27 #if defined(ARM_COMPUTE_ENABLE_SVE)
31 void sve_transpose_interleave_8VL(uint32_t *out,
const uint32_t *in,
size_t width,
size_t in_stride,
size_t height)
33 size_t out_stride = 8 * height * get_vector_length<uint8_t>();
36 "cmp %x[height], #0x2\n"
41 "mov x25, %x[width]\n"
42 "cntw x24, ALL, MUL #16\n"
43 "add x23, x26, %x[in_stride]\n"
45 "add %x[in], x23, %x[in_stride]\n"
47 "sub %x[height], %x[height], #0x2\n"
50 "ld1w { z15.s }, p1/Z, [x26]\n"
51 "ld1w { z14.s }, p1/Z, [x26, #1, MUL VL]\n"
53 "add x22, x22, %x[out_stride]\n"
54 "ld1w { z13.s }, p1/Z, [x26, #2, MUL VL]\n"
55 "ld1w { z12.s }, p1/Z, [x26, #3, MUL VL]\n"
58 "ld1w { z11.s }, p1/Z, [x26, #4, MUL VL]\n"
59 "ld1w { z10.s }, p1/Z, [x26, #5, MUL VL]\n"
61 "add x22, x22, %x[out_stride]\n"
62 "ld1w { z9.s }, p1/Z, [x26, #6, MUL VL]\n"
63 "ld1w { z8.s }, p1/Z, [x26, #7, MUL VL]\n"
64 "addvl x26, x26, #16\n"
65 "ld1w { z7.s }, p1/Z, [x23]\n"
66 "ld1w { z6.s }, p1/Z, [x23, #1, MUL VL]\n"
67 "ld1w { z5.s }, p1/Z, [x23, #2, MUL VL]\n"
68 "ld1w { z4.s }, p1/Z, [x23, #3, MUL VL]\n"
69 "ld1w { z3.s }, p1/Z, [x23, #4, MUL VL]\n"
70 "ld1w { z2.s }, p1/Z, [x23, #5, MUL VL]\n"
71 "ld1w { z1.s }, p1/Z, [x23, #6, MUL VL]\n"
72 "ld1w { z0.s }, p1/Z, [x23, #7, MUL VL]\n"
73 "addvl x23, x23, #16\n"
74 "ld1w { z31.s }, p1/Z, [x26, #-8, MUL VL]\n"
75 "ld1w { z30.s }, p1/Z, [x26, #-7, MUL VL]\n"
76 "ld1w { z29.s }, p1/Z, [x26, #-6, MUL VL]\n"
77 "ld1w { z28.s }, p1/Z, [x26, #-5, MUL VL]\n"
78 "ld1w { z27.s }, p1/Z, [x26, #-4, MUL VL]\n"
79 "ld1w { z26.s }, p1/Z, [x26, #-3, MUL VL]\n"
80 "ld1w { z25.s }, p1/Z, [x26, #-2, MUL VL]\n"
81 "ld1w { z24.s }, p1/Z, [x26, #-1, MUL VL]\n"
82 "ld1w { z23.s }, p1/Z, [x23, #-8, MUL VL]\n"
83 "ld1w { z22.s }, p1/Z, [x23, #-7, MUL VL]\n"
84 "ld1w { z21.s }, p1/Z, [x23, #-6, MUL VL]\n"
85 "ld1w { z20.s }, p1/Z, [x23, #-5, MUL VL]\n"
86 "ld1w { z19.s }, p1/Z, [x23, #-4, MUL VL]\n"
87 "ld1w { z18.s }, p1/Z, [x23, #-3, MUL VL]\n"
88 "ld1w { z17.s }, p1/Z, [x23, #-2, MUL VL]\n"
89 "ld1w { z16.s }, p1/Z, [x23, #-1, MUL VL]\n"
90 "st1w { z15.s }, p1, [x21]\n"
91 "st1w { z14.s }, p1, [x21, #1, MUL VL]\n"
92 "st1w { z13.s }, p1, [x21, #2, MUL VL]\n"
93 "st1w { z12.s }, p1, [x21, #3, MUL VL]\n"
94 "st1w { z11.s }, p1, [x21, #4, MUL VL]\n"
95 "st1w { z10.s }, p1, [x21, #5, MUL VL]\n"
