Gráfico de commits

50 Commits

Autor SHA1 Mensaje Fecha
Siu Chi Chan 6613a37b3b Fix associate code object symbols with host allocation bug (#1799)
The current implementation skips this procedure for a given device
object when a global symbol is found in the cache.  This is incorrect:

 - There could be other undefined globals that have not been previously
encountered further down the list
 - If a symbol is found in the cache, it doesn't need to be pinned again
but it still need to be defined for the current executable

Added special case for the printf buffer symbol (already pinned by HCC)

The bug was exposed by running printf on different GPUs.
2020-01-24 16:22:49 +05:30
Siu Chi Chan fc08f29735 replace std::vector for kernarg 2019-05-24 17:27:43 -04:00
Siu Chi Chan 00824be34c move executable_cache into program_state.cpp 2019-05-24 17:27:25 -04:00
Siu Chi Chan 80fec2b477 remove executables() from program_state 2019-05-24 17:27:01 -04:00
Siu Chi Chan f5eb91d53d migrate program_state logic from header into shared library (phase I) (#1077)
* Revert "Revert "Use COMgr to read Kernel Args Metadata (#1006)""

This reverts commit a3d118eaa8.

* Revert "Use COMgr to read Kernel Args Metadata (#1006)"

This reverts commit 8a548bf40b.

* Revert "improve program state commentary"

This reverts commit 7aada87cbd.

* Revert "load program state once per agent"

This reverts commit c9117de8eb.

* start moving function_names() into the hip shared lib

* start moving code_object_blobs to a new "state" object

* Consolidate various program state related static objects into a
single program_state object

* minor clean up

* move more stuffs from functional_grid_launch into program_state

* debug make_kernarg

* moving lookup for kernargs size_align into program_state

* clean up old code for kernarg size and alignment

* update hip_module to use newer api in program_state

* Create public member functions for program_state

* move most program state functions into shared library

* Pass the data buffer size to load_executable
Otherwise, it can't figure what the data size is
just from the char* (since the data is not really a string)

* turning free functions in program state into members of program_state_impl

* change the free function globals() into a member of program_state_impl

* replace the static mutex used for populating globals

* moving associate_code_object_symbols_with_host_allocation into
program_state_impl

* move load_code_object_and_freeze_executable into program_state_impl

* moving executables and functions_names into program_state_impl

* moving kernels() into program_state_impl

* moving functions() into program_state_impl

* move get_kernargs into program_state_impl

* moving kernel_descriptor into program_state_impl

* moving kernargs_size_align calculation into program_state_impl

* Changing the handle to program_state_impl to a pointer

* moving program_state_impl into a separate inline source file

* fixing/cleaning up some header file includes

* moving member function for kernargs_size_align into program_state.cpp

* moving Kernel_descriptor into program_state.inl

* add a new class to manage agent globals

* moving all agent globals processing functions into agent_globals_impl

* load program state once per agent

re-merging PR991 against other program state changes

* fix per-agent program state member initialization

* cache executables based on elf name, isa, and agent.

This avoids program state reloading executables after a shared library is dlopened.

re-merging PR1057 against other program state changes

* protect executables cache by a global mutex

* return ref to executables cache

* adapt PR#981 Make hipModuleGetGlobal be in HIP runtime
2019-05-12 19:24:03 +05:30
Alex Voicu ea0fcf3e61 dlopen() fixes (#929)
* Initial attempt to switch over to internally linked state.

* Add missing CMake update.

* hipLaunchKernelGGLImpl must be inline as well. Ensure internal linkage.

* Ensure global retrieval uses internally linked state.

* Hide HC in the implementation. Minimise ADL woes.

* Strange software exists, and must be catered to.

* Use a less spammy mechanism for ensuring internal linkage / non-export.

* Remove leftover internal detail.
2019-03-06 17:31:44 +05:30
Alex Voicu 0b57f50855 Ensure that static structures are populated 2019-01-09 17:21:53 +00:00
Alex Voicu ec14daa7ce Hook into the creaky lazy-reinit machinery. Try to minimise race-risk. 2019-01-01 11:01:16 +05:30
Alex Voicu 340674ceb6 More blobs, more problems. 2019-01-01 11:01:07 +05:30
Alex Voicu 25c7e5d609 Start re-working 731 for 2.0. 2019-01-01 11:00:57 +05:30
Siu Chi Chan 1fbf639962 Revert "Rely on code object metadat for kernarg arguments alignof and sizeof."
This reverts commit fe1e963299.
2018-11-29 11:38:37 -05:00
Siu Chi Chan 71189c10c1 Revert "Handle the very confusing dual encoding of the symbol name."
This reverts commit bce3de8162.
2018-11-29 11:38:36 -05:00
Siu Chi Chan 09f87e41d9 Revert "If we've already seen a __global__ function we do not need to re-parse"
This reverts commit f7ba987038.
2018-11-29 11:38:35 -05:00
Siu Chi Chan 8eb9b38e76 Revert "Missing handling nullary __global__ functions for mixed arity cases."
This reverts commit 4ebc229b9a.
2018-11-29 11:38:31 -05:00
Siu Chi Chan 909609773c Revert "Fix issue in kernarg metadata parsing due to early finalization"
This reverts commit 19acf86cef.
2018-11-29 11:35:00 -05:00
Siu Chi Chan 19acf86cef Fix issue in kernarg metadata parsing due to early finalization
The logic to parse the kernel metadata is unaware that enabling
of early finalization could result in multiple code blobs in a
single .kernel section.  This teaches the HIP runtime to handle
that.

