ChangeLog 32 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711
  1. # StarPU --- Runtime system for heterogeneous multicore architectures.
  2. #
  3. # Copyright (C) 2009-2014 Université de Bordeaux
  4. # Copyright (C) 2010, 2011, 2012, 2013, 2014 Centre National de la Recherche Scientifique
  5. # Copyright (C) 2014 Inria
  6. #
  7. # StarPU is free software; you can redistribute it and/or modify
  8. # it under the terms of the GNU Lesser General Public License as published by
  9. # the Free Software Foundation; either version 2.1 of the License, or (at
  10. # your option) any later version.
  11. #
  12. # StarPU is distributed in the hope that it will be useful, but
  13. # WITHOUT ANY WARRANTY; without even the implied warranty of
  14. # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  15. #
  16. # See the GNU Lesser General Public License in COPYING.LGPL for more details.
  17. StarPU 1.2.0 (svn revision xxxx)
  18. ==============================================
  19. New features:
  20. * MIC Xeon Phi support
  21. * SCC support
  22. * New function starpu_sched_ctx_exec_parallel_code to execute a
  23. parallel code on the workers of the given scheduler context
  24. * MPI:
  25. - New internal communication system : a unique tag called
  26. is now used for all communications, and a system
  27. of hashmaps on each node which stores pending receives has been
  28. implemented. Every message is now coupled with an envelope, sent
  29. before the corresponding data, which allows the receiver to
  30. allocate data correctly, and to submit the matching receive of
  31. the envelope.
  32. - New function
  33. starpu_mpi_irecv_detached_sequential_consistency which
  34. allows to enable or disable the sequential consistency for
  35. the given data handle (sequential consistency will be
  36. enabled or disabled based on the value of the function
  37. parameter and the value of the sequential consistency
  38. defined for the given data)
  39. - New functions starpu_mpi_task_build() and
  40. starpu_mpi_task_post_build()
  41. - New flag STARPU_NODE_SELECTION_POLICY to specify a policy for
  42. selecting a node to execute the codelet when several nodes
  43. own data in W mode.
  44. * New STARPU_COMMUTE flag which can be passed along STARPU_W or STARPU_RW to
  45. let starpu commute write accesses.
  46. * Out-of-core support, through registration of disk areas as additional memory
  47. nodes. It can be enabled programmatically or through the STARPU_DISK_SWAP*
  48. environment variables.
  49. * Reclaiming is now periodically done before memory becomes full. This can
  50. be controlled through the STARPU_*_AVAILABLE_MEM environment variables.
  51. * New hierarchical schedulers which allow the user to easily build
  52. its own scheduler, by coding itself each "box" it wants, or by
  53. combining existing boxes in StarPU to build it. Hierarchical
  54. schedulers have very interesting scalability properties.
  55. * Add STARPU_CUDA_ASYNC and STARPU_OPENCL_ASYNC flags to allow asynchronous
  56. CUDA and OpenCL kernel execution.
  57. * Add STARPU_CUDA_PIPELINE and STARPU_OPENCL_PIPELINE to specify how
  58. many asynchronous tasks are submitted in advance on CUDA and
  59. OpenCL devices. Setting the value to 0 forces a synchronous
  60. execution of all tasks.
  61. * Add CUDA concurrent kernel execution support through
  62. the STARPU_NWORKER_PER_CUDA environment variable.
  63. * Add CUDA and OpenCL kernel submission pipelining, to overlap costs and allow
  64. concurrent kernel execution on Fermi cards.
  65. * New locality work stealing scheduler (lws).
  66. * Add STARPU_VARIABLE_NBUFFERS to be set in cl.nbuffers, and nbuffers and
  67. modes field to the task structure, which permit to define codelets taking a
  68. variable number of data.
  69. * Add support for implementing OpenMP runtimes on top of StarPU
  70. * New performance model format to better represent parallel tasks.
  71. Used to provide estimations for the execution times of the
  72. parallel tasks on scheduling contexts or combined workers.
