27 #if defined(ARM_COMPUTE_ENABLE_SME)
31 void sme_transpose_interleave_2VL_2x2_fp32bf16(
bfloat16 *out,
const float *in,
size_t width,
size_t in_stride,
size_t height)
33 float *pad_row =
reinterpret_cast<float *
>(alloca(width *
sizeof(
float)));
36 memset(pad_row, 0, width *
sizeof(
float));
39 size_t out_stride = 2 * roundup<size_t>(height, 2) * sme::get_vector_length<uint16_t>();
42 ".inst 0xd503477f // SMSTART ZA\n"
43 "cmp %x[height], #0x4\n"
48 "add x25, x26, %x[in_stride]\n"
49 "add x24, x25, %x[in_stride]\n"
50 "mov x23, %x[width]\n"
51 "cnth x20, ALL, MUL #2\n"
52 "add x21, x24, %x[in_stride]\n"
54 "add %x[in], x21, %x[in_stride]\n"
56 "sub %x[height], %x[height], #0x4\n"
59 "ld1w { z16.s }, p2/Z, [x26]\n"
60 ".inst 0x658aaa18 // bfcvt z24.h, p2/M, z16.s\n"
63 "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n"
64 ".inst 0x658aaa17 // bfcvt z23.h, p2/M, z16.s\n"
65 "ld1w { z16.s }, p2/Z, [x24]\n"
66 ".inst 0x658aaa16 // bfcvt z22.h, p2/M, z16.s\n"
67 "ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n"
68 ".inst 0x658aaa15 // bfcvt z21.h, p2/M, z16.s\n"
69 "ld1w { z16.s }, p2/Z, [x26, #2, MUL VL]\n"
70 ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n"
71 "ld1w { z16.s }, p2/Z, [x26, #3, MUL VL]\n"
72 ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n"
73 "addvl x26, x26, #4\n"
74 "ld1w { z16.s }, p2/Z, [x24, #2, MUL VL]\n"
75 ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
76 "ld1w { z16.s }, p2/Z, [x24, #3, MUL VL]\n"
77 ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
78 "addvl x24, x24, #4\n"
79 "ld1w { z16.s }, p2/Z, [x25]\n"
80 ".inst 0x648aaa18 // bfcvtnt z24.h, p2/M, z16.s\n"
81 "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
82 ".inst 0x648aaa17 // bfcvtnt z23.h, p2/M, z16.s\n"
83 "ld1w { z16.s }, p2/Z, [x21]\n"
84 ".inst 0x648aaa16 // bfcvtnt z22.h, p2/M, z16.s\n"
85 "ld1w { z16.s }, p2/Z, [x21, #1, MUL VL]\n"
86 ".inst 0x648aaa15 // bfcvtnt z21.h, p2/M, z16.s\n"
87 "ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n"
88 ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n"
89 "ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n"
90 "addvl x25, x25, #4\n"
91 ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n"
92 "ld1w { z16.s }, p2/Z, [x21, #2, MUL VL]\n"
93 ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
94 "ld1w { z16.s }, p2/Z, [x21, #3, MUL VL]\n"
95 "st1h { z24.h }, p2, [x22]\n"
96 "addvl x21, x21, #4\n"
97 ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
98 "st1h { z23.h }, p2, [x22, #1, MUL VL]\n"
99 "st1h { z22.h }, p2, [x22, #2, MUL VL]\n"
100 "st1h { z21.h }, p2, [x22, #3, MUL VL]\n"
101 "add x22, x22, %x[out_stride]\n"
102 "st1h { z20.h }, p2, [x22]\n"
103 "st1h { z19.h }, p2, [x22, #1, MUL VL]\n"
104 "st1h { z18.h }, p2, [x22, #2, MUL VL]\n"
105 "st1h { z17.h }, p2, [x22, #3, MUL VL]\n"
106 "add x22, x22, %x[out_stride]\n"
112 "whilelt p1.s, XZR, x20\n"
113 "ld1w { z16.s }, p1/Z, [x26]\n"
114 ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n"
116 "whilelt p0.s, XZR, x20\n"
117 "ld1w { z16.s }, p0/Z, [x26, #1, MUL VL]\n"
118 ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n"
119 "ld1w { z16.s }, p1/Z, [x24]\n"
120 ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
121 "decw x23, ALL, MUL #2\n"
123 "ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n"
124 ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
125 "addvl x26, x26, #2\n"
126 "addvl x24, x24, #2\n"
127 "ld1w { z16.s }, p1/Z, [x25]\n"
