diff --git a/Android.bp b/Android.bp
index bb04864..c84e9e5 100644
--- a/Android.bp
+++ b/Android.bp
@@ -1214,6 +1214,7 @@
                 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp",
                 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp",
                 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
+                "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
                 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp",
                 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp",
                 "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp",
diff --git a/filelist.json b/filelist.json
index ab7f16b..06a0913 100644
--- a/filelist.json
+++ b/filelist.json
@@ -1695,6 +1695,7 @@
               "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp",
diff --git a/src/BUILD.bazel b/src/BUILD.bazel
index d4a3b61..dd19f38 100644
--- a/src/BUILD.bazel
+++ b/src/BUILD.bazel
@@ -524,6 +524,7 @@
 	"core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp",
 	"core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp",
 	"core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
+	"core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
 	"core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp",
 	"core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp",
 	"core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp",
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index c641071..1a085a0 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -515,6 +515,7 @@
 	core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp
 	core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp
 	core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp
+	core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp
 	core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp
 	core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp
 	core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
index 44a7bb8..e85dd59 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2023 Arm Limited.
+ * Copyright (c) 2017-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -34,6 +34,7 @@
 #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
 #include "kernels/a64_ffhybrid_fp32_mla_6x16.hpp"
 #include "kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp"
+#include "kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp"
 #include "kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp"
 #include "kernels/a64_ffinterleaved_fp32_mla_8x12.hpp"
 #endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
@@ -350,6 +351,14 @@
     [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_4x24, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_4x24, float, float>(args); }
 ),
+GemmImplementation<float, float>::with_estimate(
+    GemmMethod::GEMM_HYBRID,
+    "a64_ffhybrid_fp32bf16fp32_mmla_6x16",
+    KernelWeightFormat::VL256_BL64_BF16,
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); },
+    [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_6x16, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_6x16, float, float>(args); }
+),
 #endif // BF16
 GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp
index 923d008..ac3cbf9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022-2023 Arm Limited.
+ * Copyright (c) 2022-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -88,8 +88,10 @@
     {
         if (std::is_same<T, float>::value) {
             switch (ci->get_cpu_model()) {
+                case CPUModel::V1:
+                    return { 23.64 };
                 default:
-                    return { 28.48 };
+                    return { 16.89 };
             }
         }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp
new file mode 100644
index 0000000..98f7fc9
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2024 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.
+ */
+#pragma once
+#ifdef __aarch64__
+
+#include "../std_transforms_fixed.hpp"
+#include "../bfloat.hpp"
+#include "../kernel_weight_format.hpp"
+#include "../performance_parameters.hpp"
+
+#define ARGLIST  \
+    unsigned int, const unsigned int *, \
+    IndirectInputArg<float>, \
+    size_t, size_t, \
+    const bfloat16 *, \
+    size_t, \
+    IndirectOutputArg<float>, \
+    const float *, Activation, bool
+
+namespace arm_gemm
+{
+// Actual kernel implementations
+void a64_ffhybrid_fp32bf16fp32_mmla_6x16( ARGLIST );
+
+class cls_a64_ffhybrid_fp32bf16fp32_mmla_6x16
+{
+public:
+    typedef float lhs_operand_type;
+    typedef bfloat16 rhs_operand_type;
+    typedef float result_type;
+
+    typedef void (*kern_type)( ARGLIST );
+
+    /* Kernel blocking parameters */
+    static constexpr unsigned int out_height()
+    {
+        return 6;
+    }
+    static unsigned int stripe_width()
+    {
+        return 4;
+    }
+
+    static KernelWeightFormat kernel_weight_format()
+    {
+        return KernelWeightFormat::VL256_BL64_BF16;
+    }
+
+    static unsigned int out_width()
+    {
+        return 16;
+    }
+
+    static constexpr unsigned int k_unroll()
+    {
+        return 4;
+    }
+
+    static constexpr bool supports_accumulate()
+    {
+        return true;
+    }
+
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 4> transforms = {};
+    template<typename T>
+    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
+    {
+        if (std::is_same<T, float>::value) {
+            switch (ci->get_cpu_model()) {
+                case CPUModel::V1:
+                    return { 21.05 };
+                default:
+                    return { 15.27 };
+            }
+        }
+
+        return { 1.0 };
+    }
+
+    // Default to the generic kernel
+    kern_type kernel=a64_ffhybrid_fp32bf16fp32_mmla_6x16;
+    cls_a64_ffhybrid_fp32bf16fp32_mmla_6x16(const CPUInfo *)
+    {
+    }
+};
+
+} // namespace arm_gemm
+
+#undef ARGLIST
+#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp
new file mode 100644
index 0000000..0bf0d2a
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp
@@ -0,0 +1,3240 @@
+/*
+ * Copyright (c) 2024 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 "../../bfloat.hpp"
+
+#include <cassert>
+#include <limits>
+
+namespace arm_gemm {
+
+void a64_ffhybrid_fp32bf16fp32_mmla_6x16 (
+    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<float> A_arg,
+    size_t M, size_t N, const bfloat16 *B_ptr, size_t B_stride, 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 bfloat16 *B_ptr = {};
+        const bfloat16 *cur_B_ptr = {};
+        size_t B_stride = {};
+        size_t output_offset = {};
+        size_t input_initial_col = {};
+        size_t input_offset = {};
+        void *output_ptr;
+        const float *bias;
+    } ka;
+
+    unsigned long flags=0;
+    void *input_ptr;
+
+    if (output_arg.is_indirect) {
+        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        ka.output_offset=output_arg.indirect.offset;
+        flags |= 0x4;
+    } else {
+        ka.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;
+    ka.bias = bias;
+    ka.B_stride = B_stride;
+    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 181f\n"
+      "cmp %x[M], #0x4\n"
+      "bgt 145f\n"
+      "beq 109f\n"
+      "cmp %x[M], #0x2\n"
+      "bgt 73f\n"
+      "beq 37f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "2:"  // Height 1: Column loop
+      "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
+      "cmp x14, #0xc\n"
+      "add x11, x12, x20, LSL #1\n"
+      "add x10, x11, x20, LSL #1\n"
+      "add x9, x10, x20, LSL #1\n"
+      "add x20, x9, x20, LSL #1\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "bgt 3f\n"
+      "cmp x14, #0x8\n"
+      "mov x9, x12\n"
+      "bgt 3f\n"
+      "cmp x14, #0x4\n"
+      "mov x10, x12\n"
+      "bgt 3f\n"
+      "mov x11, x12\n"
+      "3:"  // Height 1: B setup done
+      "cbz x15, 4f\n"
+      "ldr q8, [x15, #0x0]\n"
+      "ldr q9, [x15, #0x10]\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
+      "add x15, x15, #0x40\n"
+      "zip2 v12.2d, v8.2d, v8.2d\n"
+      "zip1 v8.2d, v8.2d, v8.2d\n"
+      "zip2 v13.2d, v9.2d, v9.2d\n"
+      "zip1 v9.2d, v9.2d, v9.2d\n"
+      "zip2 v14.2d, v10.2d, v10.2d\n"
+      "zip1 v10.2d, v10.2d, v10.2d\n"
+      "zip2 v15.2d, v11.2d, v11.2d\n"
+      "zip1 v11.2d, v11.2d, v11.2d\n"
+      "b 16f\n"
+      "4:"  // Height 1: no bias
+      "tbz %x[flags], #0, 15f\n"
+      "cmp x14, #0x10\n"
+      "bge 13f\n"
+      "tbz x14, #3, 8f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "tbz x14, #2, 6f\n"
+      "ld1 { v11.4s }, [x13], #0x10\n"
+      "tbz x14, #1, 5f\n"
+      "ldr d16, [x13], #0x8\n"
+      "mov x20, #0x38\n"
+      "tbz x14, #0, 12f\n"
+      "ld1 { v16.s }[2], [x13]\n"
+      "b 12f\n"
+      "5:"  // Height 1: Partial accumulate: partial_1_12
+      "mov x20, #0x30\n"
+      "tbz x14, #0, 12f\n"
+      "ldr s16, [x13, #0x0]\n"
+      "b 12f\n"
+      "6:"  // Height 1: Partial accumulate: partial_2_8
+      "tbz x14, #1, 7f\n"
+      "ldr d11, [x13], #0x8\n"
+      "mov x20, #0x28\n"
+      "tbz x14, #0, 12f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "b 12f\n"
+      "7:"  // Height 1: Partial accumulate: partial_1_8
+      "mov x20, #0x20\n"
+      "tbz x14, #0, 12f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "b 12f\n"
+      "8:"  // Height 1: Partial accumulate: partial_4_0
+      "tbz x14, #2, 10f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "tbz x14, #1, 9f\n"
+      "ldr d10, [x13], #0x8\n"
+      "mov x20, #0x18\n"
+      "tbz x14, #0, 12f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "b 12f\n"
+      "9:"  // Height 1: Partial accumulate: partial_1_4
+      "mov x20, #0x10\n"
+      "tbz x14, #0, 12f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "b 12f\n"
+      "10:"  // Height 1: Partial accumulate: partial_2_0
+      "tbz x14, #1, 11f\n"
+      "ldr d9, [x13], #0x8\n"
+      "mov x20, #0x8\n"
+      "tbz x14, #0, 12f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "b 12f\n"
+      "11:"  // Height 1: Partial accumulate: partial_1_0
+      "ldr s9, [x13, #0x0]\n"
+      "mov x20, #0x0\n"
+      "12:"  // Height 1: Partial accumulate: Done
+      "sub x13, x13, x20\n"
+      "b 14f\n"
+      "13:"  // Height 1: full accumulate
+      "ldr q9, [x13, #0x0]\n"
+      "ldr q10, [x13, #0x10]\n"
+      "ldr q11, [x13, #0x20]\n"
+      "ldr q16, [x13, #0x30]\n"
+      "14:"  // Height 1: MMLA fixup
+      "zip1 v8.2d, v9.2d, v12.2d\n"
+      "zip2 v12.2d, v9.2d, v12.2d\n"
+      "zip1 v9.2d, v10.2d, v13.2d\n"
+      "zip2 v13.2d, v10.2d, v13.2d\n"
+      "zip1 v10.2d, v11.2d, v14.2d\n"
+      "zip2 v14.2d, v11.2d, v14.2d\n"
+      "zip1 v11.2d, v16.2d, v15.2d\n"
+      "zip2 v15.2d, v16.2d, v15.2d\n"
+      "b 16f\n"
+      "15:"  // Height 1: 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"
+      "16:"  // Height 1: setup done
+      "mov x28, #0x0\n"
+      "17:"  // Height 1: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w27, [x20, x28, LSL #0x2]\n"
+      "tbz %x[flags], #3, 18f\n"
+      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+      "add x20, x20, x21, LSL #3\n"
+      "ldr x26, [x20, #0x0]\n"
+      "cbnz x28, 19f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x26, x26, x20, LSL #2\n"
+      "b 19f\n"
+      "18:"  // Height 1: setup direct input
+      "mov x26, %x[input_ptr]\n"
+      "19:"  // Height 1: input setup done
+      "cmp x27, #0x4\n"
+      "blt 22f\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      "ldr q6, [x12, #0x0]\n"
+      "cmp x27, #0x8\n"
+      "ldr q7, [x12, #0x10]\n"
+      "blt 21f\n"
+      "20:"  // Height 1: Multiply loop: Main loop head
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "ldr q18, [x11, #0x0]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "ldr q17, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e52ec09  // bfmmla v9.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x10, #0x0]\n"
+      ".inst 0x6e51ec0d  // bfmmla v13.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e52ec0a  // bfmmla v10.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x9, #0x0]\n"
+      ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
+      "ldr q6, [x12, #0x0]\n"
+      ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      "ldr q7, [x12, #0x10]\n"
+      "bge 20b\n"
+      "21:"  // Height 1: Multiply loop: Single iteration only
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "ldr q18, [x11, #0x0]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "ldr q17, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e52ec09  // bfmmla v9.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x10, #0x0]\n"
+      ".inst 0x6e51ec0d  // bfmmla v13.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e52ec0a  // bfmmla v10.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x9, #0x0]\n"
+      ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
+      ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "22:"  // Height 1: Multiply loop: Main loop skip
+      "cbz x27, 25f\n"
+      "cbz x27, 25f\n"
+      "tbz x27, #1, 23f\n"
+      "ldr d0, [x26], #0x8\n"
+      "tbz x27, #0, 24f\n"
+      "ld1 { v0.s }[2], [x26]\n"
+      "b 24f\n"
+      "23:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
+      "ldr s0, [x26, #0x0]\n"
+      "24:"  // Height 1: Multiply loop: Ragged operand read: Done
+      "ldr q18, [x12, #0x0]\n"
+      "ldr q17, [x12, #0x10]\n"
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x6e52ec08  // bfmmla v8.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x11, #0x0]\n"
