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_2VL(uint16_t * out,const uint16_t * in,size_t width,size_t in_stride,size_t height)32 void sme_transpose_interleave_2VL(uint16_t *out, const uint16_t *in, size_t width, size_t in_stride, size_t height)
33 {
34 size_t out_stride = 2 * 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 p2.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 x19, ALL, MUL #4\n"
47 "add x20, x23, %x[in_stride]\n"
48 "cmp x22, x19\n"
49 "add %x[in], x20, %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 "sub x22, x22, x19\n"
55 "ld1h { z31.h }, p2/Z, [x25]\n"
56 "cmp x22, x19\n"
57 "ld1h { z30.h }, p2/Z, [x25, #1, MUL VL]\n"
58 "ld1h { z29.h }, p2/Z, [x25, #2, MUL VL]\n"
59 "ld1h { z28.h }, p2/Z, [x25, #3, MUL VL]\n"
60 "addvl x25, x25, #4\n"
61 "ld1h { z27.h }, p2/Z, [x24]\n"
62 "ld1h { z26.h }, p2/Z, [x24, #1, MUL VL]\n"
63 "ld1h { z25.h }, p2/Z, [x24, #2, MUL VL]\n"
64 "ld1h { z24.h }, p2/Z, [x24, #3, MUL VL]\n"
65 "addvl x24, x24, #4\n"
66 "ld1h { z23.h }, p2/Z, [x23]\n"
67 "ld1h { z22.h }, p2/Z, [x23, #1, MUL VL]\n"
68 "ld1h { z21.h }, p2/Z, [x23, #2, MUL VL]\n"
69 "ld1h { z20.h }, p2/Z, [x23, #3, MUL VL]\n"
70 "addvl x23, x23, #4\n"
71 "ld1h { z19.h }, p2/Z, [x20]\n"
72 "ld1h { z18.h }, p2/Z, [x20, #1, MUL VL]\n"
73 "ld1h { z17.h }, p2/Z, [x20, #2, MUL VL]\n"
74 "ld1h { z16.h }, p2/Z, [x20, #3, MUL VL]\n"
75 "st1h { z31.h }, p2, [x21]\n"
76 "addvl x20, x20, #4\n"
77 "st1h { z30.h }, p2, [x21, #1, MUL VL]\n"
78 "st1h { z27.h }, p2, [x21, #2, MUL VL]\n"
79 "st1h { z26.h }, p2, [x21, #3, MUL VL]\n"
80 "st1h { z23.h }, p2, [x21, #4, MUL VL]\n"
81 "st1h { z22.h }, p2, [x21, #5, MUL VL]\n"
82 "st1h { z19.h }, p2, [x21, #6, MUL VL]\n"
83 "st1h { z18.h }, p2, [x21, #7, MUL VL]\n"
84 "add x21, x21, %x[out_stride]\n"
85 "st1h { z29.h }, p2, [x21]\n"
86 "st1h { z28.h }, p2, [x21, #1, MUL VL]\n"
87 "st1h { z25.h }, p2, [x21, #2, MUL VL]\n"
88 "st1h { z24.h }, p2, [x21, #3, MUL VL]\n"
89 "st1h { z21.h }, p2, [x21, #4, MUL VL]\n"
90 "st1h { z20.h }, p2, [x21, #5, MUL VL]\n"
91 "st1h { z17.h }, p2, [x21, #6, MUL VL]\n"
92 "st1h { z16.h }, p2, [x21, #7, 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 "mov x19, x22\n"
99 "whilelt p1.h, XZR, x19\n"
100 "ld1h { z23.h }, p1/Z, [x25]\n"
101 "dech x19\n"
102 "dech x22, ALL, MUL #2\n"
103 "ld1h { z22.h }, p1/Z, [x24]\n"
104 "whilelt p0.h, XZR, x19\n"
105 "cmp x22, #0x0\n"
106 "ld1h { z21.h }, p0/Z, [x25, #1, MUL VL]\n"
107 "addvl x25, x25, #2\n"
108 "ld1h { z20.h }, p0/Z, [x24, #1, MUL VL]\n"
109 "addvl x24, x24, #2\n"
110 "ld1h { z19.h }, p1/Z, [x23]\n"
111 "ld1h { z18.h }, p0/Z, [x23, #1, MUL VL]\n"
112 "addvl x23, x23, #2\n"
113 "ld1h { z17.h }, p1/Z, [x20]\n"
114 "ld1h { z16.h }, p0/Z, [x20, #1, MUL VL]\n"
115 "addvl x20, x20, #2\n"
116 "st1h { z23.h }, p2, [x21]\n"
117 "st1h { z21.h }, p2, [x21, #1, MUL VL]\n"
118 "st1h { z22.h }, p2, [x21, #2, MUL VL]\n"
119 "st1h { z20.h }, p2, [x21, #3, MUL VL]\n"
120 "st1h { z19.h }, p2, [x21, #4, MUL VL]\n"
121 "st1h { z18.h }, p2, [x21, #5, MUL VL]\n"
122 "st1h { z17.h }, p2, [x21, #6, MUL VL]\n"
123 "st1h { z16.h }, p2, [x21, #7, MUL VL]\n"
124 "add x21, x21, %x[out_stride]\n"
125 "bgt 4b\n"
126 "5:" // Main row loop: Column loop skip
