diff --git a/nanors/deps/obl/autoshim.h b/nanors/deps/obl/autoshim.h index 21b1a03..297dc7d 100644 --- a/nanors/deps/obl/autoshim.h +++ b/nanors/deps/obl/autoshim.h @@ -6,10 +6,6 @@ #else #if defined(__SSSE3__) || (defined(_MSC_VER) && defined(_M_X64) && !defined(_M_ARM64)) #define OBLAS_SSE3 -#else -#if defined(__aarch64__) || (defined(_MSC_VER) && defined(_M_ARM64)) -#define OBLAS_NEON -#endif #endif #endif #endif diff --git a/nanors/deps/obl/oblas_lite.c b/nanors/deps/obl/oblas_lite.c index a29032e..bd1c47e 100644 --- a/nanors/deps/obl/oblas_lite.c +++ b/nanors/deps/obl/oblas_lite.c @@ -124,8 +124,7 @@ void obl_axpyb32_ref(u8 *a, u32 *b, u8 u, unsigned k) } \ } while (0) -#else -#if defined(OBLAS_AVX2) +#elif defined(OBLAS_AVX2) #include #undef OBLAS_ALIGN @@ -174,13 +173,13 @@ void obl_axpyb32_ref(u8 *a, u32 *b, u8 u, unsigned k) } while (0) #else -#if defined(OBLAS_SSE3) || defined(OBLAS_NEON) -#if defined(OBLAS_NEON) -#define SIMDE_ENABLE_NATIVE_ALIASES -#include -#else + +#if defined(OBLAS_SSE3) #include #include +#else +#define SIMDE_ENABLE_NATIVE_ALIASES +#include #endif #undef OBLAS_ALIGN @@ -231,26 +230,6 @@ void obl_axpyb32_ref(u8 *a, u32 *b, u8 u, unsigned k) _mm_storeu_si128(ap, _mm_xor_si128(_mm_loadu_si128(ap), ret_hi)); \ } \ } while (0) - -#else - -#undef OBLAS_ALIGN -#define OBLAS_ALIGN (sizeof(void *)) - -#undef OBL_SHUF -#define OBL_SHUF(op, a, b, f) \ - do { \ - op##_ref(a, b, u, k); \ - } while (0) - -#undef OBL_SHUF_XOR -#define OBL_SHUF_XOR - -#undef OBL_AXPYB32 -#define OBL_AXPYB32 obl_axpyb32_ref - -#endif -#endif #endif #define OBL_NOOP(a, b) (b)