Use ruy's newly curated default set of paths.
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
diff --git a/tensorflow/lite/kernels/cpu_backend_gemm_ruy.h b/tensorflow/lite/kernels/cpu_backend_gemm_ruy.h
index 3769479..a5bcccd 100644
--- a/tensorflow/lite/kernels/cpu_backend_gemm_ruy.h
+++ b/tensorflow/lite/kernels/cpu_backend_gemm_ruy.h
@@ -78,19 +78,8 @@
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);
}
};