96 "st1w { z9.s }, p1, [x21, #6, MUL VL]\n"
97 "st1w { z8.s }, p1, [x21, #7, MUL VL]\n"
98 "addvl x21, x21, #16\n"
99 "st1w { z7.s }, p1, [x21, #-8, MUL VL]\n"
100 "st1w { z6.s }, p1, [x21, #-7, MUL VL]\n"
101 "st1w { z5.s }, p1, [x21, #-6, MUL VL]\n"
102 "st1w { z4.s }, p1, [x21, #-5, MUL VL]\n"
103 "st1w { z3.s }, p1, [x21, #-4, MUL VL]\n"
104 "st1w { z2.s }, p1, [x21, #-3, MUL VL]\n"
105 "st1w { z1.s }, p1, [x21, #-2, MUL VL]\n"
106 "st1w { z0.s }, p1, [x21, #-1, MUL VL]\n"
107 "st1w { z31.s }, p1, [x20]\n"
108 "st1w { z30.s }, p1, [x20, #1, MUL VL]\n"
109 "st1w { z29.s }, p1, [x20, #2, MUL VL]\n"
110 "st1w { z28.s }, p1, [x20, #3, MUL VL]\n"
111 "st1w { z27.s }, p1, [x20, #4, MUL VL]\n"
112 "st1w { z26.s }, p1, [x20, #5, MUL VL]\n"
113 "st1w { z25.s }, p1, [x20, #6, MUL VL]\n"
114 "st1w { z24.s }, p1, [x20, #7, MUL VL]\n"
115 "addvl x20, x20, #16\n"
116 "st1w { z23.s }, p1, [x20, #-8, MUL VL]\n"
117 "st1w { z22.s }, p1, [x20, #-7, MUL VL]\n"
118 "st1w { z21.s }, p1, [x20, #-6, MUL VL]\n"
119 "st1w { z20.s }, p1, [x20, #-5, MUL VL]\n"
120 "st1w { z19.s }, p1, [x20, #-4, MUL VL]\n"
121 "st1w { z18.s }, p1, [x20, #-3, MUL VL]\n"
122 "st1w { z17.s }, p1, [x20, #-2, MUL VL]\n"
123 "st1w { z16.s }, p1, [x20, #-1, MUL VL]\n"
129 "whilelt p0.s, XZR, x20\n"
130 "ld1w { z31.s }, p0/Z, [x26]\n"
131 "ld1w { z30.s }, p0/Z, [x23]\n"
133 "whilelt p0.s, XZR, x20\n"
134 "ld1w { z29.s }, p0/Z, [x26, #1, MUL VL]\n"
135 "ld1w { z28.s }, p0/Z, [x23, #1, MUL VL]\n"
137 "whilelt p0.s, XZR, x20\n"
138 "ld1w { z27.s }, p0/Z, [x26, #2, MUL VL]\n"
139 "ld1w { z26.s }, p0/Z, [x23, #2, MUL VL]\n"
141 "whilelt p0.s, XZR, x20\n"
142 "ld1w { z25.s }, p0/Z, [x26, #3, MUL VL]\n"
143 "ld1w { z24.s }, p0/Z, [x23, #3, MUL VL]\n"
145 "whilelt p0.s, XZR, x20\n"
146 "ld1w { z23.s }, p0/Z, [x26, #4, MUL VL]\n"
147 "ld1w { z22.s }, p0/Z, [x23, #4, MUL VL]\n"
149 "whilelt p0.s, XZR, x20\n"
150 "ld1w { z21.s }, p0/Z, [x26, #5, MUL VL]\n"
151 "ld1w { z20.s }, p0/Z, [x23, #5, MUL VL]\n"
153 "whilelt p0.s, XZR, x20\n"
154 "ld1w { z19.s }, p0/Z, [x26, #6, MUL VL]\n"
155 "ld1w { z18.s }, p0/Z, [x23, #6, MUL VL]\n"
157 "whilelt p0.s, XZR, x20\n"
158 "ld1w { z17.s }, p0/Z, [x26, #7, MUL VL]\n"
159 "ld1w { z16.s }, p0/Z, [x23, #7, MUL VL]\n"
161 "decw x25, ALL, MUL #8\n"
162 "st1w { z31.s }, p1, [x20]\n"
