27 #if defined(ARM_COMPUTE_ENABLE_SVE)
31 void sve_transpose_interleave_6VL_2x4_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 = 6 * roundup<size_t>(height, 4) * get_vector_length<uint32_t>();
45 "add x25, x26, %x[in_stride]\n"
46 "add x24, x25, %x[in_stride]\n"
47 "mov x23, %x[width]\n"
48 "cnth x20, ALL, MUL #3\n"
49 "add x22, x24, %x[in_stride]\n"
50 "cmp %x[height], #0x3\n"
51 "add %x[in], x22, %x[in_stride]\n"
52 "csel x22, x22, %x[pad_row], GT\n"
53 "csel x24, x24, %x[pad_row], GE\n"
54 "cmp %x[height], #0x1\n"
55 "csel x25, x25, %x[pad_row], GT\n"
58 "sub %x[height], %x[height], #0x4\n"
61 "ld1w { z17.s }, p3/Z, [x26]\n"
62 "ld1w { z18.s }, p3/Z, [x26, #1, MUL VL]\n"
65 "ld1w { z19.s }, p3/Z, [x26, #2, MUL VL]\n"
66 "ld1w { z16.s }, p3/Z, [x24]\n"
67 "zip1 z21.s, z17.s, z16.s\n"
68 "zip2 z20.s, z17.s, z16.s\n"
69 "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n"
70 "ld1w { z16.s }, p3/Z, [x24, #2, MUL VL]\n"
71 "zip1 z29.s, z18.s, z17.s\n"
72 "zip2 z28.s, z18.s, z17.s\n"
73 "ld1w { z17.s }, p3/Z, [x26, #3, MUL VL]\n"
74 "ld1w { z18.s }, p3/Z, [x26, #4, MUL VL]\n"
75 "zip1 z27.s, z19.s, z16.s\n"
76 "zip2 z26.s, z19.s, z16.s\n"
77 "ld1w { z19.s }, p3/Z, [x26, #5, MUL VL]\n"
78 "ld1w { z16.s }, p3/Z, [x24, #3, MUL VL]\n"
79 "zip1 z25.s, z17.s, z16.s\n"
80 "zip2 z24.s, z17.s, z16.s\n"
81 "ld1w { z17.s }, p3/Z, [x24, #4, MUL VL]\n"
82 "ld1w { z16.s }, p3/Z, [x24, #5, MUL VL]\n"
83 "zip1 z12.s, z18.s, z17.s\n"
84 "zip2 z11.s, z18.s, z17.s\n"
85 "ld1w { z18.s }, p3/Z, [x25]\n"
86 "ld1w { z23.s }, p3/Z, [x25, #1, MUL VL]\n"
87 "zip1 z10.s, z19.s, z16.s\n"
88 "zip2 z9.s, z19.s, z16.s\n"
89 "ld1w { z22.s }, p3/Z, [x25, #2, MUL VL]\n"
90 "ld1w { z17.s }, p3/Z, [x22]\n"
91 ".inst 0x658aaea8 // bfcvt z8.h, p3/M, z21.s\n"
92 "zip1 z7.s, z18.s, z17.s\n"
93 "ld1w { z16.s }, p3/Z, [x22, #1, MUL VL]\n"
94 "ld1w { z21.s }, p3/Z, [x22, #2, MUL VL]\n"
95 ".inst 0x658aae86 // bfcvt z6.h, p3/M, z20.s\n"
96 "zip2 z5.s, z18.s, z17.s\n"
97 "ld1w { z20.s }, p3/Z, [x25, #3, MUL VL]\n"
98 "ld1w { z19.s }, p3/Z, [x25, #4, MUL VL]\n"
99 ".inst 0x658aafa4 // bfcvt z4.h, p3/M, z29.s\n"
100 "zip1 z3.s, z23.s, z16.s\n"
101 "ld1w { z2.s }, p3/Z, [x25, #5, MUL VL]\n"
102 "ld1w { z18.s }, p3/Z, [x22, #3, MUL VL]\n"
103 ".inst 0x658aaf81 // bfcvt z1.h, p3/M, z28.s\n"
104 "zip2 z0.s, z23.s, z16.s\n"
105 "ld1w { z17.s }, p3/Z, [x22, #4, MUL VL]\n"
