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 #pragma once
26 
27 #ifdef __ARM_FEATURE_SVE
28 
29 
30 namespace {
31 
sme_transpose_interleave_1VL_2x2(uint16_t * out,const uint16_t * in,size_t width,size_t in_stride,size_t height)32 void sme_transpose_interleave_1VL_2x2(uint16_t *out, const uint16_t *in, size_t width, size_t in_stride, size_t height)
33 {
34     uint16_t *pad_row = reinterpret_cast<uint16_t *>(alloca(width * sizeof(uint16_t)));
35 
36     if (height % 2) {
37         memset(pad_row, 0, width * sizeof(uint16_t));
38     }
39 
40     size_t out_stride = 1 * roundup<size_t>(height, 2) * sme::get_vector_length<uint16_t>();
41 
42     __asm__ __volatile__(
43       ".inst 0xd503477f  // SMSTART ZA\n"
44       "cmp %x[height], #0x4\n"
45       "ptrue p1.b\n"
46       "blt 6f\n"
47       "1:"  // Main row loop: Head
48       "mov x25, %x[in]\n"
49       "add x24, x25, %x[in_stride]\n"
50       "add x23, x24, %x[in_stride]\n"
51       "mov x22, %x[width]\n"
52       "cnth x20, ALL, MUL #2\n"
53       "add x19, x23, %x[in_stride]\n"
54       "cmp x22, x20\n"
55       "add %x[in], x19, %x[in_stride]\n"
56       "mov x21, %x[out]\n"
57       "sub %x[height], %x[height], #0x4\n"
58       "blt 3f\n"
59       "2:"  // Main row loop: Unroll column loop
60       "ld1h { z17.h }, p1/Z, [x25]\n"
61       "sub x22, x22, x20\n"
62       "cmp x22, x20\n"
63       "ld1h { z16.h }, p1/Z, [x24]\n"
64       "zip1 z24.h, z17.h, z16.h\n"
65       "zip2 z23.h, z17.h, z16.h\n"
66       "ld1h { z17.h }, p1/Z, [x23]\n"
67       "ld1h { z16.h }, p1/Z, [x19]\n"
68       "zip1 z22.h, z17.h, z16.h\n"
69       "zip2 z21.h, z17.h, z16.h\n"
70       "ld1h { z17.h }, p1/Z, [x25, #1, MUL VL]\n"
71       "addvl x25, x25, #2\n"
72       "ld1h { z16.h }, p1/Z, [x24, #1, MUL VL]\n"
73       "zip1 z20.h, z17.h, z16.h\n"
74       "addvl x24, x24, #2\n"
75       "zip2 z19.h, z17.h, z16.h\n"
76       "ld1h { z18.h }, p1/Z, [x23, #1, MUL VL]\n"
77       "addvl x23, x23, #2\n"
78       "ld1h { z16.h }, p1/Z, [x19, #1, MUL VL]\n"
79       "st1h { z24.h }, p1, [x21]\n"
80       "zip1 z17.h, z18.h, z16.h\n"
81       "addvl x19, x19, #2\n"
82       "st1h { z22.h }, p1, [x21, #1, MUL VL]\n"
83       "add x21, x21, %x[out_stride]\n"
84       "zip2 z16.h, z18.h, z16.h\n"
85       "st1h { z23.h }, p1, [x21]\n"
86       "st1h { z21.h }, p1, [x21, #1, MUL VL]\n"
87       "add x21, x21, %x[out_stride]\n"
88       "st1h { z20.h }, p1, [x21]\n"
89       "st1h { z17.h }, p1, [x21, #1, MUL VL]\n"
90       "add x21, x21, %x[out_stride]\n"
91       "st1h { z19.h }, p1, [x21]\n"
92       "st1h { z16.h }, p1, [x21, #1, MUL VL]\n"
93       "add x21, x21, %x[out_stride]\n"
94       "bge 2b\n"
95       "3:"  // Main row loop: Unroll column loop skip
96       "cbz x22, 5f\n"
97       "4:"  // Main row loop: Column loop
98       "whilelt p0.h, XZR, x22\n"
99       "ld1h { z17.h }, p0/Z, [x25]\n"
100       "decw x22\n"
101       "ld1h { z16.h }, p0/Z, [x24]\n"
102       "cmp x22, #0x0\n"
103       "incd x25, ALL, MUL #4\n"
104       "zip1 z18.h, z17.h, z16.h\n"
105       "ld1h { z17.h }, p0/Z, [x23]\n"
106       "incd x24, ALL, MUL #4\n"
107       "incd x23, ALL, MUL #4\n"
108       "ld1h { z16.h }, p0/Z, [x19]\n"
109       "incd x19, ALL, MUL #4\n"
110       "zip1 z16.h, z17.h, z16.h\n"