163 "st1w { z29.s }, p1, [x20, #1, MUL VL]\n"
165 "addvl x26, x26, #8\n"
166 "st1w { z27.s }, p1, [x20, #2, MUL VL]\n"
167 "addvl x23, x23, #8\n"
168 "add x22, x22, %x[out_stride]\n"
169 "st1w { z25.s }, p1, [x20, #3, MUL VL]\n"
170 "st1w { z23.s }, p1, [x20, #4, MUL VL]\n"
171 "st1w { z21.s }, p1, [x20, #5, MUL VL]\n"
172 "st1w { z19.s }, p1, [x20, #6, MUL VL]\n"
173 "st1w { z17.s }, p1, [x20, #7, MUL VL]\n"
174 "addvl x20, x20, #16\n"
175 "st1w { z30.s }, p1, [x20, #-8, MUL VL]\n"
176 "st1w { z28.s }, p1, [x20, #-7, MUL VL]\n"
177 "st1w { z26.s }, p1, [x20, #-6, MUL VL]\n"
178 "st1w { z24.s }, p1, [x20, #-5, MUL VL]\n"
179 "st1w { z22.s }, p1, [x20, #-4, MUL VL]\n"
180 "st1w { z20.s }, p1, [x20, #-3, MUL VL]\n"
181 "st1w { z18.s }, p1, [x20, #-2, MUL VL]\n"
182 "st1w { z16.s }, p1, [x20, #-1, MUL VL]\n"
185 "cmp %x[height], #0x2\n"
186 "addvl %x[out], %x[out], #16\n"
188 "cbz %x[height], 12f\n"
191 "mov x21, %x[width]\n"
192 "cntw x20, ALL, MUL #16\n"
195 "add %x[in], x26, %x[in_stride]\n"
197 "sub %x[height], %x[height], #0x1\n"
200 "ld1w { z31.s }, p1/Z, [x26]\n"
201 "ld1w { z30.s }, p1/Z, [x26, #1, MUL VL]\n"
202 "sub x21, x21, x20\n"
204 "ld1w { z29.s }, p1/Z, [x26, #2, MUL VL]\n"
205 "ld1w { z28.s }, p1/Z, [x26, #3, MUL VL]\n"
206 "ld1w { z27.s }, p1/Z, [x26, #4, MUL VL]\n"
207 "ld1w { z26.s }, p1/Z, [x26, #5, MUL VL]\n"
208 "ld1w { z25.s }, p1/Z, [x26, #6, MUL VL]\n"
209 "ld1w { z24.s }, p1/Z, [x26, #7, MUL VL]\n"
210 "addvl x26, x26, #16\n"
211 "ld1w { z23.s }, p1/Z, [x26, #-8, MUL VL]\n"
212 "ld1w { z22.s }, p1/Z, [x26, #-7, MUL VL]\n"
213 "ld1w { z21.s }, p1/Z, [x26, #-6, MUL VL]\n"
214 "ld1w { z20.s }, p1/Z, [x26, #-5, MUL VL]\n"
215 "ld1w { z19.s }, p1/Z, [x26, #-4, MUL VL]\n"
216 "ld1w { z18.s }, p1/Z, [x26, #-3, MUL VL]\n"
217 "ld1w { z17.s }, p1/Z, [x26, #-2, MUL VL]\n"
218 "ld1w { z16.s }, p1/Z, [x26, #-1, MUL VL]\n"
219 "st1w { z31.s }, p1, [x22]\n"
220 "st1w { z30.s }, p1, [x22, #1, MUL VL]\n"
221 "st1w { z29.s }, p1, [x22, #2, MUL VL]\n"
222 "st1w { z28.s }, p1, [x22, #3, MUL VL]\n"
223 "st1w { z27.s }, p1, [x22, #4, MUL VL]\n"
224 "st1w { z26.s }, p1, [x22, #5, MUL VL]\n"
225 "st1w { z25.s }, p1, [x22, #6, MUL VL]\n"
226 "st1w { z24.s }, p1, [x22, #7, MUL VL]\n"
227 "add x22, x22, %x[out_stride]\n"
228 "st1w { z23.s }, p1, [x22]\n"
