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(uint16_t * out,const uint16_t * in,size_t width,size_t in_stride,size_t height)32 void sme_transpose_interleave_1VL(uint16_t *out, const uint16_t *in, size_t width, size_t in_stride, size_t height)
33 {
34 size_t out_stride = 1 * height * sme::get_vector_length<uint8_t>();
35
36 __asm__ __volatile__(
37 ".inst 0xd503477f // SMSTART ZA\n"
38 "cmp %x[height], #0x4\n"
39 "ptrue p1.b\n"
40 "blt 6f\n"
41 "1:" // Main row loop: Head
42 "mov x25, %x[in]\n"
43 "add x24, x25, %x[in_stride]\n"
44 "add x23, x24, %x[in_stride]\n"
45 "mov x22, %x[width]\n"
46 "cnth x20, ALL, MUL #4\n"
47 "add x19, x23, %x[in_stride]\n"
48 "cmp x22, x20\n"
49 "add %x[in], x19, %x[in_stride]\n"
50 "mov x21, %x[out]\n"
51 "sub %x[height], %x[height], #0x4\n"
52 "blt 3f\n"
53 "2:" // Main row loop: Unroll column loop
54 "ld1h { z31.h }, p1/Z, [x25]\n"
55 "sub x22, x22, x20\n"
56 "cmp x22, x20\n"
57 "ld1h { z30.h }, p1/Z, [x25, #1, MUL VL]\n"
58 "ld1h { z29.h }, p1/Z, [x25, #2, MUL VL]\n"
59 "ld1h { z28.h }, p1/Z, [x25, #3, MUL VL]\n"
60 "addvl x25, x25, #4\n"
61 "ld1h { z27.h }, p1/Z, [x24]\n"
62 "ld1h { z26.h }, p1/Z, [x24, #1, MUL VL]\n"
63 "ld1h { z25.h }, p1/Z, [x24, #2, MUL VL]\n"
64 "ld1h { z24.h }, p1/Z, [x24, #3, MUL VL]\n"
65 "addvl x24, x24, #4\n"
66 "ld1h { z23.h }, p1/Z, [x23]\n"
67 "ld1h { z22.h }, p1/Z, [x23, #1, MUL VL]\n"
68 "ld1h { z21.h }, p1/Z, [x23, #2, MUL VL]\n"
69 "ld1h { z20.h }, p1/Z, [x23, #3, MUL VL]\n"
70 "addvl x23, x23, #4\n"
71 "ld1h { z19.h }, p1/Z, [x19]\n"
72 "ld1h { z18.h }, p1/Z, [x19, #1, MUL VL]\n"
73 "ld1h { z17.h }, p1/Z, [x19, #2, MUL VL]\n"
74 "ld1h { z16.h }, p1/Z, [x19, #3, MUL VL]\n"
75 "st1h { z31.h }, p1, [x21]\n"
76 "addvl x19, x19, #4\n"
77 "st1h { z27.h }, p1, [x21, #1, MUL VL]\n"
78 "st1h { z23.h }, p1, [x21, #2, MUL VL]\n"
79 "st1h { z19.h }, p1, [x21, #3, MUL VL]\n"
80 "add x21, x21, %x[out_stride]\n"
81 "st1h { z30.h }, p1, [x21]\n"
82 "st1h { z26.h }, p1, [x21, #1, MUL VL]\n"
83 "st1h { z22.h }, p1, [x21, #2, MUL VL]\n"
84 "st1h { z18.h }, p1, [x21, #3, MUL VL]\n"
85 "add x21, x21, %x[out_stride]\n"
86 "st1h { z29.h }, p1, [x21]\n"
87 "st1h { z25.h }, p1, [x21, #1, MUL VL]\n"
88 "st1h { z21.h }, p1, [x21, #2, MUL VL]\n"
89 "st1h { z17.h }, p1, [x21, #3, MUL VL]\n"
90 "add x21, x21, %x[out_stride]\n"
91 "st1h { z28.h }, p1, [x21]\n"
92 "st1h { z24.h }, p1, [x21, #1, MUL VL]\n"
93 "st1h { z20.h }, p1, [x21, #2, MUL VL]\n"
94 "st1h { z16.h }, p1, [x21, #3, MUL VL]\n"
95 "add x21, x21, %x[out_stride]\n"
96 "bge 2b\n"
97 "3:" // Main row loop: Unroll column loop skip
98 "cbz x22, 5f\n"
99 "4:" // Main row loop: Column loop
100 "whilelt p0.h, XZR, x22\n"
101 "dech x22\n"
102 "ld1h { z19.h }, p0/Z, [x25]\n"
103 "cmp x22, #0x0\n"
104 "addvl x25, x25, #1\n"
105 "ld1h { z18.h }, p0/Z, [x24]\n"
106 "addvl x24, x24, #1\n"
107 "ld1h { z17.h }, p0/Z, [x23]\n"
108 "addvl x23, x23, #1\n"
109 "ld1h { z16.h }, p0/Z, [x19]\n"
110 "addvl x19, x19, #1\n"
111 "st1h { z19.h }, p1, [x21]\n"
112 "st1h { z18.h }, p1, [x21, #1, MUL VL]\n"
113 "st1h { z17.h }, p1, [x21, #2, MUL VL]\n"
114 "st1h { z16.h }, p1, [x21, #3, MUL VL]\n"
115 "add x21, x21, %x[out_stride]\n"
