aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp174
1 files changed, 87 insertions, 87 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp
index a1a530b94e..75e4ddca15 100644
--- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -83,121 +83,121 @@ void sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst_impl(
__asm__ __volatile__(
"ldr x3, [%x[args], %[offsetof_n_channels]]\n"
- "mov x4, #0x0\n"
"ldr x20, [%x[args], %[offsetof_outptrs]]\n"
- "mov x5, #0x0\n"
- "ldr x6, [%x[args], %[offsetof_inptrs]]\n"
+ "mov x4, #0x0\n"
"mov x19, #0x4\n"
- "add x7, %x[args], %[offsetof_rescale]\n"
- "ldp x8, x17, [x20, #0x0]\n"
- "ldp x16, x15, [x20, #0x10]\n"
+ "ldr x5, [%x[args], %[offsetof_inptrs]]\n"
+ "ldp x6, x7, [x20, #0x0]\n"
"whilelt p0.h, XZR, x19\n"
- "ldp x14, x13, [x6, #0x0]\n"
"whilelt p1.h, x4, x3\n"
- "ldp x12, x11, [x6, #0x10]\n"
- "ldp x10, x9, [x6, #0x20]\n"
- "ldp x28, x27, [x6, #0x30]\n"
- "ldp x26, x25, [x6, #0x40]\n"
- "ldp x24, x23, [x6, #0x50]\n"
- "ldp x22, x21, [x6, #0x60]\n"
- "ldp x20, x19, [x6, #0x70]\n"
- "ld1rqh { z7.h }, p0/Z, [x7]\n"
- "ld1h { z8.h }, p1/Z, [x9, x4, LSL #1]\n"
+ "ldp x8, x17, [x20, #0x10]\n"
+ "ldp x16, x15, [x5, #0x0]\n"
+ "add x14, %x[args], %[offsetof_rescale]\n"
+ "mov x13, #0x0\n"
+ "ldp x12, x11, [x5, #0x10]\n"
+ "ldp x10, x9, [x5, #0x20]\n"
+ "ldp x28, x27, [x5, #0x30]\n"
+ "ldp x26, x25, [x5, #0x40]\n"
+ "ldp x24, x23, [x5, #0x50]\n"
+ "ldp x22, x21, [x5, #0x60]\n"
+ "ldp x20, x19, [x5, #0x70]\n"
+ "ld1h { z7.h }, p1/Z, [x9, x4, LSL #1]\n"
"ld1h { z6.h }, p1/Z, [x28, x4, LSL #1]\n"
"ld1h { z5.h }, p1/Z, [x25, x4, LSL #1]\n"
"ld1h { z4.h }, p1/Z, [x24, x4, LSL #1]\n"
- "ld1h { z3.h }, p1/Z, [x13, x4, LSL #1]\n"
+ "ld1h { z3.h }, p1/Z, [x15, x4, LSL #1]\n"
"ld1h { z2.h }, p1/Z, [x12, x4, LSL #1]\n"
"ld1h { z1.h }, p1/Z, [x10, x4, LSL #1]\n"
- "ld1h { z0.h }, p1/Z, [x26, x4, LSL #1]\n"
- "ld1h { z31.h }, p1/Z, [x27, x4, LSL #1]\n"
- "ld1h { z30.h }, p1/Z, [x23, x4, LSL #1]\n"
- "ld1h { z29.h }, p1/Z, [x21, x4, LSL #1]\n"
- "ld1h { z28.h }, p1/Z, [x20, x4, LSL #1]\n"
- "ld1h { z27.h }, p1/Z, [x14, x4, LSL #1]\n"
- "ld1h { z26.h }, p1/Z, [x11, x4, LSL #1]\n"
- "ld1h { z25.h }, p1/Z, [x22, x4, LSL #1]\n"
- "ld1h { z24.h }, p1/Z, [x19, x4, LSL #1]\n"
+ "ld1h { z31.h }, p1/Z, [x26, x4, LSL #1]\n"
+ "ld1h { z30.h }, p1/Z, [x27, x4, LSL #1]\n"
+ "ld1h { z29.h }, p1/Z, [x23, x4, LSL #1]\n"
+ "ld1h { z28.h }, p1/Z, [x21, x4, LSL #1]\n"
+ "ld1h { z27.h }, p1/Z, [x20, x4, LSL #1]\n"
+ "ld1h { z26.h }, p1/Z, [x16, x4, LSL #1]\n"
+ "ld1h { z25.h }, p1/Z, [x11, x4, LSL #1]\n"
+ "ld1h { z24.h }, p1/Z, [x22, x4, LSL #1]\n"
