Wook Song [Wed, 20 Nov 2019 08:59:25 +0000 (17:59 +0900)]
[Tools/GenTestData] Fix a warning about sign-compare
This patch fixes the sign-compare warning in testdata_gen.cpp.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Wed, 20 Nov 2019 08:08:28 +0000 (17:08 +0900)]
[Tools/GenTestData] Fix a warning about unused-result
This patch fixes the unused-result warning in testdata_gen.cpp.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Wed, 20 Nov 2019 07:52:33 +0000 (16:52 +0900)]
[Tools/GenTestData] Remove unneeded white spaces
This patch removes unneeded white spaces in testdata_gen.cpp.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Wed, 20 Nov 2019 06:55:23 +0000 (15:55 +0900)]
[Test] Move files related to testdata generation to the tools directory
This patch moves files related to testdata generation, which are
orinally placed in ./src/test/testdata, to the tools directory, i.e.,
./tools/gen-testdata.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Thu, 21 Nov 2019 06:48:16 +0000 (15:48 +0900)]
[Dist/Debian] Add omitted build-dependency on libiniparser-dev
This patch adds the omitted build-dependency on libiniparser-dev to the
debian control file.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Dongju Chae [Wed, 20 Nov 2019 04:40:48 +0000 (13:40 +0900)]
[EMUL/BUG] Bug fix: when model_gen.py makes esum data (ESUM vs. CONVE)
This commit fixes the bug when model generator makes esum data, which should have handled differently depending on OPCODE (i.e., ESUM vs. CONVE).
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 19 Nov 2019 07:52:06 +0000 (16:52 +0900)]
[EMUL/BUG] Bug fix: Add a missing line when setting esum data
This commit adds a missing line when setting esum data.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 19 Nov 2019 07:25:45 +0000 (16:25 +0900)]
[CONF] Set the default resv mem size to 0 (Kernel CMA)
This commit change the default resv mem size to 0 (Kernel CMA).
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 19 Nov 2019 07:17:32 +0000 (16:17 +0900)]
[FIX/GEM] fix the size checking in gem_create() to prevent overflow
This commit fixes the size checking in gem_create(), which caused size
overflow due to alignment.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Wook Song [Mon, 18 Nov 2019 07:14:38 +0000 (16:14 +0900)]
[Core/NPU/Emul] Fill the function body for CONVE emulation
This patch fills the function body for emulation of the CONVE
instruction and enables the emulation using it.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Mon, 18 Nov 2019 05:01:57 +0000 (14:01 +0900)]
[Core/NPU/Emul] Fill the function body for ReLU emulation
This function fills the function body for emulation of the ReLU
instruction. Using this function, the emulation for ReLU is enabled.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Mon, 18 Nov 2019 04:42:20 +0000 (13:42 +0900)]
[Core/NPU/Emul] Fill the function body for AVGP emulation
This function fills the function body for emulation of the Average
Pooling instruction. The emulation for AVGP is enabled via this
function.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Dongju Chae [Mon, 18 Nov 2019 02:26:17 +0000 (11:26 +0900)]
[README/DOC] Revise README.md and add some documentation
This commit revises README.md and add some documentation.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Mon, 18 Nov 2019 13:01:55 +0000 (22:01 +0900)]
[BUG] Bug fix in CONV emulation and esum/weight data setup
Thix commit fixes the bug in CONV emulation and esum/weight data setup.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 19 Nov 2019 03:13:25 +0000 (12:13 +0900)]
[TESTDATA] Add a testcase for multi-layer model to test CONVE/ReLU/AVGP
This commit adds new testcase, testcase3, to test CONVE/ReLU/AVGP.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 19 Nov 2019 03:08:57 +0000 (12:08 +0900)]
[TESTDATA/FIX] Fix the model generation to use specified addresses
This commit fixes the model generation to use specified addresses
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Wook Song [Fri, 15 Nov 2019 05:17:19 +0000 (14:17 +0900)]
[Core/NPU/Emul] Add skeleton code for ReLU, AVGP, CONVE, TCNV, TCNVE
This patch adds skeleton code for running operators, ReLU, AVGP, CONVE,
TCNV, and TCNVE, in the test mode.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Fri, 15 Nov 2019 02:31:26 +0000 (11:31 +0900)]
