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(bfloat16 * & out,const bfloat16 * const * in,size_t width,size_t height,size_t row_offset,bool first)28 void interleave_block<1, 2, VLType::SME, false>(
29   bfloat16 * &out, const bfloat16 * 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       "cnth x20\n"
38       "cntw x10\n"
39       "mov x19, %x[width]\n"
40       "inch x19\n"
41       "sub x19, x19, #0x1\n"
42       "udiv x19, x19, x20\n" // n_passes = ceildiv(width, VL<T>)
43       "sub x9, x19, #0x1\n"
44       "lsr x9, x9, #0x1\n" // n_loops = (n_passes - 1) / 2
45       "and x28, x19, #0x1\n" // odd_tail = bool(n_passes & 0x1)
46       "mov x19, %x[width]\n"
47       "sub x27, x20, #0x1\n"
48       "ands x27, x19, x27\n"
49       "csel x27, x27, x20, NE\n"
50       "add x27, x27, #0x1\n"
51       "lsr x27, x27, #0x1\n"
52       "sub x26, x10, #0x2\n"
53       "ptrue p11.s\n"
54       "lsl x19, %x[height], #0x1\n" // height * 2
55       "whilelt p10.h, XZR, x19\n"
56       "mov x25, %x[row_offset]\n"
57       "mov x24, %x[out]\n"
58       "mov x23, #0x0\n"
59       "whilelt p9.h, x23, %x[width]\n"
60       "whilelt p8.h, x23, %x[width]\n"
61       "mov x22, %x[in]\n"
62       "ldr x21, [x22, #0x0]\n"
63       "ldr x20, [x22, #0x8]\n"
64       "add x22, x22, #0x10\n"
65       "mov x12, #0x0\n"
66       "cbz x26, 2f\n"
67       "1:"  // K loop: Charge: Loop
68       ".inst 0x25286140  // dup p0.h, p8.h/Z, p10.h[w12]\n"
69       ".inst 0xe05902a0  // ld1h { za0h.h[x12] }, p0/Z, [x21, x25, LSL #1]\n"
70       ".inst 0x25686140  // dup p0.h, p8.h/Z, p10.h[w12, #2]\n"
71       "ldr x21, [x22, #0x0]\n"
72       ".inst 0xe0590282  // ld1h { za0h.h[x12, #2] }, p0/Z, [x20, x25, LSL #1]\n"
73       "ldr x20, [x22, #0x8]\n"
74       "add x22, x22, #0x10\n"
75       "add x12, x12, #0x4\n"
76       "cmp x12, x26, LSL #1\n"
77       "blt 1b\n"
78       "2:"  // K loop: Charge: End
79       ".inst 0x25286140  // dup p0.h, p8.h/Z, p10.h[w12]\n"
80       ".inst 0xe05902a0  // ld1h { za0h.h[x12] }, p0/Z, [x21, x25, LSL #1]\n"
81       ".inst 0x25686140  // dup p0.h, p8.h/Z, p10.h[w12, #2]\n"
82       "mov x22, %x[in]\n"
83       ".inst 0xe0590282  // ld1h { za0h.h[x12, #2] }, p0/Z, [x20, x25, LSL #1]\n"
84       "ldr x21, [x22, #0x0]\n"
85       "ldr x20, [x22, #0x8]\n"
86       "add x22, x22, #0x10\n"
87       "inch x25\n"
88       "inch x23\n"
89       "cbz x9, 8f\n"
90       "mov x19, x9\n"
91       "3:"  // K loop: Main loop
92       "whilelt p8.h, x23, %x[width]\n"
93       "mov x13, #0x0\n"
94       "mov x12, #0x0\n"
95       "cbz x26, 5f\n"
96       "4:"  // K loop: Main loop: First: Loop
97       ".inst 0x25396140  // dup p0.h, p8.h/Z, p10.h[w13, #1]\n"
98       ".inst 0xe05922a1  // ld1h { za0h.h[x13, #1] }, p0/Z, [x21, x25, LSL #1]\n"
99       ".inst 0x25796140  // dup p0.h, p8.h/Z, p10.h[w13, #3]\n"
100       "ldr x21, [x22, #0x0]\n"
101       ".inst 0xe0592283  // ld1h { za0h.h[x13, #3] }, p0/Z, [x20, x25, LSL #1]\n"
102       "ldr x20, [x22, #0x8]\n"
103       "add x22, x22, #0x10\n"
104       ".inst 0x25306d20  // dup p0.s, p11.s/Z, p9.s[w12]\n"
105       ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
106       ".inst 0x25706d20  // dup p0.s, p11.s/Z, p9.s[w12, #1]\n"
107       "add x13, x13, #0x4\n"
108       ".inst 0xe0aa8301  // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x10, LSL #2]\n"
109       "addvl x24, x24, #2\n"
110       "add x12, x12, #0x2\n"
111       "cmp x12, x26\n"
112       "blt 4b\n"
113       "5:"  // K loop: Main loop: First: Tail
114       "mov x22, %x[in]\n"
115       ".inst 0x25396140  // dup p0.h, p8.h/Z, p10.h[w13, #1]\n"
116       ".inst 0xe05922a1  // ld1h { za0h.h[x13, #1] }, p0/Z, [x21, x25, LSL #1]\n"
117       ".inst 0x25796140  // dup p0.h, p8.h/Z, p10.h[w13, #3]\n"
118       "ldr x21, [x22, #0x0]\n"
119       ".inst 0xe0592283  // ld1h { za0h.h[x13, #3] }, p0/Z, [x20, x25, LSL #1]\n"
120       "ldr x20, [x22, #0x8]\n"
121       "add x22, x22, #0x10\n"
122       ".inst 0x25306d20  // dup p0.s, p11.s/Z, p9.s[w12]\n"
123       ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
124       ".inst 0x25706d20  // dup p0.s, p11.s/Z, p9.s[w12, #1]\n"
125       "whilelt p9.h, x23, %x[width]\n"
