27 #if defined(ARM_COMPUTE_ENABLE_SME)
31 void sme_transpose_interleave_2VL(uint16_t *out,
const uint16_t *in,
size_t width,
size_t in_stride,
size_t height)
33 size_t out_stride = 2 * height * sme::get_vector_length<uint8_t>();
36 ".inst 0xd503477f // SMSTART ZA\n"
37 "cmp %x[height], #0x4\n"
42 "add x25, x26, %x[in_stride]\n"
43 "add x24, x25, %x[in_stride]\n"
44 "mov x23, %x[width]\n"
45 "cnth x20, ALL, MUL #4\n"
46 "add x21, x24, %x[in_stride]\n"
48 "add %x[in], x21, %x[in_stride]\n"
50 "sub %x[height], %x[height], #0x4\n"
54 "ld1h { z31.h }, p2/Z, [x26]\n"
56 "ld1h { z30.h }, p2/Z, [x26, #1, MUL VL]\n"
57 "ld1h { z29.h }, p2/Z, [x26, #2, MUL VL]\n"
58 "ld1h { z28.h }, p2/Z, [x26, #3, MUL VL]\n"
59 "addvl x26, x26, #4\n"
60 "ld1h { z27.h }, p2/Z, [x25]\n"
61 "ld1h { z26.h }, p2/Z, [x25, #1, MUL VL]\n"
62 "ld1h { z25.h }, p2/Z, [x25, #2, MUL VL]\n"
63 "ld1h { z24.h }, p2/Z, [x25, #3, MUL VL]\n"
64 "addvl x25, x25, #4\n"
65 "ld1h { z23.h }, p2/Z, [x24]\n"
66 "ld1h { z22.h }, p2/Z, [x24, #1, MUL VL]\n"
67 "ld1h { z21.h }, p2/Z, [x24, #2, MUL VL]\n"
68 "ld1h { z20.h }, p2/Z, [x24, #3, MUL VL]\n"
69 "addvl x24, x24, #4\n"
70 "ld1h { z19.h }, p2/Z, [x21]\n"
71 "ld1h { z18.h }, p2/Z, [x21, #1, MUL VL]\n"
72 "ld1h { z17.h }, p2/Z, [x21, #2, MUL VL]\n"
73 "ld1h { z16.h }, p2/Z, [x21, #3, MUL VL]\n"
74 "st1h { z31.h }, p2, [x22]\n"
75 "addvl x21, x21, #4\n"
76 "st1h { z30.h }, p2, [x22, #1, MUL VL]\n"
77 "st1h { z27.h }, p2, [x22, #2, MUL VL]\n"
78 "st1h { z26.h }, p2, [x22, #3, MUL VL]\n"
79 "st1h { z23.h }, p2, [x22, #4, MUL VL]\n"
80 "st1h { z22.h }, p2, [x22, #5, MUL VL]\n"
81 "st1h { z19.h }, p2, [x22, #6, MUL VL]\n"
82 "st1h { z18.h }, p2, [x22, #7, MUL VL]\n"
83 "add x22, x22, %x[out_stride]\n"
84 "st1h { z29.h }, p2, [x22]\n"
85 "st1h { z28.h }, p2, [x22, #1, MUL VL]\n"
86 "st1h { z25.h }, p2, [x22, #2, MUL VL]\n"
87 "st1h { z24.h }, p2, [x22, #3, MUL VL]\n"
88 "st1h { z21.h }, p2, [x22, #4, MUL VL]\n"
89 "st1h { z20.h }, p2, [x22, #5, MUL VL]\n"
90 "st1h { z17.h }, p2, [x22, #6, MUL VL]\n"
91 "st1h { z16.h }, p2, [x22, #7, MUL VL]\n"
92 "add x22, x22, %x[out_stride]\n"
98 "whilelt p1.h, XZR, x20\n"
99 "ld1h { z23.h }, p1/Z, [x26]\n"
101 "dech x23, ALL, MUL #2\n"
102 "ld1h { z22.h }, p1/Z, [x25]\n"
103 "whilelt p0.h, XZR, x20\n"
105 "ld1h { z21.h }, p0/Z, [x26, #1, MUL VL]\n"
106 "addvl x26, x26, #2\n"
107 "ld1h { z20.h }, p0/Z, [x25, #1, MUL VL]\n"
108 "addvl x25, x25, #2\n"
109 "ld1h { z19.h }, p1/Z, [x24]\n"
110 "ld1h { z18.h }, p0/Z, [x24, #1, MUL VL]\n"
111 "addvl x24, x24, #2\n"
112 "ld1h { z17.h }, p1/Z, [x21]\n"
113 "ld1h { z16.h }, p0/Z, [x21, #1, MUL VL]\n"
114 "addvl x21, x21, #2\n"
