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_1x4(uint8_t * out,const uint8_t * in,size_t width,size_t in_stride,size_t height)32 void sme_transpose_interleave_1VL_1x4(uint8_t *out, const uint8_t *in, size_t width, size_t in_stride, size_t height)
33 {
34     uint8_t *pad_row = reinterpret_cast<uint8_t *>(alloca(width * sizeof(uint8_t)));
35 
36     if (height % 4) {
37         memset(pad_row, 0, width * sizeof(uint8_t));
38     }
39 
40     size_t out_stride = 1 * roundup<size_t>(height, 4) * sme::get_vector_length<uint32_t>();
41 
42     __asm__ __volatile__(
43       ".inst 0xd503477f  // SMSTART ZA\n"
44       "ptrue p1.b\n"
45       "1:"  // Main row loop: Head
46       "mov x25, %x[in]\n"
47       "add x24, x25, %x[in_stride]\n"
48       "add x23, x24, %x[in_stride]\n"
49       "add x22, x23, %x[in_stride]\n"
50       "cmp %x[height], #0x3\n"
51       "add %x[in], x22, %x[in_stride]\n"
52       "csel x22, x22, %x[pad_row], GT\n"
53       "csel x23, x23, %x[pad_row], GE\n"
54       "cmp %x[height], #0x1\n"
55       "mov x21, %x[width]\n"
56       "cntb x20\n"
57       "csel x24, x24, %x[pad_row], GT\n"
58       "cmp x21, x20\n"
59       "mov x19, %x[out]\n"
60       "sub %x[height], %x[height], #0x4\n"
61       "blt 3f\n"
62       "2:"  // Main row loop: Unroll column loop
63       "ld1b { z17.b }, p1/Z, [x25]\n"
64       "sub x21, x21, x20\n"
65       "cmp x21, x20\n"
66       "ld1b { z18.b }, p1/Z, [x24]\n"
67       "addvl x25, x25, #1\n"
68       "addvl x24, x24, #1\n"
69       "ld1b { z16.b }, p1/Z, [x23]\n"
70       "zip1 z20.b, z17.b, z16.b\n"
71       "zip2 z19.b, z17.b, z16.b\n"
72       "addvl x23, x23, #1\n"
73       "ld1b { z16.b }, p1/Z, [x22]\n"
74       "zip1 z17.b, z18.b, z16.b\n"
75       "zip2 z18.b, z18.b, z16.b\n"
76       "addvl x22, x22, #1\n"
77       "zip1 z16.b, z20.b, z17.b\n"
78       "st1b { z16.b }, p1, [x19]\n"
79       "add x19, x19, %x[out_stride]\n"
80       "zip2 z16.b, z20.b, z17.b\n"
81       "st1b { z16.b }, p1, [x19]\n"
82       "add x19, x19, %x[out_stride]\n"
83       "zip1 z17.b, z19.b, z18.b\n"
84       "zip2 z16.b, z19.b, z18.b\n"
85       "st1b { z17.b }, p1, [x19]\n"
86       "add x19, x19, %x[out_stride]\n"
87       "st1b { z16.b }, p1, [x19]\n"
88       "add x19, x19, %x[out_stride]\n"
89       "bge 2b\n"
90       "3:"  // Main row loop: Unroll column loop skip
91       "cbz x21, 5f\n"
92       "4:"  // Main row loop: Column loop
93       "whilelt p0.b, XZR, x21\n"
94       "ld1b { z17.b }, p0/Z, [x25]\n"
95       "decw x21\n"
96       "ld1b { z18.b }, p0/Z, [x24]\n"
97       "cmp x21, #0x0\n"
98       "incd x25, ALL, MUL #2\n"
99       "ld1b { z16.b }, p0/Z, [x23]\n"
100       "zip1 z17.b, z17.b, z16.b\n"
101       "incd x24, ALL, MUL #2\n"
102       "incd x23, ALL, MUL #2\n"
103       "ld1b { z16.b }, p0/Z, [x22]\n"
104       "zip1 z16.b, z18.b, z16.b\n"
105       "incd x22, ALL, MUL #2\n"
106       "zip1 z16.b, z17.b, z16.b\n"
107       "st1b { z16.b }, p1, [x19]\n"
108       "add x19, x19, %x[out_stride]\n"
109       "bgt 4b\n"
110       "5:"  // Main row loop: Column loop skip
111       "cmp %x[height], #0x1\n"
112       "addvl %x[out], %x[out], #1\n"
113       "bge 1b\n"
114       ".inst 0xd503467f  // SMSTOP\n"
115       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
116       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
117       : "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"
118     );
119 }
120 
121 } // anonymous namespace
122 
123 template<>
Transform(uint8_t * out,const uint8_t * in,int stride,int x0,int xmax,int k0,int kmax)124 void Transform<1, 4, true, VLType::SME>(
125     uint8_t *out, const uint8_t *in, int stride, int x0, int xmax, int k0, int kmax)
126 {
127     sme_transpose_interleave_1VL_1x4(
128         reinterpret_cast<uint8_t *>(out),
129         reinterpret_cast<const uint8_t *>(in + k0 * stride + x0),
130         (xmax-x0) * sizeof(uint8_t) / 1,
131         stride * sizeof(uint8_t),
132         (kmax-k0)
133     );
134 }
135 
136 template<>
Transform(int8_t * out,const int8_t * in,int stride,int x0,int xmax,int k0,int kmax)137 void Transform<1, 4, true, VLType::SME>(
138     int8_t *out, const int8_t *in, int stride, int x0, int xmax, int k0, int kmax)
139 {
140     sme_transpose_interleave_1VL_1x4(
141         reinterpret_cast<uint8_t *>(out),
142         reinterpret_cast<const uint8_t *>(in + k0 * stride + x0),
143         (xmax-x0) * sizeof(int8_t) / 1,
144         stride * sizeof(int8_t),
145         (kmax-k0)
146     );
147 }
148 
149 #endif
150