106 "ld1w { z16.s }, p3/Z, [x22, #5, MUL VL]\n"
107 ".inst 0x658aaf7f // bfcvt z31.h, p3/M, z27.s\n"
108 "zip1 z30.s, z22.s, z21.s\n"
109 ".inst 0x658aaf5d // bfcvt z29.h, p3/M, z26.s\n"
110 "zip2 z28.s, z22.s, z21.s\n"
111 "addvl x26, x26, #6\n"
112 "addvl x25, x25, #6\n"
113 ".inst 0x658aaf3b // bfcvt z27.h, p3/M, z25.s\n"
114 "zip1 z26.s, z20.s, z18.s\n"
115 "addvl x24, x24, #6\n"
116 "addvl x22, x22, #6\n"
117 ".inst 0x658aaf19 // bfcvt z25.h, p3/M, z24.s\n"
118 "zip2 z24.s, z20.s, z18.s\n"
119 ".inst 0x658aad97 // bfcvt z23.h, p3/M, z12.s\n"
120 "zip1 z22.s, z19.s, z17.s\n"
121 ".inst 0x658aad75 // bfcvt z21.h, p3/M, z11.s\n"
122 "zip2 z20.s, z19.s, z17.s\n"
123 ".inst 0x658aad53 // bfcvt z19.h, p3/M, z10.s\n"
124 "zip1 z18.s, z2.s, z16.s\n"
125 ".inst 0x658aad31 // bfcvt z17.h, p3/M, z9.s\n"
126 "zip2 z16.s, z2.s, z16.s\n"
127 ".inst 0x648aace8 // bfcvtnt z8.h, p3/M, z7.s\n"
128 ".inst 0x648aaca6 // bfcvtnt z6.h, p3/M, z5.s\n"
129 "st1h { z8.h }, p3, [x21]\n"
130 ".inst 0x648aac64 // bfcvtnt z4.h, p3/M, z3.s\n"
131 ".inst 0x648aac01 // bfcvtnt z1.h, p3/M, z0.s\n"
132 "st1h { z6.h }, p3, [x21, #1, MUL VL]\n"
133 ".inst 0x648aafdf // bfcvtnt z31.h, p3/M, z30.s\n"
134 ".inst 0x648aaf9d // bfcvtnt z29.h, p3/M, z28.s\n"
135 "st1h { z4.h }, p3, [x21, #2, MUL VL]\n"
136 "st1h { z1.h }, p3, [x21, #3, MUL VL]\n"
137 ".inst 0x648aaf5b // bfcvtnt z27.h, p3/M, z26.s\n"
138 ".inst 0x648aaf19 // bfcvtnt z25.h, p3/M, z24.s\n"
139 "st1h { z31.h }, p3, [x21, #4, MUL VL]\n"
140 ".inst 0x648aaed7 // bfcvtnt z23.h, p3/M, z22.s\n"
141 ".inst 0x648aae95 // bfcvtnt z21.h, p3/M, z20.s\n"
142 "st1h { z29.h }, p3, [x21, #5, MUL VL]\n"
143 "add x21, x21, %x[out_stride]\n"
144 ".inst 0x648aae53 // bfcvtnt z19.h, p3/M, z18.s\n"
145 ".inst 0x648aae11 // bfcvtnt z17.h, p3/M, z16.s\n"
146 "st1h { z27.h }, p3, [x21]\n"
147 "st1h { z25.h }, p3, [x21, #1, MUL VL]\n"
148 "st1h { z23.h }, p3, [x21, #2, MUL VL]\n"
149 "st1h { z21.h }, p3, [x21, #3, MUL VL]\n"
150 "st1h { z19.h }, p3, [x21, #4, MUL VL]\n"
151 "st1h { z17.h }, p3, [x21, #5, MUL VL]\n"
152 "add x21, x21, %x[out_stride]\n"
158 "whilelt p2.s, XZR, x20\n"
159 "ld1w { z20.s }, p2/Z, [x26]\n"
160 "ld1w { z19.s }, p2/Z, [x24]\n"
162 "whilelt p1.s, XZR, x20\n"
163 "ld1w { z18.s }, p1/Z, [x26, #1, MUL VL]\n"
164 "ld1w { z17.s }, p1/Z, [x24, #1, MUL VL]\n"
166 "whilelt p0.s, XZR, x20\n"
167 "ld1w { z25.s }, p0/Z, [x26, #2, MUL VL]\n"
