XNNPACK
XNNPACK copied to clipboard
High-efficiency floating-point neural network inference operators for mobile, server, and Web
- Added XNN s16 multiply op
AVX2 vpmadd GEMM move constants to functions to params
Remove unused templates. `generate-f16-vtanh.sh && generate-f32-vtanh.sh` runs without error without these templates.
QB4 AVX256SKX Kernels
vpmaddubsw version of avx512 mimic to vnni - unsigned value (input) is broken into lower 7 bits and upper 1 bit, and multiplied by signed int8 (weights)
`XNN_ARCH_WASMSIMDRELAXED` implies ` XNN_ARCH_WASMSIMD`.
Fix duplicate test and benchmark generation - combine tests and benchmarks into single script
Remove duplicate test generator genxnn calls all scripts Was generate-tests.sh, generate-benchmarks.sh and generate-tests-and-benchmarks.sh Now generate-tests.sh, generate-benchmarks.sh
Don't rely on inferred shapes in expand dims test
S8/U8_VCLAMP microkernels for AVX512