29 #if defined(ARM_COMPUTE_ENABLE_SVE)
34 void sve_fp32_nhwc_max_2x2_s1_output2x2_depthfirst_impl(
35 const unsigned int n_channels,
36 const float *
const *
const inptrs,
37 float *
const *
const outptrs,
38 const bool exclude_padding,
39 const unsigned int pad_left,
40 const unsigned int pad_top,
41 const unsigned int pad_right,
42 const unsigned int pad_bottom
47 const uint64_t n_channels;
48 const float *
const *
const inptrs;
49 float *
const *
const outptrs;
51 unsigned int channels,
52 const float *
const *input_ptrs,
53 float *
const * output_ptrs,
54 bool,
unsigned int,
unsigned int,
unsigned int,
unsigned int
55 ) : n_channels(channels),
62 const KernelArgs
args(n_channels, inptrs, outptrs, exclude_padding,
63 pad_left, pad_top, pad_right, pad_bottom);
66 "ldr x15, [%x[args], %[offsetof_n_channels]]\n"
67 "ldr x21, [%x[args], %[offsetof_outptrs]]\n"
69 "whilelt p0.s, x14, x15\n"
70 "ldr x20, [%x[args], %[offsetof_inptrs]]\n"
71 "ldp x13, x12, [x21, #0x0]\n"
74 "ldp x10, x9, [x21, #0x10]\n"
75 "ldp x28, x27, [x20, #0x0]\n"
76 "ldp x26, x25, [x20, #0x10]\n"
77 "ldp x24, x23, [x20, #0x20]\n"
78 "ldp x22, x21, [x20, #0x30]\n"
79 "ldr x20, [x20, #0x40]\n"
80 "ld1w { z31.s }, p0/Z, [x27, x14, LSL #2]\n"
81 "ld1w { z30.s }, p0/Z, [x24, x14, LSL #2]\n"
82 "ld1w { z29.s }, p0/Z, [x21, x14, LSL #2]\n"
83 "ld1w { z28.s }, p0/Z, [x25, x14, LSL #2]\n"
84 "ld1w { z27.s }, p0/Z, [x28, x14, LSL #2]\n"
85 "ld1w { z26.s }, p0/Z, [x26, x14, LSL #2]\n"
86 "ld1w { z25.s }, p0/Z, [x23, x14, LSL #2]\n"
87 "ld1w { z24.s }, p0/Z, [x22, x14, LSL #2]\n"
88 "ld1w { z23.s }, p0/Z, [x20, x14, LSL #2]\n"
90 "whilelt p1.s, x14, x15\n"
93 "movprfx z22, z31\n fmax z22.s, p2/M, z22.s, z30.s\n"
94 "movprfx z21, z30\n fmax z21.s, p2/M, z21.s, z29.s\n"
95 "ld1w { z31.s }, p1/Z, [x27, x14, LSL #2]\n"
96 "ld1w { z30.s }, p1/Z, [x24, x14, LSL #2]\n"
97 "movprfx z20, z28\n fmax z20.s, p2/M, z20.s, z27.s\n"
98 "movprfx z19, z26\n fmax z19.s, p2/M, z19.s, z25.s\n"
99 "ld1w { z29.s }, p1/Z, [x21, x14, LSL #2]\n"
100 "ld1w { z27.s }, p1/Z, [x28, x14, LSL #2]\n"
101 "movprfx z17, z28\n fmax z17.s, p2/M, z17.s, z24.s\n"
102 "movprfx z18, z25\n fmax z18.s, p2/M, z18.s, z23.s\n"
103 "ld1w { z28.s }, p1/Z, [x25, x14, LSL #2]\n"
104 "ld1w { z26.s }, p1/Z, [x26, x14, LSL #2]\n"
105 "ld1w { z25.s }, p1/Z, [x23, x14, LSL #2]\n"
106 "ld1w { z24.s }, p1/Z, [x22, x14, LSL #2]\n"
107 "whilelt p0.s, x11, x15\n"
108 "movprfx z16, z22\n fmax z16.s, p2/M, z16.s, z20.s\n"
109 "ld1w { z23.s }, p1/Z, [x20, x14, LSL #2]\n"
111 "whilelt p1.s, x14, x15\n"
112 "st1w { z16.s }, p0, [x13, x11, LSL #2]\n"
113 "movprfx z16, z19\n fmax z16.s, p2/M, z16.s, z22.s\n"
114 "fmax z17.s, p2/M, z17.s, z21.s\n"
115 "st1w { z16.s }, p0, [x12, x11, LSL #2]\n"
116 "movprfx z16, z21\n fmax z16.s, p2/M, z16.s, z18.s\n"
117 "st1w { z17.s }, p0, [x10, x11, LSL #2]\n"
118 "st1w { z16.s }, p0, [x9, x11, LSL #2]\n"
122 "movprfx z22, z31\n fmax z22.s, p2/M, z22.s, z30.s\n"
123 "movprfx z21, z30\n fmax z21.s, p2/M, z21.s, z29.s\n"
124 "movprfx z20, z28\n fmax z20.s, p2/M, z20.s, z27.s\n"
125 "movprfx z19, z26\n fmax z19.s, p2/M, z19.s, z25.s\n"
126 "movprfx z17, z28\n fmax z17.s, p2/M, z17.s, z24.s\n"
127 "movprfx z18, z25\n fmax z18.s, p2/M, z18.s, z23.s\n"
128 "whilelt p0.s, x11, x15\n"
129 "movprfx z16, z22\n fmax z16.s, p2/M, z16.s, z20.s\n"
130 "st1w { z16.s }, p0, [x13, x11, LSL #2]\n"
131 "movprfx z16, z19\n fmax z16.s, p2/M, z16.s, z22.s\n"
132 "fmax z17.s, p2/M, z17.s, z21.s\n"
133 "st1w { z16.s }, p0, [x12, x11, LSL #2]\n"
134 "movprfx z16, z21\n fmax z16.s, p2/M, z16.s, z18.s\n"
135 "st1w { z17.s }, p0, [x10, x11, LSL #2]\n"
136 "st1w { z16.s }, p0, [x9, x11, LSL #2]\n"
138 : [
args]
"r" (&
args), [offsetof_inptrs]
"I" (offsetof(KernelArgs, inptrs)), [offsetof_n_channels]
"I" (offsetof(KernelArgs, n_channels)), [offsetof_outptrs]
"I" (offsetof(KernelArgs, outptrs))
139 :
"cc",
"memory",
"p0",
"p1",
"p2",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28",
"z16",
"z17",
"z18",
"z19",
"z20",
"z21",
"z22",
"z23",
"z24",
"z25",
"z26",
"z27",
"z28",
"z29",
"z30",
"z31"
146 #endif // defined(ARM_COMPUTE_ENABLE_SVE)