+      ".inst 0x6e51ec0c  // bfmmla v12.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e52ec09  // bfmmla v9.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x10, #0x0]\n"
+      ".inst 0x6e51ec0d  // bfmmla v13.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e52ec0a  // bfmmla v10.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x9, #0x0]\n"
+      ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
+      ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "25:"  // Height 1: Multiply loop: No odd multiplies
+      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x28, x28, #0x1\n"
+      "cmp x28, x20\n"
+      "bne 17b\n"
+      "uzp1 v8.2d, v8.2d, v12.2d\n"
+      "uzp1 v9.2d, v9.2d, v13.2d\n"
+      "uzp1 v10.2d, v10.2d, v14.2d\n"
+      "uzp1 v11.2d, v11.2d, v15.2d\n"
+      "tbz %x[flags], #1, 26f\n"
+      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v18.4s }, [x21]\n"
+      "ld1r { v17.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v18.4s\n"
+      "fmin v9.4s, v9.4s, v18.4s\n"
+      "fmin v10.4s, v10.4s, v18.4s\n"
+      "fmin v11.4s, v11.4s, v18.4s\n"
+      "fmax v8.4s, v8.4s, v17.4s\n"
+      "fmax v9.4s, v9.4s, v17.4s\n"
+      "fmax v10.4s, v10.4s, v17.4s\n"
+      "fmax v11.4s, v11.4s, v17.4s\n"
+      "26:"  // Height 1: No activation
+      "cmp x14, #0x10\n"
+      "bge 35f\n"
+      "tbz x14, #3, 30f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "st1 { v9.4s }, [x13], #0x10\n"
+      "tbz x14, #2, 28f\n"
+      "st1 { v10.4s }, [x13], #0x10\n"
+      "tbz x14, #1, 27f\n"
+      "str d11, [x13], #0x8\n"
+      "tbz x14, #0, 34f\n"
+      "st1 { v11.s }[2], [x13]\n"
+      "b 34f\n"
+      "27:"  // Height 1: Partial direct writeback: partial_1_12
+      "tbz x14, #0, 34f\n"
+      "str s11, [x13, #0x0]\n"
+      "b 34f\n"
+      "28:"  // Height 1: Partial direct writeback: partial_2_8
+      "tbz x14, #1, 29f\n"
+      "str d10, [x13], #0x8\n"
+      "tbz x14, #0, 34f\n"
+      "st1 { v10.s }[2], [x13]\n"
+      "b 34f\n"
+      "29:"  // Height 1: Partial direct writeback: partial_1_8
+      "tbz x14, #0, 34f\n"
+      "str s10, [x13, #0x0]\n"
+      "b 34f\n"
+      "30:"  // Height 1: Partial direct writeback: partial_4_0
+      "tbz x14, #2, 32f\n"
+      "st1 { v8.4s }, [x13], #0x10\n"
+      "tbz x14, #1, 31f\n"
+      "str d9, [x13], #0x8\n"
+      "tbz x14, #0, 34f\n"
+      "st1 { v9.s }[2], [x13]\n"
+      "b 34f\n"
+      "31:"  // Height 1: Partial direct writeback: partial_1_4
+      "tbz x14, #0, 34f\n"
+      "str s9, [x13, #0x0]\n"
+      "b 34f\n"
+      "32:"  // Height 1: Partial direct writeback: partial_2_0
+      "tbz x14, #1, 33f\n"
+      "str d8, [x13], #0x8\n"
+      "tbz x14, #0, 34f\n"
+      "st1 { v8.s }[2], [x13]\n"
+      "b 34f\n"
+      "33:"  // Height 1: Partial direct writeback: partial_1_0
+      "str s8, [x13, #0x0]\n"
+      "34:"  // Height 1: Partial direct writeback: Done
+      "b 36f\n"
+      "35:"  // 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"
+      "36:"  // Height 1: Writeback done
+      "subs x14, x14, #0x10\n"
+      "bgt 2b\n"
+      "b 218f\n"
+      "37:"  // Height 2
+      "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "38:"  // Height 2: Column loop
+      "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
+      "cmp x14, #0xc\n"
+      "add x11, x12, x20, LSL #1\n"
+      "add x10, x11, x20, LSL #1\n"
+      "add x9, x10, x20, LSL #1\n"
+      "add x20, x9, x20, LSL #1\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "bgt 39f\n"
+      "cmp x14, #0x8\n"
+      "mov x9, x12\n"
+      "bgt 39f\n"
+      "cmp x14, #0x4\n"
+      "mov x10, x12\n"
+      "bgt 39f\n"
+      "mov x11, x12\n"
+      "39:"  // Height 2: B setup done
+      "cbz x15, 40f\n"
+      "ldr q8, [x15, #0x0]\n"
+      "ldr q9, [x15, #0x10]\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
+      "add x15, x15, #0x40\n"
+      "zip2 v12.2d, v8.2d, v8.2d\n"
+      "zip1 v8.2d, v8.2d, v8.2d\n"
+      "zip2 v13.2d, v9.2d, v9.2d\n"
+      "zip1 v9.2d, v9.2d, v9.2d\n"
+      "zip2 v14.2d, v10.2d, v10.2d\n"
+      "zip1 v10.2d, v10.2d, v10.2d\n"
+      "zip2 v15.2d, v11.2d, v11.2d\n"
+      "zip1 v11.2d, v11.2d, v11.2d\n"
+      "b 52f\n"
+      "40:"  // Height 2: no bias
+      "tbz %x[flags], #0, 51f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "cmp x14, #0x10\n"
+      "add x26, x13, x20, LSL #2\n"
+      "bge 49f\n"
+      "tbz x14, #3, 44f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x26], #0x10\n"
+      "tbz x14, #2, 42f\n"
+      "ld1 { v11.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x26], #0x10\n"
+      "tbz x14, #1, 41f\n"
+      "ldr d16, [x13], #0x8\n"
+      "ldr d15, [x26], #0x8\n"
+      "mov x20, #0x38\n"
+      "tbz x14, #0, 48f\n"
+      "ld1 { v16.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x26]\n"
+      "b 48f\n"
+      "41:"  // Height 2: Partial accumulate: partial_1_12
+      "mov x20, #0x30\n"
+      "tbz x14, #0, 48f\n"
+      "ldr s16, [x13, #0x0]\n"
+      "ldr s15, [x26, #0x0]\n"
+      "b 48f\n"
+      "42:"  // Height 2: Partial accumulate: partial_2_8
+      "tbz x14, #1, 43f\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d14, [x26], #0x8\n"
+      "mov x20, #0x28\n"
+      "tbz x14, #0, 48f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x26]\n"
+      "b 48f\n"
+      "43:"  // Height 2: Partial accumulate: partial_1_8
+      "mov x20, #0x20\n"
+      "tbz x14, #0, 48f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s14, [x26, #0x0]\n"
+      "b 48f\n"
+      "44:"  // Height 2: Partial accumulate: partial_4_0
+      "tbz x14, #2, 46f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "tbz x14, #1, 45f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d13, [x26], #0x8\n"
+      "mov x20, #0x18\n"
+      "tbz x14, #0, 48f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x26]\n"
+      "b 48f\n"
+      "45:"  // Height 2: Partial accumulate: partial_1_4
+      "mov x20, #0x10\n"
+      "tbz x14, #0, 48f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s13, [x26, #0x0]\n"
+      "b 48f\n"
+      "46:"  // Height 2: Partial accumulate: partial_2_0
+      "tbz x14, #1, 47f\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d12, [x26], #0x8\n"
+      "mov x20, #0x8\n"
+      "tbz x14, #0, 48f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x26]\n"
+      "b 48f\n"
+      "47:"  // Height 2: Partial accumulate: partial_1_0
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s12, [x26, #0x0]\n"
+      "mov x20, #0x0\n"
+      "48:"  // Height 2: Partial accumulate: Done
+      "sub x13, x13, x20\n"
+      "b 50f\n"
+      "49:"  // Height 2: full accumulate
+      "ldr q9, [x13, #0x0]\n"
+      "ldr q10, [x13, #0x10]\n"
+      "ldr q11, [x13, #0x20]\n"
+      "ldr q16, [x13, #0x30]\n"
+      "ldr q12, [x26, #0x0]\n"
+      "ldr q13, [x26, #0x10]\n"
+      "ldr q14, [x26, #0x20]\n"
+      "ldr q15, [x26, #0x30]\n"
+      "50:"  // Height 2: MMLA fixup
+      "zip1 v8.2d, v9.2d, v12.2d\n"
+      "zip2 v12.2d, v9.2d, v12.2d\n"
+      "zip1 v9.2d, v10.2d, v13.2d\n"
+      "zip2 v13.2d, v10.2d, v13.2d\n"
+      "zip1 v10.2d, v11.2d, v14.2d\n"
+      "zip2 v14.2d, v11.2d, v14.2d\n"
+      "zip1 v11.2d, v16.2d, v15.2d\n"
+      "zip2 v15.2d, v16.2d, v15.2d\n"
+      "b 52f\n"
+      "51:"  // 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"
+      "52:"  // Height 2: setup done
+      "mov x28, #0x0\n"
+      "53:"  // Height 2: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w27, [x20, x28, LSL #0x2]\n"
+      "tbz %x[flags], #3, 54f\n"
+      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+      "add x20, x20, x21, LSL #3\n"
+      "ldr x26, [x20, #0x0]\n"
+      "ldr x25, [x20, #0x8]\n"
+      "cbnz x28, 55f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x26, x26, x20, LSL #2\n"
+      "add x25, x25, x20, LSL #2\n"
+      "b 55f\n"
+      "54:"  // Height 2: setup direct input
+      "mov x26, %x[input_ptr]\n"
+      "add x25, x26, x21, LSL #2\n"
+      "55:"  // Height 2: input setup done
+      "cmp x27, #0x4\n"
+      "blt 58f\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      "cmp x27, #0x8\n"
+      "ldr q6, [x12, #0x0]\n"
+      "ldr q7, [x12, #0x10]\n"
+      "blt 57f\n"
+      "56:"  // Height 2: Multiply loop: Main loop head
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "ldr q18, [x11, #0x0]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "ldr q17, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e52ec09  // bfmmla v9.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x10, #0x0]\n"
+      ".inst 0x6e51ec0d  // bfmmla v13.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e52ec0a  // bfmmla v10.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x9, #0x0]\n"
+      ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
+      "ldr q6, [x12, #0x0]\n"
+      ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      "ldr q7, [x12, #0x10]\n"
+      "bge 56b\n"
+      "57:"  // Height 2: Multiply loop: Single iteration only
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "ldr q18, [x11, #0x0]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "ldr q17, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e52ec09  // bfmmla v9.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x10, #0x0]\n"
+      ".inst 0x6e51ec0d  // bfmmla v13.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e52ec0a  // bfmmla v10.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x9, #0x0]\n"
+      ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
+      ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "58:"  // Height 2: Multiply loop: Main loop skip
+      "cbz x27, 61f\n"
+      "cbz x27, 61f\n"
+      "tbz x27, #1, 59f\n"
+      "ldr d0, [x26], #0x8\n"
+      "ldr d1, [x25], #0x8\n"
+      "tbz x27, #0, 60f\n"
+      "ld1 { v0.s }[2], [x26]\n"
+      "ld1 { v1.s }[2], [x25]\n"
+      "b 60f\n"
+      "59:"  // Height 2: Multiply loop: Ragged operand read: partial_1_0
+      "ldr s0, [x26, #0x0]\n"
+      "ldr s1, [x25, #0x0]\n"
+      "60:"  // Height 2: Multiply loop: Ragged operand read: Done
+      "ldr q18, [x12, #0x0]\n"
+      "ldr q17, [x12, #0x10]\n"
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x6e52ec08  // bfmmla v8.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x11, #0x0]\n"
+      ".inst 0x6e51ec0c  // bfmmla v12.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e52ec09  // bfmmla v9.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x10, #0x0]\n"
+      ".inst 0x6e51ec0d  // bfmmla v13.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e52ec0a  // bfmmla v10.4s, v0.8h, v18.8h\n"
+      "ldr q18, [x9, #0x0]\n"
+      ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
+      "ldr q17, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
+      ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "61:"  // Height 2: Multiply loop: No odd multiplies
+      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x28, x28, #0x1\n"
+      "cmp x28, x20\n"
+      "bne 53b\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "uzp1 v6.2d, v8.2d, v12.2d\n"
+      "uzp2 v8.2d, v8.2d, v12.2d\n"
+      "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "uzp2 v9.2d, v9.2d, v13.2d\n"
+      "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "uzp2 v10.2d, v10.2d, v14.2d\n"
+      "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "uzp2 v11.2d, v11.2d, v15.2d\n"
+      "add x26, x13, x20, LSL #2\n"
+      "tbz %x[flags], #1, 62f\n"
+      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v18.4s }, [x21]\n"
+      "ld1r { v17.4s }, [x20]\n"
+      "fmin v6.4s, v6.4s, v18.4s\n"
+      "fmin v12.4s, v12.4s, v18.4s\n"
+      "fmin v13.4s, v13.4s, v18.4s\n"
+      "fmin v14.4s, v14.4s, v18.4s\n"
+      "fmin v8.4s, v8.4s, v18.4s\n"
+      "fmin v9.4s, v9.4s, v18.4s\n"
+      "fmin v10.4s, v10.4s, v18.4s\n"
+      "fmin v11.4s, v11.4s, v18.4s\n"
+      "fmax v6.4s, v6.4s, v17.4s\n"
+      "fmax v12.4s, v12.4s, v17.4s\n"
+      "fmax v13.4s, v13.4s, v17.4s\n"
+      "fmax v14.4s, v14.4s, v17.4s\n"
+      "fmax v8.4s, v8.4s, v17.4s\n"
+      "fmax v9.4s, v9.4s, v17.4s\n"
+      "fmax v10.4s, v10.4s, v17.4s\n"
+      "fmax v11.4s, v11.4s, v17.4s\n"
+      "62:"  // Height 2: No activation
+      "cmp x14, #0x10\n"
+      "bge 71f\n"
+      "tbz x14, #3, 66f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v9.4s }, [x26], #0x10\n"
+      "tbz x14, #2, 64f\n"
+      "st1 { v13.4s }, [x13], #0x10\n"
+      "st1 { v10.4s }, [x26], #0x10\n"
+      "tbz x14, #1, 63f\n"
+      "str d14, [x13], #0x8\n"
+      "str d11, [x26], #0x8\n"
+      "tbz x14, #0, 70f\n"
+      "st1 { v14.s }[2], [x13]\n"
+      "st1 { v11.s }[2], [x26]\n"
+      "b 70f\n"
+      "63:"  // Height 2: Partial direct writeback: partial_1_12
+      "tbz x14, #0, 70f\n"
+      "str s14, [x13, #0x0]\n"
+      "str s11, [x26, #0x0]\n"
+      "b 70f\n"
+      "64:"  // Height 2: Partial direct writeback: partial_2_8
+      "tbz x14, #1, 65f\n"
+      "str d13, [x13], #0x8\n"
+      "str d10, [x26], #0x8\n"
+      "tbz x14, #0, 70f\n"
+      "st1 { v13.s }[2], [x13]\n"
+      "st1 { v10.s }[2], [x26]\n"
+      "b 70f\n"
+      "65:"  // Height 2: Partial direct writeback: partial_1_8
+      "tbz x14, #0, 70f\n"
+      "str s13, [x13, #0x0]\n"
+      "str s10, [x26, #0x0]\n"
+      "b 70f\n"
+      "66:"  // Height 2: Partial direct writeback: partial_4_0
+      "tbz x14, #2, 68f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "tbz x14, #1, 67f\n"
+      "str d12, [x13], #0x8\n"
