From c0b6f76561580414f08633a804fc548ccad65659 Mon Sep 17 00:00:00 2001
From: Georgios Pinitas <georgios.pinitas@arm.com>
Date: Mon, 2 Nov 2020 01:37:17 +0000
Subject: COMPMID-3776: Indirect GEMM

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I51a1b0f098bc3a8c408c50c92221e4df3061e12c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4343
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
---
 .../kernels/a64_hybrid_fp32_mla_6x16/generic.cpp   | 3430 ++++++++++++++++++++
 1 file changed, 3430 insertions(+)
 create mode 100644 src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp

(limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp')

diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp
new file mode 100644
index 0000000000..884e8986c8
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp
@@ -0,0 +1,3430 @@
+/*
+ * Copyright (c) 2019-2020 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 __aarch64__
+
+#include "arm_gemm.hpp"
+#include "../../utils.hpp"
+
+#include <cassert>
+
+namespace arm_gemm {
+
+void a64_hybrid_fp32_mla_6x16 (
+    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<float> A_arg,
+    size_t M, size_t N, const float *B_ptr, IndirectOutputArg<float> output_arg,
+    const float *bias, Activation act, bool accumulate
+)
+{
+    struct KernelArgs {
+        float maxval = static_cast<float>(std::numeric_limits<float>::infinity());
+        float minval = - static_cast<float>(std::numeric_limits<float>::infinity());
+        unsigned int num_strings = {};
+        const unsigned int *string_lengths = {};
+        size_t N = {};
+        const float *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;
+    }
+    if (accumulate) {
+        flags |= 0x1;
+    }
+    ka.num_strings = num_strings;
+    ka.string_lengths = string_lengths;
+    ka.N = N;
+    ka.B_ptr = B_ptr;
+    switch(act.type) {
+        default:
+        case Activation::Type::None:
+            break;
+        case Activation::Type::BoundedReLU:
+            ka.maxval = static_cast<float>(act.param1);
+            /* fall through */
+        case Activation::Type::ReLU:
+            ka.minval = 0;
+            flags |= 0x2;
+            break;
+    }
+    __asm__ __volatile__(
+
+      "1:"  // Row loop
+      "cmp %x[M], #0x6\n"
+      "bge 171f\n"
+      "cmp %x[M], #0x4\n"
+      "bgt 137f\n"
+      "beq 103f\n"
+      "cmp %x[M], #0x2\n"
+      "bgt 69f\n"
+      "beq 35f\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x14, %x[bias]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "tbz %x[flags], #2, 2f\n"
+      "ldr x13, [%x[output_ptr], #0x0]\n"
+      "add x13, x13, x19, LSL #2\n"
+      "b 3f\n"
+      "2:"  // Height 1: setup direct output
+      "mov x13, %x[output_ptr]\n"
+      "3:"  // Height 1: Column loop
+      "cbz x14, 4f\n"
+      "ldr q8, [x14, #0x0]\n"
+      "ldr q9, [x14, #0x10]\n"
+      "ldr q10, [x14, #0x20]\n"
+      "ldr q11, [x14, #0x30]\n"
+      "add x14, x14, #0x40\n"
+      "b 15f\n"
+      "4:"  // Height 1: no bias
+      "tbz %x[flags], #0, 14f\n"
+      "cmp x16, #0x10\n"
+      "bge 13f\n"
+      "tbz x16, #3, 8f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "tbz x16, #2, 6f\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "tbz x16, #1, 5f\n"
+      "mov x19, #0x38\n"
+      "ldr d11, [x13], #0x8\n"
+      "tbz x16, #0, 12f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "b 12f\n"
+      "5:"  // Height 1: Partial accumulate: partial_1_12
+      "mov x19, #0x30\n"
+      "tbz x16, #0, 12f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "b 12f\n"
+      "6:"  // Height 1: Partial accumulate: partial_2_8
+      "tbz x16, #1, 7f\n"
+      "ldr d10, [x13], #0x8\n"
+      "mov x19, #0x28\n"
+      "tbz x16, #0, 12f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "b 12f\n"
+      "7:"  // Height 1: Partial accumulate: partial_1_8
+      "mov x19, #0x20\n"
+      "tbz x16, #0, 12f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "b 12f\n"
+      "8:"  // Height 1: Partial accumulate: partial_4_0
+      "tbz x16, #2, 10f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "tbz x16, #1, 9f\n"
+      "mov x19, #0x18\n"
+      "ldr d9, [x13], #0x8\n"
+      "tbz x16, #0, 12f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "b 12f\n"
+      "9:"  // Height 1: Partial accumulate: partial_1_4
+      "mov x19, #0x10\n"
+      "tbz x16, #0, 12f\n"
+      "ldr s9, [x13, #0x0]\n"
+      "b 12f\n"
+      "10:"  // Height 1: Partial accumulate: partial_2_0
+      "tbz x16, #1, 11f\n"
+      "ldr d8, [x13], #0x8\n"
+      "mov x19, #0x8\n"
+      "tbz x16, #0, 12f\n"
+      "ld1 { v8.s }[2], [x13]\n"
+      "b 12f\n"
+      "11:"  // Height 1: Partial accumulate: partial_1_0
+      "mov x19, #0x0\n"
+      "ldr s8, [x13, #0x0]\n"
+      "12:"  // Height 1: Partial accumulate: Done
+      "sub x13, x13, x19\n"
+      "b 15f\n"
+      "13:"  // Height 1: full accumulate
+      "ldr q8, [x13, #0x0]\n"
+      "ldr q9, [x13, #0x10]\n"
+      "ldr q10, [x13, #0x20]\n"
+      "ldr q11, [x13, #0x30]\n"
+      "b 15f\n"
+      "14:"  // Height 1: no accumulate
+      "movi v8.16b, #0x0\n"
+      "movi v9.16b, #0x0\n"
+      "movi v10.16b, #0x0\n"
+      "movi v11.16b, #0x0\n"
+      "15:"  // Height 1: setup done
+      "mov x12, #0x0\n"
+      "16:"  // Height 1: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w11, [x20, x12, LSL #0x2]\n"
+      "tbz %x[flags], #3, 17f\n"
+      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "add x20, x20, x19, LSL #3\n"
+      "ldr x10, [x20, #0x0]\n"
+      "cbnz x12, 18f\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x10, x10, x19, LSL #2\n"
+      "b 18f\n"
+      "17:"  // Height 1: setup direct input
+      "mov x10, %x[input_ptr]\n"
+      "18:"  // Height 1: input setup done
+      "cmp x11, #0x4\n"
+      "blt 21f\n"
+      "cmp x11, #0x8\n"
+      "blt 20f\n"
+      "19:"  // Height 1: Multiply loop: Main loop head
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "add x10, x10, #0x10\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "sub x11, x11, #0x4\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "cmp x11, #0x8\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "bge 19b\n"
+      "20:"  // Height 1: Multiply loop: Single iteration only
+      "sub x11, x11, #0x4\n"
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "add x10, x10, #0x10\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "21:"  // Height 1: Multiply loop: Main loop skip
+      "cbz x11, 23f\n"
+      "22:"  // Height 1: Multiply loop: Odd block loop
+      "ldr s0, [x10], #0x4\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "sub x11, x11, #0x1\n"
+      "add x15, x15, #0x40\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "cbnz x11, 22b\n"
+      "23:"  // Height 1: Multiply loop: No odd multiplies
+      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x19\n"
+      "bne 16b\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "tbz %x[flags], #1, 24f\n"
+      "add x19, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v1.4s }, [x19]\n"
+      "add x19, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v0.4s }, [x19]\n"
+      "fmin v8.4s, v8.4s, v0.4s\n"
+      "fmin v9.4s, v9.4s, v0.4s\n"
+      "fmin v10.4s, v10.4s, v0.4s\n"
+      "fmin v11.4s, v11.4s, v0.4s\n"
+      "fmax v8.4s, v8.4s, v1.4s\n"
+      "fmax v9.4s, v9.4s, v1.4s\n"
+      "fmax v10.4s, v10.4s, v1.4s\n"
+      "fmax v11.4s, v11.4s, v1.4s\n"
+      "24:"  // Height 1: No activation
+      "cmp x16, #0x10\n"
+      "bge 33f\n"
+      "tbz x16, #3, 28f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v9.4s }, [x13], #0x10\n"
+      "tbz x16, #2, 26f\n"
+      "st1 { v10.4s }, [x13], #0x10\n"
+      "tbz x16, #1, 25f\n"
+      "str d11, [x13], #0x8\n"
+      "tbz x16, #0, 32f\n"
+      "st1 { v11.s }[2], [x13]\n"
+      "b 32f\n"
+      "25:"  // Height 1: Partial direct writeback: partial_1_12
+      "tbz x16, #0, 32f\n"
+      "str s11, [x13, #0x0]\n"
+      "b 32f\n"
+      "26:"  // Height 1: Partial direct writeback: partial_2_8
+      "tbz x16, #1, 27f\n"
+      "str d10, [x13], #0x8\n"
+      "tbz x16, #0, 32f\n"
+      "st1 { v10.s }[2], [x13]\n"
+      "b 32f\n"
+      "27:"  // Height 1: Partial direct writeback: partial_1_8
+      "tbz x16, #0, 32f\n"
+      "str s10, [x13, #0x0]\n"
+      "b 32f\n"
+      "28:"  // Height 1: Partial direct writeback: partial_4_0
+      "tbz x16, #2, 30f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "tbz x16, #1, 29f\n"
+      "str d9, [x13], #0x8\n"
+      "tbz x16, #0, 32f\n"
+      "st1 { v9.s }[2], [x13]\n"
+      "b 32f\n"
+      "29:"  // Height 1: Partial direct writeback: partial_1_4
+      "tbz x16, #0, 32f\n"
+      "str s9, [x13, #0x0]\n"
+      "b 32f\n"
+      "30:"  // Height 1: Partial direct writeback: partial_2_0
+      "tbz x16, #1, 31f\n"
+      "str d8, [x13], #0x8\n"
+      "tbz x16, #0, 32f\n"
+      "st1 { v8.s }[2], [x13]\n"
+      "b 32f\n"
+      "31:"  // Height 1: Partial direct writeback: partial_1_0
+      "str s8, [x13, #0x0]\n"
+      "32:"  // Height 1: Partial direct writeback: Done
+      "b 34f\n"
+      "33:"  // Height 1: Full writeback
+      "str q8, [x13, #0x0]\n"
+      "str q9, [x13, #0x10]\n"
+      "str q10, [x13, #0x20]\n"
+      "str q11, [x13, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "34:"  // Height 1: Writeback done
+      "subs x16, x16, #0x10\n"
+      "bgt 3b\n"
+      "b 206f\n"
+      "35:"  // Height 2
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x14, %x[bias]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "tbz %x[flags], #2, 36f\n"
+      "ldr x13, [%x[output_ptr], #0x0]\n"
+      "add x13, x13, x19, LSL #2\n"
+      "ldr x9, [%x[output_ptr], #0x8]\n"
+      "add x9, x9, x19, LSL #2\n"
+      "b 37f\n"
+      "36:"  // Height 2: setup direct output
+      "mov x13, %x[output_ptr]\n"
+      "add x9, x13, x19, LSL #2\n"
+      "37:"  // Height 2: Column loop
+      "cbz x14, 38f\n"
+      "ldr q8, [x14, #0x0]\n"
+      "mov v12.16b, v8.16b\n"
+      "ldr q9, [x14, #0x10]\n"
+      "ldr q10, [x14, #0x20]\n"
+      "mov v13.16b, v9.16b\n"
+      "ldr q11, [x14, #0x30]\n"
+      "mov v14.16b, v10.16b\n"
+      "add x14, x14, #0x40\n"
+      "mov v15.16b, v11.16b\n"
+      "b 49f\n"
+      "38:"  // Height 2: no bias
+      "tbz %x[flags], #0, 48f\n"
+      "cmp x16, #0x10\n"
+      "bge 47f\n"
+      "tbz x16, #3, 42f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x9], #0x10\n"
+      "tbz x16, #2, 40f\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x9], #0x10\n"
+      "tbz x16, #1, 39f\n"
+      "mov x19, #0x38\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d15, [x9], #0x8\n"
+      "tbz x16, #0, 46f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x9]\n"
+      "b 46f\n"
+      "39:"  // Height 2: Partial accumulate: partial_1_12
+      "mov x19, #0x30\n"
+      "tbz x16, #0, 46f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s15, [x9, #0x0]\n"
+      "b 46f\n"
+      "40:"  // Height 2: Partial accumulate: partial_2_8
+      "tbz x16, #1, 41f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d14, [x9], #0x8\n"
+      "mov x19, #0x28\n"
+      "tbz x16, #0, 46f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x9]\n"
+      "b 46f\n"
+      "41:"  // Height 2: Partial accumulate: partial_1_8
+      "mov x19, #0x20\n"
+      "tbz x16, #0, 46f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s14, [x9, #0x0]\n"
+      "b 46f\n"
+      "42:"  // Height 2: Partial accumulate: partial_4_0
+      "tbz x16, #2, 44f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "tbz x16, #1, 43f\n"
+      "mov x19, #0x18\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d13, [x9], #0x8\n"
+      "tbz x16, #0, 46f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x9]\n"
+      "b 46f\n"
+      "43:"  // Height 2: Partial accumulate: partial_1_4
+      "mov x19, #0x10\n"
+      "tbz x16, #0, 46f\n"
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s13, [x9, #0x0]\n"
+      "b 46f\n"
+      "44:"  // Height 2: Partial accumulate: partial_2_0
+      "tbz x16, #1, 45f\n"
+      "ldr d8, [x13], #0x8\n"
+      "ldr d12, [x9], #0x8\n"
+      "mov x19, #0x8\n"
+      "tbz x16, #0, 46f\n"
+      "ld1 { v8.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x9]\n"
+      "b 46f\n"
+      "45:"  // Height 2: Partial accumulate: partial_1_0