229 "st1w { z22.s }, p1, [x22, #1, MUL VL]\n"
230 "st1w { z21.s }, p1, [x22, #2, MUL VL]\n"
231 "st1w { z20.s }, p1, [x22, #3, MUL VL]\n"
232 "st1w { z19.s }, p1, [x22, #4, MUL VL]\n"
233 "st1w { z18.s }, p1, [x22, #5, MUL VL]\n"
234 "st1w { z17.s }, p1, [x22, #6, MUL VL]\n"
235 "st1w { z16.s }, p1, [x22, #7, MUL VL]\n"
236 "add x22, x22, %x[out_stride]\n"
242 "whilelt p0.s, XZR, x20\n"
243 "ld1w { z23.s }, p0/Z, [x26]\n"
245 "whilelt p0.s, XZR, x20\n"
246 "ld1w { z22.s }, p0/Z, [x26, #1, MUL VL]\n"
248 "whilelt p0.s, XZR, x20\n"
249 "ld1w { z21.s }, p0/Z, [x26, #2, MUL VL]\n"
251 "whilelt p0.s, XZR, x20\n"
252 "ld1w { z20.s }, p0/Z, [x26, #3, MUL VL]\n"
254 "whilelt p0.s, XZR, x20\n"
255 "ld1w { z19.s }, p0/Z, [x26, #4, MUL VL]\n"
257 "whilelt p0.s, XZR, x20\n"
258 "ld1w { z18.s }, p0/Z, [x26, #5, MUL VL]\n"
260 "whilelt p0.s, XZR, x20\n"
261 "ld1w { z17.s }, p0/Z, [x26, #6, MUL VL]\n"
263 "decw x21, ALL, MUL #8\n"
264 "whilelt p0.s, XZR, x20\n"
266 "ld1w { z16.s }, p0/Z, [x26, #7, MUL VL]\n"
267 "st1w { z23.s }, p1, [x22]\n"
268 "addvl x26, x26, #8\n"
269 "st1w { z22.s }, p1, [x22, #1, MUL VL]\n"
270 "st1w { z21.s }, p1, [x22, #2, MUL VL]\n"
271 "st1w { z20.s }, p1, [x22, #3, MUL VL]\n"
272 "st1w { z19.s }, p1, [x22, #4, MUL VL]\n"
273 "st1w { z18.s }, p1, [x22, #5, MUL VL]\n"
274 "st1w { z17.s }, p1, [x22, #6, MUL VL]\n"
275 "st1w { z16.s }, p1, [x22, #7, MUL VL]\n"
276 "add x22, x22, %x[out_stride]\n"
279 "cmp %x[height], #0x1\n"
280 "addvl %x[out], %x[out], #8\n"
283 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
284 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [width]
"r" (width)
285 :
"cc",
"memory",
"p0",
"p1",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"z0",
"z1",
"z2",
"z3",
"z4",
"z5",
"z6",
"z7",
"z8",
"z9",
"z10",
"z11",
"z12",
"z13",
"z14",
"z15",
"z16",
"z17",
"z18",
"z19",
"z20",
"z21",
"z22",
"z23",
"z24",
"z25",
"z26",
"z27",
"z28",
"z29",
"z30",
"z31"
292 void Transform<8, 1, true, VLType::SVE>(
293 float *out,
const float *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
295 sve_transpose_interleave_8VL(
296 reinterpret_cast<uint32_t *
>(out),
297 reinterpret_cast<const uint32_t *
>(in + k0 * stride + x0),
298 (xmax-x0) *
sizeof(
float) / 4,
299 stride *
sizeof(
float),
305 #endif // defined(ARM_COMPUTE_ENABLE_SVE)