Avoid atomics for CUDA elementwise shadow updates

[minansys]

Summary

This centralizes Enzyme's reverse-mode shadow accumulation atomic decision in DiffeGradientUtils::shouldUseAtomicShadowUpdate and makes the CUDA elementwise-read optimization explicit.

For enzyme_elementwise_read functions, each CUDA work item is expected to accumulate into a distinct shadow location, so Enzyme can emit a normal load/add/store instead of a generated atomicrmw fadd. Conservative cases still keep atomics, including unannotated CUDA paths, shared-memory reductions, and unknown aliasing cases.

Changes

• Add a named helper for deciding when generated shadow updates need atomics.
• Add cuda-elementwise-atomic.ll to check both sides of the behavior:
  • annotated CUDA elementwise read emits no atomicrmw
  • unannotated CUDA read still emits atomicrmw fadd
• Add a CUDA benchmark input that validates gradient correctness and compares unannotated atomic versus annotated elementwise kernels.

Validation

• Formatted changed C++/CUDA files with /home/minxu/code/enzyme/.tools/bin/clang-format-16.
• Rebuilt Enzyme:
  • cmake --build build/enzyme-llvmorg-19.1.7 -- -j 8
• New focused lit regression:
  • python3 deps/llvm-project/llvm/utils/lit/lit.py -sv build/enzyme-llvmorg-19.1.7/test --filter 'Enzyme/ReverseMode/cuda-elementwise-atomic\\.ll$'
• Direct generated-IR checks for existing behavior:
  • elementwise-read.ll: no atomicrmw
  • cuda.ll: atomicrmw fadd present
  • sharedmem.ll: atomicrmw fadd present
  • atomicfadd.ll: atomicrmw volatile fadd present
• CUDA benchmark/smoke on NVIDIA GeForce RTX 3070 Laptop GPU, sm_86:
  • scripts/build_enzyme_cuda.sh --skip-checkout --skip-build --skip-enzyme-tests
  • Run 1: n=4194304 reps=50 atomic_ms=0.592671 elementwise_ms=0.572068 speedup=1.036x
  • Run 2: n=4194304 reps=50 atomic_ms=0.590989 elementwise_ms=0.546058 speedup=1.082x

Notes

• The local check-enzyme target is misconfigured in this workspace and invokes /llvm-lit, which does not exist.
• After symlinking the built FileCheck into the LLVM install prefix, a broad .ll lit attempt with --max-failures=20 still failed on existing LLVM 19 opaque-pointer FileCheck spelling mismatches and a missing not helper in the local lit environment. The new regression passed independently.

[minansys]

Superseded by #3, which implements the repeated CUDA global-load root fix instead of the earlier elementwise atomic approach.