| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 4x16c2-minmax-neon-mlal-padal-dup.c | 125 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
126 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
127 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
128 …int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
130 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
131 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
132 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
133 …vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
138 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
139 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
[all …]
|
| D | 3x16c2-minmax-neon-mlal-padal-dup.c | 111 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
112 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
113 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
115 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
116 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
117 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
121 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
122 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
123 …int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
125 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
[all …]
|
| D | 4x8c2-minmax-neon-mlal-padal-dup.c | 109 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
110 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
111 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
112 …int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
114 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
115 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
116 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
117 …vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
122 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
123 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
[all …]
|
| D | 3x8c2-minmax-neon-mlal-padal-dup.c | 97 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
98 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
99 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
101 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
102 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
103 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
107 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
108 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
109 …int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
111 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
[all …]
|
| D | 2x16c2-minmax-neon-mlal-padal-dup.c | 97 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
98 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
100 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
101 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
104 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
105 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
107 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
108 …vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
111 …int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
112 …int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
[all …]
|
| D | 4x16c2-minmax-neon-mull-padal-dup.c | 122 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
123 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
124 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
125 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
130 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
131 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
132 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
133 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
138 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
139 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
[all …]
|
| D | 2x8c2-minmax-neon-mlal-padal-dup.c | 85 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
86 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
88 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
89 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
92 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
93 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
95 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
96 …vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
99 …int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
100 …int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
[all …]
|
| D | 3x16c2-minmax-neon-mull-padal-dup.c | 109 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
110 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
111 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
112 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
117 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
118 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
119 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
120 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
125 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
126 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
[all …]
|
| D | 2x16c2-minmax-neon-mull-padal-dup.c | 96 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
97 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
98 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
99 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
104 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
105 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
106 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
107 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
112 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
113 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
[all …]
|
| D | 4x8c2-minmax-neon-mull-padal-dup.c | 106 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
107 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
108 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
109 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
114 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
115 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
116 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
117 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
122 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
123 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
[all …]
|
| D | 1x16c2-minmax-neon-mlal-padal-dup.c | 83 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
85 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
87 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
89 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
91 …int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
93 …vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
95 …int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
97 …vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
99 …int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
101 …vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
[all …]
|
| D | 3x8c2-minmax-neon-mull-padal-dup.c | 95 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
96 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
97 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
98 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
103 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
104 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
105 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
106 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
111 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
112 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
[all …]
|
| D | 1x8c2-minmax-neon-mlal-padal-dup.c | 73 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
75 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
77 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
79 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
81 …int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
83 …vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
85 …int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
87 …vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
89 …int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
91 …vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_1x8c2__neon_mlal_padal_dup()
[all …]
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 4x16c2-minmax-neon-mlal-padal-dup.c | 108 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
109 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
110 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
111 …int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
113 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
114 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
115 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
116 …vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
121 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
122 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
[all …]
|
| D | 3x16c2-minmax-neon-mlal-padal-dup.c | 96 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
97 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
98 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
100 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
101 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
102 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
106 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
107 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
108 …int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
110 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
[all …]
|
| D | 2x16c2-minmax-neon-mlal-padal-dup.c | 84 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
85 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
87 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
88 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
91 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
92 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
94 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
95 …vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
98 …int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
99 …int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
[all …]
|
| D | 4x8c2-minmax-neon-mlal-padal-dup.c | 92 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
93 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
94 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
95 …int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
97 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
98 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
99 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
100 …vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
105 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
106 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
[all …]
|
| D | 3x8c2-minmax-neon-mlal-padal-dup.c | 82 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
83 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
84 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
86 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
87 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
88 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
92 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
93 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
94 …int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
96 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
[all …]
|
| D | 4x16c2-minmax-neon-mull-padal-dup.c | 105 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
106 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
107 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
108 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
113 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
114 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
115 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
116 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
121 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
122 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
[all …]
|
| D | 2x8c2-minmax-neon-mlal-padal-dup.c | 72 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
73 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
75 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
76 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
79 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
80 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
82 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
83 …vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
86 …int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
87 …int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
[all …]
|
| D | 3x16c2-minmax-neon-mull-padal-dup.c | 94 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
95 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
96 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
97 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
102 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
103 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
104 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
105 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
110 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
111 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mull_padal_dup()
[all …]
|
| D | 4x8c2-minmax-neon-mull-padal-dup.c | 89 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
90 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
91 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
92 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
97 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
98 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
99 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
100 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
105 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
106 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
[all …]
|
| D | 2x16c2-minmax-neon-mull-padal-dup.c | 83 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
84 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
85 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
86 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
91 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
92 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
93 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
94 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
99 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
100 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
[all …]
|
| D | 1x16c2-minmax-neon-mlal-padal-dup.c | 72 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
74 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
76 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
78 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
80 …int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
82 …vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
84 …int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
86 …vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
88 …int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
90 …vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
[all …]
|
| D | 3x8c2-minmax-neon-mull-padal-dup.c | 80 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
81 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
82 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
83 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
88 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
89 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
90 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
91 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
96 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
97 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mull_padal_dup()
[all …]
|