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_2x2_fp32bf16(bfloat16 * out,const float * in,size_t width,size_t in_stride,size_t height)32 void sme_transpose_interleave_1VL_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 = 1 * 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 p1.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 x20, ALL, MUL #2\n"
53       "add x19, x23, %x[in_stride]\n"
54       "cmp x22, x20\n"
55       "add %x[in], x19, %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 }, p1/Z, [x25]\n"
61       ".inst 0x658aa618  // bfcvt z24.h, p1/M, z16.s\n"
62       "sub x22, x22, x20\n"
63       "cmp x22, x20\n"
64       "ld1w { z16.s }, p1/Z, [x23]\n"
65       ".inst 0x658aa617  // bfcvt z23.h, p1/M, z16.s\n"
66       "ld1w { z16.s }, p1/Z, [x25, #1, MUL VL]\n"
67       ".inst 0x658aa616  // bfcvt z22.h, p1/M, z16.s\n"
68       "ld1w { z16.s }, p1/Z, [x23, #1, MUL VL]\n"
69       ".inst 0x658aa615  // bfcvt z21.h, p1/M, z16.s\n"
70       "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n"
71       ".inst 0x658aa614  // bfcvt z20.h, p1/M, z16.s\n"
72       "ld1w { z16.s }, p1/Z, [x23, #2, MUL VL]\n"
73       ".inst 0x658aa613  // bfcvt z19.h, p1/M, z16.s\n"
74       "ld1w { z16.s }, p1/Z, [x25, #3, MUL VL]\n"
75       ".inst 0x658aa612  // bfcvt z18.h, p1/M, z16.s\n"
76       "addvl x25, x25, #4\n"
77       "ld1w { z16.s }, p1/Z, [x23, #3, MUL VL]\n"
78       ".inst 0x658aa611  // bfcvt z17.h, p1/M, z16.s\n"
79       "addvl x23, x23, #4\n"
80       "ld1w { z16.s }, p1/Z, [x24]\n"
81       ".inst 0x648aa618  // bfcvtnt z24.h, p1/M, z16.s\n"
82       "ld1w { z16.s }, p1/Z, [x19]\n"
83       ".inst 0x648aa617  // bfcvtnt z23.h, p1/M, z16.s\n"
84       "ld1w { z16.s }, p1/Z, [x24, #1, MUL VL]\n"
85       ".inst 0x648aa616  // bfcvtnt z22.h, p1/M, z16.s\n"
86       "ld1w { z16.s }, p1/Z, [x19, #1, MUL VL]\n"
87       ".inst 0x648aa615  // bfcvtnt z21.h, p1/M, z16.s\n"
88       "ld1w { z16.s }, p1/Z, [x24, #2, MUL VL]\n"
89       ".inst 0x648aa614  // bfcvtnt z20.h, p1/M, z16.s\n"
90       "ld1w { z16.s }, p1/Z, [x19, #2, MUL VL]\n"
91       ".inst 0x648aa613  // bfcvtnt z19.h, p1/M, z16.s\n"
92       "ld1w { z16.s }, p1/Z, [x24, #3, MUL VL]\n"
93       "addvl x24, x24, #4\n"
94       ".inst 0x648aa612  // bfcvtnt z18.h, p1/M, z16.s\n"
95       "ld1w { z16.s }, p1/Z, [x19, #3, MUL VL]\n"
96       "st1h { z24.h }, p1, [x21]\n"
97       "addvl x19, x19, #4\n"
98       ".inst 0x648aa611  // bfcvtnt z17.h, p1/M, z16.s\n"
99       "st1h { z23.h }, p1, [x21, #1, MUL VL]\n"
100       "add x21, x21, %x[out_stride]\n"
101       "st1h { z22.h }, p1, [x21]\n"
102       "st1h { z21.h }, p1, [x21, #1, MUL VL]\n"
103       "add x21, x21, %x[out_stride]\n"
104       "st1h { z20.h }, p1, [x21]\n"
105       "st1h { z19.h }, p1, [x21, #1, MUL VL]\n"
106       "add x21, x21, %x[out_stride]\n"
107       "st1h { z18.h }, p1, [x21]\n"
108       "st1h { z17.h }, p1, [x21, #1, MUL VL]\n"
109       "add x21, x21, %x[out_stride]\n"
110       "bge 2b\n"
111       "3:"  // Main row loop: Unroll column loop skip
112       "cbz x22, 5f\n"
113       "4:"  // Main row loop: Column loop
114       "whilelt p0.s, XZR, x22\n"
115       "ld1w { z16.s }, p0/Z, [x25]\n"
116       ".inst 0x658aa612  // bfcvt z18.h, p1/M, z16.s\n"
117       "decw x22\n"
118       "ld1w { z16.s }, p0/Z, [x23]\n"
119       ".inst 0x658aa611  // bfcvt z17.h, p1/M, z16.s\n"
120       "cmp x22, #0x0\n"
121       "addvl x25, x25, #1\n"
122       "ld1w { z16.s }, p0/Z, [x24]\n"
123       "addvl x24, x24, #1\n"
124       "addvl x23, x23, #1\n"
125       ".inst 0x648aa612  // bfcvtnt z18.h, p1/M, z16.s\n"
126       "ld1w { z16.s }, p0/Z, [x19]\n"
127       "addvl x19, x19, #1\n"
128       ".inst 0x648aa611  // bfcvtnt z17.h, p1/M, z16.s\n"
