Ye Luo

Make hip-lang-config.cmake.in more robust.

1

Still need to address #2433 Make the following two cases work with https://github.com/ye-luo/cmake_gpu/tree/master/test_rocm Case 1. ``` cmake -DCMAKE_HIP_COMPILER=/opt/rocm/bin/amdclang++ .. ``` Case 2 AOMP or upstream clang doesn't use ROCm layout....

unrobust CMake search logic

2

ROCM 4.5 and hip-lang package assume hip is installed as `/hip` and tries to access paths outside `/hip` and inside ``` for example https://github.com/ROCm-Developer-Tools/HIP/blob/cddb52549b4f4fce9165cac8b2ccf25173ba3157/hip-lang-config.cmake.in#L108 reproducer ``` cmake_minimum_required(VERSION 3.21.0) project(test_hip CXX)...

Reduce synchronization needed

11

See #160 figure. Currently every asynchronous transfer and kernel_dispatch within a target offload has single_wait_scaquire (synchronization) right after. It should be nice to optimize them by enqueuing H2D, kernel_dispatch, D2H...

Runtime overhead on H2D and D2H tranfers

16

This figure comes from a single offload region. https://github.com/ye-luo/miniqmc/blob/46073436a432fc0472bf793784f9f87b5f8fdfcb/src/QMCWaveFunctions/einspline_spo_omp.cpp#L407 The map to/from arrays are already mapped with "target enter data" Why there are still memory_pool_allocate and agents_allow access and memory_pool_free...

OpenMP HIP Memory interop issue

11

HIP to OpenMP hipMalloc and pass the pointer to OpenMP target kernel via `is_device_ptr` works. OpenMP to HIP `omp_target_alloc` and pass the pointer to HIP API/kernel print error. For example,...

The addtional thread for GPU

2

It seems that offload plugin or down to hsa creates additional thread to communicate with the GPU. Such additional thread seems floating around and competing with regular OpenMP threads. It...

device plugin dead-lock at a atmi_malloc failure

7

On my apu laptop when graphic memory is set low, the memory allocation failure caused a deadlock in the device plugin. ``` [/home/estewart/git/aomp11/amd-llvm-project/openmp/libomptarget/plugins/hsa/impl/data.cpp:99] atmi_malloc failed: HSA_STATUS_ERROR_INVALID_ALLOCATION ``` backtrace ``` __lll_lock_wait...

std::complex support needed

8

https://github.com/ye-luo/openmp-target/blob/master/hands-on/tests/complex/complex.cpp ``` $ clang++ -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -march=native -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906 complex.cpp lld: error: undefined symbol: __mulsc3 >>> referenced by /tmp/complex-gfx906-72c03e-gfx906-c2b83e.o:(__omp_offloading_10304_2920ae4__Z8test_mulIfSt7complexIfES1_EvT0_T1__l59) >>> referenced by /tmp/complex-gfx906-72c03e-gfx906-c2b83e.o:(__omp_offloading_10304_2920ae4__Z8test_mulIfSt7complexIfES1_EvT0_T1__l59) lld: error: undefined symbol: __divsc3 >>> referenced...

More register used when multiple target regions are compiled together

3

The source code I'm using has multiple offload regions in different member functions of a class. If I enable individual target region and comment the other target pragma Kernel 1...

linker error

7

Using 0.7-7. The AOMP linker works on more complicated miniQMC but failed in linking the following test case. https://github.com/ye-luo/openmp-target/tree/master/hands-on/tests/link_static_fat_bin ``` /usr/lib/aomp/bin/clang++ -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906 -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -c classA.cpp rm -f mylib.a...