Home
last modified time | relevance | path

Searched refs:vk21c1x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x4-neonfma-2x2.c224 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
226 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
227 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
229 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
230 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
476 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
478 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
479 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
667 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
669 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x4-neon-2x2.c222 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
224 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
225 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
227 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
228 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
476 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
478 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
479 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
669 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
671 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x2.c222 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
224 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
225 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
227 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
228 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
476 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
478 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
479 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
669 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
671 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neonfma-2x2.c224 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
226 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
227 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
229 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
230 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
476 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
478 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
479 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
667 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
669 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x1.c171 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
173 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
174 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
365 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
367 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
368 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vi4x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c173 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
175 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
176 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
365 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
367 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
368 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vi4x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neonfma-2x1.c173 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
175 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
176 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
365 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
367 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
368 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c171 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
173 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
174 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
365 const float32x4_t vk21c1x0123 = vld1q_f32(w + 60); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
367 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
368 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p1c3x8-neonfma-2x2.c298 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
301 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
302 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
306 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
307 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
682 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
685 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
686 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
968 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
971 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c296 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
299 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
300 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
304 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
305 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
682 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
685 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
686 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
970 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
973 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c298 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
301 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
302 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
306 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
307 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
682 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
685 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
686 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
968 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
971 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c296 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
299 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
300 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
304 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
305 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
682 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
685 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
686 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
970 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
973 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c217 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
220 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
221 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
504 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
507 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
508 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vi4x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c215 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
218 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
219 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
504 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
507 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
508 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c215 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
218 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
219 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
504 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
507 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
508 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vi4x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c217 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
220 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
221 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
504 const float32x4_t vk21c1x0123 = vld1q_f32(w + 120); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
507 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
508 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c137 const __m128 vk21c1x0123 = _mm_load_ps(w + 60); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
139 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk21c1x0123, vi21c1)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
264 const __m128 vk21c1x0123 = _mm_load_ps(w + 60); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
266 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk21c1x0123, vi21c1)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()