Home
last modified time | relevance | path

Searched refs:vk11c1x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x4-neonfma-2x2.c216 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
218 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
219 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
221 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
222 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
471 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
473 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
474 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
662 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
664 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vi1x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
D3x3s2p1c3x4-neon-2x2.c214 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
216 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
217 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
219 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
220 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
471 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
473 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
474 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
664 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
666 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x2.c214 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
216 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
217 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
219 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
220 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
471 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
473 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
474 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
664 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
666 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vi1x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
D3x3s2p1c3x4-neonfma-2x2.c216 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
218 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
219 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
221 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
222 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
471 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
473 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
474 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
662 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
664 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
D3x3s2p0p1c3x4-neon-2x1.c166 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
168 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
169 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
360 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
362 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vi1x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
363 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vi3x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p1c3x4-neonfma-2x1.c168 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
170 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
171 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
360 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
362 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
363 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c166 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
168 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
169 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
360 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
362 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
363 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c168 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
170 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
171 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
360 const float32x4_t vk11c1x0123 = vld1q_f32(w + 56); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
362 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vi1x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
363 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vi3x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x2.c283 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
286 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
287 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
291 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
292 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
674 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
677 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
678 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
962 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
965 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vi1x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neonfma-2x2.c285 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
288 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
289 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
293 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
294 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
674 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
677 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
678 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
960 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
963 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c285 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
288 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
289 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
293 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
294 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
674 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
677 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
678 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
960 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
963 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vi1x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c283 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
286 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
287 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
291 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
292 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
674 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
677 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
678 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
962 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
965 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p1c3x8-neon-2x1.c207 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
210 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
211 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
496 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
499 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
500 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c209 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
212 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
213 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
496 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
499 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vi1x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
500 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vi3x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c209 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
212 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
213 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
496 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
499 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
500 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c207 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
210 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
211 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
496 const float32x4_t vk11c1x0123 = vld1q_f32(w + 112); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
499 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c1x0123, vi1x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
500 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vi3x1, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c133 const __m128 vk11c1x0123 = _mm_load_ps(w + 56); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
135 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c1x0123, vi11c1)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
260 const __m128 vk11c1x0123 = _mm_load_ps(w + 56); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
262 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c1x0123, vi11c1)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()