[Core/NPU/Emul] Add a struct representing instruction format for TCNVE
This patch adds a data struct representing the instruction format for
TCNVE.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Fri, 15 Nov 2019 02:27:58 +0000 (11:27 +0900)]
[Core/NPU/Emul] Add a struct representing instruction format for CONVE
This patch adds a data struct representing the instruction format for
CONVE.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Fri, 15 Nov 2019 02:15:07 +0000 (11:15 +0900)]
[Core/NPU/Emul] Add a struct representing instruction format for TCNV
This patch adds a data struct representing the instruction format for
TCNV.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Fri, 15 Nov 2019 02:06:49 +0000 (11:06 +0900)]
[Core/NPU/Emul] Add a struct representing instruction format for AVGP
This patch adds a data struct representing the instruction format for
AVGP.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Fri, 15 Nov 2019 01:51:16 +0000 (10:51 +0900)]
[Core/NPU/Emul] Add a struct representing instruction format for RELU
This patch adds a data struct representing the instruction format for
RELU.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Dongju Chae [Thu, 14 Nov 2019 10:03:56 +0000 (19:03 +0900)]
[NPU/EMUL] Add NPU emulation engine (support CONV/MAXP/ESUM)
This commit adds NPU emulation engine.
Currently, it supports CONV/MAXP/ESUM.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 15 Nov 2019 05:19:32 +0000 (14:19 +0900)]
[TESTDATA] Add enum for trinity opcodes
This commit adds enum for trinity opcodes.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Thu, 14 Nov 2019 10:48:28 +0000 (19:48 +0900)]
[TESTDATA] Fix multi-layer testdata generation when setting their shamt
This commit fixes the bug in multi-layer testdata generation.
The shamt is differently set depending on an operation type.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Wook Song [Thu, 14 Nov 2019 08:19:57 +0000 (17:19 +0900)]
[Local/Dist/Debian] Run meson with PKG_CONFIG_PATH pointing libnpuvision
NPU Engine requires dependencies on libnpuvision, which is placed under
/opt/trinity. In order to provide those dependencies, this patch adds
PKG_CONFIG_PATH pointing when running meson.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Thu, 14 Nov 2019 07:11:00 +0000 (16:11 +0900)]
[Local/Dist/Debian] Initial debianize
This patch debianizes this NPU Engine for Ubuntu 16.04 running on
x86_64.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Thu, 14 Nov 2019 07:05:51 +0000 (16:05 +0900)]
[Local/Meson] Add a mode to build/run NPU Engine without FM simulator
This patch adds a mode, build_test_only_mode, to build NPU Engine on
a Ubuntu x86_64 machine and to run it without the FastModel simulator.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Dongju Chae [Tue, 12 Nov 2019 02:29:39 +0000 (11:29 +0900)]
[NPU/EMUL] Add a skeleton for NPU ioctl emulations
This commit adds a skeleton for NPU ioctl emulations. It replaces the existing NPU ioctls to dummy ones, and will eventually perform emulations (TODO).
Note that this source file is written in C++. The reason is the existing emulation codes for VISA are written in C++, and we need their C wrappers.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Wook Song [Mon, 4 Nov 2019 07:52:42 +0000 (16:52 +0900)]
[Core/Conf] Make load_conf_default static
In order to avoid confusion between load_conf and load_conf_default,
this patch makes the load_conf_default function static.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Dongju Chae [Mon, 11 Nov 2019 03:03:49 +0000 (12:03 +0900)]
[NPU] add a missing header, typedef.h, to npubinfmt.h
This commit adds a missing header, tpyedef.h, to npubinfmt.h
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 8 Nov 2019 11:51:57 +0000 (20:51 +0900)]
[FIX] Make model generator generate final input/output fmap
This commit fixes the model generator to generate final input/output fmap
Its dependent testdata is also updated.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 8 Nov 2019 07:41:29 +0000 (16:41 +0900)]
[MEM] Add .reset_buffer() to prevent deadlocks on various error cases
This commit adds new API, .reset_buffer(), to prevent deadlocks on various error cases.
When some errors occurs, ne-host-inputservice.c and ne-inf.c used to emulate normal execution by explicitly calling return_buffer().