  73. Small features:
  74. * Tasks can now have a name (via the field const char *name of
  75. struct starpu_task)
  76. * New functions starpu_data_acquire_cb_sequential_consistency() and
  77. starpu_data_acquire_on_node_cb_sequential_consistency() which allows
  78. to enable or disable sequential consistency
  79. * New configure option --enable-fxt-lock which enables additional
  80. trace events focused on locks behaviour during the execution
  81. * Functions starpu_insert_task and starpu_mpi_insert_task are
  82. renamed in starpu_task_insert and starpu_mpi_task_insert. Old
  83. names are kept to avoid breaking old codes.
  84. * New configure option --enable-calibration-heuristic which allows
  85. the user to set the maximum authorized deviation of the
  86. history-based calibrator.
  87. * Allow application to provide the task footprint itself.
  88. * New function starpu_sched_ctx_display_workers() to display worker
  89. information belonging to a given scheduler context
  90. * The option --enable-verbose can be called with
  91. --enable-verbose=extra to increase the verbosity
  92. * Add codelet size, footprint and tag id in the paje trace.
  93. * Add STARPU_TAG_ONLY, to specify a tag for traces without making StarPU
  94. manage the tag.
  95. * On Linux x86, spinlocks now block after a hundred tries. This avoids
  96. typical 10ms pauses when the application thread tries to submit tasks.
  97. * New function char *starpu_worker_get_type_as_string(enum starpu_worker_archtype type)
  98. * Improve static scheduling by adding support for specifying the task
  99. execution order.
  100. * Add starpu_worker_can_execute_task_impl and
  101. starpu_worker_can_execute_task_first_impl to optimize getting the
  102. working implementations
  103. * Add STARPU_MALLOC_NORECLAIM flag to allocate without running a reclaim if
  104. the node is out of memory.
  105. * Add starpu_memory_allocate and _deallocate to let the application declare
  106. its own allocation to the reclaiming engine.
  107. * New flag STARPU_DATA_MODE_ARRAY for the function family
  108. starpu_task_insert to allow to define a array of data handles
  109. along with their access modes.
  110. * New configure option --enable-new-check to enable new testcases
  111. which are known to fail
  112. Changes:
  113. * Data interfaces (variable, vector, matrix and block) now define
  114. pack und unpack functions
  115. * StarPU-MPI: Fix for being able to receive data which have not yet
  116. been registered by the application (i.e it did not call
  117. starpu_data_set_tag(), data are received as a raw memory)
  118. * StarPU-MPI: Fix for being able to receive data with the same tag
  119. from several nodes (see mpi/tests/gather.c)
  120. * Remove the long-deprecated cost_model fields and task->buffers field.
  121. * Fix complexity of implicit task/data dependency, from quadratic to linear.
  122. Small changes:
  123. * Rename function starpu_trace_user_event() as
  124. starpu_fxt_trace_user_event()
  125. StarPU 1.1.4 (svn revision xxx)
  126. ==============================================
  127. The scheduling context release
  128. New features:
  129. * Fix and actually enable the cache allocation.
  130. * Enable allocation cache in main RAM when STARPU_LIMIT_CPU_MEM is set by
  131. the user.
  132. * New MPI functions starpu_mpi_issend and starpu_mpi_issend_detached
  133. to send data using a synchronous and non-blocking mode (internally
  134. uses MPI_Issend)
  135. Changes:
  136. * Fix complexity of implicit task/data dependency, from quadratic to linear.
  137. StarPU 1.1.3 (svn revision 13450)
  138. ==============================================
  139. The scheduling context release
  140. New features:
  141. * One can register an existing on-GPU buffer to be used by a handle.
  142. * Add the starpu_paje_summary statistics tool.
  143. * Enable gpu-gpu transfers for matrices.
  144. * Let interfaces declare which transfers they allow with the can_copy
  145. methode.
  146. Small changes:
  147. * Lock performance model files while writing and reading them to avoid
  148. issues on parallel launches, MPI runs notably.
  149. * Lots of build fixes for icc on Windows.
  150. StarPU 1.1.2 (svn revision 13011)
  151. ==============================================
  152. The scheduling context release
  153. New features:
  154. * The reduction init codelet is automatically used to initialize temporary
  155. buffers.
  156. * Traces now include a "scheduling" state, to show the overhead of the
  157. scheduler.
  158. * Add STARPU_CALIBRATE_MINIMUM environment variable to specify the minimum
  159. number of calibration measurements.