+      "str d9, [x26], #0x8\n"
+      "tbz x14, #0, 70f\n"
+      "st1 { v12.s }[2], [x13]\n"
+      "st1 { v9.s }[2], [x26]\n"
+      "b 70f\n"
+      "67:"  // Height 2: Partial direct writeback: partial_1_4
+      "tbz x14, #0, 70f\n"
+      "str s12, [x13, #0x0]\n"
+      "str s9, [x26, #0x0]\n"
+      "b 70f\n"
+      "68:"  // Height 2: Partial direct writeback: partial_2_0
+      "tbz x14, #1, 69f\n"
+      "str d6, [x13], #0x8\n"
+      "str d8, [x26], #0x8\n"
+      "tbz x14, #0, 70f\n"
+      "st1 { v6.s }[2], [x13]\n"
+      "st1 { v8.s }[2], [x26]\n"
+      "b 70f\n"
+      "69:"  // Height 2: Partial direct writeback: partial_1_0
+      "str s6, [x13, #0x0]\n"
+      "str s8, [x26, #0x0]\n"
+      "70:"  // Height 2: Partial direct writeback: Done
+      "b 72f\n"
+      "71:"  // Height 2: Full writeback
+      "str q6, [x13, #0x0]\n"
+      "str q12, [x13, #0x10]\n"
+      "str q13, [x13, #0x20]\n"
+      "str q14, [x13, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "str q8, [x26, #0x0]\n"
+      "str q9, [x26, #0x10]\n"
+      "str q10, [x26, #0x20]\n"
+      "str q11, [x26, #0x30]\n"
+      "72:"  // Height 2: Writeback done
+      "subs x14, x14, #0x10\n"
+      "bgt 38b\n"
+      "b 218f\n"
+      "73:"  // Height 3
+      "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "74:"  // Height 3: Column loop
+      "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
+      "cmp x14, #0xc\n"
+      "add x11, x12, x20, LSL #1\n"
+      "add x10, x11, x20, LSL #1\n"
+      "add x9, x10, x20, LSL #1\n"
+      "add x20, x9, x20, LSL #1\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "bgt 75f\n"
+      "cmp x14, #0x8\n"
+      "mov x9, x12\n"
+      "bgt 75f\n"
+      "cmp x14, #0x4\n"
+      "mov x10, x12\n"
+      "bgt 75f\n"
+      "mov x11, x12\n"
+      "75:"  // Height 3: B setup done
+      "cbz x15, 76f\n"
+      "ldr q8, [x15, #0x0]\n"
+      "ldr q9, [x15, #0x10]\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
+      "add x15, x15, #0x40\n"
+      "zip2 v12.2d, v8.2d, v8.2d\n"
+      "zip1 v8.2d, v8.2d, v8.2d\n"
+      "zip2 v13.2d, v9.2d, v9.2d\n"
+      "zip1 v9.2d, v9.2d, v9.2d\n"
+      "zip2 v14.2d, v10.2d, v10.2d\n"
+      "zip1 v10.2d, v10.2d, v10.2d\n"
+      "zip2 v15.2d, v11.2d, v11.2d\n"
+      "zip1 v11.2d, v11.2d, v11.2d\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v20.16b, v12.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "mov v21.16b, v13.16b\n"
+      "mov v18.16b, v10.16b\n"
+      "mov v22.16b, v14.16b\n"
+      "mov v19.16b, v11.16b\n"
+      "mov v23.16b, v15.16b\n"
+      "b 88f\n"
+      "76:"  // Height 3: no bias
+      "tbz %x[flags], #0, 87f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "cmp x14, #0x10\n"
+      "add x26, x13, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
+      "bge 85f\n"
+      "tbz x14, #3, 80f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v18.4s }, [x25], #0x10\n"
+      "tbz x14, #2, 78f\n"
+      "ld1 { v11.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v19.4s }, [x25], #0x10\n"
+      "tbz x14, #1, 77f\n"
+      "ldr d16, [x13], #0x8\n"
+      "ldr d15, [x26], #0x8\n"
+      "mov x20, #0x38\n"
+      "ldr d24, [x25], #0x8\n"
+      "tbz x14, #0, 84f\n"
+      "ld1 { v16.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v24.s }[2], [x25]\n"
+      "b 84f\n"
+      "77:"  // Height 3: Partial accumulate: partial_1_12
+      "mov x20, #0x30\n"
+      "tbz x14, #0, 84f\n"
+      "ldr s16, [x13, #0x0]\n"
+      "ldr s15, [x26, #0x0]\n"
+      "ldr s24, [x25, #0x0]\n"
+      "b 84f\n"
+      "78:"  // Height 3: Partial accumulate: partial_2_8
+      "tbz x14, #1, 79f\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d14, [x26], #0x8\n"
+      "mov x20, #0x28\n"
+      "ldr d19, [x25], #0x8\n"
+      "tbz x14, #0, 84f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v19.s }[2], [x25]\n"
+      "b 84f\n"
+      "79:"  // Height 3: Partial accumulate: partial_1_8
+      "mov x20, #0x20\n"
+      "tbz x14, #0, 84f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s14, [x26, #0x0]\n"
+      "ldr s19, [x25, #0x0]\n"
+      "b 84f\n"
+      "80:"  // Height 3: Partial accumulate: partial_4_0
+      "tbz x14, #2, 82f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v17.4s }, [x25], #0x10\n"
+      "tbz x14, #1, 81f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d13, [x26], #0x8\n"
+      "mov x20, #0x18\n"
+      "ldr d18, [x25], #0x8\n"
+      "tbz x14, #0, 84f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v18.s }[2], [x25]\n"
+      "b 84f\n"
+      "81:"  // Height 3: Partial accumulate: partial_1_4
+      "mov x20, #0x10\n"
+      "tbz x14, #0, 84f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s13, [x26, #0x0]\n"
+      "ldr s18, [x25, #0x0]\n"
+      "b 84f\n"
+      "82:"  // Height 3: Partial accumulate: partial_2_0
+      "tbz x14, #1, 83f\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d12, [x26], #0x8\n"
+      "mov x20, #0x8\n"
+      "ldr d17, [x25], #0x8\n"
+      "tbz x14, #0, 84f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v17.s }[2], [x25]\n"
+      "b 84f\n"
+      "83:"  // Height 3: Partial accumulate: partial_1_0
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s12, [x26, #0x0]\n"
+      "mov x20, #0x0\n"
+      "ldr s17, [x25, #0x0]\n"
+      "84:"  // Height 3: Partial accumulate: Done
+      "sub x13, x13, x20\n"
+      "b 86f\n"
+      "85:"  // Height 3: full accumulate
+      "ldr q9, [x13, #0x0]\n"
+      "ldr q10, [x13, #0x10]\n"
+      "ldr q11, [x13, #0x20]\n"
+      "ldr q16, [x13, #0x30]\n"
+      "ldr q12, [x26, #0x0]\n"
+      "ldr q13, [x26, #0x10]\n"
+      "ldr q14, [x26, #0x20]\n"
+      "ldr q15, [x26, #0x30]\n"
+      "ldr q17, [x25, #0x0]\n"
+      "ldr q18, [x25, #0x10]\n"
+      "ldr q19, [x25, #0x20]\n"
+      "ldr q24, [x25, #0x30]\n"
+      "86:"  // Height 3: MMLA fixup
+      "zip1 v8.2d, v9.2d, v12.2d\n"
+      "zip2 v12.2d, v9.2d, v12.2d\n"
+      "zip1 v9.2d, v10.2d, v13.2d\n"
+      "zip2 v13.2d, v10.2d, v13.2d\n"
+      "zip1 v10.2d, v11.2d, v14.2d\n"
+      "zip2 v14.2d, v11.2d, v14.2d\n"
+      "zip1 v11.2d, v16.2d, v15.2d\n"
+      "zip2 v15.2d, v16.2d, v15.2d\n"
+      "zip1 v16.2d, v17.2d, v20.2d\n"
+      "zip2 v20.2d, v17.2d, v20.2d\n"
+      "zip1 v17.2d, v18.2d, v21.2d\n"
+      "zip2 v21.2d, v18.2d, v21.2d\n"
+      "zip1 v18.2d, v19.2d, v22.2d\n"
+      "zip2 v22.2d, v19.2d, v22.2d\n"
+      "zip1 v19.2d, v24.2d, v23.2d\n"
+      "zip2 v23.2d, v24.2d, v23.2d\n"
+      "b 88f\n"
+      "87:"  // 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"
+      "movi v20.16b, #0x0\n"
+      "movi v21.16b, #0x0\n"
+      "movi v22.16b, #0x0\n"
+      "movi v23.16b, #0x0\n"
+      "88:"  // Height 3: setup done
+      "mov x28, #0x0\n"
+      "89:"  // Height 3: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w27, [x20, x28, LSL #0x2]\n"
+      "tbz %x[flags], #3, 90f\n"
+      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+      "add x20, x20, x21, LSL #3\n"
+      "ldr x26, [x20, #0x0]\n"
+      "ldr x25, [x20, #0x8]\n"
+      "ldr x24, [x20, #0x10]\n"
+      "cbnz x28, 91f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x26, x26, x20, LSL #2\n"
+      "add x25, x25, x20, LSL #2\n"
+      "add x24, x24, x20, LSL #2\n"
+      "b 91f\n"
+      "90:"  // Height 3: setup direct input
+      "mov x26, %x[input_ptr]\n"
+      "add x25, x26, x21, LSL #2\n"
+      "add x24, x25, x21, LSL #2\n"
+      "91:"  // Height 3: input setup done
+      "cmp x27, #0x4\n"
+      "blt 94f\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      "cmp x27, #0x8\n"
+      "ld1 { v2.4s }, [x24], #0x10\n"
+      "ldr q6, [x12, #0x0]\n"
+      "ldr q7, [x12, #0x10]\n"
+      "blt 93f\n"
+      "92:"  // Height 3: Multiply loop: Main loop head
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "ldr q25, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x10, #0x0]\n"
+      ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec55  // bfmmla v21.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e5aec0a  // bfmmla v10.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec52  // bfmmla v18.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x9, #0x0]\n"
+      ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec56  // bfmmla v22.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e5aec0b  // bfmmla v11.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec53  // bfmmla v19.4s, v2.8h, v26.8h\n"
+      "ldr q6, [x12, #0x0]\n"
+      ".inst 0x6e59ec0f  // bfmmla v15.4s, v0.8h, v25.8h\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      ".inst 0x6e59ec57  // bfmmla v23.4s, v2.8h, v25.8h\n"
+      "ld1 { v2.4s }, [x24], #0x10\n"
+      "ldr q7, [x12, #0x10]\n"
+      "bge 92b\n"
+      "93:"  // Height 3: Multiply loop: Single iteration only
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "ldr q25, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x10, #0x0]\n"
+      ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec55  // bfmmla v21.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e5aec0a  // bfmmla v10.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec52  // bfmmla v18.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x9, #0x0]\n"
+      ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec56  // bfmmla v22.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e5aec0b  // bfmmla v11.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec53  // bfmmla v19.4s, v2.8h, v26.8h\n"
+      ".inst 0x6e59ec0f  // bfmmla v15.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec57  // bfmmla v23.4s, v2.8h, v25.8h\n"
+      "94:"  // Height 3: Multiply loop: Main loop skip
+      "cbz x27, 97f\n"
+      "cbz x27, 97f\n"
+      "tbz x27, #1, 95f\n"
+      "ldr d0, [x26], #0x8\n"
+      "ldr d1, [x25], #0x8\n"
+      "ldr d2, [x24], #0x8\n"
+      "tbz x27, #0, 96f\n"
+      "ld1 { v0.s }[2], [x26]\n"
+      "ld1 { v1.s }[2], [x25]\n"
+      "ld1 { v2.s }[2], [x24]\n"
+      "b 96f\n"
+      "95:"  // Height 3: Multiply loop: Ragged operand read: partial_1_0
+      "ldr s0, [x26, #0x0]\n"
+      "ldr s1, [x25, #0x0]\n"
+      "ldr s2, [x24, #0x0]\n"
+      "96:"  // Height 3: Multiply loop: Ragged operand read: Done
+      "ldr q26, [x12, #0x0]\n"
+      "ldr q25, [x12, #0x10]\n"
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x6e5aec50  // bfmmla v16.4s, v2.8h, v26.8h\n"
+      ".inst 0x6e59ec54  // bfmmla v20.4s, v2.8h, v25.8h\n"
+      ".inst 0x6e5aec08  // bfmmla v8.4s, v0.8h, v26.8h\n"
+      "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e59ec0c  // bfmmla v12.4s, v0.8h, v25.8h\n"
+      "ldr q25, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x10, #0x0]\n"
+      ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec55  // bfmmla v21.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e5aec0a  // bfmmla v10.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec52  // bfmmla v18.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x9, #0x0]\n"
+      ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec56  // bfmmla v22.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e5aec0b  // bfmmla v11.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec53  // bfmmla v19.4s, v2.8h, v26.8h\n"
+      ".inst 0x6e59ec0f  // bfmmla v15.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec57  // bfmmla v23.4s, v2.8h, v25.8h\n"
+      "97:"  // Height 3: Multiply loop: No odd multiplies
+      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x28, x28, #0x1\n"
+      "cmp x28, x20\n"
+      "bne 89b\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "uzp1 v6.2d, v8.2d, v12.2d\n"
+      "uzp2 v8.2d, v8.2d, v12.2d\n"
+      "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "uzp2 v9.2d, v9.2d, v13.2d\n"
+      "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "uzp2 v10.2d, v10.2d, v14.2d\n"
+      "add x26, x13, x20, LSL #2\n"
+      "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "uzp2 v11.2d, v11.2d, v15.2d\n"
+      "add x25, x26, x20, LSL #2\n"
+      "uzp1 v16.2d, v16.2d, v20.2d\n"
+      "uzp1 v17.2d, v17.2d, v21.2d\n"
+      "uzp1 v18.2d, v18.2d, v22.2d\n"
+      "uzp1 v19.2d, v19.2d, v23.2d\n"
+      "tbz %x[flags], #1, 98f\n"
+      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v26.4s }, [x21]\n"
+      "ld1r { v25.4s }, [x20]\n"
+      "fmin v6.4s, v6.4s, v26.4s\n"
+      "fmin v12.4s, v12.4s, v26.4s\n"
+      "fmin v13.4s, v13.4s, v26.4s\n"
+      "fmin v14.4s, v14.4s, v26.4s\n"
+      "fmin v8.4s, v8.4s, v26.4s\n"
+      "fmin v9.4s, v9.4s, v26.4s\n"
+      "fmin v10.4s, v10.4s, v26.4s\n"
+      "fmin v11.4s, v11.4s, v26.4s\n"
+      "fmin v16.4s, v16.4s, v26.4s\n"
+      "fmin v17.4s, v17.4s, v26.4s\n"
+      "fmin v18.4s, v18.4s, v26.4s\n"
+      "fmin v19.4s, v19.4s, v26.4s\n"
+      "fmax v6.4s, v6.4s, v25.4s\n"
+      "fmax v12.4s, v12.4s, v25.4s\n"
+      "fmax v13.4s, v13.4s, v25.4s\n"
+      "fmax v14.4s, v14.4s, v25.4s\n"
+      "fmax v8.4s, v8.4s, v25.4s\n"
+      "fmax v9.4s, v9.4s, v25.4s\n"
+      "fmax v10.4s, v10.4s, v25.4s\n"
+      "fmax v11.4s, v11.4s, v25.4s\n"
+      "fmax v16.4s, v16.4s, v25.4s\n"
+      "fmax v17.4s, v17.4s, v25.4s\n"
+      "fmax v18.4s, v18.4s, v25.4s\n"
+      "fmax v19.4s, v19.4s, v25.4s\n"
+      "98:"  // Height 3: No activation
+      "cmp x14, #0x10\n"
+      "bge 107f\n"
+      "tbz x14, #3, 102f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v9.4s }, [x26], #0x10\n"
+      "st1 { v16.4s }, [x25], #0x10\n"
+      "st1 { v17.4s }, [x25], #0x10\n"
+      "tbz x14, #2, 100f\n"
+      "st1 { v13.4s }, [x13], #0x10\n"