Change-Id: I1581b42f0da8b30233d7898014f7468728c1d489
2018-11-21 12:09:46 -05:00
Alex Voicu 4ebc229b9a Missing handling nullary __global__ functions for mixed arity cases. 2018-11-12 01:51:59 +00:00
Alex Voicu f28cde3f98 Merge branch 'master' of https://github.com/ROCm-Developer-Tools/HIP into feature_get_alignment_and_size_from_metadata 2018-11-01 23:17:27 +00:00
Alex Voicu f7ba987038 If we've already seen a __global__ function we do not need to re-parse 2018-10-30 23:34:27 +00:00
Alex Voicu bce3de8162 Handle the very confusing dual encoding of the symbol name. 2018-10-30 01:55:09 +00:00
Alex Voicu fe1e963299 Rely on code object metadat for kernarg arguments alignof and sizeof. 2018-10-28 17:01:00 +00:00
Siu Chi Chan 30ce25e627 Support more than one bundles in a single .kernel section
When compiling with Early Finalization enabled in HCC,
the resulting .kernel section of the host object now may
contain more than one device code bundles.  This is to
teach the HIP runtime to correctly extract all the
bundles from the .kernel section.
2018-10-18 17:00:27 -04:00
Wen-Heng (Jack) Chung dd56942335 Keep the map which tracks GPU kernel symbols to grow monotonically 2018-09-26 19:49:02 +00:00
Wen-Heng (Jack) Chung 35c542b70f Improve performance of re-initialization logic
Keep track of shared libaries already discovered. Do not build HSA executables
for them.
2018-09-26 19:48:56 +00:00
Wen-Heng (Jack) Chung dab1a0f9db HIP program state re-initialization logic
This commit is to support kernels dynamically loaded thru means such as
dlopen() after HIP runtime initializes.
2018-09-26 19:48:47 +00:00
Maneesh Gupta 4cf851f416 Merge pull request #621 from ROCm-Developer-Tools/disable_startup_loader
Disable startup loader by default and guard with env var
2018-08-10 10:18:25 +05:30
sunway513 17f38937e0 resolve a segfault bug when env var not set; remove startup_kernel_loader class 2018-08-09 16:40:26 +00:00
sunway513 30dfa6f129 Add more check to ensure the startup loader only be enabled with the env var set to 1 2018-08-04 01:52:27 +00:00
sunway513 3a68ab4919 Add startup loader under HIP_STARTUP_LOADER env var, disable by default 2018-08-04 01:48:06 +00:00
Wen-Heng (Jack) Chung 2604f33930 Revert "HIP program state re-initialization logic"
This reverts commit 379b7a2241.
2018-08-03 17:03:04 -05:00
Wen-Heng (Jack) Chung 3426f15171 Revert "Improve performance of re-initialization logic"
This reverts commit ece4539c1d.
2018-08-03 17:02:58 -05:00
Wen-Heng (Jack) Chung 136bcc2981 Revert "Keep the map which tracks GPU kernel symbols to grow monotonically"
This reverts commit 32789a8b7d.
2018-08-03 17:02:50 -05:00
Wen-Heng (Jack) Chung 32789a8b7d Keep the map which tracks GPU kernel symbols to grow monotonically 2018-06-18 16:54:18 -05:00
Wen-Heng (Jack) Chung ece4539c1d Improve performance of re-initialization logic
Keep track of shared libaries already discovered. Do not build HSA executables
for them.
2018-06-15 18:07:33 -05:00
Wen-Heng (Jack) Chung 379b7a2241 HIP program state re-initialization logic
This commit is to support kernels dynamically loaded thru means such as
dlopen() after HIP runtime initializes.
2018-06-14 15:46:49 +00:00
Alex Voicu 13274ce559 Add support for the hipFuncGetAttributes interface. 2018-05-11 03:35:10 +01:00
Maneesh Gupta 1ba06f63c4 Apply .clangformat to all repo source files
Change-Id: I7e79c6058f0303f9a98911e3b7dd2e8596079344
2018-03-12 11:29:03 +05:30
Alex Voicu 49ef24c042 Merge branch 'master' of https://github.com/ROCm-Developer-Tools/HIP into feature_purge_genco 2018-02-22 12:08:52 +00:00
Siu Chi Chan ebc92e5975 make HIP to load the GPU objects and to setup the function symbol map on startup 2018-02-18 14:26:44 -05:00
Alex Voicu baf50a5311 Re-sync with upstream. 2018-02-12 20:20:24 +00:00
Rahul Garg 24ab820a11 Fixed host allocated globals address lookup for host usage
Fixed texture driver APIs failure
2018-01-30 18:06:31 +05:30
Alex Voicu b842394957 This introduces LipoProteinLipase (lpl), a simple tool for creating fat binaries. It represents a direct replacement of the creaky hccgenco.sh script, which had various issues. The format it uses is that of a code object bundle, generated by the Clang Offload Bundler. The output is always suffixed with the ".adipose" extension. It is shared with HCC. The hipcc script and associated tests are modified to use lpl. Help can be obtained by invoking lpl --help. A more computer-sciency / corporate friendly name is likely to be beneficial, which is a reason for choosing easily searchable/replaceable names such as lpl or adipose. 2017-12-08 04:22:57 +00:00