  160. * Add STARPU_TRACE_BUFFER_SIZE environment variable to specify the size of
  161. the trace buffer.
  162. StarPU 1.1.1 (svn revision 12638)
  163. ==============================================
  164. The scheduling context release
  165. New features:
  166. * MPI:
  167. - New variable STARPU_MPI_CACHE_STATS to print statistics on
  168. cache holding received data.
  169. - New function starpu_mpi_data_register() which sets the rank
  170. and tag of a data, and also allows to automatically clear
  171. the MPI communication cache when unregistering the data. It
  172. should be called instead of both calling
  173. starpu_data_set_tag() and starpu_data_set_rank()
  174. * Use streams for all CUDA transfers, even initiated by CPUs.
  175. * Add paje traces statistics tools.
  176. * Use streams for GPUA->GPUB and GPUB->GPUA transfers.
  177. Small features:
  178. * New STARPU_EXECUTE_ON_WORKER flag to specify the worker on which
  179. to execute the task.
  180. * New STARPU_DISABLE_PINNING environment variable to disable host memory
  181. pinning.
  182. * New STARPU_DISABLE_KERNELS environment variable to disable actual kernel
  183. execution.
  184. * New starpu_memory_get_total function to get the size of a memory node.
  185. * New starpu_parallel_task_barrier_init_n function to let a scheduler decide
  186. a set of workers without going through combined workers.
  187. Changes:
  188. * Fix simgrid execution.
  189. * Rename starpu_get_nready_tasks_of_sched_ctx to starpu_sched_ctx_get_nready_tasks
  190. * Rename starpu_get_nready_flops_of_sched_ctx to starpu_sched_ctx_get_nready_flops
  191. * New functions starpu_pause() and starpu_resume()
  192. * New codelet specific_nodes field to specify explicit target nodes for data.
  193. * StarPU-MPI: Fix overzealous allocation of memory.
  194. * Interfaces: Allow interface implementation to change pointers at will, in
  195. unpack notably.
  196. Small changes:
  197. * Use big fat abortions when one tries to make a task or callback
  198. sleep, instead of just returning EDEADLCK which few people will test
  199. * By default, StarPU FFT examples are not compiled and checked, the
  200. configure option --enable-starpufft-examples needs to be specified
  201. to change this behaviour.
  202. StarPU 1.1.0 (svn revision 11960)
  203. ==============================================
  204. The scheduling context release
  205. New features:
  206. * OpenGL interoperability support.
  207. * Capability to store compiled OpenCL kernels on the file system
  208. * Capability to load compiled OpenCL kernels
  209. * Performance models measurements can now be provided explicitly by
  210. applications.
  211. * Capability to emit communication statistics when running MPI code
  212. * Add starpu_data_unregister_submit, starpu_data_acquire_on_node and
  213. starpu_data_invalidate_submit
  214. * New functionnality to wrapper starpu_insert_task to pass a array of
  215. data_handles via the parameter STARPU_DATA_ARRAY
  216. * Enable GPU-GPU direct transfers.
  217. * GCC plug-in
  218. - Add `registered' attribute
  219. - A new pass was added that warns about the use of possibly
  220. unregistered memory buffers.
  221. * SOCL
  222. - Manual mapping of commands on specific devices is now
  223. possible
  224. - SOCL does not require StarPU CPU tasks anymore. CPU workers
  225. are automatically disabled to enhance performance of OpenCL
  226. CPU devices
  227. * New interface: COO matrix.
  228. * Data interfaces: The pack operation of user-defined data interface
  229. defines a new parameter count which should be set to the size of
  230. the buffer created by the packing of the data.
  231. * MPI:
  232. - Communication statistics for MPI can only be enabled at
  233. execution time by defining the environment variable
  234. STARPU_COMM_STATS
  235. - Communication cache mechanism is enabled by default, and can
  236. only be disabled at execution time by setting the
  237. environment variable STARPU_MPI_CACHE to 0.
  238. - New variable STARPU_MPI_CACHE_STATS to print statistics on
  239. cache holding received data.
  240. - Initialisation functions starpu_mpi_initialize_extended()
  241. and starpu_mpi_initialize() have been made deprecated. One
  242. should now use starpu_mpi_init(int *, char ***, int). The
  243. last parameter indicates if MPI should be initialised.
