1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
|
/*
* Copyright (c) 2022 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#pragma once
#ifdef __ARM_FEATURE_SVE
namespace {
void sme_transpose_interleave_2VL_2x2_fp32bf16(bfloat16 *out, const float *in, size_t width, size_t in_stride, size_t height)
{
float *pad_row = reinterpret_cast<float *>(alloca(width * sizeof(float)));
if (height % 2) {
memset(pad_row, 0, width * sizeof(float));
}
size_t out_stride = 2 * roundup<size_t>(height, 2) * sme::get_vector_length<uint16_t>();
__asm__ __volatile__(
".inst 0xd503477f // SMSTART ZA\n"
"cmp %x[height], #0x4\n"
"ptrue p2.b\n"
"blt 6f\n"
"1:" // Main row loop: Head
"mov x25, %x[in]\n"
"add x24, x25, %x[in_stride]\n"
"add x23, x24, %x[in_stride]\n"
"mov x22, %x[width]\n"
"cnth x19, ALL, MUL #2\n"
"add x20, x23, %x[in_stride]\n"
"cmp x22, x19\n"
"add %x[in], x20, %x[in_stride]\n"
"mov x21, %x[out]\n"
"sub %x[height], %x[height], #0x4\n"
"blt 3f\n"
"2:" // Main row loop: Unroll column loop
"ld1w { z16.s }, p2/Z, [x25]\n"
".inst 0x658aaa18 // bfcvt z24.h, p2/M, z16.s\n"
"sub x22, x22, x19\n"
"cmp x22, x19\n"
"ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
".inst 0x658aaa17 // bfcvt z23.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x23]\n"
".inst 0x658aaa16 // bfcvt z22.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x23, #1, MUL VL]\n"
".inst 0x658aaa15 // bfcvt z21.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n"
".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n"
".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n"
"addvl x25, x25, #4\n"
"ld1w { z16.s }, p2/Z, [x23, #2, MUL VL]\n"
".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x23, #3, MUL VL]\n"
".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
"addvl x23, x23, #4\n"
"ld1w { z16.s }, p2/Z, [x24]\n"
".inst 0x648aaa18 // bfcvtnt z24.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n"
".inst 0x648aaa17 // bfcvtnt z23.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x20]\n"
".inst 0x648aaa16 // bfcvtnt z22.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x20, #1, MUL VL]\n"
".inst 0x648aaa15 // bfcvtnt z21.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x24, #2, MUL VL]\n"
".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x24, #3, MUL VL]\n"
"addvl x24, x24, #4\n"
".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x20, #2, MUL VL]\n"
".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x20, #3, MUL VL]\n"
"st1h { z24.h }, p2, [x21]\n"
"addvl x20, x20, #4\n"
".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
"st1h { z23.h }, p2, [x21, #1, MUL VL]\n"
"st1h { z22.h }, p2, [x21, #2, MUL VL]\n"
"st1h { z21.h }, p2, [x21, #3, MUL VL]\n"
"add x21, x21, %x[out_stride]\n"
"st1h { z20.h }, p2, [x21]\n"
"st1h { z19.h }, p2, [x21, #1, MUL VL]\n"
"st1h { z18.h }, p2, [x21, #2, MUL VL]\n"
"st1h { z17.h }, p2, [x21, #3, MUL VL]\n"
"add x21, x21, %x[out_stride]\n"
"bge 2b\n"
"3:" // Main row loop: Unroll column loop skip
"cbz x22, 5f\n"
"4:" // Main row loop: Column loop
"mov x19, x22\n"
"whilelt p1.s, XZR, x19\n"
"ld1w { z16.s }, p1/Z, [x25]\n"
".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n"
"decw x19\n"
"whilelt p0.s, XZR, x19\n"
"ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n"
"ld1w { z16.s }, p1/Z, [x23]\n"
".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
"decw x22, ALL, MUL #2\n"
"cmp x22, #0x0\n"
