hipGetLastError/hipPeekAtLastError should really return the last error
produced by any of the runtime calls that did not successfuly complete,
not just the error code of the last runtime call.
For example, in this snippet of code:
1: hipLaunchKernelGGL(MyKernel, dim3(2), dim3(128), 0, 0, 10000,x);
2: hipDeviceSynchronize();
3: hipError_t error = hipGetLastError();
If the call to hipLaunchKernelGGL fails, hipDeviceSynchronize (which
succeeds) should not reset the last error to hipSuccess. hipGetLastError
should still return the non-success error code returned by
hipLaunchKernelGGL.
The last error is reset to hipSuccess after calling hipGetLastError.
Change-Id: Ib7c039067c53c94c99c8ecd83f54212bcef06f81
[ROCm/clr commit: 3dc47bc839]
hipGetLastError/hipPeekAtLastError should really return the last error
produced by any of the runtime calls that did not successfuly complete,
not just the error code of the last runtime call.
For example, in this snippet of code:
1: hipLaunchKernelGGL(MyKernel, dim3(2), dim3(128), 0, 0, 10000,x);
2: hipDeviceSynchronize();
3: hipError_t error = hipGetLastError();
If the call to hipLaunchKernelGGL fails, hipDeviceSynchronize (which
succeeds) should not reset the last error to hipSuccess. hipGetLastError
should still return the non-success error code returned by
hipLaunchKernelGGL.
The last error is reset to hipSuccess after calling hipGetLastError.
Change-Id: Ib7c039067c53c94c99c8ecd83f54212bcef06f81
HIP_RETURN(x) is not a value, it is a statement ending with a return,
this commit replaces all instances of "return HIP_RETURN(x)" with simply
"HIP_RETURN(x)"
Change-Id: I03293b2684a65367ff55e02b3a71ea49ec7a517a
[ROCm/clr commit: 3911184607]
HIP_RETURN(x) is not a value, it is a statement ending with a return,
this commit replaces all instances of "return HIP_RETURN(x)" with simply
"HIP_RETURN(x)"
Change-Id: I03293b2684a65367ff55e02b3a71ea49ec7a517a
Add new API for creating a reduction operation which multiplies the input by a rank-specific scalar before doing an inter-rank summation (see: ncclRedOpCreatePreMulSum).
Improve CollNet (SHARP) performance of ncclAllReduce when captured in a CUDA Graph via user buffer registration.
Add environment variable NCCL_NET_PLUGIN="<suffix>" to allow user to choose among multiple NCCL net plugins by substituting into "libnccl-net-<suffix>.so".
Fix memory leak of NVB connections.
Fix topology detection of IB Virtual Functions (SR-IOV).
[ROCm/rccl commit: e11238b302]
Add new API for creating a reduction operation which multiplies the input by a rank-specific scalar before doing an inter-rank summation (see: ncclRedOpCreatePreMulSum).
Improve CollNet (SHARP) performance of ncclAllReduce when captured in a CUDA Graph via user buffer registration.
Add environment variable NCCL_NET_PLUGIN="<suffix>" to allow user to choose among multiple NCCL net plugins by substituting into "libnccl-net-<suffix>.so".
Fix memory leak of NVB connections.
Fix topology detection of IB Virtual Functions (SR-IOV).
std::mem_fun() and std::bind2nd() are removed in c++17. Switch to
simpler logic that does not require those functions.
Change-Id: I19a31f076e1813e367615bd377b424046ce144c7
[ROCm/clr commit: d934612948]
std::mem_fun() and std::bind2nd() are removed in c++17. Switch to
simpler logic that does not require those functions.
Change-Id: I19a31f076e1813e367615bd377b424046ce144c7
CMake does not provide a way to query the NUMA library, hence we need
to find it manually.
Change-Id: I370b286acdee75cbebc21340da3c432c79f8ffa7
[ROCm/clr commit: dd23379ac8]
std: :mem_fun() is removed in c++17. Simplify logic to not require it.
Change-Id: Ic9a4753b48dd13fcb20cd5b90ff73c3df3211b9f
[ROCm/clr commit: c68f024b35]
For the fillBuffer shader, if there are two 32bit writes to a MMIO
register, it can get dropped. It has to be a single 64bit write.
Add optimization to fillBuffer to write 64bit and 16bit writes.
Change-Id: I3aa78e027898f8ae01e9c8f09004615673720c2b
[ROCm/clr commit: 21ba34d0fe]
For the fillBuffer shader, if there are two 32bit writes to a MMIO
register, it can get dropped. It has to be a single 64bit write.
Add optimization to fillBuffer to write 64bit and 16bit writes.
Change-Id: I3aa78e027898f8ae01e9c8f09004615673720c2b
Jonathan R. Madsen
Laurent Morichetti
Jason Tang
Chauncey Hui
Wenkai Du
Ke Wen
Wenkai Du
Vladislav Sytchenko
Chun Yang
Saleel Kudchadker
Sarbojit Sarkar
Jatin Chaudhary