1 /*
2  * Copyright (c) 2022 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22  * IN THE SOFTWARE.
23  */
24 
25 #if defined(__ARM_FEATURE_SVE)
26 
27 template <>
interleave_block(int8_t * & out,const int8_t * const * in,size_t width,size_t height,size_t row_offset,bool first)28 void interleave_block<1, 4, VLType::SME, false>(
29   int8_t * &out, const int8_t * const *in,
30   size_t width, size_t height, size_t row_offset, bool first
31 )
32 {
33   ARM_COMPUTE_UNUSED(first);
34 
35   __asm__ __volatile__(
36       ".inst 0xd503477f  // SMSTART ZA\n"
37       "cntb x20\n"
38       "mov x22, %x[width]\n"
39       "incb x22\n"
40       "mov x19, %x[width]\n"
41       "sub x9, x20, #0x1\n"
42       "cntw x28\n"
43       "sub x22, x22, #0x1\n"
44       "ands x9, x19, x9\n"
45       "udiv x22, x22, x20\n"  // n_passes = ceildiv(width, VL<T>)
46       "csel x9, x9, x20, NE\n"
47       "lsl x21, %x[height], #0x1\n"  // height * 2
48       "lsl x20, x28, #0x1\n"
49       "sub x19, x22, #0x1\n"
50       "add x9, x9, #0x3\n"
51       "sub x27, x28, #0x2\n"
52       "whilelt p9.b, XZR, x21\n"
53       "whilelt p8.b, x20, x21\n"
54       "mov x26, #0x0\n"
55       "mov x25, %x[in]\n"
56       "lsr x19, x19, #0x1\n"  // n_loops = (n_passes - 1) / 2
57       "ldr x24, [x25, #0x0]\n"
58       "and x23, x22, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
59       "lsr x9, x9, #0x2\n"
60       "ldr x22, [x25, #0x8]\n"
61       "ptrue p11.s\n"
62       "zip1 p10.b, p9.b, p8.b\n"
63       "mov x21, %x[row_offset]\n"
64       "mov x20, %x[out]\n"
65       "whilelt p9.b, x26, %x[width]\n"
66       "whilelt p8.b, x26, %x[width]\n"
67       "add x25, x25, #0x10\n"
68       "mov x12, #0x0\n"
69       "cbz x27, 2f\n"
70       "1:"  // K loop: Charge: Loop
71       ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
72       ".inst 0xe0150300  // ld1b { za0h.b[x12] }, p0/Z, [x24, x21]\n"
73       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
74       "ldr x24, [x25, #0x0]\n"
75       ".inst 0xe01502c4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x22, x21]\n"
76       "add x12, x12, #0x8\n"
77       "cmp x12, x27, LSL #2\n"
78       "ldr x22, [x25, #0x8]\n"
79       "add x25, x25, #0x10\n"
80       "blt 1b\n"
81       "2:"  // K loop: Charge: End
82       ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
83       ".inst 0xe0150300  // ld1b { za0h.b[x12] }, p0/Z, [x24, x21]\n"
84       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
85       "mov x25, %x[in]\n"
86       ".inst 0xe01502c4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x22, x21]\n"
87       "ldr x24, [x25, #0x0]\n"
88       "incb x21\n"
89       "ldr x22, [x25, #0x8]\n"
90       "add x25, x25, #0x10\n"
91       "incb x26\n"
92       "cbz x19, 8f\n"
93       "mov x19, x19\n"
94       "3:"  // K loop: Main loop
95       "whilelt p8.b, x26, %x[width]\n"
96       "mov x13, #0x0\n"
97       "mov x12, #0x0\n"
98       "cbz x27, 5f\n"
99       "4:"  // K loop: Main loop: First: Loop
100       ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
101       ".inst 0xe0152302  // ld1b { za0h.b[x13, #2] }, p0/Z, [x24, x21]\n"
102       ".inst 0x25756141  // psel p1.b, p8.b/Z, p10.b[w13, #6]\n"
103       "ldr x24, [x25, #0x0]\n"
104       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
105       ".inst 0xe01526c6  // ld1b { za0h.b[x13, #6] }, p1/Z, [x22, x21]\n"
106       "ldr x22, [x25, #0x8]\n"
107       ".inst 0xe0bf8280  // st1w { za0v.s[x12] }, p0/Z, [x20, XZR, LSL #2]\n"
108       ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
109       ".inst 0xe0bc8281  // st1w { za0v.s[x12, #1] }, p0/Z, [x20, x28, LSL #2]\n"
110       "add x12, x12, #0x2\n"
111       "cmp x12, x27\n"
112       "add x25, x25, #0x10\n"
113       "addvl x20, x20, #2\n"
114       "add x13, x13, #0x8\n"
115       "blt 4b\n"
116       "5:"  // K loop: Main loop: First: Tail
117       ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
118       ".inst 0xe0152302  // ld1b { za0h.b[x13, #2] }, p0/Z, [x24, x21]\n"
119       "mov x25, %x[in]\n"
120       "ldr x24, [x25, #0x0]\n"
121       ".inst 0x25756141  // psel p1.b, p8.b/Z, p10.b[w13, #6]\n"
122       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
123       ".inst 0xe01526c6  // ld1b { za0h.b[x13, #6] }, p1/Z, [x22, x21]\n"
124       "ldr x22, [x25, #0x8]\n"
125       ".inst 0xe0bf8280  // st1w { za0v.s[x12] }, p0/Z, [x20, XZR, LSL #2]\n"
