| /external/XNNPACK/src/f32-conv-hwc/gen/ |
| D | 3x3s2p1c3x4-neonfma-2x2.c | 247 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
249 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
250 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
252 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_low_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
253 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_low_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
486 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
488 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
489 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
677 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
679 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neon-2x2.c | 245 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
247 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
248 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
250 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_low_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
251 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_low_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
486 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
488 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
489 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
679 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
681 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x4-neon-2x2.c | 238 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
240 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
241 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
243 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
244 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
486 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
488 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
489 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
679 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
681 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x4-neonfma-2x2.c | 240 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
242 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
243 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
245 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
246 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
486 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
488 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
489 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
677 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
679 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
|
| D | 3x3s2p0p1c3x4-neon-2x1.c | 181 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
183 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
184 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
375 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
377 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
378 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vi3x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
|
| D | 3x3s2p0p1c3x4-neonfma-2x1.c | 183 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
185 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
186 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
375 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
377 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
378 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vi3x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
|
| D | 3x3s2p1c3x4-neonfma-2x1.c | 183 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
185 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
186 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
375 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
377 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
378 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
|
| D | 3x3s2p1c3x4-neon-2x1.c | 181 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
183 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
184 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
375 const float32x4_t vk11c2x0123 = vld1q_f32(w + 68); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
377 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
378 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
|
| D | 3x3s2p1c3x8-neonfma-2x2.c | 331 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
334 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
335 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
339 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_low_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
340 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_low_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
698 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
701 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
702 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
984 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
987 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x8-neon-2x2.c | 329 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
332 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
333 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
337 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_low_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
338 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_low_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
698 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
701 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
702 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
986 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
989 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x8-neonfma-2x2.c | 324 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
327 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
328 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
332 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
333 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
698 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
701 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
702 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
984 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
987 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p0p1c3x8-neon-2x2.c | 322 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
325 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
326 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
330 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_high_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
331 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_high_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
698 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
701 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
702 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
986 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
989 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x8-neonfma-2x1.c | 233 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
236 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
237 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
520 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
523 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
524 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vi3x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
| D | 3x3s2p1c3x8-neon-2x1.c | 231 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
234 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
235 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
520 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
523 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
524 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
| D | 3x3s2p0p1c3x8-neon-2x1.c | 231 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
234 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
235 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
520 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
523 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c2x0123, vi1x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
524 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vi3x1, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|
| D | 3x3s2p1c3x8-neonfma-2x1.c | 233 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
236 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
237 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
520 const float32x4_t vk11c2x0123 = vld1q_f32(w + 136); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
523 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
524 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
| /external/XNNPACK/src/f32-conv-hwc2chw/ |
| D | 3x3s2p1c3x4-sse-1x1.c | 145 const __m128 vk11c2x0123 = _mm_load_ps(w + 68); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
147 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c2x0123, vi11c2)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
272 const __m128 vk11c2x0123 = _mm_load_ps(w + 68); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
274 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c2x0123, vi11c2)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
|