[CPU/CUDA ep] Improve DeformConv op performance#27824

Merged

tianleiwu merged 54 commits into

microsoft:mainfrom

ShirasawaSama:feature/improve-deform-conv-pref

Apr 9, 2026

Commits on Apr 2, 2026

Remove DeformConvCopyGemmOutputRowMajorToNCHW
ShirasawaSama
committed
Adjust parallel cost for DeformableIm2col
ShirasawaSama
committed
Refactor deform conv bilinear with plan
ShirasawaSama
committed
Simplify DeformConv im2col plan paths and fix mask indexing bug
ShirasawaSama
committed
Refactor deform conv im2col to use a unified tiled path with context structs and remove redundant dispatch branches
ShirasawaSama
committed
Refactor DeformConv sampling plan to AoSoA layout and use Eigen for bilinear accumulation
ShirasawaSama
committed
Refine DeformConv naming clarity and avoid redundant workspace size recomputation
ShirasawaSama
committed
Optimize DeformConv by removing streaming plan logic and making bilinear sampling branchless
ShirasawaSama
committed
Refactor Deformconv cpu op
ShirasawaSama
committed
Harden DeformConv integer bounds checks and streamline hot-path casts with branchless AoSoA sampling plan execution.
ShirasawaSama
committed
Refactor DeformConv bounds validation
ShirasawaSama
committed
Add compute-time bounds checks with size_t-safe indexing
ShirasawaSama
committed
Optimize CPU DeformConv plan generation with kernel meta precompute
ShirasawaSama
committed
Refactor DeformConv kernel meta setup into a params-based cached
ShirasawaSama
committed
Refactor CPU DeformConv bias add to avoid div/mod and extract DeformConvCpuAddBias
ShirasawaSama
committed
Annotate DeformConv CPU bias/col paths with ORT_CPU_RESTRICT and force-inline bias row helper
ShirasawaSama
committed
CPU DeformConv bilinear sampling uses fast floor and inverted bounds check
ShirasawaSama
committed
Flatten CPU DeformConv bilinear sampling plan build tasks across spatial pixels for better thread pool scaling
ShirasawaSama
committed
Optimize CPU DeformConv sampling and bias parallelism with flattened work and N equals 1 fast path
ShirasawaSama
committed
Add detailed comments for DeformConv CPU implementation
ShirasawaSama
committed
Reformat codes
ShirasawaSama
committed
Optimize DeformConv CPU kernel by removing mutex and heap allocations
ShirasawaSama
committed
Optimize CUDA DeformConv kernel with static mask branching and tuned loop unrolling
ShirasawaSama
committed
CUDA DeformConv reduce 64 bit index pressure in im2col hot path
ShirasawaSama
committed
Increase InlinedVector capacity in DeformConv for 7x7 kernels
ShirasawaSama
committed
Optimize DeformConv bias indexing with int32/int64 dispatch and cleaner 64-bit gating checks
ShirasawaSama
committed
Optimize CUDA DeformConv bias add with 2D launch fast path and int32/int64 index dispatch
ShirasawaSama
committed
Optimize CUDA DeformConv by using 32-bit index arithmetic when safe and refactoring fallback check
ShirasawaSama
committed
Refactor path indexing
ShirasawaSama
committed
optimize deformconv bilinear sampling with interior fast path
ShirasawaSama
committed
Rduce deformconv address math in dynamic im2col path
ShirasawaSama
committed
Tune deform conv im2col addressing and bilinear sampling
ShirasawaSama
committed
Cuda deform conv replace 5x5 im2col launch specialization with 7x7
ShirasawaSama
committed
Pick chunk size by min rounds then balanced ceil
ShirasawaSama
committed
Fix CUDA DeformConv im2col mask stride unused-variable warning
ShirasawaSama
committed
Document and tidy CUDA DeformConv
ShirasawaSama
committed
Make deform conv bilinear sampling branchless with masked safe loads
ShirasawaSama
committed
Improve comments and code styles
ShirasawaSama
committed
Improve deform conv im2col load balance for offset_group=1
ShirasawaSama
committed
Harden DeformConv index-width guard and align mask test comment
ShirasawaSama
committed
Optimize BilinearInterpolate with one-sided bounds and float mask selp
ShirasawaSama
committed
Make deform_conv_attributes.h self-contained for numeric_limits
ShirasawaSama
committed
Clarify bilinear index int32 safety comments
ShirasawaSama
committed

Commits on Apr 6, 2026

Commits on Apr 7, 2026