126       ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
127       "whilelt p9.b, x26, %x[width]\n"
128       "incb x26\n"
129       "add x25, x25, #0x10\n"
130       ".inst 0xe0bc8281  // st1w { za0v.s[x12, #1] }, p0/Z, [x20, x28, LSL #2]\n"
131       "addvl x20, x20, #2\n"
132       "incb x21\n"
133       "whilelt p8.b, x26, %x[width]\n"
134       "mov x13, #0x0\n"
135       "mov x12, #0x0\n"
136       "cbz x27, 7f\n"
137       "6:"  // K loop: Main loop: Second: Loop
138       ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
139       ".inst 0xe0152300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x21]\n"
140       ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
141       "ldr x24, [x25, #0x0]\n"
142       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
143       ".inst 0xe01526c4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x22, x21]\n"
144       "ldr x22, [x25, #0x8]\n"
145       ".inst 0xe0bf8288  // st1w { za2v.s[x12] }, p0/Z, [x20, XZR, LSL #2]\n"
146       ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
147       ".inst 0xe0bc8289  // st1w { za2v.s[x12, #1] }, p0/Z, [x20, x28, LSL #2]\n"
148       "add x12, x12, #0x2\n"
149       "cmp x12, x27\n"
150       "add x25, x25, #0x10\n"
151       "addvl x20, x20, #2\n"
152       "add x13, x13, #0x8\n"
153       "blt 6b\n"
154       "7:"  // K loop: Main loop: Second: Tail
155       ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
156       ".inst 0xe0152300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x21]\n"
157       "mov x25, %x[in]\n"
158       "ldr x24, [x25, #0x0]\n"
159       ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
160       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
161       ".inst 0xe01526c4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x22, x21]\n"
162       "ldr x22, [x25, #0x8]\n"
163       ".inst 0xe0bf8288  // st1w { za2v.s[x12] }, p0/Z, [x20, XZR, LSL #2]\n"
164       ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
165       "whilelt p9.b, x26, %x[width]\n"
166       "subs x19, x19, #0x1\n"
167       "add x25, x25, #0x10\n"
168       ".inst 0xe0bc8289  // st1w { za2v.s[x12, #1] }, p0/Z, [x20, x28, LSL #2]\n"
169       "addvl x20, x20, #2\n"
170       "incb x26\n"
171       "incb x21\n"
172       "bgt 3b\n"
173       "8:"  // K loop: Tails
174       "cbnz x23, 11f\n"
175       "mov x25, %x[in]\n"
176       "whilelt p8.b, x26, %x[width]\n"
177       "mov x13, #0x0\n"
178       "mov x12, #0x0\n"
179       "9:"  // K loop: Tails: Even: First
180       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
181       ".inst 0xe0bf8280  // st1w { za0v.s[x12] }, p0/Z, [x20, XZR, LSL #2]\n"
182       "ldr x24, [x25, #0x0]\n"
183       "add x12, x12, #0x1\n"
184       ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
185       "cmp x12, x28\n"
186       ".inst 0xe0152302  // ld1b { za0h.b[x13, #2] }, p0/Z, [x24, x21]\n"
187       "add x25, x25, #0x8\n"
188       "addvl x20, x20, #1\n"
189       "add x13, x13, #0x4\n"
190       "blt 9b\n"
191       "whilelt p9.b, x26, %x[width]\n"
192       "whilelt p8.b, x26, %x[width]\n"
193       "mov x19, #0x0\n"
194       "mov x12, #0x0\n"
195       "10:"  // K loop: Tails: Even: Second
196       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
197       ".inst 0xe0bf8288  // st1w { za2v.s[x12] }, p0/Z, [x20, XZR, LSL #2]\n"
198       "add x12, x12, #0x1\n"
199       "cmp x12, x9\n"
200       "addvl x20, x20, #1\n"
201       "add x19, x19, #0x4\n"
202       "blt 10b\n"
203       "whilelt p9.b, x26, %x[width]\n"
204       "b 13f\n"
205       "11:"  // K loop: Tails: Odd
206       "mov x12, #0x0\n"
207       "12:"  // K loop: Tails: Odd: Loop
208       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
209       ".inst 0xe0bf8280  // st1w { za0v.s[x12] }, p0/Z, [x20, XZR, LSL #2]\n"
210       "add x12, x12, #0x1\n"
211       "cmp x12, x9\n"
212       "addvl x20, x20, #1\n"
213       "blt 12b\n"
214       "13:"  // K loop: End
215       "mov %x[out], x20\n"
216       ".inst 0xd503467f  // SMSTOP\n"
217       : [out] "+&r" (out)
218       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
219       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x12", "x13", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "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"
220     );
221 }
222 
223 #endif  // defined(__ARM_FEATURE_SVE)
224