+      "st1 { v10.4s }, [x26], #0x10\n"
+      "st1 { v18.4s }, [x25], #0x10\n"
+      "tbz x14, #1, 99f\n"
+      "str d14, [x13], #0x8\n"
+      "str d11, [x26], #0x8\n"
+      "str d19, [x25], #0x8\n"
+      "tbz x14, #0, 106f\n"
+      "st1 { v14.s }[2], [x13]\n"
+      "st1 { v11.s }[2], [x26]\n"
+      "st1 { v19.s }[2], [x25]\n"
+      "b 106f\n"
+      "99:"  // Height 3: Partial direct writeback: partial_1_12
+      "tbz x14, #0, 106f\n"
+      "str s14, [x13, #0x0]\n"
+      "str s11, [x26, #0x0]\n"
+      "str s19, [x25, #0x0]\n"
+      "b 106f\n"
+      "100:"  // Height 3: Partial direct writeback: partial_2_8
+      "tbz x14, #1, 101f\n"
+      "str d13, [x13], #0x8\n"
+      "str d10, [x26], #0x8\n"
+      "str d18, [x25], #0x8\n"
+      "tbz x14, #0, 106f\n"
+      "st1 { v13.s }[2], [x13]\n"
+      "st1 { v10.s }[2], [x26]\n"
+      "st1 { v18.s }[2], [x25]\n"
+      "b 106f\n"
+      "101:"  // Height 3: Partial direct writeback: partial_1_8
+      "tbz x14, #0, 106f\n"
+      "str s13, [x13, #0x0]\n"
+      "str s10, [x26, #0x0]\n"
+      "str s18, [x25, #0x0]\n"
+      "b 106f\n"
+      "102:"  // Height 3: Partial direct writeback: partial_4_0
+      "tbz x14, #2, 104f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v16.4s }, [x25], #0x10\n"
+      "tbz x14, #1, 103f\n"
+      "str d12, [x13], #0x8\n"
+      "str d9, [x26], #0x8\n"
+      "str d17, [x25], #0x8\n"
+      "tbz x14, #0, 106f\n"
+      "st1 { v12.s }[2], [x13]\n"
+      "st1 { v9.s }[2], [x26]\n"
+      "st1 { v17.s }[2], [x25]\n"
+      "b 106f\n"
+      "103:"  // Height 3: Partial direct writeback: partial_1_4
+      "tbz x14, #0, 106f\n"
+      "str s12, [x13, #0x0]\n"
+      "str s9, [x26, #0x0]\n"
+      "str s17, [x25, #0x0]\n"
+      "b 106f\n"
+      "104:"  // Height 3: Partial direct writeback: partial_2_0
+      "tbz x14, #1, 105f\n"
+      "str d6, [x13], #0x8\n"
+      "str d8, [x26], #0x8\n"
+      "str d16, [x25], #0x8\n"
+      "tbz x14, #0, 106f\n"
+      "st1 { v6.s }[2], [x13]\n"
+      "st1 { v8.s }[2], [x26]\n"
+      "st1 { v16.s }[2], [x25]\n"
+      "b 106f\n"
+      "105:"  // Height 3: Partial direct writeback: partial_1_0
+      "str s6, [x13, #0x0]\n"
+      "str s8, [x26, #0x0]\n"
+      "str s16, [x25, #0x0]\n"
+      "106:"  // Height 3: Partial direct writeback: Done
+      "b 108f\n"
+      "107:"  // Height 3: Full writeback
+      "str q6, [x13, #0x0]\n"
+      "str q12, [x13, #0x10]\n"
+      "str q13, [x13, #0x20]\n"
+      "str q14, [x13, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "str q8, [x26, #0x0]\n"
+      "str q9, [x26, #0x10]\n"
+      "str q10, [x26, #0x20]\n"
+      "str q11, [x26, #0x30]\n"
+      "str q16, [x25, #0x0]\n"
+      "str q17, [x25, #0x10]\n"
+      "str q18, [x25, #0x20]\n"
+      "str q19, [x25, #0x30]\n"
+      "108:"  // Height 3: Writeback done
+      "subs x14, x14, #0x10\n"
+      "bgt 74b\n"
+      "b 218f\n"
+      "109:"  // Height 4
+      "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "110:"  // Height 4: Column loop
+      "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
+      "cmp x14, #0xc\n"
+      "add x11, x12, x20, LSL #1\n"
+      "add x10, x11, x20, LSL #1\n"
+      "add x9, x10, x20, LSL #1\n"
+      "add x20, x9, x20, LSL #1\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "bgt 111f\n"
+      "cmp x14, #0x8\n"
+      "mov x9, x12\n"
+      "bgt 111f\n"
+      "cmp x14, #0x4\n"
+      "mov x10, x12\n"
+      "bgt 111f\n"
+      "mov x11, x12\n"
+      "111:"  // Height 4: B setup done
+      "cbz x15, 112f\n"
+      "ldr q8, [x15, #0x0]\n"
+      "ldr q9, [x15, #0x10]\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
+      "add x15, x15, #0x40\n"
+      "zip2 v12.2d, v8.2d, v8.2d\n"
+      "zip1 v8.2d, v8.2d, v8.2d\n"
+      "zip2 v13.2d, v9.2d, v9.2d\n"
+      "zip1 v9.2d, v9.2d, v9.2d\n"
+      "zip2 v14.2d, v10.2d, v10.2d\n"
+      "zip1 v10.2d, v10.2d, v10.2d\n"
+      "zip2 v15.2d, v11.2d, v11.2d\n"
+      "zip1 v11.2d, v11.2d, v11.2d\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v20.16b, v12.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "mov v21.16b, v13.16b\n"
+      "mov v18.16b, v10.16b\n"
+      "mov v22.16b, v14.16b\n"
+      "mov v19.16b, v11.16b\n"
+      "mov v23.16b, v15.16b\n"
+      "b 124f\n"
+      "112:"  // Height 4: no bias
+      "tbz %x[flags], #0, 123f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "cmp x14, #0x10\n"
+      "add x26, x13, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
+      "bge 121f\n"
+      "tbz x14, #3, 116f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v21.4s }, [x24], #0x10\n"
+      "tbz x14, #2, 114f\n"
+      "ld1 { v11.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v19.4s }, [x25], #0x10\n"
+      "ld1 { v22.4s }, [x24], #0x10\n"
+      "tbz x14, #1, 113f\n"
+      "ldr d16, [x13], #0x8\n"
+      "ldr d15, [x26], #0x8\n"
+      "mov x20, #0x38\n"
+      "ldr d24, [x25], #0x8\n"
+      "ldr d23, [x24], #0x8\n"
+      "tbz x14, #0, 120f\n"
+      "ld1 { v16.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v24.s }[2], [x25]\n"
+      "ld1 { v23.s }[2], [x24]\n"
+      "b 120f\n"
+      "113:"  // Height 4: Partial accumulate: partial_1_12
+      "mov x20, #0x30\n"
+      "tbz x14, #0, 120f\n"
+      "ldr s16, [x13, #0x0]\n"
+      "ldr s15, [x26, #0x0]\n"
+      "ldr s24, [x25, #0x0]\n"
+      "ldr s23, [x24, #0x0]\n"
+      "b 120f\n"
+      "114:"  // Height 4: Partial accumulate: partial_2_8
+      "tbz x14, #1, 115f\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d14, [x26], #0x8\n"
+      "mov x20, #0x28\n"
+      "ldr d19, [x25], #0x8\n"
+      "ldr d22, [x24], #0x8\n"
+      "tbz x14, #0, 120f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v22.s }[2], [x24]\n"
+      "b 120f\n"
+      "115:"  // Height 4: Partial accumulate: partial_1_8
+      "mov x20, #0x20\n"
+      "tbz x14, #0, 120f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s14, [x26, #0x0]\n"
+      "ldr s19, [x25, #0x0]\n"
+      "ldr s22, [x24, #0x0]\n"
+      "b 120f\n"
+      "116:"  // Height 4: Partial accumulate: partial_4_0
+      "tbz x14, #2, 118f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v20.4s }, [x24], #0x10\n"
+      "tbz x14, #1, 117f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d13, [x26], #0x8\n"
+      "mov x20, #0x18\n"
+      "ldr d18, [x25], #0x8\n"
+      "ldr d21, [x24], #0x8\n"
+      "tbz x14, #0, 120f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v21.s }[2], [x24]\n"
+      "b 120f\n"
+      "117:"  // Height 4: Partial accumulate: partial_1_4
+      "mov x20, #0x10\n"
+      "tbz x14, #0, 120f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s13, [x26, #0x0]\n"
+      "ldr s18, [x25, #0x0]\n"
+      "ldr s21, [x24, #0x0]\n"
+      "b 120f\n"
+      "118:"  // Height 4: Partial accumulate: partial_2_0
+      "tbz x14, #1, 119f\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d12, [x26], #0x8\n"
+      "mov x20, #0x8\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d20, [x24], #0x8\n"
+      "tbz x14, #0, 120f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v20.s }[2], [x24]\n"
+      "b 120f\n"
+      "119:"  // Height 4: Partial accumulate: partial_1_0
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s12, [x26, #0x0]\n"
+      "mov x20, #0x0\n"
+      "ldr s17, [x25, #0x0]\n"
+      "ldr s20, [x24, #0x0]\n"
+      "120:"  // Height 4: Partial accumulate: Done
+      "sub x13, x13, x20\n"
+      "b 122f\n"
+      "121:"  // Height 4: full accumulate
+      "ldr q9, [x13, #0x0]\n"
+      "ldr q10, [x13, #0x10]\n"
+      "ldr q11, [x13, #0x20]\n"
+      "ldr q16, [x13, #0x30]\n"
+      "ldr q12, [x26, #0x0]\n"
+      "ldr q13, [x26, #0x10]\n"
+      "ldr q14, [x26, #0x20]\n"
+      "ldr q15, [x26, #0x30]\n"
+      "ldr q17, [x25, #0x0]\n"
+      "ldr q18, [x25, #0x10]\n"
+      "ldr q19, [x25, #0x20]\n"
+      "ldr q24, [x25, #0x30]\n"
+      "ldr q20, [x24, #0x0]\n"
+      "ldr q21, [x24, #0x10]\n"
+      "ldr q22, [x24, #0x20]\n"
+      "ldr q23, [x24, #0x30]\n"
+      "122:"  // Height 4: MMLA fixup
+      "zip1 v8.2d, v9.2d, v12.2d\n"
+      "zip2 v12.2d, v9.2d, v12.2d\n"
+      "zip1 v9.2d, v10.2d, v13.2d\n"
+      "zip2 v13.2d, v10.2d, v13.2d\n"
+      "zip1 v10.2d, v11.2d, v14.2d\n"
+      "zip2 v14.2d, v11.2d, v14.2d\n"
+      "zip1 v11.2d, v16.2d, v15.2d\n"
+      "zip2 v15.2d, v16.2d, v15.2d\n"
+      "zip1 v16.2d, v17.2d, v20.2d\n"
+      "zip2 v20.2d, v17.2d, v20.2d\n"
+      "zip1 v17.2d, v18.2d, v21.2d\n"
+      "zip2 v21.2d, v18.2d, v21.2d\n"
+      "zip1 v18.2d, v19.2d, v22.2d\n"
+      "zip2 v22.2d, v19.2d, v22.2d\n"
+      "zip1 v19.2d, v24.2d, v23.2d\n"
+      "zip2 v23.2d, v24.2d, v23.2d\n"
+      "b 124f\n"
+      "123:"  // 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"
+      "124:"  // Height 4: setup done
+      "mov x28, #0x0\n"
+      "125:"  // Height 4: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w27, [x20, x28, LSL #0x2]\n"
+      "tbz %x[flags], #3, 126f\n"
+      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+      "add x20, x20, x21, LSL #3\n"
+      "ldr x26, [x20, #0x0]\n"
+      "ldr x25, [x20, #0x8]\n"
+      "ldr x24, [x20, #0x10]\n"
+      "ldr x23, [x20, #0x18]\n"
+      "cbnz x28, 127f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x26, x26, x20, LSL #2\n"
+      "add x25, x25, x20, LSL #2\n"
+      "add x24, x24, x20, LSL #2\n"
+      "add x23, x23, x20, LSL #2\n"
+      "b 127f\n"
+      "126:"  // Height 4: setup direct input
+      "mov x26, %x[input_ptr]\n"
+      "add x25, x26, x21, LSL #2\n"
+      "add x24, x25, x21, LSL #2\n"
+      "add x23, x24, x21, LSL #2\n"
+      "127:"  // Height 4: input setup done
+      "cmp x27, #0x4\n"
+      "blt 130f\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      "ld1 { v2.4s }, [x24], #0x10\n"
+      "cmp x27, #0x8\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      "ld1 { v3.4s }, [x23], #0x10\n"
+      "ldr q6, [x12, #0x0]\n"
+      "ldr q7, [x12, #0x10]\n"
+      "blt 129f\n"
+      "128:"  // Height 4: Multiply loop: Main loop head
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
+      "ld1 { v3.4s }, [x23], #0x10\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      "ldr q25, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x10, #0x0]\n"
+      ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec55  // bfmmla v21.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e5aec0a  // bfmmla v10.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec52  // bfmmla v18.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x9, #0x0]\n"
+      ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec56  // bfmmla v22.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e5aec0b  // bfmmla v11.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec53  // bfmmla v19.4s, v2.8h, v26.8h\n"
+      "ldr q6, [x12, #0x0]\n"
+      ".inst 0x6e59ec0f  // bfmmla v15.4s, v0.8h, v25.8h\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      ".inst 0x6e59ec57  // bfmmla v23.4s, v2.8h, v25.8h\n"
+      "ld1 { v2.4s }, [x24], #0x10\n"
+      "ldr q7, [x12, #0x10]\n"
+      "bge 128b\n"
+      "129:"  // Height 4: Multiply loop: Single iteration only
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      "ldr q25, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x10, #0x0]\n"
+      ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec55  // bfmmla v21.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e5aec0a  // bfmmla v10.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec52  // bfmmla v18.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x9, #0x0]\n"
+      ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec56  // bfmmla v22.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e5aec0b  // bfmmla v11.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec53  // bfmmla v19.4s, v2.8h, v26.8h\n"
+      ".inst 0x6e59ec0f  // bfmmla v15.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec57  // bfmmla v23.4s, v2.8h, v25.8h\n"
+      "130:"  // Height 4: Multiply loop: Main loop skip
+      "cbz x27, 133f\n"
+      "cbz x27, 133f\n"
+      "tbz x27, #1, 131f\n"
+      "ldr d0, [x26], #0x8\n"
+      "ldr d1, [x25], #0x8\n"
+      "ldr d2, [x24], #0x8\n"
+      "ldr d3, [x23], #0x8\n"
+      "tbz x27, #0, 132f\n"
+      "ld1 { v0.s }[2], [x26]\n"
+      "ld1 { v1.s }[2], [x25]\n"
+      "ld1 { v2.s }[2], [x24]\n"
+      "ld1 { v3.s }[2], [x23]\n"
+      "b 132f\n"
+      "131:"  // Height 4: Multiply loop: Ragged operand read: partial_1_0
+      "ldr s0, [x26, #0x0]\n"
+      "ldr s1, [x25, #0x0]\n"
+      "ldr s2, [x24, #0x0]\n"
+      "ldr s3, [x23, #0x0]\n"
+      "132:"  // Height 4: Multiply loop: Ragged operand read: Done
+      "ldr q26, [x12, #0x0]\n"
+      "ldr q25, [x12, #0x10]\n"
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
+      ".inst 0x6e5aec08  // bfmmla v8.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec50  // bfmmla v16.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e59ec0c  // bfmmla v12.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec54  // bfmmla v20.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x10, #0x0]\n"
+      ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec55  // bfmmla v21.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e5aec0a  // bfmmla v10.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec52  // bfmmla v18.4s, v2.8h, v26.8h\n"
+      "ldr q26, [x9, #0x0]\n"
+      ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec56  // bfmmla v22.4s, v2.8h, v25.8h\n"
