Previously, we used the following approach and Comgr actions
for device lib linking:
AMD_COMGR_COMPILE_SOURCE_TO_BC (compile with clang driver)
AMD_COMGR_ADD_DEVICE_LIBRARIES (link in device libs with
llvm-link API)
However, the clang driver can link in device libraries as part
of compilation, assuming a --rocm-path is set. In this context,
this is accomplished by using the following Comgr action instead:
AMD_COMGR_COMPILE_SOURCE_WITH_DEVICE_LIBS_TO_BC (compile and
link in device libs with clang driver)
Change-Id: I661465865365afecc44aa15d4df91bfab361af8d
hipcc and clang++ both have logic to detect the installed hardware
and to automatically select the appropriate AMDGPU target when it is
left unspecified. When the AMDGPU_TARGETS property is initialized with
a set of default values, it results in the addition of an explicit set
of --offload-arch flags being passed. These explicit architecture flags
disable the architecture autodetection in the compiler.
The resulting behaviour from setting fixed defaults makes it unpleasant
to compile with CMake because they increase the build times for projects
unless they are overriden (as most users do not need to build for all
five default architectures). The fixed defaults are also troublesome for
users with hardware not included in the default set (e.g., gfx1011,
gfx1031, gfx1100).
A possible alternative might be to detect the architecture within
hip-config.cmake rather than running the detection logic on each
compiler invocation. However, this approach is simpler.
Change-Id: I9495d766b7eed03852eb4dc72b0aabe4100bc32c
Signed-off-by: Cordell Bloor <Cordell.Bloor@amd.com>
The previous implementation using std::copy() resulted in
differences between the in-memory and on-disk representations.
With the updated implementation, we get the same contents.
Change-Id: Iadfae3cd7f7ba99538da2ac4f11f30f5a78260d8
The change enables VM support in graphs on Windows. That allows
to avoid caching of all allocations at the cost of map/unmap
overhead during memory create/destroy.
Change-Id: I792be00fba099e5e5d3cd44a963e1dfd6976a86d
hipStreamPerThrdCompilerOptn.cc test fails to build with cudaStreamGetCaptureInfo_v2
in CUDA 12.0.
fix was to change runtime API cudaStreamGetCaptureInfo_v2
to Driver cuStreamGetCaptureInfo_v2
Change-Id: I44a0110770d3246f5345092acae301c9a2f6d520
- Introduce a state variable to indicate if HwProfiling is enabled to
eliminate a possible data race of vector<> signals_.
Change-Id: Id504cc76d7fa9f7e6455587dd232b60ccbbb735b
- correct error for hipStreamWaitEvent when event recorded before
capture
- correct hipEventSync when event is synced during capture
Signed-off-by: sdashmiz <shadi.dashmiz@amd.com>
Change-Id: I7ecbed5621eaf323846d4ccb20ec112aaa8a5757
- The implementation in mempool graphs requires refcounting VA object.
That requires release() to update the map only on the actual destruction.
- Add GPU event tracking for paging operation. Otherwise, runtime
may not always flush IB.
Change-Id: Idf99ffb894321a38e04b490116a7ca435635918d
GraphMemcpyNodeSetParamsFrom/ToSymbol APIs neew to check device id for
original src/dst is same as what is passed in while set.
Change-Id: If0b610808223dce9115562bb5e9b31c8eaa2df22
Remove wrong functions
__hipMapFromNativeFloat4()
__hipMapToNativeFloat4()
Replace them with mapFrom() and mapTo()
Change-Id: I1692b92d397bfd732d562778f918ebf8ca532e93
pghafari
Jacob Lambert
sdashmiz
AravindanC
Cordell Bloor
Satyanvesh Dittakavi
Jatin Chaudhary
Ioannis Assiouras
Saleel Kudchadker
Jaydeep Patel
taosang2
German
kjayapra-amd
Ajay
Sourabh Betigeri
victzhan
Alex Xie
Maneesh Gupta