+ "ld1h { z23.h }, p1/Z, [x19, x4, LSL #1]\n"
"incw x4\n"
"whilelt p1.h, x4, x3\n"
+ "ld1rqh { z0.h }, p0/Z, [x14]\n"
"b.none 2f\n"
"1:" // Vector: Loop
- "fadd z17.h, z8.h, z6.h\n"
- "ld1h { z8.h }, p1/Z, [x9, x4, LSL #1]\n"
- "whilelt p0.h, x5, x3\n"
+ "fadd z17.h, z7.h, z6.h\n"
"fadd z16.h, z5.h, z4.h\n"
+ "ld1h { z7.h }, p1/Z, [x9, x4, LSL #1]\n"
"ld1h { z6.h }, p1/Z, [x28, x4, LSL #1]\n"
+ "fadd z19.h, z17.h, z16.h\n"
"fadd z18.h, z3.h, z2.h\n"
"ld1h { z5.h }, p1/Z, [x25, x4, LSL #1]\n"
- "fadd z23.h, z1.h, z0.h\n"
"ld1h { z4.h }, p1/Z, [x24, x4, LSL #1]\n"
- "fadd z22.h, z31.h, z30.h\n"
- "ld1h { z3.h }, p1/Z, [x13, x4, LSL #1]\n"
- "fadd z17.h, z17.h, z16.h\n"
+ "fadd z17.h, z1.h, z31.h\n"
+ "fadd z22.h, z30.h, z29.h\n"
+ "ld1h { z3.h }, p1/Z, [x15, x4, LSL #1]\n"
"ld1h { z2.h }, p1/Z, [x12, x4, LSL #1]\n"
- "fadd z16.h, z29.h, z28.h\n"
+ "fadd z16.h, z28.h, z27.h\n"
+ "fadd z21.h, z18.h, z19.h\n"
"ld1h { z1.h }, p1/Z, [x10, x4, LSL #1]\n"
- "fadd z19.h, z27.h, z23.h\n"
- "ld1h { z0.h }, p1/Z, [x26, x4, LSL #1]\n"
- "fadd z21.h, z18.h, z17.h\n"
- "ld1h { z31.h }, p1/Z, [x27, x4, LSL #1]\n"
- "fadd z20.h, z16.h, z17.h\n"
- "ld1h { z30.h }, p1/Z, [x23, x4, LSL #1]\n"
- "fadd z18.h, z26.h, z22.h\n"
- "ld1h { z29.h }, p1/Z, [x21, x4, LSL #1]\n"
- "fadd z17.h, z25.h, z23.h\n"
- "ld1h { z28.h }, p1/Z, [x20, x4, LSL #1]\n"
- "fadd z16.h, z24.h, z22.h\n"
- "ld1h { z27.h }, p1/Z, [x14, x4, LSL #1]\n"
- "fadd z19.h, z21.h, z19.h\n"
- "ld1h { z26.h }, p1/Z, [x11, x4, LSL #1]\n"
- "fadd z18.h, z21.h, z18.h\n"
- "ld1h { z25.h }, p1/Z, [x22, x4, LSL #1]\n"
- "fadd z17.h, z17.h, z20.h\n"
- "ld1h { z24.h }, p1/Z, [x19, x4, LSL #1]\n"
+ "ld1h { z31.h }, p1/Z, [x26, x4, LSL #1]\n"
+ "fadd z20.h, z16.h, z19.h\n"
+ "fadd z19.h, z26.h, z17.h\n"
+ "ld1h { z30.h }, p1/Z, [x27, x4, LSL #1]\n"
+ "ld1h { z29.h }, p1/Z, [x23, x4, LSL #1]\n"
+ "fadd z18.h, z25.h, z22.h\n"
+ "fadd z17.h, z24.h, z17.h\n"
+ "ld1h { z28.h }, p1/Z, [x21, x4, LSL #1]\n"
+ "ld1h { z27.h }, p1/Z, [x20, x4, LSL #1]\n"
+ "fadd z16.h, z23.h, z22.h\n"
+ "ld1h { z26.h }, p1/Z, [x16, x4, LSL #1]\n"
+ "ld1h { z25.h }, p1/Z, [x11, x4, LSL #1]\n"
+ "fadd z19.h, z19.h, z21.h\n"
+ "ld1h { z24.h }, p1/Z, [x22, x4, LSL #1]\n"
+ "ld1h { z23.h }, p1/Z, [x19, x4, LSL #1]\n"
"incw x4\n"
- "fadd z16.h, z20.h, z16.h\n"
+ "fadd z18.h, z18.h, z21.h\n"
+ "fadd z17.h, z17.h, z20.h\n"
+ "fadd z16.h, z16.h, z20.h\n"
+ "whilelt p0.h, x13, x3\n"
"whilelt p1.h, x4, x3\n"
- "fmul z19.h, z19.h, z7.h[0]\n"
- "st1h { z19.h }, p0, [x8, x5, LSL #1]\n"
- "fmul z18.h, z18.h, z7.h[1]\n"
- "fmul z17.h, z17.h, z7.h[2]\n"
- "st1h { z18.h }, p0, [x17, x5, LSL #1]\n"
