| /external/XNNPACK/src/f32-conv-hwc/gen/ |
| D | 3x3s2p0p1c3x4-neon-2x1.c | 266 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
271 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
272 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
276 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
402 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
407 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
408 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
412 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
|
| D | 3x3s2p0p1c3x4-neonfma-2x1.c | 266 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
271 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
272 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
276 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
400 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
405 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
406 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
410 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
|
| D | 3x3s2p1c3x4-neonfma-2x1.c | 266 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
271 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
272 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
276 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
400 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
405 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
406 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
410 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
|
| D | 3x3s2p1c3x4-neon-2x1.c | 266 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
271 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
272 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
276 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
402 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
407 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
408 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
412 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
|
| D | 3x3s2p0p1c3x8-neonfma-2x1.c | 361 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
366 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
367 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
371 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
563 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
568 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
569 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
573 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
| D | 3x3s2p1c3x8-neon-2x1.c | 361 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
366 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
367 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
371 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
565 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
570 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
571 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
575 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
| D | 3x3s2p0p1c3x8-neon-2x1.c | 361 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
366 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
367 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
371 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
565 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
570 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
571 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
575 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|
| D | 3x3s2p1c3x8-neonfma-2x1.c | 361 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
366 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
367 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
371 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
563 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
568 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
569 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
573 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
| D | 3x3s2p1c3x4-neonfma-2x2.c | 569 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
574 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
575 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
579 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
702 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
707 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
708 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
712 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
|
| D | 3x3s2p1c3x4-neon-2x2.c | 571 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
576 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
577 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
581 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
706 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
711 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
712 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
716 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
|
| D | 3x3s2p0p1c3x4-neon-2x2.c | 571 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
576 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
577 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
581 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
706 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
711 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
712 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
716 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
|
| D | 3x3s2p0p1c3x4-neonfma-2x2.c | 569 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
574 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
575 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
579 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
702 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
707 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
708 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
712 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
|
| D | 3x3s2p1c3x8-neonfma-2x2.c | 826 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
831 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
832 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
836 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
1027 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
1032 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
1033 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
1037 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
| D | 3x3s2p1c3x8-neon-2x2.c | 828 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
833 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
834 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
838 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1031 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
1036 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1037 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1041 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
| D | 3x3s2p0p1c3x8-neonfma-2x2.c | 826 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
831 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
832 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
836 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
1027 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
1032 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
1033 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
1037 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
| D | 3x3s2p0p1c3x8-neon-2x2.c | 828 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
833 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
834 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
838 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
1031 float32x2_t vo0c01 = vget_low_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
1036 vst1_f32(o0_tmp, vo0c01); o0_tmp += 2; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
1037 vo0c01 = vget_high_f32(vo0c0123); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
1041 vst1_lane_f32(o0_tmp, vo0c01, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
| /external/XNNPACK/src/f32-conv-hwc2chw/ |
| D | 3x3s2p1c3x4-wasmsimd-2x2.c | 336 const v128_t vo0c01 = wasm_v32x4_shuffle(vo0x0, vo0x1, 0, 4 + 0, 1, 4 + 1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
346 *((double*) o0c0) = wasm_f64x2_extract_lane(vo0c01, 0); o0c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
347 …*((double*) o0c1) = wasm_f64x2_extract_lane(wasm_v32x4_shuffle(vo0c01, vo0c01, 2, 3, 2, 3), 0); o0… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
596 const v128_t vo0c01 = wasm_v32x4_shuffle(vo0x0, vo0x1, 0, 4 + 0, 1, 4 + 1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
606 *((double*) o0c0) = wasm_f64x2_extract_lane(vo0c01, 0); o0c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
607 …*((double*) o0c1) = wasm_f64x2_extract_lane(wasm_v32x4_shuffle(vo0c01, vo0c01, 2, 3, 2, 3), 0); o0… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
|
| D | 3x3s2p1c3x4-neonfma-2x2.c | 335 const float32x4_t vo0c01 = vzip1q_f32(vo0x0, vo0x1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
346 vst1_f32(o0c0, vget_low_f32(vo0c01)); o0c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
347 vst1_f32(o0c1, vget_high_f32(vo0c01)); o0c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
596 const float32x4_t vo0c01 = vzip1q_f32(vo0x0, vo0x1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
606 vst1_f32(o0c0, vget_low_f32(vo0c01)); o0c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
607 vst1_f32(o0c1, vget_high_f32(vo0c01)); o0c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
|
| D | 3x3s2p1c3x4-sse-2x2.c | 335 const __m128 vo0c01 = _mm_unpacklo_ps(vo0x0, vo0x1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2() local
346 _mm_storel_pi((__m64 *)o0c0, vo0c01); o0c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
347 … _mm_storel_pi((__m64 *)o0c1, _mm_shuffle_ps(vo0c01, vo0c01, _MM_SHUFFLE(3, 2, 3, 2))); o0c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
596 const __m128 vo0c01 = _mm_unpacklo_ps(vo0x0, vo0x1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2() local
606 _mm_storel_pi((__m64 *)o0c0, vo0c01); o0c0 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
607 … _mm_storel_pi((__m64 *)o0c1, _mm_shuffle_ps(vo0c01, vo0c01, _MM_SHUFFLE(3, 2, 3, 2))); o0c1 += 2; in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
|