The heap must be cleared once per device, but ROCclr doesn't
create a queue per device in HIP. Hence, the clear operation will
be performed during the first queue creation.
Change-Id: I52ceb06d67d11cde6d019c5ab510059f426a9bfb
HIP uses icd_dispatch.h, which is not a public header. Using the public
CL/cl_icd.h removes inclusion of the private header.
This allows compiling HIP without relying on OCL ICD source, which makes
distributing a standalone ROCclr easier.
For WIN32, a few other headers are required.
Signed-off-by: Jeremy Newton <Jeremy.Newton@amd.com>
Change-Id: I0cab173b9753c5b0ab3be9222aa8bb0545cdeaab
cl_icd.h includes cl_egl.h, but cl_egl.h is excluded from install. It
was reported by the community that including cl_icd.h would error due to
the missing cl_egl.h file.
I'm assuming that cl_egl.h was excluded because EGL isn't important for
ROCm. Since cl_icd.h has some important typedefs, it makes sense just
to include cl_egl.h, which seems to be a common practice in open source
projects.
Signed-off-by: Jeremy Newton <Jeremy.Newton@amd.com>
Change-Id: I506257c2dc51512ec8e11b1e5dadbe6e48ad785b
Prior to this change, when Git_FOUND was false, HIP_VERSION_BUILD_ID
would be undefined in the CMake code. The value of HIP_VERSION_BUILD_ID
in <hip/hip_version.h> is taken from the CMake variable, so it was being
defined as nothing in those cases. That would cause compilation failures,
as src/hip_global.cpp contains the function:
size_t amd_dbgapi_get_build_id() {
return HIP_VERSION_BUILD_ID;
}
which would become
size_t amd_dbgapi_get_build_id() {
return ;
}
after preprocessing. To prevent this, we can define the version
information to a default value when Git is not found.
A related problem was reported by Harmen Stoppels in
https://github.com/ROCm-Developer-Tools/HIP/pull/2218. When Git
is not available (or if the library is being built from a tarball),
the HIP_VERSION_GITHASH is not defined. This causes trouble because
HIP_LIB_VERSION_STRING is defined as "X.Y.Z-${HIP_VERSION_GITHASH}"
and therefore becomes "X.Y.Z-".
The incomplete version string becomes a problem when it is appended
to the shared library file name. File names that end with '-' confuse
the linker. They cause strange errors when attempting to link to the
HIP library. This problem can be prevented by dropping the trailing
dash and using "X.Y.Z" as the version string when HIP_VERSION_GITHASH
is not defined.
Change-Id: I6e290c1f1b603ba30c9ded885e125d9ca9a2e688
Signed-off-by: Cordell Bloor <Cordell.Bloor@amd.com>
OCLCreateImage test passed when the image was not supported at all.
Add a warning.
For ASIC without image support, let the test pass.
Change-Id: I13c015561a51879f052c2c2d56c672ffbdbe0641
- channel format should be consistent accross parameters
Signed-off-by: sdashmiz <shadi.dashmiz@amd.com>
Change-Id: I1e7680ab9262a66e7bea4aaa21f061044b0f7956
Update HIP's unsafeAtomicAdd to:
- Compile properly even when not compiling for gfx90a
- Fall back to safe atomic add on non-gfx90a architectures
- use flat atomic add for FP64 on gfx90a, instead of dynamically
checking memory spaces.
In addition, when the compiler is passed -munsafe-fp-atomics, it
will define __AMDGCN_UNSAFE_FP_ATOMICS__. When this happens, the
compiler is requesting that the HIP headers force all HIP
atomicAdd() calls on floats or doubles to use their unsafe versions.
This patch thus causes unsafeAtomicAdd() calls when that define
is seen. This call to unsafeAtomicAdd() is also done for atomicSub(),
since that calls atomicAdd underneath. This is not done for
system-scope atomicAdd because, on gfx90a, system-scope atomic FP
add instructions would need to target fine-grained memory, which is
always unsafe.
This patch also creates safeAtomicAdd() functions for float and double.
These functions will create a standalone safe atomic, even when the
application is compiled with -munsafe-fp-atomics.
Finally, this patch adds wrappers in the Nvidia path of HIP so that
these HIP functions call through to atomicAdd there as well.
Change-Id: I8af0621d3d28ea30c9278bfeea7393d03bbdac6d
Enqueue a handler callback for hipEventRecords(aka marker_ts_) for every
64 submits, This recycles the memory if we dont end up calling
synchronize for the longest time.
Change-Id: I3d39fe76d52a5d81387927edd85b5663b563682c
Disable hostcall buffer in OCL for now. COv5 can add hostcallbuffer
metadata for unknown reason. OCL may fail the buffer allocation
and kernel launch.
Change-Id: I34a6a45bac86c57422b764c0d69760c96920d6c5
German Andryeyev
Jason Tang
Maneesh Gupta
Chauncey Hui
Sarbojit Sarkar
Rakesh Roy
Satyanvesh Dittakavi
pghafari
Christophe Paquot
Todd tiantuo Li
Cordell Bloor
Tao Sang
Jeremy Newton
Sourabh Betigeri
Alex Xie
Ajay
kjayapra-amd
Jaydeep Patel
Saleel Kudchadker
sdashmiz
Joseph Greathouse