27 #if defined(ARM_COMPUTE_ENABLE_SME)
31 void sme_transpose_interleave_4VL_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 = 4 * 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 "add x23, x24, %x[in_stride]\n"
51 "add %x[in], x23, %x[in_stride]\n"
53 "sub %x[height], %x[height], #0x4\n"
54 "mov x21, %x[width]\n"
57 "whilelt p3.s, XZR, x20\n"
58 "ld1w { z16.s }, p3/Z, [x26]\n"
59 ".inst 0x658ab218 // bfcvt z24.h, p4/M, z16.s\n"
61 "whilelt p2.s, XZR, x20\n"
62 "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n"
63 ".inst 0x658ab217 // bfcvt z23.h, p4/M, z16.s\n"
65 "whilelt p1.s, XZR, x20\n"
66 "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n"
67 ".inst 0x658ab216 // bfcvt z22.h, p4/M, z16.s\n"
69 "whilelt p0.s, XZR, x20\n"
70 "ld1w { z16.s }, p0/Z, [x26, #3, MUL VL]\n"
71 ".inst 0x658ab215 // bfcvt z21.h, p4/M, z16.s\n"
72 "ld1w { z16.s }, p3/Z, [x24]\n"
73 ".inst 0x658ab214 // bfcvt z20.h, p4/M, z16.s\n"
74 "decw x21, ALL, MUL #4\n"
76 "ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n"
77 ".inst 0x658ab213 // bfcvt z19.h, p4/M, z16.s\n"
78 "addvl x26, x26, #4\n"
79 "ld1w { z16.s }, p1/Z, [x24, #2, MUL VL]\n"
80 ".inst 0x658ab212 // bfcvt z18.h, p4/M, z16.s\n"
81 "ld1w { z16.s }, p0/Z, [x24, #3, MUL VL]\n"
82 ".inst 0x658ab211 // bfcvt z17.h, p4/M, z16.s\n"
83 "addvl x24, x24, #4\n"
84 "ld1w { z16.s }, p3/Z, [x25]\n"
85 ".inst 0x648ab218 // bfcvtnt z24.h, p4/M, z16.s\n"
86 "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
87 ".inst 0x648ab217 // bfcvtnt z23.h, p4/M, z16.s\n"
88 "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n"
89 ".inst 0x648ab216 // bfcvtnt z22.h, p4/M, z16.s\n"
90 "ld1w { z16.s }, p0/Z, [x25, #3, MUL VL]\n"
91 "addvl x25, x25, #4\n"
92 ".inst 0x648ab215 // bfcvtnt z21.h, p4/M, z16.s\n"
93 "ld1w { z16.s }, p3/Z, [x23]\n"
94 ".inst 0x648ab214 // bfcvtnt z20.h, p4/M, z16.s\n"
95 "ld1w { z16.s }, p2/Z, [x23, #1, MUL VL]\n"
96 ".inst 0x648ab213 // bfcvtnt z19.h, p4/M, z16.s\n"
97 "ld1w { z16.s }, p1/Z, [x23, #2, MUL VL]\n"
98 ".inst 0x648ab212 // bfcvtnt z18.h, p4/M, z16.s\n"
99 "ld1w { z16.s }, p0/Z, [x23, #3, MUL VL]\n"
100 "addvl x23, x23, #4\n"
101 ".inst 0x648ab211 // bfcvtnt z17.h, p4/M, z16.s\n"
102 "st1h { z24.h }, p4, [x22]\n"
103 "st1h { z23.h }, p4, [x22, #1, MUL VL]\n"
104 "st1h { z22.h }, p4, [x22, #2, MUL VL]\n"
105 "st1h { z21.h }, p4, [x22, #3, MUL VL]\n"
106 "st1h { z20.h }, p4, [x22, #4, MUL VL]\n"
107 "st1h { z19.h }, p4, [x22, #5, MUL VL]\n"
108 "st1h { z18.h }, p4, [x22, #6, MUL VL]\n"
109 "st1h { z17.h }, p4, [x22, #7, MUL VL]\n"
110 "add x22, x22, %x[out_stride]\n"
113 "cmp %x[height], #0x4\n"
114 "addvl %x[out], %x[out], #8\n"
116 "cbz %x[height], 8f\n"
120 "add x25, x26, %x[in_stride]\n"
121 "cmp %x[height], #0x1\n"
122 "add %x[in], x25, %x[in_stride]\n"
124 "csel x25, x25, %x[pad_row], GT\n"
125 "sub %x[height], %x[height], #0x2\n"
126 "mov x21, %x[width]\n"
129 "whilelt p3.s, XZR, x20\n"
130 "ld1w { z16.s }, p3/Z, [x26]\n"
131 ".inst 0x658ab214 // bfcvt z20.h, p4/M, z16.s\n"
133 "whilelt p2.s, XZR, x20\n"
134 "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n"
135 ".inst 0x658ab213 // bfcvt z19.h, p4/M, z16.s\n"
137 "whilelt p1.s, XZR, x20\n"
138 "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n"
139 ".inst 0x658ab212 // bfcvt z18.h, p4/M, z16.s\n"
141 "whilelt p0.s, XZR, x20\n"
142 "ld1w { z16.s }, p0/Z, [x26, #3, MUL VL]\n"
143 ".inst 0x658ab211 // bfcvt z17.h, p4/M, z16.s\n"
144 "ld1w { z16.s }, p3/Z, [x25]\n"
145 "decw x21, ALL, MUL #4\n"
147 ".inst 0x648ab214 // bfcvtnt z20.h, p4/M, z16.s\n"
148 "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
149 "addvl x26, x26, #4\n"
150 ".inst 0x648ab213 // bfcvtnt z19.h, p4/M, z16.s\n"
151 "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n"
152 ".inst 0x648ab212 // bfcvtnt z18.h, p4/M, z16.s\n"
153 "ld1w { z16.s }, p0/Z, [x25, #3, MUL VL]\n"
154 "addvl x25, x25, #4\n"
155 ".inst 0x648ab211 // bfcvtnt z17.h, p4/M, z16.s\n"
156 "st1h { z20.h }, p4, [x22]\n"
157 "st1h { z19.h }, p4, [x22, #1, MUL VL]\n"
158 "st1h { z18.h }, p4, [x22, #2, MUL VL]\n"
159 "st1h { z17.h }, p4, [x22, #3, MUL VL]\n"
160 "add x22, x22, %x[out_stride]\n"
163 "cmp %x[height], #0x1\n"
164 "addvl %x[out], %x[out], #4\n"
167 ".inst 0xd503467f // SMSTOP\n"
168 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
169 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
170 :
"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"
176 void Transform<4, 2, true, VLType::SME>(
177 bfloat16 *out,
const float *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
179 sme_transpose_interleave_4VL_2x2_fp32bf16(
181 in + k0 * stride + x0,
183 stride *
sizeof(
float),
189 #endif // defined(ARM_COMPUTE_ENABLE_SME)