115 "st1h { z23.h }, p2, [x22]\n"
116 "st1h { z21.h }, p2, [x22, #1, MUL VL]\n"
117 "st1h { z22.h }, p2, [x22, #2, MUL VL]\n"
118 "st1h { z20.h }, p2, [x22, #3, MUL VL]\n"
119 "st1h { z19.h }, p2, [x22, #4, MUL VL]\n"
120 "st1h { z18.h }, p2, [x22, #5, MUL VL]\n"
121 "st1h { z17.h }, p2, [x22, #6, MUL VL]\n"
122 "st1h { z16.h }, p2, [x22, #7, MUL VL]\n"
123 "add x22, x22, %x[out_stride]\n"
126 "cmp %x[height], #0x4\n"
127 "addvl %x[out], %x[out], #8\n"
129 "cbz %x[height], 12f\n"
132 "mov x21, %x[width]\n"
133 "cnth x20, ALL, MUL #4\n"
136 "add %x[in], x26, %x[in_stride]\n"
138 "sub %x[height], %x[height], #0x1\n"
141 "sub x21, x21, x20\n"
142 "ld1h { z19.h }, p2/Z, [x26]\n"
144 "ld1h { z18.h }, p2/Z, [x26, #1, MUL VL]\n"
145 "ld1h { z17.h }, p2/Z, [x26, #2, MUL VL]\n"
146 "ld1h { z16.h }, p2/Z, [x26, #3, MUL VL]\n"
147 "st1h { z19.h }, p2, [x22]\n"
148 "addvl x26, x26, #4\n"
149 "st1h { z18.h }, p2, [x22, #1, MUL VL]\n"
150 "add x22, x22, %x[out_stride]\n"
151 "st1h { z17.h }, p2, [x22]\n"
152 "st1h { z16.h }, p2, [x22, #1, MUL VL]\n"
153 "add x22, x22, %x[out_stride]\n"
159 "whilelt p0.h, XZR, x20\n"
160 "ld1h { z17.h }, p0/Z, [x26]\n"
162 "dech x21, ALL, MUL #2\n"
163 "whilelt p0.h, XZR, x20\n"
165 "ld1h { z16.h }, p0/Z, [x26, #1, MUL VL]\n"
166 "st1h { z17.h }, p2, [x22]\n"
167 "addvl x26, x26, #2\n"
168 "st1h { z16.h }, p2, [x22, #1, MUL VL]\n"
169 "add x22, x22, %x[out_stride]\n"
172 "cmp %x[height], #0x1\n"
173 "addvl %x[out], %x[out], #2\n"
176 ".inst 0xd503467f // SMSTOP\n"
177 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
178 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [width]
"r" (width)
179 :
"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"
186 void Transform<2, 1, true, VLType::SME>(
187 float *out,
const float *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
189 sme_transpose_interleave_2VL(
190 reinterpret_cast<uint16_t *
>(out),
191 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
192 (xmax-x0) *
sizeof(
float) / 2,
193 stride *
sizeof(
float),
199 void Transform<2, 1, true, VLType::SME>(
200 bfloat16 *out,
const bfloat16 *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
202 sme_transpose_interleave_2VL(
203 reinterpret_cast<uint16_t *
>(out),
204 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
212 void Transform<2, 1, true, VLType::SME>(
213 __fp16 *out,
const __fp16 *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
215 sme_transpose_interleave_2VL(
216 reinterpret_cast<uint16_t *
>(out),
217 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
218 (xmax-x0) *
sizeof(__fp16) / 2,
219 stride *
sizeof(__fp16),
225 #endif // defined(ARM_COMPUTE_ENABLE_SME)