+      "mov x19, #0x0\n"
+      "ldr s8, [x13, #0x0]\n"
+      "ldr s12, [x9, #0x0]\n"
+      "46:"  // Height 2: Partial accumulate: Done
+      "sub x13, x13, x19\n"
+      "sub x9, x9, x19\n"
+      "b 49f\n"
+      "47:"  // Height 2: full accumulate
+      "ldr q8, [x13, #0x0]\n"
+      "ldr q9, [x13, #0x10]\n"
+      "ldr q10, [x13, #0x20]\n"
+      "ldr q11, [x13, #0x30]\n"
+      "ldr q12, [x9, #0x0]\n"
+      "ldr q13, [x9, #0x10]\n"
+      "ldr q14, [x9, #0x20]\n"
+      "ldr q15, [x9, #0x30]\n"
+      "b 49f\n"
+      "48:"  // Height 2: no accumulate
+      "movi v8.16b, #0x0\n"
+      "movi v9.16b, #0x0\n"
+      "movi v10.16b, #0x0\n"
+      "movi v11.16b, #0x0\n"
+      "movi v12.16b, #0x0\n"
+      "movi v13.16b, #0x0\n"
+      "movi v14.16b, #0x0\n"
+      "movi v15.16b, #0x0\n"
+      "49:"  // Height 2: setup done
+      "mov x12, #0x0\n"
+      "50:"  // Height 2: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w11, [x20, x12, LSL #0x2]\n"
+      "tbz %x[flags], #3, 51f\n"
+      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "add x20, x20, x19, LSL #3\n"
+      "ldr x10, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "cbnz x12, 52f\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x10, x10, x19, LSL #2\n"
+      "add x28, x28, x19, LSL #2\n"
+      "b 52f\n"
+      "51:"  // Height 2: setup direct input
+      "mov x10, %x[input_ptr]\n"
+      "add x28, x10, x19, LSL #2\n"
+      "52:"  // Height 2: input setup done
+      "cmp x11, #0x4\n"
+      "blt 55f\n"
+      "cmp x11, #0x8\n"
+      "blt 54f\n"
+      "53:"  // Height 2: Multiply loop: Main loop head
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "add x10, x10, #0x10\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "add x28, x28, #0x10\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "sub x11, x11, #0x4\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "cmp x11, #0x8\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "bge 53b\n"
+      "54:"  // Height 2: Multiply loop: Single iteration only
+      "sub x11, x11, #0x4\n"
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "add x10, x10, #0x10\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "add x28, x28, #0x10\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "55:"  // Height 2: Multiply loop: Main loop skip
+      "cbz x11, 57f\n"
+      "56:"  // Height 2: Multiply loop: Odd block loop
+      "ldr s0, [x10], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "sub x11, x11, #0x1\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "add x15, x15, #0x40\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "cbnz x11, 56b\n"
+      "57:"  // Height 2: Multiply loop: No odd multiplies
+      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x19\n"
+      "bne 50b\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "tbz %x[flags], #1, 58f\n"
+      "add x19, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v1.4s }, [x19]\n"
+      "add x19, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v0.4s }, [x19]\n"
+      "fmin v8.4s, v8.4s, v0.4s\n"
+      "fmin v9.4s, v9.4s, v0.4s\n"
+      "fmin v10.4s, v10.4s, v0.4s\n"
+      "fmin v11.4s, v11.4s, v0.4s\n"
+      "fmax v8.4s, v8.4s, v1.4s\n"
+      "fmax v9.4s, v9.4s, v1.4s\n"
+      "fmax v10.4s, v10.4s, v1.4s\n"
+      "fmax v11.4s, v11.4s, v1.4s\n"
+      "fmin v12.4s, v12.4s, v0.4s\n"
+      "fmin v13.4s, v13.4s, v0.4s\n"
+      "fmin v14.4s, v14.4s, v0.4s\n"
+      "fmax v12.4s, v12.4s, v1.4s\n"
+      "fmax v13.4s, v13.4s, v1.4s\n"
+      "fmax v14.4s, v14.4s, v1.4s\n"
+      "fmin v15.4s, v15.4s, v0.4s\n"
+      "fmax v15.4s, v15.4s, v1.4s\n"
+      "58:"  // Height 2: No activation
+      "cmp x16, #0x10\n"
+      "bge 67f\n"
+      "tbz x16, #3, 62f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v9.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "st1 { v13.4s }, [x9], #0x10\n"
+      "tbz x16, #2, 60f\n"
+      "st1 { v10.4s }, [x13], #0x10\n"
+      "st1 { v14.4s }, [x9], #0x10\n"
+      "tbz x16, #1, 59f\n"
+      "str d11, [x13], #0x8\n"
+      "str d15, [x9], #0x8\n"
+      "tbz x16, #0, 66f\n"
+      "st1 { v11.s }[2], [x13]\n"
+      "st1 { v15.s }[2], [x9]\n"
+      "b 66f\n"
+      "59:"  // Height 2: Partial direct writeback: partial_1_12
+      "tbz x16, #0, 66f\n"
+      "str s11, [x13, #0x0]\n"
+      "str s15, [x9, #0x0]\n"
+      "b 66f\n"
+      "60:"  // Height 2: Partial direct writeback: partial_2_8
+      "tbz x16, #1, 61f\n"
+      "str d10, [x13], #0x8\n"
+      "str d14, [x9], #0x8\n"
+      "tbz x16, #0, 66f\n"
+      "st1 { v10.s }[2], [x13]\n"
+      "st1 { v14.s }[2], [x9]\n"
+      "b 66f\n"
+      "61:"  // Height 2: Partial direct writeback: partial_1_8
+      "tbz x16, #0, 66f\n"
+      "str s10, [x13, #0x0]\n"
+      "str s14, [x9, #0x0]\n"
+      "b 66f\n"
+      "62:"  // Height 2: Partial direct writeback: partial_4_0
+      "tbz x16, #2, 64f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "tbz x16, #1, 63f\n"
+      "str d9, [x13], #0x8\n"
+      "str d13, [x9], #0x8\n"
+      "tbz x16, #0, 66f\n"
+      "st1 { v9.s }[2], [x13]\n"
+      "st1 { v13.s }[2], [x9]\n"
+      "b 66f\n"
+      "63:"  // Height 2: Partial direct writeback: partial_1_4
+      "tbz x16, #0, 66f\n"
+      "str s9, [x13, #0x0]\n"
+      "str s13, [x9, #0x0]\n"
+      "b 66f\n"
+      "64:"  // Height 2: Partial direct writeback: partial_2_0
+      "tbz x16, #1, 65f\n"
+      "str d8, [x13], #0x8\n"
+      "str d12, [x9], #0x8\n"
+      "tbz x16, #0, 66f\n"
+      "st1 { v8.s }[2], [x13]\n"
+      "st1 { v12.s }[2], [x9]\n"
+      "b 66f\n"
+      "65:"  // Height 2: Partial direct writeback: partial_1_0
+      "str s8, [x13, #0x0]\n"
+      "str s12, [x9, #0x0]\n"
+      "66:"  // Height 2: Partial direct writeback: Done
+      "b 68f\n"
+      "67:"  // Height 2: Full writeback
+      "str q8, [x13, #0x0]\n"
+      "str q9, [x13, #0x10]\n"
+      "str q10, [x13, #0x20]\n"
+      "str q11, [x13, #0x30]\n"
+      "str q12, [x9, #0x0]\n"
+      "str q13, [x9, #0x10]\n"
+      "str q14, [x9, #0x20]\n"
+      "str q15, [x9, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "add x9, x9, #0x40\n"
+      "68:"  // Height 2: Writeback done
+      "subs x16, x16, #0x10\n"
+      "bgt 37b\n"
+      "b 206f\n"
+      "69:"  // Height 3
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x14, %x[bias]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "tbz %x[flags], #2, 70f\n"
+      "ldr x13, [%x[output_ptr], #0x0]\n"
+      "add x13, x13, x19, LSL #2\n"
+      "ldr x9, [%x[output_ptr], #0x8]\n"
+      "ldr x27, [%x[output_ptr], #0x10]\n"
+      "add x9, x9, x19, LSL #2\n"
+      "add x27, x27, x19, LSL #2\n"
+      "b 71f\n"
+      "70:"  // Height 3: setup direct output
+      "mov x13, %x[output_ptr]\n"
+      "add x9, x13, x19, LSL #2\n"
+      "add x27, x9, x19, LSL #2\n"
+      "71:"  // Height 3: Column loop
+      "cbz x14, 72f\n"
+      "ldr q8, [x14, #0x0]\n"
+      "mov v12.16b, v8.16b\n"
+      "ldr q9, [x14, #0x10]\n"
+      "mov v16.16b, v8.16b\n"
+      "ldr q10, [x14, #0x20]\n"
+      "ldr q11, [x14, #0x30]\n"
+      "mov v13.16b, v9.16b\n"
+      "add x14, x14, #0x40\n"
+      "mov v17.16b, v9.16b\n"
+      "mov v14.16b, v10.16b\n"
+      "mov v15.16b, v11.16b\n"
+      "mov v18.16b, v10.16b\n"
+      "mov v19.16b, v11.16b\n"
+      "b 83f\n"
+      "72:"  // Height 3: no bias
+      "tbz %x[flags], #0, 82f\n"
+      "cmp x16, #0x10\n"
+      "bge 81f\n"
+      "tbz x16, #3, 76f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "ld1 { v16.4s }, [x27], #0x10\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x9], #0x10\n"
+      "ld1 { v17.4s }, [x27], #0x10\n"
+      "tbz x16, #2, 74f\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x9], #0x10\n"
+      "ld1 { v18.4s }, [x27], #0x10\n"
+      "tbz x16, #1, 73f\n"
+      "mov x19, #0x38\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d15, [x9], #0x8\n"
+      "ldr d19, [x27], #0x8\n"
+      "tbz x16, #0, 80f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x9]\n"
+      "ld1 { v19.s }[2], [x27]\n"
+      "b 80f\n"
+      "73:"  // Height 3: Partial accumulate: partial_1_12
+      "mov x19, #0x30\n"
+      "tbz x16, #0, 80f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s15, [x9, #0x0]\n"
+      "ldr s19, [x27, #0x0]\n"
+      "b 80f\n"
+      "74:"  // Height 3: Partial accumulate: partial_2_8
+      "tbz x16, #1, 75f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d14, [x9], #0x8\n"
+      "ldr d18, [x27], #0x8\n"
+      "mov x19, #0x28\n"
+      "tbz x16, #0, 80f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x9]\n"
+      "ld1 { v18.s }[2], [x27]\n"
+      "b 80f\n"
+      "75:"  // Height 3: Partial accumulate: partial_1_8
+      "mov x19, #0x20\n"
+      "tbz x16, #0, 80f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s14, [x9, #0x0]\n"
+      "ldr s18, [x27, #0x0]\n"
+      "b 80f\n"
+      "76:"  // Height 3: Partial accumulate: partial_4_0
+      "tbz x16, #2, 78f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "ld1 { v16.4s }, [x27], #0x10\n"
+      "tbz x16, #1, 77f\n"
+      "mov x19, #0x18\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d13, [x9], #0x8\n"
+      "ldr d17, [x27], #0x8\n"
+      "tbz x16, #0, 80f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x9]\n"
+      "ld1 { v17.s }[2], [x27]\n"
+      "b 80f\n"
+      "77:"  // Height 3: Partial accumulate: partial_1_4
+      "mov x19, #0x10\n"
+      "tbz x16, #0, 80f\n"
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s13, [x9, #0x0]\n"
+      "ldr s17, [x27, #0x0]\n"
+      "b 80f\n"
+      "78:"  // Height 3: Partial accumulate: partial_2_0
+      "tbz x16, #1, 79f\n"
+      "ldr d8, [x13], #0x8\n"
+      "ldr d12, [x9], #0x8\n"
+      "ldr d16, [x27], #0x8\n"
+      "mov x19, #0x8\n"
+      "tbz x16, #0, 80f\n"
+      "ld1 { v8.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x9]\n"
+      "ld1 { v16.s }[2], [x27]\n"
+      "b 80f\n"
+      "79:"  // Height 3: Partial accumulate: partial_1_0
+      "mov x19, #0x0\n"
+      "ldr s8, [x13, #0x0]\n"
+      "ldr s12, [x9, #0x0]\n"
+      "ldr s16, [x27, #0x0]\n"
+      "80:"  // Height 3: Partial accumulate: Done
+      "sub x13, x13, x19\n"
+      "sub x9, x9, x19\n"
+      "sub x27, x27, x19\n"
+      "b 83f\n"
+      "81:"  // Height 3: full accumulate
+      "ldr q8, [x13, #0x0]\n"
+      "ldr q9, [x13, #0x10]\n"
+      "ldr q10, [x13, #0x20]\n"
+      "ldr q11, [x13, #0x30]\n"
+      "ldr q12, [x9, #0x0]\n"
+      "ldr q13, [x9, #0x10]\n"
+      "ldr q14, [x9, #0x20]\n"
+      "ldr q15, [x9, #0x30]\n"
+      "ldr q16, [x27, #0x0]\n"
+      "ldr q17, [x27, #0x10]\n"
+      "ldr q18, [x27, #0x20]\n"
+      "ldr q19, [x27, #0x30]\n"
+      "b 83f\n"
+      "82:"  // Height 3: no accumulate
+      "movi v8.16b, #0x0\n"
+      "movi v9.16b, #0x0\n"
+      "movi v10.16b, #0x0\n"
+      "movi v11.16b, #0x0\n"
+      "movi v12.16b, #0x0\n"
+      "movi v13.16b, #0x0\n"
+      "movi v14.16b, #0x0\n"
+      "movi v15.16b, #0x0\n"
+      "movi v16.16b, #0x0\n"
+      "movi v17.16b, #0x0\n"
+      "movi v18.16b, #0x0\n"
+      "movi v19.16b, #0x0\n"
+      "83:"  // Height 3: setup done
+      "mov x12, #0x0\n"
+      "84:"  // Height 3: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w11, [x20, x12, LSL #0x2]\n"
+      "tbz %x[flags], #3, 85f\n"
+      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "add x20, x20, x19, LSL #3\n"
+      "ldr x10, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "ldr x26, [x20, #0x10]\n"
+      "cbnz x12, 86f\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x10, x10, x19, LSL #2\n"
+      "add x28, x28, x19, LSL #2\n"
+      "add x26, x26, x19, LSL #2\n"
+      "b 86f\n"
+      "85:"  // Height 3: setup direct input
+      "mov x10, %x[input_ptr]\n"
+      "add x28, x10, x19, LSL #2\n"
+      "add x26, x28, x19, LSL #2\n"
+      "86:"  // Height 3: input setup done
+      "cmp x11, #0x4\n"
+      "blt 89f\n"
+      "cmp x11, #0x8\n"
+      "blt 88f\n"
+      "87:"  // Height 3: Multiply loop: Main loop head