  244. - Collective detached operations have new parameters, a
  245. callback function and a argument. This is to be consistent
  246. with the detached point-to-point communications.
  247. - When exchanging user-defined data interfaces, the size of
  248. the data is the size returned by the pack operation, i.e
  249. data with dynamic size can now be exchanged with StarPU-MPI.
  250. - New function starpu_mpi_data_register() which sets the rank
  251. and tag of a data, and also allows to automatically clear
  252. the MPI communication cache when unregistering the data. It
  253. should be called instead of both calling
  254. starpu_data_set_tag() and starpu_data_set_rank()
  255. * Add experimental simgrid support, to simulate execution with various
  256. number of CPUs, GPUs, amount of memory, etc.
  257. * Add support for OpenCL simulators (which provide simulated execution time)
  258. * Add support for Temanejo, a task graph debugger
  259. * Theoretical bound lp output now includes data transfer time.
  260. * Update OpenCL driver to only enable CPU devices (the environment
  261. variable STARPU_OPENCL_ONLY_ON_CPUS must be set to a positive
  262. value when executing an application)
  263. * Add Scheduling contexts to separate computation resources
  264. - Scheduling policies take into account the set of resources corresponding
  265. to the context it belongs to
  266. - Add support to dynamically change scheduling contexts
  267. (Create and Delete a context, Add Workers to a context, Remove workers from a context)
  268. - Add support to indicate to which contexts the tasks are submitted
  269. * Add the Hypervisor to manage the Scheduling Contexts automatically
  270. - The Contexts can be registered to the Hypervisor
  271. - Only the registered contexts are managed by the Hypervisor
  272. - The Hypervisor can detect the initial distribution of resources of
  273. a context and constructs it consequently (the cost of execution is required)
  274. - Several policies can adapt dynamically the distribution of resources
  275. in contexts if the initial one was not appropriate
  276. - Add a platform to implement new policies of redistribution
  277. of resources
  278. * Implement a memory manager which checks the global amount of
  279. memory available on devices, and checks there is enough memory
  280. before doing an allocation on the device.
  281. * Discard environment variable STARPU_LIMIT_GPU_MEM and define
  282. instead STARPU_LIMIT_CUDA_MEM and STARPU_LIMIT_OPENCL_MEM
  283. * Introduce new variables STARPU_LIMIT_CUDA_devid_MEM and
  284. STARPU_LIMIT_OPENCL_devid_MEM to limit memory per specific device
  285. * Introduce new variable STARPU_LIMIT_CPU_MEM to limit memory for
  286. the CPU devices
  287. * New function starpu_malloc_flags to define a memory allocation with
  288. constraints based on the following values:
  289. - STARPU_MALLOC_PINNED specifies memory should be pinned
  290. - STARPU_MALLOC_COUNT specifies the memory allocation should be in
  291. the limits defined by the environment variables STARPU_LIMIT_xxx
  292. (see above). When no memory is left, starpu_malloc_flag tries
  293. to reclaim memory from StarPU and returns -ENOMEM on failure.
  294. * starpu_malloc calls starpu_malloc_flags with a value of flag set
  295. to STARPU_MALLOC_PINNED
  296. * Define new function starpu_free_flags similarly to starpu_malloc_flags
  297. * Define new public API starpu_pthread which is similar to the
  298. pthread API. It is provided with 2 implementations: a pthread one
  299. and a Simgrid one. Applications using StarPU and wishing to use
  300. the Simgrid StarPU features should use it.
  301. * Allow to have a dynamically allocated number of buffers per task,
  302. and so overwrite the value defined --enable-maxbuffers=XXX
  303. * Performance models files are now stored in a directory whose name
  304. include the version of the performance model format. The version
  305. number is also written in the file itself.
  306. When updating the format, the internal variable
  307. _STARPU_PERFMODEL_VERSION should be updated. It is then possible
  308. to switch easily between differents versions of StarPU having
  309. different performance model formats.
  310. * Tasks can now define a optional prologue callback which is executed
  311. on the host when the task becomes ready for execution, before getting
  312. scheduled.
  313. * Small CUDA allocations (<= 4MiB) are now batched to avoid the huge
  314. cudaMalloc overhead.
  315. * Prefetching is now done for all schedulers when it can be done whatever
  316. the scheduling decision.
