opencv/cmake/checks/cpu_fp16.cpp

35 lines
953 B
C++
Raw Normal View History

2016-07-18 22:45:16 +08:00
#include <stdio.h>
#if defined __F16C__ || (defined _MSC_VER && _MSC_VER >= 1700 && defined __AVX__) || (defined __INTEL_COMPILER && defined __AVX__)
2016-07-18 22:45:16 +08:00
#include <immintrin.h>
int test()
{
const float src[] = { 0.0f, 0.0f, 0.0f, 0.0f };
short dst[8];
__m128 v_src = _mm_load_ps(src);
__m128i v_dst = _mm_cvtps_ph(v_src, 0);
_mm_storel_epi64((__m128i*)dst, v_dst);
return (int)dst[0];
}
#elif (defined __GNUC__ && (defined __arm__ || defined __aarch64__)) /*|| (defined _MSC_VER && defined _M_ARM64)*/
// Windows + ARM64 case disabled: https://github.com/opencv/opencv/issues/25052
2016-07-18 22:45:16 +08:00
#include "arm_neon.h"
int test()
{
2023-12-09 09:32:21 +08:00
const float src[] = { 0.0f, 0.0f, 0.0f, 0.0f };
short dst[8];
float32x4_t v_src = *(float32x4_t*)src;
2016-07-18 22:45:16 +08:00
float16x4_t v_dst = vcvt_f16_f32(v_src);
2023-12-09 09:32:21 +08:00
*(float16x4_t*)dst = v_dst;
return (int)dst[0];
2016-07-18 22:45:16 +08:00
}
#else
#error "FP16 is not supported"
#endif
int main()
{
printf("%d\n", test());
return 0;
}