However, as output callback can be not called, deadlock situations often happen when the exact transition of buffer states is not performed.
So, this patch makes the buffer state to initial one on all error cases to reduce the complexity and prevent deadlocks.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Thu, 7 Nov 2019 08:34:33 +0000 (17:34 +0900)]
[TESTDATA] Add testdata generator with multi-layer models
This commit adds testdata generator with multi-layer models.
The meson build script will be updated soon.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 5 Nov 2019 02:25:01 +0000 (11:25 +0900)]
[UNITTEST] Fix dummy model metadata and callback check
This commit fixes the metadata of dummy models and callback check routine.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 8 Nov 2019 07:11:40 +0000 (16:11 +0900)]
[TESTDATA] Update the NPU binary generator for multi-layer models
This commit updates the NPU binary generator for multi-layer models
It also updates testdata (i.e., *.tar.gz)
Signed-off-by: Dongju Chae dongju.chae@samsung.com
Wook Song [Tue, 5 Nov 2019 04:58:02 +0000 (13:58 +0900)]
[Test/AppTests] Reduce number of tests when running on the simulator
Currently, there are two modes in NPU-Engine, one is test mode which
means it runs without interaction with the fastmodel simulator, and the
other mode which means it runs with actual interaction with the
simulator. To shorten the test time, this patch reduces the number of
tests in async_callbacks when the test runs in the not-test mode.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Tue, 5 Nov 2019 04:29:00 +0000 (13:29 +0900)]
[Core/Conf] Add an API to get the current running mode
Although the config data has a global scope, it would be removed after
all and replaced with APIs to get information from it. This patch adds
an API to get the current running mode from the global config data.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Dongju Chae [Mon, 4 Nov 2019 05:24:32 +0000 (14:24 +0900)]
[NPU/CONF] Bug fix due to wrong config loads in NPU driver API
This commit fixes the bug due to wrong config loads in NPU driver API.
We should not use `load_conf_default()` because it may overlap user-provided configs.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Mon, 4 Nov 2019 07:13:41 +0000 (16:13 +0900)]
[MEM/CMA] Bug fix in cma wrappers for dmabuf map/unmap
This commit fixes the bug in cma wrappers for dmabuf map/unmap.
Because buffers are often resized, we need to carefully maintain the dmabuf/handle mapping.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Thu, 31 Oct 2019 08:17:10 +0000 (17:17 +0900)]
[BUG] Fix the wrong encoding for input/output offset when generating a NPU model
This commit fixes the wrong encoding for input/output offset when generating a NPU model.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Thu, 31 Oct 2019 08:14:14 +0000 (17:14 +0900)]
[FIX] fix the bug in example_visa apptest when it handles fgetc
This commit fixes the bug in example_visa apptest.
As fgetc() returns integer type not charactor type, it should be fixed.
Otherwise, the EOF check can be failed.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 30 Oct 2019 12:47:11 +0000 (21:47 +0900)]
[BUG] fix srnpu unittests to use valid dmabuf id
This commit fixes srnpu unittests to use valid dmabuf ids with cgem/srtlb.
In NE_TEST_MODE=0 (i.e., real mode in FastModel), we need to carefuly set ioctl parameters.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 30 Oct 2019 12:26:56 +0000 (21:26 +0900)]
[BUG] Fix the wrong ioctl arguments in srnpu APIs
This commit fixes the wrong ioctl arguments in srnpu APIs
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
채동주/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 30 Oct 2019 10:25:17 +0000 (19:25 +0900)]
Revert "[BUG/NPU] bug fix in npu_open() for fd return"
This reverts commit
d273a2553b48f73d29070aadd7725752ae74089e.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 30 Oct 2019 09:44:52 +0000 (18:44 +0900)]
[BUG] Fix wrong ioctl argument for dmabuf mapping
This commit fixes wrong ioctl argument for dmabuf mapping
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 30 Oct 2019 03:07:09 +0000 (12:07 +0900)]
[BUG] Fix bugs in the unittest for srtlb APIs
This commit fixes bugs in the unittest for srtlb APIs.
When dmabuf fds are mapped to srtlb driver, they should be valid.