  317. * Add a watchdog which permits to easily trigger a crash when StarPU gets
  318. stuck.
  319. * Document how to migrate data over MPI.
  320. * New function starpu_wakeup_worker() to be used by schedulers to
  321. wake up a single worker (instead of all workers) when submitting a
  322. single task.
  323. * The functions starpu_sched_set/get_min/max_priority set/get the
  324. priorities of the current scheduling context, i.e the one which
  325. was set by a call to starpu_sched_ctx_set_context() or the initial
  326. context if the function has not been called yet.
  327. * Fix for properly dealing with NAN on windows systems
  328. Small features:
  329. * Add starpu_worker_get_by_type and starpu_worker_get_by_devid
  330. * Add starpu_fxt_stop_profiling/starpu_fxt_start_profiling which permits to
  331. pause trace recording.
  332. * Add trace_buffer_size configuration field to permit to specify the tracing
  333. buffer size.
  334. * Add starpu_codelet_profile and starpu_codelet_histo_profile, tools which draw
  335. the profile of a codelet.
  336. * File STARPU-REVISION --- containing the SVN revision number from which
  337. StarPU was compiled --- is installed in the share/doc/starpu directory
  338. * starpu_perfmodel_plot can now directly draw GFlops curves.
  339. * New configure option --enable-mpi-progression-hook to enable the
  340. activity polling method for StarPU-MPI.
  341. * Permit to disable sequential consistency for a given task.
  342. * New macro STARPU_RELEASE_VERSION
  343. * New function starpu_get_version() to return as 3 integers the
  344. release version of StarPU.
  345. * Enable by default data allocation cache
  346. * New function starpu_perfmodel_directory() to print directory
  347. storing performance models. Available through the new option -d of
  348. the tool starpu_perfmodel_display
  349. * New batch files to execute StarPU applications under Microsoft
  350. Visual Studio (They are installed in path_to_starpu/bin/msvc)/
  351. * Add cl_arg_free, callback_arg_free, prologue_callback_arg_free fields to
  352. enable automatic free(cl_arg); free(callback_arg);
  353. free(prologue_callback_arg) on task destroy.
  354. * New function starpu_task_build
  355. * New configure options --with-simgrid-dir
  356. --with-simgrid-include-dir and --with-simgrid-lib-dir to specify
  357. the location of the SimGrid library
  358. Changes:
  359. * Rename all filter functions to follow the pattern
  360. starpu_DATATYPE_filter_FILTERTYPE. The script
  361. tools/dev/rename_filter.sh is provided to update your existing
  362. applications to use new filters function names.
  363. * Renaming of diverse functions and datatypes. The script
  364. tools/dev/rename.sh is provided to update your existing
  365. applications to use the new names. It is also possible to compile
  366. with the pkg-config package starpu-1.0 to keep using the old
  367. names. It is however recommended to update your code and to use
  368. the package starpu-1.1.
  369. * Fix the block filter functions.
  370. * Fix StarPU-MPI on Darwin.
  371. * The FxT code can now be used on systems other than Linux.
  372. * Keep only one hashtable implementation common/uthash.h
  373. * The cache of starpu_mpi_insert_task is fixed and thus now enabled by
  374. default.
  375. * Improve starpu_machine_display output.
  376. * Standardize objects name in the performance model API
  377. * SOCL
  378. - Virtual SOCL device has been removed
  379. - Automatic scheduling still available with command queues not
  380. assigned to any device
  381. - Remove modified OpenCL headers. ICD is now the only supported
  382. way to use SOCL.
  383. - SOCL test suite is only run when environment variable
  384. SOCL_OCL_LIB_OPENCL is defined. It should contain the location
  385. of the libOpenCL.so file of the OCL ICD implementation.
  386. * Fix main memory leak on multiple unregister/re-register.
  387. * Improve hwloc detection by configure
  388. * Cell:
  389. - It is no longer possible to enable the cell support via the
  390. gordon driver
  391. - Data interfaces no longer define functions to copy to and from
  392. SPU devices
  393. - Codelet no longer define pointer for Gordon implementations
  394. - Gordon workers are no longer enabled
  395. - Gordon performance models are no longer enabled
  396. * Fix data transfer arrows in paje traces
  397. * The "heft" scheduler no longer exists. Users should now pick "dmda"
  398. instead.
