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_16VL_2x2_fp32bf16(bfloat16 * out,const float * in,size_t width,size_t in_stride,size_t height)32 void sme_transpose_interleave_16VL_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 = 16 * roundup<size_t>(height, 2) * sme::get_vector_length<uint16_t>();
41
42 __asm__ __volatile__(
43 ".inst 0xd503477f // SMSTART ZA\n"
44 "ptrue p7.b\n"
45 "1:" // Main row loop: Head
46 "mov x24, %x[in]\n"
47 "add x23, x24, %x[in_stride]\n"
48 "cmp %x[height], #0x1\n"
49 "add %x[in], x23, %x[in_stride]\n"
50 "mov x22, %x[out]\n"
51 "csel x23, x23, %x[pad_row], GT\n"
52 "sub %x[height], %x[height], #0x2\n"
53 "mov x21, %x[width]\n"
54 "2:" // Main row loop: Column loop
55 "mov x20, x21\n"
56 "whilelt p1.s, XZR, x20\n"
57 "ld1w { z16.s }, p1/Z, [x24]\n"
58 ".inst 0x658abe00 // bfcvt z0.h, p7/M, z16.s\n"
59 "decw x20\n"
60 "whilelt p0.s, XZR, x20\n"
61 "ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n"
62 ".inst 0x658abe1f // bfcvt z31.h, p7/M, z16.s\n"
63 "decw x20\n"
64 "whilelt p6.s, XZR, x20\n"
65 "ld1w { z16.s }, p6/Z, [x24, #2, MUL VL]\n"
66 ".inst 0x658abe1e // bfcvt z30.h, p7/M, z16.s\n"
67 "decw x20\n"
68 "whilelt p5.s, XZR, x20\n"
69 "ld1w { z16.s }, p5/Z, [x24, #3, MUL VL]\n"
70 ".inst 0x658abe1d // bfcvt z29.h, p7/M, z16.s\n"
71 "decw x20\n"
72 "whilelt p4.s, XZR, x20\n"
73 "ld1w { z16.s }, p4/Z, [x24, #4, MUL VL]\n"
74 ".inst 0x658abe1c // bfcvt z28.h, p7/M, z16.s\n"
75 "decw x20\n"
76 "whilelt p3.s, XZR, x20\n"
77 "ld1w { z16.s }, p3/Z, [x24, #5, MUL VL]\n"
78 ".inst 0x658abe1b // bfcvt z27.h, p7/M, z16.s\n"
79 "decw x20\n"
80 "whilelt p2.s, XZR, x20\n"
81 "ld1w { z16.s }, p2/Z, [x24, #6, MUL VL]\n"
82 ".inst 0x658abe1a // bfcvt z26.h, p7/M, z16.s\n"
83 "decw x20\n"
84 "ld1w { z16.s }, p1/Z, [x23]\n"
85 "whilelt p1.s, XZR, x20\n"
86 ".inst 0x648abe00 // bfcvtnt z0.h, p7/M, z16.s\n"
87 "decw x20\n"
88 "ld1w { z16.s }, p1/Z, [x24, #7, MUL VL]\n"
89 "addvl x24, x24, #16\n"
90 ".inst 0x658abe19 // bfcvt z25.h, p7/M, z16.s\n"
91 "ld1w { z16.s }, p0/Z, [x23, #1, MUL VL]\n"
92 "whilelt p0.s, XZR, x20\n"
93 "decw x20\n"
94 ".inst 0x648abe1f // bfcvtnt z31.h, p7/M, z16.s\n"
95 "ld1w { z16.s }, p0/Z, [x24, #-8, MUL VL]\n"
96 ".inst 0x658abe18 // bfcvt z24.h, p7/M, z16.s\n"
97 "mov x19, x22\n"
98 "decw x21, ALL, MUL #16\n"
99 "ld1w { z16.s }, p6/Z, [x23, #2, MUL VL]\n"
100 "whilelt p6.s, XZR, x20\n"
101 "decw x20\n"
102 ".inst 0x648abe1e // bfcvtnt z30.h, p7/M, z16.s\n"
103 "ld1w { z16.s }, p6/Z, [x24, #-7, MUL VL]\n"
104 ".inst 0x658abe17 // bfcvt z23.h, p7/M, z16.s\n"
105 "add x22, x22, %x[out_stride]\n"
106 "ld1w { z16.s }, p5/Z, [x23, #3, MUL VL]\n"
107 "whilelt p5.s, XZR, x20\n"
108 "decw x20\n"
109 ".inst 0x648abe1d // bfcvtnt z29.h, p7/M, z16.s\n"
110 "ld1w { z16.s }, p5/Z, [x24, #-6, MUL VL]\n"
111 ".inst 0x658abe16 // bfcvt z22.h, p7/M, z16.s\n"
112 "ld1w { z16.s }, p4/Z, [x23, #4, MUL VL]\n"
113 "whilelt p4.s, XZR, x20\n"
114 "decw x20\n"
115 ".inst 0x648abe1c // bfcvtnt z28.h, p7/M, z16.s\n"
116 "ld1w { z16.s }, p4/Z, [x24, #-5, MUL VL]\n"
117 ".inst 0x658abe15 // bfcvt z21.h, p7/M, z16.s\n"
118 "ld1w { z16.s }, p3/Z, [x23, #5, MUL VL]\n"
119 "whilelt p3.s, XZR, x20\n"
120 "decw x20\n"
121 ".inst 0x648abe1b // bfcvtnt z27.h, p7/M, z16.s\n"