So, this patch uses the exsiting APIs providing dmabuf-sharing, cgem APIs in NE.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Mon, 28 Oct 2019 11:15:38 +0000 (20:15 +0900)]
[BUG/NPU] bug fix in npu_open() for fd return
This commit fixes the bug in npu_open(), which returns id not fd.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
MyungJoo Ham [Tue, 3 Sep 2019 07:39:48 +0000 (16:39 +0900)]
[ASR] Add skeleton for ASR support
ASR (Automatic Speech Recognition), which uses SRP for Audio, support is required.
Signed-off-by: MyungJoo Ham <myungjoo.ham@samsung.com>
Parichay Kapoor [Mon, 14 Oct 2019 07:13:29 +0000 (16:13 +0900)]
[TLBdrvAPI] sr TLB API driver implementation
Added sr TLB driver implementation
Added unittests for TLB api driver
Bug fix for hash table hashing generating constant
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Fri, 25 Oct 2019 07:45:47 +0000 (16:45 +0900)]
[n4/n3] return buffer in case of error
Return the buffer in case of error from NPU state to output state
This avoids hung state when error occurs or stop called in between processing the buffer
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Dongju Chae [Thu, 24 Oct 2019 10:54:02 +0000 (19:54 +0900)]
[APPTEST] Add the apptest to test example visa binaries
This commit adds the apptest to test example visa binaries.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 25 Oct 2019 05:23:23 +0000 (14:23 +0900)]
[TEST] Add a script to generete a NPU model with visa data
This commit adds a script to generate NPU model with visa data files.
Currently, we don't have compiled binaries, which means we have to generate test models with this.
Also, it adds testdata which contains both visa data and pre-generated binaries.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Mon, 14 Oct 2019 05:22:01 +0000 (14:22 +0900)]
[DMABUF] Add new APIs for sync features on dmabuf
This commit adds new APIs for sync features on dmabuf.
Also, its clients (such as ne-mem) should make sure that mmap() for dmabuf is being used for sync.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Parichay Kapoor [Fri, 27 Sep 2019 10:57:56 +0000 (19:57 +0900)]
[mem/npu] Add register of dmabuf when alloc memory
Added registration of dmabuf fd when memory is allocated in ne-mem
Correspondingly de-registration is also added
Registration is performed for all hwmem_type allocated by ne-mem
Added TLB user side driver along with interface to register/deregister dmabuf
Using dmabuf id in host input service when calling NPU driver
V2:
Renamed functions in GEMdrvAPI
gem_prime_export() -> gem_prime_handle_to_fd()
gem_prime_import() -> gem_prime_fd_to_handle()
add close(dmabuf_fd) for dmabuf_fd received via gem_prime_handle_to_fd()
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Tue, 1 Oct 2019 07:45:55 +0000 (16:45 +0900)]
[TLBdrvAPI] Added TLB user-side API skeleton
Added user side API skeleton for translation lookaside buffer
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Dongju Chae [Tue, 1 Oct 2019 08:00:06 +0000 (17:00 +0900)]
[MEM] Resolve the race condition bug in buffer resizing
This commit resolves the race condition bug in buffer resizing.
It introduces reference counters to track of buffer usage.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 1 Oct 2019 03:59:34 +0000 (12:59 +0900)]
[MEM] Revise enum hwmem_type for readability
This commit revised `enum hwmem_type` for readability.
Originally, HWMEM_TYPE_BUFFER/MODEL/USER are used but they don't represent their properties.
Instead of them, I add three types
- HWMEM_TYPE_INTERNAL_POOL: hwmem mananged in memory pool
- HWMEM_TYPE_INTERNAL_CMA: hwmem managed in kernel cma
- HWMEM_TYPE_EXTERNAL: external hwmem that cannot be mananged in NE
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Parichay Kapoor [Fri, 20 Sep 2019 05:02:13 +0000 (14:02 +0900)]
[NPUdrvAPI] Updated interface
Updated the interface of NPU user side driver based on updated memory interface
and remove the support of icam/other modes
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Dongju Chae [Wed, 25 Sep 2019 06:12:23 +0000 (15:12 +0900)]
[MEM] Rename the API 'get_size' to 'get_pool_size'
This commit renames the API 'get_size', which is an ambiguous name, to 'get_pool_size'.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Mon, 23 Sep 2019 06:13:19 +0000 (15:13 +0900)]
[GEM] Add an extra argument for cacheable GEM buffer
This commit adds an extra argument for cacheable GEM buffer.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Mon, 23 Sep 2019 05:55:34 +0000 (14:55 +0900)]
[AppTest] add AppTest to test dma-buf sharing for I/O buffers
This commit adds new AppTest to test dma-buf sharing for I/O buffers.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 20 Sep 2019 10:12:28 +0000 (19:12 +0900)]
[LIBHOST] Add a user-level API to allocate a generic_buffer
This commit adds a user-level API to allocate a gneneric_buffer for each buffer type
In case of BUFFER_DMABUF, KERNEL_CMA mode should be enabled (i.e., resv_mem_size == 0).
