aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
diff options
context:
space:
mode:
authorMichael Tyler <michael.tyler@arm.com>2023-01-17 11:04:14 +0000
committerGian Marco Iodice <gianmarco.iodice@arm.com>2023-01-18 09:43:38 +0000
commitbe13cead34e566bdd561ad3ffc3f645b460e482e (patch)
treecdc086de205d5a07fdd816afa6333d0b2f38d4e9 /src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
parent13bab71a76096985752a9e12711507021e25858d (diff)
downloadComputeLibrary-be13cead34e566bdd561ad3ffc3f645b460e482e.tar.gz
Revert "Update CPU kernels to remove x19"
This reverts commit 3c59f01c209d2732a15d97d65565ead964787a8b. Resolves: COMPMID-5817 Change-Id: Ie2443a21854a95db1e3d0cafa2121c0187a5e237 Signed-off-by: Michael Tyler <michael.tyler@arm.com> Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8974 Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Tested-by: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com> Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp620
1 files changed, 314 insertions, 306 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
index 9f21401840..e1b2d257b0 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021-2023 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -100,348 +100,356 @@ void sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
requant_muls, requant_shifts, outptrs);
__asm__ __volatile__(
- "mov x7, #0x0\n"
- "ldr x25, [%x[params], %[offsetof_Params_requant]]\n"
+ "ldr x5, [%x[params], %[offsetof_Params_n_channels]]\n"
"ptrue p4.b\n"
- "ldr x24, [%x[params], %[offsetof_Params_outptrs]]\n"
- "mov x23, x7\n"
- "add x21, x25, %[offsetof_Requantize32_a_offset]\n"
- "ldr x8, [%x[params], %[offsetof_Params_n_channels]]\n"
- "ldr x17, [%x[params], %[offsetof_Params_weights]]\n"
- "add x20, x25, %[offsetof_Requantize32_b_offset]\n"
- "add x22, x25, %[offsetof_Requantize32_c_offset]\n"
- "ld1rb { z23.b }, p4/Z, [x21]\n"
+ "ldr x6, [%x[params], %[offsetof_Params_weights]]\n"
+ "mov x7, #0x0\n"
+ "ldr x22, [%x[params], %[offsetof_Params_requant]]\n"
+ "mov x8, #0x0\n"
+ "ldr x17, [%x[params], %[offsetof_Params_requant_muls]]\n"
+ "add x16, %x[params], %[offsetof_Params_inptrs]\n"
+ "ldr x15, [%x[params], %[offsetof_Params_requant_shifts]]\n"
+ "add x19, x22, %[offsetof_Requantize32_a_offset]\n"
+ "ldr x21, [%x[params], %[offsetof_Params_outptrs]]\n"
+ "add x20, x22, %[offsetof_Requantize32_b_offset]\n"
+ "ld1rb { z19.b }, p4/Z, [x19]\n"
+ "add x19, x22, %[offsetof_Requantize32_c_offset]\n"
"ld1rb { z12.b }, p4/Z, [x20]\n"
- "add x21, x25, %[offsetof_Requantize32_minval]\n"
- "add x20, x25, %[offsetof_Requantize32_maxval]\n"
- "ld1rh { z14.h }, p4/Z, [x22]\n"
- "ld1rh { z16.h }, p4/Z, [x21]\n"
- "ld1rh { z15.h }, p4/Z, [x20]\n"
- "ldp x16, x15, [x24, #0x0]\n"
- "incw x23\n"
- "whilelt p3.h, x7, x8\n"
- "ldp x14, x13, [x24, #0x10]\n"
- "whilelt p2.s, x7, x8\n"
- "whilelt p1.s, x23, x8\n"
- "ldr x12, [%x[params], %[offsetof_Params_bias]]\n"
- "ld1b { z0.h }, p4/Z, [x17]\n"
- "ld1b { z1.h }, p4/Z, [x17, #1, MUL VL]\n"
- "add x11, %x[params], %[offsetof_Params_inptrs]\n"
- "mov x10, #0x0\n"
- "ld1b { z2.h }, p4/Z, [x17, #2, MUL VL]\n"
- "ld1b { z3.h }, p4/Z, [x17, #3, MUL VL]\n"
+ "add x20, x22, %[offsetof_Requantize32_minval]\n"
+ "ld1rw { z14.s }, p4/Z, [x19]\n"
+ "add x19, x22, %[offsetof_Requantize32_maxval]\n"
+ "ld1rw { z20.s }, p4/Z, [x20]\n"
+ "whilelt p3.h, x7, x5\n"
+ "ld1rw { z15.s }, p4/Z, [x19]\n"
+ "whilelt p2.s, x7, x5\n"
+ "ldp x14, x13, [x21, #0x0]\n"
+ "mov x19, x7\n"
+ "incw x19\n"
+ "ldp x12, x11, [x21, #0x10]\n"
+ "whilelt p1.s, x19, x5\n"
+ "ldr x19, [%x[params], %[offsetof_Params_bias]]\n"
+ "ld1w { z18.s }, p2/Z, [x19]\n"
+ "ld1w { z16.s }, p1/Z, [x19, #1, MUL VL]\n"
+ "uzp1 z13.s, z18.s, z16.s\n"
+ "addvl x19, x19, #2\n"
