27 #if defined(ARM_COMPUTE_ENABLE_SVE)
31 void sve_transpose_interleave_3VL(uint16_t *out,
const uint16_t *in,
size_t width,
size_t in_stride,
size_t height)
33 size_t out_stride = 3 * height * get_vector_length<uint8_t>();
36 "cmp %x[height], #0x4\n"
41 "add x25, x26, %x[in_stride]\n"
42 "add x24, x25, %x[in_stride]\n"
43 "add x23, x24, %x[in_stride]\n"
44 "add %x[in], x23, %x[in_stride]\n"
46 "sub %x[height], %x[height], #0x4\n"
47 "mov x21, %x[width]\n"
50 "whilelt p2.h, XZR, x20\n"
51 "ld1h { z27.h }, p2/Z, [x26]\n"
52 "ld1h { z26.h }, p2/Z, [x25]\n"
54 "whilelt p1.h, XZR, x20\n"
55 "ld1h { z25.h }, p1/Z, [x26, #1, MUL VL]\n"
56 "ld1h { z24.h }, p1/Z, [x25, #1, MUL VL]\n"
58 "whilelt p0.h, XZR, x20\n"
59 "ld1h { z23.h }, p0/Z, [x26, #2, MUL VL]\n"
60 "ld1h { z22.h }, p0/Z, [x25, #2, MUL VL]\n"
62 "dech x21, ALL, MUL #3\n"
63 "ld1h { z21.h }, p2/Z, [x24]\n"
64 "ld1h { z20.h }, p1/Z, [x24, #1, MUL VL]\n"
65 "ld1h { z19.h }, p0/Z, [x24, #2, MUL VL]\n"
66 "ld1h { z18.h }, p2/Z, [x23]\n"
68 "addvl x26, x26, #3\n"
69 "ld1h { z17.h }, p1/Z, [x23, #1, MUL VL]\n"
70 "ld1h { z16.h }, p0/Z, [x23, #2, MUL VL]\n"
71 "st1h { z27.h }, p3, [x20]\n"
72 "addvl x25, x25, #3\n"
73 "st1h { z25.h }, p3, [x20, #1, MUL VL]\n"
74 "addvl x24, x24, #3\n"
75 "addvl x23, x23, #3\n"
76 "st1h { z23.h }, p3, [x20, #2, MUL VL]\n"
77 "add x22, x22, %x[out_stride]\n"
78 "st1h { z26.h }, p3, [x20, #3, MUL VL]\n"
79 "st1h { z24.h }, p3, [x20, #4, MUL VL]\n"
80 "st1h { z22.h }, p3, [x20, #5, MUL VL]\n"
81 "st1h { z21.h }, p3, [x20, #6, MUL VL]\n"
82 "st1h { z20.h }, p3, [x20, #7, MUL VL]\n"
83 "addvl x20, x20, #12\n"
84 "st1h { z19.h }, p3, [x20, #-4, MUL VL]\n"
85 "st1h { z18.h }, p3, [x20, #-3, MUL VL]\n"
86 "st1h { z17.h }, p3, [x20, #-2, MUL VL]\n"
87 "st1h { z16.h }, p3, [x20, #-1, MUL VL]\n"
90 "cmp %x[height], #0x4\n"
91 "addvl %x[out], %x[out], #12\n"
93 "cbz %x[height], 8f\n"
97 "add %x[in], x26, %x[in_stride]\n"
99 "sub %x[height], %x[height], #0x1\n"
100 "mov x21, %x[width]\n"
103 "whilelt p0.h, XZR, x20\n"
104 "ld1h { z18.h }, p0/Z, [x26]\n"
106 "whilelt p0.h, XZR, x20\n"
107 "ld1h { z17.h }, p0/Z, [x26, #1, MUL VL]\n"
109 "dech x21, ALL, MUL #3\n"
110 "whilelt p0.h, XZR, x20\n"
112 "ld1h { z16.h }, p0/Z, [x26, #2, MUL VL]\n"
113 "st1h { z18.h }, p3, [x22]\n"
114 "addvl x26, x26, #3\n"
115 "st1h { z17.h }, p3, [x22, #1, MUL VL]\n"
116 "st1h { z16.h }, p3, [x22, #2, MUL VL]\n"
117 "add x22, x22, %x[out_stride]\n"
120 "cmp %x[height], #0x1\n"
121 "addvl %x[out], %x[out], #3\n"
124 : [height]
"+&r" (height), [in]
"+&r" (in), [out]
"+&r" (out)
125 : [in_stride]
"r" (in_stride), [out_stride]
"r" (out_stride), [width]
"r" (width)
126 :
"cc",
"memory",
"p0",
"p1",
"p2",
"p3",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"z16",
"z17",
"z18",
"z19",
"z20",
"z21",
"z22",
"z23",
"z24",
"z25",
"z26",
"z27"
133 void Transform<3, 1, true, VLType::SVE>(
134 float *out,
const float *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
136 sve_transpose_interleave_3VL(
137 reinterpret_cast<uint16_t *
>(out),
138 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
139 (xmax-x0) *
sizeof(
float) / 2,
140 stride *
sizeof(
float),
146 void Transform<3, 1, true, VLType::SVE>(
147 __fp16 *out,
const __fp16 *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
149 sve_transpose_interleave_3VL(
150 reinterpret_cast<uint16_t *
>(out),
151 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
152 (xmax-x0) *
sizeof(__fp16) / 2,
153 stride *
sizeof(__fp16),
159 void Transform<3, 1, true, VLType::SVE>(
160 double *out,
const double *in,
int stride,
int x0,
int xmax,
int k0,
int kmax)
162 sve_transpose_interleave_3VL(
163 reinterpret_cast<uint16_t *
>(out),
164 reinterpret_cast<const uint16_t *
>(in + k0 * stride + x0),
165 (xmax-x0) *
sizeof(
double) / 2,
166 stride *
sizeof(
double),
172 #endif // defined(ARM_COMPUTE_ENABLE_SVE)