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_2x2_fp32bf16(bfloat16 * out,const float * in,size_t width,size_t in_stride,size_t height)32 void sme_transpose_interleave_2VL_2x2_fp32bf16(bfloat16 *out, const float *in, size_t width, size_t in_stride, size_t height)
33 {
34 float *pad_row = reinterpret_cast<float *>(alloca(width * sizeof(float)));
35
36 if (height % 2) {
37 memset(pad_row, 0, width * sizeof(float));
38 }
39
40 size_t out_stride = 2 * 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 p2.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 x19, ALL, MUL #2\n"
53 "add x20, x23, %x[in_stride]\n"
54 "cmp x22, x19\n"
55 "add %x[in], x20, %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 "ld1w { z16.s }, p2/Z, [x25]\n"
61 ".inst 0x658aaa18 // bfcvt z24.h, p2/M, z16.s\n"
62 "sub x22, x22, x19\n"
63 "cmp x22, x19\n"
64 "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
65 ".inst 0x658aaa17 // bfcvt z23.h, p2/M, z16.s\n"
66 "ld1w { z16.s }, p2/Z, [x23]\n"
67 ".inst 0x658aaa16 // bfcvt z22.h, p2/M, z16.s\n"
68 "ld1w { z16.s }, p2/Z, [x23, #1, MUL VL]\n"
69 ".inst 0x658aaa15 // bfcvt z21.h, p2/M, z16.s\n"
70 "ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n"
71 ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n"
72 "ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n"
73 ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n"
74 "addvl x25, x25, #4\n"
75 "ld1w { z16.s }, p2/Z, [x23, #2, MUL VL]\n"
76 ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
77 "ld1w { z16.s }, p2/Z, [x23, #3, MUL VL]\n"
78 ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
79 "addvl x23, x23, #4\n"
80 "ld1w { z16.s }, p2/Z, [x24]\n"
81 ".inst 0x648aaa18 // bfcvtnt z24.h, p2/M, z16.s\n"
82 "ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n"
83 ".inst 0x648aaa17 // bfcvtnt z23.h, p2/M, z16.s\n"
84 "ld1w { z16.s }, p2/Z, [x20]\n"
85 ".inst 0x648aaa16 // bfcvtnt z22.h, p2/M, z16.s\n"
86 "ld1w { z16.s }, p2/Z, [x20, #1, MUL VL]\n"
87 ".inst 0x648aaa15 // bfcvtnt z21.h, p2/M, z16.s\n"
88 "ld1w { z16.s }, p2/Z, [x24, #2, MUL VL]\n"
89 ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n"
90 "ld1w { z16.s }, p2/Z, [x24, #3, MUL VL]\n"
91 "addvl x24, x24, #4\n"
92 ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n"
93 "ld1w { z16.s }, p2/Z, [x20, #2, MUL VL]\n"
94 ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
95 "ld1w { z16.s }, p2/Z, [x20, #3, MUL VL]\n"
96 "st1h { z24.h }, p2, [x21]\n"
97 "addvl x20, x20, #4\n"
98 ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
99 "st1h { z23.h }, p2, [x21, #1, MUL VL]\n"
100 "st1h { z22.h }, p2, [x21, #2, MUL VL]\n"
101 "st1h { z21.h }, p2, [x21, #3, MUL VL]\n"
102 "add x21, x21, %x[out_stride]\n"
103 "st1h { z20.h }, p2, [x21]\n"
104 "st1h { z19.h }, p2, [x21, #1, MUL VL]\n"
105 "st1h { z18.h }, p2, [x21, #2, MUL VL]\n"
106 "st1h { z17.h }, p2, [x21, #3, MUL VL]\n"
107 "add x21, x21, %x[out_stride]\n"
108 "bge 2b\n"
109 "3:" // Main row loop: Unroll column loop skip
110 "cbz x22, 5f\n"
111 "4:" // Main row loop: Column loop
112 "mov x19, x22\n"
113 "whilelt p1.s, XZR, x19\n"
114 "ld1w { z16.s }, p1/Z, [x25]\n"
115 ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n"
116 "decw x19\n"
117 "whilelt p0.s, XZR, x19\n"
118 "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
119 ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n"
120 "ld1w { z16.s }, p1/Z, [x23]\n"
121 ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
122 "decw x22, ALL, MUL #2\n"
123 "cmp x22, #0x0\n"
124 "ld1w { z16.s }, p0/Z, [x23, #1, MUL VL]\n"
125 ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
126 "addvl x25, x25, #2\n"
127 "addvl x23, x23, #2\n"
128 "ld1w { z16.s }, p1/Z, [x24]\n"
129 ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n"
130 "ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n"
131 "addvl x24, x24, #2\n"
132 ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n"