+ "str x19, [%x[params], %[offsetof_Params_bias]]\n"
+ "uzp2 z16.s, z18.s, z16.s\n"
+ "mov z11.d, z13.d\n"
+ "ld1b { z0.h }, p4/Z, [x6]\n"
".inst 0x454c1800 // usublb z0.h, z0.b, z12.b\n"
+ "mov z9.d, z16.d\n"
+ "ld1b { z1.h }, p4/Z, [x6, #1, MUL VL]\n"
+ "mov z18.d, z13.d\n"
+ "ld1b { z2.h }, p4/Z, [x6, #2, MUL VL]\n"
".inst 0x454c1821 // usublb z1.h, z1.b, z12.b\n"
- "ld1b { z4.h }, p4/Z, [x17, #4, MUL VL]\n"
- "ld1b { z5.h }, p4/Z, [x17, #5, MUL VL]\n"
+ "mov z10.d, z16.d\n"
+ "ld1b { z3.h }, p4/Z, [x6, #3, MUL VL]\n"
+ "mov z22.d, z13.d\n"
+ "ld1b { z4.h }, p4/Z, [x6, #4, MUL VL]\n"
".inst 0x454c1842 // usublb z2.h, z2.b, z12.b\n"
+ "mov z23.d, z16.d\n"
+ "ld1b { z5.h }, p4/Z, [x6, #5, MUL VL]\n"
".inst 0x454c1863 // usublb z3.h, z3.b, z12.b\n"
- "ld1b { z6.h }, p4/Z, [x17, #6, MUL VL]\n"
- "ld1b { z7.h }, p4/Z, [x17, #7, MUL VL]\n"
- "inch x17, ALL, MUL #8\n"
+ "ld1b { z6.h }, p4/Z, [x6, #6, MUL VL]\n"
+ "ld1b { z7.h }, p4/Z, [x6, #7, MUL VL]\n"
".inst 0x454c1884 // usublb z4.h, z4.b, z12.b\n"
- "ld1w { z18.s }, p2/Z, [x12]\n"
- "ld1w { z8.s }, p1/Z, [x12, #1, MUL VL]\n"
- "uzp1 z13.s, z18.s, z8.s\n"
- "uzp2 z17.s, z18.s, z8.s\n"
- "ld1b { z8.h }, p4/Z, [x17]\n"
- "ldp x9, x28, [x11, #0x0]\n"
- "addvl x12, x12, #2\n"
- "mov z9.d, z13.d\n"
- "ldp x25, x24, [x11, #0x10]\n"
- "ldp x23, x22, [x11, #0x20]\n"
- "mov z10.d, z17.d\n"
- "mov z11.d, z13.d\n"
- "ldp x21, x20, [x11, #0x30]\n"
- "ld1b { z31.h }, p3/Z, [x9, x7]\n"
- "mov z22.d, z17.d\n"
- "mov z21.d, z13.d\n"
- "ld1b { z30.h }, p3/Z, [x28, x7]\n"
- "ld1b { z29.h }, p3/Z, [x25, x7]\n"
- "mov z18.d, z17.d\n"
+ "inch x6, ALL, MUL #8\n"
+ "ld1b { z8.h }, p4/Z, [x6]\n"
+ "ldp x26, x25, [x16, #0x0]\n"
".inst 0x454c18a5 // usublb z5.h, z5.b, z12.b\n"
- "ld1b { z28.h }, p3/Z, [x24, x7]\n"
- "ld1b { z27.h }, p3/Z, [x23, x7]\n"
".inst 0x454c18c6 // usublb z6.h, z6.b, z12.b\n"
+ "ldp x24, x23, [x16, #0x10]\n"
".inst 0x454c18e7 // usublb z7.h, z7.b, z12.b\n"
- "ld1b { z26.h }, p3/Z, [x22, x7]\n"
- "ld1b { z25.h }, p3/Z, [x21, x7]\n"
".inst 0x454c1908 // usublb z8.h, z8.b, z12.b\n"
- ".inst 0x45571bff // usublb z31.h, z31.b, z23.b\n"
- "ld1b { z24.h }, p3/Z, [x20, x7]\n"
- "ldr x27, [%x[params], %[offsetof_Params_requant_muls]]\n"
- ".inst 0x45571bde // usublb z30.h, z30.b, z23.b\n"
- ".inst 0x45571bbd // usublb z29.h, z29.b, z23.b\n"
- "ldr x26, [%x[params], %[offsetof_Params_requant_shifts]]\n"
- "str x12, [%x[params], %[offsetof_Params_bias]]\n"
- ".inst 0x45571b9c // usublb z28.h, z28.b, z23.b\n"
- ".inst 0x45571b7b // usublb z27.h, z27.b, z23.b\n"
- ".inst 0x45571b5a // usublb z26.h, z26.b, z23.b\n"
- ".inst 0x45571b39 // usublb z25.h, z25.b, z23.b\n"
- ".inst 0x45571b18 // usublb z24.h, z24.b, z23.b\n"
+ "ldp x22, x21, [x16, #0x20]\n"
+ "ldp x20, x19, [x16, #0x30]\n"
+ "ld1b { z31.h }, p3/Z, [x26, x7]\n"
+ ".inst 0x45531bff // usublb z31.h, z31.b, z19.b\n"
+ "ld1b { z30.h }, p3/Z, [x25, x7]\n"
+ "ld1b { z29.h }, p3/Z, [x24, x7]\n"
+ ".inst 0x45531bde // usublb z30.h, z30.b, z19.b\n"
+ "ld1b { z28.h }, p3/Z, [x23, x7]\n"
+ "ld1b { z27.h }, p3/Z, [x22, x7]\n"
+ ".inst 0x45531bbd // usublb z29.h, z29.b, z19.b\n"
+ "ld1b { z26.h }, p3/Z, [x21, x7]\n"
+ ".inst 0x45531b9c // usublb z28.h, z28.b, z19.b\n"
+ "ld1b { z25.h }, p3/Z, [x20, x7]\n"
+ "ld1b { z24.h }, p3/Z, [x19, x7]\n"
+ ".inst 0x45531b7b // usublb z27.h, z27.b, z19.b\n"
+ ".inst 0x45531b5a // usublb z26.h, z26.b, z19.b\n"
+ ".inst 0x45531b39 // usublb z25.h, z25.b, z19.b\n"
+ ".inst 0x45531b18 // usublb z24.h, z24.b, z19.b\n"
"1:" // Loop
".inst 0x448843ed // smlalb z13.s, p4/M, z31.h, z8.h\n"
- ".inst 0x448847f1 // smlalt z17.s, p4/M, z31.h, z8.h\n"
- "ldr x25, [x11, #0x40]\n"
- "ldr x24, [x11, #0x48]\n"
- ".inst 0x448643e9 // smlalb z9.s, p4/M, z31.h, z6.h\n"
- ".inst 0x448647ea // smlalt z10.s, p4/M, z31.h, z6.h\n"