129       "st1h { z18.h }, p1, [x21]\n"
130       "st1h { z17.h }, p1, [x21, #1, MUL VL]\n"
131       "add x21, x21, %x[out_stride]\n"
132       "bgt 4b\n"
133       "5:"  // Main row loop: Column loop skip
134       "cmp %x[height], #0x4\n"
135       "addvl %x[out], %x[out], #2\n"
136       "bge 1b\n"
137       "cbz %x[height], 12f\n"
138       "6:"  // Main loop skip
139       "7:"  // Tail row loop: Head
140       "mov x25, %x[in]\n"
141       "add x24, x25, %x[in_stride]\n"
142       "cmp %x[height], #0x1\n"
143       "mov x20, %x[width]\n"
144       "cnth x19, ALL, MUL #2\n"
145       "add %x[in], x24, %x[in_stride]\n"
146       "csel x24, x24, %x[pad_row], GT\n"
147       "cmp x20, x19\n"
148       "mov x21, %x[out]\n"
149       "sub %x[height], %x[height], #0x2\n"
150       "blt 9f\n"
151       "8:"  // Tail row loop: Unroll column loop
152       "ld1w { z16.s }, p1/Z, [x25]\n"
153       ".inst 0x658aa614  // bfcvt z20.h, p1/M, z16.s\n"
154       "sub x20, x20, x19\n"
155       "cmp x20, x19\n"
156       "ld1w { z16.s }, p1/Z, [x25, #1, MUL VL]\n"
157       ".inst 0x658aa613  // bfcvt z19.h, p1/M, z16.s\n"
158       "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n"
159       ".inst 0x658aa612  // bfcvt z18.h, p1/M, z16.s\n"
160       "ld1w { z16.s }, p1/Z, [x25, #3, MUL VL]\n"
161       ".inst 0x658aa611  // bfcvt z17.h, p1/M, z16.s\n"
162       "addvl x25, x25, #4\n"
163       "ld1w { z16.s }, p1/Z, [x24]\n"
164       ".inst 0x648aa614  // bfcvtnt z20.h, p1/M, z16.s\n"
165       "ld1w { z16.s }, p1/Z, [x24, #1, MUL VL]\n"
166       ".inst 0x648aa613  // bfcvtnt z19.h, p1/M, z16.s\n"
167       "ld1w { z16.s }, p1/Z, [x24, #2, MUL VL]\n"
168       ".inst 0x648aa612  // bfcvtnt z18.h, p1/M, z16.s\n"
169       "ld1w { z16.s }, p1/Z, [x24, #3, MUL VL]\n"
170       "st1h { z20.h }, p1, [x21]\n"
171       "add x21, x21, %x[out_stride]\n"
172       "addvl x24, x24, #4\n"
173       "st1h { z19.h }, p1, [x21]\n"
174       "add x21, x21, %x[out_stride]\n"
175       ".inst 0x648aa611  // bfcvtnt z17.h, p1/M, z16.s\n"
176       "st1h { z18.h }, p1, [x21]\n"
177       "add x21, x21, %x[out_stride]\n"
178       "st1h { z17.h }, p1, [x21]\n"
179       "add x21, x21, %x[out_stride]\n"
180       "bge 8b\n"
181       "9:"  // Tail row loop: Unroll column loop skip
182       "cbz x20, 11f\n"
183       "10:"  // Tail row loop: Column loop
184       "whilelt p0.s, XZR, x20\n"
185       "ld1w { z16.s }, p0/Z, [x25]\n"
186       ".inst 0x658aa611  // bfcvt z17.h, p1/M, z16.s\n"
187       "decw x20\n"
188       "ld1w { z16.s }, p0/Z, [x24]\n"
189       "cmp x20, #0x0\n"
190       "addvl x25, x25, #1\n"
191       ".inst 0x648aa611  // bfcvtnt z17.h, p1/M, z16.s\n"
192       "addvl x24, x24, #1\n"
193       "st1h { z17.h }, p1, [x21]\n"
194       "add x21, x21, %x[out_stride]\n"
195       "bgt 10b\n"
196       "11:"  // Tail row loop: Column loop skip
197       "cmp %x[height], #0x1\n"
198       "addvl %x[out], %x[out], #1\n"
199       "bge 7b\n"
200       "12:"  // Done
201       ".inst 0xd503467f  // SMSTOP\n"
202       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
203       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
204       : "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"
205     );
206 }
207 
208 } // anonymous namespace
209 template<>
Transform(bfloat16 * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)210 void Transform<1, 2, true, VLType::SME>(
211     bfloat16 *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
212 {
213     sme_transpose_interleave_1VL_2x2_fp32bf16(
214         out,
215         in + k0 * stride + x0,
216         (xmax-x0),
217         stride * sizeof(float),
218         (kmax-k0)
219     );
220 }
221 
222 #endif
223