intrinsics.c (2096B)
1 #define FORCE_INLINE inline __attribute__((always_inline)) 2 3 /* TODO(rnp): msvc probably won't build this but there are other things preventing that as well */ 4 #define clz_u32(a) __builtin_clz(a) 5 #define ctz_u32(a) __builtin_ctz(a) 6 #define sqrt_f32(a) __builtin_sqrtf(a) 7 8 #ifdef __ARM_ARCH_ISA_A64 9 /* TODO? debuggers just loop here forever and need a manual PC increment (step over) */ 10 #define debugbreak() asm volatile ("brk 0xf000") 11 12 /* NOTE(rnp): we are only doing a handful of f32x4 operations so we will just use NEON and do 13 * the macro renaming thing. If you are implementing a serious wide vector operation you should 14 * use SVE(2) instead. The semantics are different however and the code will be written for an 15 * arbitrary vector bit width. In that case you will also need x86_64 code for determining 16 * the supported vector width (ideally at runtime though that may not be possible). 17 */ 18 #include <arm_neon.h> 19 typedef float32x4_t f32x4; 20 typedef int32x4_t i32x4; 21 22 #define cvt_i32x4_f32x4(a) vcvtq_f32_s32(a) 23 #define cvt_f32x4_i32x4(a) vcvtq_s32_f32(a) 24 #define dup_f32x4(f) vdupq_n_f32(f) 25 #define load_f32x4(a) vld1q_f32(a) 26 #define load_i32x4(a) vld1q_s32(a) 27 #define mul_f32x4(a, b) vmulq_f32(a, b) 28 #define set_f32x4(a, b, c, d) vld1q_f32((f32 []){d, c, b, a}) 29 #define sqrt_f32x4(a) vsqrtq_f32(a) 30 #define store_f32x4(a, o) vst1q_f32(o, a) 31 #define store_i32x4(a, o) vst1q_s32(o, a) 32 33 #elif __x86_64__ 34 #include <immintrin.h> 35 typedef __m128 f32x4; 36 typedef __m128i i32x4; 37 38 #define cvt_i32x4_f32x4(a) _mm_cvtepi32_ps(a) 39 #define cvt_f32x4_i32x4(a) _mm_cvtps_epi32(a) 40 #define dup_f32x4(f) _mm_set1_ps(f) 41 #define load_f32x4(a) _mm_loadu_ps(a) 42 #define load_i32x4(a) _mm_loadu_si128((i32x4 *)a) 43 #define mul_f32x4(a, b) _mm_mul_ps(a, b) 44 #define set_f32x4(a, b, c, d) _mm_set_ps(a, b, c, d) 45 #define sqrt_f32x4(a) _mm_sqrt_ps(a) 46 #define store_f32x4(a, o) _mm_storeu_ps(o, a) 47 #define store_i32x4(a, o) _mm_storeu_si128((i32x4 *)o, a) 48 49 #define debugbreak() asm volatile ("int3; nop") 50 51 #endif