#include "sn_target.h"

#if SN_IS_ARCH(SN_ARCH_X86_64)

#include "sn_intrinsic.h"

DECLARE_KER_PROC_RIGHT_ANS(float, sse, , quality)
{
        static const __m128 abs_mask_source = { .m128_i32 = { 0x80000000, 0x80000000, 0x80000000,
                                                              0x80000000 } };
        register __m128 abs_mask = _mm_load_ps(abs_mask_source.m128_i32);

        register __m128 col0 = _mm_andnot_ps(abs_mask, _mm_load_ps((float *)&right->f32_x[0]));
        register __m128 col1 = _mm_andnot_ps(abs_mask, _mm_load_ps((float *)&right->f32_x[1]));
        register __m128 col2 = _mm_andnot_ps(abs_mask, _mm_load_ps((float *)&right->f32_x[2]));
        register __m128 col3 = _mm_andnot_ps(abs_mask, _mm_load_ps((float *)&right->f32_x[3]));

        // norm
        register __m128 row_sum = _mm_add_ps(col0, col1);
        row_sum = _mm_add_ps(row_sum, col2);
        row_sum = _mm_add_ps(row_sum, col3);

        // max(0,2) and max(1,3)
        register __m128 max = _mm_max_ps(row_sum, _mm_movehl_ps(row_sum, row_sum));
        max = _mm_max_ps(max, _mm_shuffle_ps(max, max, _MM_SHUFFLE(2, 3, 0, 1)));
        float norm = _mm_cvtss_f32(max);

        ans->f32_x[0].m128_f32[0] = norm;

        // min on cross
        col1 = _mm_shuffle_ps(col1, col1, _MM_SHUFFLE(0, 3, 2, 1));
        col2 = _mm_shuffle_ps(col2, col2, _MM_SHUFFLE(1, 0, 3, 2));
        col3 = _mm_shuffle_ps(col3, col3, _MM_SHUFFLE(2, 1, 0, 3));
        max = _mm_min_ps(col0, col2);
        register max0 = _mm_min_ps(max, col1);
        max = _mm_min_ps(max0, max);

        float min_val = _mm_cvtss_f32(max);
        ans->f32_x[0].m128_f32[1] = min_val;

        float quality = min_val / (norm + 1e-20f);
        ans->f32_x[0].m128_f32[2] = quality;
}

#endif