Home
last modified time | relevance | path

Searched refs:vgetq_lane_u64 (Results 1 – 25 of 137) sorted by relevance

123456

/external/XNNPACK/src/f32-sigmoid/gen/
Dneonfma-rr1-lut64-p2-div-x24.c73 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
74 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
77 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
78 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
81 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
82 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
85 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
86 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
89 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
90 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
[all …]
Dneonfma-rr1-lut2048-p1-div-x24.c72 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
73 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
76 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
77 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
80 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
81 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
84 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
85 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
88 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
89 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
[all …]
Dneonfma-rr1-lut64-p2-div-x20.c68 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
69 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
72 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
73 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
76 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
77 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
80 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
81 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
84 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
85 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
[all …]
Dneonfma-rr1-lut2048-p1-div-x20.c67 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
68 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
71 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
72 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
75 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
76 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
79 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
80 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
83 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
84 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
[all …]
Dneonfma-rr1-lut2048-p1-div-x16.c62 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
63 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
66 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
67 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
70 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
71 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
74 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
75 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
156 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
157 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
[all …]
Dneonfma-rr1-lut64-p2-div-x16.c63 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
64 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
67 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
68 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
71 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
72 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
75 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
76 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
162 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
163 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
[all …]
Dneonfma-rr1-lut2048-p1-nr2recps-x20.c67 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
68 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
71 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
72 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
75 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
76 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
79 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
80 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
83 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
84 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
[all …]
Dneonfma-rr1-lut2048-p1-nr2recps-x24.c72 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
73 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
76 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
77 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
80 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
81 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
84 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
85 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
88 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
89 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
[all …]
Dneonfma-rr1-lut64-p2-nr2recps-x24.c73 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
74 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
77 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
78 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
81 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
82 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
85 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
86 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
89 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
90 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
[all …]
Dneonfma-rr1-lut64-p2-nr2fma-x24.c73 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
74 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
77 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
78 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
81 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
82 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
85 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
86 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
89 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
90 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
[all …]
Dneonfma-rr1-lut2048-p1-nr1recps1fma-x24.c72 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
73 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
76 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
77 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
80 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
81 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
84 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
85 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
88 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
89 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
[all …]
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dneon-lut64-p2-x20.c90 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
91 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
93 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
94 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
96 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
97 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
99 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
100 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
102 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
103 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
[all …]
Dneonfma-lut64-p2-x20.c89 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
90 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
92 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
93 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
95 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
96 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
98 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
99 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
101 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
102 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
[all …]
Dneonfma-lut64-p2-x20-acc5.c93 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
94 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
96 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
97 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
99 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
100 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
102 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
103 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
105 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
106 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
[all …]
Dneon-lut64-p2-x20-acc2.c91 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
92 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
94 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
95 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
97 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
98 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
100 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
101 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
103 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
104 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
[all …]
Dneon-lut64-p2-x20-acc5.c94 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
95 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
97 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
98 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
100 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
101 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
103 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
104 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
106 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
107 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
[all …]
Dneonfma-lut64-p2-x20-acc2.c90 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
91 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
93 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
94 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
96 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
97 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
99 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
100 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
102 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
103 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
[all …]
Dneonfma-lut64-p2-x16.c85 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
86 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
88 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
89 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
91 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
92 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
94 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
95 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
215 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
216 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
[all …]
Dneon-lut64-p2-x16-acc2.c87 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
88 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
90 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
91 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
93 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
94 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
96 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
97 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
219 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
220 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
[all …]
Dneonfma-lut64-p2-x16-acc4.c88 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
89 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
91 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
92 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
94 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
95 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
97 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
98 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
222 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
223 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
[all …]
Dneon-lut64-p2-x16-acc4.c89 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
90 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
92 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
93 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
95 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
96 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
98 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
99 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
223 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
224 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
[all …]
Dneon-lut64-p2-x16.c86 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
87 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
89 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
90 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
92 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
93 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
95 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
96 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
216 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
217 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
[all …]
Dneonfma-lut64-p2-x16-acc2.c86 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
87 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
89 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
90 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
92 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
93 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
95 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
96 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
218 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
219 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
[all …]
/external/XNNPACK/src/f32-velu/gen/
Dvelu-neon-rr2-lut16-p3-x24.c80 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
81 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
87 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
88 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
94 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
95 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
101 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
102 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
108 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
109 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
[all …]
Dvelu-neonfma-rr1-lut16-p3-x24.c79 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
80 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
86 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
87 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
93 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
94 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
100 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
101 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
107 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
108 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
[all …]

123456