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x26, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "add x10, x10, #0x10\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "add x28, x28, #0x10\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "add x26, x26, #0x10\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "sub x11, x11, #0x4\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "cmp x11, #0x8\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "fmla v16.4s, v6.4s, v2.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "fmla v17.4s, v7.4s, v2.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "fmla v18.4s, v6.4s, v2.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "fmla v19.4s, v7.4s, v2.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "fmla v16.4s, v6.4s, v2.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "fmla v17.4s, v7.4s, v2.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "fmla v18.4s, v6.4s, v2.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "fmla v19.4s, v7.4s, v2.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "fmla v16.4s, v6.4s, v2.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "fmla v17.4s, v7.4s, v2.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v18.4s, v6.4s, v2.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "fmla v19.4s, v7.4s, v2.s[3]\n"
+      "bge 87b\n"
+      "88:"  // Height 3: Multiply loop: Single iteration only
+      "sub x11, x11, #0x4\n"
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x26, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "add x10, x10, #0x10\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "add x28, x28, #0x10\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "add x26, x26, #0x10\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "fmla v16.4s, v6.4s, v2.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "fmla v17.4s, v7.4s, v2.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "fmla v18.4s, v6.4s, v2.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "fmla v19.4s, v7.4s, v2.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "fmla v16.4s, v6.4s, v2.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "fmla v17.4s, v7.4s, v2.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "fmla v18.4s, v6.4s, v2.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "fmla v19.4s, v7.4s, v2.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "fmla v16.4s, v6.4s, v2.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "fmla v17.4s, v7.4s, v2.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v18.4s, v6.4s, v2.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "fmla v19.4s, v7.4s, v2.s[3]\n"
+      "89:"  // Height 3: Multiply loop: Main loop skip
+      "cbz x11, 91f\n"
+      "90:"  // Height 3: Multiply loop: Odd block loop
+      "ldr s0, [x10], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
+      "ldr s2, [x26], #0x4\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "sub x11, x11, #0x1\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "add x15, x15, #0x40\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "cbnz x11, 90b\n"
+      "91:"  // Height 3: Multiply loop: No odd multiplies
+      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x19\n"
+      "bne 84b\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "tbz %x[flags], #1, 92f\n"
+      "add x19, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v1.4s }, [x19]\n"
+      "add x19, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v0.4s }, [x19]\n"
+      "fmin v8.4s, v8.4s, v0.4s\n"
+      "fmin v9.4s, v9.4s, v0.4s\n"
+      "fmin v10.4s, v10.4s, v0.4s\n"
+      "fmin v11.4s, v11.4s, v0.4s\n"
+      "fmax v8.4s, v8.4s, v1.4s\n"
+      "fmax v9.4s, v9.4s, v1.4s\n"
+      "fmax v10.4s, v10.4s, v1.4s\n"
+      "fmax v11.4s, v11.4s, v1.4s\n"
+      "fmin v12.4s, v12.4s, v0.4s\n"
+      "fmin v13.4s, v13.4s, v0.4s\n"
+      "fmin v14.4s, v14.4s, v0.4s\n"
+      "fmax v12.4s, v12.4s, v1.4s\n"
+      "fmax v13.4s, v13.4s, v1.4s\n"
+      "fmax v14.4s, v14.4s, v1.4s\n"
+      "fmin v15.4s, v15.4s, v0.4s\n"
+      "fmin v16.4s, v16.4s, v0.4s\n"
+      "fmin v17.4s, v17.4s, v0.4s\n"
+      "fmax v15.4s, v15.4s, v1.4s\n"
+      "fmax v16.4s, v16.4s, v1.4s\n"
+      "fmax v17.4s, v17.4s, v1.4s\n"
+      "fmin v18.4s, v18.4s, v0.4s\n"
+      "fmin v19.4s, v19.4s, v0.4s\n"
+      "fmax v18.4s, v18.4s, v1.4s\n"
+      "fmax v19.4s, v19.4s, v1.4s\n"
+      "92:"  // Height 3: No activation
+      "cmp x16, #0x10\n"
+      "bge 101f\n"
+      "tbz x16, #3, 96f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v9.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "st1 { v13.4s }, [x9], #0x10\n"
+      "st1 { v16.4s }, [x27], #0x10\n"
+      "st1 { v17.4s }, [x27], #0x10\n"
+      "tbz x16, #2, 94f\n"
+      "st1 { v10.4s }, [x13], #0x10\n"
+      "st1 { v14.4s }, [x9], #0x10\n"
+      "st1 { v18.4s }, [x27], #0x10\n"
+      "tbz x16, #1, 93f\n"
+      "str d11, [x13], #0x8\n"
+      "str d15, [x9], #0x8\n"
+      "str d19, [x27], #0x8\n"
+      "tbz x16, #0, 100f\n"
+      "st1 { v11.s }[2], [x13]\n"
+      "st1 { v15.s }[2], [x9]\n"
+      "st1 { v19.s }[2], [x27]\n"
+      "b 100f\n"
+      "93:"  // Height 3: Partial direct writeback: partial_1_12
+      "tbz x16, #0, 100f\n"
+      "str s11, [x13, #0x0]\n"
+      "str s15, [x9, #0x0]\n"
+      "str s19, [x27, #0x0]\n"
+      "b 100f\n"
+      "94:"  // Height 3: Partial direct writeback: partial_2_8
+      "tbz x16, #1, 95f\n"
+      "str d10, [x13], #0x8\n"
+      "str d14, [x9], #0x8\n"
+      "str d18, [x27], #0x8\n"
+      "tbz x16, #0, 100f\n"
+      "st1 { v10.s }[2], [x13]\n"
+      "st1 { v14.s }[2], [x9]\n"
+      "st1 { v18.s }[2], [x27]\n"
+      "b 100f\n"
+      "95:"  // Height 3: Partial direct writeback: partial_1_8
+      "tbz x16, #0, 100f\n"
+      "str s10, [x13, #0x0]\n"
+      "str s14, [x9, #0x0]\n"
+      "str s18, [x27, #0x0]\n"
+      "b 100f\n"
+      "96:"  // Height 3: Partial direct writeback: partial_4_0
+      "tbz x16, #2, 98f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "st1 { v16.4s }, [x27], #0x10\n"
+      "tbz x16, #1, 97f\n"
+      "str d9, [x13], #0x8\n"
+      "str d13, [x9], #0x8\n"
+      "str d17, [x27], #0x8\n"
+      "tbz x16, #0, 100f\n"
+      "st1 { v9.s }[2], [x13]\n"
+      "st1 { v13.s }[2], [x9]\n"
+      "st1 { v17.s }[2], [x27]\n"
+      "b 100f\n"
+      "97:"  // Height 3: Partial direct writeback: partial_1_4
+      "tbz x16, #0, 100f\n"
+      "str s9, [x13, #0x0]\n"
+      "str s13, [x9, #0x0]\n"
+      "str s17, [x27, #0x0]\n"
+      "b 100f\n"
+      "98:"  // Height 3: Partial direct writeback: partial_2_0
+      "tbz x16, #1, 99f\n"
+      "str d8, [x13], #0x8\n"
+      "str d12, [x9], #0x8\n"
+      "str d16, [x27], #0x8\n"
+      "tbz x16, #0, 100f\n"
+      "st1 { v8.s }[2], [x13]\n"
+      "st1 { v12.s }[2], [x9]\n"
+      "st1 { v16.s }[2], [x27]\n"
+      "b 100f\n"
+      "99:"  // Height 3: Partial direct writeback: partial_1_0
+      "str s8, [x13, #0x0]\n"
+      "str s12, [x9, #0x0]\n"
+      "str s16, [x27, #0x0]\n"
+      "100:"  // Height 3: Partial direct writeback: Done
+      "b 102f\n"
+      "101:"  // Height 3: Full writeback
+      "str q8, [x13, #0x0]\n"
+      "str q9, [x13, #0x10]\n"
+      "str q10, [x13, #0x20]\n"
+      "str q11, [x13, #0x30]\n"
+      "str q12, [x9, #0x0]\n"
+      "str q13, [x9, #0x10]\n"
+      "str q14, [x9, #0x20]\n"
+      "str q15, [x9, #0x30]\n"
+      "str q16, [x27, #0x0]\n"
+      "str q17, [x27, #0x10]\n"
+      "str q18, [x27, #0x20]\n"
+      "str q19, [x27, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "add x9, x9, #0x40\n"
+      "add x27, x27, #0x40\n"
+      "102:"  // Height 3: Writeback done
+      "subs x16, x16, #0x10\n"
+      "bgt 71b\n"
+      "b 206f\n"
+      "103:"  // Height 4
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x14, %x[bias]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "tbz %x[flags], #2, 104f\n"
+      "ldr x13, [%x[output_ptr], #0x0]\n"
+      "add x13, x13, x19, LSL #2\n"
+      "ldr x9, [%x[output_ptr], #0x8]\n"
+      "ldr x27, [%x[output_ptr], #0x10]\n"
+      "add x9, x9, x19, LSL #2\n"
+      "ldr x25, [%x[output_ptr], #0x18]\n"
+      "add x27, x27, x19, LSL #2\n"
+      "add x25, x25, x19, LSL #2\n"
+      "b 105f\n"
+      "104:"  // Height 4: setup direct output
+      "mov x13, %x[output_ptr]\n"
+      "add x9, x13, x19, LSL #2\n"
+      "add x27, x9, x19, LSL #2\n"
+      "add x25, x27, x19, LSL #2\n"
+      "105:"  // Height 4: Column loop
+      "cbz x14, 106f\n"
+      "ldr q8, [x14, #0x0]\n"
+      "mov v12.16b, v8.16b\n"
+      "ldr q9, [x14, #0x10]\n"
+      "mov v16.16b, v8.16b\n"
+      "ldr q10, [x14, #0x20]\n"
+      "mov v20.16b, v8.16b\n"
+      "ldr q11, [x14, #0x30]\n"
+      "add x14, x14, #0x40\n"
+      "mov v13.16b, v9.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "mov v14.16b, v10.16b\n"
+      "mov v15.16b, v11.16b\n"
+      "mov v18.16b, v10.16b\n"
+      "mov v19.16b, v11.16b\n"
+      "mov v21.16b, v9.16b\n"
+      "mov v22.16b, v10.16b\n"
+      "mov v23.16b, v11.16b\n"
+      "b 117f\n"
+      "106:"  // Height 4: no bias
+      "tbz %x[flags], #0, 116f\n"
+      "cmp x16, #0x10\n"
+      "bge 115f\n"
+      "tbz x16, #3, 110f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "ld1 { v16.4s }, [x27], #0x10\n"
+      "ld1 { v20.4s }, [x25], #0x10\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x9], #0x10\n"
+      "ld1 { v17.4s }, [x27], #0x10\n"
+      "ld1 { v21.4s }, [x25], #0x10\n"
+      "tbz x16, #2, 108f\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x9], #0x10\n"
+      "ld1 { v18.4s }, [x27], #0x10\n"
+      "ld1 { v22.4s }, [x25], #0x10\n"
+      "tbz x16, #1, 107f\n"
+      "mov x19, #0x38\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d15, [x9], #0x8\n"
+      "ldr d19, [x27], #0x8\n"
+      "ldr d23, [x25], #0x8\n"
+      "tbz x16, #0, 114f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x9]\n"
+      "ld1 { v19.s }[2], [x27]\n"
+      "ld1 { v23.s }[2], [x25]\n"
+      "b 114f\n"
+      "107:"  // Height 4: Partial accumulate: partial_1_12
+      "mov x19, #0x30\n"
+      "tbz x16, #0, 114f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s15, [x9, #0x0]\n"
+      "ldr s19, [x27, #0x0]\n"
+      "ldr s23, [x25, #0x0]\n"
+      "b 114f\n"
+      "108:"  // Height 4: Partial accumulate: partial_2_8
+      "tbz x16, #1, 109f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d14, [x9], #0x8\n"
+      "ldr d18, [x27], #0x8\n"
+      "ldr d22, [x25], #0x8\n"
+      "mov x19, #0x28\n"
+      "tbz x16, #0, 114f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x9]\n"
+      "ld1 { v18.s }[2], [x27]\n"
+      "ld1 { v22.s }[2], [x25]\n"
+      "b 114f\n"
+      "109:"  // Height 4: Partial accumulate: partial_1_8
+      "mov x19, #0x20\n"
+      "tbz x16, #0, 114f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s14, [x9, #0x0]\n"
+      "ldr s18, [x27, #0x0]\n"
+      "ldr s22, [x25, #0x0]\n"
+      "b 114f\n"
+      "110:"  // Height 4: Partial accumulate: partial_4_0
+      "tbz x16, #2, 112f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "ld1 { v16.4s }, [x27], #0x10\n"
+      "ld1 { v20.4s }, [x25], #0x10\n"
+      "tbz x16, #1, 111f\n"
+      "mov x19, #0x18\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d13, [x9], #0x8\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d21, [x25], #0x8\n"
+      "tbz x16, #0, 114f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x9]\n"
+      "ld1 { v17.s }[2], [x27]\n"
+      "ld1 { v21.s }[2], [x25]\n"
+      "b 114f\n"
+      "111:"  // Height 4: Partial accumulate: partial_1_4
+      "mov x19, #0x10\n"
+      "tbz x16, #0, 114f\n"
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s13, [x9, #0x0]\n"
+      "ldr s17, [x27, #0x0]\n"
+      "ldr s21, [x25, #0x0]\n"
+      "b 114f\n"
+      "112:"  // Height 4: Partial accumulate: partial_2_0
+      "tbz x16, #1, 113f\n"
+      "ldr d8, [x13], #0x8\n"
+      "ldr d12, [x9], #0x8\n"
+      "ldr d16, [x27], #0x8\n"
+      "ldr d20, [x25], #0x8\n"
+      "mov x19, #0x8\n"
+      "tbz x16, #0, 114f\n"
+      "ld1 { v8.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x9]\n"
+      "ld1 { v16.s }[2], [x27]\n"
+      "ld1 { v20.s }[2], [x25]\n"
+      "b 114f\n"
+      "113:"  // Height 4: Partial accumulate: partial_1_0
+      "mov x19, #0x0\n"
+      "ldr s8, [x13, #0x0]\n"
+      "ldr s12, [x9, #0x0]\n"
+      "ldr s16, [x27, #0x0]\n"
+      "ldr s20, [x25, #0x0]\n"
+      "114:"  // Height 4: Partial accumulate: Done
+      "sub x13, x13, x19\n"
+      "sub x9, x9, x19\n"
+      "sub x27, x27, x19\n"
+      "sub x25, x25, x19\n"
+      "b 117f\n"