  399. * StarPU can now use poti to generate paje traces.
  400. * Rename scheduling policy "parallel greedy" to "parallel eager"
  401. * starpu_scheduler.h is no longer automatically included by
  402. starpu.h, it has to be manually included when needed
  403. * New batch files to run StarPU applications with Microsoft Visual C
  404. * Add examples/release/Makefile to test StarPU examples against an
  405. installed version of StarPU. That can also be used to test
  406. examples using a previous API.
  407. * Tutorial is installed in ${docdir}/tutorial
  408. * Schedulers eager_central_policy, dm and dmda no longer erroneously respect
  409. priorities. dmdas has to be used to respect priorities.
  410. * StarPU-MPI: Fix potential bug for user-defined datatypes. As MPI
  411. can reorder messages, we need to make sure the sending of the size
  412. of the data has been completed.
  413. * Documentation is now generated through doxygen.
  414. * Modification of perfmodels output format for future improvements.
  415. * Fix for properly dealing with NAN on windows systems
  416. * Function starpu_sched_ctx_create() now takes a variable argument
  417. list to define the scheduler to be used, and the minimum and
  418. maximum priority values
  419. * The functions starpu_sched_set/get_min/max_priority set/get the
  420. priorities of the current scheduling context, i.e the one which
  421. was set by a call to starpu_sched_ctx_set_context() or the initial
  422. context if the function was not called yet.
  423. * MPI: Fix of the livelock issue discovered while executing applications
  424. on a CPU+GPU cluster of machines by adding a maximum trylock
  425. threshold before a blocking lock.
  426. Small changes:
  427. * STARPU_NCPU should now be used instead of STARPU_NCPUS. STARPU_NCPUS is
  428. still available for compatibility reasons.
  429. * include/starpu.h includes all include/starpu_*.h files, applications
  430. therefore only need to have #include <starpu.h>
  431. * Active task wait is now included in blocked time.
  432. * Fix GCC plugin linking issues starting with GCC 4.7.
  433. * Fix forcing calibration of never-calibrated archs.
  434. * CUDA applications are no longer compiled with the "-arch sm_13"
  435. option. It is specifically added to applications which need it.
  436. * Explicitly name the non-sleeping-non-running time "Overhead", and use
  437. another color in vite traces.
  438. * Use C99 variadic macro support, not GNU.
  439. * Fix performance regression: dmda queues were inadvertently made
  440. LIFOs in r9611.
  441. * Use big fat abortions when one tries to make a task or callback
  442. sleep, instead of just returning EDEADLCK which few people will test
  443. * By default, StarPU FFT examples are not compiled and checked, the
  444. configure option --enable-starpufft-examples needs to be specified
  445. to change this behaviour.
  446. StarPU 1.0.3 (svn revision 7379)
  447. ==============================================
  448. Changes:
  449. * Several bug fixes in the build system
  450. * Bug fixes in source code for non-Linux systems
  451. * Fix generating FXT traces bigger than 64MiB.
  452. * Improve ENODEV error detections in StarPU FFT
  453. StarPU 1.0.2 (svn revision 7210)
  454. ==============================================
  455. Changes:
  456. * Add starpu_block_shadow_filter_func_vector and an example.
  457. * Add tag dependency in trace-generated DAG.
  458. * Fix CPU binding for optimized CPU-GPU transfers.
  459. * Fix parallel tasks CPU binding and combined worker generation.
  460. * Fix generating FXT traces bigger than 64MiB.
  461. StarPU 1.0.1 (svn revision 6659)
  462. ==============================================
  463. Changes:
  464. * hwloc support. Warn users when hwloc is not found on the system and
  465. produce error when not explicitely disabled.
  466. * Several bug fixes
  467. * GCC plug-in
  468. - Add `#pragma starpu release'
  469. - Fix bug when using `acquire' pragma with function parameters
  470. - Slightly improve test suite coverage
  471. - Relax the GCC version check
  472. * Update SOCL to use new API
  473. * Documentation improvement.
  474. StarPU 1.0.0 (svn revision 6306)
  475. ==============================================
  476. The extensions-again release
  477. New features:
  478. * Add SOCL, an OpenCL interface on top of StarPU.