116 "bgt 4b\n"
117 "5:" // Main row loop: Column loop skip
118 "cmp %x[height], #0x4\n"
119 "addvl %x[out], %x[out], #4\n"
120 "bge 1b\n"
121 "cbz %x[height], 12f\n"
122 "6:" // Main loop skip
123 "7:" // Tail row loop: Head
124 "mov x20, %x[width]\n"
125 "cnth x19, ALL, MUL #4\n"
126 "mov x25, %x[in]\n"
127 "cmp x20, x19\n"
128 "add %x[in], x25, %x[in_stride]\n"
129 "mov x21, %x[out]\n"
130 "sub %x[height], %x[height], #0x1\n"
131 "blt 9f\n"
132 "8:" // Tail row loop: Unroll column loop
133 "ld1h { z19.h }, p1/Z, [x25]\n"
134 "sub x20, x20, x19\n"
135 "cmp x20, x19\n"
136 "ld1h { z18.h }, p1/Z, [x25, #1, MUL VL]\n"
137 "ld1h { z17.h }, p1/Z, [x25, #2, MUL VL]\n"
138 "ld1h { z16.h }, p1/Z, [x25, #3, MUL VL]\n"
139 "st1h { z19.h }, p1, [x21]\n"
140 "add x21, x21, %x[out_stride]\n"
141 "addvl x25, x25, #4\n"
142 "st1h { z18.h }, p1, [x21]\n"
143 "add x21, x21, %x[out_stride]\n"
144 "st1h { z17.h }, p1, [x21]\n"
145 "add x21, x21, %x[out_stride]\n"
146 "st1h { z16.h }, p1, [x21]\n"
147 "add x21, x21, %x[out_stride]\n"
148 "bge 8b\n"
149 "9:" // Tail row loop: Unroll column loop skip
150 "cbz x20, 11f\n"
151 "10:" // Tail row loop: Column loop
152 "whilelt p0.h, XZR, x20\n"
153 "dech x20\n"
154 "ld1h { z16.h }, p0/Z, [x25]\n"
155 "st1h { z16.h }, p1, [x21]\n"
156 "cmp x20, #0x0\n"
157 "addvl x25, x25, #1\n"
158 "add x21, x21, %x[out_stride]\n"
159 "bgt 10b\n"
160 "11:" // Tail row loop: Column loop skip
161 "cmp %x[height], #0x1\n"
162 "addvl %x[out], %x[out], #1\n"
163 "bge 7b\n"
164 "12:" // Done
165 ".inst 0xd503467f // SMSTOP\n"
166 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
167 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
168 : "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"
169 );
170 }
171
172 } // anonymous namespace
173
174 template<>
Transform(float * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)175 void Transform<1, 1, true, VLType::SME>(
176 float *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
177 {
178 sme_transpose_interleave_1VL(
179 reinterpret_cast<uint16_t *>(out),
180 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
181 (xmax-x0) * sizeof(float) / 2,
182 stride * sizeof(float),
183 (kmax-k0)
184 );
185 }
186
187 template<>
Transform(bfloat16 * out,const bfloat16 * in,int stride,int x0,int xmax,int k0,int kmax)188 void Transform<1, 1, true, VLType::SME>(
189 bfloat16 *out, const bfloat16 *in, int stride, int x0, int xmax, int k0, int kmax)
190 {
191 sme_transpose_interleave_1VL(
192 reinterpret_cast<uint16_t *>(out),
193 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
194 (xmax-x0) * sizeof(bfloat16) / 2,
195 stride * sizeof(bfloat16),
196 (kmax-k0)
197 );
198 }
199
200 template<>
Transform(__fp16 * out,const __fp16 * in,int stride,int x0,int xmax,int k0,int kmax)201 void Transform<1, 1, true, VLType::SME>(
202 __fp16 *out, const __fp16 *in, int stride, int x0, int xmax, int k0, int kmax)
203 {
204 sme_transpose_interleave_1VL(
205 reinterpret_cast<uint16_t *>(out),
206 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
207 (xmax-x0) * sizeof(__fp16) / 2,
208 stride * sizeof(__fp16),
209 (kmax-k0)
210 );
211 }
212
213 #endif
214