126       ".inst 0xe0aa8301  // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x10, LSL #2]\n"
127       "addvl x24, x24, #2\n"
128       "inch x23\n"
129       "inch x25\n"
130       "whilelt p8.h, x23, %x[width]\n"
131       "mov x13, #0x0\n"
132       "mov x12, #0x0\n"
133       "cbz x26, 7f\n"
134       "6:"  // K loop: Main loop: Second: Loop
135       ".inst 0x25296140  // dup p0.h, p8.h/Z, p10.h[w13]\n"
136       ".inst 0xe05922a0  // ld1h { za0h.h[x13] }, p0/Z, [x21, x25, LSL #1]\n"
137       ".inst 0x25696140  // dup p0.h, p8.h/Z, p10.h[w13, #2]\n"
138       "ldr x21, [x22, #0x0]\n"
139       ".inst 0xe0592282  // ld1h { za0h.h[x13, #2] }, p0/Z, [x20, x25, LSL #1]\n"
140       "ldr x20, [x22, #0x8]\n"
141       "add x22, x22, #0x10\n"
142       ".inst 0x25306d20  // dup p0.s, p11.s/Z, p9.s[w12]\n"
143       ".inst 0xe0bf8308  // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
144       ".inst 0x25706d20  // dup p0.s, p11.s/Z, p9.s[w12, #1]\n"
145       "add x13, x13, #0x4\n"
146       ".inst 0xe0aa8309  // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x10, LSL #2]\n"
147       "addvl x24, x24, #2\n"
148       "add x12, x12, #0x2\n"
149       "cmp x12, x26\n"
150       "blt 6b\n"
151       "7:"  // K loop: Main loop: Second: Tail
152       "mov x22, %x[in]\n"
153       ".inst 0x25296140  // dup p0.h, p8.h/Z, p10.h[w13]\n"
154       ".inst 0xe05922a0  // ld1h { za0h.h[x13] }, p0/Z, [x21, x25, LSL #1]\n"
155       ".inst 0x25696140  // dup p0.h, p8.h/Z, p10.h[w13, #2]\n"
156       "ldr x21, [x22, #0x0]\n"
157       ".inst 0xe0592282  // ld1h { za0h.h[x13, #2] }, p0/Z, [x20, x25, LSL #1]\n"
158       "ldr x20, [x22, #0x8]\n"
159       "add x22, x22, #0x10\n"
160       ".inst 0x25306d20  // dup p0.s, p11.s/Z, p9.s[w12]\n"
161       ".inst 0xe0bf8308  // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
162       ".inst 0x25706d20  // dup p0.s, p11.s/Z, p9.s[w12, #1]\n"
163       "whilelt p9.h, x23, %x[width]\n"
164       ".inst 0xe0aa8309  // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x10, LSL #2]\n"
165       "addvl x24, x24, #2\n"
166       "inch x23\n"
167       "inch x25\n"
168       "subs x19, x19, #0x1\n"
169       "bgt 3b\n"
170       "8:"  // K loop: Tails
171       "cbnz x28, 11f\n"
172       "mov x22, %x[in]\n"
173       "whilelt p8.h, x23, %x[width]\n"
174       "mov x13, #0x0\n"
175       "mov x12, #0x0\n"
176       "9:"  // K loop: Tails: Even: First
177       ".inst 0x25306d20  // dup p0.s, p11.s/Z, p9.s[w12]\n"
178       ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
179       ".inst 0x25396140  // dup p0.h, p8.h/Z, p10.h[w13, #1]\n"
180       "addvl x24, x24, #1\n"
181       "ldr x21, [x22, #0x0]\n"
182       ".inst 0xe05922a1  // ld1h { za0h.h[x13, #1] }, p0/Z, [x21, x25, LSL #1]\n"
183       "add x22, x22, #0x8\n"
184       "add x13, x13, #0x2\n"
185       "add x12, x12, #0x1\n"
186       "cmp x12, x10\n"
187       "blt 9b\n"
188       "whilelt p9.h, x23, %x[width]\n"
189       "whilelt p8.h, x23, %x[width]\n"
190       "mov x19, #0x0\n"
191       "mov x12, #0x0\n"
192       "10:"  // K loop: Tails: Even: Second
193       ".inst 0x25306d20  // dup p0.s, p11.s/Z, p9.s[w12]\n"
194       ".inst 0xe0bf8308  // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
195       "addvl x24, x24, #1\n"
196       "add x19, x19, #0x2\n"
197       "add x12, x12, #0x1\n"
198       "cmp x12, x27\n"
199       "blt 10b\n"
200       "whilelt p9.h, x23, %x[width]\n"
201       "b 13f\n"
202       "11:"  // K loop: Tails: Odd
203       "mov x12, #0x0\n"
204       "12:"  // K loop: Tails: Odd: Loop
205       ".inst 0x25306d20  // dup p0.s, p11.s/Z, p9.s[w12]\n"
206       ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
207       "addvl x24, x24, #1\n"
208       "add x12, x12, #0x1\n"
209       "cmp x12, x27\n"
210       "blt 12b\n"
211       "13:"  // K loop: End
212       "mov %x[out], x24\n"
213       ".inst 0xd503467f  // SMSTOP\n"
214       : [out] "+&r" (out)
215       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
216       : "cc", "memory", "p0", "p8", "p9", "p10", "p11", "x9", "x10", "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"
217     );
218 }
219 
220 #endif  // defined(__ARM_FEATURE_SVE)
221