+      "ldr q25, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e5aec0b  // bfmmla v11.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec53  // bfmmla v19.4s, v2.8h, v26.8h\n"
+      ".inst 0x6e59ec0f  // bfmmla v15.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec57  // bfmmla v23.4s, v2.8h, v25.8h\n"
+      "133:"  // Height 4: Multiply loop: No odd multiplies
+      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x28, x28, #0x1\n"
+      "cmp x28, x20\n"
+      "bne 125b\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "uzp1 v6.2d, v8.2d, v12.2d\n"
+      "uzp2 v8.2d, v8.2d, v12.2d\n"
+      "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "uzp2 v9.2d, v9.2d, v13.2d\n"
+      "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "uzp2 v10.2d, v10.2d, v14.2d\n"
+      "add x26, x13, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
+      "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "uzp2 v11.2d, v11.2d, v15.2d\n"
+      "add x24, x25, x20, LSL #2\n"
+      "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "uzp2 v16.2d, v16.2d, v20.2d\n"
+      "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "uzp2 v17.2d, v17.2d, v21.2d\n"
+      "uzp1 v21.2d, v18.2d, v22.2d\n"
+      "uzp2 v18.2d, v18.2d, v22.2d\n"
+      "uzp1 v22.2d, v19.2d, v23.2d\n"
+      "uzp2 v19.2d, v19.2d, v23.2d\n"
+      "tbz %x[flags], #1, 134f\n"
+      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v26.4s }, [x21]\n"
+      "ld1r { v25.4s }, [x20]\n"
+      "fmin v6.4s, v6.4s, v26.4s\n"
+      "fmin v12.4s, v12.4s, v26.4s\n"
+      "fmin v13.4s, v13.4s, v26.4s\n"
+      "fmin v14.4s, v14.4s, v26.4s\n"
+      "fmin v8.4s, v8.4s, v26.4s\n"
+      "fmin v9.4s, v9.4s, v26.4s\n"
+      "fmin v10.4s, v10.4s, v26.4s\n"
+      "fmin v11.4s, v11.4s, v26.4s\n"
+      "fmin v15.4s, v15.4s, v26.4s\n"
+      "fmin v20.4s, v20.4s, v26.4s\n"
+      "fmin v21.4s, v21.4s, v26.4s\n"
+      "fmin v22.4s, v22.4s, v26.4s\n"
+      "fmin v16.4s, v16.4s, v26.4s\n"
+      "fmin v17.4s, v17.4s, v26.4s\n"
+      "fmin v18.4s, v18.4s, v26.4s\n"
+      "fmin v19.4s, v19.4s, v26.4s\n"
+      "fmax v6.4s, v6.4s, v25.4s\n"
+      "fmax v12.4s, v12.4s, v25.4s\n"
+      "fmax v13.4s, v13.4s, v25.4s\n"
+      "fmax v14.4s, v14.4s, v25.4s\n"
+      "fmax v8.4s, v8.4s, v25.4s\n"
+      "fmax v9.4s, v9.4s, v25.4s\n"
+      "fmax v10.4s, v10.4s, v25.4s\n"
+      "fmax v11.4s, v11.4s, v25.4s\n"
+      "fmax v15.4s, v15.4s, v25.4s\n"
+      "fmax v20.4s, v20.4s, v25.4s\n"
+      "fmax v21.4s, v21.4s, v25.4s\n"
+      "fmax v22.4s, v22.4s, v25.4s\n"
+      "fmax v16.4s, v16.4s, v25.4s\n"
+      "fmax v17.4s, v17.4s, v25.4s\n"
+      "fmax v18.4s, v18.4s, v25.4s\n"
+      "fmax v19.4s, v19.4s, v25.4s\n"
+      "134:"  // Height 4: No activation
+      "cmp x14, #0x10\n"
+      "bge 143f\n"
+      "tbz x14, #3, 138f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v9.4s }, [x26], #0x10\n"
+      "st1 { v15.4s }, [x25], #0x10\n"
+      "st1 { v20.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "tbz x14, #2, 136f\n"
+      "st1 { v13.4s }, [x13], #0x10\n"
+      "st1 { v10.4s }, [x26], #0x10\n"
+      "st1 { v21.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "tbz x14, #1, 135f\n"
+      "str d14, [x13], #0x8\n"
+      "str d11, [x26], #0x8\n"
+      "str d22, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "tbz x14, #0, 142f\n"
+      "st1 { v14.s }[2], [x13]\n"
+      "st1 { v11.s }[2], [x26]\n"
+      "st1 { v22.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "b 142f\n"
+      "135:"  // Height 4: Partial direct writeback: partial_1_12
+      "tbz x14, #0, 142f\n"
+      "str s14, [x13, #0x0]\n"
+      "str s11, [x26, #0x0]\n"
+      "str s22, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "b 142f\n"
+      "136:"  // Height 4: Partial direct writeback: partial_2_8
+      "tbz x14, #1, 137f\n"
+      "str d13, [x13], #0x8\n"
+      "str d10, [x26], #0x8\n"
+      "str d21, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "tbz x14, #0, 142f\n"
+      "st1 { v13.s }[2], [x13]\n"
+      "st1 { v10.s }[2], [x26]\n"
+      "st1 { v21.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "b 142f\n"
+      "137:"  // Height 4: Partial direct writeback: partial_1_8
+      "tbz x14, #0, 142f\n"
+      "str s13, [x13, #0x0]\n"
+      "str s10, [x26, #0x0]\n"
+      "str s21, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "b 142f\n"
+      "138:"  // Height 4: Partial direct writeback: partial_4_0
+      "tbz x14, #2, 140f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v15.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "tbz x14, #1, 139f\n"
+      "str d12, [x13], #0x8\n"
+      "str d9, [x26], #0x8\n"
+      "str d20, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "tbz x14, #0, 142f\n"
+      "st1 { v12.s }[2], [x13]\n"
+      "st1 { v9.s }[2], [x26]\n"
+      "st1 { v20.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "b 142f\n"
+      "139:"  // Height 4: Partial direct writeback: partial_1_4
+      "tbz x14, #0, 142f\n"
+      "str s12, [x13, #0x0]\n"
+      "str s9, [x26, #0x0]\n"
+      "str s20, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "b 142f\n"
+      "140:"  // Height 4: Partial direct writeback: partial_2_0
+      "tbz x14, #1, 141f\n"
+      "str d6, [x13], #0x8\n"
+      "str d8, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "tbz x14, #0, 142f\n"
+      "st1 { v6.s }[2], [x13]\n"
+      "st1 { v8.s }[2], [x26]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "b 142f\n"
+      "141:"  // Height 4: Partial direct writeback: partial_1_0
+      "str s6, [x13, #0x0]\n"
+      "str s8, [x26, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "142:"  // Height 4: Partial direct writeback: Done
+      "b 144f\n"
+      "143:"  // Height 4: Full writeback
+      "str q6, [x13, #0x0]\n"
+      "str q12, [x13, #0x10]\n"
+      "str q13, [x13, #0x20]\n"
+      "str q14, [x13, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "str q8, [x26, #0x0]\n"
+      "str q9, [x26, #0x10]\n"
+      "str q10, [x26, #0x20]\n"
+      "str q11, [x26, #0x30]\n"
+      "str q15, [x25, #0x0]\n"
+      "str q20, [x25, #0x10]\n"
+      "str q21, [x25, #0x20]\n"
+      "str q22, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "144:"  // Height 4: Writeback done
+      "subs x14, x14, #0x10\n"
+      "bgt 110b\n"
+      "b 218f\n"
+      "145:"  // Height 5
+      "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "146:"  // Height 5: Column loop
+      "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
+      "cmp x14, #0xc\n"
+      "add x11, x12, x20, LSL #1\n"
+      "add x10, x11, x20, LSL #1\n"
+      "add x9, x10, x20, LSL #1\n"
+      "add x20, x9, x20, LSL #1\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "bgt 147f\n"
+      "cmp x14, #0x8\n"
+      "mov x9, x12\n"
+      "bgt 147f\n"
+      "cmp x14, #0x4\n"
+      "mov x10, x12\n"
+      "bgt 147f\n"
+      "mov x11, x12\n"
+      "147:"  // Height 5: B setup done
+      "cbz x15, 148f\n"
+      "ldr q8, [x15, #0x0]\n"
+      "ldr q9, [x15, #0x10]\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
+      "add x15, x15, #0x40\n"
+      "zip2 v12.2d, v8.2d, v8.2d\n"
+      "zip1 v8.2d, v8.2d, v8.2d\n"
+      "zip2 v13.2d, v9.2d, v9.2d\n"
+      "zip1 v9.2d, v9.2d, v9.2d\n"
+      "zip2 v14.2d, v10.2d, v10.2d\n"
+      "zip1 v10.2d, v10.2d, v10.2d\n"
+      "zip2 v15.2d, v11.2d, v11.2d\n"
+      "zip1 v11.2d, v11.2d, v11.2d\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v20.16b, v12.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "mov v21.16b, v13.16b\n"
+      "mov v18.16b, v10.16b\n"
+      "mov v22.16b, v14.16b\n"
+      "mov v19.16b, v11.16b\n"
+      "mov v23.16b, v15.16b\n"
+      "mov v24.16b, v8.16b\n"
+      "mov v28.16b, v12.16b\n"
+      "mov v25.16b, v9.16b\n"
+      "mov v29.16b, v13.16b\n"
+      "mov v26.16b, v10.16b\n"
+      "mov v30.16b, v14.16b\n"
+      "mov v27.16b, v11.16b\n"
+      "mov v31.16b, v15.16b\n"
+      "b 160f\n"
+      "148:"  // Height 5: no bias
+      "tbz %x[flags], #0, 159f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "cmp x14, #0x10\n"
+      "add x26, x13, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
+      "bge 157f\n"
+      "tbz x14, #3, 152f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v21.4s }, [x24], #0x10\n"
+      "ld1 { v26.4s }, [x23], #0x10\n"
+      "tbz x14, #2, 150f\n"
+      "ld1 { v11.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v19.4s }, [x25], #0x10\n"
+      "ld1 { v22.4s }, [x24], #0x10\n"
+      "ld1 { v27.4s }, [x23], #0x10\n"
+      "tbz x14, #1, 149f\n"
+      "ldr d16, [x13], #0x8\n"
+      "ldr d15, [x26], #0x8\n"
+      "mov x20, #0x38\n"
+      "ldr d24, [x25], #0x8\n"
+      "ldr d23, [x24], #0x8\n"
+      "ldr d6, [x23], #0x8\n"
+      "tbz x14, #0, 156f\n"
+      "ld1 { v16.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v24.s }[2], [x25]\n"
+      "ld1 { v23.s }[2], [x24]\n"
+      "ld1 { v6.s }[2], [x23]\n"
+      "b 156f\n"
+      "149:"  // Height 5: Partial accumulate: partial_1_12
+      "mov x20, #0x30\n"
+      "tbz x14, #0, 156f\n"
+      "ldr s16, [x13, #0x0]\n"
+      "ldr s15, [x26, #0x0]\n"
+      "ldr s24, [x25, #0x0]\n"
+      "ldr s23, [x24, #0x0]\n"
+      "ldr s6, [x23, #0x0]\n"
+      "b 156f\n"
+      "150:"  // Height 5: Partial accumulate: partial_2_8
+      "tbz x14, #1, 151f\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d14, [x26], #0x8\n"
+      "mov x20, #0x28\n"
+      "ldr d19, [x25], #0x8\n"
+      "ldr d22, [x24], #0x8\n"
+      "ldr d27, [x23], #0x8\n"
+      "tbz x14, #0, 156f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v22.s }[2], [x24]\n"
+      "ld1 { v27.s }[2], [x23]\n"
+      "b 156f\n"
+      "151:"  // Height 5: Partial accumulate: partial_1_8
+      "mov x20, #0x20\n"
+      "tbz x14, #0, 156f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s14, [x26, #0x0]\n"
+      "ldr s19, [x25, #0x0]\n"
+      "ldr s22, [x24, #0x0]\n"
+      "ldr s27, [x23, #0x0]\n"
+      "b 156f\n"
+      "152:"  // Height 5: Partial accumulate: partial_4_0
+      "tbz x14, #2, 154f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v25.4s }, [x23], #0x10\n"
+      "tbz x14, #1, 153f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d13, [x26], #0x8\n"
+      "mov x20, #0x18\n"
+      "ldr d18, [x25], #0x8\n"
+      "ldr d21, [x24], #0x8\n"
+      "ldr d26, [x23], #0x8\n"
+      "tbz x14, #0, 156f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v26.s }[2], [x23]\n"
+      "b 156f\n"
+      "153:"  // Height 5: Partial accumulate: partial_1_4
+      "mov x20, #0x10\n"
+      "tbz x14, #0, 156f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s13, [x26, #0x0]\n"
+      "ldr s18, [x25, #0x0]\n"
+      "ldr s21, [x24, #0x0]\n"
+      "ldr s26, [x23, #0x0]\n"
+      "b 156f\n"
+      "154:"  // Height 5: Partial accumulate: partial_2_0
+      "tbz x14, #1, 155f\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d12, [x26], #0x8\n"
+      "mov x20, #0x8\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d20, [x24], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "tbz x14, #0, 156f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v20.s }[2], [x24]\n"
+      "ld1 { v25.s }[2], [x23]\n"
+      "b 156f\n"
+      "155:"  // Height 5: Partial accumulate: partial_1_0
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s12, [x26, #0x0]\n"
+      "mov x20, #0x0\n"
+      "ldr s17, [x25, #0x0]\n"
+      "ldr s20, [x24, #0x0]\n"
+      "ldr s25, [x23, #0x0]\n"
+      "156:"  // Height 5: Partial accumulate: Done
+      "sub x13, x13, x20\n"
+      "b 158f\n"
+      "157:"  // Height 5: full accumulate
+      "ldr q9, [x13, #0x0]\n"
+      "ldr q10, [x13, #0x10]\n"
+      "ldr q11, [x13, #0x20]\n"
+      "ldr q16, [x13, #0x30]\n"
+      "ldr q12, [x26, #0x0]\n"
+      "ldr q13, [x26, #0x10]\n"
+      "ldr q14, [x26, #0x20]\n"
+      "ldr q15, [x26, #0x30]\n"
+      "ldr q17, [x25, #0x0]\n"
+      "ldr q18, [x25, #0x10]\n"
+      "ldr q19, [x25, #0x20]\n"
+      "ldr q24, [x25, #0x30]\n"
+      "ldr q20, [x24, #0x0]\n"
+      "ldr q21, [x24, #0x10]\n"
+      "ldr q22, [x24, #0x20]\n"
+      "ldr q23, [x24, #0x30]\n"
+      "ldr q25, [x23, #0x0]\n"
+      "ldr q26, [x23, #0x10]\n"
+      "ldr q27, [x23, #0x20]\n"
+      "ldr q6, [x23, #0x30]\n"
+      "158:"  // Height 5: MMLA fixup
+      "zip1 v8.2d, v9.2d, v12.2d\n"
+      "zip2 v12.2d, v9.2d, v12.2d\n"
+      "zip1 v9.2d, v10.2d, v13.2d\n"
+      "zip2 v13.2d, v10.2d, v13.2d\n"
+      "zip1 v10.2d, v11.2d, v14.2d\n"
+      "zip2 v14.2d, v11.2d, v14.2d\n"
+      "zip1 v11.2d, v16.2d, v15.2d\n"
+      "zip2 v15.2d, v16.2d, v15.2d\n"
+      "zip1 v16.2d, v17.2d, v20.2d\n"
+      "zip2 v20.2d, v17.2d, v20.2d\n"
+      "zip1 v17.2d, v18.2d, v21.2d\n"
+      "zip2 v21.2d, v18.2d, v21.2d\n"
+      "zip1 v18.2d, v19.2d, v22.2d\n"
+      "zip2 v22.2d, v19.2d, v22.2d\n"
+      "zip1 v19.2d, v24.2d, v23.2d\n"
+      "zip2 v23.2d, v24.2d, v23.2d\n"
+      "zip1 v24.2d, v25.2d, v28.2d\n"
+      "zip2 v28.2d, v25.2d, v28.2d\n"
+      "zip1 v25.2d, v26.2d, v29.2d\n"
+      "zip2 v29.2d, v26.2d, v29.2d\n"
+      "zip1 v26.2d, v27.2d, v30.2d\n"
+      "zip2 v30.2d, v27.2d, v30.2d\n"
+      "zip1 v27.2d, v6.2d, v31.2d\n"
+      "zip2 v31.2d, v6.2d, v31.2d\n"
+      "b 160f\n"
+      "159:"  // 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"
+      "movi v28.16b, #0x0\n"
+      "movi v29.16b, #0x0\n"
+      "movi v30.16b, #0x0\n"
+      "movi v31.16b, #0x0\n"
+      "160:"  // Height 5: setup done
+      "mov x28, #0x0\n"
