25 #if defined(__aarch64__)
29 #include "depthwise.hpp"
35 struct interleave_a64_u8q_3x3_dot
54 "cmp %x[ld_weight_col], XZR\n"
55 "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n"
57 "movi v31.16b, #0x0\n"
59 "mul x21, %x[ld_weight_col], x21\n"
60 "add x20, %x[qp], %[offsetof_input_offset]\n"
61 "ld1r { v30.4s }, [x20]\n"
62 "add x20, %x[qp], %[offsetof_weights_offset]\n"
63 "ld1r { v29.4s }, [x20]\n"
64 "cmp %x[ld_weight_row], XZR\n"
65 "mul v29.4s, v29.4s, v30.4s\n"
66 "csel %x[ld_weight_row], %x[ld_weight_row], x21, NE\n"
67 "lsr x21, %x[n_channels], #0x2\n"
68 "movi v28.16b, #0x1\n"
69 "mul v29.4s, v29.4s, v16.4s\n"
70 "add x25, %x[weights], %x[ld_weight_row]\n"
71 "add x20, %x[qp], %[offsetof_per_layer_mul]\n"
72 "ld1r { v27.4s }, [x20]\n"
73 "add x20, %x[qp], %[offsetof_per_layer_right_shift]\n"
74 "ld1r { v26.4s }, [x20]\n"
75 "add x24, x25, %x[ld_weight_row]\n"
76 "add x23, %x[ld_weight_col], %x[ld_weight_col]\n"
82 "ldr q25, [%x[bias], x22]\n"
84 "ldr s19, [%x[weights], #0x0]\n"
85 "ldr s16, [%x[weights], %x[ld_weight_col]]\n"
86 "zip1 v17.16b, v16.16b, v31.16b\n"
88 "ldr s16, [%x[weights], x23]\n"
89 "ldr s18, [x25, #0x0]\n"
90 "zip1 v16.16b, v19.16b, v16.16b\n"
91 "zip1 v20.16b, v16.16b, v17.16b\n"
92 "ldr s17, [x25, %x[ld_weight_col]]\n"
93 "ldr s16, [x25, x23]\n"
94 "zip1 v18.16b, v18.16b, v16.16b\n"
95 "zip1 v16.16b, v17.16b, v31.16b\n"
96 "ldr s17, [x24, #0x0]\n"
97 "ldr s19, [x24, %x[ld_weight_col]]\n"
98 ".inst 0x6e949795 // udot v21.4s, v28.16b, v20.16b\n"
99 "zip1 v18.16b, v18.16b, v16.16b\n"
100 "ldr s16, [x24, x23]\n"
101 "zip1 v17.16b, v17.16b, v16.16b\n"
102 "zip1 v16.16b, v19.16b, v31.16b\n"
103 ".inst 0x6e929795 // udot v21.4s, v28.16b, v18.16b\n"
104 "zip1 v16.16b, v17.16b, v16.16b\n"
105 ".inst 0x6e909795 // udot v21.4s, v28.16b, v16.16b\n"
106 "add %x[weights], %x[weights], #0x4\n"
107 "add x25, x25, #0x4\n"
108 "mls v25.4s, v21.4s, v30.4s\n"
109 "add x24, x24, #0x4\n"
110 "add v25.4s, v25.4s, v29.4s\n"
111 "str q25, [%x[outptr], #0x0]\n"
112 "str q20, [%x[outptr], #0x10]\n"
113 "str q18, [%x[outptr], #0x20]\n"
114 "str q16, [%x[outptr], #0x30]\n"
115 "add %x[outptr], %x[outptr], #0x40\n"
116 "cbz %x[rq_mul_perchannel], 3f\n"
117 "ldr q27, [%x[rq_mul_perchannel], x22]\n"
118 "ldr q26, [%x[rq_shift_perchannel], x22]\n"
120 "subs x21, x21, #0x1\n"
121 "str q27, [%x[outptr], #0x0]\n"
122 "add x22, x22, #0x10\n"
123 "str q26, [%x[outptr], #0x10]\n"
124 "add %x[outptr], %x[outptr], #0x20\n"
126 "tst %x[n_channels], #0x3\n"
129 "movi v25.4s, #0x0\n"
131 "add %x[bias], %x[bias], x22\n"
132 "tbz %x[n_channels], #1, 5f\n"
133 "ld1 { v25.d }[0], [%x[bias]], #0x8\n"
134 "tbz %x[n_channels], #0, 6f\n"
135 "ld1 { v25.s }[2], [%x[bias]], #0x4\n"
138 "ld1 { v25.s }[0], [%x[bias]], #0x4\n"
141 "tbz %x[n_channels], #1, 8f\n"
142 "ld1 { v17.h }[0], [%x[weights]]\n"
143 "ld1 { v24.h }[0], [x25]\n"
144 "add x21, %x[weights], %x[ld_weight_col]\n"
145 "add x20, %x[weights], x23\n"
146 "ld1 { v20.h }[0], [x21]\n"
147 "ld1 { v16.h }[0], [x20]\n"
148 "add x21, x25, %x[ld_weight_col]\n"
149 "add x20, x25, x23\n"
150 "ld1 { v19.h }[0], [x21]\n"
151 "ld1 { v18.h }[0], [x20]\n"
152 "add x21, x24, %x[ld_weight_col]\n"
153 "add x20, x24, x23\n"
154 "ld1 { v23.h }[0], [x24]\n"
155 "ld1 { v22.h }[0], [x21]\n"
156 "add %x[weights], %x[weights], #0x2\n"