122 "ld1w { z16.s }, p3/Z, [x24, #-4, MUL VL]\n"
123 ".inst 0x658abe14 // bfcvt z20.h, p7/M, z16.s\n"
124 "ld1w { z16.s }, p2/Z, [x23, #6, MUL VL]\n"
125 "whilelt p2.s, XZR, x20\n"
126 "decw x20\n"
127 ".inst 0x648abe1a // bfcvtnt z26.h, p7/M, z16.s\n"
128 "ld1w { z16.s }, p2/Z, [x24, #-3, MUL VL]\n"
129 ".inst 0x658abe13 // bfcvt z19.h, p7/M, z16.s\n"
130 "ld1w { z16.s }, p1/Z, [x23, #7, MUL VL]\n"
131 "whilelt p1.s, XZR, x20\n"
132 "decw x20\n"
133 ".inst 0x648abe19 // bfcvtnt z25.h, p7/M, z16.s\n"
134 "ld1w { z16.s }, p1/Z, [x24, #-2, MUL VL]\n"
135 "addvl x23, x23, #16\n"
136 ".inst 0x658abe12 // bfcvt z18.h, p7/M, z16.s\n"
137 "ld1w { z16.s }, p0/Z, [x23, #-8, MUL VL]\n"
138 "whilelt p0.s, XZR, x20\n"
139 "cmp x21, #0x0\n"
140 ".inst 0x648abe18 // bfcvtnt z24.h, p7/M, z16.s\n"
141 "ld1w { z16.s }, p0/Z, [x24, #-1, MUL VL]\n"
142 ".inst 0x658abe11 // bfcvt z17.h, p7/M, z16.s\n"
143 "ld1w { z16.s }, p6/Z, [x23, #-7, MUL VL]\n"
144 ".inst 0x648abe17 // bfcvtnt z23.h, p7/M, z16.s\n"
145 "ld1w { z16.s }, p5/Z, [x23, #-6, MUL VL]\n"
146 ".inst 0x648abe16 // bfcvtnt z22.h, p7/M, z16.s\n"
147 "ld1w { z16.s }, p4/Z, [x23, #-5, MUL VL]\n"
148 ".inst 0x648abe15 // bfcvtnt z21.h, p7/M, z16.s\n"
149 "ld1w { z16.s }, p3/Z, [x23, #-4, MUL VL]\n"
150 ".inst 0x648abe14 // bfcvtnt z20.h, p7/M, z16.s\n"
151 "ld1w { z16.s }, p2/Z, [x23, #-3, MUL VL]\n"
152 ".inst 0x648abe13 // bfcvtnt z19.h, p7/M, z16.s\n"
153 "ld1w { z16.s }, p1/Z, [x23, #-2, MUL VL]\n"
154 ".inst 0x648abe12 // bfcvtnt z18.h, p7/M, z16.s\n"
155 "ld1w { z16.s }, p0/Z, [x23, #-1, MUL VL]\n"
156 "st1h { z0.h }, p7, [x19]\n"
157 ".inst 0x648abe11 // bfcvtnt z17.h, p7/M, z16.s\n"
158 "st1h { z31.h }, p7, [x19, #1, MUL VL]\n"
159 "st1h { z30.h }, p7, [x19, #2, MUL VL]\n"
160 "st1h { z29.h }, p7, [x19, #3, MUL VL]\n"
161 "st1h { z28.h }, p7, [x19, #4, MUL VL]\n"
162 "st1h { z27.h }, p7, [x19, #5, MUL VL]\n"
163 "st1h { z26.h }, p7, [x19, #6, MUL VL]\n"
164 "st1h { z25.h }, p7, [x19, #7, MUL VL]\n"
165 "addvl x19, x19, #16\n"
166 "st1h { z24.h }, p7, [x19, #-8, MUL VL]\n"
167 "st1h { z23.h }, p7, [x19, #-7, MUL VL]\n"
168 "st1h { z22.h }, p7, [x19, #-6, MUL VL]\n"
169 "st1h { z21.h }, p7, [x19, #-5, MUL VL]\n"
170 "st1h { z20.h }, p7, [x19, #-4, MUL VL]\n"
171 "st1h { z19.h }, p7, [x19, #-3, MUL VL]\n"
172 "st1h { z18.h }, p7, [x19, #-2, MUL VL]\n"
173 "st1h { z17.h }, p7, [x19, #-1, MUL VL]\n"
174 "bgt 2b\n"
175 "3:" // Main row loop: Column loop skip
176 "cmp %x[height], #0x1\n"
177 "addvl %x[out], %x[out], #16\n"
178 "bge 1b\n"
179 ".inst 0xd503467f // SMSTOP\n"
180 : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
181 : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
182 : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x19", "x20", "x21", "x22", "x23", "x24", "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"
183 );
184 }
185
186 } // anonymous namespace
187 template<>
Transform(bfloat16 * out,const float * in,int stride,int x0,int xmax,int k0,int kmax)188 void Transform<16, 2, true, VLType::SME>(
189 bfloat16 *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
190 {
191 sme_transpose_interleave_16VL_2x2_fp32bf16(
192 out,
193 in + k0 * stride + x0,
194 (xmax-x0),
195 stride * sizeof(float),
196 (kmax-k0)
197 );
198 }
199
200 #endif
201