| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7p7x-minmax-fp32-neon-c8.c | 48 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
51 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
53 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
55 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
57 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
58 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
60 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
61 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
83 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
86 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
[all …]
|
| D | 7p7x-minmax-rndnu-neon-c8.c | 48 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
51 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
53 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
55 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
57 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
58 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
60 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
61 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
83 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
86 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
[all …]
|
| D | 7p7x-minmax-fp32-neonv8-c8.c | 49 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
52 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
54 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
56 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
58 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
59 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
61 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
62 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
84 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
87 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
[all …]
|
| D | 7p7x-minmax-rndnu-neon-c24.c | 52 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
59 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
65 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
71 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
77 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
82 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
86 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
87 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
105 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
108 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
[all …]
|
| D | 7p7x-minmax-fp32-neon-c24.c | 52 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local
59 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
65 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
71 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
77 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
82 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
86 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
87 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
105 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local
108 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
[all …]
|
| D | 7p7x-minmax-fp32-neonv8-c24.c | 53 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
60 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
66 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
72 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
78 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
83 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
87 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
88 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
106 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
109 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
[all …]
|
| D | 7p7x-minmax-rndnu-neon-c16.c | 50 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
55 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
59 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
63 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
67 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
70 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
73 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
74 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
102 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
107 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
[all …]
|
| D | 7p7x-minmax-fp32-neon-c16.c | 50 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
55 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
59 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
63 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
67 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
70 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
73 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
74 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
102 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
107 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
[all …]
|
| D | 7p7x-minmax-fp32-neonv8-c16.c | 51 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
56 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
60 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
64 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
68 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
71 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
74 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
75 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
103 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
108 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
[all …]
|
| D | 7x-minmax-fp32-neonv8-c8.c | 67 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8() local
70 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
72 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
74 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
76 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
77 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
79 int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
80 int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
116 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8() local
119 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c8()
[all …]
|
| D | 7p7x-minmax-rndnu-neon-c32.c | 54 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local
63 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
71 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
79 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
87 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
94 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
99 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
100 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
122 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local
125 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
[all …]
|
| D | 7p7x-minmax-fp32-neonv8-c32.c | 55 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local
64 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
72 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
80 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
88 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
95 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
100 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
101 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
123 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local
126 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
[all …]
|
| D | 7p7x-minmax-fp32-neon-c32.c | 54 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local
63 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
71 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
79 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
87 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
94 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
99 const int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
100 const int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
122 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local
125 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
[all …]
|
| D | 7x-minmax-rndnu-neon-c8.c | 68 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
71 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
73 vsum01234567 = vaddw_s8(vsum01234567, vi3x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
75 vsum01234567 = vaddw_s8(vsum01234567, vi4x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
77 vsum01234567 = vaddw_s8(vsum01234567, vi5x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
78 vsum01234567 = vaddw_s8(vsum01234567, vi6x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
80 int32x4_t vacc0123 = vaddw_s16(vinit_bias, vget_low_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
81 int32x4_t vacc4567 = vaddw_s16(vinit_bias, vget_high_s16(vsum01234567)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
117 int16x8_t vsum01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
120 vsum01234567 = vaddw_s8(vsum01234567, vi2x01234567); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
[all …]
|
| /external/XNNPACK/src/qu8-gavgpool/gen/ |
| D | 7p7x-minmax-rndnu-neon-c8.c | 48 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
51 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
53 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
55 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
57 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
58 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
60 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
61 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
83 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
86 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8()
[all …]
|
| D | 7p7x-minmax-fp32-neonv8-c8.c | 49 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
52 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
54 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
56 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
58 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
59 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
61 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
62 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
84 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
87 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8()
[all …]
|
| D | 7p7x-minmax-fp32-neon-c8.c | 48 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
51 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
53 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
55 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
57 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
58 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
60 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
61 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
83 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
86 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8()
[all …]
|
| D | 7p7x-minmax-rndnu-neon-c24.c | 52 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
59 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
65 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
71 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
77 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
82 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
86 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
87 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
105 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
108 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
[all …]
|
| D | 7p7x-minmax-fp32-neon-c24.c | 52 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local
59 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
65 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
71 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
77 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
82 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
86 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
87 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
105 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local
108 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
[all …]
|
| D | 7p7x-minmax-fp32-neonv8-c24.c | 53 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
60 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
66 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
72 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
78 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
83 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
87 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
88 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
106 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
109 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
[all …]
|
| D | 7p7x-minmax-fp32-neon-c16.c | 50 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
55 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
59 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
63 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
67 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
70 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
73 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
74 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
102 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
107 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
[all …]
|
| D | 7p7x-minmax-rndnu-neon-c16.c | 50 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
55 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
59 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
63 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
67 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
70 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
73 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
74 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
102 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
107 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
[all …]
|
| D | 7p7x-minmax-fp32-neonv8-c16.c | 51 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
56 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
60 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
64 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
68 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
71 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
74 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
75 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
103 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
108 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
[all …]
|
| D | 7p7x-minmax-fp32-neonv8-c32.c | 55 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local
64 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
72 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
80 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
88 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
95 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
100 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
101 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
123 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local
126 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
[all …]
|
| D | 7x-minmax-rndnu-neon-c8.c | 68 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
71 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
73 vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
75 vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
77 vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
78 vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
80 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
81 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
117 uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8() local
120 vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8()
[all …]
|