Alex Voicu 5127ce67e8 This is primarily intended as an additional cleanup of the module functionality, in the aftermath of adopting module based dispatch. The main effort was associated with refactoring the questionable ihipModuleGetSymbol. It was quaintly written and misleading, in that it had little to do with getting symbols, and was exactly retrieving a kernel object. Error handling is modified so as to reduce branching depth. Functions which serve as interfaces to the HSA RT are moved in a separate helper header. Code object readers are properly deleted. Some leftover dead functionality pertaining to associating namespace scope variables with their allocated memory is removed. Executable loading is changed to use a string which holds the ELF image of the code object being loaded, thus avoiding some corner cases where using a istream would fail. 2017-12-03 23:09:06 +00:00
Alex Voicu 7c0b9a005b Fix legacy mode detection of the address of an agent allocated variable. In this mode, there exist two executables per each code object, one created by HCC and one created by HIP. Since we dispatch through HCC in legacy mode, we should obtain the address for an agent allocated variable from the latter's executable. Also add two omitted validity checks, whose absence could lead to segfaults when the current process had no .kernel section and / or when an invalid or empty blob was extracted from the latter. 2017-11-30 03:29:04 +00:00
Alex Voicu 02c2bfc7ef Re-sync with upstream and re-factor platform global management for texture references. 2017-11-28 19:15:29 +00:00
Alex Voicu 5e16ee0d1f This corrects how addresses are formed for symbols which reside in shared objects. For this case, the .value component of an ELF symbol holds the offset from the base VA where the shared object was loaded. Thus, to correctly obtain the VA of the object refered by the symbol, we must add the offset to the VA where the shared object is loaded. We were already doing this correctly for symbols denoting functions, but we were incorrect for those denoting objects. 2017-11-21 13:15:13 +00:00
Alex Voicu 9d088d2283 Refactor the __device__ versions of memset and memcpy to be less awkward i.e. not return nullptr as opposed to the destination pointer (it can only be assumed it was done for maximum confusion) and actually unroll as they claim to. Change all of the {to, from}Symbol functions to use hipModuleGetGlobal, as opposed to hc::accelerator::get_symbol_address which is no longer valid with module based dispatch. 2017-11-21 02:40:34 +00:00
Alex Voicu 1824fb7698 Clean-up some remaining noise in program_state.cpp. 2017-11-20 22:41:46 +00:00
Alex Voicu 2cacda91bb Correctly deal with functions from shared objects, wherein the program visible VA == so_base_va + st_value(function_symbol). Remove quaint usage of pfe for hipMemset (which is actually fill_n). 2017-11-01 22:33:13 +00:00
Alex Voicu c2482d1255 This switches HIP from its currently convoluted macro + pfe based dispatch mechanism to a more natural one partially based on the existing module API. The basic idea is that HCC will always correctly emit __global__ functions: as empty-bodied stubs, on host, and as kernels, on device. It then becomes trivial to obtain the mangled name on host, at dispatch, from the function's address, and then to use the mangled name to retrieve the kernel. This should address all problems stemming from serialisation, dubious mismatches due to the manufactured functor, macro-isms et al. It also immediately enables support for generalised globals as a consequence of that being available in the module API. Finally, it will make debug much easier, since the actual names of the __global__ functions will automatically be used in traces etc. One detail is that due to how dispatch works now (hipLaunchKernel and hipLaunchKernelGGL are themselves variadic function templates which deduce the function type of the callee), in certain cases it may be necesssary to insert explicit casts to ensure that the variadic argument list selects a viable overload - this can be observed in some unit tests. Eventually we may be able to remove this limitation, but for now it does not appear terribly onerous. The code is not extremely HIPpie, nor is it fully optimised, but rather is intended as a starting point for the HIP team to make its own. 2017-11-01 15:09:59 +00:00