+      "115:"  // Height 4: full accumulate
+      "ldr q8, [x13, #0x0]\n"
+      "ldr q9, [x13, #0x10]\n"
+      "ldr q10, [x13, #0x20]\n"
+      "ldr q11, [x13, #0x30]\n"
+      "ldr q12, [x9, #0x0]\n"
+      "ldr q13, [x9, #0x10]\n"
+      "ldr q14, [x9, #0x20]\n"
+      "ldr q15, [x9, #0x30]\n"
+      "ldr q16, [x27, #0x0]\n"
+      "ldr q17, [x27, #0x10]\n"
+      "ldr q18, [x27, #0x20]\n"
+      "ldr q19, [x27, #0x30]\n"
+      "ldr q20, [x25, #0x0]\n"
+      "ldr q21, [x25, #0x10]\n"
+      "ldr q22, [x25, #0x20]\n"
+      "ldr q23, [x25, #0x30]\n"
+      "b 117f\n"
+      "116:"  // Height 4: no accumulate
+      "movi v8.16b, #0x0\n"
+      "movi v9.16b, #0x0\n"
+      "movi v10.16b, #0x0\n"
+      "movi v11.16b, #0x0\n"
+      "movi v12.16b, #0x0\n"
+      "movi v13.16b, #0x0\n"
+      "movi v14.16b, #0x0\n"
+      "movi v15.16b, #0x0\n"
+      "movi v16.16b, #0x0\n"
+      "movi v17.16b, #0x0\n"
+      "movi v18.16b, #0x0\n"
+      "movi v19.16b, #0x0\n"
+      "movi v20.16b, #0x0\n"
+      "movi v21.16b, #0x0\n"
+      "movi v22.16b, #0x0\n"
+      "movi v23.16b, #0x0\n"
+      "117:"  // Height 4: setup done
+      "mov x12, #0x0\n"
+      "118:"  // Height 4: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w11, [x20, x12, LSL #0x2]\n"
+      "tbz %x[flags], #3, 119f\n"
+      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "add x20, x20, x19, LSL #3\n"
+      "ldr x10, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "ldr x26, [x20, #0x10]\n"
+      "ldr x24, [x20, #0x18]\n"
+      "cbnz x12, 120f\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x10, x10, x19, LSL #2\n"
+      "add x28, x28, x19, LSL #2\n"
+      "add x26, x26, x19, LSL #2\n"
+      "add x24, x24, x19, LSL #2\n"
+      "b 120f\n"
+      "119:"  // Height 4: setup direct input
+      "mov x10, %x[input_ptr]\n"
+      "add x28, x10, x19, LSL #2\n"
+      "add x26, x28, x19, LSL #2\n"
+      "add x24, x26, x19, LSL #2\n"
+      "120:"  // Height 4: input setup done
+      "cmp x11, #0x4\n"
+      "blt 123f\n"
+      "cmp x11, #0x8\n"
+      "blt 122f\n"
+      "121:"  // Height 4: Multiply loop: Main loop head
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x26, #0x0]\n"
+      "ldr q3, [x24, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "add x10, x10, #0x10\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "add x28, x28, #0x10\n"
+      "fmla v20.4s, v6.4s, v3.s[0]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "add x26, x26, #0x10\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "add x24, x24, #0x10\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "sub x11, x11, #0x4\n"
+      "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "cmp x11, #0x8\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v22.4s, v6.4s, v3.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "fmla v23.4s, v7.4s, v3.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "fmla v16.4s, v6.4s, v2.s[1]\n"
+      "fmla v20.4s, v6.4s, v3.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "fmla v17.4s, v7.4s, v2.s[1]\n"
+      "fmla v21.4s, v7.4s, v3.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "fmla v18.4s, v6.4s, v2.s[1]\n"
+      "fmla v22.4s, v6.4s, v3.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "fmla v19.4s, v7.4s, v2.s[1]\n"
+      "fmla v23.4s, v7.4s, v3.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "fmla v16.4s, v6.4s, v2.s[2]\n"
+      "fmla v20.4s, v6.4s, v3.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "fmla v17.4s, v7.4s, v2.s[2]\n"
+      "fmla v21.4s, v7.4s, v3.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "fmla v18.4s, v6.4s, v2.s[2]\n"
+      "fmla v22.4s, v6.4s, v3.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "fmla v19.4s, v7.4s, v2.s[2]\n"
+      "fmla v23.4s, v7.4s, v3.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "fmla v16.4s, v6.4s, v2.s[3]\n"
+      "fmla v20.4s, v6.4s, v3.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "fmla v17.4s, v7.4s, v2.s[3]\n"
+      "fmla v21.4s, v7.4s, v3.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v18.4s, v6.4s, v2.s[3]\n"
+      "fmla v22.4s, v6.4s, v3.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "fmla v19.4s, v7.4s, v2.s[3]\n"
+      "fmla v23.4s, v7.4s, v3.s[3]\n"
+      "bge 121b\n"
+      "122:"  // Height 4: Multiply loop: Single iteration only
+      "sub x11, x11, #0x4\n"
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x26, #0x0]\n"
+      "ldr q3, [x24, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "add x10, x10, #0x10\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "add x28, x28, #0x10\n"
+      "fmla v20.4s, v6.4s, v3.s[0]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "add x26, x26, #0x10\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "add x24, x24, #0x10\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v22.4s, v6.4s, v3.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "fmla v23.4s, v7.4s, v3.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "fmla v16.4s, v6.4s, v2.s[1]\n"
+      "fmla v20.4s, v6.4s, v3.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "fmla v17.4s, v7.4s, v2.s[1]\n"
+      "fmla v21.4s, v7.4s, v3.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "fmla v18.4s, v6.4s, v2.s[1]\n"
+      "fmla v22.4s, v6.4s, v3.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "fmla v19.4s, v7.4s, v2.s[1]\n"
+      "fmla v23.4s, v7.4s, v3.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "fmla v16.4s, v6.4s, v2.s[2]\n"
+      "fmla v20.4s, v6.4s, v3.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "fmla v17.4s, v7.4s, v2.s[2]\n"
+      "fmla v21.4s, v7.4s, v3.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "fmla v18.4s, v6.4s, v2.s[2]\n"
+      "fmla v22.4s, v6.4s, v3.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "fmla v19.4s, v7.4s, v2.s[2]\n"
+      "fmla v23.4s, v7.4s, v3.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "fmla v16.4s, v6.4s, v2.s[3]\n"
+      "fmla v20.4s, v6.4s, v3.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "fmla v17.4s, v7.4s, v2.s[3]\n"
+      "fmla v21.4s, v7.4s, v3.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v18.4s, v6.4s, v2.s[3]\n"
+      "fmla v22.4s, v6.4s, v3.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "fmla v19.4s, v7.4s, v2.s[3]\n"
+      "fmla v23.4s, v7.4s, v3.s[3]\n"
+      "123:"  // Height 4: Multiply loop: Main loop skip
+      "cbz x11, 125f\n"
+      "124:"  // Height 4: Multiply loop: Odd block loop
+      "ldr s0, [x10], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
+      "ldr s2, [x26], #0x4\n"
+      "ldr s3, [x24], #0x4\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "sub x11, x11, #0x1\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "fmla v20.4s, v6.4s, v3.s[0]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "add x15, x15, #0x40\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v22.4s, v6.4s, v3.s[0]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "fmla v23.4s, v7.4s, v3.s[0]\n"
+      "cbnz x11, 124b\n"
+      "125:"  // Height 4: Multiply loop: No odd multiplies
+      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x19\n"
+      "bne 118b\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "tbz %x[flags], #1, 126f\n"
+      "add x19, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v1.4s }, [x19]\n"
+      "add x19, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v0.4s }, [x19]\n"
+      "fmin v8.4s, v8.4s, v0.4s\n"
+      "fmin v9.4s, v9.4s, v0.4s\n"
+      "fmin v10.4s, v10.4s, v0.4s\n"
+      "fmin v11.4s, v11.4s, v0.4s\n"
+      "fmax v8.4s, v8.4s, v1.4s\n"
+      "fmax v9.4s, v9.4s, v1.4s\n"
+      "fmax v10.4s, v10.4s, v1.4s\n"
+      "fmax v11.4s, v11.4s, v1.4s\n"
+      "fmin v12.4s, v12.4s, v0.4s\n"
+      "fmin v13.4s, v13.4s, v0.4s\n"
+      "fmin v14.4s, v14.4s, v0.4s\n"
+      "fmax v12.4s, v12.4s, v1.4s\n"
+      "fmax v13.4s, v13.4s, v1.4s\n"
+      "fmax v14.4s, v14.4s, v1.4s\n"
+      "fmin v15.4s, v15.4s, v0.4s\n"
+      "fmin v16.4s, v16.4s, v0.4s\n"
+      "fmin v17.4s, v17.4s, v0.4s\n"
+      "fmax v15.4s, v15.4s, v1.4s\n"
+      "fmax v16.4s, v16.4s, v1.4s\n"
+      "fmax v17.4s, v17.4s, v1.4s\n"
+      "fmin v18.4s, v18.4s, v0.4s\n"
+      "fmin v19.4s, v19.4s, v0.4s\n"
+      "fmin v20.4s, v20.4s, v0.4s\n"
+      "fmax v18.4s, v18.4s, v1.4s\n"
+      "fmax v19.4s, v19.4s, v1.4s\n"
+      "fmax v20.4s, v20.4s, v1.4s\n"
+      "fmin v21.4s, v21.4s, v0.4s\n"
+      "fmin v22.4s, v22.4s, v0.4s\n"
+      "fmin v23.4s, v23.4s, v0.4s\n"
+      "fmax v21.4s, v21.4s, v1.4s\n"
+      "fmax v22.4s, v22.4s, v1.4s\n"
+      "fmax v23.4s, v23.4s, v1.4s\n"
+      "126:"  // Height 4: No activation
+      "cmp x16, #0x10\n"
+      "bge 135f\n"
+      "tbz x16, #3, 130f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v9.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "st1 { v13.4s }, [x9], #0x10\n"
+      "st1 { v16.4s }, [x27], #0x10\n"
+      "st1 { v17.4s }, [x27], #0x10\n"
+      "st1 { v20.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x25], #0x10\n"
+      "tbz x16, #2, 128f\n"
+      "st1 { v10.4s }, [x13], #0x10\n"
+      "st1 { v14.4s }, [x9], #0x10\n"
+      "st1 { v18.4s }, [x27], #0x10\n"
+      "st1 { v22.4s }, [x25], #0x10\n"
+      "tbz x16, #1, 127f\n"
+      "str d11, [x13], #0x8\n"
+      "str d15, [x9], #0x8\n"
+      "str d19, [x27], #0x8\n"
+      "str d23, [x25], #0x8\n"
+      "tbz x16, #0, 134f\n"
+      "st1 { v11.s }[2], [x13]\n"
+      "st1 { v15.s }[2], [x9]\n"
+      "st1 { v19.s }[2], [x27]\n"
+      "st1 { v23.s }[2], [x25]\n"
+      "b 134f\n"
+      "127:"  // Height 4: Partial direct writeback: partial_1_12
+      "tbz x16, #0, 134f\n"
+      "str s11, [x13, #0x0]\n"
+      "str s15, [x9, #0x0]\n"
+      "str s19, [x27, #0x0]\n"
+      "str s23, [x25, #0x0]\n"
+      "b 134f\n"
+      "128:"  // Height 4: Partial direct writeback: partial_2_8
+      "tbz x16, #1, 129f\n"
+      "str d10, [x13], #0x8\n"
+      "str d14, [x9], #0x8\n"
+      "str d18, [x27], #0x8\n"
+      "str d22, [x25], #0x8\n"
+      "tbz x16, #0, 134f\n"
+      "st1 { v10.s }[2], [x13]\n"
+      "st1 { v14.s }[2], [x9]\n"
+      "st1 { v18.s }[2], [x27]\n"
+      "st1 { v22.s }[2], [x25]\n"
+      "b 134f\n"
+      "129:"  // Height 4: Partial direct writeback: partial_1_8
+      "tbz x16, #0, 134f\n"
+      "str s10, [x13, #0x0]\n"
+      "str s14, [x9, #0x0]\n"
+      "str s18, [x27, #0x0]\n"
+      "str s22, [x25, #0x0]\n"
+      "b 134f\n"
+      "130:"  // Height 4: Partial direct writeback: partial_4_0
+      "tbz x16, #2, 132f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "st1 { v16.4s }, [x27], #0x10\n"
+      "st1 { v20.4s }, [x25], #0x10\n"
+      "tbz x16, #1, 131f\n"
+      "str d9, [x13], #0x8\n"
+      "str d13, [x9], #0x8\n"
+      "str d17, [x27], #0x8\n"
+      "str d21, [x25], #0x8\n"
+      "tbz x16, #0, 134f\n"
+      "st1 { v9.s }[2], [x13]\n"
+      "st1 { v13.s }[2], [x9]\n"
+      "st1 { v17.s }[2], [x27]\n"
+      "st1 { v21.s }[2], [x25]\n"
+      "b 134f\n"
+      "131:"  // Height 4: Partial direct writeback: partial_1_4
+      "tbz x16, #0, 134f\n"
+      "str s9, [x13, #0x0]\n"
+      "str s13, [x9, #0x0]\n"
+      "str s17, [x27, #0x0]\n"
+      "str s21, [x25, #0x0]\n"
+      "b 134f\n"
+      "132:"  // Height 4: Partial direct writeback: partial_2_0
+      "tbz x16, #1, 133f\n"
+      "str d8, [x13], #0x8\n"
+      "str d12, [x9], #0x8\n"
+      "str d16, [x27], #0x8\n"
+      "str d20, [x25], #0x8\n"
+      "tbz x16, #0, 134f\n"
+      "st1 { v8.s }[2], [x13]\n"
+      "st1 { v12.s }[2], [x9]\n"
+      "st1 { v16.s }[2], [x27]\n"
+      "st1 { v20.s }[2], [x25]\n"
+      "b 134f\n"
+      "133:"  // Height 4: Partial direct writeback: partial_1_0
+      "str s8, [x13, #0x0]\n"
+      "str s12, [x9, #0x0]\n"
+      "str s16, [x27, #0x0]\n"
+      "str s20, [x25, #0x0]\n"
+      "134:"  // Height 4: Partial direct writeback: Done
+      "b 136f\n"
+      "135:"  // Height 4: Full writeback
+      "str q8, [x13, #0x0]\n"
+      "str q9, [x13, #0x10]\n"
+      "str q10, [x13, #0x20]\n"
+      "str q11, [x13, #0x30]\n"
+      "str q12, [x9, #0x0]\n"
+      "str q13, [x9, #0x10]\n"
+      "str q14, [x9, #0x20]\n"
+      "str q15, [x9, #0x30]\n"
+      "str q16, [x27, #0x0]\n"
+      "str q17, [x27, #0x10]\n"
+      "str q18, [x27, #0x20]\n"
+      "str q19, [x27, #0x30]\n"
+      "str q20, [x25, #0x0]\n"
