Some workloads running on multi-GPU create 1 process per GPU. So each
process creates a GPU agent on every GPU, but will only create queues on
one GPU. This would cause un-necessary scratch reservation.
Change-Id: I50a216f0bcc0b5f707f3943147390b0ecec1ac22
If the required scratch allocation is too large, ROCr will attempt to
reduce it by lowering the dispatch's targeted occupancy. The reduction
loop however was prone to overflow if waves_per_cu was not a multiple of
waves_per_group. Ensure no overflow by aligning waves_per_cu to
waves_per_group.
On GC 9.4.3 dGPU, dispatches with a large grid size and a
waves_per_group of e.g. 16 may require to reduce occupancy such that
waves_per_cu is less than waves_per_group to ensure the allocation size
is small enough. Allow this while also ensuring the tmpring scratch wave
count is kept divisible by the number of SEs per XCC.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: Ie4016dcd8166a9ae69e9decc26a3eec882b49480
Scratch cache reserved memory is only available for scratch memory use
so do not report this memory as available to the user via the
HSA_AMD_AGENT_INFO_MEMORY_AVAIL api.
Change-Id: I52f96e62536458bcaa52b9f4be5de856d5680dc4
Negative queue validation tests were doing many redundant from-file
kernel object loads in a loop. This was creating many simulataneous open
file handles within many dynamically allocated CodeObject objects. While
the CodeObject class implements RAII on the file handles to cleanup on
destruction, clear_code_object() only gets called on the destruction of
the TestBase-derived test objects (these being a suite abstraction).
Due to this we were hitting file open() EMFILE errors (too many open
files) in gfx94x CPX mode. Move LoadKernelFromObjFile outside of the
test loops and clear_code_object() for each test on each agent.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: I6f9d23fd122720c49a58c22698f097906d2fc97c
Add HSA_ENABLE_SRAMECC environment variable that can be used to
override SRAM ECC mode reported by KFD
Change-Id: I2b95511820a2d3d146a76b03070659c0695b61fd
The gfx940 does not support IMAGE instructions. Any get_info with
IMAGE attributes should return failure.
Signed-off-by: Mike Li <Tianxinmike.Li@amd.com>
Change-Id: I12005628f92780f551ab6f8b41526c66b54c6a59
The function IDs used to be 0 on previous asics but on gfx94x and newer
asics, these bits are set. These bits are used by user applications to
uniquely identify the locations of GPU nodes. These exta bits break
hwloc and are not needed for rocrtst.
Signed-off-by: Mike Li <Tianxinmike.Li@amd.com>
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
Change-Id: I1202f504645b0662d009b9c0926eebb7ddc08d73
gfx940 uses ttmp11 to hold the queue packet index so the first level
trap handler uses ttmp13 instead to save ib_sts.
Repurpose ttmp11[31] to mean that the ttmps are initialized. The issue
was that the debugger could not tell whether ttmp6 was written by the
trap handler when determining the stop reason.
If ttmp11[31]=0, then the trap handler has not been executed and ttmp6
should be assumed to be 0. If ttmp11[31]=1, then ttmp6 holds the
trap_id, if an s_trap instruction caused the exception.
Signed-off-by: Laurent Morichetti <laurent.morichetti@amd.com>
Signed-off-by: Lancelot Six <lancelot.six@amd.com>
Change-Id: I9af903abae044b9ec530306229caf3b883f3ee46
On Linux, the os_thread abstraction is built on top of pthread. Many of
the pthread calls might fail and return error codes. The error
conditions are only checked via assertions (if ever checked) which means
that when doing a release build, no error condition is checked. The
same goes for dlsym/dlinfo and clock_gettime.
This commit improves the situation this by checking the error conditions
and acting accordingly. When the error condition is detected in a
function with a mean to indicate some error to its caller, then this
patch prints some error message and returns. If there is no way to
propagate the error up the call stack, print some error message and
abort the process.
For the os_info::os_info ctor, the only user is CreateThread, which
checks that the built thread is Valid(). If not, nullptr is returned to
the caller.
It could be possible to use exceptions when functions cannot pass
errors, but for now I only use abort as it is what abort would do with
debug build.
Change-Id: I815703c3b95777cc29bb89a7d654ac879c14a759
When building with g++-11.3.0, I have the following warning:
/home/.../core/runtime/runtime.cpp: In member function ‘hsa_status_t rocr::core::Runtime::GetSystemInfo(hsa_system_info_t, void*)’:
/home/.../core/runtime/runtime.cpp:693:56: warning: suggest parentheses around ‘&&’ within ‘||’ [-Wparentheses]
693 | kfd_version.KernelInterfaceMajorVersion == 1 &&
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~
694 | kfd_version.KernelInterfaceMinorVersion >= 12)
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
This patch adds the parenthesis as suggested. This silences the
compiler warning.
No functional change expected.
Change-Id: I69c1a73a432b0f2393dbaf36d4424cf0056c535f
We used to report HSA_STATUS_ERROR_INVALID_ISA when receiving error code
128, but there are several other reasons why we could be exceeding
number of VGPRs, so updating the error code.
Change-Id: I6a6980d5b07b09c93d00dee5207a0d52399bc77e
One some platforms, e.g Arch Linux, -D_GLIBCXX_ASSERTIONS compile flag
is enabled by default, causing a runtime assertion.
Avoid assertion by using std::vector accessor function data().
Change-Id: I118cdf102c3e353f32c618823e363ee1059f3453
Some applications will keep trying to allocate device memory until the
allocation fails. This causes all device memory to be used up and we are
then unable to allocate scratch memory for dispatches. Reserve enough
memory for 1 small scratch allocation.
Change-Id: I968400d41540ba1aca8f28581f229693eec02225
Wait on completion signal for amd_aql_pm4_ib processing
on ASICs with gfx version >= 9.
Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Change-Id: Ia704d9cc5b2535dcf8564a30f694262b113f77a2
Engine offset that is the maximum number of engines is still valid
as offset enum 0 is occupied by blit copies so raise the limit by 1.
Change-Id: I6fcab106290e6647702efe297a4281861da4e0b8
Using backward compatibility paths will provide an #error message. Compile time option added to enable/disable the #error message.
Disabling the same will provide a #warning message
Change-Id: Ib48e361b72176e2845c8f74f980f0234e7eb4a7d
Adds hsa_amd_portable_export_dmabuf and hsa_amd_portable_close_dmabuf
which allow obtaining dmabuf handles to rocr allocations. These handles
may be shared with other APIs to support cross vendor & cross device
memory sharing.
Adds query to return whether dmabuf export is supported
Signed-off-by: Jonathan Kim <Jonathan.Kim@amd.com>
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
Change-Id: I7f98501087d9563d07fc2cb428cc886b1e518b1e
Use mwaitx instructions when busy waiting for signals to reduce CPU
energy usage.
This can be disabled by setting HSA_ENABLE_MWAITX=0
Change-Id: Ic207895a491b2bf6dacba47ef0921df3faad5b5a
Copying memory from device to host with a CPU agent
would cause a poor performance due to the reading of
uncahced device memory by CPU.
Fix it by using a GPU agent.
Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Change-Id: Ia3b562758fe73ef9efaa284f47e67bf569cc7b7b
David Yat Sin
Graham Sider
Saleel Kudchadker
Ranjith Ramakrishnan
Sam Wu
Mike Li
Laurent Morichetti
Alex Sierra
Lancelot SIX
Konstantin Zhuravlyov
Lang Yu
Jonathan Kim
Shweta Khatri
Sean Keely
Yifan Zhang
Cordell Bloor