127 "cmp %x[height], #0x4\n"
128 "addvl %x[out], %x[out], #8\n"
129 "bge 1b\n"
130 "cbz %x[height], 12f\n"
131 "6:" // Main loop skip
132 "7:" // Tail row loop: Head
133 "mov x20, %x[width]\n"
134 "cnth x19, ALL, MUL #4\n"
135 "mov x25, %x[in]\n"
136 "cmp x20, x19\n"
137 "add %x[in], x25, %x[in_stride]\n"
138 "mov x21, %x[out]\n"
139 "sub %x[height], %x[height], #0x1\n"
140 "blt 9f\n"
141 "8:" // Tail row loop: Unroll column loop
142 "sub x20, x20, x19\n"
143 "ld1h { z19.h }, p2/Z, [x25]\n"
144 "cmp x20, x19\n"
145 "ld1h { z18.h }, p2/Z, [x25, #1, MUL VL]\n"
146 "ld1h { z17.h }, p2/Z, [x25, #2, MUL VL]\n"
147 "ld1h { z16.h }, p2/Z, [x25, #3, MUL VL]\n"
148 "st1h { z19.h }, p2, [x21]\n"
149 "addvl x25, x25, #4\n"
150 "st1h { z18.h }, p2, [x21, #1, MUL VL]\n"
151 "add x21, x21, %x[out_stride]\n"
152 "st1h { z17.h }, p2, [x21]\n"
153 "st1h { z16.h }, p2, [x21, #1, MUL VL]\n"
154 "add x21, x21, %x[out_stride]\n"
155 "bge 8b\n"
156 "9:" // Tail row loop: Unroll column loop skip
157 "cbz x20, 11f\n"
158 "10:" // Tail row loop: Column loop
159 "mov x19, x20\n"
160 "whilelt p0.h, XZR, x19\n"
161 "ld1h { z17.h }, p0/Z, [x25]\n"
162 "dech x19\n"
163 "dech x20, ALL, MUL #2\n"
164 "whilelt p0.h, XZR, x19\n"
165 "cmp x20, #0x0\n"
166 "ld1h { z16.h }, p0/Z, [x25, #1, MUL VL]\n"
167 "st1h { z17.h }, p2, [x21]\n"
168 "addvl x25, x25, #2\n"
169 "st1h { z16.h }, p2, [x21, #1, MUL VL]\n"
170 "add x21, x21, %x[out_stride]\n"
171 "bgt 10b\n"
172 "11:" // Tail row loop: Column loop skip
173 "cmp %x[height], #0x1\n"
174 "addvl %x[out], %x[out], #2\n"
175 "bge 7b\n"
176 "12:" // Done
177 ".inst 0xd503467f // SMSTOP\n"
178 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
179 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
180 : "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"
181 );
182 }
183
184 } // anonymous namespace
185
186 template<>
Transform(float * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)187 void Transform<2, 1, true, VLType::SME>(
188 float *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
189 {
190 sme_transpose_interleave_2VL(
191 reinterpret_cast<uint16_t *>(out),
192 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
193 (xmax-x0) * sizeof(float) / 2,
194 stride * sizeof(float),
195 (kmax-k0)
196 );
197 }
198
199 template<>
Transform(bfloat16 * out,const bfloat16 * in,int stride,int x0,int xmax,int k0,int kmax)200 void Transform<2, 1, true, VLType::SME>(
201 bfloat16 *out, const bfloat16 *in, int stride, int x0, int xmax, int k0, int kmax)
202 {
203 sme_transpose_interleave_2VL(
204 reinterpret_cast<uint16_t *>(out),
205 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
206 (xmax-x0) * sizeof(bfloat16) / 2,
207 stride * sizeof(bfloat16),
208 (kmax-k0)
209 );
210 }
211
212 template<>
Transform(__fp16 * out,const __fp16 * in,int stride,int x0,int xmax,int k0,int kmax)213 void Transform<2, 1, true, VLType::SME>(
214 __fp16 *out, const __fp16 *in, int stride, int x0, int xmax, int k0, int kmax)
215 {
216 sme_transpose_interleave_2VL(
217 reinterpret_cast<uint16_t *>(out),
218 reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
219 (xmax-x0) * sizeof(__fp16) / 2,
220 stride * sizeof(__fp16),
221 (kmax-k0)
222 );
223 }
224
225 #endif
226