+      "str q21, [x25, #0x10]\n"
+      "str q22, [x25, #0x20]\n"
+      "str q23, [x25, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "add x9, x9, #0x40\n"
+      "add x27, x27, #0x40\n"
+      "add x25, x25, #0x40\n"
+      "136:"  // Height 4: Writeback done
+      "subs x16, x16, #0x10\n"
+      "bgt 105b\n"
+      "b 206f\n"
+      "137:"  // Height 5
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x14, %x[bias]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "tbz %x[flags], #2, 138f\n"
+      "ldr x13, [%x[output_ptr], #0x0]\n"
+      "add x13, x13, x19, LSL #2\n"
+      "ldr x9, [%x[output_ptr], #0x8]\n"
+      "ldr x27, [%x[output_ptr], #0x10]\n"
+      "add x9, x9, x19, LSL #2\n"
+      "ldr x25, [%x[output_ptr], #0x18]\n"
+      "ldr x23, [%x[output_ptr], #0x20]\n"
+      "add x27, x27, x19, LSL #2\n"
+      "add x25, x25, x19, LSL #2\n"
+      "add x23, x23, x19, LSL #2\n"
+      "b 139f\n"
+      "138:"  // Height 5: setup direct output
+      "mov x13, %x[output_ptr]\n"
+      "add x9, x13, x19, LSL #2\n"
+      "add x27, x9, x19, LSL #2\n"
+      "add x25, x27, x19, LSL #2\n"
+      "add x23, x25, x19, LSL #2\n"
+      "139:"  // Height 5: Column loop
+      "cbz x14, 140f\n"
+      "ldr q8, [x14, #0x0]\n"
+      "mov v12.16b, v8.16b\n"
+      "ldr q9, [x14, #0x10]\n"
+      "mov v16.16b, v8.16b\n"
+      "ldr q10, [x14, #0x20]\n"
+      "mov v20.16b, v8.16b\n"
+      "ldr q11, [x14, #0x30]\n"
+      "mov v24.16b, v8.16b\n"
+      "add x14, x14, #0x40\n"
+      "mov v13.16b, v9.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "mov v14.16b, v10.16b\n"
+      "mov v15.16b, v11.16b\n"
+      "mov v18.16b, v10.16b\n"
+      "mov v19.16b, v11.16b\n"
+      "mov v21.16b, v9.16b\n"
+      "mov v22.16b, v10.16b\n"
+      "mov v23.16b, v11.16b\n"
+      "mov v25.16b, v9.16b\n"
+      "mov v26.16b, v10.16b\n"
+      "mov v27.16b, v11.16b\n"
+      "b 151f\n"
+      "140:"  // Height 5: no bias
+      "tbz %x[flags], #0, 150f\n"
+      "cmp x16, #0x10\n"
+      "bge 149f\n"
+      "tbz x16, #3, 144f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "ld1 { v16.4s }, [x27], #0x10\n"
+      "ld1 { v20.4s }, [x25], #0x10\n"
+      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x9], #0x10\n"
+      "ld1 { v17.4s }, [x27], #0x10\n"
+      "ld1 { v21.4s }, [x25], #0x10\n"
+      "ld1 { v25.4s }, [x23], #0x10\n"
+      "tbz x16, #2, 142f\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x9], #0x10\n"
+      "ld1 { v18.4s }, [x27], #0x10\n"
+      "ld1 { v22.4s }, [x25], #0x10\n"
+      "ld1 { v26.4s }, [x23], #0x10\n"
+      "tbz x16, #1, 141f\n"
+      "mov x19, #0x38\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d15, [x9], #0x8\n"
+      "ldr d19, [x27], #0x8\n"
+      "ldr d23, [x25], #0x8\n"
+      "ldr d27, [x23], #0x8\n"
+      "tbz x16, #0, 148f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x9]\n"
+      "ld1 { v19.s }[2], [x27]\n"
+      "ld1 { v23.s }[2], [x25]\n"
+      "ld1 { v27.s }[2], [x23]\n"
+      "b 148f\n"
+      "141:"  // Height 5: Partial accumulate: partial_1_12
+      "mov x19, #0x30\n"
+      "tbz x16, #0, 148f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s15, [x9, #0x0]\n"
+      "ldr s19, [x27, #0x0]\n"
+      "ldr s23, [x25, #0x0]\n"
+      "ldr s27, [x23, #0x0]\n"
+      "b 148f\n"
+      "142:"  // Height 5: Partial accumulate: partial_2_8
+      "tbz x16, #1, 143f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d14, [x9], #0x8\n"
+      "ldr d18, [x27], #0x8\n"
+      "ldr d22, [x25], #0x8\n"
+      "ldr d26, [x23], #0x8\n"
+      "mov x19, #0x28\n"
+      "tbz x16, #0, 148f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x9]\n"
+      "ld1 { v18.s }[2], [x27]\n"
+      "ld1 { v22.s }[2], [x25]\n"
+      "ld1 { v26.s }[2], [x23]\n"
+      "b 148f\n"
+      "143:"  // Height 5: Partial accumulate: partial_1_8
+      "mov x19, #0x20\n"
+      "tbz x16, #0, 148f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s14, [x9, #0x0]\n"
+      "ldr s18, [x27, #0x0]\n"
+      "ldr s22, [x25, #0x0]\n"
+      "ldr s26, [x23, #0x0]\n"
+      "b 148f\n"
+      "144:"  // Height 5: Partial accumulate: partial_4_0
+      "tbz x16, #2, 146f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "ld1 { v16.4s }, [x27], #0x10\n"
+      "ld1 { v20.4s }, [x25], #0x10\n"
+      "ld1 { v24.4s }, [x23], #0x10\n"
+      "tbz x16, #1, 145f\n"
+      "mov x19, #0x18\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d13, [x9], #0x8\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d21, [x25], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "tbz x16, #0, 148f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x9]\n"
+      "ld1 { v17.s }[2], [x27]\n"
+      "ld1 { v21.s }[2], [x25]\n"
+      "ld1 { v25.s }[2], [x23]\n"
+      "b 148f\n"
+      "145:"  // Height 5: Partial accumulate: partial_1_4
+      "mov x19, #0x10\n"
+      "tbz x16, #0, 148f\n"
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s13, [x9, #0x0]\n"
+      "ldr s17, [x27, #0x0]\n"
+      "ldr s21, [x25, #0x0]\n"
+      "ldr s25, [x23, #0x0]\n"
+      "b 148f\n"
+      "146:"  // Height 5: Partial accumulate: partial_2_0
+      "tbz x16, #1, 147f\n"
+      "ldr d8, [x13], #0x8\n"
+      "ldr d12, [x9], #0x8\n"
+      "ldr d16, [x27], #0x8\n"
+      "ldr d20, [x25], #0x8\n"
+      "ldr d24, [x23], #0x8\n"
+      "mov x19, #0x8\n"
+      "tbz x16, #0, 148f\n"
+      "ld1 { v8.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x9]\n"
+      "ld1 { v16.s }[2], [x27]\n"
+      "ld1 { v20.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x23]\n"
+      "b 148f\n"
+      "147:"  // Height 5: Partial accumulate: partial_1_0
+      "mov x19, #0x0\n"
+      "ldr s8, [x13, #0x0]\n"
+      "ldr s12, [x9, #0x0]\n"
+      "ldr s16, [x27, #0x0]\n"
+      "ldr s20, [x25, #0x0]\n"
+      "ldr s24, [x23, #0x0]\n"
+      "148:"  // Height 5: Partial accumulate: Done
+      "sub x13, x13, x19\n"
+      "sub x9, x9, x19\n"
+      "sub x27, x27, x19\n"
+      "sub x25, x25, x19\n"
+      "sub x23, x23, x19\n"
+      "b 151f\n"
+      "149:"  // Height 5: full accumulate
+      "ldr q8, [x13, #0x0]\n"
+      "ldr q9, [x13, #0x10]\n"
+      "ldr q10, [x13, #0x20]\n"
+      "ldr q11, [x13, #0x30]\n"
+      "ldr q12, [x9, #0x0]\n"
+      "ldr q13, [x9, #0x10]\n"
+      "ldr q14, [x9, #0x20]\n"
+      "ldr q15, [x9, #0x30]\n"
+      "ldr q16, [x27, #0x0]\n"
+      "ldr q17, [x27, #0x10]\n"
+      "ldr q18, [x27, #0x20]\n"
+      "ldr q19, [x27, #0x30]\n"
+      "ldr q20, [x25, #0x0]\n"
+      "ldr q21, [x25, #0x10]\n"
+      "ldr q22, [x25, #0x20]\n"
+      "ldr q23, [x25, #0x30]\n"
+      "ldr q24, [x23, #0x0]\n"
+      "ldr q25, [x23, #0x10]\n"
+      "ldr q26, [x23, #0x20]\n"
+      "ldr q27, [x23, #0x30]\n"
+      "b 151f\n"
+      "150:"  // Height 5: no accumulate
+      "movi v8.16b, #0x0\n"
+      "movi v9.16b, #0x0\n"
+      "movi v10.16b, #0x0\n"
+      "movi v11.16b, #0x0\n"
+      "movi v12.16b, #0x0\n"
+      "movi v13.16b, #0x0\n"
+      "movi v14.16b, #0x0\n"
+      "movi v15.16b, #0x0\n"
+      "movi v16.16b, #0x0\n"
+      "movi v17.16b, #0x0\n"
+      "movi v18.16b, #0x0\n"
+      "movi v19.16b, #0x0\n"
+      "movi v20.16b, #0x0\n"
+      "movi v21.16b, #0x0\n"
+      "movi v22.16b, #0x0\n"
+      "movi v23.16b, #0x0\n"
+      "movi v24.16b, #0x0\n"
+      "movi v25.16b, #0x0\n"
+      "movi v26.16b, #0x0\n"
+      "movi v27.16b, #0x0\n"
+      "151:"  // Height 5: setup done
+      "mov x12, #0x0\n"
+      "152:"  // Height 5: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w11, [x20, x12, LSL #0x2]\n"
+      "tbz %x[flags], #3, 153f\n"
+      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "add x20, x20, x19, LSL #3\n"
+      "ldr x10, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "ldr x26, [x20, #0x10]\n"
+      "ldr x24, [x20, #0x18]\n"
+      "ldr x22, [x20, #0x20]\n"
+      "cbnz x12, 154f\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x10, x10, x19, LSL #2\n"
+      "add x28, x28, x19, LSL #2\n"
+      "add x26, x26, x19, LSL #2\n"
+      "add x24, x24, x19, LSL #2\n"
+      "add x22, x22, x19, LSL #2\n"
+      "b 154f\n"
+      "153:"  // Height 5: setup direct input
+      "mov x10, %x[input_ptr]\n"
+      "add x28, x10, x19, LSL #2\n"
+      "add x26, x28, x19, LSL #2\n"
+      "add x24, x26, x19, LSL #2\n"
+      "add x22, x24, x19, LSL #2\n"
+      "154:"  // Height 5: input setup done
+      "cmp x11, #0x4\n"
+      "blt 157f\n"
+      "cmp x11, #0x8\n"
+      "blt 156f\n"
+      "155:"  // Height 5: Multiply loop: Main loop head
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x26, #0x0]\n"
+      "ldr q3, [x24, #0x0]\n"
+      "ldr q4, [x22, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "add x10, x10, #0x10\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "add x28, x28, #0x10\n"
+      "fmla v20.4s, v6.4s, v3.s[0]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "add x26, x26, #0x10\n"
+      "fmla v24.4s, v6.4s, v4.s[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "add x24, x24, #0x10\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "add x22, x22, #0x10\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
+      "sub x11, x11, #0x4\n"
+      "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "cmp x11, #0x8\n"
+      "fmla v25.4s, v7.4s, v4.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v22.4s, v6.4s, v3.s[0]\n"
+      "fmla v26.4s, v6.4s, v4.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "fmla v23.4s, v7.4s, v3.s[0]\n"
+      "fmla v27.4s, v7.4s, v4.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "fmla v16.4s, v6.4s, v2.s[1]\n"
+      "fmla v20.4s, v6.4s, v3.s[1]\n"
+      "fmla v24.4s, v6.4s, v4.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "fmla v17.4s, v7.4s, v2.s[1]\n"
+      "fmla v21.4s, v7.4s, v3.s[1]\n"
+      "fmla v25.4s, v7.4s, v4.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "fmla v18.4s, v6.4s, v2.s[1]\n"
+      "fmla v22.4s, v6.4s, v3.s[1]\n"
+      "fmla v26.4s, v6.4s, v4.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "fmla v19.4s, v7.4s, v2.s[1]\n"
+      "fmla v23.4s, v7.4s, v3.s[1]\n"
+      "fmla v27.4s, v7.4s, v4.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "fmla v16.4s, v6.4s, v2.s[2]\n"
+      "fmla v20.4s, v6.4s, v3.s[2]\n"
+      "fmla v24.4s, v6.4s, v4.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "fmla v17.4s, v7.4s, v2.s[2]\n"
+      "fmla v21.4s, v7.4s, v3.s[2]\n"
+      "fmla v25.4s, v7.4s, v4.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "fmla v18.4s, v6.4s, v2.s[2]\n"
+      "fmla v22.4s, v6.4s, v3.s[2]\n"
+      "fmla v26.4s, v6.4s, v4.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "fmla v19.4s, v7.4s, v2.s[2]\n"
+      "fmla v23.4s, v7.4s, v3.s[2]\n"
+      "fmla v27.4s, v7.4s, v4.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "fmla v16.4s, v6.4s, v2.s[3]\n"
+      "fmla v20.4s, v6.4s, v3.s[3]\n"
+      "fmla v24.4s, v6.4s, v4.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "fmla v17.4s, v7.4s, v2.s[3]\n"
+      "fmla v21.4s, v7.4s, v3.s[3]\n"
+      "fmla v25.4s, v7.4s, v4.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v18.4s, v6.4s, v2.s[3]\n"
+      "fmla v22.4s, v6.4s, v3.s[3]\n"
+      "fmla v26.4s, v6.4s, v4.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "fmla v19.4s, v7.4s, v2.s[3]\n"
+      "fmla v23.4s, v7.4s, v3.s[3]\n"
+      "fmla v27.4s, v7.4s, v4.s[3]\n"
+      "bge 155b\n"
+      "156:"  // Height 5: Multiply loop: Single iteration only
+      "sub x11, x11, #0x4\n"
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x26, #0x0]\n"
+      "ldr q3, [x24, #0x0]\n"
+      "ldr q4, [x22, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "add x10, x10, #0x10\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "add x28, x28, #0x10\n"
+      "fmla v20.4s, v6.4s, v3.s[0]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "add x26, x26, #0x10\n"
+      "fmla v24.4s, v6.4s, v4.s[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "add x24, x24, #0x10\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "add x22, x22, #0x10\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
+      "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "fmla v25.4s, v7.4s, v4.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v22.4s, v6.4s, v3.s[0]\n"
+      "fmla v26.4s, v6.4s, v4.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "fmla v23.4s, v7.4s, v3.s[0]\n"