Also, some test codes are modified due to APIs are changed.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Wook Song [Thu, 19 Sep 2019 07:40:48 +0000 (16:40 +0900)]
[ReadMe] Fix unshown badges in README.md
After github.sec is upgraded, the badges in README.md has been not shown
anymore. In order to fix this issue (#226), the links pointing the svg
files are changed to another repository in github.sec.
coverage-generator.sh in the CI system used by this project is also
modified to fix this issue.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Parichay Kapoor [Fri, 20 Sep 2019 04:46:40 +0000 (13:46 +0900)]
[packaging] Dsiable track-origins for unittests
Disabling track origins for unittest
ne_core_ip_test_default_conf.registerNPUmodel_large_size ends up consuming more than **30GB**
creating problems when running with track-origins
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Parichay Kapoor [Wed, 18 Sep 2019 06:50:11 +0000 (15:50 +0900)]
[host-input-service] bugfix for base address
Corrected using hwmem_get_offset instead of hwmem_get_data
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Dongju Chae [Tue, 17 Sep 2019 06:20:07 +0000 (15:20 +0900)]
[LIBHOST] Fix the linking problem for APIs in CPP applications
This patch fixes the linking problem in CPP applications by appending cpp check macros.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 17 Sep 2019 12:37:50 +0000 (21:37 +0900)]
[BUG FIX] Fix the assertion error of 'DummyInference' on aarch64 using GBS build
This commit fixes the bug related to assertion error of 'DummyInference' on aarch64 using GBS build.
The patch includes
1) Change the destructor order between input service and ip plugin.
-- The destructor of input service should be called before that of IP plugin because input service's callback threads still acccess memory allocator's functions.
2) Ensure the callback threads are finished before freeing resources in the destructor of host input service.
-- Otherwise, it can cause the deadlock issue when the destructor frees the resources early.
3) Add check routine in `return_buffer()` whether buffers are valid or not.
-- Actually, it's an ad-hoc patch because the fundamental solution needs the sync feature in `resize_buffers()` to ensure that resizing buffers is safe. This will be updated in later commits.
It resolves #228.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Wook Song [Mon, 16 Sep 2019 07:21:14 +0000 (16:21 +0900)]
[Test/Utils] Clean code and handle BUFFER_UNDEFINED
This patch refactors the test utility functions.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Wook Song [Mon, 16 Sep 2019 06:25:30 +0000 (15:25 +0900)]
[Tests/IP] Revise the test case named 'registerNPUmodel_large_size'
This patch revises the test case, 'registerNPUmodel_large_size' in the
'ne_core_ip_test_default_conf' test group.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Dongju Chae [Mon, 16 Sep 2019 12:00:40 +0000 (21:00 +0900)]
[README] Revise README.md in root dir to add useful info
This commit revises README.md in root dir to add useful info.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Mon, 16 Sep 2019 09:42:09 +0000 (18:42 +0900)]
[MEM/IP] Support dma-buf sharing for I/O buffers
This commit enables dma-buf shring for I/O buffers.
New APIs, buffer_get_dmabuf and bufferget_dmabuf, are added.
It also changes some codes in IP plugin to use these APIs.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Mon, 16 Sep 2019 06:19:37 +0000 (15:19 +0900)]
[README] Add README.md for user APIs
This commit adds `README.md` which describes how users can build their application with NPU Engine library (i.e., `libhosthost.h`) and the API's detail usage.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 11 Sep 2019 08:02:44 +0000 (17:02 +0900)]
[Handler] Minor fix on ne-handler.c
This commit applies some minor fix such as removing uncessary lines and adding error checking.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Wook Song [Mon, 16 Sep 2019 08:00:40 +0000 (17:00 +0900)]
[Mem] Refactor mem_alloc()
This patch refactors the mem_alloc function to improve readability.