  479. * Add a gcc plugin to extend the C interface with pragmas which allows to
  480. easily define codelets and issue tasks.
  481. * Add reduction mode to starpu_mpi_insert_task.
  482. * A new multi-format interface permits to use different binary formats
  483. on CPUs & GPUs, the conversion functions being provided by the
  484. application and called by StarPU as needed (and as less as
  485. possible).
  486. * Deprecate cost_model, and introduce cost_function, which is provided
  487. with the whole task structure, the target arch and implementation
  488. number.
  489. * Permit the application to provide its own size base for performance
  490. models.
  491. * Applications can provide several implementations of a codelet for the
  492. same architecture.
  493. * Add a StarPU-Top feedback and steering interface.
  494. * Permit to specify MPI tags for more efficient starpu_mpi_insert_task
  495. Changes:
  496. * Fix several memory leaks and race conditions
  497. * Make environment variables take precedence over the configuration
  498. passed to starpu_init()
  499. * Libtool interface versioning has been included in libraries names
  500. (libstarpu-1.0.so, libstarpumpi-1.0.so,
  501. libstarpufft-1.0.so, libsocl-1.0.so)
  502. * Install headers under $includedir/starpu/1.0.
  503. * Make where field for struct starpu_codelet optional. When unset, its
  504. value will be automatically set based on the availability of the
  505. different XXX_funcs fields of the codelet.
  506. * Define access modes for data handles into starpu_codelet and no longer
  507. in starpu_task. Hence mark (struct starpu_task).buffers as
  508. deprecated, and add (struct starpu_task).handles and (struct
  509. starpu_codelet).modes
  510. * Fields xxx_func of struct starpu_codelet are made deprecated. One
  511. should use fields xxx_funcs instead.
  512. * Some types were renamed for consistency. when using pkg-config libstarpu,
  513. starpu_deprecated_api.h is automatically included (after starpu.h) to
  514. keep compatibility with existing software. Other changes are mentioned
  515. below, compatibility is also preserved for them.
  516. To port code to use new names (this is not mandatory), the
  517. tools/dev/rename.sh script can be used, and pkg-config starpu-1.0 should
  518. be used.
  519. * The communication cost in the heft and dmda scheduling strategies now
  520. take into account the contention brought by the number of GPUs. This
  521. changes the meaning of the beta factor, whose default 1.0 value should
  522. now be good enough in most case.
  523. Small features:
  524. * Allow users to disable asynchronous data transfers between CPUs and
  525. GPUs.
  526. * Update OpenCL driver to enable CPU devices (the environment variable
  527. STARPU_OPENCL_ON_CPUS must be set to a positive value when
  528. executing an application)
  529. * struct starpu_data_interface_ops --- operations on a data
  530. interface --- define a new function pointer allocate_new_data
  531. which creates a new data interface of the given type based on
  532. an existing handle
  533. * Add a field named magic to struct starpu_task which is set when
  534. initialising the task. starpu_task_submit will fail if the
  535. field does not have the right value. This will hence avoid
  536. submitting tasks which have not been properly initialised.
  537. * Add a hook function pre_exec_hook in struct starpu_sched_policy.
  538. The function is meant to be called in drivers. Schedulers
  539. can use it to be notified when a task is about being computed.
  540. * Add codelet execution time statistics plot.
  541. * Add bus speed in starpu_machine_display.
  542. * Add a STARPU_DATA_ACQUIRE_CB which permits to inline the code to be
  543. done.
  544. * Add gdb functions.
  545. * Add complex support to LU example.
  546. * Permit to use the same data several times in write mode in the
  547. parameters of the same task.
  548. Small changes:
  549. * Increase default value for STARPU_MAXCPUS -- Maximum number of
  550. CPUs supported -- to 64.
  551. * Add man pages for some of the tools
  552. * Add C++ application example in examples/cpp/
  553. * Add an OpenMP fork-join example.
  554. * Documentation improvement.
  555. StarPU 0.9 (svn revision 3721)
  556. ==============================================
  557. The extensions release
  558. * Provide the STARPU_REDUX data access mode
  559. * Externalize the scheduler API.
