl_mic_enable() MKL Xeon Phi mkl_mic_disable() _Cilk_offload #pragma offload Xeon Phi MKL 34

35 Xeon Phi MICROSOFT* WINDOWS* _Cilk_offload #pragma offload (_Cilk_offload #pragma offload ) <install-dir>\mkl\examples\examples_ mic\mic_ao\blasc (C ) <install-dir>\mkl\examples\examples_mic\mic_ao\blasf (Fortran ) Xeon Phi MIC <install-dir> \Documentation\ja_JP\debugger\gdb\pdf\vsmigdb_config_guide.pdf Xeon Phi VTune Amplifier XE 2015 for Windows* Xeon Phi 2: 35

36 Xeon Phi MICROSOFT* WINDOWS* A: Linux* Xeon Phi Linux* 1. exit: > exit 2. ls: ls l > ls a.out libioomp5.so 3. pwd: > pwd /root 4. cd <path>: > cd /tmp 5. ps: > ps 5847 root 0:00 /sbin/sshd 5914 micuser 2:47 /bin/coi_daemon coiuser=micuser 6. kill -9 <pid>: ID > kill top: CPU > top 8. cp <sourcefilename> <destinationfilename>: > cp file1 file2 9. rm <filename>: > rm file1 10. less <filename>: > less file2 36

37 Xeon Phi MICROSOFT* WINDOWS* 11. grep <pattern>: coi grep ps > ps grep coi 5914 micuser 2:47 /bin/coi_daemon coiuser=micuser 12. export: > export LD_LIBRARY_PATH=/tmp 13. Ctrl + C: 37

38 Xeon Phi MICROSOFT* WINDOWS* Loc Q Nguyen MBA & 38

39 Xeon Phi MICROSOFT* WINDOWS* Intel's Terms and Conditions of Sale ( ) ( ) Web ( IntelIntel Cilk Xeon Phi VTune Xeon / Intel Corporation * 2015 Intel Corporation. 39

40 Xeon Phi MICROSOFT* WINDOWS* SIMD 2 ( SSE2) SIMD 3 ( SSE3) SIMD 3 (SSE3) #