Lines Matching refs:xsum1
46 __m128 xsum1, xsum2; in xcorr_kernel_sse() local
47 xsum1 = _mm_loadu_ps(sum); in xcorr_kernel_sse()
56 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0x00),yj)); in xcorr_kernel_sse()
59 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xaa), in xcorr_kernel_sse()
65 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j))); in xcorr_kernel_sse()
71 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j))); in xcorr_kernel_sse()
75 _mm_storeu_ps(sum,_mm_add_ps(xsum1,xsum2)); in xcorr_kernel_sse()
83 __m128 xsum1, xsum2; in dual_inner_prod_sse() local
84 xsum1 = _mm_setzero_ps(); in dual_inner_prod_sse()
91 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(xi, y1i)); in dual_inner_prod_sse()
95 xsum1 = _mm_add_ps(xsum1, _mm_movehl_ps(xsum1, xsum1)); in dual_inner_prod_sse()
96 xsum1 = _mm_add_ss(xsum1, _mm_shuffle_ps(xsum1, xsum1, 0x55)); in dual_inner_prod_sse()
97 _mm_store_ss(xy1, xsum1); in dual_inner_prod_sse()