61         __m128 va0 = _mm_loadu_ps(a0);  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()  local
68         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
69         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
71         va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
76         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
77         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
79         va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
84         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
85         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
87         va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
92         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
93         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c3));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
105           const __m128 va0 = _mm_load1_ps(a0);  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()  local
108           vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()
109           vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));  in xnn_f32_igemm_minmax_ukernel_1x8s4__sse()