+      "161:"  // Height 5: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w27, [x20, x28, LSL #0x2]\n"
+      "tbz %x[flags], #3, 162f\n"
+      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+      "add x20, x20, x21, LSL #3\n"
+      "ldr x26, [x20, #0x0]\n"
+      "ldr x25, [x20, #0x8]\n"
+      "ldr x24, [x20, #0x10]\n"
+      "ldr x23, [x20, #0x18]\n"
+      "ldr x22, [x20, #0x20]\n"
+      "cbnz x28, 163f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x26, x26, x20, LSL #2\n"
+      "add x25, x25, x20, LSL #2\n"
+      "add x24, x24, x20, LSL #2\n"
+      "add x23, x23, x20, LSL #2\n"
+      "add x22, x22, x20, LSL #2\n"
+      "b 163f\n"
+      "162:"  // Height 5: setup direct input
+      "mov x26, %x[input_ptr]\n"
+      "add x25, x26, x21, LSL #2\n"
+      "add x24, x25, x21, LSL #2\n"
+      "add x23, x24, x21, LSL #2\n"
+      "add x22, x23, x21, LSL #2\n"
+      "163:"  // Height 5: input setup done
+      "cmp x27, #0x4\n"
+      "blt 166f\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      "ld1 { v2.4s }, [x24], #0x10\n"
+      "cmp x27, #0x8\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      "ld1 { v3.4s }, [x23], #0x10\n"
+      "ld1 { v4.4s }, [x22], #0x10\n"
+      "ldr q6, [x12, #0x0]\n"
+      "ldr q7, [x12, #0x10]\n"
+      "blt 165f\n"
+      "164:"  // Height 5: Multiply loop: Main loop head
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
+      "ld1 { v3.4s }, [x23], #0x10\n"
+      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
+      ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      "ldr q6, [x11, #0x0]\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      "ldr q5, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec51  // bfmmla v17.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec99  // bfmmla v25.4s, v4.8h, v6.8h\n"
+      "ldr q6, [x10, #0x0]\n"
+      ".inst 0x6e45ec0d  // bfmmla v13.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec55  // bfmmla v21.4s, v2.8h, v5.8h\n"
+      ".inst 0x6e45ec9d  // bfmmla v29.4s, v4.8h, v5.8h\n"
+      "ldr q5, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e46ec0a  // bfmmla v10.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec52  // bfmmla v18.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec9a  // bfmmla v26.4s, v4.8h, v6.8h\n"
+      "ldr q6, [x9, #0x0]\n"
+      ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec56  // bfmmla v22.4s, v2.8h, v5.8h\n"
+      ".inst 0x6e45ec9e  // bfmmla v30.4s, v4.8h, v5.8h\n"
+      "ldr q5, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e46ec0b  // bfmmla v11.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec53  // bfmmla v19.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec9b  // bfmmla v27.4s, v4.8h, v6.8h\n"
+      "ldr q6, [x12, #0x0]\n"
+      ".inst 0x6e45ec0f  // bfmmla v15.4s, v0.8h, v5.8h\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      ".inst 0x6e45ec57  // bfmmla v23.4s, v2.8h, v5.8h\n"
+      "ld1 { v2.4s }, [x24], #0x10\n"
+      ".inst 0x6e45ec9f  // bfmmla v31.4s, v4.8h, v5.8h\n"
+      "ld1 { v4.4s }, [x22], #0x10\n"
+      "ldr q7, [x12, #0x10]\n"
+      "bge 164b\n"
+      "165:"  // Height 5: Multiply loop: Single iteration only
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
+      ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      "ldr q3, [x11, #0x0]\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      "ldr q1, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec51  // bfmmla v17.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec99  // bfmmla v25.4s, v4.8h, v3.8h\n"
+      "ldr q3, [x10, #0x0]\n"
+      ".inst 0x6e41ec0d  // bfmmla v13.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec55  // bfmmla v21.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9d  // bfmmla v29.4s, v4.8h, v1.8h\n"
+      "ldr q1, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e43ec0a  // bfmmla v10.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec52  // bfmmla v18.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec9a  // bfmmla v26.4s, v4.8h, v3.8h\n"
+      "ldr q3, [x9, #0x0]\n"
+      ".inst 0x6e41ec0e  // bfmmla v14.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec56  // bfmmla v22.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9e  // bfmmla v30.4s, v4.8h, v1.8h\n"
+      "ldr q1, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e43ec0b  // bfmmla v11.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec53  // bfmmla v19.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec9b  // bfmmla v27.4s, v4.8h, v3.8h\n"
+      ".inst 0x6e41ec0f  // bfmmla v15.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec57  // bfmmla v23.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9f  // bfmmla v31.4s, v4.8h, v1.8h\n"
+      "166:"  // Height 5: Multiply loop: Main loop skip
+      "cbz x27, 169f\n"
+      "cbz x27, 169f\n"
+      "tbz x27, #1, 167f\n"
+      "ldr d0, [x26], #0x8\n"
+      "ldr d1, [x25], #0x8\n"
+      "ldr d2, [x24], #0x8\n"
+      "ldr d3, [x23], #0x8\n"
+      "ldr d4, [x22], #0x8\n"
+      "tbz x27, #0, 168f\n"
+      "ld1 { v0.s }[2], [x26]\n"
+      "ld1 { v1.s }[2], [x25]\n"
+      "ld1 { v2.s }[2], [x24]\n"
+      "ld1 { v3.s }[2], [x23]\n"
+      "ld1 { v4.s }[2], [x22]\n"
+      "b 168f\n"
+      "167:"  // Height 5: Multiply loop: Ragged operand read: partial_1_0
+      "ldr s0, [x26, #0x0]\n"
+      "ldr s1, [x25, #0x0]\n"
+      "ldr s2, [x24, #0x0]\n"
+      "ldr s3, [x23, #0x0]\n"
+      "ldr s4, [x22, #0x0]\n"
+      "168:"  // Height 5: Multiply loop: Ragged operand read: Done
+      "ldr q6, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
+      ".inst 0x6e45ec0c  // bfmmla v12.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec9c  // bfmmla v28.4s, v4.8h, v5.8h\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      "ldr q3, [x11, #0x0]\n"
+      ".inst 0x6e45ec54  // bfmmla v20.4s, v2.8h, v5.8h\n"
+      "ldr q1, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec51  // bfmmla v17.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec99  // bfmmla v25.4s, v4.8h, v3.8h\n"
+      "ldr q3, [x10, #0x0]\n"
+      ".inst 0x6e41ec0d  // bfmmla v13.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec55  // bfmmla v21.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9d  // bfmmla v29.4s, v4.8h, v1.8h\n"
+      "ldr q1, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e43ec0a  // bfmmla v10.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec52  // bfmmla v18.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec9a  // bfmmla v26.4s, v4.8h, v3.8h\n"
+      "ldr q3, [x9, #0x0]\n"
+      ".inst 0x6e41ec0e  // bfmmla v14.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec56  // bfmmla v22.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9e  // bfmmla v30.4s, v4.8h, v1.8h\n"
+      "ldr q1, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e43ec0b  // bfmmla v11.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec53  // bfmmla v19.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec9b  // bfmmla v27.4s, v4.8h, v3.8h\n"
+      ".inst 0x6e41ec0f  // bfmmla v15.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec57  // bfmmla v23.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9f  // bfmmla v31.4s, v4.8h, v1.8h\n"
+      "169:"  // Height 5: Multiply loop: No odd multiplies
+      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x28, x28, #0x1\n"
+      "cmp x28, x20\n"
+      "bne 161b\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "uzp1 v6.2d, v8.2d, v12.2d\n"
+      "uzp2 v8.2d, v8.2d, v12.2d\n"
+      "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "uzp2 v9.2d, v9.2d, v13.2d\n"
+      "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "uzp2 v10.2d, v10.2d, v14.2d\n"
+      "add x26, x13, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
+      "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "uzp2 v11.2d, v11.2d, v15.2d\n"
+      "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "uzp2 v16.2d, v16.2d, v20.2d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "uzp2 v17.2d, v17.2d, v21.2d\n"
+      "uzp1 v21.2d, v18.2d, v22.2d\n"
+      "uzp2 v18.2d, v18.2d, v22.2d\n"
+      "uzp1 v22.2d, v19.2d, v23.2d\n"
+      "uzp2 v19.2d, v19.2d, v23.2d\n"
+      "uzp1 v24.2d, v24.2d, v28.2d\n"
+      "uzp1 v25.2d, v25.2d, v29.2d\n"
+      "uzp1 v26.2d, v26.2d, v30.2d\n"
+      "uzp1 v27.2d, v27.2d, v31.2d\n"
+      "tbz %x[flags], #1, 170f\n"
+      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v1.4s }, [x21]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "fmin v6.4s, v6.4s, v1.4s\n"
+      "fmin v12.4s, v12.4s, v1.4s\n"
+      "fmin v13.4s, v13.4s, v1.4s\n"
+      "fmin v14.4s, v14.4s, v1.4s\n"
+      "fmin v8.4s, v8.4s, v1.4s\n"
+      "fmin v9.4s, v9.4s, v1.4s\n"
+      "fmin v10.4s, v10.4s, v1.4s\n"
+      "fmin v11.4s, v11.4s, v1.4s\n"
+      "fmin v15.4s, v15.4s, v1.4s\n"
+      "fmin v20.4s, v20.4s, v1.4s\n"
+      "fmin v21.4s, v21.4s, v1.4s\n"
+      "fmin v22.4s, v22.4s, v1.4s\n"
+      "fmin v16.4s, v16.4s, v1.4s\n"
+      "fmin v17.4s, v17.4s, v1.4s\n"
+      "fmin v18.4s, v18.4s, v1.4s\n"
+      "fmin v19.4s, v19.4s, v1.4s\n"
+      "fmin v24.4s, v24.4s, v1.4s\n"
+      "fmin v25.4s, v25.4s, v1.4s\n"
+      "fmin v26.4s, v26.4s, v1.4s\n"
+      "fmin v27.4s, v27.4s, v1.4s\n"
+      "fmax v6.4s, v6.4s, v0.4s\n"
+      "fmax v12.4s, v12.4s, v0.4s\n"
+      "fmax v13.4s, v13.4s, v0.4s\n"
+      "fmax v14.4s, v14.4s, v0.4s\n"
+      "fmax v8.4s, v8.4s, v0.4s\n"
+      "fmax v9.4s, v9.4s, v0.4s\n"
+      "fmax v10.4s, v10.4s, v0.4s\n"
+      "fmax v11.4s, v11.4s, v0.4s\n"
+      "fmax v15.4s, v15.4s, v0.4s\n"
+      "fmax v20.4s, v20.4s, v0.4s\n"
+      "fmax v21.4s, v21.4s, v0.4s\n"
+      "fmax v22.4s, v22.4s, v0.4s\n"
+      "fmax v16.4s, v16.4s, v0.4s\n"
+      "fmax v17.4s, v17.4s, v0.4s\n"
+      "fmax v18.4s, v18.4s, v0.4s\n"
+      "fmax v19.4s, v19.4s, v0.4s\n"
+      "fmax v24.4s, v24.4s, v0.4s\n"
+      "fmax v25.4s, v25.4s, v0.4s\n"
+      "fmax v26.4s, v26.4s, v0.4s\n"
+      "fmax v27.4s, v27.4s, v0.4s\n"
+      "170:"  // Height 5: No activation
+      "cmp x14, #0x10\n"
+      "bge 179f\n"
+      "tbz x14, #3, 174f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v9.4s }, [x26], #0x10\n"
+      "st1 { v15.4s }, [x25], #0x10\n"
+      "st1 { v20.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v25.4s }, [x23], #0x10\n"
+      "tbz x14, #2, 172f\n"
+      "st1 { v13.4s }, [x13], #0x10\n"
+      "st1 { v10.4s }, [x26], #0x10\n"
+      "st1 { v21.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v26.4s }, [x23], #0x10\n"
+      "tbz x14, #1, 171f\n"
+      "str d14, [x13], #0x8\n"
+      "str d11, [x26], #0x8\n"
+      "str d22, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d27, [x23], #0x8\n"
+      "tbz x14, #0, 178f\n"
+      "st1 { v14.s }[2], [x13]\n"
+      "st1 { v11.s }[2], [x26]\n"
+      "st1 { v22.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v27.s }[2], [x23]\n"
+      "b 178f\n"
+      "171:"  // Height 5: Partial direct writeback: partial_1_12
+      "tbz x14, #0, 178f\n"
+      "str s14, [x13, #0x0]\n"
+      "str s11, [x26, #0x0]\n"
+      "str s22, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s27, [x23, #0x0]\n"
+      "b 178f\n"
+      "172:"  // Height 5: Partial direct writeback: partial_2_8
+      "tbz x14, #1, 173f\n"
+      "str d13, [x13], #0x8\n"
+      "str d10, [x26], #0x8\n"
+      "str d21, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d26, [x23], #0x8\n"
+      "tbz x14, #0, 178f\n"
+      "st1 { v13.s }[2], [x13]\n"
+      "st1 { v10.s }[2], [x26]\n"
+      "st1 { v21.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v26.s }[2], [x23]\n"
+      "b 178f\n"
+      "173:"  // Height 5: Partial direct writeback: partial_1_8
+      "tbz x14, #0, 178f\n"
+      "str s13, [x13, #0x0]\n"
+      "str s10, [x26, #0x0]\n"
+      "str s21, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s26, [x23, #0x0]\n"
+      "b 178f\n"
+      "174:"  // Height 5: Partial direct writeback: partial_4_0
+      "tbz x14, #2, 176f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v15.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v24.4s }, [x23], #0x10\n"
+      "tbz x14, #1, 175f\n"
+      "str d12, [x13], #0x8\n"
+      "str d9, [x26], #0x8\n"
+      "str d20, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d25, [x23], #0x8\n"
+      "tbz x14, #0, 178f\n"
+      "st1 { v12.s }[2], [x13]\n"
+      "st1 { v9.s }[2], [x26]\n"
+      "st1 { v20.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v25.s }[2], [x23]\n"
+      "b 178f\n"
+      "175:"  // Height 5: Partial direct writeback: partial_1_4
+      "tbz x14, #0, 178f\n"
+      "str s12, [x13, #0x0]\n"
+      "str s9, [x26, #0x0]\n"
+      "str s20, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s25, [x23, #0x0]\n"
+      "b 178f\n"
+      "176:"  // Height 5: Partial direct writeback: partial_2_0
+      "tbz x14, #1, 177f\n"
+      "str d6, [x13], #0x8\n"
+      "str d8, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d24, [x23], #0x8\n"
+      "tbz x14, #0, 178f\n"
+      "st1 { v6.s }[2], [x13]\n"
+      "st1 { v8.s }[2], [x26]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v24.s }[2], [x23]\n"
+      "b 178f\n"
+      "177:"  // Height 5: Partial direct writeback: partial_1_0
+      "str s6, [x13, #0x0]\n"
+      "str s8, [x26, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s24, [x23, #0x0]\n"
+      "178:"  // Height 5: Partial direct writeback: Done
+      "b 180f\n"
+      "179:"  // Height 5: Full writeback
+      "str q6, [x13, #0x0]\n"
