1 /*
2 * Copyright (c) 2021 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
sve_transpose_interleave_6VL_2x4_fp32bf16(bfloat16 * out,const float * in,size_t width,size_t in_stride,size_t height)32 void sve_transpose_interleave_6VL_2x4_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 % 4) {
37 memset(pad_row, 0, width * sizeof(float));
38 }
39
40 size_t out_stride = 6 * roundup<size_t>(height, 4) * get_vector_length<uint32_t>();
41
42 __asm__ __volatile__(
43 "ptrue p3.b\n"
44 "1:" // Main row loop: Head
45 "mov x25, %x[in]\n"
46 "add x24, x25, %x[in_stride]\n"
47 "add x23, x24, %x[in_stride]\n"
48 "mov x22, %x[width]\n"
49 "cnth x19, ALL, MUL #3\n"
50 "add x21, x23, %x[in_stride]\n"
51 "cmp %x[height], #0x3\n"
52 "add %x[in], x21, %x[in_stride]\n"
53 "csel x21, x21, %x[pad_row], GT\n"
54 "csel x23, x23, %x[pad_row], GE\n"
55 "cmp %x[height], #0x1\n"
56 "csel x24, x24, %x[pad_row], GT\n"
57 "cmp x22, x19\n"
58 "mov x20, %x[out]\n"
59 "sub %x[height], %x[height], #0x4\n"
60 "blt 3f\n"
61 "2:" // Main row loop: Unroll column loop
62 "ld1w { z17.s }, p3/Z, [x25]\n"
63 "ld1w { z18.s }, p3/Z, [x25, #1, MUL VL]\n"
64 "sub x22, x22, x19\n"
65 "cmp x22, x19\n"
66 "ld1w { z19.s }, p3/Z, [x25, #2, MUL VL]\n"
67 "ld1w { z16.s }, p3/Z, [x23]\n"
68 "zip1 z21.s, z17.s, z16.s\n"
69 "zip2 z20.s, z17.s, z16.s\n"
70 "ld1w { z17.s }, p3/Z, [x23, #1, MUL VL]\n"
71 "ld1w { z16.s }, p3/Z, [x23, #2, MUL VL]\n"
72 "zip1 z29.s, z18.s, z17.s\n"
73 "zip2 z28.s, z18.s, z17.s\n"
74 "ld1w { z17.s }, p3/Z, [x25, #3, MUL VL]\n"
75 "ld1w { z18.s }, p3/Z, [x25, #4, MUL VL]\n"
76 "zip1 z27.s, z19.s, z16.s\n"
77 "zip2 z26.s, z19.s, z16.s\n"
78 "ld1w { z19.s }, p3/Z, [x25, #5, MUL VL]\n"
79 "ld1w { z16.s }, p3/Z, [x23, #3, MUL VL]\n"
80 "zip1 z25.s, z17.s, z16.s\n"
81 "zip2 z24.s, z17.s, z16.s\n"
82 "ld1w { z17.s }, p3/Z, [x23, #4, MUL VL]\n"
83 "ld1w { z16.s }, p3/Z, [x23, #5, MUL VL]\n"
84 "zip1 z12.s, z18.s, z17.s\n"
85 "zip2 z11.s, z18.s, z17.s\n"
86 "ld1w { z18.s }, p3/Z, [x24]\n"
87 "ld1w { z23.s }, p3/Z, [x24, #1, MUL VL]\n"
88 "zip1 z10.s, z19.s, z16.s\n"
89 "zip2 z9.s, z19.s, z16.s\n"
90 "ld1w { z22.s }, p3/Z, [x24, #2, MUL VL]\n"
91 "ld1w { z17.s }, p3/Z, [x21]\n"
92 ".inst 0x658aaea8 // bfcvt z8.h, p3/M, z21.s\n"
93 "zip1 z7.s, z18.s, z17.s\n"
94 "ld1w { z16.s }, p3/Z, [x21, #1, MUL VL]\n"
95 "ld1w { z21.s }, p3/Z, [x21, #2, MUL VL]\n"
96 ".inst 0x658aae86 // bfcvt z6.h, p3/M, z20.s\n"
97 "zip2 z5.s, z18.s, z17.s\n"
98 "ld1w { z20.s }, p3/Z, [x24, #3, MUL VL]\n"
99 "ld1w { z19.s }, p3/Z, [x24, #4, MUL VL]\n"
100 ".inst 0x658aafa4 // bfcvt z4.h, p3/M, z29.s\n"
101 "zip1 z3.s, z23.s, z16.s\n"
102 "ld1w { z2.s }, p3/Z, [x24, #5, MUL VL]\n"
