27 #if defined(ARM_COMPUTE_ENABLE_SVE)
31 void sve_transpose_interleave_2VL_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 = 2 * roundup<size_t>(height, 4) * get_vector_length<uint32_t>();
45 "mov x25, %x[width]\n"
47 "cmp %x[height], #0x3\n"
49 "add x22, x26, %x[in_stride]\n"
50 "add x21, x22, %x[in_stride]\n"
51 "add x20, x21, %x[in_stride]\n"
52 "add %x[in], x20, %x[in_stride]\n"
53 "csel x20, x20, %x[pad_row], GT\n"
54 "csel x21, x21, %x[pad_row], GE\n"
55 "cmp %x[height], #0x1\n"
56 "csel x22, x22, %x[pad_row], GT\n"
58 "sub %x[height], %x[height], #0x4\n"
61 "ld1w { z18.s }, p1/Z, [x26]\n"
62 "ld1w { z17.s }, p1/Z, [x21]\n"
64 "ld1w { z21.s }, p1/Z, [x26, #1, MUL VL]\n"
65 "ld1w { z16.s }, p1/Z, [x21, #1, MUL VL]\n"
67 "addvl x26, x26, #2\n"
68 "ld1w { z26.s }, p1/Z, [x22]\n"
69 "ld1w { z20.s }, p1/Z, [x20]\n"
70 "addvl x21, x21, #2\n"
71 "zip1 z19.s, z18.s, z17.s\n"
72 "zip2 z18.s, z18.s, z17.s\n"
73 "ld1w { z25.s }, p1/Z, [x22, #1, MUL VL]\n"
74 "ld1w { z24.s }, p1/Z, [x20, #1, MUL VL]\n"
75 "addvl x22, x22, #2\n"
76 "zip1 z17.s, z21.s, z16.s\n"
77 "zip2 z16.s, z21.s, z16.s\n"
78 "addvl x20, x20, #2\n"
79 ".inst 0x658aa677 // bfcvt z23.h, p1/M, z19.s\n"
80 "zip1 z22.s, z26.s, z20.s\n"
81 ".inst 0x658aa655 // bfcvt z21.h, p1/M, z18.s\n"
82 "zip2 z20.s, z26.s, z20.s\n"
83 ".inst 0x658aa633 // bfcvt z19.h, p1/M, z17.s\n"
84 "zip1 z18.s, z25.s, z24.s\n"
85 ".inst 0x658aa611 // bfcvt z17.h, p1/M, z16.s\n"
86 "zip2 z16.s, z25.s, z24.s\n"
87 ".inst 0x648aa6d7 // bfcvtnt z23.h, p1/M, z22.s\n"
88 ".inst 0x648aa695 // bfcvtnt z21.h, p1/M, z20.s\n"
89 ".inst 0x648aa653 // bfcvtnt z19.h, p1/M, z18.s\n"
90 ".inst 0x648aa611 // bfcvtnt z17.h, p1/M, z16.s\n"
91 "st1h { z23.h }, p1, [x23]\n"
92 "st1h { z21.h }, p1, [x23, #1, MUL VL]\n"
93 "add x23, x23, %x[out_stride]\n"
94 "st1h { z19.h }, p1, [x23]\n"
95 "st1h { z17.h }, p1, [x23, #1, MUL VL]\n"
96 "add x23, x23, %x[out_stride]\n"
101 "whilelt p0.s, XZR, x25\n"
102 "decd x25, ALL, MUL #2\n"
103 "ld1w { z19.s }, p0/Z, [x26]\n"
104 "addvl x26, x26, #1\n"
105 "ld1w { z16.s }, p0/Z, [x21]\n"
106 "addvl x21, x21, #1\n"
107 "ld1w { z20.s }, p0/Z, [x22]\n"
108 "addvl x22, x22, #1\n"
109 "ld1w { z18.s }, p0/Z, [x20]\n"
110 "addvl x20, x20, #1\n"
112 "zip1 z17.s, z19.s, z16.s\n"
113 "zip2 z16.s, z19.s, z16.s\n"
114 "zip1 z19.s, z20.s, z18.s\n"
115 "zip2 z18.s, z20.s, z18.s\n"
116 ".inst 0x658aa631 // bfcvt z17.h, p1/M, z17.s\n"
117 ".inst 0x658aa610 // bfcvt z16.h, p1/M, z16.s\n"
118 ".inst 0x648aa671 // bfcvtnt z17.h, p1/M, z19.s\n"
119 ".inst 0x648aa650 // bfcvtnt z16.h, p1/M, z18.s\n"
120 "st1h { z17.h }, p1, [x23]\n"
121 "st1h { z16.h }, p1, [x23, #1, MUL VL]\n"
122 "add x23, x23, %x[out_stride]\n"
125 "cmp %x[height], #0x1\n"
126 "addvl %x[out], %x[out], #2\n"
128 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
129 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [pad_row]
"r" (pad_row), [width]
"r" (width)
130 :
"cc",
"memory",
"p0",
"p1",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"z16",
"z17",
"z18",
"z19",
"z20",
"z21",
"z22",
"z23",
"z24",
"z25",
"z26"
136 void Transform<2, 4, true, VLType::SVE>(
137 bfloat16 *out,
const float *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
139 sve_transpose_interleave_2VL_2x4_fp32bf16(
141 in + k0 * stride + x0,
143 stride *
sizeof(
float),
149 #endif // defined(ARM_COMPUTE_ENABLE_SVE)