29 #if defined(ARM_COMPUTE_ENABLE_SME) && defined(__ARM_FP16_ARGS)
34 void sme_fp16_nhwc_max_2x2_s1_output2x2_depthfirst_impl(
35 const unsigned int n_channels,
36 const __fp16 *
const *
const inptrs,
37 __fp16 *
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 __fp16 *
const *
const inptrs;
49 __fp16 *
const *
const outptrs;
51 unsigned int channels,
52 const __fp16 *
const *input_ptrs,
53 __fp16 *
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 x21, [%x[args], %[offsetof_outptrs]]\n"
67 ".inst 0xd503477f // SMSTART ZA\n"
70 "ldr x20, [%x[args], %[offsetof_inptrs]]\n"
72 "ldr x13, [%x[args], %[offsetof_n_channels]]\n"
73 "whilelt p0.h, x15, x13\n"
74 "ldp x12, x11, [x21, #0x0]\n"
75 "ldp x10, x9, [x21, #0x10]\n"
76 "ldp x28, x27, [x20, #0x0]\n"
77 "ld1h { z30.h }, p0/Z, [x27, x15, LSL #1]\n"
78 "ldp x26, x25, [x20, #0x10]\n"
79 "ld1h { z29.h }, p0/Z, [x25, x15, LSL #1]\n"
80 "ldp x24, x23, [x20, #0x20]\n"
81 "ld1h { z28.h }, p0/Z, [x24, x15, LSL #1]\n"
82 "ldp x22, x21, [x20, #0x30]\n"
83 "ld1h { z27.h }, p0/Z, [x21, x15, LSL #1]\n"
84 "ldr x20, [x20, #0x40]\n"
85 "ld1h { z26.h }, p0/Z, [x28, x15, LSL #1]\n"
86 "ld1h { z25.h }, p0/Z, [x26, x15, LSL #1]\n"
87 "ld1h { z24.h }, p0/Z, [x23, x15, LSL #1]\n"
88 "ld1h { z19.h }, p0/Z, [x22, x15, LSL #1]\n"
89 "ld1h { z23.h }, p0/Z, [x20, x15, LSL #1]\n"
91 "whilelt p1.h, x15, x13\n"
94 "movprfx z22, z30\n fmax z22.h, p2/M, z22.h, z28.h\n"
95 "movprfx z21, z28\n fmax z21.h, p2/M, z21.h, z27.h\n"
96 "ld1h { z30.h }, p1/Z, [x27, x15, LSL #1]\n"
97 "whilelt p0.h, x14, x13\n"
98 "movprfx z18, z29\n fmax z18.h, p2/M, z18.h, z26.h\n"
99 "movprfx z17, z25\n fmax z17.h, p2/M, z17.h, z24.h\n"
100 "ld1h { z28.h }, p1/Z, [x24, x15, LSL #1]\n"
101 "movprfx z16, z29\n fmax z16.h, p2/M, z16.h, z19.h\n"
102 "movprfx z20, z24\n fmax z20.h, p2/M, z20.h, z23.h\n"
103 "ld1h { z27.h }, p1/Z, [x21, x15, LSL #1]\n"
104 "ld1h { z29.h }, p1/Z, [x25, x15, LSL #1]\n"
105 "movprfx z19, z22\n fmax z19.h, p2/M, z19.h, z18.h\n"
106 "movprfx z18, z17\n fmax z18.h, p2/M, z18.h, z22.h\n"
107 "ld1h { z26.h }, p1/Z, [x28, x15, LSL #1]\n"
108 "movprfx z17, z16\n fmax z17.h, p2/M, z17.h, z21.h\n"
109 "movprfx z16, z21\n fmax z16.h, p2/M, z16.h, z20.h\n"
110 "ld1h { z25.h }, p1/Z, [x26, x15, LSL #1]\n"
111 "st1h { z19.h }, p0, [x12, x14, LSL #1]\n"
112 "ld1h { z24.h }, p1/Z, [x23, x15, LSL #1]\n"
113 "st1h { z18.h }, p0, [x11, x14, LSL #1]\n"
114 "ld1h { z19.h }, p1/Z, [x22, x15, LSL #1]\n"
115 "st1h { z17.h }, p0, [x10, x14, LSL #1]\n"
116 "ld1h { z23.h }, p1/Z, [x20, x15, LSL #1]\n"
118 "whilelt p1.h, x15, x13\n"
119 "st1h { z16.h }, p0, [x9, x14, LSL #1]\n"
123 "movprfx z22, z30\n fmax z22.h, p2/M, z22.h, z28.h\n"
124 "movprfx z21, z28\n fmax z21.h, p2/M, z21.h, z27.h\n"
125 "whilelt p0.h, x14, x13\n"
126 "movprfx z20, z29\n fmax z20.h, p2/M, z20.h, z26.h\n"
127 "movprfx z18, z25\n fmax z18.h, p2/M, z18.h, z24.h\n"
128 "movprfx z17, z29\n fmax z17.h, p2/M, z17.h, z19.h\n"
129 "movprfx z19, z24\n fmax z19.h, p2/M, z19.h, z23.h\n"
130 "movprfx z16, z22\n fmax z16.h, p2/M, z16.h, z20.h\n"
131 "fmax z18.h, p2/M, z18.h, z22.h\n"
132 "st1h { z16.h }, p0, [x12, x14, LSL #1]\n"
133 "fmax z17.h, p2/M, z17.h, z21.h\n"
134 "movprfx z16, z21\n fmax z16.h, p2/M, z16.h, z19.h\n"
135 "st1h { z18.h }, p0, [x11, x14, LSL #1]\n"
136 "st1h { z17.h }, p0, [x10, x14, LSL #1]\n"
137 "st1h { z16.h }, p0, [x9, x14, LSL #1]\n"
138 ".inst 0xd503467f // SMSTOP\n"
140 : [
args]
"r" (&
args), [offsetof_inptrs]
"I" (offsetof(KernelArgs, inptrs)), [offsetof_n_channels]
"I" (offsetof(KernelArgs, n_channels)), [offsetof_outptrs]
"I" (offsetof(KernelArgs, outptrs))
141 :
"cc",
"memory",
"p0",
"p1",
"p2",
"p3",
"p4",
"p5",
"p6",
"p7",
"p8",
"p9",
"p10",
"p11",
"p12",
"p13",
"p14",
"p15",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28",
"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"
148 #endif // defined(ARM_COMPUTE_ENABLE_SME) && defined(__ARM_FP16_ARGS)