| /external/webrtc/common_audio/third_party/ooura/fft_size_128/ |
| D | ooura_fft_neon.cc | 35 float32x4_t a01v = vcombine_f32(vget_low_f32(a00v), vget_low_f32(a08v)); in cft1st_128_neon()
36 float32x4_t a23v = vcombine_f32(vget_high_f32(a00v), vget_high_f32(a08v)); in cft1st_128_neon()
37 float32x4_t a45v = vcombine_f32(vget_low_f32(a04v), vget_low_f32(a12v)); in cft1st_128_neon()
38 float32x4_t a67v = vcombine_f32(vget_high_f32(a04v), vget_high_f32(a12v)); in cft1st_128_neon()
64 a00v = vcombine_f32(vget_low_f32(a01v), vget_low_f32(a23v)); in cft1st_128_neon()
65 a04v = vcombine_f32(vget_low_f32(a45v), vget_low_f32(a67v)); in cft1st_128_neon()
66 a08v = vcombine_f32(vget_high_f32(a01v), vget_high_f32(a23v)); in cft1st_128_neon()
67 a12v = vcombine_f32(vget_high_f32(a45v), vget_high_f32(a67v)); in cft1st_128_neon()
86 const float32x4_t a_00_32 = vcombine_f32(a_00, a_32); in cftmdl_128_neon()
87 const float32x4_t a_08_40 = vcombine_f32(a_08, a_40); in cftmdl_128_neon()
[all …]
|
| /external/XNNPACK/src/f32-ibilinear-chw/gen/ |
| D | neonfma-p8.c | 79 const float32x4_t vtltr01 = vcombine_f32(vtltr0, vtltr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
80 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
81 const float32x4_t vtltr23 = vcombine_f32(vtltr2, vtltr3); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
82 const float32x4_t vblbr23 = vcombine_f32(vblbr2, vblbr3); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
83 const float32x4_t vtltr45 = vcombine_f32(vtltr4, vtltr5); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
84 const float32x4_t vblbr45 = vcombine_f32(vblbr4, vblbr5); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
85 const float32x4_t vtltr67 = vcombine_f32(vtltr6, vtltr7); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
86 const float32x4_t vblbr67 = vcombine_f32(vblbr6, vblbr7); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
149 const float32x4_t vtltr01 = vcombine_f32(vtltr0, vtltr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
150 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
[all …]
|
| D | neon-p8.c | 79 const float32x4_t vtltr01 = vcombine_f32(vtltr0, vtltr1); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
80 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
81 const float32x4_t vtltr23 = vcombine_f32(vtltr2, vtltr3); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
82 const float32x4_t vblbr23 = vcombine_f32(vblbr2, vblbr3); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
83 const float32x4_t vtltr45 = vcombine_f32(vtltr4, vtltr5); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
84 const float32x4_t vblbr45 = vcombine_f32(vblbr4, vblbr5); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
85 const float32x4_t vtltr67 = vcombine_f32(vtltr6, vtltr7); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
86 const float32x4_t vblbr67 = vcombine_f32(vblbr6, vblbr7); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
149 const float32x4_t vtltr01 = vcombine_f32(vtltr0, vtltr1); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
150 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
[all …]
|
| D | neon-p4.c | 61 const float32x4_t vtltr01 = vcombine_f32(vtltr0, vtltr1); in xnn_f32_ibilinear_chw_ukernel__neon_p4()
62 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neon_p4()
63 const float32x4_t vtltr23 = vcombine_f32(vtltr2, vtltr3); in xnn_f32_ibilinear_chw_ukernel__neon_p4()
64 const float32x4_t vblbr23 = vcombine_f32(vblbr2, vblbr3); in xnn_f32_ibilinear_chw_ukernel__neon_p4()
|
| D | neonfma-p4.c | 61 const float32x4_t vtltr01 = vcombine_f32(vtltr0, vtltr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p4()
62 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p4()
63 const float32x4_t vtltr23 = vcombine_f32(vtltr2, vtltr3); in xnn_f32_ibilinear_chw_ukernel__neonfma_p4()
64 const float32x4_t vblbr23 = vcombine_f32(vblbr2, vblbr3); in xnn_f32_ibilinear_chw_ukernel__neonfma_p4()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | neonfma-rr1-lut64-p2-div-x24.c | 100 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
103 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
106 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
109 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
115 const float32x4_t vlKLMN = vcombine_f32(vlKL, vlMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
216 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
250 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
|
| D | neonfma-rr1-lut2048-p1-div-x24.c | 99 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
102 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
105 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
108 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
114 const float32x4_t vlKLMN = vcombine_f32(vlKL, vlMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
208 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
241 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
|
| D | neonfma-rr1-lut64-p2-div-x20.c | 91 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
94 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
97 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
100 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
192 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
226 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
|
| D | neonfma-rr1-lut2048-p1-div-x20.c | 90 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
93 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
96 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
99 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
185 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
218 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
|
| D | neonfma-rr1-lut2048-p1-div-x16.c | 81 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
84 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
87 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
90 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
162 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
195 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
|
| D | neonfma-rr1-lut64-p2-div-x16.c | 82 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
85 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
88 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
91 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
168 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
202 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
|
| D | neonfma-rr1-lut2048-p1-nr2recps-x20.c | 90 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
93 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
96 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
99 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
203 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
240 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
|
| D | neonfma-rr1-lut2048-p1-nr2recps-x24.c | 99 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
102 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
105 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
108 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
114 const float32x4_t vlKLMN = vcombine_f32(vlKL, vlMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
229 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
266 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | neon-lut64-p2-x20.c | 118 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
121 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
124 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
127 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
244 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
320 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20()
|
| D | neonfma-lut64-p2-x20.c | 117 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
120 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
123 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
126 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
243 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
319 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20()
|
| D | neonfma-lut64-p2-x20-acc5.c | 121 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
124 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
127 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
130 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
252 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
328 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5()
|
| D | neon-lut64-p2-x20-acc2.c | 119 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
122 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
125 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
128 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
247 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
323 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2()
|
| D | neon-lut64-p2-x20-acc5.c | 122 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
125 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
128 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
131 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
253 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
329 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5()
|
| D | neonfma-lut64-p2-x20-acc2.c | 118 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
121 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
124 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
127 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); 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()
246 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
322 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2()
|
| D | neonfma-lut64-p2-x16.c | 108 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
111 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
114 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
117 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
221 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
297 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
|
| D | neon-lut64-p2-x16-acc2.c | 110 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
113 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
116 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
119 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
225 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
301 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
|
| D | neonfma-lut64-p2-x16-acc4.c | 111 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
114 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
117 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
120 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
228 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
304 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
|
| D | neon-lut64-p2-x16-acc4.c | 112 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
115 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
118 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
121 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
229 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
305 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
|
| D | neon-lut64-p2-x16.c | 109 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
112 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
115 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
118 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
222 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
298 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
|
| D | neonfma-lut64-p2-x16-acc2.c | 109 const float32x4_t vl0123 = vcombine_f32(vl01, vl23); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
112 const float32x4_t vl4567 = vcombine_f32(vl45, vl67); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
115 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
118 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
224 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
300 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
|