Signed-off-by: Wook Song <wook16.song@samsung.com>
Dongju Chae [Mon, 16 Sep 2019 04:35:56 +0000 (13:35 +0900)]
[AppTest] add AppTest to test dma-buf sharing for model data
This PR adds new AppTest to test dma-buf sharing for model data.
Also, there are minor changes in other unittests due to the updated common API.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 11 Sep 2019 08:18:48 +0000 (17:18 +0900)]
[NPU] Fix ioctl() warnings in NPU driver API
This commit fixes ioctl() warnings in NPU driver API.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 10 Sep 2019 05:45:32 +0000 (14:45 +0900)]
[MEM] Revise .get_dmabuf() to support KERNEL_CMA
This commit revise .get_dmabuf() to support KERNEL_CMA mode.
Currently, get_dmabuf() returns the global dmabuf fd for internal memory pool.
As KERNEL_CMA mode handles multiple CMA buffers, it should be revised to support this.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 11 Sep 2019 07:28:53 +0000 (16:28 +0900)]
[MEM] Fix error return statements
This commit fixes error return statements in some functions to keep the same format.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 11 Sep 2019 04:52:14 +0000 (13:52 +0900)]
[CONF] Replace sscanf() to strtoull() for parsing 'resv_mem_size'
This commit replaces sscanf() to strtoull() for parsing `resv_mem_size` to avoid unexpected errors.
It resolves #215.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 10 Sep 2019 05:36:21 +0000 (14:36 +0900)]
[IP] Change IP-plugin to use dmabuf_register() for BUFFER_DMABUF
This commit changes IP-plugin to use `dmabuf_register()` for BUFFER_DMABUF type.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 6 Sep 2019 10:39:25 +0000 (19:39 +0900)]
[CONF] Enable KERNEL_CMA as default in FastModel
This commit enables KERNEL_CMA as default in FastModel.
A zero-sized reserved memory indicates that the memory allocator used Kernel's CMA allocation.
In order to accept the zero size, ne-conf.c should be changed.
Also, some minor bugs are fixed
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 6 Sep 2019 08:01:04 +0000 (17:01 +0900)]
[MEM] Add new API, register_dmabuf, to register pre-allocated dmabuf as a hwmem
This commit adds new API in memory allocator to register a pre-allocated dmabuf as a hwmem for user-provided mem.
A user has its ownership to allocate and free, and memory allocator may not manage such memory explicitly.
As long as it can be exported as dmabuf fd, any memory source will be fine because it's physically contiguous.
Users can allocate such memory from NPU engine or other third-party libaries.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 6 Sep 2019 03:41:08 +0000 (12:41 +0900)]
[BUG] Fix the bug caused by a non-terminated thread in destructor
This commit fixes the bug in `ne-host-input-service.c`.
Sometimes, `cb_thread` is not terminiated until the destructor is called.
In this case, `pthread_cond_destroy()` generates unexpected bugs because cb_thread is still using this cond (e.g., `pthread_cond_broadcast()` at exit).
So, the destructor needs to make sure that `cb_thread` is terminated or not.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Fri, 6 Sep 2019 03:54:07 +0000 (12:54 +0900)]
[Handler] Revise APIs in ne-handler
This commit revises APIs in ne-handler such as
1) remove getAvailableModelMemory()
--> it does not provide useful information because we cannot guarantee the available memory size due to compaction. Also, in Kernel-CMA mode, it's impossible to provide this API.
2) add getSizeModels/getSizeBuffers()
--> it provides memory usage information of model and buffers to users.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
MyungJoo Ham [Thu, 5 Sep 2019 10:28:02 +0000 (19:28 +0900)]
[ASR] Add device-type for ASR
1. Rename "npu_type" to "dev_type" support ASR IP.
2. Add ASR types.
This is the prerequisite of #199
Signed-off-by: MyungJoo Ham <myungjoo.ham@samsung.com>
Suggested-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 4 Sep 2019 10:13:23 +0000 (19:13 +0900)]
[MEM] Revise N71 to support KERNEL_CMA
This commit revises the impl. of memory allocator to support kernel CMA-based memory allocation.