+      "fmla v27.4s, v7.4s, v4.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "fmla v16.4s, v6.4s, v2.s[1]\n"
+      "fmla v20.4s, v6.4s, v3.s[1]\n"
+      "fmla v24.4s, v6.4s, v4.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "fmla v17.4s, v7.4s, v2.s[1]\n"
+      "fmla v21.4s, v7.4s, v3.s[1]\n"
+      "fmla v25.4s, v7.4s, v4.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "fmla v18.4s, v6.4s, v2.s[1]\n"
+      "fmla v22.4s, v6.4s, v3.s[1]\n"
+      "fmla v26.4s, v6.4s, v4.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "fmla v19.4s, v7.4s, v2.s[1]\n"
+      "fmla v23.4s, v7.4s, v3.s[1]\n"
+      "fmla v27.4s, v7.4s, v4.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "fmla v16.4s, v6.4s, v2.s[2]\n"
+      "fmla v20.4s, v6.4s, v3.s[2]\n"
+      "fmla v24.4s, v6.4s, v4.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "fmla v17.4s, v7.4s, v2.s[2]\n"
+      "fmla v21.4s, v7.4s, v3.s[2]\n"
+      "fmla v25.4s, v7.4s, v4.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "fmla v18.4s, v6.4s, v2.s[2]\n"
+      "fmla v22.4s, v6.4s, v3.s[2]\n"
+      "fmla v26.4s, v6.4s, v4.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "fmla v19.4s, v7.4s, v2.s[2]\n"
+      "fmla v23.4s, v7.4s, v3.s[2]\n"
+      "fmla v27.4s, v7.4s, v4.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "fmla v16.4s, v6.4s, v2.s[3]\n"
+      "fmla v20.4s, v6.4s, v3.s[3]\n"
+      "fmla v24.4s, v6.4s, v4.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "fmla v17.4s, v7.4s, v2.s[3]\n"
+      "fmla v21.4s, v7.4s, v3.s[3]\n"
+      "fmla v25.4s, v7.4s, v4.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v18.4s, v6.4s, v2.s[3]\n"
+      "fmla v22.4s, v6.4s, v3.s[3]\n"
+      "fmla v26.4s, v6.4s, v4.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "fmla v19.4s, v7.4s, v2.s[3]\n"
+      "fmla v23.4s, v7.4s, v3.s[3]\n"
+      "fmla v27.4s, v7.4s, v4.s[3]\n"
+      "157:"  // Height 5: Multiply loop: Main loop skip
+      "cbz x11, 159f\n"
+      "158:"  // Height 5: Multiply loop: Odd block loop
+      "ldr s0, [x10], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
+      "ldr s2, [x26], #0x4\n"
+      "ldr s3, [x24], #0x4\n"
+      "ldr s4, [x22], #0x4\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "sub x11, x11, #0x1\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "fmla v20.4s, v6.4s, v3.s[0]\n"
+      "fmla v24.4s, v6.4s, v4.s[0]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "fmla v25.4s, v7.4s, v4.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "add x15, x15, #0x40\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v22.4s, v6.4s, v3.s[0]\n"
+      "fmla v26.4s, v6.4s, v4.s[0]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "fmla v23.4s, v7.4s, v3.s[0]\n"
+      "fmla v27.4s, v7.4s, v4.s[0]\n"
+      "cbnz x11, 158b\n"
+      "159:"  // Height 5: Multiply loop: No odd multiplies
+      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x19\n"
+      "bne 152b\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "tbz %x[flags], #1, 160f\n"
+      "add x19, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v1.4s }, [x19]\n"
+      "add x19, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v0.4s }, [x19]\n"
+      "fmin v8.4s, v8.4s, v0.4s\n"
+      "fmin v9.4s, v9.4s, v0.4s\n"
+      "fmin v10.4s, v10.4s, v0.4s\n"
+      "fmin v11.4s, v11.4s, v0.4s\n"
+      "fmax v8.4s, v8.4s, v1.4s\n"
+      "fmax v9.4s, v9.4s, v1.4s\n"
+      "fmax v10.4s, v10.4s, v1.4s\n"
+      "fmax v11.4s, v11.4s, v1.4s\n"
+      "fmin v12.4s, v12.4s, v0.4s\n"
+      "fmin v13.4s, v13.4s, v0.4s\n"
+      "fmin v14.4s, v14.4s, v0.4s\n"
+      "fmax v12.4s, v12.4s, v1.4s\n"
+      "fmax v13.4s, v13.4s, v1.4s\n"
+      "fmax v14.4s, v14.4s, v1.4s\n"
+      "fmin v15.4s, v15.4s, v0.4s\n"
+      "fmin v16.4s, v16.4s, v0.4s\n"
+      "fmin v17.4s, v17.4s, v0.4s\n"
+      "fmax v15.4s, v15.4s, v1.4s\n"
+      "fmax v16.4s, v16.4s, v1.4s\n"
+      "fmax v17.4s, v17.4s, v1.4s\n"
+      "fmin v18.4s, v18.4s, v0.4s\n"
+      "fmin v19.4s, v19.4s, v0.4s\n"
+      "fmin v20.4s, v20.4s, v0.4s\n"
+      "fmax v18.4s, v18.4s, v1.4s\n"
+      "fmax v19.4s, v19.4s, v1.4s\n"
+      "fmax v20.4s, v20.4s, v1.4s\n"
+      "fmin v21.4s, v21.4s, v0.4s\n"
+      "fmin v22.4s, v22.4s, v0.4s\n"
+      "fmin v23.4s, v23.4s, v0.4s\n"
+      "fmax v21.4s, v21.4s, v1.4s\n"
+      "fmax v22.4s, v22.4s, v1.4s\n"
+      "fmax v23.4s, v23.4s, v1.4s\n"
+      "fmin v24.4s, v24.4s, v0.4s\n"
+      "fmin v25.4s, v25.4s, v0.4s\n"
+      "fmin v26.4s, v26.4s, v0.4s\n"
+      "fmax v24.4s, v24.4s, v1.4s\n"
+      "fmax v25.4s, v25.4s, v1.4s\n"
+      "fmax v26.4s, v26.4s, v1.4s\n"
+      "fmin v27.4s, v27.4s, v0.4s\n"
+      "fmax v27.4s, v27.4s, v1.4s\n"
+      "160:"  // Height 5: No activation
+      "cmp x16, #0x10\n"
+      "bge 169f\n"
+      "tbz x16, #3, 164f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v9.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "st1 { v13.4s }, [x9], #0x10\n"
+      "st1 { v16.4s }, [x27], #0x10\n"
+      "st1 { v17.4s }, [x27], #0x10\n"
+      "st1 { v20.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x25], #0x10\n"
+      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v25.4s }, [x23], #0x10\n"
+      "tbz x16, #2, 162f\n"
+      "st1 { v10.4s }, [x13], #0x10\n"
+      "st1 { v14.4s }, [x9], #0x10\n"
+      "st1 { v18.4s }, [x27], #0x10\n"
+      "st1 { v22.4s }, [x25], #0x10\n"
+      "st1 { v26.4s }, [x23], #0x10\n"
+      "tbz x16, #1, 161f\n"
+      "str d11, [x13], #0x8\n"
+      "str d15, [x9], #0x8\n"
+      "str d19, [x27], #0x8\n"
+      "str d23, [x25], #0x8\n"
+      "str d27, [x23], #0x8\n"
+      "tbz x16, #0, 168f\n"
+      "st1 { v11.s }[2], [x13]\n"
+      "st1 { v15.s }[2], [x9]\n"
+      "st1 { v19.s }[2], [x27]\n"
+      "st1 { v23.s }[2], [x25]\n"
+      "st1 { v27.s }[2], [x23]\n"
+      "b 168f\n"
+      "161:"  // Height 5: Partial direct writeback: partial_1_12
+      "tbz x16, #0, 168f\n"
+      "str s11, [x13, #0x0]\n"
+      "str s15, [x9, #0x0]\n"
+      "str s19, [x27, #0x0]\n"
+      "str s23, [x25, #0x0]\n"
+      "str s27, [x23, #0x0]\n"
+      "b 168f\n"
+      "162:"  // Height 5: Partial direct writeback: partial_2_8
+      "tbz x16, #1, 163f\n"
+      "str d10, [x13], #0x8\n"
+      "str d14, [x9], #0x8\n"
+      "str d18, [x27], #0x8\n"
+      "str d22, [x25], #0x8\n"
+      "str d26, [x23], #0x8\n"
+      "tbz x16, #0, 168f\n"
+      "st1 { v10.s }[2], [x13]\n"
+      "st1 { v14.s }[2], [x9]\n"
+      "st1 { v18.s }[2], [x27]\n"
+      "st1 { v22.s }[2], [x25]\n"
+      "st1 { v26.s }[2], [x23]\n"
+      "b 168f\n"
+      "163:"  // Height 5: Partial direct writeback: partial_1_8
+      "tbz x16, #0, 168f\n"
+      "str s10, [x13, #0x0]\n"
+      "str s14, [x9, #0x0]\n"
+      "str s18, [x27, #0x0]\n"
+      "str s22, [x25, #0x0]\n"
+      "str s26, [x23, #0x0]\n"
+      "b 168f\n"
+      "164:"  // Height 5: Partial direct writeback: partial_4_0
+      "tbz x16, #2, 166f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "st1 { v16.4s }, [x27], #0x10\n"
+      "st1 { v20.4s }, [x25], #0x10\n"
+      "st1 { v24.4s }, [x23], #0x10\n"
+      "tbz x16, #1, 165f\n"
+      "str d9, [x13], #0x8\n"
+      "str d13, [x9], #0x8\n"
+      "str d17, [x27], #0x8\n"
+      "str d21, [x25], #0x8\n"
+      "str d25, [x23], #0x8\n"
+      "tbz x16, #0, 168f\n"
+      "st1 { v9.s }[2], [x13]\n"
+      "st1 { v13.s }[2], [x9]\n"
+      "st1 { v17.s }[2], [x27]\n"
+      "st1 { v21.s }[2], [x25]\n"
+      "st1 { v25.s }[2], [x23]\n"
+      "b 168f\n"
+      "165:"  // Height 5: Partial direct writeback: partial_1_4
+      "tbz x16, #0, 168f\n"
+      "str s9, [x13, #0x0]\n"
+      "str s13, [x9, #0x0]\n"
+      "str s17, [x27, #0x0]\n"
+      "str s21, [x25, #0x0]\n"
+      "str s25, [x23, #0x0]\n"
+      "b 168f\n"
+      "166:"  // Height 5: Partial direct writeback: partial_2_0
+      "tbz x16, #1, 167f\n"
+      "str d8, [x13], #0x8\n"
+      "str d12, [x9], #0x8\n"
+      "str d16, [x27], #0x8\n"
+      "str d20, [x25], #0x8\n"
+      "str d24, [x23], #0x8\n"
+      "tbz x16, #0, 168f\n"
+      "st1 { v8.s }[2], [x13]\n"
+      "st1 { v12.s }[2], [x9]\n"
+      "st1 { v16.s }[2], [x27]\n"
+      "st1 { v20.s }[2], [x25]\n"
+      "st1 { v24.s }[2], [x23]\n"
+      "b 168f\n"
+      "167:"  // Height 5: Partial direct writeback: partial_1_0
+      "str s8, [x13, #0x0]\n"
+      "str s12, [x9, #0x0]\n"
+      "str s16, [x27, #0x0]\n"
+      "str s20, [x25, #0x0]\n"
+      "str s24, [x23, #0x0]\n"
+      "168:"  // Height 5: Partial direct writeback: Done
+      "b 170f\n"
+      "169:"  // Height 5: Full writeback
+      "str q8, [x13, #0x0]\n"
+      "str q9, [x13, #0x10]\n"
+      "str q10, [x13, #0x20]\n"
+      "str q11, [x13, #0x30]\n"
+      "str q12, [x9, #0x0]\n"
+      "str q13, [x9, #0x10]\n"
+      "str q14, [x9, #0x20]\n"
+      "str q15, [x9, #0x30]\n"
+      "str q16, [x27, #0x0]\n"
+      "str q17, [x27, #0x10]\n"
+      "str q18, [x27, #0x20]\n"
+      "str q19, [x27, #0x30]\n"
+      "str q20, [x25, #0x0]\n"
+      "str q21, [x25, #0x10]\n"
+      "str q22, [x25, #0x20]\n"
+      "str q23, [x25, #0x30]\n"
+      "str q24, [x23, #0x0]\n"
+      "str q25, [x23, #0x10]\n"
+      "str q26, [x23, #0x20]\n"
+      "str q27, [x23, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "add x9, x9, #0x40\n"
+      "add x27, x27, #0x40\n"
+      "add x25, x25, #0x40\n"
+      "add x23, x23, #0x40\n"
+      "170:"  // Height 5: Writeback done
+      "subs x16, x16, #0x10\n"
+      "bgt 139b\n"
+      "b 206f\n"
+      "171:"  // Height 6
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x14, %x[bias]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "tbz %x[flags], #2, 172f\n"
+      "ldr x13, [%x[output_ptr], #0x0]\n"
+      "add x13, x13, x19, LSL #2\n"
+      "ldr x9, [%x[output_ptr], #0x8]\n"
+      "ldr x27, [%x[output_ptr], #0x10]\n"
+      "add x9, x9, x19, LSL #2\n"
+      "ldr x25, [%x[output_ptr], #0x18]\n"
+      "ldr x23, [%x[output_ptr], #0x20]\n"
+      "add x27, x27, x19, LSL #2\n"
+      "ldr x21, [%x[output_ptr], #0x28]\n"
+      "add %x[output_ptr], %x[output_ptr], #0x30\n"
+      "add x25, x25, x19, LSL #2\n"
+      "add x23, x23, x19, LSL #2\n"
+      "add x21, x21, x19, LSL #2\n"
+      "b 173f\n"
+      "172:"  // Height 6: setup direct output
+      "mov x13, %x[output_ptr]\n"
+      "add x9, x13, x19, LSL #2\n"
+      "add x27, x9, x19, LSL #2\n"
+      "add x25, x27, x19, LSL #2\n"
+      "add x23, x25, x19, LSL #2\n"
+      "add x21, x23, x19, LSL #2\n"
+      "add %x[output_ptr], x21, x19, LSL #2\n"
+      "173:"  // Height 6: Column loop
+      "cbz x14, 174f\n"
+      "ldr q8, [x14, #0x0]\n"
+      "mov v12.16b, v8.16b\n"
+      "ldr q9, [x14, #0x10]\n"
+      "mov v16.16b, v8.16b\n"
+      "ldr q10, [x14, #0x20]\n"
+      "mov v20.16b, v8.16b\n"
+      "ldr q11, [x14, #0x30]\n"
+      "mov v24.16b, v8.16b\n"
+      "add x14, x14, #0x40\n"
+      "mov v28.16b, v8.16b\n"
+      "mov v13.16b, v9.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "mov v14.16b, v10.16b\n"
+      "mov v15.16b, v11.16b\n"
+      "mov v18.16b, v10.16b\n"
+      "mov v19.16b, v11.16b\n"
+      "mov v21.16b, v9.16b\n"
+      "mov v22.16b, v10.16b\n"
+      "mov v23.16b, v11.16b\n"
+      "mov v25.16b, v9.16b\n"
+      "mov v26.16b, v10.16b\n"
+      "mov v27.16b, v11.16b\n"
+      "mov v29.16b, v9.16b\n"
+      "mov v30.16b, v10.16b\n"
+      "mov v31.16b, v11.16b\n"
+      "b 185f\n"
+      "174:"  // Height 6: no bias
+      "tbz %x[flags], #0, 184f\n"
+      "cmp x16, #0x10\n"
+      "bge 183f\n"
+      "tbz x16, #3, 178f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "ld1 { v16.4s }, [x27], #0x10\n"
+      "ld1 { v20.4s }, [x25], #0x10\n"
+      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x9], #0x10\n"
+      "ld1 { v17.4s }, [x27], #0x10\n"
+      "ld1 { v21.4s }, [x25], #0x10\n"
+      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
+      "tbz x16, #2, 176f\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x9], #0x10\n"
+      "ld1 { v18.4s }, [x27], #0x10\n"
+      "ld1 { v22.4s }, [x25], #0x10\n"
+      "ld1 { v26.4s }, [x23], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
+      "tbz x16, #1, 175f\n"
+      "mov x19, #0x38\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d15, [x9], #0x8\n"
+      "ldr d19, [x27], #0x8\n"