111       "st1h { z18.h }, p1, [x21]\n"
112       "st1h { z16.h }, p1, [x21, #1, MUL VL]\n"
113       "add x21, x21, %x[out_stride]\n"
114       "bgt 4b\n"
115       "5:"  // Main row loop: Column loop skip
116       "cmp %x[height], #0x4\n"
117       "addvl %x[out], %x[out], #2\n"
118       "bge 1b\n"
119       "cbz %x[height], 12f\n"
120       "6:"  // Main loop skip
121       "7:"  // Tail row loop: Head
122       "mov x25, %x[in]\n"
123       "add x24, x25, %x[in_stride]\n"
124       "cmp %x[height], #0x1\n"
125       "mov x20, %x[width]\n"
126       "cnth x19, ALL, MUL #2\n"
127       "add %x[in], x24, %x[in_stride]\n"
128       "csel x24, x24, %x[pad_row], GT\n"
129       "cmp x20, x19\n"
130       "mov x21, %x[out]\n"
131       "sub %x[height], %x[height], #0x2\n"
132       "blt 9f\n"
133       "8:"  // Tail row loop: Unroll column loop
134       "ld1h { z18.h }, p1/Z, [x25]\n"
135       "sub x20, x20, x19\n"
136       "cmp x20, x19\n"
137       "ld1h { z16.h }, p1/Z, [x24]\n"
138       "zip1 z17.h, z18.h, z16.h\n"
139       "zip2 z19.h, z18.h, z16.h\n"
140       "ld1h { z18.h }, p1/Z, [x25, #1, MUL VL]\n"
141       "addvl x25, x25, #2\n"
142       "ld1h { z16.h }, p1/Z, [x24, #1, MUL VL]\n"
143       "st1h { z17.h }, p1, [x21]\n"
144       "add x21, x21, %x[out_stride]\n"
145       "zip1 z17.h, z18.h, z16.h\n"
146       "st1h { z19.h }, p1, [x21]\n"
147       "add x21, x21, %x[out_stride]\n"
148       "addvl x24, x24, #2\n"
149       "zip2 z16.h, z18.h, z16.h\n"
150       "st1h { z17.h }, p1, [x21]\n"
151       "add x21, x21, %x[out_stride]\n"
152       "st1h { z16.h }, p1, [x21]\n"
153       "add x21, x21, %x[out_stride]\n"
154       "bge 8b\n"
155       "9:"  // Tail row loop: Unroll column loop skip
156       "cbz x20, 11f\n"
157       "10:"  // Tail row loop: Column loop
158       "whilelt p0.h, XZR, x20\n"
159       "ld1h { z17.h }, p0/Z, [x25]\n"
160       "decw x20\n"
161       "ld1h { z16.h }, p0/Z, [x24]\n"
162       "cmp x20, #0x0\n"
163       "incd x25, ALL, MUL #4\n"
164       "zip1 z16.h, z17.h, z16.h\n"
165       "incd x24, ALL, MUL #4\n"
166       "st1h { z16.h }, p1, [x21]\n"
167       "add x21, x21, %x[out_stride]\n"
168       "bgt 10b\n"
169       "11:"  // Tail row loop: Column loop skip
170       "cmp %x[height], #0x1\n"
171       "addvl %x[out], %x[out], #1\n"
172       "bge 7b\n"
173       "12:"  // Done
174       ".inst 0xd503467f  // SMSTOP\n"
175       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
176       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
177       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "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"
178     );
179 }
180 
181 } // anonymous namespace
182 
183 template<>
Transform(bfloat16 * out,const bfloat16 * in,int stride,int x0,int xmax,int k0,int kmax)184 void Transform<1, 2, true, VLType::SME>(
185     bfloat16 *out, const bfloat16 *in, int stride, int x0, int xmax, int k0, int kmax)
186 {
187     sme_transpose_interleave_1VL_2x2(
188         reinterpret_cast<uint16_t *>(out),
189         reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
190         (xmax-x0) * sizeof(bfloat16) / 2,
191         stride * sizeof(bfloat16),
192         (kmax-k0)
193     );
194 }
195 
196 #endif
197