"ld1w { z16.s }, p0/Z, [x23, #1, MUL VL]\n"
".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
"addvl x25, x25, #2\n"
"addvl x23, x23, #2\n"
"ld1w { z16.s }, p1/Z, [x24]\n"
".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n"
"ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n"
"addvl x24, x24, #2\n"
".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
"ld1w { z16.s }, p0/Z, [x20, #1, MUL VL]\n"
"addvl x20, x20, #2\n"
".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
"st1h { z20.h }, p2, [x21]\n"
"st1h { z19.h }, p2, [x21, #1, MUL VL]\n"
"st1h { z18.h }, p2, [x21, #2, MUL VL]\n"
"st1h { z17.h }, p2, [x21, #3, MUL VL]\n"
"add x21, x21, %x[out_stride]\n"
"bgt 4b\n"
"5:" // Main row loop: Column loop skip
"cmp %x[height], #0x4\n"
"addvl %x[out], %x[out], #4\n"
"bge 1b\n"
"cbz %x[height], 12f\n"
"6:" // Main loop skip
"7:" // Tail row loop: Head
"mov x25, %x[in]\n"
"add x24, x25, %x[in_stride]\n"
"cmp %x[height], #0x1\n"
"mov x20, %x[width]\n"
"cnth x19, ALL, MUL #2\n"
"add %x[in], x24, %x[in_stride]\n"
"csel x24, x24, %x[pad_row], GT\n"
"cmp x20, x19\n"
"mov x21, %x[out]\n"
"sub %x[height], %x[height], #0x2\n"
"blt 9f\n"
"8:" // Tail row loop: Unroll column loop
"ld1w { z16.s }, p2/Z, [x25]\n"
".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n"
"sub x20, x20, x19\n"
"cmp x20, x19\n"
"ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n"
".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n"
".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
"addvl x25, x25, #4\n"
"ld1w { z16.s }, p2/Z, [x24]\n"
".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n"
".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x24, #2, MUL VL]\n"
".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
"ld1w { z16.s }, p2/Z, [x24, #3, MUL VL]\n"
"st1h { z20.h }, p2, [x21]\n"
"addvl x24, x24, #4\n"
".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
"st1h { z19.h }, p2, [x21, #1, MUL VL]\n"
"add x21, x21, %x[out_stride]\n"
"st1h { z18.h }, p2, [x21]\n"
"st1h { z17.h }, p2, [x21, #1, MUL VL]\n"
"add x21, x21, %x[out_stride]\n"
"bge 8b\n"
"9:" // Tail row loop: Unroll column loop skip
"cbz x20, 11f\n"
"10:" // Tail row loop: Column loop
"mov x19, x20\n"
"whilelt p1.s, XZR, x19\n"
"ld1w { z16.s }, p1/Z, [x25]\n"
".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
"decw x19\n"
"whilelt p0.s, XZR, x19\n"
"ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
"ld1w { z16.s }, p1/Z, [x24]\n"
"decw x20, ALL, MUL #2\n"
"cmp x20, #0x0\n"
".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
"ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n"
"addvl x25, x25, #2\n"
"addvl x24, x24, #2\n"
".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
"st1h { z18.h }, p2, [x21]\n"
"st1h { z17.h }, p2, [x21, #1, MUL VL]\n"
"add x21, x21, %x[out_stride]\n"
"bgt 10b\n"
"11:" // Tail row loop: Column loop skip
"cmp %x[height], #0x1\n"
"addvl %x[out], %x[out], #2\n"
"bge 7b\n"
"12:" // Done
".inst 0xd503467f // SMSTOP\n"
: [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
: [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
: "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"
);
}
} // anonymous namespace
template<>
void Transform<2, 2, true, VLType::SME>(
bfloat16 *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
{
sme_transpose_interleave_2VL_2x2_fp32bf16(
out,
in + k0 * stride + x0,
(xmax-x0),
stride * sizeof(float),
(kmax-k0)
);
}
#endif
|
