| /external/XNNPACK/src/f16-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-neonfp16arith-1x8-acc2.c | 61 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2() local
76 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2()
81 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2()
91 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2()
96 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2()
106 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2()
124 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2()
131 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2()
136 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2()
148 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2()
|
| D | 3x3p1-minmax-neonfp16arith-1x8-acc3.c | 61 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3() local
76 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3()
81 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3()
91 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3()
96 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3()
106 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3()
125 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3()
132 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3()
137 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3()
149 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3()
|
| D | 3x3p1-minmax-neonfp16arith-1x8.c | 61 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8() local
76 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8()
81 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8()
91 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8()
96 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8()
106 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8()
123 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8()
130 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8()
135 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8()
147 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8()
|
| D | 3x3p1-minmax-neonfp16arith-1x8-acc4.c | 61 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4() local
76 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4()
81 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4()
91 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4()
96 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4()
106 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4()
126 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4()
133 float16x8_t vo0p2 = vmulq_lane_f16(vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4()
138 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4()
150 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vzero, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4()
|
| D | 3x3p1-minmax-neonfp16arith-2x8.c | 68 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8() local
86 vo1p0 = vfmaq_lane_f16(vo1p0, vi2x89ABCDEF, vget_high_f16(vw01234567), 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8()
88 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8()
94 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8()
108 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8()
114 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8()
128 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8()
150 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8()
158 vo1p0 = vfmaq_lane_f16(vo1p0, vi2x89ABCDEF, vget_high_f16(vw01234567), 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8()
160 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8()
[all …]
|
| D | 3x3p1-minmax-neonfp16arith-2x8-acc2.c | 68 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2() local
86 float16x8_t vo1p1 = vmulq_lane_f16(vi2x89ABCDEF, vget_high_f16(vw01234567), 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2()
88 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2()
94 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2()
108 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2()
114 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2()
128 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2()
152 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2()
160 float16x8_t vo1p1 = vmulq_lane_f16(vi2x89ABCDEF, vget_high_f16(vw01234567), 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2()
162 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2()
[all …]
|
| D | 3x3p1-minmax-neonfp16arith-3x8.c | 75 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8() local
94 vo2p0 = vfmaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8()
97 vo1p0 = vfmaq_lane_f16(vo1p0, vi2x89ABCDEF, vget_high_f16(vw01234567), 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8()
100 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8()
107 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8()
125 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8()
132 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8()
150 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8()
177 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8()
184 vo2p0 = vfmaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8()
[all …]
|
| D | 3x3p1-minmax-neonfp16arith-4x8.c | 82 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8() local
104 vo2p0 = vfmaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8()
108 vo1p0 = vfmaq_lane_f16(vo1p0, vi2x89ABCDEF, vget_high_f16(vw01234567), 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8()
112 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8()
120 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8()
142 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8()
150 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8()
172 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8()
204 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8()
212 vo2p0 = vfmaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8()
[all …]
|
| D | 3x3p1-minmax-neonfp16arith-5x8.c | 89 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8() local
114 vo2p0 = vfmaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8()
119 vo1p0 = vfmaq_lane_f16(vo1p0, vi2x89ABCDEF, vget_high_f16(vw01234567), 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8()
124 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8()
133 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8()
159 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8()
168 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8()
194 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8()
231 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8()
240 vo2p0 = vfmaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8()
[all …]
|
| D | 3x3p1-minmax-neonfp16arith-6x8.c | 96 float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8() local
124 vo2p0 = vfmaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
130 vo1p0 = vfmaq_lane_f16(vo1p0, vi2x89ABCDEF, vget_high_f16(vw01234567), 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
136 vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
146 const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
176 vi2x01234567 = vi2x89ABCDEF; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
186 const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
216 vi2x89ABCDEF = vi2xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
258 vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF))); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
268 vo2p0 = vfmaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
[all …]
|
| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7p7x-minmax-fp32-sse2-c16.c | 59 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local
66 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16()
147 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local
154 … const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16()
255 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local
262 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16()
|
| D | 7p7x-minmax-rndnu-neon-c16.c | 51 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
57 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
103 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
109 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
181 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
187 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
|
| D | 7p7x-minmax-fp32-neon-c16.c | 51 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
57 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
103 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
109 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
180 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
186 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
|
| D | 7p7x-minmax-fp32-neonv8-c16.c | 52 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
58 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
104 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
110 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
180 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
186 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
|
| D | 7p7x-minmax-fp32-sse2-c24.c | 62 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local
71 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24()
231 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local
240 … const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24()
420 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local
429 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24()
|
| D | 7p7x-minmax-rndnu-neon-c24.c | 53 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
61 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
148 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
156 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
273 const int8x8_t vi2x89ABCDEF = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
281 vsum89ABCDEF = vaddw_s8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
|
| /external/XNNPACK/src/f16-gavgpool/gen/ |
| D | 7p7x-minmax-neonfp16arith-c16.c | 50 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
56 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16()
122 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
127 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16()
214 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
219 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16()
|
| D | 7p7x-minmax-neonfp16arith-c24.c | 52 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
60 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24()
140 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
147 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24()
248 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
255 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24()
|
| D | 7p7x-minmax-f16c-c16.c | 51 const __m256 vi2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
57 …vacc89ABCDEF = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc89ABCDEF), vi2x89ABCDEF), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16()
123 const __m256 vi2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
128 …vacc89ABCDEF = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc89ABCDEF), vi2x89ABCDEF), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16()
217 const __m256 vi2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
222 …vacc89ABCDEF = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc89ABCDEF), vi2x89ABCDEF), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16()
|
| D | 7p7x-minmax-neonfp16arith-c32.c | 54 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
64 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32()
158 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
167 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32()
282 const float16x8_t vi2x89ABCDEF = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
291 vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32()
|
| /external/XNNPACK/src/qu8-gavgpool/gen/ |
| D | 7p7x-minmax-fp32-neon-c16.c | 51 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
57 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
103 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
109 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
180 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
186 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16()
|
| D | 7p7x-minmax-rndnu-neon-c16.c | 51 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
57 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
103 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
109 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
181 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
187 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16()
|
| D | 7p7x-minmax-fp32-neonv8-c16.c | 52 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
58 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
104 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
110 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
180 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
186 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16()
|
| D | 7p7x-minmax-fp32-sse2-c16.c | 60 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local
67 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16()
146 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local
153 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16()
252 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local
259 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16()
|
| D | 7p7x-minmax-rndnu-neon-c24.c | 53 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
61 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
148 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
156 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
273 const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
281 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
|