It adds COMPACT_MODE_KERNEL_CMA selected when the passed memory pool size is zero (in init()).
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Thu, 5 Sep 2019 05:57:34 +0000 (14:57 +0900)]
[README] add the unittest badge into README.md
This badge shows the number of unittests (and apptests) which are successfully executed.
To generate the badge, it requires gtest output files (.xml).
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Wed, 4 Sep 2019 10:31:23 +0000 (19:31 +0900)]
[README] Change the url of coverage badge (http --> https)
This commit changes the url of coverage badge to use https instead of http.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 3 Sep 2019 12:11:36 +0000 (21:11 +0900)]
[Bug/Test] Bug fix in ne-handler unittest when obtaining data pointers
This commit fixes bugs in ne-handler unittest, which obtained data pointers without activating their hwmems.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 3 Sep 2019 03:51:26 +0000 (12:51 +0900)]
[README] Add the coverage badge in README.md
This commit adds the coverage badge in README.md
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 3 Sep 2019 12:17:08 +0000 (21:17 +0900)]
[AppTest] Make AppTest, async_callbacks, use ne_test_utils.h
This commit changes AppTest, async_callbacks, to use ne_test_utils.h.
Also, alloc_generic_buffer now has an extra argument for a buffer type.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 27 Aug 2019 10:48:46 +0000 (19:48 +0900)]
[GEM] Revise GEM APIs to support multiple malloc-ed addrs in test mode
This commit revises GEM APIs such as
- support multiple virtual malloc-ed addresses used in test mode
- remove uncessary argument in mmap()
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 3 Sep 2019 11:57:14 +0000 (20:57 +0900)]
[IP-plugin] Reuse the codes in runNPU_async() for runNPU_sync()
This commit changes the code flow of runNPU_sync() to be the same with runNPU_async().
runNPU_sync() uses the internal output callback to reuse existing codes in runNPU_async().
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Thu, 29 Aug 2019 05:57:50 +0000 (14:57 +0900)]
[Bug] Fix deadlock bug in ip plugin to handle async callbacks
This commit fixes the deadlock in ip plugin when handling async callbacks
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Tue, 3 Sep 2019 05:42:01 +0000 (14:42 +0900)]
[Bug] Replace 'npuengine' in some files to 'npu-engine'
This commit replaces 'npuengine' in some files to 'npu-engine'.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Thu, 29 Aug 2019 11:09:18 +0000 (20:09 +0900)]
[Bug] bug fix in output callback handling
This commit fixes minor bugs in output callback handling.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Dongju Chae [Thu, 29 Aug 2019 07:45:21 +0000 (16:45 +0900)]
[API] Revise buffer types defined in libnpuhost.h
This commit revises the buffer types defined in libnpuhost.h.
1) BUFFER_CONT/BUFFER_NON_CONT --> BUFFER_MAPPED
`BUFFER_CONT` cannot be used for user-provided buffers to avoid memory copies, because there is no way to find out their physical address.
For the purpose of avoiding memory copy, dma-buf sharing would be enough.
Also, `BUFFER_MAPPED` will indicate that the buffer already has a mapped buffer and users just can access this.
2) BUFFER_FD --> BUFFER_DMABUF
`BUFFER_FD` is somewhat ambiguous and confused with `BUFFER_FILE`.
So it's changed to `BUFFER_DMABUF`.
Signed-off-by: Dongju Chae <dongju.chae@samsung.com>
Parichay Kapoor [Thu, 22 Aug 2019 12:29:27 +0000 (21:29 +0900)]
[NPUdrvAPI/test] Bug fix for npu open name
Added bug fix for opening npu device name
Added check for npu test cases based on testmode or not
This fix makes npu dev_open_close work with npu fast model simulator
Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>
Wook Song [Fri, 23 Aug 2019 10:18:40 +0000 (19:18 +0900)]
[Core/Utils] Make logging utility functions be thread-safe
This patch makes logging utility functions be thread-safe.
Signed-off-by: Wook Song <wook16.song@samsung.com>