52 float *y,
float *cb,
float *cr,
ui32 repeat)
59 for (
ui32 i = (repeat + 3) >> 2; i > 0; --i)
61 __m128 mr = _mm_load_ps(r);
62 __m128 mb = _mm_load_ps(b);
63 __m128 my = _mm_mul_ps(alpha_rf, mr);
64 my = _mm_add_ps(my, _mm_mul_ps(alpha_gf, _mm_load_ps(g)));
65 my = _mm_add_ps(my, _mm_mul_ps(alpha_bf, mb));
67 _mm_store_ps(cb, _mm_mul_ps(beta_cbf, _mm_sub_ps(mb, my)));
68 _mm_store_ps(cr, _mm_mul_ps(beta_crf, _mm_sub_ps(mr, my)));
70 r += 4; g += 4; b += 4;
71 y += 4; cb += 4; cr += 4;
77 float *r,
float *g,
float *b,
ui32 repeat)
83 for (
ui32 i = (repeat + 3) >> 2; i > 0; --i)
85 __m128 my = _mm_load_ps(y);
86 __m128 mcr = _mm_load_ps(cr);
87 __m128 mcb = _mm_load_ps(cb);
88 __m128 mg = _mm_sub_ps(my, _mm_mul_ps(gamma_cr2g, mcr));
89 _mm_store_ps(g, _mm_sub_ps(mg, _mm_mul_ps(gamma_cb2g, mcb)));
90 _mm_store_ps(r, _mm_add_ps(my, _mm_mul_ps(gamma_cr2r, mcr)));
91 _mm_store_ps(b, _mm_add_ps(my, _mm_mul_ps(gamma_cb2b, mcb)));
93 y += 4; cb += 4; cr += 4;
94 r += 4; g += 4; b += 4;
void sse_ict_forward(const float *r, const float *g, const float *b, float *y, float *cb, float *cr, ui32 repeat)
void sse_ict_backward(const float *y, const float *cb, const float *cr, float *r, float *g, float *b, ui32 repeat)