experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Use ruy's newly curated default set of paths.

Browse Source 
This:
 - on all CPU architecturesremoves the reference code path, which is never used (the need for a portable non-simd fallback is already filled by the 'standard c++' path).
 - on x86, with --define=tflite_with_ruy=true (non-default), this removes a couple of experimental paths, keeping only AVX2 and AVX512 paths.
 - on x86 for per-channel-quantized ops where ruy is unconditionally used (no alternativee implementation), this enables AVX2 and AVX512 paths. This should fix slowness issues for such per-channel-quantized ops on x86.

PiperOrigin-RevId: 307861040
Change-Id: I4914522b4dc57748095efbcd61049bb3193bbd36
This commit is contained in:
Benoit Jacob 2020-04-22 11:36:21 -07:00 committed by TensorFlower Gardener
parent e853835634
commit 7eb1c830f7
1 changed files with 2 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
tensorflow/lite/kernels/cpu_backend_gemm_ruy.h
View File

@ -78,19 +78,8 @@ struct GemmImplUsingRuy {
ruy::MulParams<AccumScalar, DstScalar> ruy_mul_params;
MakeRuyMulParams(params, &ruy_mul_params);
// If Ruy is not selected intentionally (TFLITE_WITH_RUY not defined)
// and GEMMLOWP_NEON is absent, we fall back to Ruy for some quantized
// kernels. Some Ruy paths are still experimental, so we restrict to reference
// code in that case.
#if !defined(TFLITE_WITH_RUY) && !defined(GEMMLOWP_NEON)
constexpr ruy::Path kRuyPath =
ruy::Path::kReference | ruy::Path::kStandardCpp;
#else
constexpr ruy::Path kRuyPath = ruy::kAllPaths;
#endif
ruy::Mul<kRuyPath>(ruy_lhs, ruy_rhs, ruy_mul_params, context->ruy_context(),
&ruy_dst);
ruy::Mul(ruy_lhs, ruy_rhs, ruy_mul_params, context->ruy_context(),
&ruy_dst);
}
};