103 "ld1w { z18.s }, p3/Z, [x21, #3, MUL VL]\n"
104 ".inst 0x658aaf81 // bfcvt z1.h, p3/M, z28.s\n"
105 "zip2 z0.s, z23.s, z16.s\n"
106 "ld1w { z17.s }, p3/Z, [x21, #4, MUL VL]\n"
107 "ld1w { z16.s }, p3/Z, [x21, #5, MUL VL]\n"
108 ".inst 0x658aaf7f // bfcvt z31.h, p3/M, z27.s\n"
109 "zip1 z30.s, z22.s, z21.s\n"
110 ".inst 0x658aaf5d // bfcvt z29.h, p3/M, z26.s\n"
111 "zip2 z28.s, z22.s, z21.s\n"
112 "addvl x25, x25, #6\n"
113 "addvl x24, x24, #6\n"
114 ".inst 0x658aaf3b // bfcvt z27.h, p3/M, z25.s\n"
115 "zip1 z26.s, z20.s, z18.s\n"
116 "addvl x23, x23, #6\n"
117 "addvl x21, x21, #6\n"
118 ".inst 0x658aaf19 // bfcvt z25.h, p3/M, z24.s\n"
119 "zip2 z24.s, z20.s, z18.s\n"
120 ".inst 0x658aad97 // bfcvt z23.h, p3/M, z12.s\n"
121 "zip1 z22.s, z19.s, z17.s\n"
122 ".inst 0x658aad75 // bfcvt z21.h, p3/M, z11.s\n"
123 "zip2 z20.s, z19.s, z17.s\n"
124 ".inst 0x658aad53 // bfcvt z19.h, p3/M, z10.s\n"
125 "zip1 z18.s, z2.s, z16.s\n"
126 ".inst 0x658aad31 // bfcvt z17.h, p3/M, z9.s\n"
127 "zip2 z16.s, z2.s, z16.s\n"
128 ".inst 0x648aace8 // bfcvtnt z8.h, p3/M, z7.s\n"
129 ".inst 0x648aaca6 // bfcvtnt z6.h, p3/M, z5.s\n"
130 "st1h { z8.h }, p3, [x20]\n"
131 ".inst 0x648aac64 // bfcvtnt z4.h, p3/M, z3.s\n"
132 ".inst 0x648aac01 // bfcvtnt z1.h, p3/M, z0.s\n"
133 "st1h { z6.h }, p3, [x20, #1, MUL VL]\n"
134 ".inst 0x648aafdf // bfcvtnt z31.h, p3/M, z30.s\n"
135 ".inst 0x648aaf9d // bfcvtnt z29.h, p3/M, z28.s\n"
136 "st1h { z4.h }, p3, [x20, #2, MUL VL]\n"
137 "st1h { z1.h }, p3, [x20, #3, MUL VL]\n"
138 ".inst 0x648aaf5b // bfcvtnt z27.h, p3/M, z26.s\n"
139 ".inst 0x648aaf19 // bfcvtnt z25.h, p3/M, z24.s\n"
140 "st1h { z31.h }, p3, [x20, #4, MUL VL]\n"
141 ".inst 0x648aaed7 // bfcvtnt z23.h, p3/M, z22.s\n"
142 ".inst 0x648aae95 // bfcvtnt z21.h, p3/M, z20.s\n"
143 "st1h { z29.h }, p3, [x20, #5, MUL VL]\n"
144 "add x20, x20, %x[out_stride]\n"
145 ".inst 0x648aae53 // bfcvtnt z19.h, p3/M, z18.s\n"
146 ".inst 0x648aae11 // bfcvtnt z17.h, p3/M, z16.s\n"
147 "st1h { z27.h }, p3, [x20]\n"
148 "st1h { z25.h }, p3, [x20, #1, MUL VL]\n"
149 "st1h { z23.h }, p3, [x20, #2, MUL VL]\n"
150 "st1h { z21.h }, p3, [x20, #3, MUL VL]\n"
151 "st1h { z19.h }, p3, [x20, #4, MUL VL]\n"
152 "st1h { z17.h }, p3, [x20, #5, MUL VL]\n"
153 "add x20, x20, %x[out_stride]\n"
154 "bge 2b\n"
155 "3:" // Main row loop: Unroll column loop skip
156 "cbz x22, 5f\n"
157 "4:" // Main row loop: Column loop
158 "mov x19, x22\n"
159 "whilelt p2.s, XZR, x19\n"
160 "ld1w { z20.s }, p2/Z, [x25]\n"
161 "ld1w { z19.s }, p2/Z, [x23]\n"
162 "decw x19\n"
163 "whilelt p1.s, XZR, x19\n"
164 "ld1w { z18.s }, p1/Z, [x25, #1, MUL VL]\n"
165 "ld1w { z17.s }, p1/Z, [x23, #1, MUL VL]\n"
166 "decw x19\n"