+      "str q12, [x13, #0x10]\n"
+      "str q13, [x13, #0x20]\n"
+      "str q14, [x13, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "str q8, [x26, #0x0]\n"
+      "str q9, [x26, #0x10]\n"
+      "str q10, [x26, #0x20]\n"
+      "str q11, [x26, #0x30]\n"
+      "str q15, [x25, #0x0]\n"
+      "str q20, [x25, #0x10]\n"
+      "str q21, [x25, #0x20]\n"
+      "str q22, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q24, [x23, #0x0]\n"
+      "str q25, [x23, #0x10]\n"
+      "str q26, [x23, #0x20]\n"
+      "str q27, [x23, #0x30]\n"
+      "180:"  // Height 5: Writeback done
+      "subs x14, x14, #0x10\n"
+      "bgt 146b\n"
+      "b 218f\n"
+      "181:"  // Height 6
+      "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x21, #0x18\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd x21, x20, x21, x13\n"
+      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "182:"  // Height 6: Column loop
+      "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
+      "cmp x14, #0xc\n"
+      "add x11, x12, x20, LSL #1\n"
+      "add x10, x11, x20, LSL #1\n"
+      "add x9, x10, x20, LSL #1\n"
+      "add x20, x9, x20, LSL #1\n"
+      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "bgt 183f\n"
+      "cmp x14, #0x8\n"
+      "mov x9, x12\n"
+      "bgt 183f\n"
+      "cmp x14, #0x4\n"
+      "mov x10, x12\n"
+      "bgt 183f\n"
+      "mov x11, x12\n"
+      "183:"  // Height 6: B setup done
+      "cbz x15, 184f\n"
+      "ldr q8, [x15, #0x0]\n"
+      "ldr q9, [x15, #0x10]\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
+      "add x15, x15, #0x40\n"
+      "zip2 v12.2d, v8.2d, v8.2d\n"
+      "zip1 v8.2d, v8.2d, v8.2d\n"
+      "zip2 v13.2d, v9.2d, v9.2d\n"
+      "zip1 v9.2d, v9.2d, v9.2d\n"
+      "zip2 v14.2d, v10.2d, v10.2d\n"
+      "zip1 v10.2d, v10.2d, v10.2d\n"
+      "zip2 v15.2d, v11.2d, v11.2d\n"
+      "zip1 v11.2d, v11.2d, v11.2d\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v20.16b, v12.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "mov v21.16b, v13.16b\n"
+      "mov v18.16b, v10.16b\n"
+      "mov v22.16b, v14.16b\n"
+      "mov v19.16b, v11.16b\n"
+      "mov v23.16b, v15.16b\n"
+      "mov v24.16b, v8.16b\n"
+      "mov v28.16b, v12.16b\n"
+      "mov v25.16b, v9.16b\n"
+      "mov v29.16b, v13.16b\n"
+      "mov v26.16b, v10.16b\n"
+      "mov v30.16b, v14.16b\n"
+      "mov v27.16b, v11.16b\n"
+      "mov v31.16b, v15.16b\n"
+      "b 196f\n"
+      "184:"  // Height 6: no bias
+      "tbz %x[flags], #0, 195f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "cmp x14, #0x10\n"
+      "add x26, x13, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
+      "bge 193f\n"
+      "tbz x14, #3, 188f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v10.4s }, [x13], #0x10\n"
+      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v21.4s }, [x24], #0x10\n"
+      "ld1 { v26.4s }, [x23], #0x10\n"
+      "ld1 { v29.4s }, [x22], #0x10\n"
+      "tbz x14, #2, 186f\n"
+      "ld1 { v11.4s }, [x13], #0x10\n"
+      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v19.4s }, [x25], #0x10\n"
+      "ld1 { v22.4s }, [x24], #0x10\n"
+      "ld1 { v27.4s }, [x23], #0x10\n"
+      "ld1 { v30.4s }, [x22], #0x10\n"
+      "tbz x14, #1, 185f\n"
+      "ldr d16, [x13], #0x8\n"
+      "ldr d15, [x26], #0x8\n"
+      "mov x20, #0x38\n"
+      "ldr d24, [x25], #0x8\n"
+      "ldr d23, [x24], #0x8\n"
+      "ldr d6, [x23], #0x8\n"
+      "ldr d31, [x22], #0x8\n"
+      "tbz x14, #0, 192f\n"
+      "ld1 { v16.s }[2], [x13]\n"
+      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v24.s }[2], [x25]\n"
+      "ld1 { v23.s }[2], [x24]\n"
+      "ld1 { v6.s }[2], [x23]\n"
+      "ld1 { v31.s }[2], [x22]\n"
+      "b 192f\n"
+      "185:"  // Height 6: Partial accumulate: partial_1_12
+      "mov x20, #0x30\n"
+      "tbz x14, #0, 192f\n"
+      "ldr s16, [x13, #0x0]\n"
+      "ldr s15, [x26, #0x0]\n"
+      "ldr s24, [x25, #0x0]\n"
+      "ldr s23, [x24, #0x0]\n"
+      "ldr s6, [x23, #0x0]\n"
+      "ldr s31, [x22, #0x0]\n"
+      "b 192f\n"
+      "186:"  // Height 6: Partial accumulate: partial_2_8
+      "tbz x14, #1, 187f\n"
+      "ldr d11, [x13], #0x8\n"
+      "ldr d14, [x26], #0x8\n"
+      "mov x20, #0x28\n"
+      "ldr d19, [x25], #0x8\n"
+      "ldr d22, [x24], #0x8\n"
+      "ldr d27, [x23], #0x8\n"
+      "ldr d30, [x22], #0x8\n"
+      "tbz x14, #0, 192f\n"
+      "ld1 { v11.s }[2], [x13]\n"
+      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v22.s }[2], [x24]\n"
+      "ld1 { v27.s }[2], [x23]\n"
+      "ld1 { v30.s }[2], [x22]\n"
+      "b 192f\n"
+      "187:"  // Height 6: Partial accumulate: partial_1_8
+      "mov x20, #0x20\n"
+      "tbz x14, #0, 192f\n"
+      "ldr s11, [x13, #0x0]\n"
+      "ldr s14, [x26, #0x0]\n"
+      "ldr s19, [x25, #0x0]\n"
+      "ldr s22, [x24, #0x0]\n"
+      "ldr s27, [x23, #0x0]\n"
+      "ldr s30, [x22, #0x0]\n"
+      "b 192f\n"
+      "188:"  // Height 6: Partial accumulate: partial_4_0
+      "tbz x14, #2, 190f\n"
+      "ld1 { v9.4s }, [x13], #0x10\n"
+      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v28.4s }, [x22], #0x10\n"
+      "tbz x14, #1, 189f\n"
+      "ldr d10, [x13], #0x8\n"
+      "ldr d13, [x26], #0x8\n"
+      "mov x20, #0x18\n"
+      "ldr d18, [x25], #0x8\n"
+      "ldr d21, [x24], #0x8\n"
+      "ldr d26, [x23], #0x8\n"
+      "ldr d29, [x22], #0x8\n"
+      "tbz x14, #0, 192f\n"
+      "ld1 { v10.s }[2], [x13]\n"
+      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v26.s }[2], [x23]\n"
+      "ld1 { v29.s }[2], [x22]\n"
+      "b 192f\n"
+      "189:"  // Height 6: Partial accumulate: partial_1_4
+      "mov x20, #0x10\n"
+      "tbz x14, #0, 192f\n"
+      "ldr s10, [x13, #0x0]\n"
+      "ldr s13, [x26, #0x0]\n"
+      "ldr s18, [x25, #0x0]\n"
+      "ldr s21, [x24, #0x0]\n"
+      "ldr s26, [x23, #0x0]\n"
+      "ldr s29, [x22, #0x0]\n"
+      "b 192f\n"
+      "190:"  // Height 6: Partial accumulate: partial_2_0
+      "tbz x14, #1, 191f\n"
+      "ldr d9, [x13], #0x8\n"
+      "ldr d12, [x26], #0x8\n"
+      "mov x20, #0x8\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d20, [x24], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "ldr d28, [x22], #0x8\n"
+      "tbz x14, #0, 192f\n"
+      "ld1 { v9.s }[2], [x13]\n"
+      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v20.s }[2], [x24]\n"
+      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v28.s }[2], [x22]\n"
+      "b 192f\n"
+      "191:"  // Height 6: Partial accumulate: partial_1_0
+      "ldr s9, [x13, #0x0]\n"
+      "ldr s12, [x26, #0x0]\n"
+      "mov x20, #0x0\n"
+      "ldr s17, [x25, #0x0]\n"
+      "ldr s20, [x24, #0x0]\n"
+      "ldr s25, [x23, #0x0]\n"
+      "ldr s28, [x22, #0x0]\n"
+      "192:"  // Height 6: Partial accumulate: Done
+      "sub x13, x13, x20\n"
+      "b 194f\n"
+      "193:"  // Height 6: full accumulate
+      "ldr q9, [x13, #0x0]\n"
+      "ldr q10, [x13, #0x10]\n"
+      "ldr q11, [x13, #0x20]\n"
+      "ldr q16, [x13, #0x30]\n"
+      "ldr q12, [x26, #0x0]\n"
+      "ldr q13, [x26, #0x10]\n"
+      "ldr q14, [x26, #0x20]\n"
+      "ldr q15, [x26, #0x30]\n"
+      "ldr q17, [x25, #0x0]\n"
+      "ldr q18, [x25, #0x10]\n"
+      "ldr q19, [x25, #0x20]\n"
+      "ldr q24, [x25, #0x30]\n"
+      "ldr q20, [x24, #0x0]\n"
+      "ldr q21, [x24, #0x10]\n"
+      "ldr q22, [x24, #0x20]\n"
+      "ldr q23, [x24, #0x30]\n"
+      "ldr q25, [x23, #0x0]\n"
+      "ldr q26, [x23, #0x10]\n"
+      "ldr q27, [x23, #0x20]\n"
+      "ldr q6, [x23, #0x30]\n"
+      "ldr q28, [x22, #0x0]\n"
+      "ldr q29, [x22, #0x10]\n"
+      "ldr q30, [x22, #0x20]\n"
+      "ldr q31, [x22, #0x30]\n"
+      "194:"  // Height 6: MMLA fixup
+      "zip1 v8.2d, v9.2d, v12.2d\n"
+      "zip2 v12.2d, v9.2d, v12.2d\n"
+      "zip1 v9.2d, v10.2d, v13.2d\n"
+      "zip2 v13.2d, v10.2d, v13.2d\n"
+      "zip1 v10.2d, v11.2d, v14.2d\n"
+      "zip2 v14.2d, v11.2d, v14.2d\n"
+      "zip1 v11.2d, v16.2d, v15.2d\n"
+      "zip2 v15.2d, v16.2d, v15.2d\n"
+      "zip1 v16.2d, v17.2d, v20.2d\n"
+      "zip2 v20.2d, v17.2d, v20.2d\n"
+      "zip1 v17.2d, v18.2d, v21.2d\n"
+      "zip2 v21.2d, v18.2d, v21.2d\n"
+      "zip1 v18.2d, v19.2d, v22.2d\n"
+      "zip2 v22.2d, v19.2d, v22.2d\n"
+      "zip1 v19.2d, v24.2d, v23.2d\n"
+      "zip2 v23.2d, v24.2d, v23.2d\n"
+      "zip1 v24.2d, v25.2d, v28.2d\n"
+      "zip2 v28.2d, v25.2d, v28.2d\n"
+      "zip1 v25.2d, v26.2d, v29.2d\n"
+      "zip2 v29.2d, v26.2d, v29.2d\n"
+      "zip1 v26.2d, v27.2d, v30.2d\n"
+      "zip2 v30.2d, v27.2d, v30.2d\n"
+      "zip1 v27.2d, v6.2d, v31.2d\n"
+      "zip2 v31.2d, v6.2d, v31.2d\n"
+      "b 196f\n"
+      "195:"  // 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"
+      "196:"  // Height 6: setup done
+      "mov x28, #0x0\n"
+      "197:"  // Height 6: String loop
+      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr w27, [x20, x28, LSL #0x2]\n"
+      "tbz %x[flags], #3, 198f\n"
+      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+      "add x20, x20, x21, LSL #3\n"
+      "ldr x26, [x20, #0x0]\n"
+      "ldr x25, [x20, #0x8]\n"
+      "ldr x24, [x20, #0x10]\n"
+      "ldr x23, [x20, #0x18]\n"
+      "ldr x22, [x20, #0x20]\n"
+      "ldr x21, [x20, #0x28]\n"
+      "cbnz x28, 199f\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x26, x26, x20, LSL #2\n"
+      "add x25, x25, x20, LSL #2\n"
+      "add x24, x24, x20, LSL #2\n"
+      "add x23, x23, x20, LSL #2\n"
+      "add x22, x22, x20, LSL #2\n"
+      "add x21, x21, x20, LSL #2\n"
+      "b 199f\n"
+      "198:"  // Height 6: setup direct input
+      "mov x26, %x[input_ptr]\n"
+      "add x25, x26, x21, LSL #2\n"
+      "add x24, x25, x21, LSL #2\n"
+      "add x23, x24, x21, LSL #2\n"
+      "add x22, x23, x21, LSL #2\n"
+      "add x21, x22, x21, LSL #2\n"
+      "199:"  // Height 6: input setup done
+      "cmp x27, #0x4\n"
+      "blt 202f\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      "ld1 { v2.4s }, [x24], #0x10\n"
+      "cmp x27, #0x8\n"
+      "ld1 { v4.4s }, [x22], #0x10\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      "ld1 { v3.4s }, [x23], #0x10\n"
+      "ld1 { v5.4s }, [x21], #0x10\n"
+      "ldr q6, [x12, #0x0]\n"
+      "ldr q7, [x12, #0x10]\n"
+      "blt 201f\n"
+      "200:"  // Height 6: Multiply loop: Main loop head
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
+      "ld1 { v3.4s }, [x23], #0x10\n"
+      ".inst 0x4ea168a4  // bfcvtn2 v4.8h, v5.4s\n"
+      "ld1 { v5.4s }, [x21], #0x10\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
+      "ldr q6, [x11, #0x0]\n"
+      ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
+      "ldr q7, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec51  // bfmmla v17.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec99  // bfmmla v25.4s, v4.8h, v6.8h\n"
+      "ldr q6, [x10, #0x0]\n"
+      ".inst 0x6e47ec0d  // bfmmla v13.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e47ec55  // bfmmla v21.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e47ec9d  // bfmmla v29.4s, v4.8h, v7.8h\n"
+      "ldr q7, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e46ec0a  // bfmmla v10.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec52  // bfmmla v18.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec9a  // bfmmla v26.4s, v4.8h, v6.8h\n"
+      "ldr q6, [x9, #0x0]\n"
+      ".inst 0x6e47ec0e  // bfmmla v14.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e47ec56  // bfmmla v22.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e47ec9e  // bfmmla v30.4s, v4.8h, v7.8h\n"
+      "ldr q7, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e46ec0b  // bfmmla v11.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e46ec53  // bfmmla v19.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec9b  // bfmmla v27.4s, v4.8h, v6.8h\n"
+      "ldr q6, [x12, #0x0]\n"
+      ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      ".inst 0x6e47ec57  // bfmmla v23.4s, v2.8h, v7.8h\n"
+      "ld1 { v2.4s }, [x24], #0x10\n"
+      ".inst 0x6e47ec9f  // bfmmla v31.4s, v4.8h, v7.8h\n"
+      "ld1 { v4.4s }, [x22], #0x10\n"
+      "ldr q7, [x12, #0x10]\n"
+      "bge 200b\n"
+      "201:"  // Height 6: Multiply loop: Single iteration only
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "sub x27, x27, #0x4\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
+      ".inst 0x4ea168a4  // bfcvtn2 v4.8h, v5.4s\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
+      "ldr q3, [x11, #0x0]\n"
+      ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
+      "ldr q1, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec51  // bfmmla v17.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec99  // bfmmla v25.4s, v4.8h, v3.8h\n"
+      "ldr q3, [x10, #0x0]\n"
+      ".inst 0x6e41ec0d  // bfmmla v13.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec55  // bfmmla v21.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9d  // bfmmla v29.4s, v4.8h, v1.8h\n"
+      "ldr q1, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e43ec0a  // bfmmla v10.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec52  // bfmmla v18.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec9a  // bfmmla v26.4s, v4.8h, v3.8h\n"
+      "ldr q3, [x9, #0x0]\n"
+      ".inst 0x6e41ec0e  // bfmmla v14.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec56  // bfmmla v22.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9e  // bfmmla v30.4s, v4.8h, v1.8h\n"