+      "ldr d23, [x25], #0x8\n"
+      "ldr d27, [x23], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
+      "tbz x16, #0, 182f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x9]\n"
+      "ld1 { v19.s }[2], [x27]\n"
+      "ld1 { v23.s }[2], [x25]\n"
+      "ld1 { v27.s }[2], [x23]\n"
+      "ld1 { v31.s }[2], [x21]\n"
+      "b 182f\n"
+      "175:"  // Height 6: Partial accumulate: partial_1_12
+      "mov x19, #0x30\n"
+      "tbz x16, #0, 182f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s15, [x9, #0x0]\n"
+      "ldr s19, [x27, #0x0]\n"
+      "ldr s23, [x25, #0x0]\n"
+      "ldr s27, [x23, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
+      "b 182f\n"
+      "176:"  // Height 6: Partial accumulate: partial_2_8
+      "tbz x16, #1, 177f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d14, [x9], #0x8\n"
+      "ldr d18, [x27], #0x8\n"
+      "ldr d22, [x25], #0x8\n"
+      "ldr d26, [x23], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
+      "mov x19, #0x28\n"
+      "tbz x16, #0, 182f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x9]\n"
+      "ld1 { v18.s }[2], [x27]\n"
+      "ld1 { v22.s }[2], [x25]\n"
+      "ld1 { v26.s }[2], [x23]\n"
+      "ld1 { v30.s }[2], [x21]\n"
+      "b 182f\n"
+      "177:"  // Height 6: Partial accumulate: partial_1_8
+      "mov x19, #0x20\n"
+      "tbz x16, #0, 182f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s14, [x9, #0x0]\n"
+      "ldr s18, [x27, #0x0]\n"
+      "ldr s22, [x25, #0x0]\n"
+      "ldr s26, [x23, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
+      "b 182f\n"
+      "178:"  // Height 6: Partial accumulate: partial_4_0
+      "tbz x16, #2, 180f\n"
+      "ld1 { v8.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x9], #0x10\n"
+      "ld1 { v16.4s }, [x27], #0x10\n"
+      "ld1 { v20.4s }, [x25], #0x10\n"
+      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
+      "tbz x16, #1, 179f\n"
+      "mov x19, #0x18\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d13, [x9], #0x8\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d21, [x25], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
+      "tbz x16, #0, 182f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x9]\n"
+      "ld1 { v17.s }[2], [x27]\n"
+      "ld1 { v21.s }[2], [x25]\n"
+      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v29.s }[2], [x21]\n"
+      "b 182f\n"
+      "179:"  // Height 6: Partial accumulate: partial_1_4
+      "mov x19, #0x10\n"
+      "tbz x16, #0, 182f\n"
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s13, [x9, #0x0]\n"
+      "ldr s17, [x27, #0x0]\n"
+      "ldr s21, [x25, #0x0]\n"
+      "ldr s25, [x23, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
+      "b 182f\n"
+      "180:"  // Height 6: Partial accumulate: partial_2_0
+      "tbz x16, #1, 181f\n"
+      "ldr d8, [x13], #0x8\n"
+      "ldr d12, [x9], #0x8\n"
+      "ldr d16, [x27], #0x8\n"
+      "ldr d20, [x25], #0x8\n"
+      "ldr d24, [x23], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
+      "mov x19, #0x8\n"
+      "tbz x16, #0, 182f\n"
+      "ld1 { v8.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x9]\n"
+      "ld1 { v16.s }[2], [x27]\n"
+      "ld1 { v20.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x23]\n"
+      "ld1 { v28.s }[2], [x21]\n"
+      "b 182f\n"
+      "181:"  // Height 6: Partial accumulate: partial_1_0
+      "mov x19, #0x0\n"
+      "ldr s8, [x13, #0x0]\n"
+      "ldr s12, [x9, #0x0]\n"
+      "ldr s16, [x27, #0x0]\n"
+      "ldr s20, [x25, #0x0]\n"
+      "ldr s24, [x23, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
+      "182:"  // Height 6: Partial accumulate: Done
+      "sub x13, x13, x19\n"
+      "sub x9, x9, x19\n"
+      "sub x27, x27, x19\n"
+      "sub x25, x25, x19\n"
+      "sub x23, x23, x19\n"
+      "sub x21, x21, x19\n"
+      "b 185f\n"
+      "183:"  // Height 6: full accumulate
+      "ldr q8, [x13, #0x0]\n"
+      "ldr q9, [x13, #0x10]\n"
+      "ldr q10, [x13, #0x20]\n"
+      "ldr q11, [x13, #0x30]\n"
+      "ldr q12, [x9, #0x0]\n"
+      "ldr q13, [x9, #0x10]\n"
+      "ldr q14, [x9, #0x20]\n"
+      "ldr q15, [x9, #0x30]\n"
+      "ldr q16, [x27, #0x0]\n"
+      "ldr q17, [x27, #0x10]\n"
+      "ldr q18, [x27, #0x20]\n"
+      "ldr q19, [x27, #0x30]\n"
+      "ldr q20, [x25, #0x0]\n"
+      "ldr q21, [x25, #0x10]\n"
+      "ldr q22, [x25, #0x20]\n"
+      "ldr q23, [x25, #0x30]\n"
+      "ldr q24, [x23, #0x0]\n"
+      "ldr q25, [x23, #0x10]\n"
+      "ldr q26, [x23, #0x20]\n"
+      "ldr q27, [x23, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
+      "b 185f\n"
+      "184:"  // Height 6: no accumulate
+      "movi v8.16b, #0x0\n"
+      "movi v9.16b, #0x0\n"
+      "movi v10.16b, #0x0\n"
+      "movi v11.16b, #0x0\n"
+      "movi v12.16b, #0x0\n"
+      "movi v13.16b, #0x0\n"
+      "movi v14.16b, #0x0\n"
+      "movi v15.16b, #0x0\n"
+      "movi v16.16b, #0x0\n"
+      "movi v17.16b, #0x0\n"
+      "movi v18.16b, #0x0\n"
+      "movi v19.16b, #0x0\n"
+      "movi v20.16b, #0x0\n"
+      "movi v21.16b, #0x0\n"
+      "movi v22.16b, #0x0\n"
+      "movi v23.16b, #0x0\n"
+      "movi v24.16b, #0x0\n"
+      "movi v25.16b, #0x0\n"
+      "movi v26.16b, #0x0\n"
+      "movi v27.16b, #0x0\n"
+      "movi v28.16b, #0x0\n"
+      "movi v29.16b, #0x0\n"
+      "movi v30.16b, #0x0\n"
+      "movi v31.16b, #0x0\n"
+      "185:"  // Height 6: setup done
+      "mov x12, #0x0\n"
+      "186:"  // Height 6: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w11, [x20, x12, LSL #0x2]\n"
+      "tbz %x[flags], #3, 187f\n"
+      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "add x20, x20, x19, LSL #3\n"
+      "ldr x10, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "ldr x26, [x20, #0x10]\n"
+      "ldr x24, [x20, #0x18]\n"
+      "ldr x22, [x20, #0x20]\n"
+      "ldr x20, [x20, #0x28]\n"
+      "cbnz x12, 188f\n"
+      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x10, x10, x19, LSL #2\n"
+      "add x28, x28, x19, LSL #2\n"
+      "add x26, x26, x19, LSL #2\n"
+      "add x24, x24, x19, LSL #2\n"
+      "add x22, x22, x19, LSL #2\n"
+      "add x20, x20, x19, LSL #2\n"
+      "b 188f\n"
+      "187:"  // Height 6: setup direct input
+      "mov x10, %x[input_ptr]\n"
+      "add x28, x10, x19, LSL #2\n"
+      "add x26, x28, x19, LSL #2\n"
+      "add x24, x26, x19, LSL #2\n"
+      "add x22, x24, x19, LSL #2\n"
+      "add x20, x22, x19, LSL #2\n"
+      "188:"  // Height 6: input setup done
+      "cmp x11, #0x4\n"
+      "blt 191f\n"
+      "cmp x11, #0x8\n"
+      "blt 190f\n"
+      "189:"  // Height 6: Multiply loop: Main loop head
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x26, #0x0]\n"
+      "ldr q3, [x24, #0x0]\n"
+      "ldr q4, [x22, #0x0]\n"
+      "ldr q5, [x20, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "add x10, x10, #0x10\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "add x28, x28, #0x10\n"
+      "fmla v20.4s, v6.4s, v3.s[0]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "add x26, x26, #0x10\n"
+      "fmla v24.4s, v6.4s, v4.s[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x24, x24, #0x10\n"
+      "fmla v28.4s, v6.4s, v5.s[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "add x22, x22, #0x10\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "add x20, x20, #0x10\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "prfm pldl1keep, [x20, #0x80]\n"
+      "sub x11, x11, #0x4\n"
+      "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "cmp x11, #0x8\n"
+      "fmla v25.4s, v7.4s, v4.s[0]\n"
+      "fmla v29.4s, v7.4s, v5.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v22.4s, v6.4s, v3.s[0]\n"
+      "fmla v26.4s, v6.4s, v4.s[0]\n"
+      "fmla v30.4s, v6.4s, v5.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "fmla v23.4s, v7.4s, v3.s[0]\n"
+      "fmla v27.4s, v7.4s, v4.s[0]\n"
+      "fmla v31.4s, v7.4s, v5.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "fmla v16.4s, v6.4s, v2.s[1]\n"
+      "fmla v20.4s, v6.4s, v3.s[1]\n"
+      "fmla v24.4s, v6.4s, v4.s[1]\n"
+      "fmla v28.4s, v6.4s, v5.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "fmla v17.4s, v7.4s, v2.s[1]\n"
+      "fmla v21.4s, v7.4s, v3.s[1]\n"
+      "fmla v25.4s, v7.4s, v4.s[1]\n"
+      "fmla v29.4s, v7.4s, v5.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "fmla v18.4s, v6.4s, v2.s[1]\n"
+      "fmla v22.4s, v6.4s, v3.s[1]\n"
+      "fmla v26.4s, v6.4s, v4.s[1]\n"
+      "fmla v30.4s, v6.4s, v5.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "fmla v19.4s, v7.4s, v2.s[1]\n"
+      "fmla v23.4s, v7.4s, v3.s[1]\n"
+      "fmla v27.4s, v7.4s, v4.s[1]\n"
+      "fmla v31.4s, v7.4s, v5.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "fmla v16.4s, v6.4s, v2.s[2]\n"
+      "fmla v20.4s, v6.4s, v3.s[2]\n"
+      "fmla v24.4s, v6.4s, v4.s[2]\n"
+      "fmla v28.4s, v6.4s, v5.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "fmla v17.4s, v7.4s, v2.s[2]\n"
+      "fmla v21.4s, v7.4s, v3.s[2]\n"
+      "fmla v25.4s, v7.4s, v4.s[2]\n"
+      "fmla v29.4s, v7.4s, v5.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "fmla v18.4s, v6.4s, v2.s[2]\n"
+      "fmla v22.4s, v6.4s, v3.s[2]\n"
+      "fmla v26.4s, v6.4s, v4.s[2]\n"
+      "fmla v30.4s, v6.4s, v5.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "fmla v19.4s, v7.4s, v2.s[2]\n"
+      "fmla v23.4s, v7.4s, v3.s[2]\n"
+      "fmla v27.4s, v7.4s, v4.s[2]\n"
+      "fmla v31.4s, v7.4s, v5.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "fmla v16.4s, v6.4s, v2.s[3]\n"
+      "fmla v20.4s, v6.4s, v3.s[3]\n"
+      "fmla v24.4s, v6.4s, v4.s[3]\n"
+      "fmla v28.4s, v6.4s, v5.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "fmla v17.4s, v7.4s, v2.s[3]\n"
+      "fmla v21.4s, v7.4s, v3.s[3]\n"
+      "fmla v25.4s, v7.4s, v4.s[3]\n"
+      "fmla v29.4s, v7.4s, v5.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v18.4s, v6.4s, v2.s[3]\n"
+      "fmla v22.4s, v6.4s, v3.s[3]\n"
+      "fmla v26.4s, v6.4s, v4.s[3]\n"
+      "fmla v30.4s, v6.4s, v5.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "fmla v19.4s, v7.4s, v2.s[3]\n"
+      "fmla v23.4s, v7.4s, v3.s[3]\n"
+      "fmla v27.4s, v7.4s, v4.s[3]\n"
+      "fmla v31.4s, v7.4s, v5.s[3]\n"
+      "bge 189b\n"
+      "190:"  // Height 6: Multiply loop: Single iteration only
+      "sub x11, x11, #0x4\n"
+      "ldr q0, [x10, #0x0]\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x26, #0x0]\n"
+      "ldr q3, [x24, #0x0]\n"
+      "ldr q4, [x22, #0x0]\n"
+      "ldr q5, [x20, #0x0]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "add x10, x10, #0x10\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "add x28, x28, #0x10\n"
+      "fmla v20.4s, v6.4s, v3.s[0]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "add x26, x26, #0x10\n"
+      "fmla v24.4s, v6.4s, v4.s[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x24, x24, #0x10\n"
+      "fmla v28.4s, v6.4s, v5.s[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "add x22, x22, #0x10\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "add x20, x20, #0x10\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "prfm pldl1keep, [x20, #0x80]\n"
+      "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "fmla v25.4s, v7.4s, v4.s[0]\n"
+      "fmla v29.4s, v7.4s, v5.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v22.4s, v6.4s, v3.s[0]\n"
+      "fmla v26.4s, v6.4s, v4.s[0]\n"
+      "fmla v30.4s, v6.4s, v5.s[0]\n"
+      "ldr q6, [x15, #0x40]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "fmla v23.4s, v7.4s, v3.s[0]\n"
+      "fmla v27.4s, v7.4s, v4.s[0]\n"
+      "fmla v31.4s, v7.4s, v5.s[0]\n"
+      "ldr q7, [x15, #0x50]\n"
+      "fmla v8.4s, v6.4s, v0.s[1]\n"
+      "fmla v12.4s, v6.4s, v1.s[1]\n"
+      "fmla v16.4s, v6.4s, v2.s[1]\n"
+      "fmla v20.4s, v6.4s, v3.s[1]\n"
+      "fmla v24.4s, v6.4s, v4.s[1]\n"
+      "fmla v28.4s, v6.4s, v5.s[1]\n"
+      "ldr q6, [x15, #0x60]\n"
+      "fmla v9.4s, v7.4s, v0.s[1]\n"
+      "fmla v13.4s, v7.4s, v1.s[1]\n"
+      "fmla v17.4s, v7.4s, v2.s[1]\n"
+      "fmla v21.4s, v7.4s, v3.s[1]\n"
+      "fmla v25.4s, v7.4s, v4.s[1]\n"
+      "fmla v29.4s, v7.4s, v5.s[1]\n"
+      "ldr q7, [x15, #0x70]\n"
+      "fmla v10.4s, v6.4s, v0.s[1]\n"
+      "fmla v14.4s, v6.4s, v1.s[1]\n"
+      "fmla v18.4s, v6.4s, v2.s[1]\n"
+      "fmla v22.4s, v6.4s, v3.s[1]\n"