133 "ld1w { z16.s }, p1/Z, [x20]\n"
134 ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
135 "ld1w { z16.s }, p0/Z, [x20, #1, MUL VL]\n"
136 "addvl x20, x20, #2\n"
137 ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
138 "st1h { z20.h }, p2, [x21]\n"
139 "st1h { z19.h }, p2, [x21, #1, MUL VL]\n"
140 "st1h { z18.h }, p2, [x21, #2, MUL VL]\n"
141 "st1h { z17.h }, p2, [x21, #3, MUL VL]\n"
142 "add x21, x21, %x[out_stride]\n"
143 "bgt 4b\n"
144 "5:" // Main row loop: Column loop skip
145 "cmp %x[height], #0x4\n"
146 "addvl %x[out], %x[out], #4\n"
147 "bge 1b\n"
148 "cbz %x[height], 12f\n"
149 "6:" // Main loop skip
150 "7:" // Tail row loop: Head
151 "mov x25, %x[in]\n"
152 "add x24, x25, %x[in_stride]\n"
153 "cmp %x[height], #0x1\n"
154 "mov x20, %x[width]\n"
155 "cnth x19, ALL, MUL #2\n"
156 "add %x[in], x24, %x[in_stride]\n"
157 "csel x24, x24, %x[pad_row], GT\n"
158 "cmp x20, x19\n"
159 "mov x21, %x[out]\n"
160 "sub %x[height], %x[height], #0x2\n"
161 "blt 9f\n"
162 "8:" // Tail row loop: Unroll column loop
163 "ld1w { z16.s }, p2/Z, [x25]\n"
164 ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n"
165 "sub x20, x20, x19\n"
166 "cmp x20, x19\n"
167 "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
168 ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n"
169 "ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n"
170 ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
171 "ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n"
172 ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
173 "addvl x25, x25, #4\n"
174 "ld1w { z16.s }, p2/Z, [x24]\n"
175 ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n"
176 "ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n"
177 ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n"
178 "ld1w { z16.s }, p2/Z, [x24, #2, MUL VL]\n"
179 ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
180 "ld1w { z16.s }, p2/Z, [x24, #3, MUL VL]\n"
181 "st1h { z20.h }, p2, [x21]\n"
182 "addvl x24, x24, #4\n"
183 ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
184 "st1h { z19.h }, p2, [x21, #1, MUL VL]\n"
185 "add x21, x21, %x[out_stride]\n"
186 "st1h { z18.h }, p2, [x21]\n"
187 "st1h { z17.h }, p2, [x21, #1, MUL VL]\n"
188 "add x21, x21, %x[out_stride]\n"
189 "bge 8b\n"
190 "9:" // Tail row loop: Unroll column loop skip
191 "cbz x20, 11f\n"
192 "10:" // Tail row loop: Column loop
193 "mov x19, x20\n"
194 "whilelt p1.s, XZR, x19\n"
195 "ld1w { z16.s }, p1/Z, [x25]\n"
196 ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
197 "decw x19\n"
198 "whilelt p0.s, XZR, x19\n"
199 "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
200 ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
201 "ld1w { z16.s }, p1/Z, [x24]\n"
202 "decw x20, ALL, MUL #2\n"
203 "cmp x20, #0x0\n"
204 ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
205 "ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n"
206 "addvl x25, x25, #2\n"
207 "addvl x24, x24, #2\n"
208 ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
209 "st1h { z18.h }, p2, [x21]\n"
210 "st1h { z17.h }, p2, [x21, #1, MUL VL]\n"
211 "add x21, x21, %x[out_stride]\n"
212 "bgt 10b\n"
213 "11:" // Tail row loop: Column loop skip
214 "cmp %x[height], #0x1\n"
215 "addvl %x[out], %x[out], #2\n"
216 "bge 7b\n"
217 "12:" // Done
218 ".inst 0xd503467f // SMSTOP\n"
219 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
220 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
221 : "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"
222 );
223 }
224
225 } // anonymous namespace
226 template<>
Transform(bfloat16 * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)227 void Transform<2, 2, true, VLType::SME>(
228 bfloat16 *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
229 {
230 sme_transpose_interleave_2VL_2x2_fp32bf16(
231 out,
232 in + k0 * stride + x0,
233 (xmax-x0),
234 stride * sizeof(float),
235 (kmax-k0)
236 );
237 }
238
239 #endif
240