+      "ldr q1, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e43ec0b  // bfmmla v11.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec53  // bfmmla v19.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec9b  // bfmmla v27.4s, v4.8h, v3.8h\n"
+      ".inst 0x6e41ec0f  // bfmmla v15.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec57  // bfmmla v23.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9f  // bfmmla v31.4s, v4.8h, v1.8h\n"
+      "202:"  // Height 6: Multiply loop: Main loop skip
+      "cbz x27, 205f\n"
+      "cbz x27, 205f\n"
+      "tbz x27, #1, 203f\n"
+      "ldr d0, [x26], #0x8\n"
+      "ldr d1, [x25], #0x8\n"
+      "ldr d2, [x24], #0x8\n"
+      "ldr d3, [x23], #0x8\n"
+      "ldr d4, [x22], #0x8\n"
+      "ldr d5, [x21], #0x8\n"
+      "tbz x27, #0, 204f\n"
+      "ld1 { v0.s }[2], [x26]\n"
+      "ld1 { v1.s }[2], [x25]\n"
+      "ld1 { v2.s }[2], [x24]\n"
+      "ld1 { v3.s }[2], [x23]\n"
+      "ld1 { v4.s }[2], [x22]\n"
+      "ld1 { v5.s }[2], [x21]\n"
+      "b 204f\n"
+      "203:"  // Height 6: Multiply loop: Ragged operand read: partial_1_0
+      "ldr s0, [x26, #0x0]\n"
+      "ldr s1, [x25, #0x0]\n"
+      "ldr s2, [x24, #0x0]\n"
+      "ldr s3, [x23, #0x0]\n"
+      "ldr s4, [x22, #0x0]\n"
+      "ldr s5, [x21, #0x0]\n"
+      "204:"  // Height 6: Multiply loop: Ragged operand read: Done
+      "ldr q7, [x12, #0x0]\n"
+      "ldr q6, [x12, #0x10]\n"
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
+      ".inst 0x4ea168a4  // bfcvtn2 v4.8h, v5.4s\n"
+      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e46ec0c  // bfmmla v12.4s, v0.8h, v6.8h\n"
+      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
+      "ldr q3, [x11, #0x0]\n"
+      ".inst 0x6e46ec54  // bfmmla v20.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec9c  // bfmmla v28.4s, v4.8h, v6.8h\n"
+      "ldr q1, [x11, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec51  // bfmmla v17.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec99  // bfmmla v25.4s, v4.8h, v3.8h\n"
+      "ldr q3, [x10, #0x0]\n"
+      ".inst 0x6e41ec0d  // bfmmla v13.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec55  // bfmmla v21.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9d  // bfmmla v29.4s, v4.8h, v1.8h\n"
+      "ldr q1, [x10, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e43ec0a  // bfmmla v10.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec52  // bfmmla v18.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec9a  // bfmmla v26.4s, v4.8h, v3.8h\n"
+      "ldr q3, [x9, #0x0]\n"
+      ".inst 0x6e41ec0e  // bfmmla v14.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec56  // bfmmla v22.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9e  // bfmmla v30.4s, v4.8h, v1.8h\n"
+      "ldr q1, [x9, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      ".inst 0x6e43ec0b  // bfmmla v11.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e43ec53  // bfmmla v19.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec9b  // bfmmla v27.4s, v4.8h, v3.8h\n"
+      ".inst 0x6e41ec0f  // bfmmla v15.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec57  // bfmmla v23.4s, v2.8h, v1.8h\n"
+      ".inst 0x6e41ec9f  // bfmmla v31.4s, v4.8h, v1.8h\n"
+      "205:"  // Height 6: Multiply loop: No odd multiplies
+      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "add x28, x28, #0x1\n"
+      "cmp x28, x20\n"
+      "bne 197b\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "uzp1 v6.2d, v8.2d, v12.2d\n"
+      "uzp2 v8.2d, v8.2d, v12.2d\n"
+      "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "uzp2 v9.2d, v9.2d, v13.2d\n"
+      "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "uzp2 v10.2d, v10.2d, v14.2d\n"
+      "add x26, x13, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
+      "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "uzp2 v11.2d, v11.2d, v15.2d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "uzp2 v16.2d, v16.2d, v20.2d\n"
+      "add x22, x23, x20, LSL #2\n"
+      "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "uzp2 v17.2d, v17.2d, v21.2d\n"
+      "uzp1 v21.2d, v18.2d, v22.2d\n"
+      "uzp2 v18.2d, v18.2d, v22.2d\n"
+      "uzp1 v22.2d, v19.2d, v23.2d\n"
+      "uzp2 v19.2d, v19.2d, v23.2d\n"
+      "uzp1 v23.2d, v24.2d, v28.2d\n"
+      "uzp2 v24.2d, v24.2d, v28.2d\n"
+      "uzp1 v28.2d, v25.2d, v29.2d\n"
+      "uzp2 v25.2d, v25.2d, v29.2d\n"
+      "uzp1 v29.2d, v26.2d, v30.2d\n"
+      "uzp2 v26.2d, v26.2d, v30.2d\n"
+      "uzp1 v30.2d, v27.2d, v31.2d\n"
+      "uzp2 v27.2d, v27.2d, v31.2d\n"
+      "tbz %x[flags], #1, 206f\n"
+      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v1.4s }, [x21]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "fmin v6.4s, v6.4s, v1.4s\n"
+      "fmin v12.4s, v12.4s, v1.4s\n"
+      "fmin v13.4s, v13.4s, v1.4s\n"
+      "fmin v14.4s, v14.4s, v1.4s\n"
+      "fmin v8.4s, v8.4s, v1.4s\n"
+      "fmin v9.4s, v9.4s, v1.4s\n"
+      "fmin v10.4s, v10.4s, v1.4s\n"
+      "fmin v11.4s, v11.4s, v1.4s\n"
+      "fmin v15.4s, v15.4s, v1.4s\n"
+      "fmin v20.4s, v20.4s, v1.4s\n"
+      "fmin v21.4s, v21.4s, v1.4s\n"
+      "fmin v22.4s, v22.4s, v1.4s\n"
+      "fmin v16.4s, v16.4s, v1.4s\n"
+      "fmin v17.4s, v17.4s, v1.4s\n"
+      "fmin v18.4s, v18.4s, v1.4s\n"
+      "fmin v19.4s, v19.4s, v1.4s\n"
+      "fmin v23.4s, v23.4s, v1.4s\n"
+      "fmin v28.4s, v28.4s, v1.4s\n"
+      "fmin v29.4s, v29.4s, v1.4s\n"
+      "fmin v30.4s, v30.4s, v1.4s\n"
+      "fmin v24.4s, v24.4s, v1.4s\n"
+      "fmin v25.4s, v25.4s, v1.4s\n"
+      "fmin v26.4s, v26.4s, v1.4s\n"
+      "fmin v27.4s, v27.4s, v1.4s\n"
+      "fmax v6.4s, v6.4s, v0.4s\n"
+      "fmax v12.4s, v12.4s, v0.4s\n"
+      "fmax v13.4s, v13.4s, v0.4s\n"
+      "fmax v14.4s, v14.4s, v0.4s\n"
+      "fmax v8.4s, v8.4s, v0.4s\n"
+      "fmax v9.4s, v9.4s, v0.4s\n"
+      "fmax v10.4s, v10.4s, v0.4s\n"
+      "fmax v11.4s, v11.4s, v0.4s\n"
+      "fmax v15.4s, v15.4s, v0.4s\n"
+      "fmax v20.4s, v20.4s, v0.4s\n"
+      "fmax v21.4s, v21.4s, v0.4s\n"
+      "fmax v22.4s, v22.4s, v0.4s\n"
+      "fmax v16.4s, v16.4s, v0.4s\n"
+      "fmax v17.4s, v17.4s, v0.4s\n"
+      "fmax v18.4s, v18.4s, v0.4s\n"
+      "fmax v19.4s, v19.4s, v0.4s\n"
+      "fmax v23.4s, v23.4s, v0.4s\n"
+      "fmax v28.4s, v28.4s, v0.4s\n"
+      "fmax v29.4s, v29.4s, v0.4s\n"
+      "fmax v30.4s, v30.4s, v0.4s\n"
+      "fmax v24.4s, v24.4s, v0.4s\n"
+      "fmax v25.4s, v25.4s, v0.4s\n"
+      "fmax v26.4s, v26.4s, v0.4s\n"
+      "fmax v27.4s, v27.4s, v0.4s\n"
+      "206:"  // Height 6: No activation
+      "cmp x14, #0x10\n"
+      "bge 215f\n"
+      "tbz x14, #3, 210f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v12.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v9.4s }, [x26], #0x10\n"
+      "st1 { v15.4s }, [x25], #0x10\n"
+      "st1 { v20.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v23.4s }, [x23], #0x10\n"
+      "st1 { v28.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
+      "tbz x14, #2, 208f\n"
+      "st1 { v13.4s }, [x13], #0x10\n"
+      "st1 { v10.4s }, [x26], #0x10\n"
+      "st1 { v21.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v29.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
+      "tbz x14, #1, 207f\n"
+      "str d14, [x13], #0x8\n"
+      "str d11, [x26], #0x8\n"
+      "str d22, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d30, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "tbz x14, #0, 214f\n"
+      "st1 { v14.s }[2], [x13]\n"
+      "st1 { v11.s }[2], [x26]\n"
+      "st1 { v22.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v30.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
+      "b 214f\n"
+      "207:"  // Height 6: Partial direct writeback: partial_1_12
+      "tbz x14, #0, 214f\n"
+      "str s14, [x13, #0x0]\n"
+      "str s11, [x26, #0x0]\n"
+      "str s22, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s30, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
+      "b 214f\n"
+      "208:"  // Height 6: Partial direct writeback: partial_2_8
+      "tbz x14, #1, 209f\n"
+      "str d13, [x13], #0x8\n"
+      "str d10, [x26], #0x8\n"
+      "str d21, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d29, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "tbz x14, #0, 214f\n"
+      "st1 { v13.s }[2], [x13]\n"
+      "st1 { v10.s }[2], [x26]\n"
+      "st1 { v21.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v29.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
+      "b 214f\n"
+      "209:"  // Height 6: Partial direct writeback: partial_1_8
+      "tbz x14, #0, 214f\n"
+      "str s13, [x13, #0x0]\n"
+      "str s10, [x26, #0x0]\n"
+      "str s21, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s29, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
+      "b 214f\n"
+      "210:"  // Height 6: Partial direct writeback: partial_4_0
+      "tbz x14, #2, 212f\n"
+      "st1 { v6.4s }, [x13], #0x10\n"
+      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v15.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v23.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "tbz x14, #1, 211f\n"
+      "str d12, [x13], #0x8\n"
+      "str d9, [x26], #0x8\n"
+      "str d20, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d28, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "tbz x14, #0, 214f\n"
+      "st1 { v12.s }[2], [x13]\n"
+      "st1 { v9.s }[2], [x26]\n"
+      "st1 { v20.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v28.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
+      "b 214f\n"
+      "211:"  // Height 6: Partial direct writeback: partial_1_4
+      "tbz x14, #0, 214f\n"
+      "str s12, [x13, #0x0]\n"
+      "str s9, [x26, #0x0]\n"
+      "str s20, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s28, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
+      "b 214f\n"
+      "212:"  // Height 6: Partial direct writeback: partial_2_0
+      "tbz x14, #1, 213f\n"
+      "str d6, [x13], #0x8\n"
+      "str d8, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "tbz x14, #0, 214f\n"
+      "st1 { v6.s }[2], [x13]\n"
+      "st1 { v8.s }[2], [x26]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
+      "b 214f\n"
+      "213:"  // Height 6: Partial direct writeback: partial_1_0
+      "str s6, [x13, #0x0]\n"
+      "str s8, [x26, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
+      "214:"  // Height 6: Partial direct writeback: Done
+      "b 216f\n"
+      "215:"  // Height 6: Full writeback
+      "str q6, [x13, #0x0]\n"
+      "str q12, [x13, #0x10]\n"
+      "str q13, [x13, #0x20]\n"
+      "str q14, [x13, #0x30]\n"
+      "add x13, x13, #0x40\n"
+      "str q8, [x26, #0x0]\n"
+      "str q9, [x26, #0x10]\n"
+      "str q10, [x26, #0x20]\n"
+      "str q11, [x26, #0x30]\n"
+      "str q15, [x25, #0x0]\n"
+      "str q20, [x25, #0x10]\n"
+      "str q21, [x25, #0x20]\n"
+      "str q22, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q23, [x23, #0x0]\n"
+      "str q28, [x23, #0x10]\n"
+      "str q29, [x23, #0x20]\n"
+      "str q30, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
+      "216:"  // Height 6: Writeback done
+      "subs x14, x14, #0x10\n"
+      "bgt 182b\n"
+      "subs %x[M], %x[M], #0x6\n"
+      "beq 218f\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "tbz %x[flags], #3, 217f\n"
+      "add x21, x21, #0x6\n"
+      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "b 1b\n"
+      "217:"  // Update direct input
+      "mov x20, #0x18\n"
+      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
+      "b 1b\n"
+      "218:"  // Exit
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [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_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [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", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    );
+}
+
+} // namespace arm_gemm
+#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp
index cf4d742..1a8b0fd 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022-2023 Arm Limited.
+ * Copyright (c) 2022-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -88,8 +88,10 @@
 
         if (std::is_same<T, float>::value) {
             switch (ci->get_cpu_model()) {
+                case CPUModel::V1:
+                    return { 45.25, 4.29, 4.80 };
                 default:
-                    return { 38.10, 5.23, 3.15 };
+                    return { 29.85, 2.60, 5.49 };
             }
         }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp
index 887d78e..23f686a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022-2023 Arm Limited.
+ * Copyright (c) 2022-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -88,8 +88,10 @@
     {
         if (std::is_same<T, float>::value) {
             switch (ci->get_cpu_model()) {
+                case CPUModel::V1:
+                    return { 28.74 };
                 default:
-                    return { 32.35 };
+                    return { 15.27 };
             }
         }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp
index d0ef531..1fe5f48 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022-2023 Arm Limited.
+ * Copyright (c) 2022-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -88,8 +88,10 @@
 
         if (std::is_same<T, float>::value) {
             switch (ci->get_cpu_model()) {
-                default:
-                    return { 39.66, 5.18, 4.37 };
+                case CPUModel::V1:
+                    return { 53.48, 4.23, 6.53 };
+                default:                
+                    return { 29.07, 2.76, 5.39 };
             }
         }