Home
last modified time | relevance | path

Searched refs:vlIJ (Results 1 – 25 of 38) sorted by relevance

12

/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dneon-lut64-p2-x20.c114 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20() local
129 vlIJ = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
130 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
Dneonfma-lut64-p2-x20.c113 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20() local
128 vlIJ = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
129 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
Dneonfma-lut64-p2-x20-acc5.c117 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5() local
132 vlIJ = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
133 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
Dneon-lut64-p2-x20-acc2.c115 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2() local
130 vlIJ = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
131 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
Dneon-lut64-p2-x20-acc5.c118 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5() local
133 vlIJ = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
134 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
Dneonfma-lut64-p2-x20-acc2.c114 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2() local
129 vlIJ = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
130 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
/external/XNNPACK/src/f32-sigmoid/gen/
Dneonfma-rr1-lut64-p2-div-x20.c87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20() local
102 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
103 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
Dneonfma-rr1-lut2048-p1-div-x20.c86 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() local
101 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
102 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
Dneonfma-rr1-lut2048-p1-nr2recps-x20.c86 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20() local
101 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
102 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
Dneonfma-rr1-lut64-p2-nr2fma-x20.c87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local
102 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
103 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
Dneonfma-rr1-lut64-p2-nr2recps-x20.c87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20() local
102 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
103 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
Dneonfma-rr1-lut64-p2-div-x24.c92 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24() local
111 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
112 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
Dneon-rr2-lut64-p2-nr2recps-x20.c88 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20() local
103 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20()
104 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20()
Dneonfma-rr1-lut2048-p1-div-x24.c91 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24() local
110 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
111 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
Dneonfma-rr1-lut64-p2-nr1recps1fma-x20.c87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() local
102 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
103 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
Dneon-rr2-lut2048-p1-nr2recps-x20.c87 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20() local
102 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
103 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
Dneonfma-rr1-lut2048-p1-nr1recps1fma-x20.c86 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() local
101 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
102 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
Dneonfma-rr1-lut2048-p1-nr2fma-x20.c86 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() local
101 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
102 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
Dneonfma-rr1-lut2048-p1-nr2recps-x24.c91 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24() local
110 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
111 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
Dneonfma-rr1-lut64-p2-nr2recps-x24.c92 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24() local
111 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
112 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
Dneonfma-rr1-lut64-p2-nr2fma-x24.c92 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() local
111 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
112 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
/external/XNNPACK/src/f32-velu/gen/
Dvelu-neonfma-rr1-lut16-p3-x20.c105 …int32x2_t vlIJ = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x20() local
107 …vlIJ = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x20()
108 const int32x4_t vlGHIJ = vcombine_s32(vlGH, vlIJ); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x20()
Dvelu-neon-rr2-lut16-p3-x20.c106 …int32x2_t vlIJ = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x20() local
108 …vlIJ = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x20()
109 const int32x4_t vlGHIJ = vcombine_s32(vlGH, vlIJ); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x20()
Dvelu-neon-rr2-lut16-p3-x24.c111 …int32x2_t vlIJ = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() local
113 …vlIJ = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
114 const int32x4_t vlGHIJ = vcombine_s32(vlGH, vlIJ); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24()
Dvelu-neonfma-rr1-lut16-p3-x24.c110 …int32x2_t vlIJ = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() local
112 …vlIJ = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()
113 const int32x4_t vlGHIJ = vcombine_s32(vlGH, vlIJ); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24()

12