[MPS Dev] 2026.07.01

[Manfredss]

PR Category

User Experience

PR Types

New features

Description

Expands the Apple MPS (MetalPerformanceShadersGraph) backend from 51 → 79
forward kernels by porting the missing "siblings" of the op categories already
implemented — the ops that map cleanly onto MPSGraph primitives. All kernels are
forward-only and float32-only, one self-contained .mm file each, following the
existing MPS kernel template.

What's added (28 kernels)

Category	Ops
Unary math	asinh acosh atanh expm1 log1p trunc
Activations (no attr)	tanh_shrink softsign logsigmoid hardswish relu6 swish
Activations (1 attr)	elu celu mish softshrink hard_shrink
Activations (2 attr)	hardtanh hardsigmoid thresholded_relu softplus
Reduction activation	log_softmax
Binary elementwise	atan2 floor_divide remainder fmax fmin heaviside

Implementation notes

• Each kernel builds an MPSGraph, runs it via runWithMTLCommandQueue:…, and
  registers float only, matching the existing MPS kernels. New files are picked
  up automatically by the mps/*.mm glob in paddle/phi/kernels/CMakeLists.txt
  (no build-file changes).
• Semantics are matched to the CPU reference functors (funcs/activation_functor.h,
  funcs/elementwise_functor.h) so MPS output equals CPU output:
  • floor_divide = floor(x/y), remainder = x - floor(x/y)*y (sign of divisor)
  • logsigmoid / softplus / mish use the numerically-stable functor forms
  • hardswish bakes threshold/scale/offset = 6/6/3; relu6 clamps to 6; swish β=1
  • elu = $α(e^x−1)$ for x<0; celu = $α(e^{x/α}−1)$ for x<0
• log_softmax relies on MPSGraph reductions keeping the reduced axis (size 1),
  so the max/sum subtractions broadcast without an explicit reshape.
• New MPSGraph APIs used: selectWithPredicateTensor:…,
  atan2WithPrimaryTensor:secondaryTensor:, and the inverse-hyperbolics
  asinh/acosh/atanhWithTensor:.

Testing

Adds test/test_mps_math_activation_kernels.py, which exercises every new op on
the MPS backend against the CPU backend (same public paddle API) plus a set of
numpy known-value checks. Guarded by paddle.is_compiled_with_mps() /
paddle.mps.is_available(), so it skips cleanly where MPS is unavailable.

Build + run on macOS:

make sure you have cmake and gcc installed on your Mac. For more information, check this guidance:
macOS 下使用 make 从源码编译

rm -rf build && mkdir build && cd build
cmake .. -DPY_VERSION=3.10 -DWITH_GPU=OFF -DWITH_MPS=ON
make -j$(sysctl -n hw.ncpu)
python test/test_mps_math_activation_kernels.py

Scope

Forward, float32 only. Reductions (prod/all/any/argmax/argmin/
logsumexp), selection/util (where/clip/cast/isnan/bitwise), and shape
ops remain as a follow-up.

🤖 Generated with Claude Code

是否引起精度变化

否