[WIP] Generate LLVM IR for target parallel kernels #226

DominikAdamski · 2023-08-22T11:26:21Z

This MR provides patches for compilation of simple parallel OpenMP Fortran kernel:

program main
   use omp_lib
   integer :: e(256)
   integer :: i
   integer :: d
   !$omp target parallel do map(tofrom: e)
   do i = 1, 256
        e(i) = i + 2
   end do
        !$omp end target parallel do
   print *,"parallel array"
   print *, e

   !$omp target parallel map(from: d)
         d = omp_get_num_threads()
    !$omp end target parallel
   print *,"parallel"
   print *, d
   !$omp target  map(from: d)
         d = omp_get_num_threads()
    !$omp end target
   print *,"generic"
   print *, d

   print *,"passed"
end program main

Known limitations:

Parallel region has to cover target region. If target region contains code which is not covered by parallel pragma then execution mode isn't correctly set up.
Allocas which are used only inside loop body regions are not moved to loop body. This problem should be fixed on MLIR level.
The CodeExtractor contains hardcoded allocation space for struct arguments
No support for additional parallel clauses (like num_threads, etc.) for offloaded code.
Skipped compilation of the new device RTL function for target teams distribute pragma

This patch adds initial support for detecting type of launched kernel. If there is a parallel pragma inside target region, then we assume that we have SPMD kernel. This assumption is not valid for target regions for which parallel region does not cover the whole target region.

Proposed implementation of the new device OpenMP functions relies on internal functions which are not used in current version of the OpenMP runtime.

TODO: Modify CodeExtractor so that it will take alloca address space from constructor parameter

TODO: Add missing definitions for the newwq OpenMP device RTL functions TODO: Add missing attributes for the new OpenMP device RTL functions

DominikAdamski and others added 9 commits August 22, 2023 03:33

finalization_fix

9facef1

work_kmpc_parallel_51

cf41abd

Sergio's 227 PR

4e8013e

Johannes proposition of the new OpenMP runtime

c26cf26

Do not build the new Device RTL functions which cannot be compiled

adb419d

Proposed implementation of the new device OpenMP functions relies on internal functions which are not used in current version of the OpenMP runtime.

[WIP][CodeExtractor] Modify alloca address space for struct

b4b571c

TODO: Modify CodeExtractor so that it will take alloca address space from constructor parameter

[WIP] Add OpenMP definitions

a0703ec

TODO: Add missing definitions for the newwq OpenMP device RTL functions TODO: Add missing attributes for the new OpenMP device RTL functions

[OpenMP] Generate code for target worksharing loop

470dde0

DominikAdamski force-pushed the gpu_parallel_51_main_21_08 branch from 6203694 to 470dde0 Compare September 5, 2023 12:16

DominikAdamski closed this Sep 29, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP] Generate LLVM IR for target parallel kernels #226

[WIP] Generate LLVM IR for target parallel kernels #226

DominikAdamski commented Aug 22, 2023 •

edited

Loading

[WIP] Generate LLVM IR for target parallel kernels #226

[WIP] Generate LLVM IR for target parallel kernels #226

Conversation

DominikAdamski commented Aug 22, 2023 • edited Loading

DominikAdamski commented Aug 22, 2023 •

edited

Loading