157 "add x25, x25, #0x2\n"
158 "ld1 { v21.h }[0], [x20]\n"
159 "add x24, x24, #0x2\n"
160 "tbz %x[n_channels], #0, 9f\n"
161 "ld1 { v17.b }[2], [%x[weights]]\n"
162 "ld1 { v24.b }[2], [x25]\n"
163 "add x21, %x[weights], %x[ld_weight_col]\n"
164 "add x20, %x[weights], x23\n"
165 "ld1 { v20.b }[2], [x21]\n"
166 "ld1 { v16.b }[2], [x20]\n"
167 "add x21, x25, %x[ld_weight_col]\n"
168 "add x20, x25, x23\n"
169 "ld1 { v19.b }[2], [x21]\n"
170 "ld1 { v18.b }[2], [x20]\n"
171 "add x21, x24, %x[ld_weight_col]\n"
172 "add x20, x24, x23\n"
173 "ld1 { v23.b }[2], [x24]\n"
174 "ld1 { v22.b }[2], [x21]\n"
175 "add %x[weights], %x[weights], #0x1\n"
176 "ld1 { v21.b }[2], [x20]\n"
179 "ld1 { v17.b }[0], [%x[weights]]\n"
180 "ld1 { v24.b }[0], [x25]\n"
181 "add x21, %x[weights], %x[ld_weight_col]\n"
182 "add x20, %x[weights], x23\n"
183 "ld1 { v20.b }[0], [x21]\n"
184 "ld1 { v16.b }[0], [x20]\n"
185 "add x21, x25, %x[ld_weight_col]\n"
186 "add x20, x25, x23\n"
187 "ld1 { v19.b }[0], [x21]\n"
188 "ld1 { v18.b }[0], [x20]\n"
189 "add x21, x24, %x[ld_weight_col]\n"
190 "add x20, x24, x23\n"
191 "ld1 { v23.b }[0], [x24]\n"
192 "ld1 { v22.b }[0], [x21]\n"
193 "add %x[weights], %x[weights], #0x1\n"
194 "ld1 { v21.b }[0], [x20]\n"
196 "zip1 v17.16b, v17.16b, v16.16b\n"
197 "zip1 v16.16b, v20.16b, v31.16b\n"
198 "zip1 v20.16b, v17.16b, v16.16b\n"
199 "zip1 v17.16b, v24.16b, v18.16b\n"
200 "zip1 v16.16b, v19.16b, v31.16b\n"
201 "movi v19.4s, #0x0\n"
202 ".inst 0x6e949793 // udot v19.4s, v28.16b, v20.16b\n"
203 "zip1 v18.16b, v17.16b, v16.16b\n"
204 "zip1 v17.16b, v23.16b, v21.16b\n"
205 ".inst 0x6e929793 // udot v19.4s, v28.16b, v18.16b\n"
206 "zip1 v16.16b, v22.16b, v31.16b\n"
207 "zip1 v16.16b, v17.16b, v16.16b\n"
208 ".inst 0x6e909793 // udot v19.4s, v28.16b, v16.16b\n"
209 "mls v25.4s, v19.4s, v30.4s\n"
210 "add v25.4s, v25.4s, v29.4s\n"
211 "str q25, [%x[outptr], #0x0]\n"
212 "str q20, [%x[outptr], #0x10]\n"
213 "str q18, [%x[outptr], #0x20]\n"
214 "str q16, [%x[outptr], #0x30]\n"
215 "add %x[outptr], %x[outptr], #0x40\n"
216 "cbz %x[rq_mul_perchannel], 12f\n"
217 "add x21, %x[rq_mul_perchannel], x22\n"
218 "add x20, %x[rq_shift_perchannel], x22\n"
219 "tbz %x[n_channels], #1, 10f\n"
220 "ld1 { v27.d }[0], [x21], #0x8\n"
221 "ld1 { v26.d }[0], [x20], #0x8\n"
222 "tbz %x[n_channels], #0, 11f\n"
223 "ld1 { v27.s }[2], [x21], #0x4\n"
224 "ld1 { v26.s }[2], [x20], #0x4\n"
227 "ld1 { v27.s }[0], [x21], #0x4\n"
228 "ld1 { v26.s }[0], [x20], #0x4\n"
231 "str q27, [%x[outptr], #0x0]\n"
232 "str q26, [%x[outptr], #0x10]\n"
233 "add %x[outptr], %x[outptr], #0x20\n"
235 : [
bias]
"+&r" (
bias), [ld_weight_col]
"+&r" (ld_weight_col), [ld_weight_row]
"+&r" (ld_weight_row), [outptr]
"+&r" (outptr), [weights]
"+&r" (weights)
236 : [n_channels]
"r" (n_channels), [offsetof_input_offset]
"I" (offsetof(
arm_gemm::Requantize32, a_offset)), [offsetof_per_layer_mul]
"I" (offsetof(
arm_gemm::Requantize32, per_layer_mul)), [offsetof_per_layer_right_shift]
"I" (offsetof(
arm_gemm::Requantize32, per_layer_right_shift)), [offsetof_weights_offset]
"I" (offsetof(
arm_gemm::Requantize32, b_offset)), [qp]
"r" (&qp), [rq_mul_perchannel]
"r" (qp.
per_channel_muls), [rq_shift_perchannel]
"r" (qp.
per_channel_right_shifts)
237 :
"cc",
"memory",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25"
244 #endif // defined(__aarch64__)