+      "fmla v26.4s, v6.4s, v4.s[1]\n"
+      "fmla v30.4s, v6.4s, v5.s[1]\n"
+      "ldr q6, [x15, #0x80]\n"
+      "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "fmla v19.4s, v7.4s, v2.s[1]\n"
+      "fmla v23.4s, v7.4s, v3.s[1]\n"
+      "fmla v27.4s, v7.4s, v4.s[1]\n"
+      "fmla v31.4s, v7.4s, v5.s[1]\n"
+      "ldr q7, [x15, #0x90]\n"
+      "fmla v8.4s, v6.4s, v0.s[2]\n"
+      "fmla v12.4s, v6.4s, v1.s[2]\n"
+      "fmla v16.4s, v6.4s, v2.s[2]\n"
+      "fmla v20.4s, v6.4s, v3.s[2]\n"
+      "fmla v24.4s, v6.4s, v4.s[2]\n"
+      "fmla v28.4s, v6.4s, v5.s[2]\n"
+      "ldr q6, [x15, #0xa0]\n"
+      "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "fmla v17.4s, v7.4s, v2.s[2]\n"
+      "fmla v21.4s, v7.4s, v3.s[2]\n"
+      "fmla v25.4s, v7.4s, v4.s[2]\n"
+      "fmla v29.4s, v7.4s, v5.s[2]\n"
+      "ldr q7, [x15, #0xb0]\n"
+      "fmla v10.4s, v6.4s, v0.s[2]\n"
+      "fmla v14.4s, v6.4s, v1.s[2]\n"
+      "fmla v18.4s, v6.4s, v2.s[2]\n"
+      "fmla v22.4s, v6.4s, v3.s[2]\n"
+      "fmla v26.4s, v6.4s, v4.s[2]\n"
+      "fmla v30.4s, v6.4s, v5.s[2]\n"
+      "ldr q6, [x15, #0xc0]\n"
+      "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "fmla v19.4s, v7.4s, v2.s[2]\n"
+      "fmla v23.4s, v7.4s, v3.s[2]\n"
+      "fmla v27.4s, v7.4s, v4.s[2]\n"
+      "fmla v31.4s, v7.4s, v5.s[2]\n"
+      "ldr q7, [x15, #0xd0]\n"
+      "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "fmla v12.4s, v6.4s, v1.s[3]\n"
+      "fmla v16.4s, v6.4s, v2.s[3]\n"
+      "fmla v20.4s, v6.4s, v3.s[3]\n"
+      "fmla v24.4s, v6.4s, v4.s[3]\n"
+      "fmla v28.4s, v6.4s, v5.s[3]\n"
+      "ldr q6, [x15, #0xe0]\n"
+      "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "fmla v13.4s, v7.4s, v1.s[3]\n"
+      "fmla v17.4s, v7.4s, v2.s[3]\n"
+      "fmla v21.4s, v7.4s, v3.s[3]\n"
+      "fmla v25.4s, v7.4s, v4.s[3]\n"
+      "fmla v29.4s, v7.4s, v5.s[3]\n"
+      "ldr q7, [x15, #0xf0]\n"
+      "fmla v10.4s, v6.4s, v0.s[3]\n"
+      "add x15, x15, #0x100\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "fmla v18.4s, v6.4s, v2.s[3]\n"
+      "fmla v22.4s, v6.4s, v3.s[3]\n"
+      "fmla v26.4s, v6.4s, v4.s[3]\n"
+      "fmla v30.4s, v6.4s, v5.s[3]\n"
+      "fmla v11.4s, v7.4s, v0.s[3]\n"
+      "fmla v15.4s, v7.4s, v1.s[3]\n"
+      "fmla v19.4s, v7.4s, v2.s[3]\n"
+      "fmla v23.4s, v7.4s, v3.s[3]\n"
+      "fmla v27.4s, v7.4s, v4.s[3]\n"
+      "fmla v31.4s, v7.4s, v5.s[3]\n"
+      "191:"  // Height 6: Multiply loop: Main loop skip
+      "cbz x11, 193f\n"
+      "192:"  // Height 6: Multiply loop: Odd block loop
+      "ldr s0, [x10], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
+      "ldr s2, [x26], #0x4\n"
+      "ldr s3, [x24], #0x4\n"
+      "ldr s4, [x22], #0x4\n"
+      "ldr s5, [x20], #0x4\n"
+      "ldr q6, [x15, #0x0]\n"
+      "fmla v8.4s, v6.4s, v0.s[0]\n"
+      "ldr q7, [x15, #0x10]\n"
+      "fmla v12.4s, v6.4s, v1.s[0]\n"
+      "sub x11, x11, #0x1\n"
+      "fmla v16.4s, v6.4s, v2.s[0]\n"
+      "fmla v20.4s, v6.4s, v3.s[0]\n"
+      "fmla v24.4s, v6.4s, v4.s[0]\n"
+      "fmla v28.4s, v6.4s, v5.s[0]\n"
+      "ldr q6, [x15, #0x20]\n"
+      "fmla v9.4s, v7.4s, v0.s[0]\n"
+      "fmla v13.4s, v7.4s, v1.s[0]\n"
+      "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "fmla v25.4s, v7.4s, v4.s[0]\n"
+      "fmla v29.4s, v7.4s, v5.s[0]\n"
+      "ldr q7, [x15, #0x30]\n"
+      "fmla v10.4s, v6.4s, v0.s[0]\n"
+      "add x15, x15, #0x40\n"
+      "fmla v14.4s, v6.4s, v1.s[0]\n"
+      "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "fmla v22.4s, v6.4s, v3.s[0]\n"
+      "fmla v26.4s, v6.4s, v4.s[0]\n"
+      "fmla v30.4s, v6.4s, v5.s[0]\n"
+      "fmla v11.4s, v7.4s, v0.s[0]\n"
+      "fmla v15.4s, v7.4s, v1.s[0]\n"
+      "fmla v19.4s, v7.4s, v2.s[0]\n"
+      "fmla v23.4s, v7.4s, v3.s[0]\n"
+      "fmla v27.4s, v7.4s, v4.s[0]\n"
+      "fmla v31.4s, v7.4s, v5.s[0]\n"
+      "cbnz x11, 192b\n"
+      "193:"  // Height 6: Multiply loop: No odd multiplies
+      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x19\n"
+      "bne 186b\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
+      "tbz %x[flags], #1, 194f\n"
+      "add x19, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v1.4s }, [x19]\n"
+      "add x19, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v0.4s }, [x19]\n"
+      "fmin v8.4s, v8.4s, v0.4s\n"
+      "fmin v9.4s, v9.4s, v0.4s\n"
+      "fmin v10.4s, v10.4s, v0.4s\n"
+      "fmin v11.4s, v11.4s, v0.4s\n"
+      "fmax v8.4s, v8.4s, v1.4s\n"
+      "fmax v9.4s, v9.4s, v1.4s\n"
+      "fmax v10.4s, v10.4s, v1.4s\n"
+      "fmax v11.4s, v11.4s, v1.4s\n"
+      "fmin v12.4s, v12.4s, v0.4s\n"
+      "fmin v13.4s, v13.4s, v0.4s\n"
+      "fmin v14.4s, v14.4s, v0.4s\n"
+      "fmax v12.4s, v12.4s, v1.4s\n"
+      "fmax v13.4s, v13.4s, v1.4s\n"
+      "fmax v14.4s, v14.4s, v1.4s\n"
+      "fmin v15.4s, v15.4s, v0.4s\n"
+      "fmin v16.4s, v16.4s, v0.4s\n"
+      "fmin v17.4s, v17.4s, v0.4s\n"
+      "fmax v15.4s, v15.4s, v1.4s\n"
+      "fmax v16.4s, v16.4s, v1.4s\n"
+      "fmax v17.4s, v17.4s, v1.4s\n"
+      "fmin v18.4s, v18.4s, v0.4s\n"
+      "fmin v19.4s, v19.4s, v0.4s\n"
+      "fmin v20.4s, v20.4s, v0.4s\n"
+      "fmax v18.4s, v18.4s, v1.4s\n"
+      "fmax v19.4s, v19.4s, v1.4s\n"
+      "fmax v20.4s, v20.4s, v1.4s\n"
+      "fmin v21.4s, v21.4s, v0.4s\n"
+      "fmin v22.4s, v22.4s, v0.4s\n"
+      "fmin v23.4s, v23.4s, v0.4s\n"
+      "fmax v21.4s, v21.4s, v1.4s\n"
+      "fmax v22.4s, v22.4s, v1.4s\n"
+      "fmax v23.4s, v23.4s, v1.4s\n"
+      "fmin v24.4s, v24.4s, v0.4s\n"
+      "fmin v25.4s, v25.4s, v0.4s\n"
+      "fmin v26.4s, v26.4s, v0.4s\n"
+      "fmax v24.4s, v24.4s, v1.4s\n"
+      "fmax v25.4s, v25.4s, v1.4s\n"
+      "fmax v26.4s, v26.4s, v1.4s\n"
+      "fmin v27.4s, v27.4s, v0.4s\n"
+      "fmin v28.4s, v28.4s, v0.4s\n"
+      "fmin v29.4s, v29.4s, v0.4s\n"
+      "fmax v27.4s, v27.4s, v1.4s\n"
+      "fmax v28.4s, v28.4s, v1.4s\n"
+      "fmax v29.4s, v29.4s, v1.4s\n"
+      "fmin v30.4s, v30.4s, v0.4s\n"
+      "fmin v31.4s, v31.4s, v0.4s\n"
+      "fmax v30.4s, v30.4s, v1.4s\n"
+      "fmax v31.4s, v31.4s, v1.4s\n"
+      "194:"  // Height 6: No activation
+      "cmp x16, #0x10\n"
+      "bge 203f\n"
+      "tbz x16, #3, 198f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v9.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "st1 { v13.4s }, [x9], #0x10\n"
+      "st1 { v16.4s }, [x27], #0x10\n"
+      "st1 { v17.4s }, [x27], #0x10\n"
+      "st1 { v20.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x25], #0x10\n"
+      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v25.4s }, [x23], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
+      "st1 { v29.4s }, [x21], #0x10\n"
+      "tbz x16, #2, 196f\n"
+      "st1 { v10.4s }, [x13], #0x10\n"
+      "st1 { v14.4s }, [x9], #0x10\n"
+      "st1 { v18.4s }, [x27], #0x10\n"
+      "st1 { v22.4s }, [x25], #0x10\n"
+      "st1 { v26.4s }, [x23], #0x10\n"
+      "st1 { v30.4s }, [x21], #0x10\n"
+      "tbz x16, #1, 195f\n"
+      "str d11, [x13], #0x8\n"
+      "str d15, [x9], #0x8\n"
+      "str d19, [x27], #0x8\n"
+      "str d23, [x25], #0x8\n"
+      "str d27, [x23], #0x8\n"
+      "str d31, [x21], #0x8\n"
+      "tbz x16, #0, 202f\n"
+      "st1 { v11.s }[2], [x13]\n"
+      "st1 { v15.s }[2], [x9]\n"
+      "st1 { v19.s }[2], [x27]\n"
+      "st1 { v23.s }[2], [x25]\n"
+      "st1 { v27.s }[2], [x23]\n"
+      "st1 { v31.s }[2], [x21]\n"
+      "b 202f\n"
+      "195:"  // Height 6: Partial direct writeback: partial_1_12
+      "tbz x16, #0, 202f\n"
+      "str s11, [x13, #0x0]\n"
+      "str s15, [x9, #0x0]\n"
+      "str s19, [x27, #0x0]\n"
+      "str s23, [x25, #0x0]\n"
+      "str s27, [x23, #0x0]\n"
+      "str s31, [x21, #0x0]\n"
+      "b 202f\n"
+      "196:"  // Height 6: Partial direct writeback: partial_2_8
+      "tbz x16, #1, 197f\n"
+      "str d10, [x13], #0x8\n"
+      "str d14, [x9], #0x8\n"
+      "str d18, [x27], #0x8\n"
+      "str d22, [x25], #0x8\n"
+      "str d26, [x23], #0x8\n"
+      "str d30, [x21], #0x8\n"
+      "tbz x16, #0, 202f\n"
+      "st1 { v10.s }[2], [x13]\n"
+      "st1 { v14.s }[2], [x9]\n"
+      "st1 { v18.s }[2], [x27]\n"
+      "st1 { v22.s }[2], [x25]\n"
+      "st1 { v26.s }[2], [x23]\n"
+      "st1 { v30.s }[2], [x21]\n"
+      "b 202f\n"
+      "197:"  // Height 6: Partial direct writeback: partial_1_8
+      "tbz x16, #0, 202f\n"
+      "str s10, [x13, #0x0]\n"
+      "str s14, [x9, #0x0]\n"
+      "str s18, [x27, #0x0]\n"
+      "str s22, [x25, #0x0]\n"
+      "str s26, [x23, #0x0]\n"
+      "str s30, [x21, #0x0]\n"
+      "b 202f\n"
+      "198:"  // Height 6: Partial direct writeback: partial_4_0
+      "tbz x16, #2, 200f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x9], #0x10\n"
+      "st1 { v16.4s }, [x27], #0x10\n"
+      "st1 { v20.4s }, [x25], #0x10\n"
+      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
+      "tbz x16, #1, 199f\n"
+      "str d9, [x13], #0x8\n"
+      "str d13, [x9], #0x8\n"
+      "str d17, [x27], #0x8\n"
+      "str d21, [x25], #0x8\n"
+      "str d25, [x23], #0x8\n"
+      "str d29, [x21], #0x8\n"
+      "tbz x16, #0, 202f\n"
+      "st1 { v9.s }[2], [x13]\n"
+      "st1 { v13.s }[2], [x9]\n"
+      "st1 { v17.s }[2], [x27]\n"
+      "st1 { v21.s }[2], [x25]\n"
+      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v29.s }[2], [x21]\n"
+      "b 202f\n"
+      "199:"  // Height 6: Partial direct writeback: partial_1_4
+      "tbz x16, #0, 202f\n"
+      "str s9, [x13, #0x0]\n"
+      "str s13, [x9, #0x0]\n"
+      "str s17, [x27, #0x0]\n"
+      "str s21, [x25, #0x0]\n"
+      "str s25, [x23, #0x0]\n"
+      "str s29, [x21, #0x0]\n"
+      "b 202f\n"
+      "200:"  // Height 6: Partial direct writeback: partial_2_0
+      "tbz x16, #1, 201f\n"
+      "str d8, [x13], #0x8\n"
+      "str d12, [x9], #0x8\n"
+      "str d16, [x27], #0x8\n"
+      "str d20, [x25], #0x8\n"
+      "str d24, [x23], #0x8\n"
+      "str d28, [x21], #0x8\n"
+      "tbz x16, #0, 202f\n"
+      "st1 { v8.s }[2], [x13]\n"
+      "st1 { v12.s }[2], [x9]\n"
+      "st1 { v16.s }[2], [x27]\n"
+      "st1 { v20.s }[2], [x25]\n"
+      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v28.s }[2], [x21]\n"
+      "b 202f\n"
+      "201:"  // Height 6: Partial direct writeback: partial_1_0
+      "str s8, [x13, #0x0]\n"
+      "str s12, [x9, #0x0]\n"
+      "str s16, [x27, #0x0]\n"
+      "str s20, [x25, #0x0]\n"
+      "str s24, [x23, #0x0]\n"
+      "str s28, [x21, #0x0]\n"
+      "202:"  // Height 6: Partial direct writeback: Done
+      "b 204f\n"
+      "203:"  // Height 6: Full writeback
+      "str q8, [x13, #0x0]\n"
+      "str q9, [x13, #0x10]\n"
+      "str q10, [x13, #0x20]\n"
+      "str q11, [x13, #0x30]\n"
+      "str q12, [x9, #0x0]\n"
+      "str q13, [x9, #0x10]\n"
+      "str q14, [x9, #0x20]\n"
+      "str q15, [x9, #0x30]\n"
+      "str q16, [x27, #0x0]\n"
+      "str q17, [x27, #0x10]\n"
+      "str q18, [x27, #0x20]\n"
+      "str q19, [x27, #0x30]\n"
+      "str q20, [x25, #0x0]\n"
+      "str q21, [x25, #0x10]\n"
+      "str q22, [x25, #0x20]\n"
+      "str q23, [x25, #0x30]\n"
+      "str q24, [x23, #0x0]\n"
+      "str q25, [x23, #0x10]\n"
+      "str q26, [x23, #0x20]\n"
+      "str q27, [x23, #0x30]\n"
+      "str q28, [x21, #0x0]\n"
+      "str q29, [x21, #0x10]\n"
+      "str q30, [x21, #0x20]\n"
+      "str q31, [x21, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "add x9, x9, #0x40\n"
+      "add x27, x27, #0x40\n"
+      "add x25, x25, #0x40\n"
+      "add x23, x23, #0x40\n"
+      "add x21, x21, #0x40\n"
+      "204:"  // Height 6: Writeback done
+      "subs x16, x16, #0x10\n"
+      "bgt 173b\n"
+      "subs %x[M], %x[M], #0x6\n"
+      "beq 206f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "tbz %x[flags], #3, 205f\n"
+      "add x20, x20, #0x6\n"
+      "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "b 1b\n"
+      "205:"  // Update direct input
+      "mov x19, #0x18\n"
+      "madd %x[input_ptr], x19, x20, %x[input_ptr]\n"
+      "b 1b\n"
+      "206:"  // Exit
+
+      : [M] "+r" (M), [input_ptr] "+r" (input_ptr), [output_ptr] "+r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, 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_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    );
+}
+
+} // namespace arm_gemm
+#endif // __aarch64__
-- 
cgit v1.2.1