168 "ld1w { z16.s }, p0/Z, [x24, #2, MUL VL]\n"
169 "ld1w { z24.s }, p2/Z, [x25]\n"
170 "ld1w { z30.s }, p1/Z, [x25, #1, MUL VL]\n"
171 "zip1 z23.s, z20.s, z19.s\n"
172 "zip2 z22.s, z20.s, z19.s\n"
173 "ld1w { z29.s }, p0/Z, [x25, #2, MUL VL]\n"
174 "ld1w { z21.s }, p2/Z, [x22]\n"
175 "zip1 z20.s, z18.s, z17.s\n"
176 "zip2 z19.s, z18.s, z17.s\n"
177 "ld1w { z18.s }, p1/Z, [x22, #1, MUL VL]\n"
178 "ld1w { z28.s }, p0/Z, [x22, #2, MUL VL]\n"
179 "zip1 z17.s, z25.s, z16.s\n"
180 "zip2 z16.s, z25.s, z16.s\n"
181 "decd x23, ALL, MUL #6\n"
182 ".inst 0x658aaefb // bfcvt z27.h, p3/M, z23.s\n"
183 "zip1 z26.s, z24.s, z21.s\n"
185 ".inst 0x658aaed9 // bfcvt z25.h, p3/M, z22.s\n"
186 "zip2 z24.s, z24.s, z21.s\n"
187 "addvl x26, x26, #3\n"
188 "addvl x25, x25, #3\n"
189 ".inst 0x658aae97 // bfcvt z23.h, p3/M, z20.s\n"
190 "zip1 z22.s, z30.s, z18.s\n"
191 "addvl x24, x24, #3\n"
192 "addvl x22, x22, #3\n"
193 ".inst 0x658aae75 // bfcvt z21.h, p3/M, z19.s\n"
194 "zip2 z20.s, z30.s, z18.s\n"
195 ".inst 0x658aae33 // bfcvt z19.h, p3/M, z17.s\n"
196 "zip1 z18.s, z29.s, z28.s\n"
197 ".inst 0x658aae11 // bfcvt z17.h, p3/M, z16.s\n"
198 "zip2 z16.s, z29.s, z28.s\n"
199 ".inst 0x648aaf5b // bfcvtnt z27.h, p3/M, z26.s\n"
200 ".inst 0x648aaf19 // bfcvtnt z25.h, p3/M, z24.s\n"
201 "st1h { z27.h }, p3, [x21]\n"
202 ".inst 0x648aaed7 // bfcvtnt z23.h, p3/M, z22.s\n"
203 ".inst 0x648aae95 // bfcvtnt z21.h, p3/M, z20.s\n"
204 "st1h { z25.h }, p3, [x21, #1, MUL VL]\n"
205 ".inst 0x648aae53 // bfcvtnt z19.h, p3/M, z18.s\n"
206 ".inst 0x648aae11 // bfcvtnt z17.h, p3/M, z16.s\n"
207 "st1h { z23.h }, p3, [x21, #2, MUL VL]\n"
208 "st1h { z21.h }, p3, [x21, #3, MUL VL]\n"
209 "st1h { z19.h }, p3, [x21, #4, MUL VL]\n"
210 "st1h { z17.h }, p3, [x21, #5, MUL VL]\n"
211 "add x21, x21, %x[out_stride]\n"
214 "cmp %x[height], #0x1\n"
215 "addvl %x[out], %x[out], #6\n"
217 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
218 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
219 :
"cc",
"memory",
"p0",
"p1",
"p2",
"p3",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"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"
225 void Transform<6, 4, true, VLType::SVE>(
226 bfloat16 *out,
const float *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
228 sve_transpose_interleave_6VL_2x4_fp32bf16(
230 in + k0 * stride + x0,
232 stride *
sizeof(
float),
238 #endif // defined(ARM_COMPUTE_ENABLE_SVE)