167 "whilelt p0.s, XZR, x19\n"
168 "ld1w { z25.s }, p0/Z, [x25, #2, MUL VL]\n"
169 "ld1w { z16.s }, p0/Z, [x23, #2, MUL VL]\n"
170 "ld1w { z24.s }, p2/Z, [x24]\n"
171 "ld1w { z30.s }, p1/Z, [x24, #1, MUL VL]\n"
172 "zip1 z23.s, z20.s, z19.s\n"
173 "zip2 z22.s, z20.s, z19.s\n"
174 "ld1w { z29.s }, p0/Z, [x24, #2, MUL VL]\n"
175 "ld1w { z21.s }, p2/Z, [x21]\n"
176 "zip1 z20.s, z18.s, z17.s\n"
177 "zip2 z19.s, z18.s, z17.s\n"
178 "ld1w { z18.s }, p1/Z, [x21, #1, MUL VL]\n"
179 "ld1w { z28.s }, p0/Z, [x21, #2, MUL VL]\n"
180 "zip1 z17.s, z25.s, z16.s\n"
181 "zip2 z16.s, z25.s, z16.s\n"
182 "decd x22, ALL, MUL #6\n"
183 ".inst 0x658aaefb // bfcvt z27.h, p3/M, z23.s\n"
184 "zip1 z26.s, z24.s, z21.s\n"
185 "cmp x22, #0x0\n"
186 ".inst 0x658aaed9 // bfcvt z25.h, p3/M, z22.s\n"
187 "zip2 z24.s, z24.s, z21.s\n"
188 "addvl x25, x25, #3\n"
189 "addvl x24, x24, #3\n"
190 ".inst 0x658aae97 // bfcvt z23.h, p3/M, z20.s\n"
191 "zip1 z22.s, z30.s, z18.s\n"
192 "addvl x23, x23, #3\n"
193 "addvl x21, x21, #3\n"
194 ".inst 0x658aae75 // bfcvt z21.h, p3/M, z19.s\n"
195 "zip2 z20.s, z30.s, z18.s\n"
196 ".inst 0x658aae33 // bfcvt z19.h, p3/M, z17.s\n"
197 "zip1 z18.s, z29.s, z28.s\n"
198 ".inst 0x658aae11 // bfcvt z17.h, p3/M, z16.s\n"
199 "zip2 z16.s, z29.s, z28.s\n"
200 ".inst 0x648aaf5b // bfcvtnt z27.h, p3/M, z26.s\n"
201 ".inst 0x648aaf19 // bfcvtnt z25.h, p3/M, z24.s\n"
202 "st1h { z27.h }, p3, [x20]\n"
203 ".inst 0x648aaed7 // bfcvtnt z23.h, p3/M, z22.s\n"
204 ".inst 0x648aae95 // bfcvtnt z21.h, p3/M, z20.s\n"
205 "st1h { z25.h }, p3, [x20, #1, MUL VL]\n"
206 ".inst 0x648aae53 // bfcvtnt z19.h, p3/M, z18.s\n"
207 ".inst 0x648aae11 // bfcvtnt z17.h, p3/M, z16.s\n"
208 "st1h { z23.h }, p3, [x20, #2, MUL VL]\n"
209 "st1h { z21.h }, p3, [x20, #3, MUL VL]\n"
210 "st1h { z19.h }, p3, [x20, #4, MUL VL]\n"
211 "st1h { z17.h }, p3, [x20, #5, MUL VL]\n"
212 "add x20, x20, %x[out_stride]\n"
213 "bgt 4b\n"
214 "5:" // Main row loop: Column loop skip
215 "cmp %x[height], #0x1\n"
216 "addvl %x[out], %x[out], #6\n"
217 "bge 1b\n"
218 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
219 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
220 : "cc", "memory", "p0", "p1", "p2", "p3", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
221 );
222 }
223
224 } // anonymous namespace
225 template<>
Transform(bfloat16 * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)226 void Transform<6, 4, true, VLType::SVE>(
227 bfloat16 *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
228 {
229 sve_transpose_interleave_6VL_2x4_fp32bf16(
230 out,
231 in + k0 * stride + x0,
232 (xmax-x0),
233 stride * sizeof(float),
234 (kmax-k0)
235 );
236 }
237
238 #endif
239