- "fmul z16.h, z16.h, z7.h[3]\n"
- "st1h { z17.h }, p0, [x16, x5, LSL #1]\n"
- "st1h { z16.h }, p0, [x15, x5, LSL #1]\n"
- "incw x5\n"
+ "fmul z19.h, z19.h, z0.h[0]\n"
+ "fmul z18.h, z18.h, z0.h[1]\n"
+ "st1h { z19.h }, p0, [x6, x13, LSL #1]\n"
+ "fmul z17.h, z17.h, z0.h[2]\n"
+ "fmul z16.h, z16.h, z0.h[3]\n"
+ "st1h { z18.h }, p0, [x7, x13, LSL #1]\n"
+ "st1h { z17.h }, p0, [x8, x13, LSL #1]\n"
+ "st1h { z16.h }, p0, [x17, x13, LSL #1]\n"
+ "incw x13\n"
"b.any 1b\n"
"2:" // Vector: Tail
- "fadd z17.h, z8.h, z6.h\n"
- "whilelt p0.h, x5, x3\n"
+ "fadd z17.h, z7.h, z6.h\n"
"fadd z16.h, z5.h, z4.h\n"
+ "whilelt p0.h, x13, x3\n"
+ "fadd z19.h, z17.h, z16.h\n"
"fadd z18.h, z3.h, z2.h\n"
- "fadd z23.h, z1.h, z0.h\n"
- "fadd z17.h, z17.h, z16.h\n"
- "fadd z22.h, z31.h, z30.h\n"
- "fadd z16.h, z29.h, z28.h\n"
- "fadd z21.h, z18.h, z17.h\n"
- "fadd z19.h, z27.h, z23.h\n"
- "fadd z20.h, z16.h, z17.h\n"
- "fadd z18.h, z26.h, z22.h\n"
- "fadd z17.h, z25.h, z23.h\n"
- "fadd z16.h, z24.h, z22.h\n"
- "fadd z19.h, z21.h, z19.h\n"
- "fadd z18.h, z21.h, z18.h\n"
+ "fadd z17.h, z1.h, z31.h\n"
+ "fadd z22.h, z30.h, z29.h\n"
+ "fadd z16.h, z28.h, z27.h\n"
+ "fadd z21.h, z18.h, z19.h\n"
+ "fadd z20.h, z16.h, z19.h\n"
+ "fadd z19.h, z26.h, z17.h\n"
+ "fadd z18.h, z25.h, z22.h\n"
+ "fadd z17.h, z24.h, z17.h\n"
+ "fadd z16.h, z23.h, z22.h\n"
+ "fadd z19.h, z19.h, z21.h\n"
+ "fmul z19.h, z19.h, z0.h[0]\n"
+ "st1h { z19.h }, p0, [x6, x13, LSL #1]\n"
+ "fadd z18.h, z18.h, z21.h\n"
"fadd z17.h, z17.h, z20.h\n"
- "fadd z16.h, z20.h, z16.h\n"
- "fmul z19.h, z19.h, z7.h[0]\n"
- "st1h { z19.h }, p0, [x8, x5, LSL #1]\n"
- "fmul z18.h, z18.h, z7.h[1]\n"
- "fmul z17.h, z17.h, z7.h[2]\n"
- "st1h { z18.h }, p0, [x17, x5, LSL #1]\n"
- "fmul z16.h, z16.h, z7.h[3]\n"
- "st1h { z17.h }, p0, [x16, x5, LSL #1]\n"
- "st1h { z16.h }, p0, [x15, x5, LSL #1]\n"
+ "fmul z18.h, z18.h, z0.h[1]\n"
+ "fmul z17.h, z17.h, z0.h[2]\n"
+ "fadd z16.h, z16.h, z20.h\n"
+ "fmul z16.h, z16.h, z0.h[3]\n"
+ "st1h { z18.h }, p0, [x7, x13, LSL #1]\n"
+ "st1h { z17.h }, p0, [x8, x13, LSL #1]\n"
+ "st1h { z16.h }, p0, [x17, x13, LSL #1]\n"
:
: [args] "r" (&args), [offsetof_inptrs] "I" (offsetof(KernelArgs, inptrs)), [offsetof_n_channels] "I" (offsetof(KernelArgs, n_channels)), [offsetof_outptrs] "I" (offsetof(KernelArgs, outptrs)), [offsetof_rescale] "I" (offsetof(KernelArgs, rescale_vals))
- : "cc", "memory", "p0", "p1", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+ : "cc", "memory", "p0", "p1", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
);
}
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-20 19:37:58 +0000