128 ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n"
129 "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
130 "addvl x25, x25, #2\n"
131 ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n"
132 "ld1w { z16.s }, p1/Z, [x21]\n"
133 ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
134 "ld1w { z16.s }, p0/Z, [x21, #1, MUL VL]\n"
135 "addvl x21, x21, #2\n"
136 ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
137 "st1h { z20.h }, p2, [x22]\n"
138 "st1h { z19.h }, p2, [x22, #1, MUL VL]\n"
139 "st1h { z18.h }, p2, [x22, #2, MUL VL]\n"
140 "st1h { z17.h }, p2, [x22, #3, MUL VL]\n"
141 "add x22, x22, %x[out_stride]\n"
144 "cmp %x[height], #0x4\n"
145 "addvl %x[out], %x[out], #4\n"
147 "cbz %x[height], 12f\n"
151 "add x25, x26, %x[in_stride]\n"
152 "cmp %x[height], #0x1\n"
153 "mov x21, %x[width]\n"
154 "cnth x20, ALL, MUL #2\n"
155 "add %x[in], x25, %x[in_stride]\n"
156 "csel x25, x25, %x[pad_row], GT\n"
159 "sub %x[height], %x[height], #0x2\n"
162 "ld1w { z16.s }, p2/Z, [x26]\n"
163 ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n"
164 "sub x21, x21, x20\n"
166 "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n"
167 ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n"
168 "ld1w { z16.s }, p2/Z, [x26, #2, MUL VL]\n"
169 ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
170 "ld1w { z16.s }, p2/Z, [x26, #3, MUL VL]\n"
171 ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
172 "addvl x26, x26, #4\n"
173 "ld1w { z16.s }, p2/Z, [x25]\n"
174 ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n"
175 "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
176 ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n"
177 "ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n"
178 ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
179 "ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n"
180 "st1h { z20.h }, p2, [x22]\n"
181 "addvl x25, x25, #4\n"
182 ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
183 "st1h { z19.h }, p2, [x22, #1, MUL VL]\n"
184 "add x22, x22, %x[out_stride]\n"
185 "st1h { z18.h }, p2, [x22]\n"
186 "st1h { z17.h }, p2, [x22, #1, MUL VL]\n"
187 "add x22, x22, %x[out_stride]\n"
193 "whilelt p1.s, XZR, x20\n"
194 "ld1w { z16.s }, p1/Z, [x26]\n"
195 ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n"
197 "whilelt p0.s, XZR, x20\n"
198 "ld1w { z16.s }, p0/Z, [x26, #1, MUL VL]\n"
199 ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n"
200 "ld1w { z16.s }, p1/Z, [x25]\n"
201 "decw x21, ALL, MUL #2\n"
203 ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n"
204 "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
205 "addvl x26, x26, #2\n"
206 "addvl x25, x25, #2\n"
207 ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n"
208 "st1h { z18.h }, p2, [x22]\n"
209 "st1h { z17.h }, p2, [x22, #1, MUL VL]\n"
210 "add x22, x22, %x[out_stride]\n"
213 "cmp %x[height], #0x1\n"
214 "addvl %x[out], %x[out], #2\n"
217 ".inst 0xd503467f // SMSTOP\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",
"p4",
"p5",
"p6",
"p7",
"p8",
"p9",
"p10",
"p11",
"p12",
"p13",
"p14",
"p15",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"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"
226 void Transform<2, 2, true, VLType::SME>(
227 bfloat16 *out,
const float *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
229 sme_transpose_interleave_2VL_2x2_fp32bf16(
231 in + k0 * stride + x0,
233 stride *
sizeof(
float),
239 #endif // defined(ARM_COMPUTE_ENABLE_SME)