Home
last modified time | relevance | path

Searched refs:va0x01234567 (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/qs8-igemm/gen/
D1x16c4-minmax-neondot.c65 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot() local
78 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
79 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
80 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
81 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
82 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
83 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
84 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
85 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
92 const int8x8_t va0x01234567 = vld1_s8(a0); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot() local
[all …]
D1x8c4-minmax-neondot.c63 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot() local
72 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot()
73 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot()
74 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot()
75 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot()
82 const int8x8_t va0x01234567 = vld1_s8(a0); in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot() local
89 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot()
90 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot()
D4x16c4-minmax-neondot.c101 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() local
117 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
118 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
119 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
120 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
133 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
134 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
135 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
136 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
155 const int8x8_t va0x01234567 = vld1_s8(a0); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() local
[all …]
D4x8c4-minmax-neondot.c93 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot() local
105 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
106 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
113 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
114 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
127 const int8x8_t va0x01234567 = vld1_s8(a0); in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot() local
137 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
138 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot()
D6x16c4-minmax-neondot.c125 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() local
143 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
144 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
145 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
146 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
167 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
168 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
169 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
170 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
197 const int8x8_t va0x01234567 = vld1_s8(a0); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() local
[all …]
D6x8c4-minmax-neondot.c113 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot() local
127 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
128 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
139 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
140 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
157 const int8x8_t va0x01234567 = vld1_s8(a0); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot() local
169 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
170 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot()
D8x16c4-minmax-neondot.c149 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local
169 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
170 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
171 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
172 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
201 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
202 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
203 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
204 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
239 const int8x8_t va0x01234567 = vld1_s8(a0); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local
[all …]
D8x8c4-minmax-neondot.c133 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() local
149 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
150 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
165 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
166 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
187 const int8x8_t va0x01234567 = vld1_s8(a0); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() local
201 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
202 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
/external/XNNPACK/src/qs8-gemm/gen/
D1x16c4-minmax-neondot.c56 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot() local
69 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
70 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
71 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
72 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
73 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
74 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
75 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
76 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
83 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4; in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot() local
[all …]
D1x8c4-minmax-neondot.c54 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot() local
63 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot()
64 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot()
65 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot()
66 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot()
73 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4; in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot() local
80 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot()
81 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot()
D4x16c4-minmax-neondot.c86 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() local
102 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
103 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
104 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
105 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
118 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
119 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
120 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
121 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
140 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4; in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() local
[all …]
D4x8c4-minmax-neondot.c78 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot() local
90 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
91 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
98 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
99 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
112 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4; in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot() local
122 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
123 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot()
D6x16c4-minmax-neondot.c106 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() local
124 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
125 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
126 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
127 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
148 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
149 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
150 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
151 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
178 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4; in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() local
[all …]
D6x8c4-minmax-neondot.c94 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot() local
108 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
109 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
120 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
121 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
138 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4; in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot() local
150 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
151 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot()
D8x16c4-minmax-neondot.c126 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local
146 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
147 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
148 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
149 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
178 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
179 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
180 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
181 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
216 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4; in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local
[all …]
D8x8c4-minmax-neondot.c110 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() local
126 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
127 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
142 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
143 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
164 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4; in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() local
178 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
179 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()