- "ldr x22, [x11, #0x50]\n"
- "ldr x20, [x11, #0x58]\n"
+ "ldr x23, [x16, #0x40]\n"
+ "whilelt p0.h, x8, x5\n"
+ ".inst 0x448847f0 // smlalt z16.s, p4/M, z31.h, z8.h\n"
+ "ldr x22, [x16, #0x48]\n"
+ "inch x6\n"
+ ".inst 0x448643eb // smlalb z11.s, p4/M, z31.h, z6.h\n"
+ "ldr x21, [x16, #0x50]\n"
+ ".inst 0x448647e9 // smlalt z9.s, p4/M, z31.h, z6.h\n"
+ "ldr x20, [x16, #0x58]\n"
+ ".inst 0x448243f2 // smlalb z18.s, p4/M, z31.h, z2.h\n"
+ "ldr x19, [x16, #0x60]\n"
+ ".inst 0x448247ea // smlalt z10.s, p4/M, z31.h, z2.h\n"
+ "ldr x10, [x16, #0x68]\n"
+ ".inst 0x448043f6 // smlalb z22.s, p4/M, z31.h, z0.h\n"
+ "ldr x9, [x16, #0x70]\n"
+ ".inst 0x448047f7 // smlalt z23.s, p4/M, z31.h, z0.h\n"
+ "ldr x28, [x16, #0x78]\n"
".inst 0x448043cd // smlalb z13.s, p4/M, z30.h, z0.h\n"
- ".inst 0x448047d1 // smlalt z17.s, p4/M, z30.h, z0.h\n"
- "ldr x23, [x11, #0x78]\n"
- "ldr x21, [x11, #0x60]\n"
- ".inst 0x44814389 // smlalb z9.s, p4/M, z28.h, z1.h\n"
- ".inst 0x4481478a // smlalt z10.s, p4/M, z28.h, z1.h\n"
- "ld1b { z28.h }, p3/Z, [x24, x7]\n"
- ".inst 0x45571b9c // usublb z28.h, z28.b, z23.b\n"
+ "ldr x27, [x16, #0x80]\n"
+ ".inst 0x448047d0 // smlalt z16.s, p4/M, z30.h, z0.h\n"
+ "ldr x26, [x16, #0x88]\n"
+ ".inst 0x4481438b // smlalb z11.s, p4/M, z28.h, z1.h\n"
+ "ldr x25, [x16, #0x90]\n"
+ ".inst 0x44814789 // smlalt z9.s, p4/M, z28.h, z1.h\n"
+ "ld1b { z28.h }, p3/Z, [x22, x7]\n"
+ ".inst 0x45531b9c // usublb z28.h, z28.b, z19.b\n"
".inst 0x448143ad // smlalb z13.s, p4/M, z29.h, z1.h\n"
- ".inst 0x448147b1 // smlalt z17.s, p4/M, z29.h, z1.h\n"
- "ld1b { z29.h }, p3/Z, [x25, x7]\n"
- ".inst 0x45571bbd // usublb z29.h, z29.b, z23.b\n"
- ".inst 0x44824369 // smlalb z9.s, p4/M, z27.h, z2.h\n"
- ".inst 0x4482476a // smlalt z10.s, p4/M, z27.h, z2.h\n"
- "ld1b { z27.h }, p3/Z, [x22, x7]\n"
- ".inst 0x45571b7b // usublb z27.h, z27.b, z23.b\n"
+ "ldr x24, [x16, #0x98]\n"
+ ".inst 0x448147b0 // smlalt z16.s, p4/M, z29.h, z1.h\n"
+ "ld1b { z29.h }, p3/Z, [x23, x7]\n"
+ ".inst 0x45531bbd // usublb z29.h, z29.b, z19.b\n"
+ ".inst 0x4482436b // smlalb z11.s, p4/M, z27.h, z2.h\n"
+ "ldr x23, [x16, #0xa0]\n"
+ ".inst 0x44824769 // smlalt z9.s, p4/M, z27.h, z2.h\n"
+ "ld1b { z27.h }, p3/Z, [x21, x7]\n"
+ ".inst 0x45531b7b // usublb z27.h, z27.b, z19.b\n"
".inst 0x4483434d // smlalb z13.s, p4/M, z26.h, z3.h\n"
- ".inst 0x44834751 // smlalt z17.s, p4/M, z26.h, z3.h\n"
+ "ldr x22, [x16, #0xa8]\n"
+ ".inst 0x44834750 // smlalt z16.s, p4/M, z26.h, z3.h\n"
"ld1b { z26.h }, p3/Z, [x20, x7]\n"
- ".inst 0x45571b5a // usublb z26.h, z26.b, z23.b\n"
- ".inst 0x44804309 // smlalb z9.s, p4/M, z24.h, z0.h\n"
- ".inst 0x4480470a // smlalt z10.s, p4/M, z24.h, z0.h\n"
- "ldr x22, [x11, #0x80]\n"
- "ldr x20, [x11, #0x68]\n"
+ ".inst 0x45531b5a // usublb z26.h, z26.b, z19.b\n"
".inst 0x4484432d // smlalb z13.s, p4/M, z25.h, z4.h\n"
- ".inst 0x44844731 // smlalt z17.s, p4/M, z25.h, z4.h\n"
- "ld1b { z25.h }, p3/Z, [x21, x7]\n"
- ".inst 0x45571b39 // usublb z25.h, z25.b, z23.b\n"
- ".inst 0x448443a9 // smlalb z9.s, p4/M, z29.h, z4.h\n"
- ".inst 0x448447aa // smlalt z10.s, p4/M, z29.h, z4.h\n"
- "ldr x21, [x11, #0x88]\n"
- "ld1b { z29.h }, p3/Z, [x20, x7]\n"
+ "ldr x21, [x16, #0xb0]\n"
+ ".inst 0x44844730 // smlalt z16.s, p4/M, z25.h, z4.h\n"
+ "ld1b { z25.h }, p3/Z, [x19, x7]\n"
+ ".inst 0x45531b39 // usublb z25.h, z25.b, z19.b\n"
".inst 0x4482430d // smlalb z13.s, p4/M, z24.h, z2.h\n"
- ".inst 0x44824711 // smlalt z17.s, p4/M, z24.h, z2.h\n"
- "ldr x20, [x11, #0x70]\n"
- ".inst 0x45571bbd // usublb z29.h, z29.b, z23.b\n"
- ".inst 0x44854389 // smlalb z9.s, p4/M, z28.h, z5.h\n"
- ".inst 0x4485478a // smlalt z10.s, p4/M, z28.h, z5.h\n"
- "ld1b { z28.h }, p3/Z, [x22, x7]\n"
- ".inst 0x45571b9c // usublb z28.h, z28.b, z23.b\n"
- ".inst 0x448243eb // smlalb z11.s, p4/M, z31.h, z2.h\n"
- ".inst 0x448247f6 // smlalt z22.s, p4/M, z31.h, z2.h\n"
- "ldr x25, [x11, #0x98]\n"
- "ld1b { z24.h }, p3/Z, [x20, x7]\n"
+ "ldr x20, [x16, #0xb8]\n"
+ ".inst 0x44824710 // smlalt z16.s, p4/M, z24.h, z2.h\n"
+ "ldr x19, [x16, #0xc0]\n"
+ ".inst 0x4480430b // smlalb z11.s, p4/M, z24.h, z0.h\n"
+ "ld1w { z21.s }, p2/Z, [x17]\n"
+ ".inst 0x44804709 // smlalt z9.s, p4/M, z24.h, z0.h\n"
+ "ld1b { z24.h }, p3/Z, [x9, x7]\n"
+ ".inst 0x45531b18 // usublb z24.h, z24.b, z19.b\n"
+ ".inst 0x448443ab // smlalb z11.s, p4/M, z29.h, z4.h\n"
+ "ld1w { z17.s }, p1/Z, [x17, #1, MUL VL]\n"
+ ".inst 0x448447a9 // smlalt z9.s, p4/M, z29.h, z4.h\n"
+ "ld1b { z29.h }, p3/Z, [x10, x7]\n"
+ "addvl x17, x17, #2\n"
".inst 0x4485436d // smlalb z13.s, p4/M, z27.h, z5.h\n"
- ".inst 0x44854771 // smlalt z17.s, p4/M, z27.h, z5.h\n"
- ".inst 0x45571b18 // usublb z24.h, z24.b, z23.b\n"
- "ldr x24, [x11, #0x90]\n"
- ".inst 0x44834369 // smlalb z9.s, p4/M, z27.h, z3.h\n"
- ".inst 0x4483476a // smlalt z10.s, p4/M, z27.h, z3.h\n"
- "ld1b { z27.h }, p3/Z, [x23, x7]\n"
- ".inst 0x45571b7b // usublb z27.h, z27.b, z23.b\n"
- ".inst 0x448043f5 // smlalb z21.s, p4/M, z31.h, z0.h\n"
- ".inst 0x4483434b // smlalb z11.s, p4/M, z26.h, z3.h\n"
- "ldr x23, [x11, #0xa8]\n"
- "ldr x20, [x11, #0xa0]\n"
- ".inst 0x44834756 // smlalt z22.s, p4/M, z26.h, z3.h\n"
- ".inst 0x448047f2 // smlalt z18.s, p4/M, z31.h, z0.h\n"
- "ld1b { z26.h }, p3/Z, [x21, x7]\n"
- ".inst 0x45571b5a // usublb z26.h, z26.b, z23.b\n"
- ".inst 0x44844375 // smlalb z21.s, p4/M, z27.h, z4.h\n"
- ".inst 0x4480432b // smlalb z11.s, p4/M, z25.h, z0.h\n"
- "ldr x22, [x11, #0xb0]\n"
- "ldr x21, [x11, #0xb8]\n"
- ".inst 0x44804736 // smlalt z22.s, p4/M, z25.h, z0.h\n"
- ".inst 0x44844772 // smlalt z18.s, p4/M, z27.h, z4.h\n"
- "ld1b { z27.h }, p3/Z, [x20, x7]\n"
- ".inst 0x45571b7b // usublb z27.h, z27.b, z23.b\n"
- ".inst 0x44814395 // smlalb z21.s, p4/M, z28.h, z1.h\n"
+ ".inst 0x45531bbd // usublb z29.h, z29.b, z19.b\n"
+ "uzp1 z30.s, z21.s, z17.s\n"
+ "uzp2 z31.s, z21.s, z17.s\n"
+ "ld1w { z21.s }, p2/Z, [x15]\n"
+ ".inst 0x4485438b // smlalb z11.s, p4/M, z28.h, z5.h\n"
+ "ld1w { z17.s }, p1/Z, [x15, #1, MUL VL]\n"
+ "addvl x15, x15, #2\n"
+ ".inst 0x44854789 // smlalt z9.s, p4/M, z28.h, z5.h\n"
+ "ld1b { z28.h }, p3/Z, [x27, x7]\n"
+ ".inst 0x45531b9c // usublb z28.h, z28.b, z19.b\n"
+ ".inst 0x44854770 // smlalt z16.s, p4/M, z27.h, z5.h\n"
+ ".inst 0x4483436b // smlalb z11.s, p4/M, z27.h, z3.h\n"
+ ".inst 0x44834769 // smlalt z9.s, p4/M, z27.h, z3.h\n"
+ "ld1b { z27.h }, p3/Z, [x28, x7]\n"
+ ".inst 0x45531b7b // usublb z27.h, z27.b, z19.b\n"
+ ".inst 0x44834352 // smlalb z18.s, p4/M, z26.h, z3.h\n"
+ ".inst 0x4483474a // smlalt z10.s, p4/M, z26.h, z3.h\n"
+ "ld1b { z26.h }, p3/Z, [x26, x7]\n"
+ ".inst 0x45531b5a // usublb z26.h, z26.b, z19.b\n"
".inst 0x4486432d // smlalb z13.s, p4/M, z25.h, z6.h\n"
- "ldr x20, [x11, #0xc0]\n"
- "ld1w { z31.s }, p2/Z, [x27]\n"
- ".inst 0x44864731 // smlalt z17.s, p4/M, z25.h, z6.h\n"
- ".inst 0x448443ab // smlalb z11.s, p4/M, z29.h, z4.h\n"
- "ld1b { z25.h }, p3/Z, [x24, x7]\n"
- ".inst 0x45571b39 // usublb z25.h, z25.b, z23.b\n"
- ".inst 0x448447b6 // smlalt z22.s, p4/M, z29.h, z4.h\n"
- "ld1b { z29.h }, p3/Z, [x25, x7]\n"
- ".inst 0x44814792 // smlalt z18.s, p4/M, z28.h, z1.h\n"
- ".inst 0x45571bbd // usublb z29.h, z29.b, z23.b\n"
- ".inst 0x44854355 // smlalb z21.s, p4/M, z26.h, z5.h\n"
+ ".inst 0x44864730 // smlalt z16.s, p4/M, z25.h, z6.h\n"
+ ".inst 0x44804332 // smlalb z18.s, p4/M, z25.h, z0.h\n"
+ ".inst 0x4480472a // smlalt z10.s, p4/M, z25.h, z0.h\n"
+ "ld1b { z25.h }, p3/Z, [x25, x7]\n"
+ ".inst 0x45531b39 // usublb z25.h, z25.b, z19.b\n"
+ "uzp1 z0.s, z21.s, z17.s\n"
+ "uzp2 z21.s, z21.s, z17.s\n"
+ ".inst 0x448443b2 // smlalb z18.s, p4/M, z29.h, z4.h\n"
+ ".inst 0x448447aa // smlalt z10.s, p4/M, z29.h, z4.h\n"
+ "ld1b { z29.h }, p3/Z, [x24, x7]\n"
+ ".inst 0x45531bbd // usublb z29.h, z29.b, z19.b\n"
".inst 0x4487430d // smlalb z13.s, p4/M, z24.h, z7.h\n"
- "ld1w { z20.s }, p1/Z, [x27, #1, MUL VL]\n"
- "uzp1 z19.s, z31.s, z20.s\n"
- ".inst 0x44874711 // smlalt z17.s, p4/M, z24.h, z7.h\n"
- ".inst 0x4481430b // smlalb z11.s, p4/M, z24.h, z1.h\n"
- "uzp2 z30.s, z31.s, z20.s\n"
- "ld1w { z31.s }, p2/Z, [x26]\n"
- ".inst 0x44814716 // smlalt z22.s, p4/M, z24.h, z1.h\n"
- "ld1b { z24.h }, p3/Z, [x23, x7]\n"
- ".inst 0x44854752 // smlalt z18.s, p4/M, z26.h, z5.h\n"
- ".inst 0x45571b18 // usublb z24.h, z24.b, z23.b\n"
- ".inst 0x448243b5 // smlalb z21.s, p4/M, z29.h, z2.h\n"
- "ld1b { z26.h }, p3/Z, [x22, x7]\n"
- ".inst 0x448247b2 // smlalt z18.s, p4/M, z29.h, z2.h\n"
- ".inst 0x45571b5a // usublb z26.h, z26.b, z23.b\n"
- ".inst 0x4486432b // smlalb z11.s, p4/M, z25.h, z6.h\n"
- ".inst 0x44834315 // smlalb z21.s, p4/M, z24.h, z3.h\n"
- "ld1w { z20.s }, p1/Z, [x26, #1, MUL VL]\n"
- "uzp1 z1.s, z31.s, z20.s\n"
- ".inst 0x44874389 // smlalb z9.s, p4/M, z28.h, z7.h\n"
- ".inst 0x4487478a // smlalt z10.s, p4/M, z28.h, z7.h\n"
- ".inst 0x04b375ad // sqrdmulh z13.s, z13.s, z19.s\n"
- "whilelt p0.h, x10, x8\n"
- ".inst 0x44864736 // smlalt z22.s, p4/M, z25.h, z6.h\n"
- "ld1b { z25.h }, p3/Z, [x21, x7]\n"
- ".inst 0x44834712 // smlalt z18.s, p4/M, z24.h, z3.h\n"
- ".inst 0x45571b39 // usublb z25.h, z25.b, z23.b\n"
- ".inst 0x4487436b // smlalb z11.s, p4/M, z27.h, z7.h\n"
- ".inst 0x44874355 // smlalb z21.s, p4/M, z26.h, z7.h\n"
- "uzp2 z31.s, z31.s, z20.s\n"
- "inch x17\n"
- ".inst 0x448843a9 // smlalb z9.s, p4/M, z29.h, z8.h\n"
- ".inst 0x448847aa // smlalt z10.s, p4/M, z29.h, z8.h\n"
- "ld1b { z29.h }, p3/Z, [x20, x7]\n"
- ".inst 0x45571bbd // usublb z29.h, z29.b, z23.b\n"
- ".inst 0x44874776 // smlalt z22.s, p4/M, z27.h, z7.h\n"
- ".inst 0x44874752 // smlalt z18.s, p4/M, z26.h, z7.h\n"
- "and z0.d, z13.d, z1.d\n"
+ ".inst 0x44874710 // smlalt z16.s, p4/M, z24.h, z7.h\n"
+ ".inst 0x44814312 // smlalb z18.s, p4/M, z24.h, z1.h\n"
+ ".inst 0x4481470a // smlalt z10.s, p4/M, z24.h, z1.h\n"
+ "ld1b { z24.h }, p3/Z, [x22, x7]\n"
+ ".inst 0x45531b18 // usublb z24.h, z24.b, z19.b\n"
+ ".inst 0x04be75ad // sqrdmulh z13.s, z13.s, z30.s\n"
+ ".inst 0x04bf7610 // sqrdmulh z16.s, z16.s, z31.s\n"
+ ".inst 0x44844376 // smlalb z22.s, p4/M, z27.h, z4.h\n"
+ ".inst 0x44844777 // smlalt z23.s, p4/M, z27.h, z4.h\n"
+ "ld1b { z27.h }, p3/Z, [x23, x7]\n"
+ ".inst 0x45531b7b // usublb z27.h, z27.b, z19.b\n"
+ "and z4.d, z13.d, z0.d\n"
+ "and z17.d, z16.d, z21.d\n"
+ "asr z4.s, z4.s, #0x1f\n"
+ ".inst 0x4487438b // smlalb z11.s, p4/M, z28.h, z7.h\n"
+ ".inst 0x44874789 // smlalt z9.s, p4/M, z28.h, z7.h\n"
+ "asr z17.s, z17.s, #0x1f\n"
+ ".inst 0x44814396 // smlalb z22.s, p4/M, z28.h, z1.h\n"
+ ".inst 0x44814797 // smlalt z23.s, p4/M, z28.h, z1.h\n"
+ ".inst 0x44864332 // smlalb z18.s, p4/M, z25.h, z6.h\n"
+ ".inst 0x4486472a // smlalt z10.s, p4/M, z25.h, z6.h\n"
+ "ld1b { z25.h }, p3/Z, [x20, x7]\n"
+ ".inst 0x45531b39 // usublb z25.h, z25.b, z19.b\n"
+ "sqadd z13.s, z13.s, z4.s\n"
+ "sqadd z16.s, z16.s, z17.s\n"
+ ".inst 0x44854356 // smlalb z22.s, p4/M, z26.h, z5.h\n"
+ ".inst 0x44854757 // smlalt z23.s, p4/M, z26.h, z5.h\n"
+ "ld1b { z26.h }, p3/Z, [x21, x7]\n"
+ ".inst 0x45531b5a // usublb z26.h, z26.b, z19.b\n"
+ ".inst 0x448843ab // smlalb z11.s, p4/M, z29.h, z8.h\n"
+ ".inst 0x448847a9 // smlalt z9.s, p4/M, z29.h, z8.h\n"
+ ".inst 0x448243b6 // smlalb z22.s, p4/M, z29.h, z2.h\n"
+ ".inst 0x448247b7 // smlalt z23.s, p4/M, z29.h, z2.h\n"
+ "ld1b { z29.h }, p3/Z, [x19, x7]\n"
"inch x7\n"
- ".inst 0x4485430b // smlalb z11.s, p4/M, z24.h, z5.h\n"
- ".inst 0x44864335 // smlalb z21.s, p4/M, z25.h, z6.h\n"
- ".inst 0x04be7631 // sqrdmulh z17.s, z17.s, z30.s\n"
- "mov x20, x7\n"
- ".inst 0x44854716 // smlalt z22.s, p4/M, z24.h, z5.h\n"
- ".inst 0x44864732 // smlalt z18.s, p4/M, z25.h, z6.h\n"
- "asr z0.s, z0.s, #0x1f\n"
- "incw x20\n"
- ".inst 0x4488432b // smlalb z11.s, p4/M, z25.h, z8.h\n"
- ".inst 0x448843b5 // smlalb z21.s, p4/M, z29.h, z8.h\n"
- "and z20.d, z17.d, z31.d\n"
- "whilelt p2.s, x7, x8\n"
- ".inst 0x44884736 // smlalt z22.s, p4/M, z25.h, z8.h\n"
- ".inst 0x448847b2 // smlalt z18.s, p4/M, z29.h, z8.h\n"
- ".inst 0x04b37529 // sqrdmulh z9.s, z9.s, z19.s\n"
- "whilelt p1.s, x20, x8\n"
- ".inst 0x04b3756b // sqrdmulh z11.s, z11.s, z19.s\n"
- ".inst 0x04b376b5 // sqrdmulh z21.s, z21.s, z19.s\n"
- "ldr x12, [%x[params], %[offsetof_Params_bias]]\n"
- "whilelt p3.h, x7, x8\n"
- "sqadd z13.s, z13.s, z0.s\n"
- "asr z20.s, z20.s, #0x1f\n"
- ".inst 0x4482902d // srshl z13.s, p4/M, z13.s, z1.s\n"
- "addvl x27, x27, #2\n"
- "and z19.d, z9.d, z1.d\n"
- ".inst 0x04be754a // sqrdmulh z10.s, z10.s, z30.s\n"
- "addvl x26, x26, #2\n"
- "and z2.d, z11.d, z1.d\n"
- ".inst 0x04be76d6 // sqrdmulh z22.s, z22.s, z30.s\n"
- "and z0.d, z21.d, z1.d\n"
+ ".inst 0x04be756b // sqrdmulh z11.s, z11.s, z30.s\n"
+ "whilelt p2.s, x7, x5\n"
+ ".inst 0x04bf7529 // sqrdmulh z9.s, z9.s, z31.s\n"
+ "mov x19, x7\n"
+ ".inst 0x44874372 // smlalb z18.s, p4/M, z27.h, z7.h\n"
+ ".inst 0x45531bbd // usublb z29.h, z29.b, z19.b\n"
+ ".inst 0x4487476a // smlalt z10.s, p4/M, z27.h, z7.h\n"
+ "incw x19\n"
+ ".inst 0x44834316 // smlalb z22.s, p4/M, z24.h, z3.h\n"
+ "whilelt p1.s, x19, x5\n"
+ "and z1.d, z11.d, z0.d\n"
+ "whilelt p3.h, x7, x5\n"
+ "and z17.d, z9.d, z21.d\n"
+ "asr z1.s, z1.s, #0x1f\n"
+ ".inst 0x44854312 // smlalb z18.s, p4/M, z24.h, z5.h\n"
+ ".inst 0x4485470a // smlalt z10.s, p4/M, z24.h, z5.h\n"
+ "asr z17.s, z17.s, #0x1f\n"
+ ".inst 0x44834717 // smlalt z23.s, p4/M, z24.h, z3.h\n"
+ ".inst 0x44874356 // smlalb z22.s, p4/M, z26.h, z7.h\n"
+ ".inst 0x4482900d // srshl z13.s, p4/M, z13.s, z0.s\n"
+ ".inst 0x44884332 // smlalb z18.s, p4/M, z25.h, z8.h\n"
+ "sqadd z11.s, z11.s, z1.s\n"
+ "sqadd z9.s, z9.s, z17.s\n"
+ "add z13.s, z13.s, z14.s\n"
".inst 0x04be7652 // sqrdmulh z18.s, z18.s, z30.s\n"
- "sqadd z17.s, z17.s, z20.s\n"
- "asr z19.s, z19.s, #0x1f\n"
- ".inst 0x448293f1 // srshl z17.s, p4/M, z17.s, z31.s\n"
- "and z3.d, z10.d, z31.d\n"
+ ".inst 0x44874757 // smlalt z23.s, p4/M, z26.h, z7.h\n"
+ ".inst 0x4488472a // smlalt z10.s, p4/M, z25.h, z8.h\n"
+ ".inst 0x44864336 // smlalb z22.s, p4/M, z25.h, z6.h\n"
+ "and z17.d, z18.d, z0.d\n"
+ "asr z17.s, z17.s, #0x1f\n"
+ ".inst 0x04bf754a // sqrdmulh z10.s, z10.s, z31.s\n"
+ ".inst 0x44864737 // smlalt z23.s, p4/M, z25.h, z6.h\n"
+ ".inst 0x448843b6 // smlalb z22.s, p4/M, z29.h, z8.h\n"
+ "smin z13.s, p4/M, z13.s, z15.s\n"
+ ".inst 0x448292b0 // srshl z16.s, p4/M, z16.s, z21.s\n"
+ "and z1.d, z10.d, z21.d\n"
+ "asr z1.s, z1.s, #0x1f\n"
+ "add z16.s, z16.s, z14.s\n"
+ "sqadd z18.s, z18.s, z17.s\n"
+ ".inst 0x04be76d6 // sqrdmulh z22.s, z22.s, z30.s\n"
+ ".inst 0x448847b7 // smlalt z23.s, p4/M, z29.h, z8.h\n"
+ "smax z13.s, p4/M, z13.s, z20.s\n"
+ "smin z16.s, p4/M, z16.s, z15.s\n"
+ "sqadd z10.s, z10.s, z1.s\n"
+ "and z2.d, z22.d, z0.d\n"
"asr z2.s, z2.s, #0x1f\n"
- "and z26.d, z22.d, z31.d\n"
- "asr z0.s, z0.s, #0x1f\n"
- "and z20.d, z18.d, z31.d\n"
- "sqadd z9.s, z9.s, z19.s\n"
- ".inst 0x44829029 // srshl z9.s, p4/M, z9.s, z1.s\n"
- "asr z3.s, z3.s, #0x1f\n"
- "sqadd z11.s, z11.s, z2.s\n"
- ".inst 0x4482902b // srshl z11.s, p4/M, z11.s, z1.s\n"
- "asr z26.s, z26.s, #0x1f\n"
- "sqadd z21.s, z21.s, z0.s\n"
- ".inst 0x44829035 // srshl z21.s, p4/M, z21.s, z1.s\n"
- "asr z20.s, z20.s, #0x1f\n"
- "sqadd z10.s, z10.s, z3.s\n"
- ".inst 0x448293ea // srshl z10.s, p4/M, z10.s, z31.s\n"
- "sqadd z22.s, z22.s, z26.s\n"
- "sqadd z18.s, z18.s, z20.s\n"
- ".inst 0x448293f6 // srshl z22.s, p4/M, z22.s, z31.s\n"
- ".inst 0x448293f2 // srshl z18.s, p4/M, z18.s, z31.s\n"
- ".inst 0x453041ad // sqxtnb z13.h, z13.s\n"
- ".inst 0x45304129 // sqxtnb z9.h, z9.s\n"
- ".inst 0x4530416b // sqxtnb z11.h, z11.s\n"
- ".inst 0x453042b5 // sqxtnb z21.h, z21.s\n"
- ".inst 0x4530462d // sqxtnt z13.h, z17.s\n"
- ".inst 0x45304549 // sqxtnt z9.h, z10.s\n"
- ".inst 0x453046cb // sqxtnt z11.h, z22.s\n"
- ".inst 0x45304655 // sqxtnt z21.h, z18.s\n"
- "sqadd z13.h, z13.h, z14.h\n"
- "sqadd z9.h, z9.h, z14.h\n"
- "smax z13.h, p4/M, z13.h, z16.h\n"
- "smax z9.h, p4/M, z9.h, z16.h\n"
- "sqadd z11.h, z11.h, z14.h\n"
- "sqadd z21.h, z21.h, z14.h\n"
- "smax z11.h, p4/M, z11.h, z16.h\n"
- "smax z21.h, p4/M, z21.h, z16.h\n"
- "smin z13.h, p4/M, z13.h, z15.h\n"
- "smin z9.h, p4/M, z9.h, z15.h\n"
- "st1b { z13.h }, p0, [x16, x10]\n"
- "smin z11.h, p4/M, z11.h, z15.h\n"
- "smin z21.h, p4/M, z21.h, z15.h\n"
- "st1b { z9.h }, p0, [x15, x10]\n"
- "st1b { z11.h }, p0, [x14, x10]\n"
- "st1b { z21.h }, p0, [x13, x10]\n"
- "ld1b { z0.h }, p4/Z, [x17]\n"
- "ld1b { z1.h }, p4/Z, [x17, #1, MUL VL]\n"
- "inch x10\n"
- "ld1b { z2.h }, p4/Z, [x17, #2, MUL VL]\n"
- "ld1b { z3.h }, p4/Z, [x17, #3, MUL VL]\n"
+ ".inst 0x04bf76f7 // sqrdmulh z23.s, z23.s, z31.s\n"
+ "smax z16.s, p4/M, z16.s, z20.s\n"
+ ".inst 0x4482900b // srshl z11.s, p4/M, z11.s, z0.s\n"
+ ".inst 0x448292a9 // srshl z9.s, p4/M, z9.s, z21.s\n"
+ ".inst 0x44829012 // srshl z18.s, p4/M, z18.s, z0.s\n"
+ "trn1 z13.h, z13.h, z16.h\n"
+ "st1b { z13.h }, p0, [x14, x8]\n"
+ "add z11.s, z11.s, z14.s\n"
+ "add z9.s, z9.s, z14.s\n"
+ "add z18.s, z18.s, z14.s\n"
+ "sqadd z22.s, z22.s, z2.s\n"
+ "and z16.d, z23.d, z21.d\n"
+ "asr z16.s, z16.s, #0x1f\n"
+ "smin z11.s, p4/M, z11.s, z15.s\n"
+ "smin z9.s, p4/M, z9.s, z15.s\n"
+ "smin z18.s, p4/M, z18.s, z15.s\n"
+ ".inst 0x448292aa // srshl z10.s, p4/M, z10.s, z21.s\n"
+ ".inst 0x44829016 // srshl z22.s, p4/M, z22.s, z0.s\n"
+ "smax z11.s, p4/M, z11.s, z20.s\n"
+ "sqadd z23.s, z23.s, z16.s\n"
+ "add z10.s, z10.s, z14.s\n"
+ "add z22.s, z22.s, z14.s\n"
+ "smax z9.s, p4/M, z9.s, z20.s\n"
+ "smax z18.s, p4/M, z18.s, z20.s\n"
+ "smin z10.s, p4/M, z10.s, z15.s\n"
+ "smin z22.s, p4/M, z22.s, z15.s\n"
+ "trn1 z11.h, z11.h, z9.h\n"
+ "st1b { z11.h }, p0, [x13, x8]\n"
+ "smax z10.s, p4/M, z10.s, z20.s\n"
+ ".inst 0x448292b7 // srshl z23.s, p4/M, z23.s, z21.s\n"
+ "smax z22.s, p4/M, z22.s, z20.s\n"
+ "trn1 z18.h, z18.h, z10.h\n"
+ "st1b { z18.h }, p0, [x12, x8]\n"
+ "add z23.s, z23.s, z14.s\n"
+ "smin z23.s, p4/M, z23.s, z15.s\n"
+ "smax z23.s, p4/M, z23.s, z20.s\n"
+ "trn1 z22.h, z22.h, z23.h\n"
+ "st1b { z22.h }, p0, [x11, x8]\n"
+ "inch x8\n"
+ "ldr x19, [%x[params], %[offsetof_Params_bias]]\n"
+ "ld1w { z18.s }, p2/Z, [x19]\n"
+ "ld1w { z16.s }, p1/Z, [x19, #1, MUL VL]\n"
+ "uzp1 z13.s, z18.s, z16.s\n"
+ "addvl x19, x19, #2\n"
+ "str x19, [%x[params], %[offsetof_Params_bias]]\n"
+ "uzp2 z16.s, z18.s, z16.s\n"
+ "mov z11.d, z13.d\n"
+ "ld1b { z0.h }, p4/Z, [x6]\n"
".inst 0x454c1800 // usublb z0.h, z0.b, z12.b\n"
+ "mov z9.d, z16.d\n"
+ "ld1b { z1.h }, p4/Z, [x6, #1, MUL VL]\n"
+ "mov z18.d, z13.d\n"
+ "ld1b { z2.h }, p4/Z, [x6, #2, MUL VL]\n"
".inst 0x454c1821 // usublb z1.h, z1.b, z12.b\n"
- "ld1b { z4.h }, p4/Z, [x17, #4, MUL VL]\n"
- "ld1b { z5.h }, p4/Z, [x17, #5, MUL VL]\n"
+ "mov z10.d, z16.d\n"
+ "ld1b { z3.h }, p4/Z, [x6, #3, MUL VL]\n"
+ "mov z22.d, z13.d\n"
+ "ld1b { z4.h }, p4/Z, [x6, #4, MUL VL]\n"
".inst 0x454c1842 // usublb z2.h, z2.b, z12.b\n"
+ "mov z23.d, z16.d\n"
+ "ld1b { z5.h }, p4/Z, [x6, #5, MUL VL]\n"
".inst 0x454c1863 // usublb z3.h, z3.b, z12.b\n"
- "ld1b { z6.h }, p4/Z, [x17, #6, MUL VL]\n"
- "ld1b { z7.h }, p4/Z, [x17, #7, MUL VL]\n"
- "inch x17, ALL, MUL #8\n"
+ "ld1b { z6.h }, p4/Z, [x6, #6, MUL VL]\n"
+ "ld1b { z7.h }, p4/Z, [x6, #7, MUL VL]\n"
".inst 0x454c1884 // usublb z4.h, z4.b, z12.b\n"
- "ld1w { z18.s }, p2/Z, [x12]\n"
- "ld1w { z8.s }, p1/Z, [x12, #1, MUL VL]\n"
- "uzp1 z13.s, z18.s, z8.s\n"
- "uzp2 z17.s, z18.s, z8.s\n"
- "ld1b { z8.h }, p4/Z, [x17]\n"
- "ldp x9, x28, [x11, #0x0]\n"
- "addvl x12, x12, #2\n"
- "str x12, [%x[params], %[offsetof_Params_bias]]\n"
- "ldp x25, x24, [x11, #0x10]\n"
- "ldp x23, x22, [x11, #0x20]\n"
- "mov z9.d, z13.d\n"
- "mov z10.d, z17.d\n"
- "ldp x21, x20, [x11, #0x30]\n"
- "ld1b { z31.h }, p3/Z, [x9, x7]\n"
- "mov z11.d, z13.d\n"
- "mov z22.d, z17.d\n"
- "ld1b { z30.h }, p3/Z, [x28, x7]\n"
- "ld1b { z29.h }, p3/Z, [x25, x7]\n"
- "mov z21.d, z13.d\n"
- "mov z18.d, z17.d\n"
- "ld1b { z28.h }, p3/Z, [x24, x7]\n"
- "ld1b { z27.h }, p3/Z, [x23, x7]\n"
+ "inch x6, ALL, MUL #8\n"
+ "ld1b { z8.h }, p4/Z, [x6]\n"
+ "ldp x26, x25, [x16, #0x0]\n"
".inst 0x454c18a5 // usublb z5.h, z5.b, z12.b\n"
".inst 0x454c18c6 // usublb z6.h, z6.b, z12.b\n"
- "ld1b { z26.h }, p3/Z, [x22, x7]\n"
- "ld1b { z25.h }, p3/Z, [x21, x7]\n"
+ "ldp x24, x23, [x16, #0x10]\n"
".inst 0x454c18e7 // usublb z7.h, z7.b, z12.b\n"
".inst 0x454c1908 // usublb z8.h, z8.b, z12.b\n"
- "ld1b { z24.h }, p3/Z, [x20, x7]\n"
- ".inst 0x45571bff // usublb z31.h, z31.b, z23.b\n"
- ".inst 0x45571bde // usublb z30.h, z30.b, z23.b\n"
- ".inst 0x45571bbd // usublb z29.h, z29.b, z23.b\n"
- ".inst 0x45571b9c // usublb z28.h, z28.b, z23.b\n"
- ".inst 0x45571b7b // usublb z27.h, z27.b, z23.b\n"
- ".inst 0x45571b5a // usublb z26.h, z26.b, z23.b\n"
- ".inst 0x45571b39 // usublb z25.h, z25.b, z23.b\n"
- ".inst 0x45571b18 // usublb z24.h, z24.b, z23.b\n"
+ "ldp x22, x21, [x16, #0x20]\n"
+ "ldp x20, x19, [x16, #0x30]\n"
+ "ld1b { z31.h }, p3/Z, [x26, x7]\n"
+ ".inst 0x45531bff // usublb z31.h, z31.b, z19.b\n"
+ "ld1b { z30.h }, p3/Z, [x25, x7]\n"
+ "ld1b { z29.h }, p3/Z, [x24, x7]\n"
+ ".inst 0x45531bde // usublb z30.h, z30.b, z19.b\n"
+ "ld1b { z28.h }, p3/Z, [x23, x7]\n"
+ "ld1b { z27.h }, p3/Z, [x22, x7]\n"
+ ".inst 0x45531bbd // usublb z29.h, z29.b, z19.b\n"
+ "ld1b { z26.h }, p3/Z, [x21, x7]\n"
+ ".inst 0x45531b9c // usublb z28.h, z28.b, z19.b\n"
+ "ld1b { z25.h }, p3/Z, [x20, x7]\n"
+ "ld1b { z24.h }, p3/Z, [x19, x7]\n"
+ ".inst 0x45531b7b // usublb z27.h, z27.b, z19.b\n"
+ ".inst 0x45531b5a // usublb z26.h, z26.b, z19.b\n"
+ ".inst 0x45531b39 // usublb z25.h, z25.b, z19.b\n"
+ ".inst 0x45531b18 // usublb z24.h, z24.b, z19.b\n"
"b.any 1b\n"
:
: [offsetof_Params_bias] "I" (offsetof(Params, bias)), [offsetof_Params_inptrs] "I" (offsetof(Params, inptrs)), [offsetof_Params_n_channels] "I" (offsetof(Params, n_channels)), [offsetof_Params_outptrs] "I" (offsetof(Params, outptrs)), [offsetof_Params_requant] "I" (offsetof(Params, requant)), [offsetof_Params_requant_muls] "I" (offsetof(Params, requant_muls)), [offsetof_Params_requant_shifts] "I" (offsetof(Params, requant_shifts)), [offsetof_Params_weights] "I" (offsetof(Params, weights)), [offsetof_Requantize32_a_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_Requantize32_b_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [offsetof_Requantize32_c_offset] "I" (offsetof(arm_gemm::Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(arm_gemm::Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(arm_gemm::Requantize32, minval)), [params] "r" (&params)
- : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "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"
+ : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "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", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
);
}