Lines Matching refs:vacc0123p0
85 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
92 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
98 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
104 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
110 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
116 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
122 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
128 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
134 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
140 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
145 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
152 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
156 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
160 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
164 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
168 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
172 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
176 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
180 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
184 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
188 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
191 __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()