| /external/XNNPACK/src/qu8-vaddc/gen/ |
| D | minmax-neon-ld64-x8.c | 48 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8() local
50 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
52 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
54 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
70 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8() local
71 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
72 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
75 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
76 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
79 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2; in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
[all …]
|
| /external/XNNPACK/src/qs8-vaddc/gen/ |
| D | minmax-neon-ld64-x8.c | 48 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8() local
50 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
52 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
54 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
70 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8() local
71 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
72 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
75 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
76 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
79 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2; in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
[all …]
|
| D | minmax-neon-ld64-x16.c | 77 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16() local
78 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
79 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
82 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
86 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
87 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
90 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2; in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
91 vout01234567 = vext_s8(vout01234567, vout01234567, 2); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
94 vst1_lane_s8(output, vout01234567, 0); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16()
|
| /external/XNNPACK/src/qu8-vmulc/gen/ |
| D | minmax-fp32-neonv8-ld64-x8.c | 60 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
62 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
65 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
67 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
69 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
97 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
99 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
100 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
102 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
103 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
[all …]
|
| D | minmax-fp32-neon-ld64-x8.c | 62 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
64 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
67 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
69 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
71 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
101 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
103 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
104 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
106 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
107 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
[all …]
|
| D | minmax-rndnu-neon-ld64-x8.c | 61 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local
63 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local
66 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
68 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
70 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
98 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local
100 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
101 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
103 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
104 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
[all …]
|
| /external/XNNPACK/src/qs8-vmulc/gen/ |
| D | minmax-fp32-neonv8-ld64-x8.c | 60 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
62 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
65 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
67 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
69 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
97 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
99 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
100 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
102 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
103 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
[all …]
|
| D | minmax-fp32-neon-ld64-x8.c | 62 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
64 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
67 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
69 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
71 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
101 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
103 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
104 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
106 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
107 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
[all …]
|
| D | minmax-rndnu-neon-ld64-x8.c | 61 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local
63 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local
66 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
68 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
70 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
98 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local
100 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
101 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
103 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
104 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8()
[all …]
|
| /external/XNNPACK/src/qs8-vmul/gen/ |
| D | minmax-fp32-neonv8-ld64-x8.c | 60 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
62 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
65 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
67 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
69 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
99 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
101 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
102 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
104 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
105 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
[all …]
|
| D | minmax-rndnu-neon-ld64-x8.c | 61 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local
63 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local
66 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
68 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
70 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
100 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local
102 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
103 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
105 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
106 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
[all …]
|
| D | minmax-fp32-neon-ld64-x8.c | 62 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
64 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
67 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
69 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
71 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
103 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
105 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
106 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
108 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
109 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
[all …]
|
| /external/XNNPACK/src/qu8-vmul/gen/ |
| D | minmax-fp32-neonv8-ld64-x8.c | 60 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
62 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
65 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
67 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
69 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
99 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
101 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
102 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
104 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
105 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
[all …]
|
| D | minmax-rndnu-neon-ld64-x8.c | 61 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local
63 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local
66 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
68 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
70 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
100 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local
102 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
103 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
105 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
106 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8()
[all …]
|
| D | minmax-fp32-neon-ld64-x8.c | 62 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
64 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
67 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
69 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
71 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
103 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
105 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
106 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
108 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
109 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
[all …]
|
| /external/XNNPACK/src/qu8-vadd/gen/ |
| D | minmax-neon-ld64-x8.c | 51 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8() local
53 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
55 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
57 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
78 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8() local
79 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
80 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
83 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
84 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
87 vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2; in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
[all …]
|
| /external/XNNPACK/src/qs8-vadd/gen/ |
| D | minmax-neon-ld64-x8.c | 51 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8() local
53 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8()
55 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8()
57 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8()
78 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8() local
79 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8()
80 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8()
83 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8()
84 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8()
87 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2; in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8()
[all …]
|
| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7x-minmax-fp32-neonv8-c8.c | 100 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8() local
102 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8() local
105 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
107 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
109 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
147 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8() local
148 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
149 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
152 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
153 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
[all …]
|
| D | 7x-minmax-rndnu-neon-c8.c | 101 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
103 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
106 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
108 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
110 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
148 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
149 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
150 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
153 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
154 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
[all …]
|
| D | 7x-minmax-fp32-neon-c8.c | 102 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8() local
104 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8() local
107 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
109 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
111 vst1_s8(output, vout01234567); output += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
151 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8() local
152 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
153 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
156 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
157 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
[all …]
|
| /external/XNNPACK/src/qc8-dwconv/gen/ |
| D | up8x3-minmax-fp32-neonv8-mla8-ld64.c | 97 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64() local
103 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64() local
106 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64()
108 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64()
110 vst1_s8(output, vout01234567); output += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64()
155 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64() local
156 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64()
157 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64()
160 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64()
161 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neonv8_mla8_ld64()
[all …]
|
| D | up8x3-minmax-fp32-neon-mla8-ld64.c | 99 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64() local
104 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64() local
107 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64()
109 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64()
111 vst1_s8(output, vout01234567); output += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64()
158 int8x8_t vout01234567 = vqmovn_s16(vacc01234567); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64() local
159 vout01234567 = vmax_s8(vout01234567, voutput_min); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64()
160 vout01234567 = vmin_s8(vout01234567, voutput_max); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64()
163 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64()
164 vout01234567 = vext_s8(vout01234567, vout01234567, 4); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__neon_mla8_ld64()
[all …]
|
| /external/XNNPACK/src/qu8-gavgpool/gen/ |
| D | 7x-minmax-rndnu-neon-c8.c | 101 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
103 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
106 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
108 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
110 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
148 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
149 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
150 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
153 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
154 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
[all …]
|
| D | 7x-minmax-fp32-neonv8-c8.c | 100 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8() local
102 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8() local
105 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
107 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
109 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
147 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8() local
148 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
149 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
152 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
153 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
[all …]
|
| D | 7x-minmax-fp32-neon-c8.c | 102 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8() local
104 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8() local
107 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
109 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
111 vst1_u8(output, vout01234567); output += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
151 uint8x8_t vout01234567 = vqmovun_s16(vacc01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8() local
152 vout01234567 = vmax_u8(vout01234567, voutput_min); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
153 vout01234567 = vmin_u8(vout01234567, voutput_max); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
156 vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
157 vout01234567 = vext_u8(vout01234567, vout01234567, 4); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c8()
[all …]
|