  560. * Add theoretical bound computation
  561. * Add the void interface
  562. * Add power consumption optimization
  563. * Add parallel task support
  564. * Add starpu_mpi_insert_task
  565. * Add profiling information interface.
  566. * Add STARPU_LIMIT_GPU_MEM environment variable.
  567. * OpenCL fixes
  568. * MPI fixes
  569. * Improve optimization documentation
  570. * Upgrade to hwloc 1.1 interface
  571. * Add fortran example
  572. * Add mandelbrot OpenCL example
  573. * Add cg example
  574. * Add stencil MPI example
  575. * Initial support for CUDA4
  576. StarPU 0.4 (svn revision 2535)
  577. ==============================================
  578. The API strengthening release
  579. * Major API improvements
  580. - Provide the STARPU_SCRATCH data access mode
  581. - Rework data filter interface
  582. - Rework data interface structure
  583. - A script that automatically renames old functions to accomodate with the new
  584. API is available from https://scm.gforge.inria.fr/svn/starpu/scripts/renaming
  585. (login: anonsvn, password: anonsvn)
  586. * Implement dependencies between task directly (eg. without tags)
  587. * Implicit data-driven task dependencies simplifies the design of
  588. data-parallel algorithms
  589. * Add dynamic profiling capabilities
  590. - Provide per-task feedback
  591. - Provide per-worker feedback
  592. - Provide feedback about memory transfers
  593. * Provide a library to help accelerating MPI applications
  594. * Improve data transfers overhead prediction
  595. - Transparently benchmark buses to generate performance models
  596. - Bind accelerator-controlling threads with respect to NUMA locality
  597. * Improve StarPU's portability
  598. - Add OpenCL support
  599. - Add support for Windows
  600. StarPU 0.2.901 aka 0.3-rc1 (svn revision 1236)
  601. ==============================================
  602. The asynchronous heterogeneous multi-accelerator release
  603. * Many API changes and code cleanups
  604. - Implement starpu_worker_get_id
  605. - Implement starpu_worker_get_name
  606. - Implement starpu_worker_get_type
  607. - Implement starpu_worker_get_count
  608. - Implement starpu_display_codelet_stats
  609. - Implement starpu_data_prefetch_on_node
  610. - Expose the starpu_data_set_wt_mask function
  611. * Support nvidia (heterogeneous) multi-GPU
  612. * Add the data request mechanism
  613. - All data transfers use data requests now
  614. - Implement asynchronous data transfers
  615. - Implement prefetch mechanism
  616. - Chain data requests to support GPU->RAM->GPU transfers
  617. * Make it possible to bypass the scheduler and to assign a task to a specific
  618. worker
  619. * Support restartable tasks to reinstanciate dependencies task graphs
  620. * Improve performance prediction
  621. - Model data transfer overhead
  622. - One model is created for each accelerator
  623. * Support for CUDA's driver API is deprecated
  624. * The STARPU_WORKERS_CUDAID and STARPU_WORKERS_CPUID env. variables make it possible to
  625. specify where to bind the workers
  626. * Use the hwloc library to detect the actual number of cores
  627. StarPU 0.2.0 (svn revision 1013)
  628. ==============================================
  629. The Stabilizing-the-Basics release
  630. * Various API cleanups
  631. * Mac OS X is supported now
  632. * Add dynamic code loading facilities onto Cell's SPUs
  633. * Improve performance analysis/feedback tools
  634. * Application can interact with StarPU tasks
  635. - The application may access/modify data managed by the DSM
  636. - The application may wait for the termination of a (set of) task(s)
  637. * An initial documentation is added
  638. * More examples are supplied
  639. StarPU 0.1.0 (svn revision 794)
  640. ==============================================
  641. First release.
  642. Status:
  643. * Only supports Linux platforms yet
  644. * Supported architectures
  645. - multicore CPUs
  646. - NVIDIA GPUs (with CUDA 2.x)
  647. - experimental Cell/BE support
  648. Changes:
  649. * Scheduling facilities
  650. - run-time selection of the scheduling policy
  651. - basic auto-tuning facilities
  652. * Software-based DSM
  653. - transparent data coherency management
  654. - High-level expressive interface
  655. # Local Variables:
  656. # mode: text
  657. # coding: utf-8
  658. # ispell-local-dictionary: "american"
  659. # End: