aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp2430
1 files changed, 2430 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp
new file mode 100644
index 0000000000..cd5f85411c
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp
@@ -0,0 +1,2430 @@
+/*
+ * Copyright (c) 2021, 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifdef ARM_COMPUTE_ENABLE_SVE
+
+#include "arm_gemm.hpp"
+#include "../../utils.hpp"
+
+#include <cassert>
+#include <cstdint>
+
+namespace arm_gemm {
+
+void sve_hybrid_s8qs_mmla_6x4VL (
+ unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<int8_t> A_arg,
+ size_t M, size_t N, const int8_t *B_ptr, IndirectOutputArg<int8_t> output_arg,
+ const Requantize32 *qp, const int32_t *col_bias, unsigned int col_base
+)
+{
+ struct KernelArgs {
+ const int32_t *multiplier_ptr = {};
+ const int32_t *shift_ptr = {};
+ unsigned int num_strings = {};
+ const unsigned int *string_lengths = {};
+ size_t N = {};
+ const int8_t *B_ptr = {};
+ size_t output_offset = {};
+ size_t input_initial_col = {};
+ size_t input_offset = {};
+ } ka;
+
+ unsigned long flags=0;
+ void *output_ptr;
+ void *input_ptr;
+
+ if (output_arg.is_indirect) {
+ output_ptr=(void *)(output_arg.indirect.ptr);
+ ka.output_offset=output_arg.indirect.offset;
+ flags |= 0x4;
+ } else {
+ output_ptr=(void *)(output_arg.direct.base);
+ ka.output_offset=output_arg.direct.stride;
+ }
+
+ if (A_arg.is_indirect) {
+ input_ptr=(void *)(A_arg.indirect.ptr);
+ ka.input_offset=A_arg.indirect.start_row;
+ ka.input_initial_col=A_arg.indirect.start_col;
+ flags |= 0x8;
+ } else {
+ assert(num_strings==1);
+ input_ptr=(void *)(A_arg.direct.base);
+ ka.input_offset=A_arg.direct.stride;
+ }
+ ka.num_strings = num_strings;
+ ka.string_lengths = string_lengths;
+ ka.N = N;
+ ka.B_ptr = B_ptr;
+ if (qp->per_channel_requant) {
+ flags |= 0x10;
+ ka.multiplier_ptr=qp->per_channel_muls + col_base;
+ ka.shift_ptr=qp->per_channel_right_shifts + col_base;
+ }
+ if (qp->c_offset > qp->minval) {
+ flags |= 0x20;
+ }
+ __asm__ __volatile__(
+ "ptrue p2.b\n"
+ "1:" // Row loop
+ "cmp %x[M], #0x6\n"
+ "bge 66f\n"
+ "cmp %x[M], #0x4\n"
+ "bgt 53f\n"
+ "beq 40f\n"
+ "cmp %x[M], #0x2\n"
+ "bgt 27f\n"
+ "beq 14f\n"
+ "mov x14, %x[col_bias]\n"
+ "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
+ "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
+ "mov x11, %x[output_ptr]\n"
+ "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "2:" // Height 1: Column loop
+ "mov x20, #0x0\n"
+ "whilelt p1.b, x20, x10\n"
+ "mov z8.s, #0x0\n"
+ "mov z9.s, #0x0\n"
+ "mov z10.s, #0x0\n"
+ "mov z11.s, #0x0\n"
+ "mov z12.s, #0x0\n"
+ "mov z13.s, #0x0\n"
+ "mov z14.s, #0x0\n"
+ "mov z15.s, #0x0\n"
+ "3:" // Height 1: setup done
+ "mov x28, #0x0\n"
+ "4:" // Height 1: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w27, [x20, x28, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 5f\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "cbnz x28, 6f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x26, x26, x20\n"
+ "b 6f\n"
+ "5:" // Height 1: setup direct input
+ "mov x26, %x[input_ptr]\n"
+ "6:" // Height 1: input setup done
+ "cmp x27, #0x10\n"
+ "ble 8f\n"
+ "7:" // Height 1: Multiply loop: Main loop head
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z20.b }, p0/Z, [x26]\n"
+ "trn1 z18.d, z20.d, z19.d\n"
+ "ld1b { z17.b }, p2/Z, [x9]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45119a48 // smmla z8.s, z18.b, z17.b\n"
+ ".inst 0x45109a4c // smmla z12.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45119a49 // smmla z9.s, z18.b, z17.b\n"
+ ".inst 0x45109a4d // smmla z13.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x45119a4a // smmla z10.s, z18.b, z17.b\n"
+ ".inst 0x45109a4e // smmla z14.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #16\n"
+ "trn2 z20.d, z20.d, z19.d\n"
+ ".inst 0x45119a4b // smmla z11.s, z18.b, z17.b\n"
+ ".inst 0x45109a4f // smmla z15.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #-8, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #-7, MUL VL]\n"
+ ".inst 0x45119a88 // smmla z8.s, z20.b, z17.b\n"
+ ".inst 0x45109a8c // smmla z12.s, z20.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #-6, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #-5, MUL VL]\n"
+ ".inst 0x45119a89 // smmla z9.s, z20.b, z17.b\n"
+ ".inst 0x45109a8d // smmla z13.s, z20.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #-4, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #-3, MUL VL]\n"
+ ".inst 0x45119a8a // smmla z10.s, z20.b, z17.b\n"
+ ".inst 0x45109a8e // smmla z14.s, z20.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #-2, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #-1, MUL VL]\n"
+ "sub x27, x27, #0x10\n"
+ "cmp x27, #0x10\n"
+ ".inst 0x45119a8b // smmla z11.s, z20.b, z17.b\n"
+ ".inst 0x45109a8f // smmla z15.s, z20.b, z16.b\n"
+ "add x26, x26, #0x10\n"
+ "bgt 7b\n"
+ "8:" // Height 1: Multiply loop: Single iteration only
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z1.b }, p0/Z, [x26]\n"
+ "trn1 z18.d, z1.d, z19.d\n"
+ "ld1b { z17.b }, p2/Z, [x9]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45119a48 // smmla z8.s, z18.b, z17.b\n"
+ ".inst 0x45109a4c // smmla z12.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45119a49 // smmla z9.s, z18.b, z17.b\n"
+ ".inst 0x45109a4d // smmla z13.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x45119a4a // smmla z10.s, z18.b, z17.b\n"
+ ".inst 0x45109a4e // smmla z14.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "subs x27, x27, #0x8\n"
+ "trn2 z1.d, z1.d, z19.d\n"
+ ".inst 0x45119a4b // smmla z11.s, z18.b, z17.b\n"
+ ".inst 0x45109a4f // smmla z15.s, z18.b, z16.b\n"
+ "addvl x9, x9, #8\n"
+ "ble 9f\n"
+ "ld1b { z17.b }, p2/Z, [x9]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45119828 // smmla z8.s, z1.b, z17.b\n"
+ ".inst 0x4510982c // smmla z12.s, z1.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45119829 // smmla z9.s, z1.b, z17.b\n"
+ ".inst 0x4510982d // smmla z13.s, z1.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x4511982a // smmla z10.s, z1.b, z17.b\n"
+ ".inst 0x4510982e // smmla z14.s, z1.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+ ".inst 0x4511982b // smmla z11.s, z1.b, z17.b\n"
+ ".inst 0x4510982f // smmla z15.s, z1.b, z16.b\n"
+ "addvl x9, x9, #8\n"
+ "9:" // Height 1: Multiply loop: multiply skip
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x28, x28, #0x1\n"
+ "cmp x28, x20\n"
+ "bne 4b\n"
+ "uzp1 z8.d, z8.d, z12.d\n"
+ "uzp1 z9.d, z9.d, z13.d\n"
+ "ld1w { z19.s }, p2/Z, [x14]\n"
+ "ld1w { z18.s }, p2/Z, [x14, #1, MUL VL]\n"
+ "uzp1 z10.d, z10.d, z14.d\n"
+ "uzp1 z11.d, z11.d, z15.d\n"
+ "ld1w { z17.s }, p2/Z, [x14, #2, MUL VL]\n"
+ "ld1w { z16.s }, p2/Z, [x14, #3, MUL VL]\n"
+ "mov z15.d, z8.d\n"
+ "add z15.s, z15.s, z19.s\n"
+ "addvl x14, x14, #4\n"
+ "add z9.s, z9.s, z18.s\n"
+ "add z10.s, z10.s, z17.s\n"
+ "add z11.s, z11.s, z16.s\n"
+ "tbz %x[flags], #4, 10f\n"
+ "ld1w { z0.s }, p2/Z, [x12]\n"
+ "ld1w { z4.s }, p2/Z, [x13]\n"
+ "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n"
+ "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n"
+ "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n"
+ "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n"
+ "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n"
+ "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n"
+ "addvl x12, x12, #4\n"
+ "addvl x13, x13, #4\n"
+ "b 11f\n"
+ "10:" // Height 1: per layer parameters
+ "add x21, %x[qp], %[per_layer_right_shift]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1rw { z0.s }, p2/Z, [x21]\n"
+ "ld1rw { z4.s }, p2/Z, [x20]\n"
+ "mov z1.d, z0.d\n"
+ "mov z5.d, z4.d\n"
+ "mov z2.d, z0.d\n"
+ "mov z6.d, z4.d\n"
+ "mov z3.d, z0.d\n"
+ "mov z7.d, z4.d\n"
+ "11:" // Height 1: parameters loaded
+ ".inst 0x04a475ef // sqrdmulh z15.s, z15.s, z4.s\n"
+ ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n"
+ ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n"
+ ".inst 0x04a7756b // sqrdmulh z11.s, z11.s, z7.s\n"
+ "tbz %x[flags], #5, 12f\n"
+ "and z19.d, z15.d, z0.d\n"
+ "and z18.d, z9.d, z1.d\n"
+ "and z17.d, z10.d, z2.d\n"
+ "and z16.d, z11.d, z3.d\n"
+ "asr z19.s, z19.s, #0x1f\n"
+ "asr z18.s, z18.s, #0x1f\n"
+ "asr z17.s, z17.s, #0x1f\n"
+ "asr z16.s, z16.s, #0x1f\n"
+ "sqadd z15.s, z15.s, z19.s\n"
+ "sqadd z9.s, z9.s, z18.s\n"
+ "sqadd z10.s, z10.s, z17.s\n"
+ "sqadd z11.s, z11.s, z16.s\n"
+ "12:" // Height 1: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1rw { z17.s }, p2/Z, [x20]\n"
+ ".inst 0x4482880f // srshl z15.s, p2/M, z15.s, z0.s\n"
+ "add z15.s, z15.s, z17.s\n"
+ ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n"
+ ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n"
+ "add z9.s, z9.s, z17.s\n"
+ "add z10.s, z10.s, z17.s\n"
+ ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1rw { z16.s }, p2/Z, [x20]\n"
+ "add z11.s, z11.s, z17.s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1rw { z31.s }, p2/Z, [x20]\n"
+ "smin z15.s, p2/M, z15.s, z16.s\n"
+ "smin z9.s, p2/M, z9.s, z16.s\n"
+ "smin z10.s, p2/M, z10.s, z16.s\n"
+ "smin z11.s, p2/M, z11.s, z16.s\n"
+ "smax z15.s, p2/M, z15.s, z31.s\n"
+ "smax z9.s, p2/M, z9.s, z31.s\n"
+ "smax z10.s, p2/M, z10.s, z31.s\n"
+ "uzp1 z15.h, z15.h, z9.h\n"
+ "smax z11.s, p2/M, z11.s, z31.s\n"
+ "uzp1 z16.h, z10.h, z11.h\n"
+ "uzp1 z15.b, z15.b, z16.b\n"
+ "st1b { z15.b }, p1, [x11]\n"
+ "addvl x11, x11, #1\n"
+ "13:" // Height 1: Writeback done
+ "decw x10, ALL, MUL #4\n"
+ "cmp x10, XZR\n"
+ "bgt 2b\n"
+ "b 80f\n"
+ "14:" // Height 2
+ "mov x14, %x[col_bias]\n"
+ "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
+ "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
+ "mov x11, %x[output_ptr]\n"
+ "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "15:" // Height 2: Column loop
+ "mov x20, #0x0\n"
+ "whilelt p1.b, x20, x10\n"
+ "mov z8.s, #0x0\n"
+ "mov z9.s, #0x0\n"
+ "mov z10.s, #0x0\n"
+ "mov z11.s, #0x0\n"
+ "mov z12.s, #0x0\n"
+ "mov z13.s, #0x0\n"
+ "mov z14.s, #0x0\n"
+ "mov z15.s, #0x0\n"
+ "16:" // Height 2: setup done
+ "mov x28, #0x0\n"
+ "17:" // Height 2: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w27, [x20, x28, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 18f\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "cbnz x28, 19f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x26, x26, x20\n"
+ "add x25, x25, x20\n"
+ "b 19f\n"
+ "18:" // Height 2: setup direct input
+ "mov x26, %x[input_ptr]\n"
+ "add x25, x26, x21\n"
+ "19:" // Height 2: input setup done
+ "cmp x27, #0x10\n"
+ "ble 21f\n"
+ "20:" // Height 2: Multiply loop: Main loop head
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z20.b }, p0/Z, [x26]\n"
+ "ld1rqb { z19.b }, p0/Z, [x25]\n"
+ "trn1 z18.d, z20.d, z19.d\n"
+ "ld1b { z17.b }, p2/Z, [x9]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45119a48 // smmla z8.s, z18.b, z17.b\n"
+ ".inst 0x45109a4c // smmla z12.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45119a49 // smmla z9.s, z18.b, z17.b\n"
+ ".inst 0x45109a4d // smmla z13.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x45119a4a // smmla z10.s, z18.b, z17.b\n"
+ ".inst 0x45109a4e // smmla z14.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #16\n"
+ "trn2 z20.d, z20.d, z19.d\n"
+ ".inst 0x45119a4b // smmla z11.s, z18.b, z17.b\n"
+ ".inst 0x45109a4f // smmla z15.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #-8, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #-7, MUL VL]\n"
+ ".inst 0x45119a88 // smmla z8.s, z20.b, z17.b\n"
+ ".inst 0x45109a8c // smmla z12.s, z20.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #-6, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #-5, MUL VL]\n"
+ ".inst 0x45119a89 // smmla z9.s, z20.b, z17.b\n"
+ ".inst 0x45109a8d // smmla z13.s, z20.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #-4, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #-3, MUL VL]\n"
+ ".inst 0x45119a8a // smmla z10.s, z20.b, z17.b\n"
+ ".inst 0x45109a8e // smmla z14.s, z20.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #-2, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #-1, MUL VL]\n"
+ "sub x27, x27, #0x10\n"
+ "cmp x27, #0x10\n"
+ ".inst 0x45119a8b // smmla z11.s, z20.b, z17.b\n"
+ ".inst 0x45109a8f // smmla z15.s, z20.b, z16.b\n"
+ "add x26, x26, #0x10\n"
+ "add x25, x25, #0x10\n"
+ "bgt 20b\n"
+ "21:" // Height 2: Multiply loop: Single iteration only
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z1.b }, p0/Z, [x26]\n"
+ "ld1rqb { z19.b }, p0/Z, [x25]\n"
+ "trn1 z18.d, z1.d, z19.d\n"
+ "ld1b { z17.b }, p2/Z, [x9]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45119a48 // smmla z8.s, z18.b, z17.b\n"
+ ".inst 0x45109a4c // smmla z12.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45119a49 // smmla z9.s, z18.b, z17.b\n"
+ ".inst 0x45109a4d // smmla z13.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x45119a4a // smmla z10.s, z18.b, z17.b\n"
+ ".inst 0x45109a4e // smmla z14.s, z18.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "subs x27, x27, #0x8\n"
+ "trn2 z1.d, z1.d, z19.d\n"
+ ".inst 0x45119a4b // smmla z11.s, z18.b, z17.b\n"
+ ".inst 0x45109a4f // smmla z15.s, z18.b, z16.b\n"
+ "addvl x9, x9, #8\n"
+ "ble 22f\n"
+ "ld1b { z17.b }, p2/Z, [x9]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45119828 // smmla z8.s, z1.b, z17.b\n"
+ ".inst 0x4510982c // smmla z12.s, z1.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45119829 // smmla z9.s, z1.b, z17.b\n"
+ ".inst 0x4510982d // smmla z13.s, z1.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x4511982a // smmla z10.s, z1.b, z17.b\n"
+ ".inst 0x4510982e // smmla z14.s, z1.b, z16.b\n"
+ "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+ "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+ ".inst 0x4511982b // smmla z11.s, z1.b, z17.b\n"
+ ".inst 0x4510982f // smmla z15.s, z1.b, z16.b\n"
+ "addvl x9, x9, #8\n"
+ "22:" // Height 2: Multiply loop: multiply skip
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x28, x28, #0x1\n"
+ "cmp x28, x20\n"
+ "bne 17b\n"
+ "uzp1 z20.d, z8.d, z12.d\n"
+ "uzp2 z8.d, z8.d, z12.d\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "ld1w { z19.s }, p2/Z, [x14]\n"
+ "uzp1 z12.d, z9.d, z13.d\n"
+ "uzp2 z9.d, z9.d, z13.d\n"
+ "ld1w { z18.s }, p2/Z, [x14, #1, MUL VL]\n"
+ "ld1w { z17.s }, p2/Z, [x14, #2, MUL VL]\n"
+ "uzp1 z13.d, z10.d, z14.d\n"
+ "uzp2 z10.d, z10.d, z14.d\n"
+ "ld1w { z16.s }, p2/Z, [x14, #3, MUL VL]\n"
+ "add x26, x11, x20\n"
+ "uzp1 z14.d, z11.d, z15.d\n"
+ "uzp2 z11.d, z11.d, z15.d\n"
+ "addvl x14, x14, #4\n"
+ "mov z15.d, z20.d\n"
+ "add z15.s, z15.s, z19.s\n"
+ "add z12.s, z12.s, z18.s\n"
+ "add z13.s, z13.s, z17.s\n"
+ "add z14.s, z14.s, z16.s\n"
+ "add z8.s, z8.s, z19.s\n"
+ "add z9.s, z9.s, z18.s\n"
+ "add z10.s, z10.s, z17.s\n"
+ "add z11.s, z11.s, z16.s\n"
+ "tbz %x[flags], #4, 23f\n"
+ "ld1w { z0.s }, p2/Z, [x12]\n"
+ "ld1w { z4.s }, p2/Z, [x13]\n"
+ "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n"
+ "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n"
+ "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n"
+ "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n"
+ "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n"
+ "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n"
+ "addvl x12, x12, #4\n"
+ "addvl x13, x13, #4\n"
+ "b 24f\n"
+ "23:" // Height 2: per layer parameters
+ "add x21, %x[qp], %[per_layer_right_shift]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1rw { z0.s }, p2/Z, [x21]\n"
+ "ld1rw { z4.s }, p2/Z, [x20]\n"
+ "mov z1.d, z0.d\n"
+ "mov z5.d, z4.d\n"
+ "mov z2.d, z0.d\n"
+ "mov z6.d, z4.d\n"
+ "mov z3.d, z0.d\n"
+ "mov z7.d, z4.d\n"
+ "24:" // Height 2: parameters loaded
+ ".inst 0x04a475ef // sqrdmulh z15.s, z15.s, z4.s\n"
+ ".inst 0x04a5758c // sqrdmulh z12.s, z12.s, z5.s\n"
+ ".inst 0x04a675ad // sqrdmulh z13.s, z13.s, z6.s\n"
+ ".inst 0x04a775ce // sqrdmulh z14.s, z14.s, z7.s\n"
+ ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n"
+ ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n"
+ ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n"
+ ".inst 0x04a7756b // sqrdmulh z11.s, z11.s, z7.s\n"
+ "tbz %x[flags], #5, 25f\n"
+ "and z19.d, z15.d, z0.d\n"
+ "and z18.d, z12.d, z1.d\n"
+ "and z17.d, z13.d, z2.d\n"
+ "and z16.d, z14.d, z3.d\n"
+ "asr z19.s, z19.s, #0x1f\n"
+ "asr z18.s, z18.s, #0x1f\n"
+ "asr z17.s, z17.s, #0x1f\n"
+ "asr z16.s, z16.s, #0x1f\n"
+ "sqadd z15.s, z15.s, z19.s\n"
+ "sqadd z12.s, z12.s, z18.s\n"
+ "sqadd z13.s, z13.s, z17.s\n"
+ "sqadd z14.s, z14.s, z16.s\n"
+ "and z18.d, z8.d, z0.d\n"
+ "and z24.d, z9.d, z1.d\n"
+ "and z17.d, z10.d, z2.d\n"
+ "and z16.d, z11.d, z3.d\n"
+ "asr z18.s, z18.s, #0x1f\n"
+ "asr z24.s, z24.s, #0x1f\n"
+ "asr z17.s, z17.s, #0x1f\n"
+ "asr z16.s, z16.s, #0x1f\n"
+ "sqadd z8.s, z8.s, z18.s\n"
+ "sqadd z9.s, z9.s, z24.s\n"
+ "sqadd z10.s, z10.s, z17.s\n"
+ "sqadd z11.s, z11.s, z16.s\n"
+ "25:" // Height 2: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1rw { z17.s }, p2/Z, [x20]\n"
+ ".inst 0x4482880f // srshl z15.s, p2/M, z15.s, z0.s\n"
+ "add z15.s, z15.s, z17.s\n"
+ ".inst 0x4482882c // srshl z12.s, p2/M, z12.s, z1.s\n"
+ ".inst 0x4482884d // srshl z13.s, p2/M, z13.s, z2.s\n"
+ "add z12.s, z12.s, z17.s\n"
+ "add z13.s, z13.s, z17.s\n"
+ ".inst 0x4482886e // srshl z14.s, p2/M, z14.s, z3.s\n"
+ ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n"
+ "add z14.s, z14.s, z17.s\n"
+ "add z8.s, z8.s, z17.s\n"
+ ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n"
+ ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n"
+ "add z9.s, z9.s, z17.s\n"
+ "add z10.s, z10.s, z17.s\n"
+ ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1rw { z16.s }, p2/Z, [x20]\n"
+ "add z11.s, z11.s, z17.s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1rw { z17.s }, p2/Z, [x20]\n"
+ "smin z15.s, p2/M, z15.s, z16.s\n"
+ "smin z12.s, p2/M, z12.s, z16.s\n"
+ "smin z13.s, p2/M, z13.s, z16.s\n"
+ "smin z14.s, p2/M, z14.s, z16.s\n"
+ "smin z8.s, p2/M, z8.s, z16.s\n"
+ "smin z9.s, p2/M, z9.s, z16.s\n"
+ "smin z10.s, p2/M, z10.s, z16.s\n"
+ "smin z11.s, p2/M, z11.s, z16.s\n"
+ "smax z15.s, p2/M, z15.s, z17.s\n"
+ "smax z12.s, p2/M, z12.s, z17.s\n"
+ "smax z13.s, p2/M, z13.s, z17.s\n"
+ "uzp1 z15.h, z15.h, z12.h\n"
+ "smax z14.s, p2/M, z14.s, z17.s\n"
+ "smax z8.s, p2/M, z8.s, z17.s\n"
+ "uzp1 z16.h, z13.h, z14.h\n"
+ "uzp1 z15.b, z15.b, z16.b\n"
+ "smax z9.s, p2/M, z9.s, z17.s\n"
+ "smax z10.s, p2/M, z10.s, z17.s\n"
+ "uzp1 z8.h, z8.h, z9.h\n"
+ "st1b { z15.b }, p1, [x11]\n"
+ "smax z11.s, p2/M, z11.s, z17.s\n"
+ "uzp1 z16.h, z10.h, z11.h\n"
+ "uzp1 z8.b, z8.b, z16.b\n"
+ "st1b { z8.b }, p1, [x26]\n"
+ "addvl x11, x11, #1\n"
+ "26:" // Height 2: Writeback done
+ "decw x10, ALL, MUL #4\n"
+ "cmp x10, XZR\n"
+ "bgt 15b\n"
+ "b 80f\n"
+ "27:" // Height 3
+ "mov x14, %x[col_bias]\n"
+ "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
+ "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
+ "mov x11, %x[output_ptr]\n"
+ "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "28:" // Height 3: Column loop
+ "mov x20, #0x0\n"
+ "whilelt p1.b, x20, x10\n"
+ "mov z8.s, #0x0\n"
+ "mov z9.s, #0x0\n"
+ "mov z10.s, #0x0\n"
+ "mov z11.s, #0x0\n"
+ "mov z12.s, #0x0\n"
+ "mov z13.s, #0x0\n"
+ "mov z14.s, #0x0\n"
+ "mov z15.s, #0x0\n"
+ "mov z16.s, #0x0\n"
+ "mov z17.s, #0x0\n"
+ "mov z18.s, #0x0\n"
+ "mov z19.s, #0x0\n"
+ "mov z20.s, #0x0\n"
+ "mov z21.s, #0x0\n"
+ "mov z22.s, #0x0\n"
+ "mov z23.s, #0x0\n"
+ "29:" // Height 3: setup done
+ "mov x28, #0x0\n"
+ "30:" // Height 3: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w27, [x20, x28, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 31f\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "cbnz x28, 32f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x26, x26, x20\n"
+ "add x25, x25, x20\n"
+ "add x24, x24, x20\n"
+ "b 32f\n"
+ "31:" // Height 3: setup direct input
+ "mov x26, %x[input_ptr]\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "32:" // Height 3: input setup done
+ "cmp x27, #0x10\n"
+ "ble 34f\n"
+ "33:" // Height 3: Multiply loop: Main loop head
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z30.b }, p0/Z, [x26]\n"
+ "ld1rqb { z24.b }, p0/Z, [x25]\n"
+ "ld1rqb { z28.b }, p0/Z, [x24]\n"
+ "trn1 z27.d, z30.d, z24.d\n"
+ "trn2 z30.d, z30.d, z24.d\n"
+ "ld1b { z25.b }, p2/Z, [x9]\n"
+ "trn1 z26.d, z28.d, z29.d\n"
+ "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45199b68 // smmla z8.s, z27.b, z25.b\n"
+ ".inst 0x45199b50 // smmla z16.s, z26.b, z25.b\n"
+ ".inst 0x45189b6c // smmla z12.s, z27.b, z24.b\n"
+ ".inst 0x45189b54 // smmla z20.s, z26.b, z24.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45199b69 // smmla z9.s, z27.b, z25.b\n"
+ ".inst 0x45199b51 // smmla z17.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "trn2 z28.d, z28.d, z29.d\n"
+ ".inst 0x45189b6d // smmla z13.s, z27.b, z24.b\n"
+ ".inst 0x45189b55 // smmla z21.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+ "sub x27, x27, #0x10\n"
+ ".inst 0x45199b6a // smmla z10.s, z27.b, z25.b\n"
+ ".inst 0x45199b52 // smmla z18.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
+ "cmp x27, #0x10\n"
+ ".inst 0x45189b6e // smmla z14.s, z27.b, z24.b\n"
+ ".inst 0x45189b56 // smmla z22.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #16\n"
+ ".inst 0x45199b6b // smmla z11.s, z27.b, z25.b\n"
+ ".inst 0x45199b53 // smmla z19.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #-8, MUL VL]\n"
+ "add x26, x26, #0x10\n"
+ ".inst 0x45189b6f // smmla z15.s, z27.b, z24.b\n"
+ ".inst 0x45189b57 // smmla z23.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #-7, MUL VL]\n"
+ "add x25, x25, #0x10\n"
+ ".inst 0x45199bc8 // smmla z8.s, z30.b, z25.b\n"
+ ".inst 0x45199b90 // smmla z16.s, z28.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #-6, MUL VL]\n"
+ "add x24, x24, #0x10\n"
+ ".inst 0x45189bcc // smmla z12.s, z30.b, z24.b\n"
+ ".inst 0x45189b94 // smmla z20.s, z28.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #-5, MUL VL]\n"
+ ".inst 0x45199bc9 // smmla z9.s, z30.b, z25.b\n"
+ ".inst 0x45199b91 // smmla z17.s, z28.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #-4, MUL VL]\n"
+ ".inst 0x45189bcd // smmla z13.s, z30.b, z24.b\n"
+ ".inst 0x45189b95 // smmla z21.s, z28.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #-3, MUL VL]\n"
+ ".inst 0x45199bca // smmla z10.s, z30.b, z25.b\n"
+ ".inst 0x45199b92 // smmla z18.s, z28.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #-2, MUL VL]\n"
+ ".inst 0x45189bce // smmla z14.s, z30.b, z24.b\n"
+ ".inst 0x45189b96 // smmla z22.s, z28.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #-1, MUL VL]\n"
+ ".inst 0x45199bcb // smmla z11.s, z30.b, z25.b\n"
+ ".inst 0x45199b93 // smmla z19.s, z28.b, z25.b\n"
+ ".inst 0x45189bcf // smmla z15.s, z30.b, z24.b\n"
+ ".inst 0x45189b97 // smmla z23.s, z28.b, z24.b\n"
+ "bgt 33b\n"
+ "34:" // Height 3: Multiply loop: Single iteration only
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z1.b }, p0/Z, [x26]\n"
+ "ld1rqb { z24.b }, p0/Z, [x25]\n"
+ "ld1rqb { z3.b }, p0/Z, [x24]\n"
+ "trn1 z27.d, z1.d, z24.d\n"
+ "trn2 z1.d, z1.d, z24.d\n"
+ "ld1b { z25.b }, p2/Z, [x9]\n"
+ "trn1 z26.d, z3.d, z28.d\n"
+ "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45199b68 // smmla z8.s, z27.b, z25.b\n"
+ ".inst 0x45199b50 // smmla z16.s, z26.b, z25.b\n"
+ ".inst 0x45189b6c // smmla z12.s, z27.b, z24.b\n"
+ ".inst 0x45189b54 // smmla z20.s, z26.b, z24.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45199b69 // smmla z9.s, z27.b, z25.b\n"
+ ".inst 0x45199b51 // smmla z17.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "subs x27, x27, #0x8\n"
+ ".inst 0x45189b6d // smmla z13.s, z27.b, z24.b\n"
+ ".inst 0x45189b55 // smmla z21.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+ "trn2 z3.d, z3.d, z28.d\n"
+ ".inst 0x45199b6a // smmla z10.s, z27.b, z25.b\n"
+ ".inst 0x45199b52 // smmla z18.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x45189b6e // smmla z14.s, z27.b, z24.b\n"
+ ".inst 0x45189b56 // smmla z22.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #8\n"
+ ".inst 0x45199b6b // smmla z11.s, z27.b, z25.b\n"
+ ".inst 0x45199b53 // smmla z19.s, z26.b, z25.b\n"
+ ".inst 0x45189b6f // smmla z15.s, z27.b, z24.b\n"
+ ".inst 0x45189b57 // smmla z23.s, z26.b, z24.b\n"
+ "ble 35f\n"
+ "ld1b { z25.b }, p2/Z, [x9]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45199828 // smmla z8.s, z1.b, z25.b\n"
+ ".inst 0x45199870 // smmla z16.s, z3.b, z25.b\n"
+ ".inst 0x4518982c // smmla z12.s, z1.b, z24.b\n"
+ ".inst 0x45189874 // smmla z20.s, z3.b, z24.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45199829 // smmla z9.s, z1.b, z25.b\n"
+ ".inst 0x45199871 // smmla z17.s, z3.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+ ".inst 0x4518982d // smmla z13.s, z1.b, z24.b\n"
+ ".inst 0x45189875 // smmla z21.s, z3.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x4519982a // smmla z10.s, z1.b, z25.b\n"
+ ".inst 0x45199872 // smmla z18.s, z3.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x4518982e // smmla z14.s, z1.b, z24.b\n"
+ ".inst 0x45189876 // smmla z22.s, z3.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #8\n"
+ ".inst 0x4519982b // smmla z11.s, z1.b, z25.b\n"
+ ".inst 0x45199873 // smmla z19.s, z3.b, z25.b\n"
+ ".inst 0x4518982f // smmla z15.s, z1.b, z24.b\n"
+ ".inst 0x45189877 // smmla z23.s, z3.b, z24.b\n"
+ "35:" // Height 3: Multiply loop: multiply skip
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x28, x28, #0x1\n"
+ "cmp x28, x20\n"
+ "bne 30b\n"
+ "uzp1 z28.d, z8.d, z12.d\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "uzp2 z8.d, z8.d, z12.d\n"
+ "ld1w { z27.s }, p2/Z, [x14]\n"
+ "uzp1 z12.d, z9.d, z13.d\n"
+ "uzp2 z9.d, z9.d, z13.d\n"
+ "ld1w { z26.s }, p2/Z, [x14, #1, MUL VL]\n"
+ "ld1w { z25.s }, p2/Z, [x14, #2, MUL VL]\n"
+ "uzp1 z13.d, z10.d, z14.d\n"
+ "uzp2 z10.d, z10.d, z14.d\n"
+ "ld1w { z24.s }, p2/Z, [x14, #3, MUL VL]\n"
+ "add x26, x11, x20\n"
+ "uzp1 z14.d, z11.d, z15.d\n"
+ "uzp2 z11.d, z11.d, z15.d\n"
+ "add x25, x26, x20\n"
+ "addvl x14, x14, #4\n"
+ "uzp1 z16.d, z16.d, z20.d\n"
+ "uzp1 z17.d, z17.d, z21.d\n"
+ "uzp1 z18.d, z18.d, z22.d\n"
+ "uzp1 z19.d, z19.d, z23.d\n"
+ "mov z23.d, z28.d\n"
+ "add z23.s, z23.s, z27.s\n"
+ "add z12.s, z12.s, z26.s\n"
+ "add z13.s, z13.s, z25.s\n"
+ "add z14.s, z14.s, z24.s\n"
+ "add z8.s, z8.s, z27.s\n"
+ "add z9.s, z9.s, z26.s\n"
+ "add z10.s, z10.s, z25.s\n"
+ "add z11.s, z11.s, z24.s\n"
+ "add z16.s, z16.s, z27.s\n"
+ "add z17.s, z17.s, z26.s\n"
+ "add z18.s, z18.s, z25.s\n"
+ "add z19.s, z19.s, z24.s\n"
+ "tbz %x[flags], #4, 36f\n"
+ "ld1w { z0.s }, p2/Z, [x12]\n"
+ "ld1w { z4.s }, p2/Z, [x13]\n"
+ "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n"
+ "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n"
+ "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n"
+ "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n"
+ "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n"
+ "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n"
+ "addvl x12, x12, #4\n"
+ "addvl x13, x13, #4\n"
+ "b 37f\n"
+ "36:" // Height 3: per layer parameters
+ "add x21, %x[qp], %[per_layer_right_shift]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1rw { z0.s }, p2/Z, [x21]\n"
+ "ld1rw { z4.s }, p2/Z, [x20]\n"
+ "mov z1.d, z0.d\n"
+ "mov z5.d, z4.d\n"
+ "mov z2.d, z0.d\n"
+ "mov z6.d, z4.d\n"
+ "mov z3.d, z0.d\n"
+ "mov z7.d, z4.d\n"
+ "37:" // Height 3: parameters loaded
+ ".inst 0x04a476f7 // sqrdmulh z23.s, z23.s, z4.s\n"
+ ".inst 0x04a5758c // sqrdmulh z12.s, z12.s, z5.s\n"
+ ".inst 0x04a675ad // sqrdmulh z13.s, z13.s, z6.s\n"
+ ".inst 0x04a775ce // sqrdmulh z14.s, z14.s, z7.s\n"
+ ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n"
+ ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n"
+ ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n"
+ ".inst 0x04a7756b // sqrdmulh z11.s, z11.s, z7.s\n"
+ ".inst 0x04a47610 // sqrdmulh z16.s, z16.s, z4.s\n"
+ ".inst 0x04a57631 // sqrdmulh z17.s, z17.s, z5.s\n"
+ ".inst 0x04a67652 // sqrdmulh z18.s, z18.s, z6.s\n"
+ ".inst 0x04a77673 // sqrdmulh z19.s, z19.s, z7.s\n"
+ "tbz %x[flags], #5, 38f\n"
+ "and z24.d, z23.d, z0.d\n"
+ "and z22.d, z12.d, z1.d\n"
+ "and z21.d, z13.d, z2.d\n"
+ "and z20.d, z14.d, z3.d\n"
+ "asr z24.s, z24.s, #0x1f\n"
+ "asr z22.s, z22.s, #0x1f\n"
+ "asr z21.s, z21.s, #0x1f\n"
+ "asr z20.s, z20.s, #0x1f\n"
+ "sqadd z23.s, z23.s, z24.s\n"
+ "sqadd z12.s, z12.s, z22.s\n"
+ "sqadd z13.s, z13.s, z21.s\n"
+ "sqadd z14.s, z14.s, z20.s\n"
+ "and z24.d, z8.d, z0.d\n"
+ "and z22.d, z9.d, z1.d\n"
+ "and z21.d, z10.d, z2.d\n"
+ "and z20.d, z11.d, z3.d\n"
+ "asr z24.s, z24.s, #0x1f\n"
+ "asr z22.s, z22.s, #0x1f\n"
+ "asr z21.s, z21.s, #0x1f\n"
+ "asr z20.s, z20.s, #0x1f\n"
+ "sqadd z8.s, z8.s, z24.s\n"
+ "sqadd z9.s, z9.s, z22.s\n"
+ "sqadd z10.s, z10.s, z21.s\n"
+ "sqadd z11.s, z11.s, z20.s\n"
+ "and z24.d, z16.d, z0.d\n"
+ "and z22.d, z17.d, z1.d\n"
+ "and z21.d, z18.d, z2.d\n"
+ "and z20.d, z19.d, z3.d\n"
+ "asr z24.s, z24.s, #0x1f\n"
+ "asr z22.s, z22.s, #0x1f\n"
+ "asr z21.s, z21.s, #0x1f\n"
+ "asr z20.s, z20.s, #0x1f\n"
+ "sqadd z16.s, z16.s, z24.s\n"
+ "sqadd z17.s, z17.s, z22.s\n"
+ "sqadd z18.s, z18.s, z21.s\n"
+ "sqadd z19.s, z19.s, z20.s\n"
+ "38:" // Height 3: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1rw { z21.s }, p2/Z, [x20]\n"
+ ".inst 0x44828817 // srshl z23.s, p2/M, z23.s, z0.s\n"
+ "add z23.s, z23.s, z21.s\n"
+ ".inst 0x4482882c // srshl z12.s, p2/M, z12.s, z1.s\n"
+ ".inst 0x4482884d // srshl z13.s, p2/M, z13.s, z2.s\n"
+ "add z12.s, z12.s, z21.s\n"
+ "add z13.s, z13.s, z21.s\n"
+ ".inst 0x4482886e // srshl z14.s, p2/M, z14.s, z3.s\n"
+ ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n"
+ "add z14.s, z14.s, z21.s\n"
+ "add z8.s, z8.s, z21.s\n"
+ ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n"
+ ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n"
+ "add z9.s, z9.s, z21.s\n"
+ "add z10.s, z10.s, z21.s\n"
+ ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n"
+ ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n"
+ "add z11.s, z11.s, z21.s\n"
+ "add z16.s, z16.s, z21.s\n"
+ ".inst 0x44828831 // srshl z17.s, p2/M, z17.s, z1.s\n"
+ ".inst 0x44828852 // srshl z18.s, p2/M, z18.s, z2.s\n"
+ "add z17.s, z17.s, z21.s\n"
+ "add z18.s, z18.s, z21.s\n"
+ ".inst 0x44828873 // srshl z19.s, p2/M, z19.s, z3.s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1rw { z20.s }, p2/Z, [x20]\n"
+ "add z19.s, z19.s, z21.s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1rw { z21.s }, p2/Z, [x20]\n"
+ "smin z23.s, p2/M, z23.s, z20.s\n"
+ "smin z12.s, p2/M, z12.s, z20.s\n"
+ "smin z13.s, p2/M, z13.s, z20.s\n"
+ "smin z14.s, p2/M, z14.s, z20.s\n"
+ "smin z8.s, p2/M, z8.s, z20.s\n"
+ "smin z9.s, p2/M, z9.s, z20.s\n"
+ "smin z10.s, p2/M, z10.s, z20.s\n"
+ "smin z11.s, p2/M, z11.s, z20.s\n"
+ "smin z16.s, p2/M, z16.s, z20.s\n"
+ "smin z17.s, p2/M, z17.s, z20.s\n"
+ "smin z18.s, p2/M, z18.s, z20.s\n"
+ "smin z19.s, p2/M, z19.s, z20.s\n"
+ "smax z23.s, p2/M, z23.s, z21.s\n"
+ "smax z12.s, p2/M, z12.s, z21.s\n"
+ "smax z13.s, p2/M, z13.s, z21.s\n"
+ "uzp1 z23.h, z23.h, z12.h\n"
+ "smax z14.s, p2/M, z14.s, z21.s\n"
+ "smax z8.s, p2/M, z8.s, z21.s\n"
+ "uzp1 z20.h, z13.h, z14.h\n"
+ "uzp1 z23.b, z23.b, z20.b\n"
+ "smax z9.s, p2/M, z9.s, z21.s\n"
+ "smax z10.s, p2/M, z10.s, z21.s\n"
+ "uzp1 z8.h, z8.h, z9.h\n"
+ "st1b { z23.b }, p1, [x11]\n"
+ "smax z11.s, p2/M, z11.s, z21.s\n"
+ "smax z16.s, p2/M, z16.s, z21.s\n"
+ "uzp1 z20.h, z10.h, z11.h\n"
+ "uzp1 z8.b, z8.b, z20.b\n"
+ "smax z17.s, p2/M, z17.s, z21.s\n"
+ "smax z18.s, p2/M, z18.s, z21.s\n"
+ "uzp1 z16.h, z16.h, z17.h\n"
+ "st1b { z8.b }, p1, [x26]\n"
+ "smax z19.s, p2/M, z19.s, z21.s\n"
+ "uzp1 z17.h, z18.h, z19.h\n"
+ "uzp1 z16.b, z16.b, z17.b\n"
+ "st1b { z16.b }, p1, [x25]\n"
+ "addvl x11, x11, #1\n"
+ "39:" // Height 3: Writeback done
+ "decw x10, ALL, MUL #4\n"
+ "cmp x10, XZR\n"
+ "bgt 28b\n"
+ "b 80f\n"
+ "40:" // Height 4
+ "mov x14, %x[col_bias]\n"
+ "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
+ "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
+ "mov x11, %x[output_ptr]\n"
+ "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "41:" // Height 4: Column loop
+ "mov x20, #0x0\n"
+ "whilelt p1.b, x20, x10\n"
+ "mov z8.s, #0x0\n"
+ "mov z9.s, #0x0\n"
+ "mov z10.s, #0x0\n"
+ "mov z11.s, #0x0\n"
+ "mov z12.s, #0x0\n"
+ "mov z13.s, #0x0\n"
+ "mov z14.s, #0x0\n"
+ "mov z15.s, #0x0\n"
+ "mov z16.s, #0x0\n"
+ "mov z17.s, #0x0\n"
+ "mov z18.s, #0x0\n"
+ "mov z19.s, #0x0\n"
+ "mov z20.s, #0x0\n"
+ "mov z21.s, #0x0\n"
+ "mov z22.s, #0x0\n"
+ "mov z23.s, #0x0\n"
+ "42:" // Height 4: setup done
+ "mov x28, #0x0\n"
+ "43:" // Height 4: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w27, [x20, x28, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 44f\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
+ "cbnz x28, 45f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x26, x26, x20\n"
+ "add x25, x25, x20\n"
+ "add x24, x24, x20\n"
+ "add x23, x23, x20\n"
+ "b 45f\n"
+ "44:" // Height 4: setup direct input
+ "mov x26, %x[input_ptr]\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "add x23, x24, x21\n"
+ "45:" // Height 4: input setup done
+ "cmp x27, #0x10\n"
+ "ble 47f\n"
+ "46:" // Height 4: Multiply loop: Main loop head
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z30.b }, p0/Z, [x26]\n"
+ "ld1rqb { z24.b }, p0/Z, [x25]\n"
+ "trn1 z29.d, z30.d, z24.d\n"
+ "ld1rqb { z28.b }, p0/Z, [x24]\n"
+ "ld1rqb { z27.b }, p0/Z, [x23]\n"
+ "trn2 z30.d, z30.d, z24.d\n"
+ "trn1 z26.d, z28.d, z27.d\n"
+ "ld1b { z25.b }, p2/Z, [x9]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45199ba8 // smmla z8.s, z29.b, z25.b\n"
+ ".inst 0x45199b50 // smmla z16.s, z26.b, z25.b\n"
+ ".inst 0x45189bac // smmla z12.s, z29.b, z24.b\n"
+ ".inst 0x45189b54 // smmla z20.s, z26.b, z24.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45199ba9 // smmla z9.s, z29.b, z25.b\n"
+ ".inst 0x45199b51 // smmla z17.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "trn2 z28.d, z28.d, z27.d\n"
+ ".inst 0x45189bad // smmla z13.s, z29.b, z24.b\n"
+ ".inst 0x45189b55 // smmla z21.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+ "sub x27, x27, #0x10\n"
+ ".inst 0x45199baa // smmla z10.s, z29.b, z25.b\n"
+ ".inst 0x45199b52 // smmla z18.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
+ "cmp x27, #0x10\n"
+ ".inst 0x45189bae // smmla z14.s, z29.b, z24.b\n"
+ ".inst 0x45189b56 // smmla z22.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #16\n"
+ ".inst 0x45199bab // smmla z11.s, z29.b, z25.b\n"
+ ".inst 0x45199b53 // smmla z19.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #-8, MUL VL]\n"
+ "add x26, x26, #0x10\n"
+ ".inst 0x45189baf // smmla z15.s, z29.b, z24.b\n"
+ ".inst 0x45189b57 // smmla z23.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #-7, MUL VL]\n"
+ "add x25, x25, #0x10\n"
+ ".inst 0x45199bc8 // smmla z8.s, z30.b, z25.b\n"
+ ".inst 0x45199b90 // smmla z16.s, z28.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #-6, MUL VL]\n"
+ "add x24, x24, #0x10\n"
+ ".inst 0x45189bcc // smmla z12.s, z30.b, z24.b\n"
+ ".inst 0x45189b94 // smmla z20.s, z28.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #-5, MUL VL]\n"
+ "add x23, x23, #0x10\n"
+ ".inst 0x45199bc9 // smmla z9.s, z30.b, z25.b\n"
+ ".inst 0x45199b91 // smmla z17.s, z28.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #-4, MUL VL]\n"
+ ".inst 0x45189bcd // smmla z13.s, z30.b, z24.b\n"
+ ".inst 0x45189b95 // smmla z21.s, z28.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #-3, MUL VL]\n"
+ ".inst 0x45199bca // smmla z10.s, z30.b, z25.b\n"
+ ".inst 0x45199b92 // smmla z18.s, z28.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #-2, MUL VL]\n"
+ ".inst 0x45189bce // smmla z14.s, z30.b, z24.b\n"
+ ".inst 0x45189b96 // smmla z22.s, z28.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #-1, MUL VL]\n"
+ ".inst 0x45199bcb // smmla z11.s, z30.b, z25.b\n"
+ ".inst 0x45199b93 // smmla z19.s, z28.b, z25.b\n"
+ ".inst 0x45189bcf // smmla z15.s, z30.b, z24.b\n"
+ ".inst 0x45189b97 // smmla z23.s, z28.b, z24.b\n"
+ "bgt 46b\n"
+ "47:" // Height 4: Multiply loop: Single iteration only
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z1.b }, p0/Z, [x26]\n"
+ "ld1rqb { z24.b }, p0/Z, [x25]\n"
+ "trn1 z28.d, z1.d, z24.d\n"
+ "ld1rqb { z3.b }, p0/Z, [x24]\n"
+ "ld1rqb { z27.b }, p0/Z, [x23]\n"
+ "trn2 z1.d, z1.d, z24.d\n"
+ "trn1 z26.d, z3.d, z27.d\n"
+ "ld1b { z25.b }, p2/Z, [x9]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45199b88 // smmla z8.s, z28.b, z25.b\n"
+ ".inst 0x45199b50 // smmla z16.s, z26.b, z25.b\n"
+ ".inst 0x45189b8c // smmla z12.s, z28.b, z24.b\n"
+ ".inst 0x45189b54 // smmla z20.s, z26.b, z24.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45199b89 // smmla z9.s, z28.b, z25.b\n"
+ ".inst 0x45199b51 // smmla z17.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "subs x27, x27, #0x8\n"
+ ".inst 0x45189b8d // smmla z13.s, z28.b, z24.b\n"
+ ".inst 0x45189b55 // smmla z21.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+ "trn2 z3.d, z3.d, z27.d\n"
+ ".inst 0x45199b8a // smmla z10.s, z28.b, z25.b\n"
+ ".inst 0x45199b52 // smmla z18.s, z26.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x45189b8e // smmla z14.s, z28.b, z24.b\n"
+ ".inst 0x45189b56 // smmla z22.s, z26.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #8\n"
+ ".inst 0x45199b8b // smmla z11.s, z28.b, z25.b\n"
+ ".inst 0x45199b53 // smmla z19.s, z26.b, z25.b\n"
+ ".inst 0x45189b8f // smmla z15.s, z28.b, z24.b\n"
+ ".inst 0x45189b57 // smmla z23.s, z26.b, z24.b\n"
+ "ble 48f\n"
+ "ld1b { z25.b }, p2/Z, [x9]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45199828 // smmla z8.s, z1.b, z25.b\n"
+ ".inst 0x45199870 // smmla z16.s, z3.b, z25.b\n"
+ ".inst 0x4518982c // smmla z12.s, z1.b, z24.b\n"
+ ".inst 0x45189874 // smmla z20.s, z3.b, z24.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45199829 // smmla z9.s, z1.b, z25.b\n"
+ ".inst 0x45199871 // smmla z17.s, z3.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+ ".inst 0x4518982d // smmla z13.s, z1.b, z24.b\n"
+ ".inst 0x45189875 // smmla z21.s, z3.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x4519982a // smmla z10.s, z1.b, z25.b\n"
+ ".inst 0x45199872 // smmla z18.s, z3.b, z25.b\n"
+ "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x4518982e // smmla z14.s, z1.b, z24.b\n"
+ ".inst 0x45189876 // smmla z22.s, z3.b, z24.b\n"
+ "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #8\n"
+ ".inst 0x4519982b // smmla z11.s, z1.b, z25.b\n"
+ ".inst 0x45199873 // smmla z19.s, z3.b, z25.b\n"
+ ".inst 0x4518982f // smmla z15.s, z1.b, z24.b\n"
+ ".inst 0x45189877 // smmla z23.s, z3.b, z24.b\n"
+ "48:" // Height 4: Multiply loop: multiply skip
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x28, x28, #0x1\n"
+ "cmp x28, x20\n"
+ "bne 43b\n"
+ "uzp1 z28.d, z8.d, z12.d\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "uzp2 z8.d, z8.d, z12.d\n"
+ "ld1w { z27.s }, p2/Z, [x14]\n"
+ "uzp1 z12.d, z9.d, z13.d\n"
+ "uzp2 z9.d, z9.d, z13.d\n"
+ "ld1w { z26.s }, p2/Z, [x14, #1, MUL VL]\n"
+ "ld1w { z25.s }, p2/Z, [x14, #2, MUL VL]\n"
+ "uzp1 z13.d, z10.d, z14.d\n"
+ "uzp2 z10.d, z10.d, z14.d\n"
+ "ld1w { z24.s }, p2/Z, [x14, #3, MUL VL]\n"
+ "add x26, x11, x20\n"
+ "uzp1 z14.d, z11.d, z15.d\n"
+ "uzp2 z11.d, z11.d, z15.d\n"
+ "add x25, x26, x20\n"
+ "add x24, x25, x20\n"
+ "uzp1 z15.d, z16.d, z20.d\n"
+ "uzp2 z16.d, z16.d, z20.d\n"
+ "addvl x14, x14, #4\n"
+ "uzp1 z20.d, z17.d, z21.d\n"
+ "uzp2 z17.d, z17.d, z21.d\n"
+ "uzp1 z21.d, z18.d, z22.d\n"
+ "uzp2 z18.d, z18.d, z22.d\n"
+ "uzp1 z22.d, z19.d, z23.d\n"
+ "uzp2 z19.d, z19.d, z23.d\n"
+ "mov z23.d, z28.d\n"
+ "add z23.s, z23.s, z27.s\n"
+ "add z12.s, z12.s, z26.s\n"
+ "add z13.s, z13.s, z25.s\n"
+ "add z14.s, z14.s, z24.s\n"
+ "add z8.s, z8.s, z27.s\n"
+ "add z9.s, z9.s, z26.s\n"
+ "add z10.s, z10.s, z25.s\n"
+ "add z11.s, z11.s, z24.s\n"
+ "add z15.s, z15.s, z27.s\n"
+ "add z20.s, z20.s, z26.s\n"
+ "add z21.s, z21.s, z25.s\n"
+ "add z22.s, z22.s, z24.s\n"
+ "add z16.s, z16.s, z27.s\n"
+ "add z17.s, z17.s, z26.s\n"
+ "add z18.s, z18.s, z25.s\n"
+ "add z19.s, z19.s, z24.s\n"
+ "tbz %x[flags], #4, 49f\n"
+ "ld1w { z0.s }, p2/Z, [x12]\n"
+ "ld1w { z4.s }, p2/Z, [x13]\n"
+ "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n"
+ "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n"
+ "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n"
+ "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n"
+ "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n"
+ "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n"
+ "addvl x12, x12, #4\n"
+ "addvl x13, x13, #4\n"
+ "b 50f\n"
+ "49:" // Height 4: per layer parameters
+ "add x21, %x[qp], %[per_layer_right_shift]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1rw { z0.s }, p2/Z, [x21]\n"
+ "ld1rw { z4.s }, p2/Z, [x20]\n"
+ "mov z1.d, z0.d\n"
+ "mov z5.d, z4.d\n"
+ "mov z2.d, z0.d\n"
+ "mov z6.d, z4.d\n"
+ "mov z3.d, z0.d\n"
+ "mov z7.d, z4.d\n"
+ "50:" // Height 4: parameters loaded
+ ".inst 0x04a476f7 // sqrdmulh z23.s, z23.s, z4.s\n"
+ ".inst 0x04a5758c // sqrdmulh z12.s, z12.s, z5.s\n"
+ ".inst 0x04a675ad // sqrdmulh z13.s, z13.s, z6.s\n"
+ ".inst 0x04a775ce // sqrdmulh z14.s, z14.s, z7.s\n"
+ ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n"
+ ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n"
+ ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n"
+ ".inst 0x04a7756b // sqrdmulh z11.s, z11.s, z7.s\n"
+ ".inst 0x04a475ef // sqrdmulh z15.s, z15.s, z4.s\n"
+ ".inst 0x04a57694 // sqrdmulh z20.s, z20.s, z5.s\n"
+ ".inst 0x04a676b5 // sqrdmulh z21.s, z21.s, z6.s\n"
+ ".inst 0x04a776d6 // sqrdmulh z22.s, z22.s, z7.s\n"
+ ".inst 0x04a47610 // sqrdmulh z16.s, z16.s, z4.s\n"
+ ".inst 0x04a57631 // sqrdmulh z17.s, z17.s, z5.s\n"
+ ".inst 0x04a67652 // sqrdmulh z18.s, z18.s, z6.s\n"
+ ".inst 0x04a77673 // sqrdmulh z19.s, z19.s, z7.s\n"
+ "tbz %x[flags], #5, 51f\n"
+ "and z27.d, z23.d, z0.d\n"
+ "and z26.d, z12.d, z1.d\n"
+ "and z25.d, z13.d, z2.d\n"
+ "and z24.d, z14.d, z3.d\n"
+ "asr z27.s, z27.s, #0x1f\n"
+ "asr z26.s, z26.s, #0x1f\n"
+ "asr z25.s, z25.s, #0x1f\n"
+ "asr z24.s, z24.s, #0x1f\n"
+ "sqadd z23.s, z23.s, z27.s\n"
+ "sqadd z12.s, z12.s, z26.s\n"
+ "sqadd z13.s, z13.s, z25.s\n"
+ "sqadd z14.s, z14.s, z24.s\n"
+ "and z27.d, z8.d, z0.d\n"
+ "and z26.d, z9.d, z1.d\n"
+ "and z25.d, z10.d, z2.d\n"
+ "and z24.d, z11.d, z3.d\n"
+ "asr z27.s, z27.s, #0x1f\n"
+ "asr z26.s, z26.s, #0x1f\n"
+ "asr z25.s, z25.s, #0x1f\n"
+ "asr z24.s, z24.s, #0x1f\n"
+ "sqadd z8.s, z8.s, z27.s\n"
+ "sqadd z9.s, z9.s, z26.s\n"
+ "sqadd z10.s, z10.s, z25.s\n"
+ "sqadd z11.s, z11.s, z24.s\n"
+ "and z27.d, z15.d, z0.d\n"
+ "and z26.d, z20.d, z1.d\n"
+ "and z25.d, z21.d, z2.d\n"
+ "and z24.d, z22.d, z3.d\n"
+ "asr z27.s, z27.s, #0x1f\n"
+ "asr z26.s, z26.s, #0x1f\n"
+ "asr z25.s, z25.s, #0x1f\n"
+ "asr z24.s, z24.s, #0x1f\n"
+ "sqadd z15.s, z15.s, z27.s\n"
+ "sqadd z20.s, z20.s, z26.s\n"
+ "sqadd z21.s, z21.s, z25.s\n"
+ "sqadd z22.s, z22.s, z24.s\n"
+ "and z27.d, z16.d, z0.d\n"
+ "and z26.d, z17.d, z1.d\n"
+ "and z25.d, z18.d, z2.d\n"
+ "and z24.d, z19.d, z3.d\n"
+ "asr z27.s, z27.s, #0x1f\n"
+ "asr z26.s, z26.s, #0x1f\n"
+ "asr z25.s, z25.s, #0x1f\n"
+ "asr z24.s, z24.s, #0x1f\n"
+ "sqadd z16.s, z16.s, z27.s\n"
+ "sqadd z17.s, z17.s, z26.s\n"
+ "sqadd z18.s, z18.s, z25.s\n"
+ "sqadd z19.s, z19.s, z24.s\n"
+ "51:" // Height 4: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1rw { z25.s }, p2/Z, [x20]\n"
+ ".inst 0x44828817 // srshl z23.s, p2/M, z23.s, z0.s\n"
+ "add z23.s, z23.s, z25.s\n"
+ ".inst 0x4482882c // srshl z12.s, p2/M, z12.s, z1.s\n"
+ ".inst 0x4482884d // srshl z13.s, p2/M, z13.s, z2.s\n"
+ "add z12.s, z12.s, z25.s\n"
+ "add z13.s, z13.s, z25.s\n"
+ ".inst 0x4482886e // srshl z14.s, p2/M, z14.s, z3.s\n"
+ ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n"
+ "add z14.s, z14.s, z25.s\n"
+ "add z8.s, z8.s, z25.s\n"
+ ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n"
+ ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n"
+ "add z9.s, z9.s, z25.s\n"
+ "add z10.s, z10.s, z25.s\n"
+ ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n"
+ ".inst 0x4482880f // srshl z15.s, p2/M, z15.s, z0.s\n"
+ "add z11.s, z11.s, z25.s\n"
+ "add z15.s, z15.s, z25.s\n"
+ ".inst 0x44828834 // srshl z20.s, p2/M, z20.s, z1.s\n"
+ ".inst 0x44828855 // srshl z21.s, p2/M, z21.s, z2.s\n"
+ "add z20.s, z20.s, z25.s\n"
+ "add z21.s, z21.s, z25.s\n"
+ ".inst 0x44828876 // srshl z22.s, p2/M, z22.s, z3.s\n"
+ ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n"
+ "add z22.s, z22.s, z25.s\n"
+ "add z16.s, z16.s, z25.s\n"
+ ".inst 0x44828831 // srshl z17.s, p2/M, z17.s, z1.s\n"
+ ".inst 0x44828852 // srshl z18.s, p2/M, z18.s, z2.s\n"
+ "add z17.s, z17.s, z25.s\n"
+ "add z18.s, z18.s, z25.s\n"
+ ".inst 0x44828873 // srshl z19.s, p2/M, z19.s, z3.s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1rw { z24.s }, p2/Z, [x20]\n"
+ "add z19.s, z19.s, z25.s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1rw { z25.s }, p2/Z, [x20]\n"
+ "smin z23.s, p2/M, z23.s, z24.s\n"
+ "smin z12.s, p2/M, z12.s, z24.s\n"
+ "smin z13.s, p2/M, z13.s, z24.s\n"
+ "smin z14.s, p2/M, z14.s, z24.s\n"
+ "smin z8.s, p2/M, z8.s, z24.s\n"
+ "smin z9.s, p2/M, z9.s, z24.s\n"
+ "smin z10.s, p2/M, z10.s, z24.s\n"
+ "smin z11.s, p2/M, z11.s, z24.s\n"
+ "smin z15.s, p2/M, z15.s, z24.s\n"
+ "smin z20.s, p2/M, z20.s, z24.s\n"
+ "smin z21.s, p2/M, z21.s, z24.s\n"
+ "smin z22.s, p2/M, z22.s, z24.s\n"
+ "smin z16.s, p2/M, z16.s, z24.s\n"
+ "smin z17.s, p2/M, z17.s, z24.s\n"
+ "smin z18.s, p2/M, z18.s, z24.s\n"
+ "smin z19.s, p2/M, z19.s, z24.s\n"
+ "smax z23.s, p2/M, z23.s, z25.s\n"
+ "smax z12.s, p2/M, z12.s, z25.s\n"
+ "smax z13.s, p2/M, z13.s, z25.s\n"
+ "uzp1 z23.h, z23.h, z12.h\n"
+ "smax z14.s, p2/M, z14.s, z25.s\n"
+ "smax z8.s, p2/M, z8.s, z25.s\n"
+ "uzp1 z24.h, z13.h, z14.h\n"
+ "uzp1 z23.b, z23.b, z24.b\n"
+ "smax z9.s, p2/M, z9.s, z25.s\n"
+ "smax z10.s, p2/M, z10.s, z25.s\n"
+ "uzp1 z8.h, z8.h, z9.h\n"
+ "st1b { z23.b }, p1, [x11]\n"
+ "smax z11.s, p2/M, z11.s, z25.s\n"
+ "smax z15.s, p2/M, z15.s, z25.s\n"
+ "uzp1 z23.h, z10.h, z11.h\n"
+ "uzp1 z8.b, z8.b, z23.b\n"
+ "smax z20.s, p2/M, z20.s, z25.s\n"
+ "smax z21.s, p2/M, z21.s, z25.s\n"
+ "uzp1 z15.h, z15.h, z20.h\n"
+ "st1b { z8.b }, p1, [x26]\n"
+ "smax z22.s, p2/M, z22.s, z25.s\n"
+ "smax z16.s, p2/M, z16.s, z25.s\n"
+ "uzp1 z20.h, z21.h, z22.h\n"
+ "uzp1 z15.b, z15.b, z20.b\n"
+ "smax z17.s, p2/M, z17.s, z25.s\n"
+ "smax z18.s, p2/M, z18.s, z25.s\n"
+ "uzp1 z16.h, z16.h, z17.h\n"
+ "st1b { z15.b }, p1, [x25]\n"
+ "smax z19.s, p2/M, z19.s, z25.s\n"
+ "uzp1 z17.h, z18.h, z19.h\n"
+ "uzp1 z16.b, z16.b, z17.b\n"
+ "st1b { z16.b }, p1, [x24]\n"
+ "addvl x11, x11, #1\n"
+ "52:" // Height 4: Writeback done
+ "decw x10, ALL, MUL #4\n"
+ "cmp x10, XZR\n"
+ "bgt 41b\n"
+ "b 80f\n"
+ "53:" // Height 5
+ "mov x14, %x[col_bias]\n"
+ "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
+ "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
+ "mov x11, %x[output_ptr]\n"
+ "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "54:" // Height 5: Column loop
+ "mov x20, #0x0\n"
+ "whilelt p1.b, x20, x10\n"
+ "mov z8.s, #0x0\n"
+ "mov z9.s, #0x0\n"
+ "mov z10.s, #0x0\n"
+ "mov z11.s, #0x0\n"
+ "mov z12.s, #0x0\n"
+ "mov z13.s, #0x0\n"
+ "mov z14.s, #0x0\n"
+ "mov z15.s, #0x0\n"
+ "mov z16.s, #0x0\n"
+ "mov z17.s, #0x0\n"
+ "mov z18.s, #0x0\n"
+ "mov z19.s, #0x0\n"
+ "mov z20.s, #0x0\n"
+ "mov z21.s, #0x0\n"
+ "mov z22.s, #0x0\n"
+ "mov z23.s, #0x0\n"
+ "mov z24.s, #0x0\n"
+ "mov z25.s, #0x0\n"
+ "mov z26.s, #0x0\n"
+ "mov z27.s, #0x0\n"
+ "mov z28.s, #0x0\n"
+ "mov z29.s, #0x0\n"
+ "mov z30.s, #0x0\n"
+ "mov z31.s, #0x0\n"
+ "55:" // Height 5: setup done
+ "mov x28, #0x0\n"
+ "56:" // Height 5: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w27, [x20, x28, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 57f\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
+ "ldr x22, [x20, #0x20]\n"
+ "cbnz x28, 58f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x26, x26, x20\n"
+ "add x25, x25, x20\n"
+ "add x24, x24, x20\n"
+ "add x23, x23, x20\n"
+ "add x22, x22, x20\n"
+ "b 58f\n"
+ "57:" // Height 5: setup direct input
+ "mov x26, %x[input_ptr]\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "add x23, x24, x21\n"
+ "add x22, x23, x21\n"
+ "58:" // Height 5: input setup done
+ "cmp x27, #0x10\n"
+ "ble 60f\n"
+ "59:" // Height 5: Multiply loop: Main loop head
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z6.b }, p0/Z, [x26]\n"
+ "ld1rqb { z1.b }, p0/Z, [x25]\n"
+ "ld1rqb { z7.b }, p0/Z, [x24]\n"
+ "ld1rqb { z2.b }, p0/Z, [x23]\n"
+ "trn1 z5.d, z6.d, z1.d\n"
+ "trn2 z6.d, z6.d, z1.d\n"
+ "ld1rqb { z4.b }, p0/Z, [x22]\n"
+ "trn1 z3.d, z7.d, z2.d\n"
+ "trn2 z7.d, z7.d, z2.d\n"
+ "ld1b { z1.b }, p2/Z, [x9]\n"
+ "trn1 z2.d, z4.d, z0.d\n"
+ "trn2 z4.d, z4.d, z0.d\n"
+ "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x450198a8 // smmla z8.s, z5.b, z1.b\n"
+ ".inst 0x45019870 // smmla z16.s, z3.b, z1.b\n"
+ ".inst 0x45019858 // smmla z24.s, z2.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "sub x27, x27, #0x10\n"
+ ".inst 0x450098ac // smmla z12.s, z5.b, z0.b\n"
+ ".inst 0x45009874 // smmla z20.s, z3.b, z0.b\n"
+ "cmp x27, #0x10\n"
+ "add x26, x26, #0x10\n"
+ ".inst 0x4500985c // smmla z28.s, z2.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x450198a9 // smmla z9.s, z5.b, z1.b\n"
+ "add x25, x25, #0x10\n"
+ ".inst 0x45019871 // smmla z17.s, z3.b, z1.b\n"
+ ".inst 0x45019859 // smmla z25.s, z2.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "add x24, x24, #0x10\n"
+ ".inst 0x450098ad // smmla z13.s, z5.b, z0.b\n"
+ ".inst 0x45009875 // smmla z21.s, z3.b, z0.b\n"
+ "add x23, x23, #0x10\n"
+ "add x22, x22, #0x10\n"
+ ".inst 0x4500985d // smmla z29.s, z2.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x450198aa // smmla z10.s, z5.b, z1.b\n"
+ ".inst 0x45019872 // smmla z18.s, z3.b, z1.b\n"
+ ".inst 0x4501985a // smmla z26.s, z2.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x450098ae // smmla z14.s, z5.b, z0.b\n"
+ ".inst 0x45009876 // smmla z22.s, z3.b, z0.b\n"
+ ".inst 0x4500985e // smmla z30.s, z2.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #16\n"
+ ".inst 0x450198ab // smmla z11.s, z5.b, z1.b\n"
+ ".inst 0x45019873 // smmla z19.s, z3.b, z1.b\n"
+ ".inst 0x4501985b // smmla z27.s, z2.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #-8, MUL VL]\n"
+ ".inst 0x450098af // smmla z15.s, z5.b, z0.b\n"
+ ".inst 0x45009877 // smmla z23.s, z3.b, z0.b\n"
+ ".inst 0x4500985f // smmla z31.s, z2.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #-7, MUL VL]\n"
+ ".inst 0x450198c8 // smmla z8.s, z6.b, z1.b\n"
+ ".inst 0x450198f0 // smmla z16.s, z7.b, z1.b\n"
+ ".inst 0x45019898 // smmla z24.s, z4.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #-6, MUL VL]\n"
+ ".inst 0x450098cc // smmla z12.s, z6.b, z0.b\n"
+ ".inst 0x450098f4 // smmla z20.s, z7.b, z0.b\n"
+ ".inst 0x4500989c // smmla z28.s, z4.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #-5, MUL VL]\n"
+ ".inst 0x450198c9 // smmla z9.s, z6.b, z1.b\n"
+ ".inst 0x450198f1 // smmla z17.s, z7.b, z1.b\n"
+ ".inst 0x45019899 // smmla z25.s, z4.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #-4, MUL VL]\n"
+ ".inst 0x450098cd // smmla z13.s, z6.b, z0.b\n"
+ ".inst 0x450098f5 // smmla z21.s, z7.b, z0.b\n"
+ ".inst 0x4500989d // smmla z29.s, z4.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #-3, MUL VL]\n"
+ ".inst 0x450198ca // smmla z10.s, z6.b, z1.b\n"
+ ".inst 0x450198f2 // smmla z18.s, z7.b, z1.b\n"
+ ".inst 0x4501989a // smmla z26.s, z4.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #-2, MUL VL]\n"
+ ".inst 0x450098ce // smmla z14.s, z6.b, z0.b\n"
+ ".inst 0x450098f6 // smmla z22.s, z7.b, z0.b\n"
+ ".inst 0x4500989e // smmla z30.s, z4.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #-1, MUL VL]\n"
+ ".inst 0x450198cb // smmla z11.s, z6.b, z1.b\n"
+ ".inst 0x450198f3 // smmla z19.s, z7.b, z1.b\n"
+ ".inst 0x4501989b // smmla z27.s, z4.b, z1.b\n"
+ ".inst 0x450098cf // smmla z15.s, z6.b, z0.b\n"
+ ".inst 0x450098f7 // smmla z23.s, z7.b, z0.b\n"
+ ".inst 0x4500989f // smmla z31.s, z4.b, z0.b\n"
+ "bgt 59b\n"
+ "60:" // Height 5: Multiply loop: Single iteration only
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z1.b }, p0/Z, [x26]\n"
+ "ld1rqb { z4.b }, p0/Z, [x25]\n"
+ "ld1rqb { z3.b }, p0/Z, [x24]\n"
+ "ld1rqb { z2.b }, p0/Z, [x23]\n"
+ "trn1 z7.d, z1.d, z4.d\n"
+ "trn2 z1.d, z1.d, z4.d\n"
+ "ld1rqb { z5.b }, p0/Z, [x22]\n"
+ "trn1 z6.d, z3.d, z2.d\n"
+ "trn2 z3.d, z3.d, z2.d\n"
+ "ld1b { z2.b }, p2/Z, [x9]\n"
+ "trn1 z4.d, z5.d, z0.d\n"
+ "trn2 z5.d, z5.d, z0.d\n"
+ "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x450298e8 // smmla z8.s, z7.b, z2.b\n"
+ ".inst 0x450298d0 // smmla z16.s, z6.b, z2.b\n"
+ ".inst 0x45029898 // smmla z24.s, z4.b, z2.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "subs x27, x27, #0x8\n"
+ ".inst 0x450098ec // smmla z12.s, z7.b, z0.b\n"
+ ".inst 0x450098d4 // smmla z20.s, z6.b, z0.b\n"
+ ".inst 0x4500989c // smmla z28.s, z4.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x450298e9 // smmla z9.s, z7.b, z2.b\n"
+ ".inst 0x450298d1 // smmla z17.s, z6.b, z2.b\n"
+ ".inst 0x45029899 // smmla z25.s, z4.b, z2.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #4, MUL VL]\n"
+ ".inst 0x450098ed // smmla z13.s, z7.b, z0.b\n"
+ ".inst 0x450098d5 // smmla z21.s, z6.b, z0.b\n"
+ ".inst 0x4500989d // smmla z29.s, z4.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x450298ea // smmla z10.s, z7.b, z2.b\n"
+ ".inst 0x450298d2 // smmla z18.s, z6.b, z2.b\n"
+ ".inst 0x4502989a // smmla z26.s, z4.b, z2.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x450098ee // smmla z14.s, z7.b, z0.b\n"
+ ".inst 0x450098d6 // smmla z22.s, z6.b, z0.b\n"
+ ".inst 0x4500989e // smmla z30.s, z4.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+ ".inst 0x450298eb // smmla z11.s, z7.b, z2.b\n"
+ "addvl x9, x9, #8\n"
+ ".inst 0x450298d3 // smmla z19.s, z6.b, z2.b\n"
+ ".inst 0x4502989b // smmla z27.s, z4.b, z2.b\n"
+ ".inst 0x450098ef // smmla z15.s, z7.b, z0.b\n"
+ ".inst 0x450098d7 // smmla z23.s, z6.b, z0.b\n"
+ ".inst 0x4500989f // smmla z31.s, z4.b, z0.b\n"
+ "ble 61f\n"
+ "ld1b { z2.b }, p2/Z, [x9]\n"
+ "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45029828 // smmla z8.s, z1.b, z2.b\n"
+ ".inst 0x45029870 // smmla z16.s, z3.b, z2.b\n"
+ ".inst 0x450298b8 // smmla z24.s, z5.b, z2.b\n"
+ ".inst 0x4500982c // smmla z12.s, z1.b, z0.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #2, MUL VL]\n"
+ ".inst 0x45009874 // smmla z20.s, z3.b, z0.b\n"
+ ".inst 0x450098bc // smmla z28.s, z5.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45029829 // smmla z9.s, z1.b, z2.b\n"
+ ".inst 0x45029871 // smmla z17.s, z3.b, z2.b\n"
+ ".inst 0x450298b9 // smmla z25.s, z5.b, z2.b\n"
+ ".inst 0x4500982d // smmla z13.s, z1.b, z0.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #4, MUL VL]\n"
+ ".inst 0x45009875 // smmla z21.s, z3.b, z0.b\n"
+ ".inst 0x450098bd // smmla z29.s, z5.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x4502982a // smmla z10.s, z1.b, z2.b\n"
+ ".inst 0x45029872 // smmla z18.s, z3.b, z2.b\n"
+ ".inst 0x450298ba // smmla z26.s, z5.b, z2.b\n"
+ ".inst 0x4500982e // smmla z14.s, z1.b, z0.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x45009876 // smmla z22.s, z3.b, z0.b\n"
+ ".inst 0x450098be // smmla z30.s, z5.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #8\n"
+ ".inst 0x4502982b // smmla z11.s, z1.b, z2.b\n"
+ ".inst 0x45029873 // smmla z19.s, z3.b, z2.b\n"
+ ".inst 0x450298bb // smmla z27.s, z5.b, z2.b\n"
+ ".inst 0x4500982f // smmla z15.s, z1.b, z0.b\n"
+ ".inst 0x45009877 // smmla z23.s, z3.b, z0.b\n"
+ ".inst 0x450098bf // smmla z31.s, z5.b, z0.b\n"
+ "61:" // Height 5: Multiply loop: multiply skip
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x28, x28, #0x1\n"
+ "cmp x28, x20\n"
+ "bne 56b\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "uzp1 z4.d, z8.d, z12.d\n"
+ "uzp2 z8.d, z8.d, z12.d\n"
+ "add x26, x11, x20\n"
+ "uzp1 z12.d, z9.d, z13.d\n"
+ "uzp2 z9.d, z9.d, z13.d\n"
+ "ld1w { z3.s }, p2/Z, [x14]\n"
+ "ld1w { z2.s }, p2/Z, [x14, #1, MUL VL]\n"
+ "uzp1 z13.d, z10.d, z14.d\n"
+ "uzp2 z10.d, z10.d, z14.d\n"
+ "ld1w { z1.s }, p2/Z, [x14, #2, MUL VL]\n"
+ "ld1w { z0.s }, p2/Z, [x14, #3, MUL VL]\n"
+ "uzp1 z14.d, z11.d, z15.d\n"
+ "uzp2 z11.d, z11.d, z15.d\n"
+ "add x25, x26, x20\n"
+ "add x24, x25, x20\n"
+ "uzp1 z15.d, z16.d, z20.d\n"
+ "uzp2 z16.d, z16.d, z20.d\n"
+ "add x23, x24, x20\n"
+ "addvl x14, x14, #4\n"
+ "uzp1 z20.d, z17.d, z21.d\n"
+ "uzp2 z17.d, z17.d, z21.d\n"
+ "uzp1 z21.d, z18.d, z22.d\n"
+ "uzp2 z18.d, z18.d, z22.d\n"
+ "uzp1 z22.d, z19.d, z23.d\n"
+ "uzp2 z19.d, z19.d, z23.d\n"
+ "uzp1 z24.d, z24.d, z28.d\n"
+ "uzp1 z25.d, z25.d, z29.d\n"
+ "uzp1 z26.d, z26.d, z30.d\n"
+ "uzp1 z27.d, z27.d, z31.d\n"
+ "mov z31.d, z4.d\n"
+ "add z31.s, z31.s, z3.s\n"
+ "add z12.s, z12.s, z2.s\n"
+ "add z13.s, z13.s, z1.s\n"
+ "add z14.s, z14.s, z0.s\n"
+ "add z8.s, z8.s, z3.s\n"
+ "add z9.s, z9.s, z2.s\n"
+ "add z10.s, z10.s, z1.s\n"
+ "add z11.s, z11.s, z0.s\n"
+ "add z15.s, z15.s, z3.s\n"
+ "add z20.s, z20.s, z2.s\n"
+ "add z21.s, z21.s, z1.s\n"
+ "add z22.s, z22.s, z0.s\n"
+ "add z16.s, z16.s, z3.s\n"
+ "add z17.s, z17.s, z2.s\n"
+ "add z18.s, z18.s, z1.s\n"
+ "add z19.s, z19.s, z0.s\n"
+ "add z24.s, z24.s, z3.s\n"
+ "add z25.s, z25.s, z2.s\n"
+ "add z26.s, z26.s, z1.s\n"
+ "add z27.s, z27.s, z0.s\n"
+ "tbz %x[flags], #4, 62f\n"
+ "ld1w { z0.s }, p2/Z, [x12]\n"
+ "ld1w { z4.s }, p2/Z, [x13]\n"
+ "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n"
+ "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n"
+ "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n"
+ "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n"
+ "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n"
+ "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n"
+ "addvl x12, x12, #4\n"
+ "addvl x13, x13, #4\n"
+ "b 63f\n"
+ "62:" // Height 5: per layer parameters
+ "add x21, %x[qp], %[per_layer_right_shift]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1rw { z0.s }, p2/Z, [x21]\n"
+ "ld1rw { z4.s }, p2/Z, [x20]\n"
+ "mov z1.d, z0.d\n"
+ "mov z5.d, z4.d\n"
+ "mov z2.d, z0.d\n"
+ "mov z6.d, z4.d\n"
+ "mov z3.d, z0.d\n"
+ "mov z7.d, z4.d\n"
+ "63:" // Height 5: parameters loaded
+ ".inst 0x04a477ff // sqrdmulh z31.s, z31.s, z4.s\n"
+ ".inst 0x04a5758c // sqrdmulh z12.s, z12.s, z5.s\n"
+ ".inst 0x04a675ad // sqrdmulh z13.s, z13.s, z6.s\n"
+ ".inst 0x04a775ce // sqrdmulh z14.s, z14.s, z7.s\n"
+ ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n"
+ ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n"
+ ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n"
+ ".inst 0x04a7756b // sqrdmulh z11.s, z11.s, z7.s\n"
+ ".inst 0x04a475ef // sqrdmulh z15.s, z15.s, z4.s\n"
+ ".inst 0x04a57694 // sqrdmulh z20.s, z20.s, z5.s\n"
+ ".inst 0x04a676b5 // sqrdmulh z21.s, z21.s, z6.s\n"
+ ".inst 0x04a776d6 // sqrdmulh z22.s, z22.s, z7.s\n"
+ ".inst 0x04a47610 // sqrdmulh z16.s, z16.s, z4.s\n"
+ ".inst 0x04a57631 // sqrdmulh z17.s, z17.s, z5.s\n"
+ ".inst 0x04a67652 // sqrdmulh z18.s, z18.s, z6.s\n"
+ ".inst 0x04a77673 // sqrdmulh z19.s, z19.s, z7.s\n"
+ ".inst 0x04a47718 // sqrdmulh z24.s, z24.s, z4.s\n"
+ ".inst 0x04a57739 // sqrdmulh z25.s, z25.s, z5.s\n"
+ ".inst 0x04a6775a // sqrdmulh z26.s, z26.s, z6.s\n"
+ ".inst 0x04a7777b // sqrdmulh z27.s, z27.s, z7.s\n"
+ "tbz %x[flags], #5, 64f\n"
+ "and z30.d, z31.d, z0.d\n"
+ "and z29.d, z12.d, z1.d\n"
+ "and z28.d, z13.d, z2.d\n"
+ "and z23.d, z14.d, z3.d\n"
+ "asr z30.s, z30.s, #0x1f\n"
+ "asr z29.s, z29.s, #0x1f\n"
+ "asr z28.s, z28.s, #0x1f\n"
+ "asr z23.s, z23.s, #0x1f\n"
+ "sqadd z31.s, z31.s, z30.s\n"
+ "sqadd z12.s, z12.s, z29.s\n"
+ "sqadd z13.s, z13.s, z28.s\n"
+ "sqadd z14.s, z14.s, z23.s\n"
+ "and z30.d, z8.d, z0.d\n"
+ "and z29.d, z9.d, z1.d\n"
+ "and z28.d, z10.d, z2.d\n"
+ "and z23.d, z11.d, z3.d\n"
+ "asr z30.s, z30.s, #0x1f\n"
+ "asr z29.s, z29.s, #0x1f\n"
+ "asr z28.s, z28.s, #0x1f\n"
+ "asr z23.s, z23.s, #0x1f\n"
+ "sqadd z8.s, z8.s, z30.s\n"
+ "sqadd z9.s, z9.s, z29.s\n"
+ "sqadd z10.s, z10.s, z28.s\n"
+ "sqadd z11.s, z11.s, z23.s\n"
+ "and z30.d, z15.d, z0.d\n"
+ "and z29.d, z20.d, z1.d\n"
+ "and z28.d, z21.d, z2.d\n"
+ "and z23.d, z22.d, z3.d\n"
+ "asr z30.s, z30.s, #0x1f\n"
+ "asr z29.s, z29.s, #0x1f\n"
+ "asr z28.s, z28.s, #0x1f\n"
+ "asr z23.s, z23.s, #0x1f\n"
+ "sqadd z15.s, z15.s, z30.s\n"
+ "sqadd z20.s, z20.s, z29.s\n"
+ "sqadd z21.s, z21.s, z28.s\n"
+ "sqadd z22.s, z22.s, z23.s\n"
+ "and z30.d, z16.d, z0.d\n"
+ "and z29.d, z17.d, z1.d\n"
+ "and z28.d, z18.d, z2.d\n"
+ "and z23.d, z19.d, z3.d\n"
+ "asr z30.s, z30.s, #0x1f\n"
+ "asr z29.s, z29.s, #0x1f\n"
+ "asr z28.s, z28.s, #0x1f\n"
+ "asr z23.s, z23.s, #0x1f\n"
+ "sqadd z16.s, z16.s, z30.s\n"
+ "sqadd z17.s, z17.s, z29.s\n"
+ "sqadd z18.s, z18.s, z28.s\n"
+ "sqadd z19.s, z19.s, z23.s\n"
+ "and z30.d, z24.d, z0.d\n"
+ "and z29.d, z25.d, z1.d\n"
+ "and z28.d, z26.d, z2.d\n"
+ "and z23.d, z27.d, z3.d\n"
+ "asr z30.s, z30.s, #0x1f\n"
+ "asr z29.s, z29.s, #0x1f\n"
+ "asr z28.s, z28.s, #0x1f\n"
+ "asr z23.s, z23.s, #0x1f\n"
+ "sqadd z24.s, z24.s, z30.s\n"
+ "sqadd z25.s, z25.s, z29.s\n"
+ "sqadd z26.s, z26.s, z28.s\n"
+ "sqadd z27.s, z27.s, z23.s\n"
+ "64:" // Height 5: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1rw { z28.s }, p2/Z, [x20]\n"
+ ".inst 0x4482881f // srshl z31.s, p2/M, z31.s, z0.s\n"
+ "add z31.s, z31.s, z28.s\n"
+ ".inst 0x4482882c // srshl z12.s, p2/M, z12.s, z1.s\n"
+ ".inst 0x4482884d // srshl z13.s, p2/M, z13.s, z2.s\n"
+ "add z12.s, z12.s, z28.s\n"
+ "add z13.s, z13.s, z28.s\n"
+ ".inst 0x4482886e // srshl z14.s, p2/M, z14.s, z3.s\n"
+ ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n"
+ "add z14.s, z14.s, z28.s\n"
+ "add z8.s, z8.s, z28.s\n"
+ ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n"
+ ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n"
+ "add z9.s, z9.s, z28.s\n"
+ "add z10.s, z10.s, z28.s\n"
+ ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n"
+ ".inst 0x4482880f // srshl z15.s, p2/M, z15.s, z0.s\n"
+ "add z11.s, z11.s, z28.s\n"
+ "add z15.s, z15.s, z28.s\n"
+ ".inst 0x44828834 // srshl z20.s, p2/M, z20.s, z1.s\n"
+ ".inst 0x44828855 // srshl z21.s, p2/M, z21.s, z2.s\n"
+ "add z20.s, z20.s, z28.s\n"
+ "add z21.s, z21.s, z28.s\n"
+ ".inst 0x44828876 // srshl z22.s, p2/M, z22.s, z3.s\n"
+ ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n"
+ "add z22.s, z22.s, z28.s\n"
+ "add z16.s, z16.s, z28.s\n"
+ ".inst 0x44828831 // srshl z17.s, p2/M, z17.s, z1.s\n"
+ ".inst 0x44828852 // srshl z18.s, p2/M, z18.s, z2.s\n"
+ "add z17.s, z17.s, z28.s\n"
+ "add z18.s, z18.s, z28.s\n"
+ ".inst 0x44828873 // srshl z19.s, p2/M, z19.s, z3.s\n"
+ ".inst 0x44828818 // srshl z24.s, p2/M, z24.s, z0.s\n"
+ "add z19.s, z19.s, z28.s\n"
+ "add z24.s, z24.s, z28.s\n"
+ ".inst 0x44828839 // srshl z25.s, p2/M, z25.s, z1.s\n"
+ ".inst 0x4482885a // srshl z26.s, p2/M, z26.s, z2.s\n"
+ "add z25.s, z25.s, z28.s\n"
+ "add z26.s, z26.s, z28.s\n"
+ ".inst 0x4482887b // srshl z27.s, p2/M, z27.s, z3.s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1rw { z23.s }, p2/Z, [x20]\n"
+ "add z27.s, z27.s, z28.s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1rw { z28.s }, p2/Z, [x20]\n"
+ "smin z31.s, p2/M, z31.s, z23.s\n"
+ "smin z12.s, p2/M, z12.s, z23.s\n"
+ "smin z13.s, p2/M, z13.s, z23.s\n"
+ "smin z14.s, p2/M, z14.s, z23.s\n"
+ "smin z8.s, p2/M, z8.s, z23.s\n"
+ "smin z9.s, p2/M, z9.s, z23.s\n"
+ "smin z10.s, p2/M, z10.s, z23.s\n"
+ "smin z11.s, p2/M, z11.s, z23.s\n"
+ "smin z15.s, p2/M, z15.s, z23.s\n"
+ "smin z20.s, p2/M, z20.s, z23.s\n"
+ "smin z21.s, p2/M, z21.s, z23.s\n"
+ "smin z22.s, p2/M, z22.s, z23.s\n"
+ "smin z16.s, p2/M, z16.s, z23.s\n"
+ "smin z17.s, p2/M, z17.s, z23.s\n"
+ "smin z18.s, p2/M, z18.s, z23.s\n"
+ "smin z19.s, p2/M, z19.s, z23.s\n"
+ "smin z24.s, p2/M, z24.s, z23.s\n"
+ "smin z25.s, p2/M, z25.s, z23.s\n"
+ "smin z26.s, p2/M, z26.s, z23.s\n"
+ "smin z27.s, p2/M, z27.s, z23.s\n"
+ "smax z31.s, p2/M, z31.s, z28.s\n"
+ "smax z12.s, p2/M, z12.s, z28.s\n"
+ "smax z13.s, p2/M, z13.s, z28.s\n"
+ "uzp1 z31.h, z31.h, z12.h\n"
+ "smax z14.s, p2/M, z14.s, z28.s\n"
+ "smax z8.s, p2/M, z8.s, z28.s\n"
+ "uzp1 z23.h, z13.h, z14.h\n"
+ "uzp1 z31.b, z31.b, z23.b\n"
+ "smax z9.s, p2/M, z9.s, z28.s\n"
+ "smax z10.s, p2/M, z10.s, z28.s\n"
+ "uzp1 z8.h, z8.h, z9.h\n"
+ "st1b { z31.b }, p1, [x11]\n"
+ "smax z11.s, p2/M, z11.s, z28.s\n"
+ "smax z15.s, p2/M, z15.s, z28.s\n"
+ "uzp1 z23.h, z10.h, z11.h\n"
+ "uzp1 z8.b, z8.b, z23.b\n"
+ "smax z20.s, p2/M, z20.s, z28.s\n"
+ "smax z21.s, p2/M, z21.s, z28.s\n"
+ "uzp1 z15.h, z15.h, z20.h\n"
+ "st1b { z8.b }, p1, [x26]\n"
+ "smax z22.s, p2/M, z22.s, z28.s\n"
+ "smax z16.s, p2/M, z16.s, z28.s\n"
+ "uzp1 z20.h, z21.h, z22.h\n"
+ "uzp1 z15.b, z15.b, z20.b\n"
+ "smax z17.s, p2/M, z17.s, z28.s\n"
+ "smax z18.s, p2/M, z18.s, z28.s\n"
+ "uzp1 z16.h, z16.h, z17.h\n"
+ "st1b { z15.b }, p1, [x25]\n"
+ "smax z19.s, p2/M, z19.s, z28.s\n"
+ "smax z24.s, p2/M, z24.s, z28.s\n"
+ "uzp1 z17.h, z18.h, z19.h\n"
+ "uzp1 z16.b, z16.b, z17.b\n"
+ "smax z25.s, p2/M, z25.s, z28.s\n"
+ "smax z26.s, p2/M, z26.s, z28.s\n"
+ "uzp1 z24.h, z24.h, z25.h\n"
+ "st1b { z16.b }, p1, [x24]\n"
+ "smax z27.s, p2/M, z27.s, z28.s\n"
+ "uzp1 z16.h, z26.h, z27.h\n"
+ "uzp1 z24.b, z24.b, z16.b\n"
+ "st1b { z24.b }, p1, [x23]\n"
+ "addvl x11, x11, #1\n"
+ "65:" // Height 5: Writeback done
+ "decw x10, ALL, MUL #4\n"
+ "cmp x10, XZR\n"
+ "bgt 54b\n"
+ "b 80f\n"
+ "66:" // Height 6
+ "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "mov x20, #0x6\n"
+ "mov x14, %x[col_bias]\n"
+ "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
+ "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
+ "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+ "mov x11, %x[output_ptr]\n"
+ "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "67:" // Height 6: Column loop
+ "mov x20, #0x0\n"
+ "whilelt p1.b, x20, x10\n"
+ "mov z8.s, #0x0\n"
+ "mov z9.s, #0x0\n"
+ "mov z10.s, #0x0\n"
+ "mov z11.s, #0x0\n"
+ "mov z12.s, #0x0\n"
+ "mov z13.s, #0x0\n"
+ "mov z14.s, #0x0\n"
+ "mov z15.s, #0x0\n"
+ "mov z16.s, #0x0\n"
+ "mov z17.s, #0x0\n"
+ "mov z18.s, #0x0\n"
+ "mov z19.s, #0x0\n"
+ "mov z20.s, #0x0\n"
+ "mov z21.s, #0x0\n"
+ "mov z22.s, #0x0\n"
+ "mov z23.s, #0x0\n"
+ "mov z24.s, #0x0\n"
+ "mov z25.s, #0x0\n"
+ "mov z26.s, #0x0\n"
+ "mov z27.s, #0x0\n"
+ "mov z28.s, #0x0\n"
+ "mov z29.s, #0x0\n"
+ "mov z30.s, #0x0\n"
+ "mov z31.s, #0x0\n"
+ "68:" // Height 6: setup done
+ "mov x28, #0x0\n"
+ "69:" // Height 6: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w27, [x20, x28, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 70f\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
+ "ldr x22, [x20, #0x20]\n"
+ "ldr x21, [x20, #0x28]\n"
+ "cbnz x28, 71f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x26, x26, x20\n"
+ "add x25, x25, x20\n"
+ "add x24, x24, x20\n"
+ "add x23, x23, x20\n"
+ "add x22, x22, x20\n"
+ "add x21, x21, x20\n"
+ "b 71f\n"
+ "70:" // Height 6: setup direct input
+ "mov x26, %x[input_ptr]\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "add x23, x24, x21\n"
+ "add x22, x23, x21\n"
+ "add x21, x22, x21\n"
+ "71:" // Height 6: input setup done
+ "cmp x27, #0x10\n"
+ "ble 73f\n"
+ "72:" // Height 6: Multiply loop: Main loop head
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z7.b }, p0/Z, [x26]\n"
+ "ld1rqb { z0.b }, p0/Z, [x25]\n"
+ "trn1 z6.d, z7.d, z0.d\n"
+ "ld1rqb { z5.b }, p0/Z, [x24]\n"
+ "ld1rqb { z1.b }, p0/Z, [x23]\n"
+ "trn2 z7.d, z7.d, z0.d\n"
+ "trn1 z4.d, z5.d, z1.d\n"
+ "ld1rqb { z3.b }, p0/Z, [x22]\n"
+ "ld1rqb { z0.b }, p0/Z, [x21]\n"
+ "trn2 z5.d, z5.d, z1.d\n"
+ "trn1 z2.d, z3.d, z0.d\n"
+ "trn2 z3.d, z3.d, z0.d\n"
+ "ld1b { z1.b }, p2/Z, [x9]\n"
+ "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x450198c8 // smmla z8.s, z6.b, z1.b\n"
+ ".inst 0x45019890 // smmla z16.s, z4.b, z1.b\n"
+ ".inst 0x45019858 // smmla z24.s, z2.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "sub x27, x27, #0x10\n"
+ ".inst 0x450098cc // smmla z12.s, z6.b, z0.b\n"
+ ".inst 0x45009894 // smmla z20.s, z4.b, z0.b\n"
+ "cmp x27, #0x10\n"
+ "add x26, x26, #0x10\n"
+ ".inst 0x4500985c // smmla z28.s, z2.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x450198c9 // smmla z9.s, z6.b, z1.b\n"
+ "add x25, x25, #0x10\n"
+ ".inst 0x45019891 // smmla z17.s, z4.b, z1.b\n"
+ ".inst 0x45019859 // smmla z25.s, z2.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #4, MUL VL]\n"
+ "add x24, x24, #0x10\n"
+ ".inst 0x450098cd // smmla z13.s, z6.b, z0.b\n"
+ ".inst 0x45009895 // smmla z21.s, z4.b, z0.b\n"
+ "add x23, x23, #0x10\n"
+ "add x22, x22, #0x10\n"
+ ".inst 0x4500985d // smmla z29.s, z2.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x450198ca // smmla z10.s, z6.b, z1.b\n"
+ "add x21, x21, #0x10\n"
+ ".inst 0x45019892 // smmla z18.s, z4.b, z1.b\n"
+ ".inst 0x4501985a // smmla z26.s, z2.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x450098ce // smmla z14.s, z6.b, z0.b\n"
+ ".inst 0x45009896 // smmla z22.s, z4.b, z0.b\n"
+ ".inst 0x4500985e // smmla z30.s, z2.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #16\n"
+ ".inst 0x450198cb // smmla z11.s, z6.b, z1.b\n"
+ ".inst 0x45019893 // smmla z19.s, z4.b, z1.b\n"
+ ".inst 0x4501985b // smmla z27.s, z2.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #-8, MUL VL]\n"
+ ".inst 0x450098cf // smmla z15.s, z6.b, z0.b\n"
+ ".inst 0x45009897 // smmla z23.s, z4.b, z0.b\n"
+ ".inst 0x4500985f // smmla z31.s, z2.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #-7, MUL VL]\n"
+ ".inst 0x450198e8 // smmla z8.s, z7.b, z1.b\n"
+ ".inst 0x450198b0 // smmla z16.s, z5.b, z1.b\n"
+ ".inst 0x45019878 // smmla z24.s, z3.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #-6, MUL VL]\n"
+ ".inst 0x450098ec // smmla z12.s, z7.b, z0.b\n"
+ ".inst 0x450098b4 // smmla z20.s, z5.b, z0.b\n"
+ ".inst 0x4500987c // smmla z28.s, z3.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #-5, MUL VL]\n"
+ ".inst 0x450198e9 // smmla z9.s, z7.b, z1.b\n"
+ ".inst 0x450198b1 // smmla z17.s, z5.b, z1.b\n"
+ ".inst 0x45019879 // smmla z25.s, z3.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #-4, MUL VL]\n"
+ ".inst 0x450098ed // smmla z13.s, z7.b, z0.b\n"
+ ".inst 0x450098b5 // smmla z21.s, z5.b, z0.b\n"
+ ".inst 0x4500987d // smmla z29.s, z3.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #-3, MUL VL]\n"
+ ".inst 0x450198ea // smmla z10.s, z7.b, z1.b\n"
+ ".inst 0x450198b2 // smmla z18.s, z5.b, z1.b\n"
+ ".inst 0x4501987a // smmla z26.s, z3.b, z1.b\n"
+ "ld1b { z1.b }, p2/Z, [x9, #-2, MUL VL]\n"
+ ".inst 0x450098ee // smmla z14.s, z7.b, z0.b\n"
+ ".inst 0x450098b6 // smmla z22.s, z5.b, z0.b\n"
+ ".inst 0x4500987e // smmla z30.s, z3.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #-1, MUL VL]\n"
+ ".inst 0x450198eb // smmla z11.s, z7.b, z1.b\n"
+ ".inst 0x450198b3 // smmla z19.s, z5.b, z1.b\n"
+ ".inst 0x4501987b // smmla z27.s, z3.b, z1.b\n"
+ ".inst 0x450098ef // smmla z15.s, z7.b, z0.b\n"
+ ".inst 0x450098b7 // smmla z23.s, z5.b, z0.b\n"
+ ".inst 0x4500987f // smmla z31.s, z3.b, z0.b\n"
+ "bgt 72b\n"
+ "73:" // Height 6: Multiply loop: Single iteration only
+ "whilelt p0.b, XZR, x27\n"
+ "ld1rqb { z1.b }, p0/Z, [x26]\n"
+ "ld1rqb { z0.b }, p0/Z, [x25]\n"
+ "trn1 z7.d, z1.d, z0.d\n"
+ "ld1rqb { z3.b }, p0/Z, [x24]\n"
+ "ld1rqb { z2.b }, p0/Z, [x23]\n"
+ "trn2 z1.d, z1.d, z0.d\n"
+ "trn1 z6.d, z3.d, z2.d\n"
+ "ld1rqb { z5.b }, p0/Z, [x22]\n"
+ "ld1rqb { z0.b }, p0/Z, [x21]\n"
+ "trn2 z3.d, z3.d, z2.d\n"
+ "trn1 z4.d, z5.d, z0.d\n"
+ "trn2 z5.d, z5.d, z0.d\n"
+ "ld1b { z2.b }, p2/Z, [x9]\n"
+ "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x450298e8 // smmla z8.s, z7.b, z2.b\n"
+ ".inst 0x450298d0 // smmla z16.s, z6.b, z2.b\n"
+ ".inst 0x45029898 // smmla z24.s, z4.b, z2.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #2, MUL VL]\n"
+ "subs x27, x27, #0x8\n"
+ ".inst 0x450098ec // smmla z12.s, z7.b, z0.b\n"
+ ".inst 0x450098d4 // smmla z20.s, z6.b, z0.b\n"
+ ".inst 0x4500989c // smmla z28.s, z4.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x450298e9 // smmla z9.s, z7.b, z2.b\n"
+ ".inst 0x450298d1 // smmla z17.s, z6.b, z2.b\n"
+ ".inst 0x45029899 // smmla z25.s, z4.b, z2.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #4, MUL VL]\n"
+ ".inst 0x450098ed // smmla z13.s, z7.b, z0.b\n"
+ ".inst 0x450098d5 // smmla z21.s, z6.b, z0.b\n"
+ ".inst 0x4500989d // smmla z29.s, z4.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x450298ea // smmla z10.s, z7.b, z2.b\n"
+ ".inst 0x450298d2 // smmla z18.s, z6.b, z2.b\n"
+ ".inst 0x4502989a // smmla z26.s, z4.b, z2.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x450098ee // smmla z14.s, z7.b, z0.b\n"
+ ".inst 0x450098d6 // smmla z22.s, z6.b, z0.b\n"
+ ".inst 0x4500989e // smmla z30.s, z4.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+ ".inst 0x450298eb // smmla z11.s, z7.b, z2.b\n"
+ "addvl x9, x9, #8\n"
+ ".inst 0x450298d3 // smmla z19.s, z6.b, z2.b\n"
+ ".inst 0x4502989b // smmla z27.s, z4.b, z2.b\n"
+ ".inst 0x450098ef // smmla z15.s, z7.b, z0.b\n"
+ ".inst 0x450098d7 // smmla z23.s, z6.b, z0.b\n"
+ ".inst 0x4500989f // smmla z31.s, z4.b, z0.b\n"
+ "ble 74f\n"
+ "ld1b { z2.b }, p2/Z, [x9]\n"
+ "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
+ ".inst 0x45029828 // smmla z8.s, z1.b, z2.b\n"
+ ".inst 0x45029870 // smmla z16.s, z3.b, z2.b\n"
+ ".inst 0x450298b8 // smmla z24.s, z5.b, z2.b\n"
+ ".inst 0x4500982c // smmla z12.s, z1.b, z0.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #2, MUL VL]\n"
+ ".inst 0x45009874 // smmla z20.s, z3.b, z0.b\n"
+ ".inst 0x450098bc // smmla z28.s, z5.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
+ ".inst 0x45029829 // smmla z9.s, z1.b, z2.b\n"
+ ".inst 0x45029871 // smmla z17.s, z3.b, z2.b\n"
+ ".inst 0x450298b9 // smmla z25.s, z5.b, z2.b\n"
+ ".inst 0x4500982d // smmla z13.s, z1.b, z0.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #4, MUL VL]\n"
+ ".inst 0x45009875 // smmla z21.s, z3.b, z0.b\n"
+ ".inst 0x450098bd // smmla z29.s, z5.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
+ ".inst 0x4502982a // smmla z10.s, z1.b, z2.b\n"
+ ".inst 0x45029872 // smmla z18.s, z3.b, z2.b\n"
+ ".inst 0x450298ba // smmla z26.s, z5.b, z2.b\n"
+ ".inst 0x4500982e // smmla z14.s, z1.b, z0.b\n"
+ "ld1b { z2.b }, p2/Z, [x9, #6, MUL VL]\n"
+ ".inst 0x45009876 // smmla z22.s, z3.b, z0.b\n"
+ ".inst 0x450098be // smmla z30.s, z5.b, z0.b\n"
+ "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+ "addvl x9, x9, #8\n"
+ ".inst 0x4502982b // smmla z11.s, z1.b, z2.b\n"
+ ".inst 0x45029873 // smmla z19.s, z3.b, z2.b\n"
+ ".inst 0x450298bb // smmla z27.s, z5.b, z2.b\n"
+ ".inst 0x4500982f // smmla z15.s, z1.b, z0.b\n"
+ ".inst 0x45009877 // smmla z23.s, z3.b, z0.b\n"
+ ".inst 0x450098bf // smmla z31.s, z5.b, z0.b\n"
+ "74:" // Height 6: Multiply loop: multiply skip
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x28, x28, #0x1\n"
+ "cmp x28, x20\n"
+ "bne 69b\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "uzp1 z4.d, z8.d, z12.d\n"
+ "add x26, x11, x20\n"
+ "uzp2 z8.d, z8.d, z12.d\n"
+ "uzp1 z12.d, z9.d, z13.d\n"
+ "uzp2 z9.d, z9.d, z13.d\n"
+ "add x25, x26, x20\n"
+ "ld1w { z3.s }, p2/Z, [x14]\n"
+ "uzp1 z13.d, z10.d, z14.d\n"
+ "uzp2 z10.d, z10.d, z14.d\n"
+ "ld1w { z2.s }, p2/Z, [x14, #1, MUL VL]\n"
+ "ld1w { z1.s }, p2/Z, [x14, #2, MUL VL]\n"
+ "uzp1 z14.d, z11.d, z15.d\n"
+ "uzp2 z11.d, z11.d, z15.d\n"
+ "ld1w { z0.s }, p2/Z, [x14, #3, MUL VL]\n"
+ "add x24, x25, x20\n"
+ "uzp1 z15.d, z16.d, z20.d\n"
+ "uzp2 z16.d, z16.d, z20.d\n"
+ "add x23, x24, x20\n"
+ "add x22, x23, x20\n"
+ "uzp1 z20.d, z17.d, z21.d\n"
+ "uzp2 z17.d, z17.d, z21.d\n"
+ "addvl x14, x14, #4\n"
+ "uzp1 z21.d, z18.d, z22.d\n"
+ "uzp2 z18.d, z18.d, z22.d\n"
+ "uzp1 z22.d, z19.d, z23.d\n"
+ "uzp2 z19.d, z19.d, z23.d\n"
+ "uzp1 z23.d, z24.d, z28.d\n"
+ "uzp2 z24.d, z24.d, z28.d\n"
+ "uzp1 z28.d, z25.d, z29.d\n"
+ "uzp2 z25.d, z25.d, z29.d\n"
+ "uzp1 z29.d, z26.d, z30.d\n"
+ "uzp2 z26.d, z26.d, z30.d\n"
+ "uzp1 z30.d, z27.d, z31.d\n"
+ "uzp2 z27.d, z27.d, z31.d\n"
+ "mov z31.d, z4.d\n"
+ "add z31.s, z31.s, z3.s\n"
+ "add z12.s, z12.s, z2.s\n"
+ "add z13.s, z13.s, z1.s\n"
+ "add z14.s, z14.s, z0.s\n"
+ "add z8.s, z8.s, z3.s\n"
+ "add z9.s, z9.s, z2.s\n"
+ "add z10.s, z10.s, z1.s\n"
+ "add z11.s, z11.s, z0.s\n"
+ "add z15.s, z15.s, z3.s\n"
+ "add z20.s, z20.s, z2.s\n"
+ "add z21.s, z21.s, z1.s\n"
+ "add z22.s, z22.s, z0.s\n"
+ "add z16.s, z16.s, z3.s\n"
+ "add z17.s, z17.s, z2.s\n"
+ "add z18.s, z18.s, z1.s\n"
+ "add z19.s, z19.s, z0.s\n"
+ "add z23.s, z23.s, z3.s\n"
+ "add z28.s, z28.s, z2.s\n"
+ "add z29.s, z29.s, z1.s\n"
+ "add z30.s, z30.s, z0.s\n"
+ "add z24.s, z24.s, z3.s\n"
+ "add z25.s, z25.s, z2.s\n"
+ "add z26.s, z26.s, z1.s\n"
+ "add z27.s, z27.s, z0.s\n"
+ "tbz %x[flags], #4, 75f\n"
+ "ld1w { z0.s }, p2/Z, [x12]\n"
+ "ld1w { z4.s }, p2/Z, [x13]\n"
+ "ld1w { z1.s }, p2/Z, [x12, #1, MUL VL]\n"
+ "ld1w { z5.s }, p2/Z, [x13, #1, MUL VL]\n"
+ "ld1w { z2.s }, p2/Z, [x12, #2, MUL VL]\n"
+ "ld1w { z6.s }, p2/Z, [x13, #2, MUL VL]\n"
+ "ld1w { z3.s }, p2/Z, [x12, #3, MUL VL]\n"
+ "ld1w { z7.s }, p2/Z, [x13, #3, MUL VL]\n"
+ "addvl x12, x12, #4\n"
+ "addvl x13, x13, #4\n"
+ "b 76f\n"
+ "75:" // Height 6: per layer parameters
+ "add x21, %x[qp], %[per_layer_right_shift]\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1rw { z0.s }, p2/Z, [x21]\n"
+ "ld1rw { z4.s }, p2/Z, [x20]\n"
+ "mov z1.d, z0.d\n"
+ "mov z5.d, z4.d\n"
+ "mov z2.d, z0.d\n"
+ "mov z6.d, z4.d\n"
+ "mov z3.d, z0.d\n"
+ "mov z7.d, z4.d\n"
+ "76:" // Height 6: parameters loaded
+ ".inst 0x04a477ff // sqrdmulh z31.s, z31.s, z4.s\n"
+ ".inst 0x04a5758c // sqrdmulh z12.s, z12.s, z5.s\n"
+ ".inst 0x04a675ad // sqrdmulh z13.s, z13.s, z6.s\n"
+ ".inst 0x04a775ce // sqrdmulh z14.s, z14.s, z7.s\n"
+ ".inst 0x04a47508 // sqrdmulh z8.s, z8.s, z4.s\n"
+ ".inst 0x04a57529 // sqrdmulh z9.s, z9.s, z5.s\n"
+ ".inst 0x04a6754a // sqrdmulh z10.s, z10.s, z6.s\n"
+ ".inst 0x04a7756b // sqrdmulh z11.s, z11.s, z7.s\n"
+ ".inst 0x04a475ef // sqrdmulh z15.s, z15.s, z4.s\n"
+ ".inst 0x04a57694 // sqrdmulh z20.s, z20.s, z5.s\n"
+ ".inst 0x04a676b5 // sqrdmulh z21.s, z21.s, z6.s\n"
+ ".inst 0x04a776d6 // sqrdmulh z22.s, z22.s, z7.s\n"
+ ".inst 0x04a47610 // sqrdmulh z16.s, z16.s, z4.s\n"
+ ".inst 0x04a57631 // sqrdmulh z17.s, z17.s, z5.s\n"
+ ".inst 0x04a67652 // sqrdmulh z18.s, z18.s, z6.s\n"
+ ".inst 0x04a77673 // sqrdmulh z19.s, z19.s, z7.s\n"
+ ".inst 0x04a476f7 // sqrdmulh z23.s, z23.s, z4.s\n"
+ ".inst 0x04a5779c // sqrdmulh z28.s, z28.s, z5.s\n"
+ ".inst 0x04a677bd // sqrdmulh z29.s, z29.s, z6.s\n"
+ ".inst 0x04a777de // sqrdmulh z30.s, z30.s, z7.s\n"
+ ".inst 0x04a47718 // sqrdmulh z24.s, z24.s, z4.s\n"
+ ".inst 0x04a57739 // sqrdmulh z25.s, z25.s, z5.s\n"
+ ".inst 0x04a6775a // sqrdmulh z26.s, z26.s, z6.s\n"
+ ".inst 0x04a7777b // sqrdmulh z27.s, z27.s, z7.s\n"
+ "tbz %x[flags], #5, 77f\n"
+ "and z7.d, z31.d, z0.d\n"
+ "and z6.d, z12.d, z1.d\n"
+ "and z5.d, z13.d, z2.d\n"
+ "and z4.d, z14.d, z3.d\n"
+ "asr z7.s, z7.s, #0x1f\n"
+ "asr z6.s, z6.s, #0x1f\n"
+ "asr z5.s, z5.s, #0x1f\n"
+ "asr z4.s, z4.s, #0x1f\n"
+ "sqadd z31.s, z31.s, z7.s\n"
+ "sqadd z12.s, z12.s, z6.s\n"
+ "sqadd z13.s, z13.s, z5.s\n"
+ "sqadd z14.s, z14.s, z4.s\n"
+ "and z7.d, z8.d, z0.d\n"
+ "and z6.d, z9.d, z1.d\n"
+ "and z5.d, z10.d, z2.d\n"
+ "and z4.d, z11.d, z3.d\n"
+ "asr z7.s, z7.s, #0x1f\n"
+ "asr z6.s, z6.s, #0x1f\n"
+ "asr z5.s, z5.s, #0x1f\n"
+ "asr z4.s, z4.s, #0x1f\n"
+ "sqadd z8.s, z8.s, z7.s\n"
+ "sqadd z9.s, z9.s, z6.s\n"
+ "sqadd z10.s, z10.s, z5.s\n"
+ "sqadd z11.s, z11.s, z4.s\n"
+ "and z7.d, z15.d, z0.d\n"
+ "and z6.d, z20.d, z1.d\n"
+ "and z5.d, z21.d, z2.d\n"
+ "and z4.d, z22.d, z3.d\n"
+ "asr z7.s, z7.s, #0x1f\n"
+ "asr z6.s, z6.s, #0x1f\n"
+ "asr z5.s, z5.s, #0x1f\n"
+ "asr z4.s, z4.s, #0x1f\n"
+ "sqadd z15.s, z15.s, z7.s\n"
+ "sqadd z20.s, z20.s, z6.s\n"
+ "sqadd z21.s, z21.s, z5.s\n"
+ "sqadd z22.s, z22.s, z4.s\n"
+ "and z7.d, z16.d, z0.d\n"
+ "and z6.d, z17.d, z1.d\n"
+ "and z5.d, z18.d, z2.d\n"
+ "and z4.d, z19.d, z3.d\n"
+ "asr z7.s, z7.s, #0x1f\n"
+ "asr z6.s, z6.s, #0x1f\n"
+ "asr z5.s, z5.s, #0x1f\n"
+ "asr z4.s, z4.s, #0x1f\n"
+ "sqadd z16.s, z16.s, z7.s\n"
+ "sqadd z17.s, z17.s, z6.s\n"
+ "sqadd z18.s, z18.s, z5.s\n"
+ "sqadd z19.s, z19.s, z4.s\n"
+ "and z7.d, z23.d, z0.d\n"
+ "and z6.d, z28.d, z1.d\n"
+ "and z5.d, z29.d, z2.d\n"
+ "and z4.d, z30.d, z3.d\n"
+ "asr z7.s, z7.s, #0x1f\n"
+ "asr z6.s, z6.s, #0x1f\n"
+ "asr z5.s, z5.s, #0x1f\n"
+ "asr z4.s, z4.s, #0x1f\n"
+ "sqadd z23.s, z23.s, z7.s\n"
+ "sqadd z28.s, z28.s, z6.s\n"
+ "sqadd z29.s, z29.s, z5.s\n"
+ "sqadd z30.s, z30.s, z4.s\n"
+ "and z7.d, z24.d, z0.d\n"
+ "and z6.d, z25.d, z1.d\n"
+ "and z5.d, z26.d, z2.d\n"
+ "and z4.d, z27.d, z3.d\n"
+ "asr z7.s, z7.s, #0x1f\n"
+ "asr z6.s, z6.s, #0x1f\n"
+ "asr z5.s, z5.s, #0x1f\n"
+ "asr z4.s, z4.s, #0x1f\n"
+ "sqadd z24.s, z24.s, z7.s\n"
+ "sqadd z25.s, z25.s, z6.s\n"
+ "sqadd z26.s, z26.s, z5.s\n"
+ "sqadd z27.s, z27.s, z4.s\n"
+ "77:" // Height 6: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1rw { z4.s }, p2/Z, [x20]\n"
+ ".inst 0x4482881f // srshl z31.s, p2/M, z31.s, z0.s\n"
+ "add z31.s, z31.s, z4.s\n"
+ ".inst 0x4482882c // srshl z12.s, p2/M, z12.s, z1.s\n"
+ ".inst 0x4482884d // srshl z13.s, p2/M, z13.s, z2.s\n"
+ "add z12.s, z12.s, z4.s\n"
+ "add z13.s, z13.s, z4.s\n"
+ ".inst 0x4482886e // srshl z14.s, p2/M, z14.s, z3.s\n"
+ ".inst 0x44828808 // srshl z8.s, p2/M, z8.s, z0.s\n"
+ "add z14.s, z14.s, z4.s\n"
+ "add z8.s, z8.s, z4.s\n"
+ ".inst 0x44828829 // srshl z9.s, p2/M, z9.s, z1.s\n"
+ ".inst 0x4482884a // srshl z10.s, p2/M, z10.s, z2.s\n"
+ "add z9.s, z9.s, z4.s\n"
+ "add z10.s, z10.s, z4.s\n"
+ ".inst 0x4482886b // srshl z11.s, p2/M, z11.s, z3.s\n"
+ ".inst 0x4482880f // srshl z15.s, p2/M, z15.s, z0.s\n"
+ "add z11.s, z11.s, z4.s\n"
+ "add z15.s, z15.s, z4.s\n"
+ ".inst 0x44828834 // srshl z20.s, p2/M, z20.s, z1.s\n"
+ ".inst 0x44828855 // srshl z21.s, p2/M, z21.s, z2.s\n"
+ "add z20.s, z20.s, z4.s\n"
+ "add z21.s, z21.s, z4.s\n"
+ ".inst 0x44828876 // srshl z22.s, p2/M, z22.s, z3.s\n"
+ ".inst 0x44828810 // srshl z16.s, p2/M, z16.s, z0.s\n"
+ "add z22.s, z22.s, z4.s\n"
+ "add z16.s, z16.s, z4.s\n"
+ ".inst 0x44828831 // srshl z17.s, p2/M, z17.s, z1.s\n"
+ ".inst 0x44828852 // srshl z18.s, p2/M, z18.s, z2.s\n"
+ "add z17.s, z17.s, z4.s\n"
+ "add z18.s, z18.s, z4.s\n"
+ ".inst 0x44828873 // srshl z19.s, p2/M, z19.s, z3.s\n"
+ ".inst 0x44828817 // srshl z23.s, p2/M, z23.s, z0.s\n"
+ "add z19.s, z19.s, z4.s\n"
+ "add z23.s, z23.s, z4.s\n"
+ ".inst 0x4482883c // srshl z28.s, p2/M, z28.s, z1.s\n"
+ ".inst 0x4482885d // srshl z29.s, p2/M, z29.s, z2.s\n"
+ "add z28.s, z28.s, z4.s\n"
+ "add z29.s, z29.s, z4.s\n"
+ ".inst 0x4482887e // srshl z30.s, p2/M, z30.s, z3.s\n"
+ ".inst 0x44828818 // srshl z24.s, p2/M, z24.s, z0.s\n"
+ "add z30.s, z30.s, z4.s\n"
+ "add z24.s, z24.s, z4.s\n"
+ ".inst 0x44828839 // srshl z25.s, p2/M, z25.s, z1.s\n"
+ ".inst 0x4482885a // srshl z26.s, p2/M, z26.s, z2.s\n"
+ "add z25.s, z25.s, z4.s\n"
+ "add z26.s, z26.s, z4.s\n"
+ ".inst 0x4482887b // srshl z27.s, p2/M, z27.s, z3.s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1rw { z0.s }, p2/Z, [x20]\n"
+ "add z27.s, z27.s, z4.s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1rw { z1.s }, p2/Z, [x20]\n"
+ "smin z31.s, p2/M, z31.s, z0.s\n"
+ "smin z12.s, p2/M, z12.s, z0.s\n"
+ "smin z13.s, p2/M, z13.s, z0.s\n"
+ "smin z14.s, p2/M, z14.s, z0.s\n"
+ "smin z8.s, p2/M, z8.s, z0.s\n"
+ "smin z9.s, p2/M, z9.s, z0.s\n"
+ "smin z10.s, p2/M, z10.s, z0.s\n"
+ "smin z11.s, p2/M, z11.s, z0.s\n"
+ "smin z15.s, p2/M, z15.s, z0.s\n"
+ "smin z20.s, p2/M, z20.s, z0.s\n"
+ "smin z21.s, p2/M, z21.s, z0.s\n"
+ "smin z22.s, p2/M, z22.s, z0.s\n"
+ "smin z16.s, p2/M, z16.s, z0.s\n"
+ "smin z17.s, p2/M, z17.s, z0.s\n"
+ "smin z18.s, p2/M, z18.s, z0.s\n"
+ "smin z19.s, p2/M, z19.s, z0.s\n"
+ "smin z23.s, p2/M, z23.s, z0.s\n"
+ "smin z28.s, p2/M, z28.s, z0.s\n"
+ "smin z29.s, p2/M, z29.s, z0.s\n"
+ "smin z30.s, p2/M, z30.s, z0.s\n"
+ "smin z24.s, p2/M, z24.s, z0.s\n"
+ "smin z25.s, p2/M, z25.s, z0.s\n"
+ "smin z26.s, p2/M, z26.s, z0.s\n"
+ "smin z27.s, p2/M, z27.s, z0.s\n"
+ "smax z31.s, p2/M, z31.s, z1.s\n"
+ "smax z12.s, p2/M, z12.s, z1.s\n"
+ "smax z13.s, p2/M, z13.s, z1.s\n"
+ "uzp1 z31.h, z31.h, z12.h\n"
+ "smax z14.s, p2/M, z14.s, z1.s\n"
+ "smax z8.s, p2/M, z8.s, z1.s\n"
+ "uzp1 z0.h, z13.h, z14.h\n"
+ "uzp1 z31.b, z31.b, z0.b\n"
+ "smax z9.s, p2/M, z9.s, z1.s\n"
+ "smax z10.s, p2/M, z10.s, z1.s\n"
+ "uzp1 z8.h, z8.h, z9.h\n"
+ "st1b { z31.b }, p1, [x11]\n"
+ "smax z11.s, p2/M, z11.s, z1.s\n"
+ "smax z15.s, p2/M, z15.s, z1.s\n"
+ "uzp1 z31.h, z10.h, z11.h\n"
+ "uzp1 z8.b, z8.b, z31.b\n"
+ "smax z20.s, p2/M, z20.s, z1.s\n"
+ "smax z21.s, p2/M, z21.s, z1.s\n"
+ "uzp1 z15.h, z15.h, z20.h\n"
+ "st1b { z8.b }, p1, [x26]\n"
+ "smax z22.s, p2/M, z22.s, z1.s\n"
+ "smax z16.s, p2/M, z16.s, z1.s\n"
+ "uzp1 z20.h, z21.h, z22.h\n"
+ "uzp1 z15.b, z15.b, z20.b\n"
+ "smax z17.s, p2/M, z17.s, z1.s\n"
+ "smax z18.s, p2/M, z18.s, z1.s\n"
+ "uzp1 z16.h, z16.h, z17.h\n"
+ "st1b { z15.b }, p1, [x25]\n"
+ "smax z19.s, p2/M, z19.s, z1.s\n"
+ "smax z23.s, p2/M, z23.s, z1.s\n"
+ "uzp1 z17.h, z18.h, z19.h\n"
+ "uzp1 z16.b, z16.b, z17.b\n"
+ "smax z28.s, p2/M, z28.s, z1.s\n"
+ "smax z29.s, p2/M, z29.s, z1.s\n"
+ "uzp1 z23.h, z23.h, z28.h\n"
+ "st1b { z16.b }, p1, [x24]\n"
+ "smax z30.s, p2/M, z30.s, z1.s\n"
+ "smax z24.s, p2/M, z24.s, z1.s\n"
+ "uzp1 z16.h, z29.h, z30.h\n"
+ "uzp1 z23.b, z23.b, z16.b\n"
+ "smax z25.s, p2/M, z25.s, z1.s\n"
+ "smax z26.s, p2/M, z26.s, z1.s\n"
+ "uzp1 z24.h, z24.h, z25.h\n"
+ "st1b { z23.b }, p1, [x23]\n"
+ "smax z27.s, p2/M, z27.s, z1.s\n"
+ "uzp1 z16.h, z26.h, z27.h\n"
+ "uzp1 z24.b, z24.b, z16.b\n"
+ "st1b { z24.b }, p1, [x22]\n"
+ "addvl x11, x11, #1\n"
+ "78:" // Height 6: Writeback done
+ "decw x10, ALL, MUL #4\n"
+ "cmp x10, XZR\n"
+ "bgt 67b\n"
+ "subs %x[M], %x[M], #0x6\n"
+ "beq 80f\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 79f\n"
+ "add x21, x21, #0x6\n"
+ "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "b 1b\n"
+ "79:" // Update direct input
+ "mov x20, #0x6\n"
+ "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
+ "b 1b\n"
+ "80:" // Exit
+ : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+ : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+ : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x13", "x14", "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"
+ );
+}
+
+} // namespace arm_gemm
+#endif // ARM_COMPUTE_ENABLE_SVE