[ONCPUML-1451] Add matmul kernel to enable bf16 to bf16 operations via PyTorch® autocast() function
The full range of tests must be added with [MLINFSW-482] epic due to the lack of reordering kernels implemented in Acl.
Co-Authored-By: David Mansell <David.Mansell@arm.com>
Change-Id: I820d316295a1ec94fdc89c37e4144a268f914c36
Signed-off-by: Renato Arantes <renato.arantes@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11169
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/BUILD.bazel b/src/BUILD.bazel
index f9d166c..d4a3b61 100644
--- a/src/BUILD.bazel
+++ b/src/BUILD.bazel
@@ -1,4 +1,4 @@
-# Copyright (c) 2023 Arm Limited.
+# Copyright (c) 2023-2024 Arm Limited.
#
# SPDX-License-Identifier: MIT
#
@@ -509,6 +509,7 @@
"core/NEON/kernels/arm_conv/pooling/pooling_u8.cpp",
"core/NEON/kernels/arm_conv/pooling/pooling_u8q.cpp",
"core/NEON/kernels/arm_gemm/gemm_bf16.cpp",
+ "core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp",
"core/NEON/kernels/arm_gemm/gemm_fp16.cpp",
"core/NEON/kernels/arm_gemm/gemm_fp32.cpp",
"core/NEON/kernels/arm_gemm/gemm_int16.cpp",
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index c5a1721..c641071 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -1,4 +1,4 @@
-# Copyright (c) 2023 Arm Limited.
+# Copyright (c) 2023-2024 Arm Limited.
#
# SPDX-License-Identifier: MIT
#
@@ -500,6 +500,7 @@
core/NEON/kernels/arm_conv/pooling/pooling_u8.cpp
core/NEON/kernels/arm_conv/pooling/pooling_u8q.cpp
core/NEON/kernels/arm_gemm/gemm_bf16.cpp
+ core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp
core/NEON/kernels/arm_gemm/gemm_fp16.cpp
core/NEON/kernels/arm_gemm/gemm_fp32.cpp
core/NEON/kernels/arm_gemm/gemm_int16.cpp
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp
new file mode 100644
index 0000000..aa761b4
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2017-2020, 2022-2024 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.
+ */
+#include "bfloat.hpp"
+#include "gemm_implementation.hpp"
+#include "gemm_interleaved.hpp"
+
+#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+#include "kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp"
+#include "kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp"
+#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+
+namespace arm_gemm {
+
+static const GemmImplementation<bfloat16, bfloat16> gemm_bf16bf16_methods[] =
+{
+#ifdef __aarch64__
+#ifdef ARM_COMPUTE_ENABLE_BF16
+#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+GemmImplementation<bfloat16, bfloat16>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "a64_ffinterleaved_bf16fp32_mmla_8x12",
+ KernelWeightFormat::VL256_BL64,
+ [](const GemmArgs &args) { return args._ci->has_bf16(); },
+ [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, bfloat16>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, bfloat16>(args); }
+),
+GemmImplementation<bfloat16, bfloat16>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "sve_ffinterleaved_bf16fp32_mmla_8x3VL",
+ KernelWeightFormat::VL2VL_BL64,
+ [](const GemmArgs &args) { return args._ci->has_svebf16(); },
+ [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, bfloat16, bfloat16>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, bfloat16, bfloat16>(args); }
+),
+#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+#endif // ARM_COMPUTE_ENABLE_BF16
+#endif // __aarch64__
+{
+ GemmMethod::DEFAULT,
+ "",
+ nullptr,
+ nullptr,
+ nullptr
+}
+};
+
+template<>
+const GemmImplementation<bfloat16, bfloat16> *gemm_implementation_list<bfloat16, bfloat16>() {
+ return gemm_bf16bf16_methods;
+}
+
+/* Explicitly instantiate the external functions for these types. */
+template UniqueGemmCommon<bfloat16, bfloat16> gemm<bfloat16, bfloat16, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<bfloat16, bfloat16, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
+template KernelDescription get_gemm_method<bfloat16, bfloat16, Nothing>(const GemmArgs &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<bfloat16, bfloat16, Nothing>(const GemmArgs &args, const Nothing &);
+
+} // namespace arm_gemm
diff --git a/src/core/NEON/kernels/arm_gemm/mergeresults-sve.cpp b/src/core/NEON/kernels/arm_gemm/mergeresults-sve.cpp
index a4124c4..d366553 100644
--- a/src/core/NEON/kernels/arm_gemm/mergeresults-sve.cpp
+++ b/src/core/NEON/kernels/arm_gemm/mergeresults-sve.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2021, 2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -30,6 +30,7 @@
#include "arm_gemm.hpp"
#include "asmlib.hpp"
+#include "bfloat.hpp"
#include "utils.hpp"
#include "mergeresults.hpp"
diff --git a/src/core/NEON/kernels/arm_gemm/mergeresults.cpp b/src/core/NEON/kernels/arm_gemm/mergeresults.cpp
index 2b712ce..e100d9f 100644
--- a/src/core/NEON/kernels/arm_gemm/mergeresults.cpp
+++ b/src/core/NEON/kernels/arm_gemm/mergeresults.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2021, 2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -30,6 +30,7 @@
#include "arm_gemm.hpp"
#include "asmlib.hpp"
+#include "bfloat.hpp"
#include "utils.hpp"
namespace arm_gemm {
@@ -114,4 +115,8 @@
template void MergeResults<8u, 6u, false, float, __fp16>(__fp16*, float const*, int, int, int, int, int, __fp16 const*, Activation, bool);
#endif
+#if defined(__arm__) && defined(ARM_COMPUTE_ENABLE_BF16)
+template void MergeResults<8u, 6u, false, float, bfloat16>(bfloat16*, float const*, int, int, int, int, int, bfloat16 const*, Activation, bool);
+#endif
+
} // namespace arm_gemm
diff --git a/src/core/NEON/kernels/arm_gemm/merges/a64_merge_fp32_bf16_8x12.hpp b/src/core/NEON/kernels/arm_gemm/merges/a64_merge_fp32_bf16_8x12.hpp
new file mode 100644
index 0000000..a57a855
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/merges/a64_merge_fp32_bf16_8x12.hpp
@@ -0,0 +1,2809 @@
+/*
+ * 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__
+
+template<>
+void MergeResults<12, 8, false>(
+ bfloat16 *out_ptr,
+ const float * in_ptr,
+ const int ldout,
+ const int y0, const int ymax,
+ const int x0, const int xmax,
+ const bfloat16 *bias,
+ Activation act,
+ bool accumulate)
+{
+ float maxval = static_cast<float>(std::numeric_limits<float>::infinity());
+ float minval = - static_cast<float>(std::numeric_limits<float>::infinity());
+
+ switch(act.type) {
+ default:
+ case Activation::Type::None:
+ break;
+ case Activation::Type::BoundedReLU:
+ maxval = static_cast<float>(act.param1);
+ /* fall through */
+ case Activation::Type::ReLU:
+ minval = 0;
+ break;
+ }
+
+ size_t rows = ymax-y0;
+ size_t cols = xmax-x0;
+
+ out_ptr += (y0 * ldout) + x0;
+ bias = (bias == nullptr) ? nullptr : bias + x0;
+
+ __asm__ __volatile__(
+ "cbz %x[cols], 108f\n"
+ "cbz %x[rows], 108f\n"
+ "mov x11, #0x20\n"
+ "dup v13.4s, %w[maxval]\n"
+ "dup v12.4s, %w[minval]\n"
+ "mul x11, %x[ldout], x11\n"
+ "cbnz %x[accumulate], 66f\n"
+ "1:" // Initial: Row loop
+ "cmp %x[rows], #0x7\n"
+ "bgt 58f\n"
+ "beq 50f\n"
+ "cmp %x[rows], #0x5\n"
+ "bgt 42f\n"
+ "beq 34f\n"
+ "cmp %x[rows], #0x3\n"
+ "bgt 26f\n"
+ "beq 18f\n"
+ "cmp %x[rows], #0x1\n"
+ "bgt 10f\n"
+ "2:" // Initial: Height 1
+ "mov x10, %x[cols]\n"
+ "mov x9, %x[out_ptr]\n"
+ "mov x28, %x[bias]\n"
+ "cmp x10, #0xc\n"
+ "blt 6f\n"
+ "3:" // Initial: Height 1: Block loop
+ "cbnz %x[bias], 4f\n"
+ "movi v21.16b, #0x0\n"
+ "movi v20.16b, #0x0\n"
+ "movi v19.16b, #0x0\n"
+ "b 5f\n"
+ "4:" // Initial: Height 1: Width 3: bias
+ "ldr d18, [x28, #0x0]\n"
+ "ldr d17, [x28, #0x8]\n"
+ "ldr d16, [x28, #0x10]\n"
+ "shll v21.4s, v18.4h, #0x10\n"
+ "shll v20.4s, v17.4h, #0x10\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "5:" // Initial: Height 1: Width 3: init done
+ "ldr q18, [%x[in_ptr], #0x0]\n"
+ "ldr q17, [%x[in_ptr], #0x10]\n"
+ "sub x10, x10, #0xc\n"
+ "add x28, x28, #0x18\n"
+ "ldr q16, [%x[in_ptr], #0x20]\n"
+ "cmp x10, #0xc\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v18.4s, v18.4s, v21.4s\n"
+ "fadd v17.4s, v17.4s, v20.4s\n"
+ "fadd v16.4s, v16.4s, v19.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ "str d18, [x9, #0x0]\n"
+ "str d17, [x9, #0x8]\n"
+ "str d16, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ "bge 3b\n"
+ "6:" // Initial: Height 1: no full blocks
+ "cbz x10, 9f\n"
+ "mov x20, %x[in_ptr]\n"
+ "7:" // Initial: Height 1: Single loop
+ "movi v17.16b, #0x0\n"
+ "cbz %x[bias], 8f\n"
+ "ldr h16, [x28, #0x0]\n"
+ "shll v17.4s, v16.4h, #0x10\n"
+ "8:" // Initial: Height 1: Scalar: no bias
+ "ldr s16, [%x[in_ptr], #0x0]\n"
+ "subs x10, x10, #0x1\n"
+ "add x28, x28, #0x2\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v16.4s, v16.4s, v17.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ "str h16, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ "bne 7b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "9:" // Initial: Height 1: no oddments
+ "b 108f\n"
+ "10:" // Initial: Height 2
+ "mov x10, %x[cols]\n"
+ "mov x9, %x[out_ptr]\n"
+ "mov x28, %x[bias]\n"
+ "cmp x10, #0xc\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "blt 14f\n"
+ "11:" // Initial: Height 2: Block loop
+ "cbnz %x[bias], 12f\n"
+ "movi v24.16b, #0x0\n"
+ "movi v23.16b, #0x0\n"
+ "movi v22.16b, #0x0\n"
+ "b 13f\n"
+ "12:" // Initial: Height 2: Width 3: bias
+ "ldr d18, [x28, #0x0]\n"
+ "ldr d17, [x28, #0x8]\n"
+ "ldr d16, [x28, #0x10]\n"
+ "shll v24.4s, v18.4h, #0x10\n"
+ "shll v23.4s, v17.4h, #0x10\n"
+ "shll v22.4s, v16.4h, #0x10\n"
+ "13:" // Initial: Height 2: Width 3: init done
+ "ldr q16, [%x[in_ptr], #0x0]\n"
+ "ldr q20, [%x[in_ptr], #0x10]\n"
+ "sub x10, x10, #0xc\n"
+ "add x28, x28, #0x18\n"
+ "ldr q19, [%x[in_ptr], #0x20]\n"
+ "ldr q18, [%x[in_ptr], #0x30]\n"
+ "cmp x10, #0xc\n"
+ "ldr q17, [%x[in_ptr], #0x40]\n"
+ "ldr q21, [%x[in_ptr], #0x50]\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v16.4s, v16.4s, v24.4s\n"
+ "fadd v20.4s, v20.4s, v23.4s\n"
+ "fadd v19.4s, v19.4s, v22.4s\n"
+ "fadd v18.4s, v18.4s, v24.4s\n"
+ "fadd v17.4s, v17.4s, v23.4s\n"
+ "fadd v21.4s, v21.4s, v22.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ "str d16, [x9, #0x0]\n"
+ ".inst 0x0ea16ab0 // bfcvtn v16.4h, v21.4s\n"
+ "str d20, [x9, #0x8]\n"
+ "str d19, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ "str d18, [x27, #0x0]\n"
+ "str d17, [x27, #0x8]\n"
+ "str d16, [x27, #0x10]\n"
+ "add x27, x27, #0x18\n"
+ "bge 11b\n"
+ "14:" // Initial: Height 2: no full blocks
+ "cbz x10, 17f\n"
+ "mov x20, %x[in_ptr]\n"
+ "15:" // Initial: Height 2: Single loop
+ "movi v18.16b, #0x0\n"
+ "cbz %x[bias], 16f\n"
+ "ldr h16, [x28, #0x0]\n"
+ "shll v18.4s, v16.4h, #0x10\n"
+ "16:" // Initial: Height 2: Scalar: no bias
+ "ldr s17, [%x[in_ptr], #0x0]\n"
+ "ldr s16, [%x[in_ptr], #0x30]\n"
+ "subs x10, x10, #0x1\n"
+ "add x28, x28, #0x2\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v17.4s, v17.4s, v18.4s\n"
+ "fadd v16.4s, v16.4s, v18.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ "str h17, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ "str h16, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "bne 15b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "17:" // Initial: Height 2: no oddments
+ "b 108f\n"
+ "18:" // Initial: Height 3
+ "mov x10, %x[cols]\n"
+ "mov x9, %x[out_ptr]\n"
+ "mov x28, %x[bias]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "blt 22f\n"
+ "19:" // Initial: Height 3: Block loop
+ "cbnz %x[bias], 20f\n"
+ "movi v27.16b, #0x0\n"
+ "movi v26.16b, #0x0\n"
+ "movi v25.16b, #0x0\n"
+ "b 21f\n"
+ "20:" // Initial: Height 3: Width 3: bias
+ "ldr d18, [x28, #0x0]\n"
+ "ldr d17, [x28, #0x8]\n"
+ "ldr d16, [x28, #0x10]\n"
+ "shll v27.4s, v18.4h, #0x10\n"
+ "shll v26.4s, v17.4h, #0x10\n"
+ "shll v25.4s, v16.4h, #0x10\n"
+ "21:" // Initial: Height 3: Width 3: init done
+ "ldr q18, [%x[in_ptr], #0x0]\n"
+ "ldr q17, [%x[in_ptr], #0x10]\n"
+ "sub x10, x10, #0xc\n"
+ "add x28, x28, #0x18\n"
+ "ldr q16, [%x[in_ptr], #0x20]\n"
+ "ldr q21, [%x[in_ptr], #0x30]\n"
+ "cmp x10, #0xc\n"
+ "ldr q20, [%x[in_ptr], #0x40]\n"
+ "ldr q19, [%x[in_ptr], #0x50]\n"
+ "ldr q24, [%x[in_ptr], #0x60]\n"
+ "ldr q23, [%x[in_ptr], #0x70]\n"
+ "fadd v18.4s, v18.4s, v27.4s\n"
+ "fadd v17.4s, v17.4s, v26.4s\n"
+ "ldr q22, [%x[in_ptr], #0x80]\n"
+ "fadd v16.4s, v16.4s, v25.4s\n"
+ "fadd v21.4s, v21.4s, v27.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v20.4s, v20.4s, v26.4s\n"
+ "fadd v19.4s, v19.4s, v25.4s\n"
+ "fadd v24.4s, v24.4s, v27.4s\n"
+ "fadd v23.4s, v23.4s, v26.4s\n"
+ "fadd v22.4s, v22.4s, v25.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ "str d18, [x9, #0x0]\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16b12 // bfcvtn v18.4h, v24.4s\n"
+ "str d17, [x9, #0x8]\n"
+ "str d16, [x9, #0x10]\n"
+ ".inst 0x0ea16af1 // bfcvtn v17.4h, v23.4s\n"
+ ".inst 0x0ea16ad0 // bfcvtn v16.4h, v22.4s\n"
+ "add x9, x9, #0x18\n"
+ "str d21, [x27, #0x0]\n"
+ "str d20, [x27, #0x8]\n"
+ "str d19, [x27, #0x10]\n"
+ "add x27, x27, #0x18\n"
+ "str d18, [x26, #0x0]\n"
+ "str d17, [x26, #0x8]\n"
+ "str d16, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "bge 19b\n"
+ "22:" // Initial: Height 3: no full blocks
+ "cbz x10, 25f\n"
+ "mov x20, %x[in_ptr]\n"
+ "23:" // Initial: Height 3: Single loop
+ "movi v19.16b, #0x0\n"
+ "cbz %x[bias], 24f\n"
+ "ldr h16, [x28, #0x0]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "24:" // Initial: Height 3: Scalar: no bias
+ "ldr s16, [%x[in_ptr], #0x0]\n"
+ "ldr s17, [%x[in_ptr], #0x30]\n"
+ "subs x10, x10, #0x1\n"
+ "add x28, x28, #0x2\n"
+ "ldr s18, [%x[in_ptr], #0x60]\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v16.4s, v16.4s, v19.4s\n"
+ "fadd v17.4s, v17.4s, v19.4s\n"
+ "fadd v18.4s, v18.4s, v19.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ "str h16, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ ".inst 0x0ea16a50 // bfcvtn v16.4h, v18.4s\n"
+ "str h17, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "str h16, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "bne 23b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "25:" // Initial: Height 3: no oddments
+ "b 108f\n"
+ "26:" // Initial: Height 4
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "mov x28, %x[bias]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "blt 30f\n"
+ "27:" // Initial: Height 4: Block loop
+ "cbnz %x[bias], 28f\n"
+ "movi v30.16b, #0x0\n"
+ "movi v29.16b, #0x0\n"
+ "movi v28.16b, #0x0\n"
+ "b 29f\n"
+ "28:" // Initial: Height 4: Width 3: bias
+ "ldr d18, [x28, #0x0]\n"
+ "ldr d17, [x28, #0x8]\n"
+ "ldr d16, [x28, #0x10]\n"
+ "shll v30.4s, v18.4h, #0x10\n"
+ "shll v29.4s, v17.4h, #0x10\n"
+ "shll v28.4s, v16.4h, #0x10\n"
+ "29:" // Initial: Height 4: Width 3: init done
+ "ldr q19, [%x[in_ptr], #0x0]\n"
+ "ldr q18, [%x[in_ptr], #0x10]\n"
+ "sub x10, x10, #0xc\n"
+ "add x28, x28, #0x18\n"
+ "ldr q17, [%x[in_ptr], #0x20]\n"
+ "ldr q16, [%x[in_ptr], #0x30]\n"
+ "cmp x10, #0xc\n"
+ "ldr q23, [%x[in_ptr], #0x40]\n"
+ "ldr q22, [%x[in_ptr], #0x50]\n"
+ "ldr q21, [%x[in_ptr], #0x60]\n"
+ "ldr q20, [%x[in_ptr], #0x70]\n"
+ "fadd v19.4s, v19.4s, v30.4s\n"
+ "fadd v18.4s, v18.4s, v29.4s\n"
+ "ldr q27, [%x[in_ptr], #0x80]\n"
+ "ldr q26, [%x[in_ptr], #0x90]\n"
+ "fadd v17.4s, v17.4s, v28.4s\n"
+ "fadd v16.4s, v16.4s, v30.4s\n"
+ "ldr q25, [%x[in_ptr], #0xa0]\n"
+ "ldr q24, [%x[in_ptr], #0xb0]\n"
+ "fadd v23.4s, v23.4s, v29.4s\n"
+ "fadd v22.4s, v22.4s, v28.4s\n"
+ "fadd v21.4s, v21.4s, v30.4s\n"
+ "fadd v20.4s, v20.4s, v29.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v27.4s, v27.4s, v28.4s\n"
+ "fadd v26.4s, v26.4s, v30.4s\n"
+ "fadd v25.4s, v25.4s, v29.4s\n"
+ "fadd v24.4s, v24.4s, v28.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16af7 // bfcvtn v23.4h, v23.4s\n"
+ ".inst 0x0ea16ad6 // bfcvtn v22.4h, v22.4s\n"
+ "str d19, [x9, #0x0]\n"
+ "str d18, [x9, #0x8]\n"
+ ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ "str d17, [x9, #0x10]\n"
+ ".inst 0x0ea16b73 // bfcvtn v19.4h, v27.4s\n"
+ ".inst 0x0ea16b52 // bfcvtn v18.4h, v26.4s\n"
+ "add x9, x9, #0x18\n"
+ "str d16, [x27, #0x0]\n"
+ ".inst 0x0ea16b31 // bfcvtn v17.4h, v25.4s\n"
+ ".inst 0x0ea16b10 // bfcvtn v16.4h, v24.4s\n"
+ "str d23, [x27, #0x8]\n"
+ "str d22, [x27, #0x10]\n"
+ "add x27, x27, #0x18\n"
+ "str d21, [x26, #0x0]\n"
+ "str d20, [x26, #0x8]\n"
+ "str d19, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "str d18, [x25, #0x0]\n"
+ "str d17, [x25, #0x8]\n"
+ "str d16, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "bge 27b\n"
+ "30:" // Initial: Height 4: no full blocks
+ "cbz x10, 33f\n"
+ "mov x20, %x[in_ptr]\n"
+ "31:" // Initial: Height 4: Single loop
+ "movi v20.16b, #0x0\n"
+ "cbz %x[bias], 32f\n"
+ "ldr h16, [x28, #0x0]\n"
+ "shll v20.4s, v16.4h, #0x10\n"
+ "32:" // Initial: Height 4: Scalar: no bias
+ "ldr s16, [%x[in_ptr], #0x0]\n"
+ "ldr s18, [%x[in_ptr], #0x30]\n"
+ "subs x10, x10, #0x1\n"
+ "add x28, x28, #0x2\n"
+ "ldr s17, [%x[in_ptr], #0x60]\n"
+ "ldr s19, [%x[in_ptr], #0x90]\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v16.4s, v16.4s, v20.4s\n"
+ "fadd v18.4s, v18.4s, v20.4s\n"
+ "fadd v17.4s, v17.4s, v20.4s\n"
+ "fadd v19.4s, v19.4s, v20.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ "str h16, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ ".inst 0x0ea16a70 // bfcvtn v16.4h, v19.4s\n"
+ "str h18, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "str h17, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h16, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "bne 31b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "33:" // Initial: Height 4: no oddments
+ "b 108f\n"
+ "34:" // Initial: Height 5
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "mov x28, %x[bias]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "blt 38f\n"
+ "35:" // Initial: Height 5: Block loop
+ "cbnz %x[bias], 36f\n"
+ "movi v1.16b, #0x0\n"
+ "movi v0.16b, #0x0\n"
+ "movi v31.16b, #0x0\n"
+ "b 37f\n"
+ "36:" // Initial: Height 5: Width 3: bias
+ "ldr d18, [x28, #0x0]\n"
+ "ldr d17, [x28, #0x8]\n"
+ "ldr d16, [x28, #0x10]\n"
+ "shll v1.4s, v18.4h, #0x10\n"
+ "shll v0.4s, v17.4h, #0x10\n"
+ "shll v31.4s, v16.4h, #0x10\n"
+ "37:" // Initial: Height 5: Width 3: init done
+ "ldr q16, [%x[in_ptr], #0x0]\n"
+ "ldr q20, [%x[in_ptr], #0x10]\n"
+ "sub x10, x10, #0xc\n"
+ "add x28, x28, #0x18\n"
+ "ldr q19, [%x[in_ptr], #0x20]\n"
+ "ldr q18, [%x[in_ptr], #0x30]\n"
+ "cmp x10, #0xc\n"
+ "ldr q17, [%x[in_ptr], #0x40]\n"
+ "ldr q30, [%x[in_ptr], #0x50]\n"
+ "ldr q24, [%x[in_ptr], #0x60]\n"
+ "ldr q23, [%x[in_ptr], #0x70]\n"
+ "fadd v16.4s, v16.4s, v1.4s\n"
+ "fadd v20.4s, v20.4s, v0.4s\n"
+ "ldr q22, [%x[in_ptr], #0x80]\n"
+ "ldr q21, [%x[in_ptr], #0x90]\n"
+ "fadd v19.4s, v19.4s, v31.4s\n"
+ "fadd v18.4s, v18.4s, v1.4s\n"
+ "ldr q29, [%x[in_ptr], #0xa0]\n"
+ "ldr q28, [%x[in_ptr], #0xb0]\n"
+ "fadd v17.4s, v17.4s, v0.4s\n"
+ "fadd v30.4s, v30.4s, v31.4s\n"
+ "ldr q27, [%x[in_ptr], #0xc0]\n"
+ "ldr q26, [%x[in_ptr], #0xd0]\n"
+ "fadd v24.4s, v24.4s, v1.4s\n"
+ "fadd v23.4s, v23.4s, v0.4s\n"
+ "ldr q25, [%x[in_ptr], #0xe0]\n"
+ "fadd v22.4s, v22.4s, v31.4s\n"
+ "fadd v21.4s, v21.4s, v1.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v29.4s, v29.4s, v0.4s\n"
+ "fadd v28.4s, v28.4s, v31.4s\n"
+ "fadd v27.4s, v27.4s, v1.4s\n"
+ "fadd v26.4s, v26.4s, v0.4s\n"
+ "fadd v25.4s, v25.4s, v31.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v30.4s, v30.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v30.4s, v30.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ "str d16, [x9, #0x0]\n"
+ ".inst 0x0ea16bd0 // bfcvtn v16.4h, v30.4s\n"
+ ".inst 0x0ea16b18 // bfcvtn v24.4h, v24.4s\n"
+ "str d20, [x9, #0x8]\n"
+ "str d19, [x9, #0x10]\n"
+ ".inst 0x0ea16af7 // bfcvtn v23.4h, v23.4s\n"
+ ".inst 0x0ea16ad6 // bfcvtn v22.4h, v22.4s\n"
+ "add x9, x9, #0x18\n"
+ "str d18, [x27, #0x0]\n"
+ ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
+ ".inst 0x0ea16bb4 // bfcvtn v20.4h, v29.4s\n"
+ "str d17, [x27, #0x8]\n"
+ ".inst 0x0ea16b93 // bfcvtn v19.4h, v28.4s\n"
+ ".inst 0x0ea16b72 // bfcvtn v18.4h, v27.4s\n"
+ "str d16, [x27, #0x10]\n"
+ ".inst 0x0ea16b51 // bfcvtn v17.4h, v26.4s\n"
+ ".inst 0x0ea16b30 // bfcvtn v16.4h, v25.4s\n"
+ "add x27, x27, #0x18\n"
+ "str d24, [x26, #0x0]\n"
+ "str d23, [x26, #0x8]\n"
+ "str d22, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "str d21, [x25, #0x0]\n"
+ "str d20, [x25, #0x8]\n"
+ "str d19, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "str d18, [x24, #0x0]\n"
+ "str d17, [x24, #0x8]\n"
+ "str d16, [x24, #0x10]\n"
+ "add x24, x24, #0x18\n"
+ "bge 35b\n"
+ "38:" // Initial: Height 5: no full blocks
+ "cbz x10, 41f\n"
+ "mov x20, %x[in_ptr]\n"
+ "39:" // Initial: Height 5: Single loop
+ "movi v21.16b, #0x0\n"
+ "cbz %x[bias], 40f\n"
+ "ldr h16, [x28, #0x0]\n"
+ "shll v21.4s, v16.4h, #0x10\n"
+ "40:" // Initial: Height 5: Scalar: no bias
+ "ldr s16, [%x[in_ptr], #0x0]\n"
+ "ldr s19, [%x[in_ptr], #0x30]\n"
+ "subs x10, x10, #0x1\n"
+ "add x28, x28, #0x2\n"
+ "ldr s18, [%x[in_ptr], #0x60]\n"
+ "ldr s17, [%x[in_ptr], #0x90]\n"
+ "ldr s20, [%x[in_ptr], #0xc0]\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v16.4s, v16.4s, v21.4s\n"
+ "fadd v19.4s, v19.4s, v21.4s\n"
+ "fadd v18.4s, v18.4s, v21.4s\n"
+ "fadd v17.4s, v17.4s, v21.4s\n"
+ "fadd v20.4s, v20.4s, v21.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ "str h16, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ ".inst 0x0ea16a90 // bfcvtn v16.4h, v20.4s\n"
+ "str h19, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "str h18, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h17, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "str h16, [x24, #0x0]\n"
+ "add x24, x24, #0x2\n"
+ "bne 39b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "41:" // Initial: Height 5: no oddments
+ "b 108f\n"
+ "42:" // Initial: Height 6
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "mov x28, %x[bias]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "blt 46f\n"
+ "43:" // Initial: Height 6: Block loop
+ "cbnz %x[bias], 44f\n"
+ "movi v4.16b, #0x0\n"
+ "movi v3.16b, #0x0\n"
+ "movi v2.16b, #0x0\n"
+ "b 45f\n"
+ "44:" // Initial: Height 6: Width 3: bias
+ "ldr d18, [x28, #0x0]\n"
+ "ldr d17, [x28, #0x8]\n"
+ "ldr d16, [x28, #0x10]\n"
+ "shll v4.4s, v18.4h, #0x10\n"
+ "shll v3.4s, v17.4h, #0x10\n"
+ "shll v2.4s, v16.4h, #0x10\n"
+ "45:" // Initial: Height 6: Width 3: init done
+ "ldr q21, [%x[in_ptr], #0x0]\n"
+ "ldr q16, [%x[in_ptr], #0x10]\n"
+ "sub x10, x10, #0xc\n"
+ "add x28, x28, #0x18\n"
+ "ldr q20, [%x[in_ptr], #0x20]\n"
+ "ldr q19, [%x[in_ptr], #0x30]\n"
+ "cmp x10, #0xc\n"
+ "ldr q18, [%x[in_ptr], #0x40]\n"
+ "ldr q17, [%x[in_ptr], #0x50]\n"
+ "ldr q1, [%x[in_ptr], #0x60]\n"
+ "ldr q26, [%x[in_ptr], #0x70]\n"
+ "fadd v21.4s, v21.4s, v4.4s\n"
+ "fadd v16.4s, v16.4s, v3.4s\n"
+ "ldr q25, [%x[in_ptr], #0x80]\n"
+ "ldr q24, [%x[in_ptr], #0x90]\n"
+ "fadd v20.4s, v20.4s, v2.4s\n"
+ "fadd v19.4s, v19.4s, v4.4s\n"
+ "ldr q23, [%x[in_ptr], #0xa0]\n"
+ "ldr q22, [%x[in_ptr], #0xb0]\n"
+ "fadd v18.4s, v18.4s, v3.4s\n"
+ "fadd v17.4s, v17.4s, v2.4s\n"
+ "ldr q0, [%x[in_ptr], #0xc0]\n"
+ "ldr q31, [%x[in_ptr], #0xd0]\n"
+ "fadd v1.4s, v1.4s, v4.4s\n"
+ "fadd v26.4s, v26.4s, v3.4s\n"
+ "ldr q30, [%x[in_ptr], #0xe0]\n"
+ "ldr q29, [%x[in_ptr], #0xf0]\n"
+ "fadd v25.4s, v25.4s, v2.4s\n"
+ "fadd v24.4s, v24.4s, v4.4s\n"
+ "ldr q28, [%x[in_ptr], #0x100]\n"
+ "ldr q27, [%x[in_ptr], #0x110]\n"
+ "fadd v23.4s, v23.4s, v3.4s\n"
+ "fadd v22.4s, v22.4s, v2.4s\n"
+ "fadd v0.4s, v0.4s, v4.4s\n"
+ "fadd v31.4s, v31.4s, v3.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v30.4s, v30.4s, v2.4s\n"
+ "fadd v29.4s, v29.4s, v4.4s\n"
+ "fadd v28.4s, v28.4s, v3.4s\n"
+ "fadd v27.4s, v27.4s, v2.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v1.4s, v1.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmin v0.4s, v0.4s, v13.4s\n"
+ "fmin v31.4s, v31.4s, v13.4s\n"
+ "fmin v30.4s, v30.4s, v13.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v1.4s, v1.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ "fmax v0.4s, v0.4s, v12.4s\n"
+ "fmax v31.4s, v31.4s, v12.4s\n"
+ "fmax v30.4s, v30.4s, v12.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ "str d21, [x9, #0x0]\n"
+ "str d16, [x9, #0x8]\n"
+ ".inst 0x0ea16830 // bfcvtn v16.4h, v1.4s\n"
+ ".inst 0x0ea16b5a // bfcvtn v26.4h, v26.4s\n"
+ "str d20, [x9, #0x10]\n"
+ ".inst 0x0ea16b39 // bfcvtn v25.4h, v25.4s\n"
+ ".inst 0x0ea16b18 // bfcvtn v24.4h, v24.4s\n"
+ "add x9, x9, #0x18\n"
+ "str d19, [x27, #0x0]\n"
+ ".inst 0x0ea16af7 // bfcvtn v23.4h, v23.4s\n"
+ ".inst 0x0ea16ad6 // bfcvtn v22.4h, v22.4s\n"
+ "str d18, [x27, #0x8]\n"
+ ".inst 0x0ea16815 // bfcvtn v21.4h, v0.4s\n"
+ ".inst 0x0ea16bf4 // bfcvtn v20.4h, v31.4s\n"
+ "str d17, [x27, #0x10]\n"
+ ".inst 0x0ea16bd3 // bfcvtn v19.4h, v30.4s\n"
+ ".inst 0x0ea16bb2 // bfcvtn v18.4h, v29.4s\n"
+ "add x27, x27, #0x18\n"
+ "str d16, [x26, #0x0]\n"
+ ".inst 0x0ea16b91 // bfcvtn v17.4h, v28.4s\n"
+ ".inst 0x0ea16b70 // bfcvtn v16.4h, v27.4s\n"
+ "str d26, [x26, #0x8]\n"
+ "str d25, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "str d24, [x25, #0x0]\n"
+ "str d23, [x25, #0x8]\n"
+ "str d22, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "str d21, [x24, #0x0]\n"
+ "str d20, [x24, #0x8]\n"
+ "str d19, [x24, #0x10]\n"
+ "add x24, x24, #0x18\n"
+ "str d18, [x23, #0x0]\n"
+ "str d17, [x23, #0x8]\n"
+ "str d16, [x23, #0x10]\n"
+ "add x23, x23, #0x18\n"
+ "bge 43b\n"
+ "46:" // Initial: Height 6: no full blocks
+ "cbz x10, 49f\n"
+ "mov x20, %x[in_ptr]\n"
+ "47:" // Initial: Height 6: Single loop
+ "movi v22.16b, #0x0\n"
+ "cbz %x[bias], 48f\n"
+ "ldr h16, [x28, #0x0]\n"
+ "shll v22.4s, v16.4h, #0x10\n"
+ "48:" // Initial: Height 6: Scalar: no bias
+ "ldr s16, [%x[in_ptr], #0x0]\n"
+ "ldr s20, [%x[in_ptr], #0x30]\n"
+ "subs x10, x10, #0x1\n"
+ "add x28, x28, #0x2\n"
+ "ldr s19, [%x[in_ptr], #0x60]\n"
+ "ldr s18, [%x[in_ptr], #0x90]\n"
+ "ldr s17, [%x[in_ptr], #0xc0]\n"
+ "ldr s21, [%x[in_ptr], #0xf0]\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v16.4s, v16.4s, v22.4s\n"
+ "fadd v20.4s, v20.4s, v22.4s\n"
+ "fadd v19.4s, v19.4s, v22.4s\n"
+ "fadd v18.4s, v18.4s, v22.4s\n"
+ "fadd v17.4s, v17.4s, v22.4s\n"
+ "fadd v21.4s, v21.4s, v22.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ "str h16, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ ".inst 0x0ea16ab0 // bfcvtn v16.4h, v21.4s\n"
+ "str h20, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "str h19, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h18, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "str h17, [x24, #0x0]\n"
+ "add x24, x24, #0x2\n"
+ "str h16, [x23, #0x0]\n"
+ "add x23, x23, #0x2\n"
+ "bne 47b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "49:" // Initial: Height 6: no oddments
+ "b 108f\n"
+ "50:" // Initial: Height 7
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "mov x28, %x[bias]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "add x22, x23, %x[ldout], LSL #1\n"
+ "blt 54f\n"
+ "51:" // Initial: Height 7: Block loop
+ "cbnz %x[bias], 52f\n"
+ "movi v7.16b, #0x0\n"
+ "movi v6.16b, #0x0\n"
+ "movi v5.16b, #0x0\n"
+ "b 53f\n"
+ "52:" // Initial: Height 7: Width 3: bias
+ "ldr d18, [x28, #0x0]\n"
+ "ldr d17, [x28, #0x8]\n"
+ "ldr d16, [x28, #0x10]\n"
+ "shll v7.4s, v18.4h, #0x10\n"
+ "shll v6.4s, v17.4h, #0x10\n"
+ "shll v5.4s, v16.4h, #0x10\n"
+ "53:" // Initial: Height 7: Width 3: init done
+ "ldr q18, [%x[in_ptr], #0x0]\n"
+ "ldr q17, [%x[in_ptr], #0x10]\n"
+ "sub x10, x10, #0xc\n"
+ "add x28, x28, #0x18\n"
+ "ldr q16, [%x[in_ptr], #0x20]\n"
+ "ldr q21, [%x[in_ptr], #0x30]\n"
+ "cmp x10, #0xc\n"
+ "ldr q20, [%x[in_ptr], #0x40]\n"
+ "ldr q19, [%x[in_ptr], #0x50]\n"
+ "ldr q4, [%x[in_ptr], #0x60]\n"
+ "ldr q3, [%x[in_ptr], #0x70]\n"
+ "fadd v18.4s, v18.4s, v7.4s\n"
+ "fadd v17.4s, v17.4s, v6.4s\n"
+ "ldr q2, [%x[in_ptr], #0x80]\n"
+ "ldr q27, [%x[in_ptr], #0x90]\n"
+ "fadd v16.4s, v16.4s, v5.4s\n"
+ "fadd v21.4s, v21.4s, v7.4s\n"
+ "ldr q26, [%x[in_ptr], #0xa0]\n"
+ "ldr q25, [%x[in_ptr], #0xb0]\n"
+ "fadd v20.4s, v20.4s, v6.4s\n"
+ "fadd v19.4s, v19.4s, v5.4s\n"
+ "ldr q24, [%x[in_ptr], #0xc0]\n"
+ "ldr q23, [%x[in_ptr], #0xd0]\n"
+ "fadd v4.4s, v4.4s, v7.4s\n"
+ "fadd v3.4s, v3.4s, v6.4s\n"
+ "ldr q22, [%x[in_ptr], #0xe0]\n"
+ "ldr q1, [%x[in_ptr], #0xf0]\n"
+ "fadd v2.4s, v2.4s, v5.4s\n"
+ "fadd v27.4s, v27.4s, v7.4s\n"
+ "ldr q0, [%x[in_ptr], #0x100]\n"
+ "ldr q31, [%x[in_ptr], #0x110]\n"
+ "fadd v26.4s, v26.4s, v6.4s\n"
+ "fadd v25.4s, v25.4s, v5.4s\n"
+ "ldr q30, [%x[in_ptr], #0x120]\n"
+ "ldr q29, [%x[in_ptr], #0x130]\n"
+ "fadd v24.4s, v24.4s, v7.4s\n"
+ "fadd v23.4s, v23.4s, v6.4s\n"
+ "ldr q28, [%x[in_ptr], #0x140]\n"
+ "fadd v22.4s, v22.4s, v5.4s\n"
+ "fadd v1.4s, v1.4s, v7.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v0.4s, v0.4s, v6.4s\n"
+ "fadd v31.4s, v31.4s, v5.4s\n"
+ "fadd v30.4s, v30.4s, v7.4s\n"
+ "fadd v29.4s, v29.4s, v6.4s\n"
+ "fadd v28.4s, v28.4s, v5.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v4.4s, v4.4s, v13.4s\n"
+ "fmin v3.4s, v3.4s, v13.4s\n"
+ "fmin v2.4s, v2.4s, v13.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmin v1.4s, v1.4s, v13.4s\n"
+ "fmin v0.4s, v0.4s, v13.4s\n"
+ "fmin v31.4s, v31.4s, v13.4s\n"
+ "fmin v30.4s, v30.4s, v13.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v4.4s, v4.4s, v12.4s\n"
+ "fmax v3.4s, v3.4s, v12.4s\n"
+ "fmax v2.4s, v2.4s, v12.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ "fmax v1.4s, v1.4s, v12.4s\n"
+ "fmax v0.4s, v0.4s, v12.4s\n"
+ "fmax v31.4s, v31.4s, v12.4s\n"
+ "fmax v30.4s, v30.4s, v12.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ "str d18, [x9, #0x0]\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16892 // bfcvtn v18.4h, v4.4s\n"
+ "str d17, [x9, #0x8]\n"
+ "str d16, [x9, #0x10]\n"
+ ".inst 0x0ea16871 // bfcvtn v17.4h, v3.4s\n"
+ ".inst 0x0ea16850 // bfcvtn v16.4h, v2.4s\n"
+ "add x9, x9, #0x18\n"
+ "str d21, [x27, #0x0]\n"
+ ".inst 0x0ea16b7b // bfcvtn v27.4h, v27.4s\n"
+ ".inst 0x0ea16b5a // bfcvtn v26.4h, v26.4s\n"
+ "str d20, [x27, #0x8]\n"
+ ".inst 0x0ea16b39 // bfcvtn v25.4h, v25.4s\n"
+ ".inst 0x0ea16b18 // bfcvtn v24.4h, v24.4s\n"
+ "str d19, [x27, #0x10]\n"
+ ".inst 0x0ea16af7 // bfcvtn v23.4h, v23.4s\n"
+ ".inst 0x0ea16ad6 // bfcvtn v22.4h, v22.4s\n"
+ "add x27, x27, #0x18\n"
+ "str d18, [x26, #0x0]\n"
+ ".inst 0x0ea16835 // bfcvtn v21.4h, v1.4s\n"
+ ".inst 0x0ea16814 // bfcvtn v20.4h, v0.4s\n"
+ "str d17, [x26, #0x8]\n"
+ ".inst 0x0ea16bf3 // bfcvtn v19.4h, v31.4s\n"
+ ".inst 0x0ea16bd2 // bfcvtn v18.4h, v30.4s\n"
+ "str d16, [x26, #0x10]\n"
+ ".inst 0x0ea16bb1 // bfcvtn v17.4h, v29.4s\n"
+ ".inst 0x0ea16b90 // bfcvtn v16.4h, v28.4s\n"
+ "add x26, x26, #0x18\n"
+ "str d27, [x25, #0x0]\n"
+ "str d26, [x25, #0x8]\n"
+ "str d25, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "str d24, [x24, #0x0]\n"
+ "str d23, [x24, #0x8]\n"
+ "str d22, [x24, #0x10]\n"
+ "add x24, x24, #0x18\n"
+ "str d21, [x23, #0x0]\n"
+ "str d20, [x23, #0x8]\n"
+ "str d19, [x23, #0x10]\n"
+ "add x23, x23, #0x18\n"
+ "str d18, [x22, #0x0]\n"
+ "str d17, [x22, #0x8]\n"
+ "str d16, [x22, #0x10]\n"
+ "add x22, x22, #0x18\n"
+ "bge 51b\n"
+ "54:" // Initial: Height 7: no full blocks
+ "cbz x10, 57f\n"
+ "mov x20, %x[in_ptr]\n"
+ "55:" // Initial: Height 7: Single loop
+ "movi v23.16b, #0x0\n"
+ "cbz %x[bias], 56f\n"
+ "ldr h16, [x28, #0x0]\n"
+ "shll v23.4s, v16.4h, #0x10\n"
+ "56:" // Initial: Height 7: Scalar: no bias
+ "ldr s16, [%x[in_ptr], #0x0]\n"
+ "ldr s21, [%x[in_ptr], #0x30]\n"
+ "subs x10, x10, #0x1\n"
+ "add x28, x28, #0x2\n"
+ "ldr s20, [%x[in_ptr], #0x60]\n"
+ "ldr s19, [%x[in_ptr], #0x90]\n"
+ "ldr s18, [%x[in_ptr], #0xc0]\n"
+ "ldr s17, [%x[in_ptr], #0xf0]\n"
+ "ldr s22, [%x[in_ptr], #0x120]\n"
+ "fadd v16.4s, v16.4s, v23.4s\n"
+ "fadd v21.4s, v21.4s, v23.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v20.4s, v20.4s, v23.4s\n"
+ "fadd v19.4s, v19.4s, v23.4s\n"
+ "fadd v18.4s, v18.4s, v23.4s\n"
+ "fadd v17.4s, v17.4s, v23.4s\n"
+ "fadd v22.4s, v22.4s, v23.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ "str h16, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ ".inst 0x0ea16ad0 // bfcvtn v16.4h, v22.4s\n"
+ "str h21, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "str h20, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h19, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "str h18, [x24, #0x0]\n"
+ "add x24, x24, #0x2\n"
+ "str h17, [x23, #0x0]\n"
+ "add x23, x23, #0x2\n"
+ "str h16, [x22, #0x0]\n"
+ "add x22, x22, #0x2\n"
+ "bne 55b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "57:" // Initial: Height 7: no oddments
+ "b 108f\n"
+ "58:" // Initial: Height 8
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "mov x28, %x[bias]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "add x22, x23, %x[ldout], LSL #1\n"
+ "add x21, x22, %x[ldout], LSL #1\n"
+ "blt 62f\n"
+ "59:" // Initial: Height 8: Block loop
+ "cbnz %x[bias], 60f\n"
+ "movi v10.16b, #0x0\n"
+ "movi v9.16b, #0x0\n"
+ "movi v8.16b, #0x0\n"
+ "b 61f\n"
+ "60:" // Initial: Height 8: Width 3: bias
+ "ldr d18, [x28, #0x0]\n"
+ "ldr d17, [x28, #0x8]\n"
+ "ldr d16, [x28, #0x10]\n"
+ "shll v10.4s, v18.4h, #0x10\n"
+ "shll v9.4s, v17.4h, #0x10\n"
+ "shll v8.4s, v16.4h, #0x10\n"
+ "61:" // Initial: Height 8: Width 3: init done
+ "ldr q18, [%x[in_ptr], #0x0]\n"
+ "ldr q17, [%x[in_ptr], #0x10]\n"
+ "sub x10, x10, #0xc\n"
+ "add x28, x28, #0x18\n"
+ "ldr q16, [%x[in_ptr], #0x20]\n"
+ "ldr q22, [%x[in_ptr], #0x30]\n"
+ "cmp x10, #0xc\n"
+ "ldr q21, [%x[in_ptr], #0x40]\n"
+ "ldr q20, [%x[in_ptr], #0x50]\n"
+ "ldr q19, [%x[in_ptr], #0x60]\n"
+ "ldr q7, [%x[in_ptr], #0x70]\n"
+ "fadd v18.4s, v18.4s, v10.4s\n"
+ "fadd v17.4s, v17.4s, v9.4s\n"
+ "ldr q6, [%x[in_ptr], #0x80]\n"
+ "ldr q5, [%x[in_ptr], #0x90]\n"
+ "fadd v16.4s, v16.4s, v8.4s\n"
+ "fadd v22.4s, v22.4s, v10.4s\n"
+ "ldr q29, [%x[in_ptr], #0xa0]\n"
+ "ldr q28, [%x[in_ptr], #0xb0]\n"
+ "fadd v21.4s, v21.4s, v9.4s\n"
+ "fadd v20.4s, v20.4s, v8.4s\n"
+ "ldr q27, [%x[in_ptr], #0xc0]\n"
+ "ldr q26, [%x[in_ptr], #0xd0]\n"
+ "fadd v19.4s, v19.4s, v10.4s\n"
+ "fadd v7.4s, v7.4s, v9.4s\n"
+ "ldr q25, [%x[in_ptr], #0xe0]\n"
+ "ldr q24, [%x[in_ptr], #0xf0]\n"
+ "fadd v6.4s, v6.4s, v8.4s\n"
+ "fadd v5.4s, v5.4s, v10.4s\n"
+ "ldr q23, [%x[in_ptr], #0x100]\n"
+ "ldr q4, [%x[in_ptr], #0x110]\n"
+ "fadd v29.4s, v29.4s, v9.4s\n"
+ "fadd v28.4s, v28.4s, v8.4s\n"
+ "ldr q3, [%x[in_ptr], #0x120]\n"
+ "ldr q2, [%x[in_ptr], #0x130]\n"
+ "fadd v27.4s, v27.4s, v10.4s\n"
+ "fadd v26.4s, v26.4s, v9.4s\n"
+ "ldr q1, [%x[in_ptr], #0x140]\n"
+ "ldr q0, [%x[in_ptr], #0x150]\n"
+ "fadd v25.4s, v25.4s, v8.4s\n"
+ "fadd v24.4s, v24.4s, v10.4s\n"
+ "ldr q31, [%x[in_ptr], #0x160]\n"
+ "ldr q30, [%x[in_ptr], #0x170]\n"
+ "fadd v23.4s, v23.4s, v9.4s\n"
+ "fadd v4.4s, v4.4s, v8.4s\n"
+ "fadd v3.4s, v3.4s, v10.4s\n"
+ "fadd v2.4s, v2.4s, v9.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v1.4s, v1.4s, v8.4s\n"
+ "fadd v0.4s, v0.4s, v10.4s\n"
+ "fadd v31.4s, v31.4s, v9.4s\n"
+ "fadd v30.4s, v30.4s, v8.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v7.4s, v7.4s, v13.4s\n"
+ "fmin v6.4s, v6.4s, v13.4s\n"
+ "fmin v5.4s, v5.4s, v13.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v4.4s, v4.4s, v13.4s\n"
+ "fmin v3.4s, v3.4s, v13.4s\n"
+ "fmin v2.4s, v2.4s, v13.4s\n"
+ "fmin v1.4s, v1.4s, v13.4s\n"
+ "fmin v0.4s, v0.4s, v13.4s\n"
+ "fmin v31.4s, v31.4s, v13.4s\n"
+ "fmin v30.4s, v30.4s, v13.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v7.4s, v7.4s, v12.4s\n"
+ "fmax v6.4s, v6.4s, v12.4s\n"
+ "fmax v5.4s, v5.4s, v12.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v4.4s, v4.4s, v12.4s\n"
+ "fmax v3.4s, v3.4s, v12.4s\n"
+ "fmax v2.4s, v2.4s, v12.4s\n"
+ "fmax v1.4s, v1.4s, v12.4s\n"
+ "fmax v0.4s, v0.4s, v12.4s\n"
+ "fmax v31.4s, v31.4s, v12.4s\n"
+ "fmax v30.4s, v30.4s, v12.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16ad6 // bfcvtn v22.4h, v22.4s\n"
+ ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ "str d18, [x9, #0x0]\n"
+ "str d17, [x9, #0x8]\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea168f2 // bfcvtn v18.4h, v7.4s\n"
+ "str d16, [x9, #0x10]\n"
+ ".inst 0x0ea168d1 // bfcvtn v17.4h, v6.4s\n"
+ ".inst 0x0ea168b0 // bfcvtn v16.4h, v5.4s\n"
+ "add x9, x9, #0x18\n"
+ "str d22, [x27, #0x0]\n"
+ ".inst 0x0ea16bbd // bfcvtn v29.4h, v29.4s\n"
+ ".inst 0x0ea16b9c // bfcvtn v28.4h, v28.4s\n"
+ "str d21, [x27, #0x8]\n"
+ ".inst 0x0ea16b7b // bfcvtn v27.4h, v27.4s\n"
+ ".inst 0x0ea16b5a // bfcvtn v26.4h, v26.4s\n"
+ "str d20, [x27, #0x10]\n"
+ ".inst 0x0ea16b39 // bfcvtn v25.4h, v25.4s\n"
+ ".inst 0x0ea16b18 // bfcvtn v24.4h, v24.4s\n"
+ "add x27, x27, #0x18\n"
+ "str d19, [x26, #0x0]\n"
+ ".inst 0x0ea16af7 // bfcvtn v23.4h, v23.4s\n"
+ ".inst 0x0ea16896 // bfcvtn v22.4h, v4.4s\n"
+ "str d18, [x26, #0x8]\n"
+ ".inst 0x0ea16875 // bfcvtn v21.4h, v3.4s\n"
+ ".inst 0x0ea16854 // bfcvtn v20.4h, v2.4s\n"
+ "str d17, [x26, #0x10]\n"
+ ".inst 0x0ea16833 // bfcvtn v19.4h, v1.4s\n"
+ ".inst 0x0ea16812 // bfcvtn v18.4h, v0.4s\n"
+ "add x26, x26, #0x18\n"
+ "str d16, [x25, #0x0]\n"
+ ".inst 0x0ea16bf1 // bfcvtn v17.4h, v31.4s\n"
+ ".inst 0x0ea16bd0 // bfcvtn v16.4h, v30.4s\n"
+ "str d29, [x25, #0x8]\n"
+ "str d28, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "str d27, [x24, #0x0]\n"
+ "str d26, [x24, #0x8]\n"
+ "str d25, [x24, #0x10]\n"
+ "add x24, x24, #0x18\n"
+ "str d24, [x23, #0x0]\n"
+ "str d23, [x23, #0x8]\n"
+ "str d22, [x23, #0x10]\n"
+ "add x23, x23, #0x18\n"
+ "str d21, [x22, #0x0]\n"
+ "str d20, [x22, #0x8]\n"
+ "str d19, [x22, #0x10]\n"
+ "add x22, x22, #0x18\n"
+ "str d18, [x21, #0x0]\n"
+ "str d17, [x21, #0x8]\n"
+ "str d16, [x21, #0x10]\n"
+ "add x21, x21, #0x18\n"
+ "bge 59b\n"
+ "62:" // Initial: Height 8: no full blocks
+ "cbz x10, 65f\n"
+ "mov x20, %x[in_ptr]\n"
+ "63:" // Initial: Height 8: Single loop
+ "movi v24.16b, #0x0\n"
+ "cbz %x[bias], 64f\n"
+ "ldr h16, [x28, #0x0]\n"
+ "shll v24.4s, v16.4h, #0x10\n"
+ "64:" // Initial: Height 8: Scalar: no bias
+ "ldr s17, [%x[in_ptr], #0x0]\n"
+ "ldr s16, [%x[in_ptr], #0x30]\n"
+ "subs x10, x10, #0x1\n"
+ "add x28, x28, #0x2\n"
+ "ldr s21, [%x[in_ptr], #0x60]\n"
+ "ldr s20, [%x[in_ptr], #0x90]\n"
+ "ldr s19, [%x[in_ptr], #0xc0]\n"
+ "ldr s18, [%x[in_ptr], #0xf0]\n"
+ "ldr s23, [%x[in_ptr], #0x120]\n"
+ "ldr s22, [%x[in_ptr], #0x150]\n"
+ "fadd v17.4s, v17.4s, v24.4s\n"
+ "fadd v16.4s, v16.4s, v24.4s\n"
+ "fadd v21.4s, v21.4s, v24.4s\n"
+ "fadd v20.4s, v20.4s, v24.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v19.4s, v19.4s, v24.4s\n"
+ "fadd v18.4s, v18.4s, v24.4s\n"
+ "fadd v23.4s, v23.4s, v24.4s\n"
+ "fadd v22.4s, v22.4s, v24.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v16.4s, v16.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v16.4s, v16.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ "str h17, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ "str h16, [x27, #0x0]\n"
+ ".inst 0x0ea16af1 // bfcvtn v17.4h, v23.4s\n"
+ ".inst 0x0ea16ad0 // bfcvtn v16.4h, v22.4s\n"
+ "add x27, x27, #0x2\n"
+ "str h21, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h20, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "str h19, [x24, #0x0]\n"
+ "add x24, x24, #0x2\n"
+ "str h18, [x23, #0x0]\n"
+ "add x23, x23, #0x2\n"
+ "str h17, [x22, #0x0]\n"
+ "add x22, x22, #0x2\n"
+ "str h16, [x21, #0x0]\n"
+ "add x21, x21, #0x2\n"
+ "bne 63b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "65:" // Initial: Height 8: no oddments
+ "subs %x[rows], %x[rows], #0x8\n"
+ "add %x[out_ptr], %x[out_ptr], x11\n"
+ "bgt 1b\n"
+ "b 108f\n"
+ "66:" // Accumulate
+ "67:" // Accumulate: Row loop
+ "cmp %x[rows], #0x7\n"
+ "bgt 103f\n"
+ "beq 98f\n"
+ "cmp %x[rows], #0x5\n"
+ "bgt 93f\n"
+ "beq 88f\n"
+ "cmp %x[rows], #0x3\n"
+ "bgt 83f\n"
+ "beq 78f\n"
+ "cmp %x[rows], #0x1\n"
+ "bgt 73f\n"
+ "68:" // Accumulate: Height 1
+ "mov x10, %x[cols]\n"
+ "mov x9, %x[out_ptr]\n"
+ "cmp x10, #0xc\n"
+ "blt 70f\n"
+ "69:" // Accumulate: Height 1: Block loop
+ "ldr d16, [x9, #0x0]\n"
+ "ldr q19, [%x[in_ptr], #0x0]\n"
+ "sub x10, x10, #0xc\n"
+ "ldr q18, [%x[in_ptr], #0x10]\n"
+ "ldr q17, [%x[in_ptr], #0x20]\n"
+ "cmp x10, #0xc\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v19.4s, v19.4s, v16.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ ".inst 0x0ea16a70 // bfcvtn v16.4h, v19.4s\n"
+ "str d16, [x9, #0x0]\n"
+ "ldr d16, [x9, #0x8]\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v18.4s, v18.4s, v16.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ ".inst 0x0ea16a50 // bfcvtn v16.4h, v18.4s\n"
+ "str d16, [x9, #0x8]\n"
+ "ldr d16, [x9, #0x10]\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v17.4s, v17.4s, v16.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ ".inst 0x0ea16a30 // bfcvtn v16.4h, v17.4s\n"
+ "str d16, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ "bge 69b\n"
+ "70:" // Accumulate: Height 1: no full blocks
+ "cbz x10, 72f\n"
+ "mov x20, %x[in_ptr]\n"
+ "71:" // Accumulate: Height 1: Single loop
+ "ldr h16, [x9, #0x0]\n"
+ "ldr s17, [%x[in_ptr], #0x0]\n"
+ "subs x10, x10, #0x1\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v17.4s, v17.4s, v16.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ ".inst 0x0ea16a30 // bfcvtn v16.4h, v17.4s\n"
+ "str h16, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ "bne 71b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "72:" // Accumulate: Height 1: no oddments
+ "b 108f\n"
+ "73:" // Accumulate: Height 2
+ "mov x10, %x[cols]\n"
+ "mov x9, %x[out_ptr]\n"
+ "cmp x10, #0xc\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "blt 75f\n"
+ "74:" // Accumulate: Height 2: Block loop
+ "ldr d17, [x9, #0x0]\n"
+ "ldr d16, [x27, #0x0]\n"
+ "sub x10, x10, #0xc\n"
+ "ldr q23, [%x[in_ptr], #0x0]\n"
+ "ldr q22, [%x[in_ptr], #0x30]\n"
+ "cmp x10, #0xc\n"
+ "ldr q21, [%x[in_ptr], #0x10]\n"
+ "ldr q20, [%x[in_ptr], #0x40]\n"
+ "ldr q19, [%x[in_ptr], #0x20]\n"
+ "ldr q18, [%x[in_ptr], #0x50]\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v23.4s, v23.4s, v17.4s\n"
+ "fadd v22.4s, v22.4s, v16.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16af0 // bfcvtn v16.4h, v23.4s\n"
+ ".inst 0x0ea16ad1 // bfcvtn v17.4h, v22.4s\n"
+ "str d16, [x9, #0x0]\n"
+ "ldr d16, [x9, #0x8]\n"
+ "str d17, [x27, #0x0]\n"
+ "shll v17.4s, v16.4h, #0x10\n"
+ "ldr d16, [x27, #0x8]\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v21.4s, v21.4s, v17.4s\n"
+ "fadd v20.4s, v20.4s, v16.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ ".inst 0x0ea16ab0 // bfcvtn v16.4h, v21.4s\n"
+ ".inst 0x0ea16a91 // bfcvtn v17.4h, v20.4s\n"
+ "str d16, [x9, #0x8]\n"
+ "ldr d16, [x9, #0x10]\n"
+ "str d17, [x27, #0x8]\n"
+ "shll v17.4s, v16.4h, #0x10\n"
+ "ldr d16, [x27, #0x10]\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v19.4s, v19.4s, v17.4s\n"
+ "fadd v18.4s, v18.4s, v16.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ ".inst 0x0ea16a71 // bfcvtn v17.4h, v19.4s\n"
+ ".inst 0x0ea16a50 // bfcvtn v16.4h, v18.4s\n"
+ "str d17, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ "str d16, [x27, #0x10]\n"
+ "add x27, x27, #0x18\n"
+ "bge 74b\n"
+ "75:" // Accumulate: Height 2: no full blocks
+ "cbz x10, 77f\n"
+ "mov x20, %x[in_ptr]\n"
+ "76:" // Accumulate: Height 2: Single loop
+ "ldr h17, [x9, #0x0]\n"
+ "ldr h16, [x27, #0x0]\n"
+ "subs x10, x10, #0x1\n"
+ "ldr s19, [%x[in_ptr], #0x0]\n"
+ "ldr s18, [%x[in_ptr], #0x30]\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v19.4s, v19.4s, v17.4s\n"
+ "fadd v18.4s, v18.4s, v16.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ ".inst 0x0ea16a70 // bfcvtn v16.4h, v19.4s\n"
+ "str h16, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ ".inst 0x0ea16a50 // bfcvtn v16.4h, v18.4s\n"
+ "str h16, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "bne 76b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "77:" // Accumulate: Height 2: no oddments
+ "b 108f\n"
+ "78:" // Accumulate: Height 3
+ "mov x10, %x[cols]\n"
+ "mov x9, %x[out_ptr]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "blt 80f\n"
+ "79:" // Accumulate: Height 3: Block loop
+ "ldr d18, [x9, #0x0]\n"
+ "ldr d17, [x27, #0x0]\n"
+ "sub x10, x10, #0xc\n"
+ "ldr d16, [x26, #0x0]\n"
+ "ldr q27, [%x[in_ptr], #0x0]\n"
+ "cmp x10, #0xc\n"
+ "ldr q26, [%x[in_ptr], #0x30]\n"
+ "ldr q25, [%x[in_ptr], #0x60]\n"
+ "ldr q24, [%x[in_ptr], #0x10]\n"
+ "ldr q23, [%x[in_ptr], #0x40]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "ldr q22, [%x[in_ptr], #0x70]\n"
+ "ldr q21, [%x[in_ptr], #0x20]\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "ldr q20, [%x[in_ptr], #0x50]\n"
+ "ldr q19, [%x[in_ptr], #0x80]\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fadd v27.4s, v27.4s, v18.4s\n"
+ "fadd v26.4s, v26.4s, v17.4s\n"
+ "fadd v25.4s, v25.4s, v16.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ ".inst 0x0ea16b72 // bfcvtn v18.4h, v27.4s\n"
+ ".inst 0x0ea16b50 // bfcvtn v16.4h, v26.4s\n"
+ ".inst 0x0ea16b31 // bfcvtn v17.4h, v25.4s\n"
+ "str d18, [x9, #0x0]\n"
+ "str d16, [x27, #0x0]\n"
+ "ldr d16, [x9, #0x8]\n"
+ "str d17, [x26, #0x0]\n"
+ "ldr d17, [x27, #0x8]\n"
+ "shll v18.4s, v16.4h, #0x10\n"
+ "ldr d16, [x26, #0x8]\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v24.4s, v24.4s, v18.4s\n"
+ "fadd v23.4s, v23.4s, v17.4s\n"
+ "fadd v22.4s, v22.4s, v16.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16b10 // bfcvtn v16.4h, v24.4s\n"
+ ".inst 0x0ea16af2 // bfcvtn v18.4h, v23.4s\n"
+ "str d16, [x9, #0x8]\n"
+ ".inst 0x0ea16ad1 // bfcvtn v17.4h, v22.4s\n"
+ "ldr d16, [x9, #0x10]\n"
+ "str d18, [x27, #0x8]\n"
+ "str d17, [x26, #0x8]\n"
+ "shll v18.4s, v16.4h, #0x10\n"
+ "ldr d17, [x27, #0x10]\n"
+ "ldr d16, [x26, #0x10]\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v21.4s, v21.4s, v18.4s\n"
+ "fadd v20.4s, v20.4s, v17.4s\n"
+ "fadd v19.4s, v19.4s, v16.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ ".inst 0x0ea16ab0 // bfcvtn v16.4h, v21.4s\n"
+ ".inst 0x0ea16a91 // bfcvtn v17.4h, v20.4s\n"
+ "str d16, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ ".inst 0x0ea16a70 // bfcvtn v16.4h, v19.4s\n"
+ "str d17, [x27, #0x10]\n"
+ "add x27, x27, #0x18\n"
+ "str d16, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "bge 79b\n"
+ "80:" // Accumulate: Height 3: no full blocks
+ "cbz x10, 82f\n"
+ "mov x20, %x[in_ptr]\n"
+ "81:" // Accumulate: Height 3: Single loop
+ "ldr h18, [x9, #0x0]\n"
+ "ldr h17, [x27, #0x0]\n"
+ "subs x10, x10, #0x1\n"
+ "ldr h16, [x26, #0x0]\n"
+ "ldr s21, [%x[in_ptr], #0x0]\n"
+ "ldr s20, [%x[in_ptr], #0x30]\n"
+ "ldr s19, [%x[in_ptr], #0x60]\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v21.4s, v21.4s, v18.4s\n"
+ "fadd v20.4s, v20.4s, v17.4s\n"
+ "fadd v19.4s, v19.4s, v16.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ ".inst 0x0ea16ab0 // bfcvtn v16.4h, v21.4s\n"
+ ".inst 0x0ea16a91 // bfcvtn v17.4h, v20.4s\n"
+ "str h16, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ ".inst 0x0ea16a70 // bfcvtn v16.4h, v19.4s\n"
+ "str h17, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "str h16, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "bne 81b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "82:" // Accumulate: Height 3: no oddments
+ "b 108f\n"
+ "83:" // Accumulate: Height 4
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "blt 85f\n"
+ "84:" // Accumulate: Height 4: Block loop
+ "ldr d19, [x9, #0x0]\n"
+ "ldr d18, [x27, #0x0]\n"
+ "sub x10, x10, #0xc\n"
+ "ldr d17, [x26, #0x0]\n"
+ "ldr d16, [x25, #0x0]\n"
+ "cmp x10, #0xc\n"
+ "ldr q31, [%x[in_ptr], #0x0]\n"
+ "ldr q30, [%x[in_ptr], #0x30]\n"
+ "ldr q29, [%x[in_ptr], #0x60]\n"
+ "ldr q28, [%x[in_ptr], #0x90]\n"
+ "shll v19.4s, v19.4h, #0x10\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "ldr q27, [%x[in_ptr], #0x10]\n"
+ "ldr q26, [%x[in_ptr], #0x40]\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "ldr q25, [%x[in_ptr], #0x70]\n"
+ "ldr q24, [%x[in_ptr], #0xa0]\n"
+ "ldr q23, [%x[in_ptr], #0x20]\n"
+ "ldr q22, [%x[in_ptr], #0x50]\n"
+ "fadd v31.4s, v31.4s, v19.4s\n"
+ "fadd v30.4s, v30.4s, v18.4s\n"
+ "ldr q21, [%x[in_ptr], #0x80]\n"
+ "ldr q20, [%x[in_ptr], #0xb0]\n"
+ "fadd v29.4s, v29.4s, v17.4s\n"
+ "fadd v28.4s, v28.4s, v16.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fmin v31.4s, v31.4s, v13.4s\n"
+ "fmin v30.4s, v30.4s, v13.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fmax v31.4s, v31.4s, v12.4s\n"
+ "fmax v30.4s, v30.4s, v12.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ ".inst 0x0ea16bf3 // bfcvtn v19.4h, v31.4s\n"
+ ".inst 0x0ea16bd0 // bfcvtn v16.4h, v30.4s\n"
+ ".inst 0x0ea16bb2 // bfcvtn v18.4h, v29.4s\n"
+ ".inst 0x0ea16b91 // bfcvtn v17.4h, v28.4s\n"
+ "str d19, [x9, #0x0]\n"
+ "str d16, [x27, #0x0]\n"
+ "ldr d16, [x9, #0x8]\n"
+ "str d18, [x26, #0x0]\n"
+ "str d17, [x25, #0x0]\n"
+ "ldr d18, [x27, #0x8]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "ldr d17, [x26, #0x8]\n"
+ "ldr d16, [x25, #0x8]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "fadd v27.4s, v27.4s, v19.4s\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v26.4s, v26.4s, v18.4s\n"
+ "fadd v25.4s, v25.4s, v17.4s\n"
+ "fadd v24.4s, v24.4s, v16.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ ".inst 0x0ea16b71 // bfcvtn v17.4h, v27.4s\n"
+ ".inst 0x0ea16b53 // bfcvtn v19.4h, v26.4s\n"
+ ".inst 0x0ea16b30 // bfcvtn v16.4h, v25.4s\n"
+ "str d17, [x9, #0x8]\n"
+ ".inst 0x0ea16b12 // bfcvtn v18.4h, v24.4s\n"
+ "ldr d17, [x9, #0x10]\n"
+ "str d19, [x27, #0x8]\n"
+ "str d16, [x26, #0x8]\n"
+ "ldr d16, [x27, #0x10]\n"
+ "str d18, [x25, #0x8]\n"
+ "shll v19.4s, v17.4h, #0x10\n"
+ "ldr d17, [x26, #0x10]\n"
+ "shll v18.4s, v16.4h, #0x10\n"
+ "ldr d16, [x25, #0x10]\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "fadd v23.4s, v23.4s, v19.4s\n"
+ "fadd v22.4s, v22.4s, v18.4s\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v21.4s, v21.4s, v17.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fadd v20.4s, v20.4s, v16.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ ".inst 0x0ea16af1 // bfcvtn v17.4h, v23.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ ".inst 0x0ea16ad0 // bfcvtn v16.4h, v22.4s\n"
+ "str d17, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ ".inst 0x0ea16ab1 // bfcvtn v17.4h, v21.4s\n"
+ "str d16, [x27, #0x10]\n"
+ "add x27, x27, #0x18\n"
+ ".inst 0x0ea16a90 // bfcvtn v16.4h, v20.4s\n"
+ "str d17, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "str d16, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "bge 84b\n"
+ "85:" // Accumulate: Height 4: no full blocks
+ "cbz x10, 87f\n"
+ "mov x20, %x[in_ptr]\n"
+ "86:" // Accumulate: Height 4: Single loop
+ "ldr h19, [x9, #0x0]\n"
+ "ldr h18, [x27, #0x0]\n"
+ "subs x10, x10, #0x1\n"
+ "ldr h17, [x26, #0x0]\n"
+ "ldr h16, [x25, #0x0]\n"
+ "ldr s23, [%x[in_ptr], #0x0]\n"
+ "ldr s22, [%x[in_ptr], #0x30]\n"
+ "ldr s21, [%x[in_ptr], #0x60]\n"
+ "ldr s20, [%x[in_ptr], #0x90]\n"
+ "shll v19.4s, v19.4h, #0x10\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v23.4s, v23.4s, v19.4s\n"
+ "fadd v22.4s, v22.4s, v18.4s\n"
+ "fadd v21.4s, v21.4s, v17.4s\n"
+ "fadd v20.4s, v20.4s, v16.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ ".inst 0x0ea16af3 // bfcvtn v19.4h, v23.4s\n"
+ ".inst 0x0ea16ad2 // bfcvtn v18.4h, v22.4s\n"
+ ".inst 0x0ea16ab1 // bfcvtn v17.4h, v21.4s\n"
+ ".inst 0x0ea16a90 // bfcvtn v16.4h, v20.4s\n"
+ "str h19, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ "str h18, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "str h17, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h16, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "bne 86b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "87:" // Accumulate: Height 4: no oddments
+ "b 108f\n"
+ "88:" // Accumulate: Height 5
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "blt 90f\n"
+ "89:" // Accumulate: Height 5: Block loop
+ "ldr d20, [x9, #0x0]\n"
+ "ldr d19, [x27, #0x0]\n"
+ "sub x10, x10, #0xc\n"
+ "ldr d18, [x26, #0x0]\n"
+ "ldr d17, [x25, #0x0]\n"
+ "cmp x10, #0xc\n"
+ "ldr d16, [x24, #0x0]\n"
+ "ldr q3, [%x[in_ptr], #0x0]\n"
+ "ldr q2, [%x[in_ptr], #0x30]\n"
+ "ldr q1, [%x[in_ptr], #0x60]\n"
+ "shll v20.4s, v20.4h, #0x10\n"
+ "shll v19.4s, v19.4h, #0x10\n"
+ "ldr q0, [%x[in_ptr], #0x90]\n"
+ "ldr q31, [%x[in_ptr], #0xc0]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "ldr q30, [%x[in_ptr], #0x10]\n"
+ "ldr q29, [%x[in_ptr], #0x40]\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "ldr q28, [%x[in_ptr], #0x70]\n"
+ "ldr q27, [%x[in_ptr], #0xa0]\n"
+ "fadd v3.4s, v3.4s, v20.4s\n"
+ "fadd v2.4s, v2.4s, v19.4s\n"
+ "ldr q26, [%x[in_ptr], #0xd0]\n"
+ "ldr q25, [%x[in_ptr], #0x20]\n"
+ "fadd v1.4s, v1.4s, v18.4s\n"
+ "fadd v0.4s, v0.4s, v17.4s\n"
+ "ldr q24, [%x[in_ptr], #0x50]\n"
+ "ldr q23, [%x[in_ptr], #0x80]\n"
+ "fadd v31.4s, v31.4s, v16.4s\n"
+ "ldr q22, [%x[in_ptr], #0xb0]\n"
+ "ldr q21, [%x[in_ptr], #0xe0]\n"
+ "fmin v3.4s, v3.4s, v13.4s\n"
+ "fmin v2.4s, v2.4s, v13.4s\n"
+ "fmin v1.4s, v1.4s, v13.4s\n"
+ "fmin v0.4s, v0.4s, v13.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fmin v31.4s, v31.4s, v13.4s\n"
+ "fmax v3.4s, v3.4s, v12.4s\n"
+ "fmax v2.4s, v2.4s, v12.4s\n"
+ "fmax v1.4s, v1.4s, v12.4s\n"
+ "fmax v0.4s, v0.4s, v12.4s\n"
+ "fmax v31.4s, v31.4s, v12.4s\n"
+ ".inst 0x0ea16874 // bfcvtn v20.4h, v3.4s\n"
+ ".inst 0x0ea16853 // bfcvtn v19.4h, v2.4s\n"
+ ".inst 0x0ea16831 // bfcvtn v17.4h, v1.4s\n"
+ ".inst 0x0ea16810 // bfcvtn v16.4h, v0.4s\n"
+ ".inst 0x0ea16bf2 // bfcvtn v18.4h, v31.4s\n"
+ "str d20, [x9, #0x0]\n"
+ "str d19, [x27, #0x0]\n"
+ "str d17, [x26, #0x0]\n"
+ "ldr d17, [x9, #0x8]\n"
+ "str d16, [x25, #0x0]\n"
+ "ldr d16, [x27, #0x8]\n"
+ "str d18, [x24, #0x0]\n"
+ "ldr d18, [x26, #0x8]\n"
+ "shll v20.4s, v17.4h, #0x10\n"
+ "ldr d17, [x25, #0x8]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "ldr d16, [x24, #0x8]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "fadd v30.4s, v30.4s, v20.4s\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v29.4s, v29.4s, v19.4s\n"
+ "fadd v28.4s, v28.4s, v18.4s\n"
+ "fadd v27.4s, v27.4s, v17.4s\n"
+ "fmin v30.4s, v30.4s, v13.4s\n"
+ "fadd v26.4s, v26.4s, v16.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmax v30.4s, v30.4s, v12.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ ".inst 0x0ea16bd2 // bfcvtn v18.4h, v30.4s\n"
+ ".inst 0x0ea16bb3 // bfcvtn v19.4h, v29.4s\n"
+ ".inst 0x0ea16b91 // bfcvtn v17.4h, v28.4s\n"
+ ".inst 0x0ea16b70 // bfcvtn v16.4h, v27.4s\n"
+ "str d18, [x9, #0x8]\n"
+ ".inst 0x0ea16b52 // bfcvtn v18.4h, v26.4s\n"
+ "str d19, [x27, #0x8]\n"
+ "str d17, [x26, #0x8]\n"
+ "ldr d17, [x9, #0x10]\n"
+ "str d16, [x25, #0x8]\n"
+ "ldr d16, [x27, #0x10]\n"
+ "str d18, [x24, #0x8]\n"
+ "ldr d18, [x26, #0x10]\n"
+ "shll v20.4s, v17.4h, #0x10\n"
+ "ldr d17, [x25, #0x10]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "ldr d16, [x24, #0x10]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "fadd v25.4s, v25.4s, v20.4s\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v24.4s, v24.4s, v19.4s\n"
+ "fadd v23.4s, v23.4s, v18.4s\n"
+ "fadd v22.4s, v22.4s, v17.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fadd v21.4s, v21.4s, v16.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ ".inst 0x0ea16b30 // bfcvtn v16.4h, v25.4s\n"
+ ".inst 0x0ea16b13 // bfcvtn v19.4h, v24.4s\n"
+ ".inst 0x0ea16af2 // bfcvtn v18.4h, v23.4s\n"
+ ".inst 0x0ea16ad1 // bfcvtn v17.4h, v22.4s\n"
+ "str d16, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ ".inst 0x0ea16ab0 // bfcvtn v16.4h, v21.4s\n"
+ "str d19, [x27, #0x10]\n"
+ "add x27, x27, #0x18\n"
+ "str d18, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "str d17, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "str d16, [x24, #0x10]\n"
+ "add x24, x24, #0x18\n"
+ "bge 89b\n"
+ "90:" // Accumulate: Height 5: no full blocks
+ "cbz x10, 92f\n"
+ "mov x20, %x[in_ptr]\n"
+ "91:" // Accumulate: Height 5: Single loop
+ "ldr h20, [x9, #0x0]\n"
+ "ldr h19, [x27, #0x0]\n"
+ "subs x10, x10, #0x1\n"
+ "ldr h18, [x26, #0x0]\n"
+ "ldr h17, [x25, #0x0]\n"
+ "ldr h16, [x24, #0x0]\n"
+ "ldr s25, [%x[in_ptr], #0x0]\n"
+ "ldr s24, [%x[in_ptr], #0x30]\n"
+ "ldr s23, [%x[in_ptr], #0x60]\n"
+ "shll v20.4s, v20.4h, #0x10\n"
+ "shll v19.4s, v19.4h, #0x10\n"
+ "ldr s22, [%x[in_ptr], #0x90]\n"
+ "ldr s21, [%x[in_ptr], #0xc0]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v25.4s, v25.4s, v20.4s\n"
+ "fadd v24.4s, v24.4s, v19.4s\n"
+ "fadd v23.4s, v23.4s, v18.4s\n"
+ "fadd v22.4s, v22.4s, v17.4s\n"
+ "fadd v21.4s, v21.4s, v16.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmin v21.4s, v21.4s, v13.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ "fmax v21.4s, v21.4s, v12.4s\n"
+ ".inst 0x0ea16b34 // bfcvtn v20.4h, v25.4s\n"
+ ".inst 0x0ea16b13 // bfcvtn v19.4h, v24.4s\n"
+ ".inst 0x0ea16af2 // bfcvtn v18.4h, v23.4s\n"
+ ".inst 0x0ea16ad1 // bfcvtn v17.4h, v22.4s\n"
+ ".inst 0x0ea16ab0 // bfcvtn v16.4h, v21.4s\n"
+ "str h20, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ "str h19, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "str h18, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h17, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "str h16, [x24, #0x0]\n"
+ "add x24, x24, #0x2\n"
+ "bne 91b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "92:" // Accumulate: Height 5: no oddments
+ "b 108f\n"
+ "93:" // Accumulate: Height 6
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "blt 95f\n"
+ "94:" // Accumulate: Height 6: Block loop
+ "ldr d21, [x9, #0x0]\n"
+ "ldr d20, [x27, #0x0]\n"
+ "sub x10, x10, #0xc\n"
+ "ldr d19, [x26, #0x0]\n"
+ "ldr d18, [x25, #0x0]\n"
+ "cmp x10, #0xc\n"
+ "ldr d17, [x24, #0x0]\n"
+ "ldr d16, [x23, #0x0]\n"
+ "ldr q6, [%x[in_ptr], #0x0]\n"
+ "ldr q5, [%x[in_ptr], #0x30]\n"
+ "shll v22.4s, v21.4h, #0x10\n"
+ "shll v21.4s, v20.4h, #0x10\n"
+ "ldr q4, [%x[in_ptr], #0x60]\n"
+ "ldr q3, [%x[in_ptr], #0x90]\n"
+ "shll v20.4s, v19.4h, #0x10\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "ldr q2, [%x[in_ptr], #0xc0]\n"
+ "ldr q19, [%x[in_ptr], #0xf0]\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "ldr q1, [%x[in_ptr], #0x10]\n"
+ "ldr q0, [%x[in_ptr], #0x40]\n"
+ "fadd v6.4s, v6.4s, v22.4s\n"
+ "fadd v5.4s, v5.4s, v21.4s\n"
+ "ldr q31, [%x[in_ptr], #0x70]\n"
+ "ldr q30, [%x[in_ptr], #0xa0]\n"
+ "fadd v4.4s, v4.4s, v20.4s\n"
+ "fadd v3.4s, v3.4s, v18.4s\n"
+ "ldr q29, [%x[in_ptr], #0xd0]\n"
+ "ldr q28, [%x[in_ptr], #0x100]\n"
+ "fadd v2.4s, v2.4s, v17.4s\n"
+ "fadd v19.4s, v19.4s, v16.4s\n"
+ "ldr q27, [%x[in_ptr], #0x20]\n"
+ "ldr q26, [%x[in_ptr], #0x50]\n"
+ "fmin v6.4s, v6.4s, v13.4s\n"
+ "fmin v5.4s, v5.4s, v13.4s\n"
+ "ldr q25, [%x[in_ptr], #0x80]\n"
+ "ldr q24, [%x[in_ptr], #0xb0]\n"
+ "fmin v4.4s, v4.4s, v13.4s\n"
+ "fmin v3.4s, v3.4s, v13.4s\n"
+ "ldr q23, [%x[in_ptr], #0xe0]\n"
+ "ldr q22, [%x[in_ptr], #0x110]\n"
+ "fmin v2.4s, v2.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmax v6.4s, v6.4s, v12.4s\n"
+ "fmax v5.4s, v5.4s, v12.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fmax v4.4s, v4.4s, v12.4s\n"
+ "fmax v3.4s, v3.4s, v12.4s\n"
+ "fmax v2.4s, v2.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ ".inst 0x0ea168d5 // bfcvtn v21.4h, v6.4s\n"
+ ".inst 0x0ea168b4 // bfcvtn v20.4h, v5.4s\n"
+ ".inst 0x0ea16892 // bfcvtn v18.4h, v4.4s\n"
+ ".inst 0x0ea16871 // bfcvtn v17.4h, v3.4s\n"
+ ".inst 0x0ea16850 // bfcvtn v16.4h, v2.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ "str d21, [x9, #0x0]\n"
+ "str d20, [x27, #0x0]\n"
+ "str d18, [x26, #0x0]\n"
+ "ldr d18, [x9, #0x8]\n"
+ "str d17, [x25, #0x0]\n"
+ "ldr d17, [x27, #0x8]\n"
+ "str d16, [x24, #0x0]\n"
+ "ldr d16, [x26, #0x8]\n"
+ "str d19, [x23, #0x0]\n"
+ "shll v21.4s, v18.4h, #0x10\n"
+ "ldr d18, [x25, #0x8]\n"
+ "shll v20.4s, v17.4h, #0x10\n"
+ "ldr d17, [x24, #0x8]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "ldr d16, [x23, #0x8]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "fadd v1.4s, v1.4s, v21.4s\n"
+ "fadd v0.4s, v0.4s, v20.4s\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v31.4s, v31.4s, v19.4s\n"
+ "fadd v30.4s, v30.4s, v18.4s\n"
+ "fmin v1.4s, v1.4s, v13.4s\n"
+ "fmin v0.4s, v0.4s, v13.4s\n"
+ "fadd v29.4s, v29.4s, v17.4s\n"
+ "fadd v28.4s, v28.4s, v16.4s\n"
+ "fmin v31.4s, v31.4s, v13.4s\n"
+ "fmin v30.4s, v30.4s, v13.4s\n"
+ "fmax v1.4s, v1.4s, v12.4s\n"
+ "fmax v0.4s, v0.4s, v12.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fmax v31.4s, v31.4s, v12.4s\n"
+ "fmax v30.4s, v30.4s, v12.4s\n"
+ ".inst 0x0ea16832 // bfcvtn v18.4h, v1.4s\n"
+ ".inst 0x0ea16810 // bfcvtn v16.4h, v0.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ ".inst 0x0ea16bf4 // bfcvtn v20.4h, v31.4s\n"
+ ".inst 0x0ea16bd1 // bfcvtn v17.4h, v30.4s\n"
+ "str d18, [x9, #0x8]\n"
+ "str d16, [x27, #0x8]\n"
+ ".inst 0x0ea16bb3 // bfcvtn v19.4h, v29.4s\n"
+ ".inst 0x0ea16b92 // bfcvtn v18.4h, v28.4s\n"
+ "ldr d16, [x9, #0x10]\n"
+ "str d20, [x26, #0x8]\n"
+ "str d17, [x25, #0x8]\n"
+ "ldr d17, [x27, #0x10]\n"
+ "str d19, [x24, #0x8]\n"
+ "shll v21.4s, v16.4h, #0x10\n"
+ "ldr d16, [x26, #0x10]\n"
+ "str d18, [x23, #0x8]\n"
+ "ldr d18, [x25, #0x10]\n"
+ "shll v20.4s, v17.4h, #0x10\n"
+ "ldr d17, [x24, #0x10]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "fadd v27.4s, v27.4s, v21.4s\n"
+ "ldr d16, [x23, #0x10]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "fadd v26.4s, v26.4s, v20.4s\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v25.4s, v25.4s, v19.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fadd v24.4s, v24.4s, v18.4s\n"
+ "fadd v23.4s, v23.4s, v17.4s\n"
+ "fadd v22.4s, v22.4s, v16.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ ".inst 0x0ea16b71 // bfcvtn v17.4h, v27.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16b50 // bfcvtn v16.4h, v26.4s\n"
+ "str d17, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ ".inst 0x0ea16b33 // bfcvtn v19.4h, v25.4s\n"
+ ".inst 0x0ea16b12 // bfcvtn v18.4h, v24.4s\n"
+ ".inst 0x0ea16af1 // bfcvtn v17.4h, v23.4s\n"
+ "str d16, [x27, #0x10]\n"
+ "add x27, x27, #0x18\n"
+ ".inst 0x0ea16ad0 // bfcvtn v16.4h, v22.4s\n"
+ "str d19, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "str d18, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "str d17, [x24, #0x10]\n"
+ "add x24, x24, #0x18\n"
+ "str d16, [x23, #0x10]\n"
+ "add x23, x23, #0x18\n"
+ "bge 94b\n"
+ "95:" // Accumulate: Height 6: no full blocks
+ "cbz x10, 97f\n"
+ "mov x20, %x[in_ptr]\n"
+ "96:" // Accumulate: Height 6: Single loop
+ "ldr h21, [x9, #0x0]\n"
+ "ldr h20, [x27, #0x0]\n"
+ "subs x10, x10, #0x1\n"
+ "ldr h19, [x26, #0x0]\n"
+ "ldr h18, [x25, #0x0]\n"
+ "ldr h17, [x24, #0x0]\n"
+ "ldr h16, [x23, #0x0]\n"
+ "ldr s27, [%x[in_ptr], #0x0]\n"
+ "ldr s26, [%x[in_ptr], #0x30]\n"
+ "shll v21.4s, v21.4h, #0x10\n"
+ "shll v20.4s, v20.4h, #0x10\n"
+ "ldr s25, [%x[in_ptr], #0x60]\n"
+ "ldr s24, [%x[in_ptr], #0x90]\n"
+ "shll v19.4s, v19.4h, #0x10\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "ldr s23, [%x[in_ptr], #0xc0]\n"
+ "ldr s22, [%x[in_ptr], #0xf0]\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v27.4s, v27.4s, v21.4s\n"
+ "fadd v26.4s, v26.4s, v20.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v25.4s, v25.4s, v19.4s\n"
+ "fadd v24.4s, v24.4s, v18.4s\n"
+ "fadd v23.4s, v23.4s, v17.4s\n"
+ "fadd v22.4s, v22.4s, v16.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16b75 // bfcvtn v21.4h, v27.4s\n"
+ ".inst 0x0ea16b54 // bfcvtn v20.4h, v26.4s\n"
+ ".inst 0x0ea16b33 // bfcvtn v19.4h, v25.4s\n"
+ ".inst 0x0ea16b12 // bfcvtn v18.4h, v24.4s\n"
+ ".inst 0x0ea16af1 // bfcvtn v17.4h, v23.4s\n"
+ ".inst 0x0ea16ad0 // bfcvtn v16.4h, v22.4s\n"
+ "str h21, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ "str h20, [x27, #0x0]\n"
+ "add x27, x27, #0x2\n"
+ "str h19, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h18, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "str h17, [x24, #0x0]\n"
+ "add x24, x24, #0x2\n"
+ "str h16, [x23, #0x0]\n"
+ "add x23, x23, #0x2\n"
+ "bne 96b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "97:" // Accumulate: Height 6: no oddments
+ "b 108f\n"
+ "98:" // Accumulate: Height 7
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "add x22, x23, %x[ldout], LSL #1\n"
+ "blt 100f\n"
+ "99:" // Accumulate: Height 7: Block loop
+ "ldr d22, [x9, #0x0]\n"
+ "ldr d21, [x27, #0x0]\n"
+ "sub x10, x10, #0xc\n"
+ "ldr d20, [x26, #0x0]\n"
+ "ldr d19, [x25, #0x0]\n"
+ "cmp x10, #0xc\n"
+ "ldr d18, [x24, #0x0]\n"
+ "ldr d17, [x23, #0x0]\n"
+ "ldr d16, [x22, #0x0]\n"
+ "ldr q9, [%x[in_ptr], #0x0]\n"
+ "shll v24.4s, v22.4h, #0x10\n"
+ "shll v23.4s, v21.4h, #0x10\n"
+ "ldr q8, [%x[in_ptr], #0x30]\n"
+ "ldr q7, [%x[in_ptr], #0x60]\n"
+ "shll v21.4s, v20.4h, #0x10\n"
+ "shll v19.4s, v19.4h, #0x10\n"
+ "ldr q6, [%x[in_ptr], #0x90]\n"
+ "ldr q5, [%x[in_ptr], #0xc0]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "ldr q20, [%x[in_ptr], #0xf0]\n"
+ "ldr q22, [%x[in_ptr], #0x120]\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v9.4s, v9.4s, v24.4s\n"
+ "ldr q4, [%x[in_ptr], #0x10]\n"
+ "ldr q3, [%x[in_ptr], #0x40]\n"
+ "fadd v8.4s, v8.4s, v23.4s\n"
+ "fadd v7.4s, v7.4s, v21.4s\n"
+ "ldr q2, [%x[in_ptr], #0x70]\n"
+ "ldr q1, [%x[in_ptr], #0xa0]\n"
+ "fadd v6.4s, v6.4s, v19.4s\n"
+ "fadd v5.4s, v5.4s, v18.4s\n"
+ "ldr q0, [%x[in_ptr], #0xd0]\n"
+ "ldr q31, [%x[in_ptr], #0x100]\n"
+ "fadd v20.4s, v20.4s, v17.4s\n"
+ "fadd v22.4s, v22.4s, v16.4s\n"
+ "ldr q30, [%x[in_ptr], #0x130]\n"
+ "ldr q29, [%x[in_ptr], #0x20]\n"
+ "fmin v9.4s, v9.4s, v13.4s\n"
+ "fmin v8.4s, v8.4s, v13.4s\n"
+ "ldr q28, [%x[in_ptr], #0x50]\n"
+ "ldr q27, [%x[in_ptr], #0x80]\n"
+ "fmin v7.4s, v7.4s, v13.4s\n"
+ "fmin v6.4s, v6.4s, v13.4s\n"
+ "ldr q26, [%x[in_ptr], #0xb0]\n"
+ "ldr q25, [%x[in_ptr], #0xe0]\n"
+ "fmin v5.4s, v5.4s, v13.4s\n"
+ "fmin v20.4s, v20.4s, v13.4s\n"
+ "ldr q24, [%x[in_ptr], #0x110]\n"
+ "ldr q23, [%x[in_ptr], #0x140]\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v9.4s, v9.4s, v12.4s\n"
+ "fmax v8.4s, v8.4s, v12.4s\n"
+ "fmax v7.4s, v7.4s, v12.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fmax v6.4s, v6.4s, v12.4s\n"
+ "fmax v5.4s, v5.4s, v12.4s\n"
+ "fmax v20.4s, v20.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16935 // bfcvtn v21.4h, v9.4s\n"
+ ".inst 0x0ea16913 // bfcvtn v19.4h, v8.4s\n"
+ ".inst 0x0ea168f0 // bfcvtn v16.4h, v7.4s\n"
+ ".inst 0x0ea168d2 // bfcvtn v18.4h, v6.4s\n"
+ ".inst 0x0ea168b1 // bfcvtn v17.4h, v5.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ "str d21, [x9, #0x0]\n"
+ "str d19, [x27, #0x0]\n"
+ ".inst 0x0ea16ad3 // bfcvtn v19.4h, v22.4s\n"
+ "str d16, [x26, #0x0]\n"
+ "ldr d16, [x9, #0x8]\n"
+ "str d18, [x25, #0x0]\n"
+ "ldr d18, [x27, #0x8]\n"
+ "str d17, [x24, #0x0]\n"
+ "ldr d17, [x26, #0x8]\n"
+ "str d20, [x23, #0x0]\n"
+ "shll v22.4s, v16.4h, #0x10\n"
+ "ldr d16, [x25, #0x8]\n"
+ "str d19, [x22, #0x0]\n"
+ "shll v21.4s, v18.4h, #0x10\n"
+ "ldr d18, [x24, #0x8]\n"
+ "shll v20.4s, v17.4h, #0x10\n"
+ "ldr d17, [x23, #0x8]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "fadd v4.4s, v4.4s, v22.4s\n"
+ "ldr d16, [x22, #0x8]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "fadd v3.4s, v3.4s, v21.4s\n"
+ "fadd v2.4s, v2.4s, v20.4s\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v1.4s, v1.4s, v19.4s\n"
+ "fadd v0.4s, v0.4s, v18.4s\n"
+ "fmin v4.4s, v4.4s, v13.4s\n"
+ "fadd v31.4s, v31.4s, v17.4s\n"
+ "fmin v3.4s, v3.4s, v13.4s\n"
+ "fadd v30.4s, v30.4s, v16.4s\n"
+ "fmin v2.4s, v2.4s, v13.4s\n"
+ "fmin v1.4s, v1.4s, v13.4s\n"
+ "fmin v0.4s, v0.4s, v13.4s\n"
+ "fmin v31.4s, v31.4s, v13.4s\n"
+ "fmax v4.4s, v4.4s, v12.4s\n"
+ "fmin v30.4s, v30.4s, v13.4s\n"
+ "fmax v3.4s, v3.4s, v12.4s\n"
+ "fmax v2.4s, v2.4s, v12.4s\n"
+ "fmax v1.4s, v1.4s, v12.4s\n"
+ "fmax v0.4s, v0.4s, v12.4s\n"
+ "fmax v31.4s, v31.4s, v12.4s\n"
+ "fmax v30.4s, v30.4s, v12.4s\n"
+ ".inst 0x0ea16893 // bfcvtn v19.4h, v4.4s\n"
+ ".inst 0x0ea16875 // bfcvtn v21.4h, v3.4s\n"
+ ".inst 0x0ea16850 // bfcvtn v16.4h, v2.4s\n"
+ ".inst 0x0ea16832 // bfcvtn v18.4h, v1.4s\n"
+ ".inst 0x0ea16811 // bfcvtn v17.4h, v0.4s\n"
+ "str d19, [x9, #0x8]\n"
+ ".inst 0x0ea16bf4 // bfcvtn v20.4h, v31.4s\n"
+ ".inst 0x0ea16bd3 // bfcvtn v19.4h, v30.4s\n"
+ "str d21, [x27, #0x8]\n"
+ "str d16, [x26, #0x8]\n"
+ "ldr d16, [x9, #0x10]\n"
+ "str d18, [x25, #0x8]\n"
+ "ldr d18, [x27, #0x10]\n"
+ "str d17, [x24, #0x8]\n"
+ "ldr d17, [x26, #0x10]\n"
+ "str d20, [x23, #0x8]\n"
+ "shll v22.4s, v16.4h, #0x10\n"
+ "ldr d16, [x25, #0x10]\n"
+ "str d19, [x22, #0x8]\n"
+ "shll v21.4s, v18.4h, #0x10\n"
+ "ldr d18, [x24, #0x10]\n"
+ "shll v20.4s, v17.4h, #0x10\n"
+ "ldr d17, [x23, #0x10]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "fadd v29.4s, v29.4s, v22.4s\n"
+ "ldr d16, [x22, #0x10]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "fadd v28.4s, v28.4s, v21.4s\n"
+ "fadd v27.4s, v27.4s, v20.4s\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v26.4s, v26.4s, v19.4s\n"
+ "fadd v25.4s, v25.4s, v18.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fadd v24.4s, v24.4s, v17.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fadd v23.4s, v23.4s, v16.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ ".inst 0x0ea16bb0 // bfcvtn v16.4h, v29.4s\n"
+ ".inst 0x0ea16b95 // bfcvtn v21.4h, v28.4s\n"
+ ".inst 0x0ea16b74 // bfcvtn v20.4h, v27.4s\n"
+ ".inst 0x0ea16b53 // bfcvtn v19.4h, v26.4s\n"
+ ".inst 0x0ea16b32 // bfcvtn v18.4h, v25.4s\n"
+ "str d16, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ ".inst 0x0ea16b11 // bfcvtn v17.4h, v24.4s\n"
+ ".inst 0x0ea16af0 // bfcvtn v16.4h, v23.4s\n"
+ "str d21, [x27, #0x10]\n"
+ "add x27, x27, #0x18\n"
+ "str d20, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "str d19, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "str d18, [x24, #0x10]\n"
+ "add x24, x24, #0x18\n"
+ "str d17, [x23, #0x10]\n"
+ "add x23, x23, #0x18\n"
+ "str d16, [x22, #0x10]\n"
+ "add x22, x22, #0x18\n"
+ "bge 99b\n"
+ "100:" // Accumulate: Height 7: no full blocks
+ "cbz x10, 102f\n"
+ "mov x20, %x[in_ptr]\n"
+ "101:" // Accumulate: Height 7: Single loop
+ "ldr h22, [x9, #0x0]\n"
+ "ldr h21, [x27, #0x0]\n"
+ "subs x10, x10, #0x1\n"
+ "ldr h20, [x26, #0x0]\n"
+ "ldr h19, [x25, #0x0]\n"
+ "ldr h18, [x24, #0x0]\n"
+ "ldr h17, [x23, #0x0]\n"
+ "ldr h16, [x22, #0x0]\n"
+ "ldr s29, [%x[in_ptr], #0x0]\n"
+ "shll v28.4s, v22.4h, #0x10\n"
+ "shll v27.4s, v21.4h, #0x10\n"
+ "ldr s26, [%x[in_ptr], #0x30]\n"
+ "ldr s25, [%x[in_ptr], #0x60]\n"
+ "shll v21.4s, v20.4h, #0x10\n"
+ "shll v20.4s, v19.4h, #0x10\n"
+ "ldr s24, [%x[in_ptr], #0x90]\n"
+ "ldr s23, [%x[in_ptr], #0xc0]\n"
+ "shll v19.4s, v18.4h, #0x10\n"
+ "shll v18.4s, v17.4h, #0x10\n"
+ "ldr s17, [%x[in_ptr], #0xf0]\n"
+ "ldr s22, [%x[in_ptr], #0x120]\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v29.4s, v29.4s, v28.4s\n"
+ "fadd v26.4s, v26.4s, v27.4s\n"
+ "fadd v25.4s, v25.4s, v21.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v24.4s, v24.4s, v20.4s\n"
+ "fadd v23.4s, v23.4s, v19.4s\n"
+ "fadd v17.4s, v17.4s, v18.4s\n"
+ "fadd v22.4s, v22.4s, v16.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v17.4s, v17.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v17.4s, v17.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16bb5 // bfcvtn v21.4h, v29.4s\n"
+ ".inst 0x0ea16b50 // bfcvtn v16.4h, v26.4s\n"
+ ".inst 0x0ea16b34 // bfcvtn v20.4h, v25.4s\n"
+ ".inst 0x0ea16b13 // bfcvtn v19.4h, v24.4s\n"
+ ".inst 0x0ea16af2 // bfcvtn v18.4h, v23.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ "str h21, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ "str h16, [x27, #0x0]\n"
+ ".inst 0x0ea16ad0 // bfcvtn v16.4h, v22.4s\n"
+ "add x27, x27, #0x2\n"
+ "str h20, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h19, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "str h18, [x24, #0x0]\n"
+ "add x24, x24, #0x2\n"
+ "str h17, [x23, #0x0]\n"
+ "add x23, x23, #0x2\n"
+ "str h16, [x22, #0x0]\n"
+ "add x22, x22, #0x2\n"
+ "bne 101b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "102:" // Accumulate: Height 7: no oddments
+ "b 108f\n"
+ "103:" // Accumulate: Height 8
+ "mov x9, %x[out_ptr]\n"
+ "mov x10, %x[cols]\n"
+ "add x27, x9, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "cmp x10, #0xc\n"
+ "add x22, x23, %x[ldout], LSL #1\n"
+ "add x21, x22, %x[ldout], LSL #1\n"
+ "blt 105f\n"
+ "104:" // Accumulate: Height 8: Block loop
+ "ldr d23, [x9, #0x0]\n"
+ "ldr d22, [x27, #0x0]\n"
+ "sub x10, x10, #0xc\n"
+ "ldr d21, [x26, #0x0]\n"
+ "ldr d20, [x25, #0x0]\n"
+ "cmp x10, #0xc\n"
+ "ldr d19, [x24, #0x0]\n"
+ "ldr d18, [x23, #0x0]\n"
+ "ldr d17, [x22, #0x0]\n"
+ "ldr d16, [x21, #0x0]\n"
+ "shll v26.4s, v23.4h, #0x10\n"
+ "shll v25.4s, v22.4h, #0x10\n"
+ "ldr q11, [%x[in_ptr], #0x0]\n"
+ "ldr q10, [%x[in_ptr], #0x30]\n"
+ "shll v24.4s, v21.4h, #0x10\n"
+ "shll v23.4s, v20.4h, #0x10\n"
+ "ldr q9, [%x[in_ptr], #0x60]\n"
+ "ldr q8, [%x[in_ptr], #0x90]\n"
+ "shll v21.4s, v19.4h, #0x10\n"
+ "shll v20.4s, v18.4h, #0x10\n"
+ "ldr q18, [%x[in_ptr], #0xc0]\n"
+ "ldr q19, [%x[in_ptr], #0xf0]\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "ldr q7, [%x[in_ptr], #0x120]\n"
+ "ldr q22, [%x[in_ptr], #0x150]\n"
+ "fadd v11.4s, v11.4s, v26.4s\n"
+ "fadd v10.4s, v10.4s, v25.4s\n"
+ "ldr q6, [%x[in_ptr], #0x10]\n"
+ "ldr q5, [%x[in_ptr], #0x40]\n"
+ "fadd v9.4s, v9.4s, v24.4s\n"
+ "fadd v8.4s, v8.4s, v23.4s\n"
+ "ldr q4, [%x[in_ptr], #0x70]\n"
+ "ldr q3, [%x[in_ptr], #0xa0]\n"
+ "fadd v18.4s, v18.4s, v21.4s\n"
+ "fadd v19.4s, v19.4s, v20.4s\n"
+ "ldr q2, [%x[in_ptr], #0xd0]\n"
+ "ldr q1, [%x[in_ptr], #0x100]\n"
+ "fadd v7.4s, v7.4s, v17.4s\n"
+ "fadd v22.4s, v22.4s, v16.4s\n"
+ "ldr q0, [%x[in_ptr], #0x130]\n"
+ "ldr q31, [%x[in_ptr], #0x160]\n"
+ "fmin v11.4s, v11.4s, v13.4s\n"
+ "fmin v10.4s, v10.4s, v13.4s\n"
+ "ldr q30, [%x[in_ptr], #0x20]\n"
+ "ldr q29, [%x[in_ptr], #0x50]\n"
+ "fmin v9.4s, v9.4s, v13.4s\n"
+ "fmin v8.4s, v8.4s, v13.4s\n"
+ "ldr q28, [%x[in_ptr], #0x80]\n"
+ "ldr q27, [%x[in_ptr], #0xb0]\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "ldr q26, [%x[in_ptr], #0xe0]\n"
+ "ldr q25, [%x[in_ptr], #0x110]\n"
+ "fmin v7.4s, v7.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "ldr q24, [%x[in_ptr], #0x140]\n"
+ "ldr q23, [%x[in_ptr], #0x170]\n"
+ "fmax v11.4s, v11.4s, v12.4s\n"
+ "fmax v10.4s, v10.4s, v12.4s\n"
+ "fmax v9.4s, v9.4s, v12.4s\n"
+ "fmax v8.4s, v8.4s, v12.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x180\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v7.4s, v7.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16975 // bfcvtn v21.4h, v11.4s\n"
+ ".inst 0x0ea16954 // bfcvtn v20.4h, v10.4s\n"
+ ".inst 0x0ea16931 // bfcvtn v17.4h, v9.4s\n"
+ ".inst 0x0ea16910 // bfcvtn v16.4h, v8.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ "str d21, [x9, #0x0]\n"
+ "str d20, [x27, #0x0]\n"
+ ".inst 0x0ea168f5 // bfcvtn v21.4h, v7.4s\n"
+ ".inst 0x0ea16ad4 // bfcvtn v20.4h, v22.4s\n"
+ "str d17, [x26, #0x0]\n"
+ "ldr d17, [x9, #0x8]\n"
+ "str d16, [x25, #0x0]\n"
+ "ldr d16, [x27, #0x8]\n"
+ "str d18, [x24, #0x0]\n"
+ "ldr d18, [x26, #0x8]\n"
+ "str d19, [x23, #0x0]\n"
+ "shll v19.4s, v17.4h, #0x10\n"
+ "ldr d17, [x25, #0x8]\n"
+ "str d21, [x22, #0x0]\n"
+ "shll v22.4s, v16.4h, #0x10\n"
+ "ldr d16, [x24, #0x8]\n"
+ "str d20, [x21, #0x0]\n"
+ "shll v21.4s, v18.4h, #0x10\n"
+ "ldr d18, [x23, #0x8]\n"
+ "shll v20.4s, v17.4h, #0x10\n"
+ "fadd v6.4s, v6.4s, v19.4s\n"
+ "ldr d17, [x22, #0x8]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "fadd v5.4s, v5.4s, v22.4s\n"
+ "ldr d16, [x21, #0x8]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "fadd v4.4s, v4.4s, v21.4s\n"
+ "fadd v3.4s, v3.4s, v20.4s\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "fadd v2.4s, v2.4s, v19.4s\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v1.4s, v1.4s, v18.4s\n"
+ "fmin v6.4s, v6.4s, v13.4s\n"
+ "fmin v5.4s, v5.4s, v13.4s\n"
+ "fadd v0.4s, v0.4s, v17.4s\n"
+ "fmin v4.4s, v4.4s, v13.4s\n"
+ "fadd v31.4s, v31.4s, v16.4s\n"
+ "fmin v3.4s, v3.4s, v13.4s\n"
+ "fmin v2.4s, v2.4s, v13.4s\n"
+ "fmin v1.4s, v1.4s, v13.4s\n"
+ "fmin v0.4s, v0.4s, v13.4s\n"
+ "fmax v6.4s, v6.4s, v12.4s\n"
+ "fmin v31.4s, v31.4s, v13.4s\n"
+ "fmax v5.4s, v5.4s, v12.4s\n"
+ "fmax v4.4s, v4.4s, v12.4s\n"
+ "fmax v3.4s, v3.4s, v12.4s\n"
+ "fmax v2.4s, v2.4s, v12.4s\n"
+ "fmax v1.4s, v1.4s, v12.4s\n"
+ "fmax v0.4s, v0.4s, v12.4s\n"
+ "fmax v31.4s, v31.4s, v12.4s\n"
+ ".inst 0x0ea168d5 // bfcvtn v21.4h, v6.4s\n"
+ ".inst 0x0ea168b4 // bfcvtn v20.4h, v5.4s\n"
+ ".inst 0x0ea16891 // bfcvtn v17.4h, v4.4s\n"
+ ".inst 0x0ea16870 // bfcvtn v16.4h, v3.4s\n"
+ ".inst 0x0ea16852 // bfcvtn v18.4h, v2.4s\n"
+ ".inst 0x0ea16833 // bfcvtn v19.4h, v1.4s\n"
+ "str d21, [x9, #0x8]\n"
+ "str d20, [x27, #0x8]\n"
+ ".inst 0x0ea16815 // bfcvtn v21.4h, v0.4s\n"
+ ".inst 0x0ea16bf4 // bfcvtn v20.4h, v31.4s\n"
+ "str d17, [x26, #0x8]\n"
+ "ldr d17, [x9, #0x10]\n"
+ "str d16, [x25, #0x8]\n"
+ "ldr d16, [x27, #0x10]\n"
+ "str d18, [x24, #0x8]\n"
+ "ldr d18, [x26, #0x10]\n"
+ "str d19, [x23, #0x8]\n"
+ "shll v19.4s, v17.4h, #0x10\n"
+ "ldr d17, [x25, #0x10]\n"
+ "str d21, [x22, #0x8]\n"
+ "shll v22.4s, v16.4h, #0x10\n"
+ "ldr d16, [x24, #0x10]\n"
+ "str d20, [x21, #0x8]\n"
+ "shll v21.4s, v18.4h, #0x10\n"
+ "ldr d18, [x23, #0x10]\n"
+ "shll v20.4s, v17.4h, #0x10\n"
+ "fadd v30.4s, v30.4s, v19.4s\n"
+ "ldr d17, [x22, #0x10]\n"
+ "shll v19.4s, v16.4h, #0x10\n"
+ "fadd v29.4s, v29.4s, v22.4s\n"
+ "ldr d16, [x21, #0x10]\n"
+ "shll v18.4s, v18.4h, #0x10\n"
+ "fadd v28.4s, v28.4s, v21.4s\n"
+ "fadd v27.4s, v27.4s, v20.4s\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "fadd v26.4s, v26.4s, v19.4s\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "fadd v25.4s, v25.4s, v18.4s\n"
+ "fmin v30.4s, v30.4s, v13.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fadd v24.4s, v24.4s, v17.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fadd v23.4s, v23.4s, v16.4s\n"
+ "fmin v27.4s, v27.4s, v13.4s\n"
+ "fmin v26.4s, v26.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmax v30.4s, v30.4s, v12.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ "fmax v27.4s, v27.4s, v12.4s\n"
+ "fmax v26.4s, v26.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ ".inst 0x0ea16bd1 // bfcvtn v17.4h, v30.4s\n"
+ ".inst 0x0ea16bb0 // bfcvtn v16.4h, v29.4s\n"
+ ".inst 0x0ea16b95 // bfcvtn v21.4h, v28.4s\n"
+ ".inst 0x0ea16b74 // bfcvtn v20.4h, v27.4s\n"
+ ".inst 0x0ea16b53 // bfcvtn v19.4h, v26.4s\n"
+ ".inst 0x0ea16b32 // bfcvtn v18.4h, v25.4s\n"
+ "str d17, [x9, #0x10]\n"
+ "add x9, x9, #0x18\n"
+ "str d16, [x27, #0x10]\n"
+ ".inst 0x0ea16b11 // bfcvtn v17.4h, v24.4s\n"
+ ".inst 0x0ea16af0 // bfcvtn v16.4h, v23.4s\n"
+ "add x27, x27, #0x18\n"
+ "str d21, [x26, #0x10]\n"
+ "add x26, x26, #0x18\n"
+ "str d20, [x25, #0x10]\n"
+ "add x25, x25, #0x18\n"
+ "str d19, [x24, #0x10]\n"
+ "add x24, x24, #0x18\n"
+ "str d18, [x23, #0x10]\n"
+ "add x23, x23, #0x18\n"
+ "str d17, [x22, #0x10]\n"
+ "add x22, x22, #0x18\n"
+ "str d16, [x21, #0x10]\n"
+ "add x21, x21, #0x18\n"
+ "bge 104b\n"
+ "105:" // Accumulate: Height 8: no full blocks
+ "cbz x10, 107f\n"
+ "mov x20, %x[in_ptr]\n"
+ "106:" // Accumulate: Height 8: Single loop
+ "ldr h23, [x9, #0x0]\n"
+ "ldr h22, [x27, #0x0]\n"
+ "subs x10, x10, #0x1\n"
+ "ldr h21, [x26, #0x0]\n"
+ "ldr h20, [x25, #0x0]\n"
+ "ldr h19, [x24, #0x0]\n"
+ "ldr h18, [x23, #0x0]\n"
+ "ldr h17, [x22, #0x0]\n"
+ "ldr h16, [x21, #0x0]\n"
+ "shll v31.4s, v23.4h, #0x10\n"
+ "shll v30.4s, v22.4h, #0x10\n"
+ "ldr s29, [%x[in_ptr], #0x0]\n"
+ "ldr s28, [%x[in_ptr], #0x30]\n"
+ "shll v27.4s, v21.4h, #0x10\n"
+ "shll v26.4s, v20.4h, #0x10\n"
+ "ldr s25, [%x[in_ptr], #0x60]\n"
+ "ldr s24, [%x[in_ptr], #0x90]\n"
+ "shll v21.4s, v19.4h, #0x10\n"
+ "shll v20.4s, v18.4h, #0x10\n"
+ "ldr s19, [%x[in_ptr], #0xc0]\n"
+ "ldr s18, [%x[in_ptr], #0xf0]\n"
+ "shll v17.4s, v17.4h, #0x10\n"
+ "shll v16.4s, v16.4h, #0x10\n"
+ "ldr s23, [%x[in_ptr], #0x120]\n"
+ "ldr s22, [%x[in_ptr], #0x150]\n"
+ "fadd v29.4s, v29.4s, v31.4s\n"
+ "fadd v28.4s, v28.4s, v30.4s\n"
+ "fadd v25.4s, v25.4s, v27.4s\n"
+ "fadd v24.4s, v24.4s, v26.4s\n"
+ "add %x[in_ptr], %x[in_ptr], #0x4\n"
+ "fadd v19.4s, v19.4s, v21.4s\n"
+ "fadd v18.4s, v18.4s, v20.4s\n"
+ "fadd v23.4s, v23.4s, v17.4s\n"
+ "fadd v22.4s, v22.4s, v16.4s\n"
+ "fmin v29.4s, v29.4s, v13.4s\n"
+ "fmin v28.4s, v28.4s, v13.4s\n"
+ "fmin v25.4s, v25.4s, v13.4s\n"
+ "fmin v24.4s, v24.4s, v13.4s\n"
+ "fmin v19.4s, v19.4s, v13.4s\n"
+ "fmin v18.4s, v18.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v13.4s\n"
+ "fmin v22.4s, v22.4s, v13.4s\n"
+ "fmax v29.4s, v29.4s, v12.4s\n"
+ "fmax v28.4s, v28.4s, v12.4s\n"
+ "fmax v25.4s, v25.4s, v12.4s\n"
+ "fmax v24.4s, v24.4s, v12.4s\n"
+ "fmax v19.4s, v19.4s, v12.4s\n"
+ "fmax v18.4s, v18.4s, v12.4s\n"
+ "fmax v23.4s, v23.4s, v12.4s\n"
+ "fmax v22.4s, v22.4s, v12.4s\n"
+ ".inst 0x0ea16bb1 // bfcvtn v17.4h, v29.4s\n"
+ ".inst 0x0ea16b90 // bfcvtn v16.4h, v28.4s\n"
+ ".inst 0x0ea16b35 // bfcvtn v21.4h, v25.4s\n"
+ ".inst 0x0ea16b14 // bfcvtn v20.4h, v24.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ "str h17, [x9, #0x0]\n"
+ "add x9, x9, #0x2\n"
+ "str h16, [x27, #0x0]\n"
+ ".inst 0x0ea16af1 // bfcvtn v17.4h, v23.4s\n"
+ ".inst 0x0ea16ad0 // bfcvtn v16.4h, v22.4s\n"
+ "add x27, x27, #0x2\n"
+ "str h21, [x26, #0x0]\n"
+ "add x26, x26, #0x2\n"
+ "str h20, [x25, #0x0]\n"
+ "add x25, x25, #0x2\n"
+ "str h19, [x24, #0x0]\n"
+ "add x24, x24, #0x2\n"
+ "str h18, [x23, #0x0]\n"
+ "add x23, x23, #0x2\n"
+ "str h17, [x22, #0x0]\n"
+ "add x22, x22, #0x2\n"
+ "str h16, [x21, #0x0]\n"
+ "add x21, x21, #0x2\n"
+ "bne 106b\n"
+ "add %x[in_ptr], x20, #0x180\n"
+ "107:" // Accumulate: Height 8: no oddments
+ "subs %x[rows], %x[rows], #0x8\n"
+ "add %x[out_ptr], %x[out_ptr], x11\n"
+ "bgt 67b\n"
+ "108:" // Exit
+ : [in_ptr] "+&r" (in_ptr), [out_ptr] "+&r" (out_ptr), [rows] "+&r" (rows)
+ : [accumulate] "r" (accumulate), [bias] "r" (bias), [cols] "r" (cols), [ldout] "r" (ldout), [maxval] "r" (maxval), [minval] "r" (minval)
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+ );
+}
+
+
+#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/merges/list-sve.hpp b/src/core/NEON/kernels/arm_gemm/merges/list-sve.hpp
index aded4b3..d11740e 100644
--- a/src/core/NEON/kernels/arm_gemm/merges/list-sve.hpp
+++ b/src/core/NEON/kernels/arm_gemm/merges/list-sve.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021, 2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -24,5 +24,6 @@
#include "sve_merge_fp16_3VLx8.hpp"
#include "sve_merge_fp32_3VLx8.hpp"
+#include "sve_merge_fp32_bf16_8x3VL.hpp"
#include "sve_merge_s32_3VLx8.hpp"
#include "sve_merge_u32_3VLx8.hpp"
\ No newline at end of file
diff --git a/src/core/NEON/kernels/arm_gemm/merges/list.hpp b/src/core/NEON/kernels/arm_gemm/merges/list.hpp
index 3443c6f..fd6be5b 100644
--- a/src/core/NEON/kernels/arm_gemm/merges/list.hpp
+++ b/src/core/NEON/kernels/arm_gemm/merges/list.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2020, 2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -23,6 +23,7 @@
*/
#include "a32_merge_float_8x6.hpp"
#include "a64_merge_fp32_12x8.hpp"
+#include "a64_merge_fp32_bf16_8x12.hpp"
#include "a64_merge_s32_12x8.hpp"
#include "a64_merge_s32_4x4.hpp"
#include "a64_merge_u32_12x8.hpp"
diff --git a/src/core/NEON/kernels/arm_gemm/merges/sve_merge_fp32_bf16_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_fp32_bf16_8x3VL.hpp
new file mode 100644
index 0000000..5d4a8bf
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_fp32_bf16_8x3VL.hpp
@@ -0,0 +1,2137 @@
+/*
+ * 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 ARM_COMPUTE_ENABLE_SVE
+
+template<>
+void MergeResults<3, 8, true>(
+ bfloat16 *out_ptr,
+ const float * in_ptr,
+ const int ldout,
+ const int y0, const int ymax,
+ const int x0, const int xmax,
+ const bfloat16 *bias,
+ Activation act,
+ bool accumulate)
+{
+ float maxval = static_cast<float>(std::numeric_limits<float>::infinity());
+ float minval = - static_cast<float>(std::numeric_limits<float>::infinity());
+
+ switch(act.type) {
+ default:
+ case Activation::Type::None:
+ break;
+ case Activation::Type::BoundedReLU:
+ maxval = static_cast<float>(act.param1);
+ /* fall through */
+ case Activation::Type::ReLU:
+ minval = 0;
+ break;
+ }
+
+ size_t rows = ymax-y0;
+ size_t cols = xmax-x0;
+
+ out_ptr += (y0 * ldout) + x0;
+ bias = (bias == nullptr) ? nullptr : bias + x0;
+
+ __asm__ __volatile__(
+ "ptrue p3.b\n"
+ "cbz %x[cols], 52f\n"
+ "cbz %x[rows], 52f\n"
+ "mov x12, #0x20\n"
+ "dup z12.s, %w[maxval]\n"
+ "dup z11.s, %w[minval]\n"
+ "mul x12, %x[ldout], x12\n"
+ "cbnz %x[accumulate], 34f\n"
+ "1:" // Initial: Row loop
+ "cmp %x[rows], #0x7\n"
+ "bgt 30f\n"
+ "beq 26f\n"
+ "cmp %x[rows], #0x5\n"
+ "bgt 22f\n"
+ "beq 18f\n"
+ "cmp %x[rows], #0x3\n"
+ "bgt 14f\n"
+ "beq 10f\n"
+ "cmp %x[rows], #0x1\n"
+ "bgt 6f\n"
+ "2:" // Initial: Height 1
+ "mov x11, %x[cols]\n"
+ "mov x10, %x[out_ptr]\n"
+ "mov x9, %x[bias]\n"
+ "3:" // Initial: Height 1: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p0.s, x21, x11\n"
+ "incw x21\n"
+ "cbnz %x[bias], 4f\n"
+ "mov z21.b, #0x0\n"
+ "mov z20.b, #0x0\n"
+ "mov z19.b, #0x0\n"
+ "b 5f\n"
+ "4:" // Initial: Height 1: Width 3: bias
+ "ld1h { z18.s }, p2/Z, [x9]\n"
+ "ld1h { z17.s }, p1/Z, [x9, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x9, #2, MUL VL]\n"
+ "lsl z21.s, z18.s, #0x10\n"
+ "lsl z20.s, z17.s, #0x10\n"
+ "lsl z19.s, z16.s, #0x10\n"
+ "5:" // Initial: Height 1: Width 3: init done
+ "ld1w { z17.s }, p2/Z, [%x[in_ptr]]\n"
+ "ld1w { z16.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ "inch x9, ALL, MUL #3\n"
+ "ld1w { z18.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "fadd z17.s, z17.s, z21.s\n"
+ "fadd z16.s, z16.s, z20.s\n"
+ "cmp x11, XZR\n"
+ "fadd z18.s, z18.s, z19.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fmin z16.s, p3/M, z16.s, z12.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ "fmax z16.s, p3/M, z16.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ ".inst 0x658aae31 // bfcvt z17.h, p3/M, z17.s\n"
+ ".inst 0x658aae10 // bfcvt z16.h, p3/M, z16.s\n"
+ "st1h { z17.s }, p2, [x10]\n"
+ "st1h { z16.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aae50 // bfcvt z16.h, p3/M, z18.s\n"
+ "st1h { z16.s }, p0, [x10, #2, MUL VL]\n"
+ "inch x10, ALL, MUL #3\n"
+ "bgt 3b\n"
+ "b 52f\n"
+ "6:" // Initial: Height 2
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "mov x9, %x[bias]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "7:" // Initial: Height 2: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p0.s, x21, x11\n"
+ "incw x21\n"
+ "cbnz %x[bias], 8f\n"
+ "mov z24.b, #0x0\n"
+ "mov z23.b, #0x0\n"
+ "mov z22.b, #0x0\n"
+ "b 9f\n"
+ "8:" // Initial: Height 2: Width 3: bias
+ "ld1h { z18.s }, p2/Z, [x9]\n"
+ "ld1h { z17.s }, p1/Z, [x9, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x9, #2, MUL VL]\n"
+ "lsl z24.s, z18.s, #0x10\n"
+ "lsl z23.s, z17.s, #0x10\n"
+ "lsl z22.s, z16.s, #0x10\n"
+ "9:" // Initial: Height 2: Width 3: init done
+ "ld1w { z17.s }, p2/Z, [%x[in_ptr]]\n"
+ "ld1w { z16.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ "inch x9, ALL, MUL #3\n"
+ "ld1w { z19.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z18.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "ld1w { z21.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "ld1w { z20.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "fadd z17.s, z17.s, z24.s\n"
+ "fadd z16.s, z16.s, z23.s\n"
+ "cmp x11, XZR\n"
+ "fadd z19.s, z19.s, z22.s\n"
+ "fadd z18.s, z18.s, z24.s\n"
+ "fadd z21.s, z21.s, z23.s\n"
+ "fadd z20.s, z20.s, z22.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fmin z16.s, p3/M, z16.s, z12.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ "fmax z16.s, p3/M, z16.s, z11.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ ".inst 0x658aae31 // bfcvt z17.h, p3/M, z17.s\n"
+ ".inst 0x658aae10 // bfcvt z16.h, p3/M, z16.s\n"
+ ".inst 0x658aae73 // bfcvt z19.h, p3/M, z19.s\n"
+ ".inst 0x658aae52 // bfcvt z18.h, p3/M, z18.s\n"
+ "st1h { z17.s }, p2, [x10]\n"
+ "st1h { z16.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aaeb1 // bfcvt z17.h, p3/M, z21.s\n"
+ ".inst 0x658aae90 // bfcvt z16.h, p3/M, z20.s\n"
+ "st1h { z19.s }, p0, [x10, #2, MUL VL]\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z18.s }, p2, [x28]\n"
+ "st1h { z17.s }, p1, [x28, #1, MUL VL]\n"
+ "st1h { z16.s }, p0, [x28, #2, MUL VL]\n"
+ "inch x28, ALL, MUL #3\n"
+ "bgt 7b\n"
+ "b 52f\n"
+ "10:" // Initial: Height 3
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "mov x9, %x[bias]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "11:" // Initial: Height 3: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p0.s, x21, x11\n"
+ "incw x21\n"
+ "cbnz %x[bias], 12f\n"
+ "mov z27.b, #0x0\n"
+ "mov z26.b, #0x0\n"
+ "mov z25.b, #0x0\n"
+ "b 13f\n"
+ "12:" // Initial: Height 3: Width 3: bias
+ "ld1h { z18.s }, p2/Z, [x9]\n"
+ "ld1h { z17.s }, p1/Z, [x9, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x9, #2, MUL VL]\n"
+ "lsl z27.s, z18.s, #0x10\n"
+ "lsl z26.s, z17.s, #0x10\n"
+ "lsl z25.s, z16.s, #0x10\n"
+ "13:" // Initial: Height 3: Width 3: init done
+ "ld1w { z18.s }, p2/Z, [%x[in_ptr]]\n"
+ "ld1w { z17.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ "inch x9, ALL, MUL #3\n"
+ "ld1w { z16.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z21.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "ld1w { z20.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "ld1w { z19.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "ld1w { z24.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "ld1w { z23.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "fadd z18.s, z18.s, z27.s\n"
+ "fadd z17.s, z17.s, z26.s\n"
+ "ld1w { z22.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "fadd z16.s, z16.s, z25.s\n"
+ "fadd z21.s, z21.s, z27.s\n"
+ "cmp x11, XZR\n"
+ "fadd z20.s, z20.s, z26.s\n"
+ "fadd z19.s, z19.s, z25.s\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "fadd z24.s, z24.s, z27.s\n"
+ "fadd z23.s, z23.s, z26.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fadd z22.s, z22.s, z25.s\n"
+ "fmin z16.s, p3/M, z16.s, z12.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ "fmax z16.s, p3/M, z16.s, z11.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ ".inst 0x658aae52 // bfcvt z18.h, p3/M, z18.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ ".inst 0x658aae31 // bfcvt z17.h, p3/M, z17.s\n"
+ ".inst 0x658aae10 // bfcvt z16.h, p3/M, z16.s\n"
+ ".inst 0x658aaeb5 // bfcvt z21.h, p3/M, z21.s\n"
+ ".inst 0x658aae94 // bfcvt z20.h, p3/M, z20.s\n"
+ "st1h { z18.s }, p2, [x10]\n"
+ ".inst 0x658aae73 // bfcvt z19.h, p3/M, z19.s\n"
+ ".inst 0x658aaf12 // bfcvt z18.h, p3/M, z24.s\n"
+ "st1h { z17.s }, p1, [x10, #1, MUL VL]\n"
+ "st1h { z16.s }, p0, [x10, #2, MUL VL]\n"
+ ".inst 0x658aaef1 // bfcvt z17.h, p3/M, z23.s\n"
+ ".inst 0x658aaed0 // bfcvt z16.h, p3/M, z22.s\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z21.s }, p2, [x28]\n"
+ "st1h { z20.s }, p1, [x28, #1, MUL VL]\n"
+ "st1h { z19.s }, p0, [x28, #2, MUL VL]\n"
+ "inch x28, ALL, MUL #3\n"
+ "st1h { z18.s }, p2, [x27]\n"
+ "st1h { z17.s }, p1, [x27, #1, MUL VL]\n"
+ "st1h { z16.s }, p0, [x27, #2, MUL VL]\n"
+ "inch x27, ALL, MUL #3\n"
+ "bgt 11b\n"
+ "b 52f\n"
+ "14:" // Initial: Height 4
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "mov x9, %x[bias]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "15:" // Initial: Height 4: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p0.s, x21, x11\n"
+ "incw x21\n"
+ "cbnz %x[bias], 16f\n"
+ "mov z30.b, #0x0\n"
+ "mov z29.b, #0x0\n"
+ "mov z28.b, #0x0\n"
+ "b 17f\n"
+ "16:" // Initial: Height 4: Width 3: bias
+ "ld1h { z18.s }, p2/Z, [x9]\n"
+ "ld1h { z17.s }, p1/Z, [x9, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x9, #2, MUL VL]\n"
+ "lsl z30.s, z18.s, #0x10\n"
+ "lsl z29.s, z17.s, #0x10\n"
+ "lsl z28.s, z16.s, #0x10\n"
+ "17:" // Initial: Height 4: Width 3: init done
+ "ld1w { z18.s }, p2/Z, [%x[in_ptr]]\n"
+ "ld1w { z17.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ "inch x9, ALL, MUL #3\n"
+ "ld1w { z16.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z24.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "ld1w { z23.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "ld1w { z22.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "ld1w { z21.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "ld1w { z20.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "fadd z18.s, z18.s, z30.s\n"
+ "fadd z17.s, z17.s, z29.s\n"
+ "ld1w { z19.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "ld1w { z27.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "fadd z16.s, z16.s, z28.s\n"
+ "fadd z24.s, z24.s, z30.s\n"
+ "ld1w { z26.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "ld1w { z25.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "fadd z23.s, z23.s, z29.s\n"
+ "fadd z22.s, z22.s, z28.s\n"
+ "fadd z21.s, z21.s, z30.s\n"
+ "fadd z20.s, z20.s, z29.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fadd z19.s, z19.s, z28.s\n"
+ "fadd z27.s, z27.s, z30.s\n"
+ "fmin z16.s, p3/M, z16.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fadd z26.s, z26.s, z29.s\n"
+ "fadd z25.s, z25.s, z28.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ "fmax z16.s, p3/M, z16.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ ".inst 0x658aae52 // bfcvt z18.h, p3/M, z18.s\n"
+ ".inst 0x658aae31 // bfcvt z17.h, p3/M, z17.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ ".inst 0x658aae10 // bfcvt z16.h, p3/M, z16.s\n"
+ ".inst 0x658aaf18 // bfcvt z24.h, p3/M, z24.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ ".inst 0x658aaef7 // bfcvt z23.h, p3/M, z23.s\n"
+ ".inst 0x658aaed6 // bfcvt z22.h, p3/M, z22.s\n"
+ "cmp x11, XZR\n"
+ "st1h { z18.s }, p2, [x10]\n"
+ ".inst 0x658aaeb5 // bfcvt z21.h, p3/M, z21.s\n"
+ ".inst 0x658aae94 // bfcvt z20.h, p3/M, z20.s\n"
+ "st1h { z17.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aae73 // bfcvt z19.h, p3/M, z19.s\n"
+ ".inst 0x658aaf72 // bfcvt z18.h, p3/M, z27.s\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "st1h { z16.s }, p0, [x10, #2, MUL VL]\n"
+ ".inst 0x658aaf51 // bfcvt z17.h, p3/M, z26.s\n"
+ ".inst 0x658aaf30 // bfcvt z16.h, p3/M, z25.s\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z24.s }, p2, [x28]\n"
+ "st1h { z23.s }, p1, [x28, #1, MUL VL]\n"
+ "st1h { z22.s }, p0, [x28, #2, MUL VL]\n"
+ "inch x28, ALL, MUL #3\n"
+ "st1h { z21.s }, p2, [x27]\n"
+ "st1h { z20.s }, p1, [x27, #1, MUL VL]\n"
+ "st1h { z19.s }, p0, [x27, #2, MUL VL]\n"
+ "inch x27, ALL, MUL #3\n"
+ "st1h { z18.s }, p2, [x26]\n"
+ "st1h { z17.s }, p1, [x26, #1, MUL VL]\n"
+ "st1h { z16.s }, p0, [x26, #2, MUL VL]\n"
+ "inch x26, ALL, MUL #3\n"
+ "bgt 15b\n"
+ "b 52f\n"
+ "18:" // Initial: Height 5
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "mov x9, %x[bias]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "19:" // Initial: Height 5: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p0.s, x21, x11\n"
+ "incw x21\n"
+ "cbnz %x[bias], 20f\n"
+ "mov z1.b, #0x0\n"
+ "mov z0.b, #0x0\n"
+ "mov z31.b, #0x0\n"
+ "b 21f\n"
+ "20:" // Initial: Height 5: Width 3: bias
+ "ld1h { z18.s }, p2/Z, [x9]\n"
+ "ld1h { z17.s }, p1/Z, [x9, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x9, #2, MUL VL]\n"
+ "lsl z1.s, z18.s, #0x10\n"
+ "lsl z0.s, z17.s, #0x10\n"
+ "lsl z31.s, z16.s, #0x10\n"
+ "21:" // Initial: Height 5: Width 3: init done
+ "ld1w { z21.s }, p2/Z, [%x[in_ptr]]\n"
+ "ld1w { z20.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ "inch x9, ALL, MUL #3\n"
+ "ld1w { z19.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z18.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "ld1w { z17.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "ld1w { z16.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "ld1w { z24.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "ld1w { z23.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "fadd z21.s, z21.s, z1.s\n"
+ "fadd z20.s, z20.s, z0.s\n"
+ "ld1w { z22.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "ld1w { z30.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "fadd z19.s, z19.s, z31.s\n"
+ "fadd z18.s, z18.s, z1.s\n"
+ "ld1w { z29.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "ld1w { z28.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "fadd z17.s, z17.s, z0.s\n"
+ "fadd z16.s, z16.s, z31.s\n"
+ "ld1w { z27.s }, p2/Z, [x20, #-4, MUL VL]\n"
+ "ld1w { z26.s }, p1/Z, [x20, #-3, MUL VL]\n"
+ "fadd z24.s, z24.s, z1.s\n"
+ "fadd z23.s, z23.s, z0.s\n"
+ "ld1w { z25.s }, p0/Z, [x20, #-2, MUL VL]\n"
+ "fadd z22.s, z22.s, z31.s\n"
+ "fadd z30.s, z30.s, z1.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fadd z29.s, z29.s, z0.s\n"
+ "fadd z28.s, z28.s, z31.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fadd z27.s, z27.s, z1.s\n"
+ "fadd z26.s, z26.s, z0.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fadd z25.s, z25.s, z31.s\n"
+ "fmin z16.s, p3/M, z16.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmin z30.s, p3/M, z30.s, z12.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ "fmax z16.s, p3/M, z16.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ ".inst 0x658aaeb5 // bfcvt z21.h, p3/M, z21.s\n"
+ ".inst 0x658aae94 // bfcvt z20.h, p3/M, z20.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ "fmax z30.s, p3/M, z30.s, z11.s\n"
+ ".inst 0x658aae73 // bfcvt z19.h, p3/M, z19.s\n"
+ ".inst 0x658aae52 // bfcvt z18.h, p3/M, z18.s\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ ".inst 0x658aae31 // bfcvt z17.h, p3/M, z17.s\n"
+ ".inst 0x658aae10 // bfcvt z16.h, p3/M, z16.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "st1h { z21.s }, p2, [x10]\n"
+ ".inst 0x658aaf18 // bfcvt z24.h, p3/M, z24.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ "cmp x11, XZR\n"
+ "st1h { z20.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aaef7 // bfcvt z23.h, p3/M, z23.s\n"
+ "st1h { z19.s }, p0, [x10, #2, MUL VL]\n"
+ ".inst 0x658aaed6 // bfcvt z22.h, p3/M, z22.s\n"
+ ".inst 0x658aafd5 // bfcvt z21.h, p3/M, z30.s\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z18.s }, p2, [x28]\n"
+ ".inst 0x658aafb4 // bfcvt z20.h, p3/M, z29.s\n"
+ ".inst 0x658aaf93 // bfcvt z19.h, p3/M, z28.s\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "st1h { z17.s }, p1, [x28, #1, MUL VL]\n"
+ ".inst 0x658aaf72 // bfcvt z18.h, p3/M, z27.s\n"
+ ".inst 0x658aaf51 // bfcvt z17.h, p3/M, z26.s\n"
+ "st1h { z16.s }, p0, [x28, #2, MUL VL]\n"
+ ".inst 0x658aaf30 // bfcvt z16.h, p3/M, z25.s\n"
+ "inch x28, ALL, MUL #3\n"
+ "st1h { z24.s }, p2, [x27]\n"
+ "st1h { z23.s }, p1, [x27, #1, MUL VL]\n"
+ "st1h { z22.s }, p0, [x27, #2, MUL VL]\n"
+ "inch x27, ALL, MUL #3\n"
+ "st1h { z21.s }, p2, [x26]\n"
+ "st1h { z20.s }, p1, [x26, #1, MUL VL]\n"
+ "st1h { z19.s }, p0, [x26, #2, MUL VL]\n"
+ "inch x26, ALL, MUL #3\n"
+ "st1h { z18.s }, p2, [x25]\n"
+ "st1h { z17.s }, p1, [x25, #1, MUL VL]\n"
+ "st1h { z16.s }, p0, [x25, #2, MUL VL]\n"
+ "inch x25, ALL, MUL #3\n"
+ "bgt 19b\n"
+ "b 52f\n"
+ "22:" // Initial: Height 6
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "mov x9, %x[bias]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "23:" // Initial: Height 6: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p0.s, x21, x11\n"
+ "incw x21\n"
+ "cbnz %x[bias], 24f\n"
+ "mov z4.b, #0x0\n"
+ "mov z3.b, #0x0\n"
+ "mov z2.b, #0x0\n"
+ "b 25f\n"
+ "24:" // Initial: Height 6: Width 3: bias
+ "ld1h { z18.s }, p2/Z, [x9]\n"
+ "ld1h { z17.s }, p1/Z, [x9, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x9, #2, MUL VL]\n"
+ "lsl z4.s, z18.s, #0x10\n"
+ "lsl z3.s, z17.s, #0x10\n"
+ "lsl z2.s, z16.s, #0x10\n"
+ "25:" // Initial: Height 6: Width 3: init done
+ "ld1w { z17.s }, p2/Z, [%x[in_ptr]]\n"
+ "ld1w { z16.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ "inch x9, ALL, MUL #3\n"
+ "ld1w { z21.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z20.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "ld1w { z19.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "ld1w { z18.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "ld1w { z1.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "ld1w { z0.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "fadd z17.s, z17.s, z4.s\n"
+ "fadd z16.s, z16.s, z3.s\n"
+ "ld1w { z25.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "ld1w { z24.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "fadd z21.s, z21.s, z2.s\n"
+ "fadd z20.s, z20.s, z4.s\n"
+ "ld1w { z23.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "ld1w { z22.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "fadd z19.s, z19.s, z3.s\n"
+ "fadd z18.s, z18.s, z2.s\n"
+ "ld1w { z31.s }, p2/Z, [x20, #-4, MUL VL]\n"
+ "ld1w { z30.s }, p1/Z, [x20, #-3, MUL VL]\n"
+ "fadd z1.s, z1.s, z4.s\n"
+ "fadd z0.s, z0.s, z3.s\n"
+ "ld1w { z29.s }, p0/Z, [x20, #-2, MUL VL]\n"
+ "ld1w { z28.s }, p2/Z, [x20, #-1, MUL VL]\n"
+ "fadd z25.s, z25.s, z2.s\n"
+ "fadd z24.s, z24.s, z4.s\n"
+ "ld1w { z27.s }, p1/Z, [x20]\n"
+ "ld1w { z26.s }, p0/Z, [x20, #1, MUL VL]\n"
+ "fadd z23.s, z23.s, z3.s\n"
+ "fadd z22.s, z22.s, z2.s\n"
+ "fadd z31.s, z31.s, z4.s\n"
+ "fadd z30.s, z30.s, z3.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fmin z16.s, p3/M, z16.s, z12.s\n"
+ "fadd z29.s, z29.s, z2.s\n"
+ "fadd z28.s, z28.s, z4.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fadd z27.s, z27.s, z3.s\n"
+ "fadd z26.s, z26.s, z2.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmin z1.s, p3/M, z1.s, z12.s\n"
+ "fmin z0.s, p3/M, z0.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmin z31.s, p3/M, z31.s, z12.s\n"
+ "fmin z30.s, p3/M, z30.s, z12.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ "fmax z16.s, p3/M, z16.s, z11.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ "fmax z1.s, p3/M, z1.s, z11.s\n"
+ "fmax z0.s, p3/M, z0.s, z11.s\n"
+ ".inst 0x658aae31 // bfcvt z17.h, p3/M, z17.s\n"
+ ".inst 0x658aae10 // bfcvt z16.h, p3/M, z16.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ ".inst 0x658aaeb5 // bfcvt z21.h, p3/M, z21.s\n"
+ ".inst 0x658aae94 // bfcvt z20.h, p3/M, z20.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ ".inst 0x658aae73 // bfcvt z19.h, p3/M, z19.s\n"
+ ".inst 0x658aae52 // bfcvt z18.h, p3/M, z18.s\n"
+ "fmax z31.s, p3/M, z31.s, z11.s\n"
+ "fmax z30.s, p3/M, z30.s, z11.s\n"
+ "st1h { z17.s }, p2, [x10]\n"
+ ".inst 0x658aac31 // bfcvt z17.h, p3/M, z1.s\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ "st1h { z16.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aac10 // bfcvt z16.h, p3/M, z0.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "st1h { z21.s }, p0, [x10, #2, MUL VL]\n"
+ ".inst 0x658aaf39 // bfcvt z25.h, p3/M, z25.s\n"
+ "cmp x11, XZR\n"
+ "st1h { z20.s }, p2, [x28]\n"
+ ".inst 0x658aaf18 // bfcvt z24.h, p3/M, z24.s\n"
+ ".inst 0x658aaef7 // bfcvt z23.h, p3/M, z23.s\n"
+ "st1h { z19.s }, p1, [x28, #1, MUL VL]\n"
+ ".inst 0x658aaed6 // bfcvt z22.h, p3/M, z22.s\n"
+ ".inst 0x658aaff5 // bfcvt z21.h, p3/M, z31.s\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z18.s }, p0, [x28, #2, MUL VL]\n"
+ ".inst 0x658aafd4 // bfcvt z20.h, p3/M, z30.s\n"
+ ".inst 0x658aafb3 // bfcvt z19.h, p3/M, z29.s\n"
+ "inch x28, ALL, MUL #3\n"
+ "st1h { z17.s }, p2, [x27]\n"
+ ".inst 0x658aaf92 // bfcvt z18.h, p3/M, z28.s\n"
+ ".inst 0x658aaf71 // bfcvt z17.h, p3/M, z27.s\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "st1h { z16.s }, p1, [x27, #1, MUL VL]\n"
+ ".inst 0x658aaf50 // bfcvt z16.h, p3/M, z26.s\n"
+ "st1h { z25.s }, p0, [x27, #2, MUL VL]\n"
+ "inch x27, ALL, MUL #3\n"
+ "st1h { z24.s }, p2, [x26]\n"
+ "st1h { z23.s }, p1, [x26, #1, MUL VL]\n"
+ "st1h { z22.s }, p0, [x26, #2, MUL VL]\n"
+ "inch x26, ALL, MUL #3\n"
+ "st1h { z21.s }, p2, [x25]\n"
+ "st1h { z20.s }, p1, [x25, #1, MUL VL]\n"
+ "st1h { z19.s }, p0, [x25, #2, MUL VL]\n"
+ "inch x25, ALL, MUL #3\n"
+ "st1h { z18.s }, p2, [x24]\n"
+ "st1h { z17.s }, p1, [x24, #1, MUL VL]\n"
+ "st1h { z16.s }, p0, [x24, #2, MUL VL]\n"
+ "inch x24, ALL, MUL #3\n"
+ "bgt 23b\n"
+ "b 52f\n"
+ "26:" // Initial: Height 7
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "mov x9, %x[bias]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "27:" // Initial: Height 7: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p0.s, x21, x11\n"
+ "incw x21\n"
+ "cbnz %x[bias], 28f\n"
+ "mov z7.b, #0x0\n"
+ "mov z6.b, #0x0\n"
+ "mov z5.b, #0x0\n"
+ "b 29f\n"
+ "28:" // Initial: Height 7: Width 3: bias
+ "ld1h { z18.s }, p2/Z, [x9]\n"
+ "ld1h { z17.s }, p1/Z, [x9, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x9, #2, MUL VL]\n"
+ "lsl z7.s, z18.s, #0x10\n"
+ "lsl z6.s, z17.s, #0x10\n"
+ "lsl z5.s, z16.s, #0x10\n"
+ "29:" // Initial: Height 7: Width 3: init done
+ "ld1w { z19.s }, p2/Z, [%x[in_ptr]]\n"
+ "ld1w { z18.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ "inch x9, ALL, MUL #3\n"
+ "ld1w { z17.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z16.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "ld1w { z21.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "ld1w { z20.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "ld1w { z4.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "ld1w { z3.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "fadd z19.s, z19.s, z7.s\n"
+ "fadd z18.s, z18.s, z6.s\n"
+ "ld1w { z2.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "ld1w { z1.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "fadd z17.s, z17.s, z5.s\n"
+ "fadd z16.s, z16.s, z7.s\n"
+ "ld1w { z26.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "ld1w { z25.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "fadd z21.s, z21.s, z6.s\n"
+ "fadd z20.s, z20.s, z5.s\n"
+ "ld1w { z24.s }, p2/Z, [x20, #-4, MUL VL]\n"
+ "ld1w { z23.s }, p1/Z, [x20, #-3, MUL VL]\n"
+ "fadd z4.s, z4.s, z7.s\n"
+ "fadd z3.s, z3.s, z6.s\n"
+ "ld1w { z22.s }, p0/Z, [x20, #-2, MUL VL]\n"
+ "ld1w { z0.s }, p2/Z, [x20, #-1, MUL VL]\n"
+ "fadd z2.s, z2.s, z5.s\n"
+ "fadd z1.s, z1.s, z7.s\n"
+ "ld1w { z31.s }, p1/Z, [x20]\n"
+ "ld1w { z30.s }, p0/Z, [x20, #1, MUL VL]\n"
+ "fadd z26.s, z26.s, z6.s\n"
+ "fadd z25.s, z25.s, z5.s\n"
+ "ld1w { z29.s }, p2/Z, [x20, #2, MUL VL]\n"
+ "ld1w { z28.s }, p1/Z, [x20, #3, MUL VL]\n"
+ "fadd z24.s, z24.s, z7.s\n"
+ "fadd z23.s, z23.s, z6.s\n"
+ "ld1w { z27.s }, p0/Z, [x20, #4, MUL VL]\n"
+ "fadd z22.s, z22.s, z5.s\n"
+ "fadd z0.s, z0.s, z7.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fadd z31.s, z31.s, z6.s\n"
+ "fadd z30.s, z30.s, z5.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fadd z29.s, z29.s, z7.s\n"
+ "fadd z28.s, z28.s, z6.s\n"
+ "fmin z16.s, p3/M, z16.s, z12.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fadd z27.s, z27.s, z5.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fmin z4.s, p3/M, z4.s, z12.s\n"
+ "fmin z3.s, p3/M, z3.s, z12.s\n"
+ "fmin z2.s, p3/M, z2.s, z12.s\n"
+ "fmin z1.s, p3/M, z1.s, z12.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmin z0.s, p3/M, z0.s, z12.s\n"
+ "fmin z31.s, p3/M, z31.s, z12.s\n"
+ "fmin z30.s, p3/M, z30.s, z12.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ "fmax z16.s, p3/M, z16.s, z11.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ "fmax z4.s, p3/M, z4.s, z11.s\n"
+ "fmax z3.s, p3/M, z3.s, z11.s\n"
+ ".inst 0x658aae73 // bfcvt z19.h, p3/M, z19.s\n"
+ ".inst 0x658aae52 // bfcvt z18.h, p3/M, z18.s\n"
+ "fmax z2.s, p3/M, z2.s, z11.s\n"
+ "fmax z1.s, p3/M, z1.s, z11.s\n"
+ ".inst 0x658aae31 // bfcvt z17.h, p3/M, z17.s\n"
+ ".inst 0x658aae10 // bfcvt z16.h, p3/M, z16.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ ".inst 0x658aaeb5 // bfcvt z21.h, p3/M, z21.s\n"
+ ".inst 0x658aae94 // bfcvt z20.h, p3/M, z20.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "st1h { z19.s }, p2, [x10]\n"
+ ".inst 0x658aac93 // bfcvt z19.h, p3/M, z4.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ "fmax z0.s, p3/M, z0.s, z11.s\n"
+ "st1h { z18.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aac72 // bfcvt z18.h, p3/M, z3.s\n"
+ "fmax z31.s, p3/M, z31.s, z11.s\n"
+ "fmax z30.s, p3/M, z30.s, z11.s\n"
+ "st1h { z17.s }, p0, [x10, #2, MUL VL]\n"
+ ".inst 0x658aac51 // bfcvt z17.h, p3/M, z2.s\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ "st1h { z16.s }, p2, [x28]\n"
+ ".inst 0x658aac30 // bfcvt z16.h, p3/M, z1.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ "cmp x11, XZR\n"
+ "st1h { z21.s }, p1, [x28, #1, MUL VL]\n"
+ ".inst 0x658aaf5a // bfcvt z26.h, p3/M, z26.s\n"
+ "st1h { z20.s }, p0, [x28, #2, MUL VL]\n"
+ ".inst 0x658aaf39 // bfcvt z25.h, p3/M, z25.s\n"
+ ".inst 0x658aaf18 // bfcvt z24.h, p3/M, z24.s\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z19.s }, p2, [x27]\n"
+ ".inst 0x658aaef7 // bfcvt z23.h, p3/M, z23.s\n"
+ ".inst 0x658aaed6 // bfcvt z22.h, p3/M, z22.s\n"
+ "inch x28, ALL, MUL #3\n"
+ "st1h { z18.s }, p1, [x27, #1, MUL VL]\n"
+ ".inst 0x658aac15 // bfcvt z21.h, p3/M, z0.s\n"
+ ".inst 0x658aaff4 // bfcvt z20.h, p3/M, z31.s\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "st1h { z17.s }, p0, [x27, #2, MUL VL]\n"
+ ".inst 0x658aafd3 // bfcvt z19.h, p3/M, z30.s\n"
+ ".inst 0x658aafb2 // bfcvt z18.h, p3/M, z29.s\n"
+ "inch x27, ALL, MUL #3\n"
+ "st1h { z16.s }, p2, [x26]\n"
+ ".inst 0x658aaf91 // bfcvt z17.h, p3/M, z28.s\n"
+ ".inst 0x658aaf70 // bfcvt z16.h, p3/M, z27.s\n"
+ "st1h { z26.s }, p1, [x26, #1, MUL VL]\n"
+ "st1h { z25.s }, p0, [x26, #2, MUL VL]\n"
+ "inch x26, ALL, MUL #3\n"
+ "st1h { z24.s }, p2, [x25]\n"
+ "st1h { z23.s }, p1, [x25, #1, MUL VL]\n"
+ "st1h { z22.s }, p0, [x25, #2, MUL VL]\n"
+ "inch x25, ALL, MUL #3\n"
+ "st1h { z21.s }, p2, [x24]\n"
+ "st1h { z20.s }, p1, [x24, #1, MUL VL]\n"
+ "st1h { z19.s }, p0, [x24, #2, MUL VL]\n"
+ "inch x24, ALL, MUL #3\n"
+ "st1h { z18.s }, p2, [x23]\n"
+ "st1h { z17.s }, p1, [x23, #1, MUL VL]\n"
+ "st1h { z16.s }, p0, [x23, #2, MUL VL]\n"
+ "inch x23, ALL, MUL #3\n"
+ "bgt 27b\n"
+ "b 52f\n"
+ "30:" // Initial: Height 8
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "mov x9, %x[bias]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "add x22, x23, %x[ldout], LSL #1\n"
+ "31:" // Initial: Height 8: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "whilelt p0.s, x21, x11\n"
+ "incw x21\n"
+ "cbnz %x[bias], 32f\n"
+ "mov z10.b, #0x0\n"
+ "mov z9.b, #0x0\n"
+ "mov z8.b, #0x0\n"
+ "b 33f\n"
+ "32:" // Initial: Height 8: Width 3: bias
+ "ld1h { z18.s }, p2/Z, [x9]\n"
+ "ld1h { z17.s }, p1/Z, [x9, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x9, #2, MUL VL]\n"
+ "lsl z10.s, z18.s, #0x10\n"
+ "lsl z9.s, z17.s, #0x10\n"
+ "lsl z8.s, z16.s, #0x10\n"
+ "33:" // Initial: Height 8: Width 3: init done
+ "ld1w { z21.s }, p2/Z, [%x[in_ptr]]\n"
+ "ld1w { z20.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ "inch x9, ALL, MUL #3\n"
+ "ld1w { z19.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z18.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "ld1w { z17.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "ld1w { z16.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "ld1w { z7.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "ld1w { z6.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "fadd z21.s, z21.s, z10.s\n"
+ "fadd z20.s, z20.s, z9.s\n"
+ "ld1w { z5.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "ld1w { z4.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "fadd z19.s, z19.s, z8.s\n"
+ "fadd z18.s, z18.s, z10.s\n"
+ "ld1w { z3.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "ld1w { z2.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "fadd z17.s, z17.s, z9.s\n"
+ "fadd z16.s, z16.s, z8.s\n"
+ "ld1w { z27.s }, p2/Z, [x20, #-4, MUL VL]\n"
+ "ld1w { z26.s }, p1/Z, [x20, #-3, MUL VL]\n"
+ "fadd z7.s, z7.s, z10.s\n"
+ "fadd z6.s, z6.s, z9.s\n"
+ "ld1w { z25.s }, p0/Z, [x20, #-2, MUL VL]\n"
+ "ld1w { z24.s }, p2/Z, [x20, #-1, MUL VL]\n"
+ "fadd z5.s, z5.s, z8.s\n"
+ "fadd z4.s, z4.s, z10.s\n"
+ "ld1w { z23.s }, p1/Z, [x20]\n"
+ "ld1w { z22.s }, p0/Z, [x20, #1, MUL VL]\n"
+ "fadd z3.s, z3.s, z9.s\n"
+ "fadd z2.s, z2.s, z8.s\n"
+ "ld1w { z1.s }, p2/Z, [x20, #2, MUL VL]\n"
+ "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
+ "fadd z27.s, z27.s, z10.s\n"
+ "fadd z26.s, z26.s, z9.s\n"
+ "ld1w { z31.s }, p0/Z, [x20, #4, MUL VL]\n"
+ "ld1w { z30.s }, p2/Z, [x20, #5, MUL VL]\n"
+ "fadd z25.s, z25.s, z8.s\n"
+ "fadd z24.s, z24.s, z10.s\n"
+ "ld1w { z29.s }, p1/Z, [x20, #6, MUL VL]\n"
+ "ld1w { z28.s }, p0/Z, [x20, #7, MUL VL]\n"
+ "fadd z23.s, z23.s, z9.s\n"
+ "fadd z22.s, z22.s, z8.s\n"
+ "fadd z1.s, z1.s, z10.s\n"
+ "fadd z0.s, z0.s, z9.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fadd z31.s, z31.s, z8.s\n"
+ "fadd z30.s, z30.s, z10.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fadd z29.s, z29.s, z9.s\n"
+ "fadd z28.s, z28.s, z8.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fmin z16.s, p3/M, z16.s, z12.s\n"
+ "fmin z7.s, p3/M, z7.s, z12.s\n"
+ "fmin z6.s, p3/M, z6.s, z12.s\n"
+ "fmin z5.s, p3/M, z5.s, z12.s\n"
+ "fmin z4.s, p3/M, z4.s, z12.s\n"
+ "fmin z3.s, p3/M, z3.s, z12.s\n"
+ "fmin z2.s, p3/M, z2.s, z12.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmin z1.s, p3/M, z1.s, z12.s\n"
+ "fmin z0.s, p3/M, z0.s, z12.s\n"
+ "fmin z31.s, p3/M, z31.s, z12.s\n"
+ "fmin z30.s, p3/M, z30.s, z12.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ "fmax z16.s, p3/M, z16.s, z11.s\n"
+ "fmax z7.s, p3/M, z7.s, z11.s\n"
+ "fmax z6.s, p3/M, z6.s, z11.s\n"
+ ".inst 0x658aaeb5 // bfcvt z21.h, p3/M, z21.s\n"
+ ".inst 0x658aae94 // bfcvt z20.h, p3/M, z20.s\n"
+ "fmax z5.s, p3/M, z5.s, z11.s\n"
+ "fmax z4.s, p3/M, z4.s, z11.s\n"
+ ".inst 0x658aae73 // bfcvt z19.h, p3/M, z19.s\n"
+ ".inst 0x658aae52 // bfcvt z18.h, p3/M, z18.s\n"
+ "fmax z3.s, p3/M, z3.s, z11.s\n"
+ "fmax z2.s, p3/M, z2.s, z11.s\n"
+ ".inst 0x658aae31 // bfcvt z17.h, p3/M, z17.s\n"
+ ".inst 0x658aae10 // bfcvt z16.h, p3/M, z16.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "st1h { z21.s }, p2, [x10]\n"
+ ".inst 0x658aacf5 // bfcvt z21.h, p3/M, z7.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ "st1h { z20.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aacd4 // bfcvt z20.h, p3/M, z6.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ "st1h { z19.s }, p0, [x10, #2, MUL VL]\n"
+ ".inst 0x658aacb3 // bfcvt z19.h, p3/M, z5.s\n"
+ "fmax z1.s, p3/M, z1.s, z11.s\n"
+ "fmax z0.s, p3/M, z0.s, z11.s\n"
+ "st1h { z18.s }, p2, [x28]\n"
+ ".inst 0x658aac92 // bfcvt z18.h, p3/M, z4.s\n"
+ "fmax z31.s, p3/M, z31.s, z11.s\n"
+ "fmax z30.s, p3/M, z30.s, z11.s\n"
+ "st1h { z17.s }, p1, [x28, #1, MUL VL]\n"
+ ".inst 0x658aac71 // bfcvt z17.h, p3/M, z3.s\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ "st1h { z16.s }, p0, [x28, #2, MUL VL]\n"
+ ".inst 0x658aac50 // bfcvt z16.h, p3/M, z2.s\n"
+ "cmp x11, XZR\n"
+ "st1h { z21.s }, p2, [x27]\n"
+ ".inst 0x658aaf7b // bfcvt z27.h, p3/M, z27.s\n"
+ ".inst 0x658aaf5a // bfcvt z26.h, p3/M, z26.s\n"
+ "st1h { z20.s }, p1, [x27, #1, MUL VL]\n"
+ ".inst 0x658aaf39 // bfcvt z25.h, p3/M, z25.s\n"
+ ".inst 0x658aaf18 // bfcvt z24.h, p3/M, z24.s\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z19.s }, p0, [x27, #2, MUL VL]\n"
+ ".inst 0x658aaef7 // bfcvt z23.h, p3/M, z23.s\n"
+ ".inst 0x658aaed6 // bfcvt z22.h, p3/M, z22.s\n"
+ "inch x28, ALL, MUL #3\n"
+ "st1h { z18.s }, p2, [x26]\n"
+ ".inst 0x658aac35 // bfcvt z21.h, p3/M, z1.s\n"
+ ".inst 0x658aac14 // bfcvt z20.h, p3/M, z0.s\n"
+ "inch x27, ALL, MUL #3\n"
+ "st1h { z17.s }, p1, [x26, #1, MUL VL]\n"
+ ".inst 0x658aaff3 // bfcvt z19.h, p3/M, z31.s\n"
+ ".inst 0x658aafd2 // bfcvt z18.h, p3/M, z30.s\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "st1h { z16.s }, p0, [x26, #2, MUL VL]\n"
+ ".inst 0x658aafb1 // bfcvt z17.h, p3/M, z29.s\n"
+ ".inst 0x658aaf90 // bfcvt z16.h, p3/M, z28.s\n"
+ "inch x26, ALL, MUL #3\n"
+ "st1h { z27.s }, p2, [x25]\n"
+ "st1h { z26.s }, p1, [x25, #1, MUL VL]\n"
+ "st1h { z25.s }, p0, [x25, #2, MUL VL]\n"
+ "inch x25, ALL, MUL #3\n"
+ "st1h { z24.s }, p2, [x24]\n"
+ "st1h { z23.s }, p1, [x24, #1, MUL VL]\n"
+ "st1h { z22.s }, p0, [x24, #2, MUL VL]\n"
+ "inch x24, ALL, MUL #3\n"
+ "st1h { z21.s }, p2, [x23]\n"
+ "st1h { z20.s }, p1, [x23, #1, MUL VL]\n"
+ "st1h { z19.s }, p0, [x23, #2, MUL VL]\n"
+ "inch x23, ALL, MUL #3\n"
+ "st1h { z18.s }, p2, [x22]\n"
+ "st1h { z17.s }, p1, [x22, #1, MUL VL]\n"
+ "st1h { z16.s }, p0, [x22, #2, MUL VL]\n"
+ "inch x22, ALL, MUL #3\n"
+ "bgt 31b\n"
+ "subs %x[rows], %x[rows], #0x8\n"
+ "add %x[out_ptr], %x[out_ptr], x12\n"
+ "bgt 1b\n"
+ "b 52f\n"
+ "34:" // Accumulate
+ "35:" // Accumulate: Row loop
+ "cmp %x[rows], #0x7\n"
+ "bgt 50f\n"
+ "beq 48f\n"
+ "cmp %x[rows], #0x5\n"
+ "bgt 46f\n"
+ "beq 44f\n"
+ "cmp %x[rows], #0x3\n"
+ "bgt 42f\n"
+ "beq 40f\n"
+ "cmp %x[rows], #0x1\n"
+ "bgt 38f\n"
+ "36:" // Accumulate: Height 1
+ "mov x11, %x[cols]\n"
+ "mov x10, %x[out_ptr]\n"
+ "37:" // Accumulate: Height 1: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "ld1h { z16.s }, p2/Z, [x10]\n"
+ "ld1w { z19.s }, p2/Z, [%x[in_ptr]]\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "fadd z19.s, z19.s, z16.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "ld1w { z18.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "whilelt p0.s, x21, x11\n"
+ "decw x11, ALL, MUL #3\n"
+ "incw x21\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ "ld1w { z17.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "cmp x11, XZR\n"
+ ".inst 0x658aae70 // bfcvt z16.h, p3/M, z19.s\n"
+ "st1h { z16.s }, p2, [x10]\n"
+ "ld1h { z16.s }, p1/Z, [x10, #1, MUL VL]\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z18.s, z18.s, z16.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ ".inst 0x658aae50 // bfcvt z16.h, p3/M, z18.s\n"
+ "st1h { z16.s }, p1, [x10, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x10, #2, MUL VL]\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z17.s, z17.s, z16.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ ".inst 0x658aae30 // bfcvt z16.h, p3/M, z17.s\n"
+ "st1h { z16.s }, p0, [x10, #2, MUL VL]\n"
+ "inch x10, ALL, MUL #3\n"
+ "bgt 37b\n"
+ "b 52f\n"
+ "38:" // Accumulate: Height 2
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "39:" // Accumulate: Height 2: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "ld1h { z17.s }, p2/Z, [x10]\n"
+ "ld1h { z16.s }, p2/Z, [x28]\n"
+ "ld1w { z23.s }, p2/Z, [%x[in_ptr]]\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "ld1w { z22.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "fadd z23.s, z23.s, z17.s\n"
+ "fadd z22.s, z22.s, z16.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "ld1w { z21.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "ld1w { z20.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "whilelt p0.s, x21, x11\n"
+ "decw x11, ALL, MUL #3\n"
+ "incw x21\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ "ld1w { z19.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z18.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "cmp x11, XZR\n"
+ ".inst 0x658aaef1 // bfcvt z17.h, p3/M, z23.s\n"
+ ".inst 0x658aaed0 // bfcvt z16.h, p3/M, z22.s\n"
+ "st1h { z17.s }, p2, [x10]\n"
+ "st1h { z16.s }, p2, [x28]\n"
+ "ld1h { z17.s }, p1/Z, [x10, #1, MUL VL]\n"
+ "ld1h { z16.s }, p1/Z, [x28, #1, MUL VL]\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z21.s, z21.s, z17.s\n"
+ "fadd z20.s, z20.s, z16.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ ".inst 0x658aaeb0 // bfcvt z16.h, p3/M, z21.s\n"
+ "st1h { z16.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aae90 // bfcvt z16.h, p3/M, z20.s\n"
+ "ld1h { z17.s }, p0/Z, [x10, #2, MUL VL]\n"
+ "st1h { z16.s }, p1, [x28, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x28, #2, MUL VL]\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z19.s, z19.s, z17.s\n"
+ "fadd z18.s, z18.s, z16.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ ".inst 0x658aae70 // bfcvt z16.h, p3/M, z19.s\n"
+ "st1h { z16.s }, p0, [x10, #2, MUL VL]\n"
+ "inch x10, ALL, MUL #3\n"
+ ".inst 0x658aae50 // bfcvt z16.h, p3/M, z18.s\n"
+ "st1h { z16.s }, p0, [x28, #2, MUL VL]\n"
+ "inch x28, ALL, MUL #3\n"
+ "bgt 39b\n"
+ "b 52f\n"
+ "40:" // Accumulate: Height 3
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "41:" // Accumulate: Height 3: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "ld1h { z18.s }, p2/Z, [x10]\n"
+ "ld1h { z17.s }, p2/Z, [x28]\n"
+ "ld1h { z16.s }, p2/Z, [x27]\n"
+ "ld1w { z26.s }, p2/Z, [%x[in_ptr]]\n"
+ "lsl z19.s, z18.s, #0x10\n"
+ "ld1w { z25.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "ld1w { z18.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z26.s, z26.s, z19.s\n"
+ "fadd z25.s, z25.s, z17.s\n"
+ "ld1w { z24.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "ld1w { z23.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "ld1w { z22.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "fadd z18.s, z18.s, z16.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "whilelt p0.s, x21, x11\n"
+ "decw x11, ALL, MUL #3\n"
+ "incw x21\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "ld1w { z21.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z20.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "ld1w { z19.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "cmp x11, XZR\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ ".inst 0x658aaf51 // bfcvt z17.h, p3/M, z26.s\n"
+ ".inst 0x658aaf30 // bfcvt z16.h, p3/M, z25.s\n"
+ "st1h { z17.s }, p2, [x10]\n"
+ "st1h { z16.s }, p2, [x28]\n"
+ ".inst 0x658aae51 // bfcvt z17.h, p3/M, z18.s\n"
+ "ld1h { z16.s }, p1/Z, [x10, #1, MUL VL]\n"
+ "st1h { z17.s }, p2, [x27]\n"
+ "ld1h { z17.s }, p1/Z, [x28, #1, MUL VL]\n"
+ "lsl z18.s, z16.s, #0x10\n"
+ "ld1h { z16.s }, p1/Z, [x27, #1, MUL VL]\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z24.s, z24.s, z18.s\n"
+ "fadd z23.s, z23.s, z17.s\n"
+ "fadd z22.s, z22.s, z16.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ ".inst 0x658aaf10 // bfcvt z16.h, p3/M, z24.s\n"
+ "st1h { z16.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aaef2 // bfcvt z18.h, p3/M, z23.s\n"
+ ".inst 0x658aaed1 // bfcvt z17.h, p3/M, z22.s\n"
+ "ld1h { z16.s }, p0/Z, [x10, #2, MUL VL]\n"
+ "st1h { z18.s }, p1, [x28, #1, MUL VL]\n"
+ "st1h { z17.s }, p1, [x27, #1, MUL VL]\n"
+ "ld1h { z17.s }, p0/Z, [x28, #2, MUL VL]\n"
+ "lsl z18.s, z16.s, #0x10\n"
+ "ld1h { z16.s }, p0/Z, [x27, #2, MUL VL]\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z21.s, z21.s, z18.s\n"
+ "fadd z20.s, z20.s, z17.s\n"
+ "fadd z19.s, z19.s, z16.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ ".inst 0x658aaeb0 // bfcvt z16.h, p3/M, z21.s\n"
+ "st1h { z16.s }, p0, [x10, #2, MUL VL]\n"
+ "inch x10, ALL, MUL #3\n"
+ ".inst 0x658aae91 // bfcvt z17.h, p3/M, z20.s\n"
+ ".inst 0x658aae70 // bfcvt z16.h, p3/M, z19.s\n"
+ "st1h { z17.s }, p0, [x28, #2, MUL VL]\n"
+ "inch x28, ALL, MUL #3\n"
+ "st1h { z16.s }, p0, [x27, #2, MUL VL]\n"
+ "inch x27, ALL, MUL #3\n"
+ "bgt 41b\n"
+ "b 52f\n"
+ "42:" // Accumulate: Height 4
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "43:" // Accumulate: Height 4: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "ld1h { z19.s }, p2/Z, [x10]\n"
+ "ld1h { z18.s }, p2/Z, [x28]\n"
+ "ld1h { z17.s }, p2/Z, [x27]\n"
+ "ld1h { z16.s }, p2/Z, [x26]\n"
+ "ld1w { z30.s }, p2/Z, [%x[in_ptr]]\n"
+ "lsl z20.s, z19.s, #0x10\n"
+ "ld1w { z29.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "ld1w { z28.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "ld1w { z19.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z30.s, z30.s, z20.s\n"
+ "fadd z29.s, z29.s, z18.s\n"
+ "ld1w { z27.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "ld1w { z26.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "ld1w { z25.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "ld1w { z24.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "whilelt p0.s, x21, x11\n"
+ "decw x11, ALL, MUL #3\n"
+ "fadd z28.s, z28.s, z17.s\n"
+ "fadd z19.s, z19.s, z16.s\n"
+ "incw x21\n"
+ "fmin z30.s, p3/M, z30.s, z12.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "ld1w { z23.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z22.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "ld1w { z21.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "ld1w { z20.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "cmp x11, XZR\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fmin z19.s, p3/M, z19.s, z12.s\n"
+ "fmax z30.s, p3/M, z30.s, z11.s\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ "fmax z19.s, p3/M, z19.s, z11.s\n"
+ ".inst 0x658aafd2 // bfcvt z18.h, p3/M, z30.s\n"
+ ".inst 0x658aafb1 // bfcvt z17.h, p3/M, z29.s\n"
+ ".inst 0x658aaf90 // bfcvt z16.h, p3/M, z28.s\n"
+ "st1h { z18.s }, p2, [x10]\n"
+ "st1h { z17.s }, p2, [x28]\n"
+ ".inst 0x658aae71 // bfcvt z17.h, p3/M, z19.s\n"
+ "st1h { z16.s }, p2, [x27]\n"
+ "ld1h { z16.s }, p1/Z, [x10, #1, MUL VL]\n"
+ "st1h { z17.s }, p2, [x26]\n"
+ "ld1h { z18.s }, p1/Z, [x28, #1, MUL VL]\n"
+ "ld1h { z17.s }, p1/Z, [x27, #1, MUL VL]\n"
+ "lsl z19.s, z16.s, #0x10\n"
+ "ld1h { z16.s }, p1/Z, [x26, #1, MUL VL]\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "fadd z27.s, z27.s, z19.s\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z26.s, z26.s, z18.s\n"
+ "fadd z25.s, z25.s, z17.s\n"
+ "fadd z24.s, z24.s, z16.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ ".inst 0x658aaf71 // bfcvt z17.h, p3/M, z27.s\n"
+ ".inst 0x658aaf50 // bfcvt z16.h, p3/M, z26.s\n"
+ "st1h { z17.s }, p1, [x10, #1, MUL VL]\n"
+ "st1h { z16.s }, p1, [x28, #1, MUL VL]\n"
+ ".inst 0x658aaf32 // bfcvt z18.h, p3/M, z25.s\n"
+ ".inst 0x658aaf11 // bfcvt z17.h, p3/M, z24.s\n"
+ "ld1h { z16.s }, p0/Z, [x10, #2, MUL VL]\n"
+ "st1h { z18.s }, p1, [x27, #1, MUL VL]\n"
+ "st1h { z17.s }, p1, [x26, #1, MUL VL]\n"
+ "ld1h { z18.s }, p0/Z, [x28, #2, MUL VL]\n"
+ "lsl z19.s, z16.s, #0x10\n"
+ "ld1h { z17.s }, p0/Z, [x27, #2, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x26, #2, MUL VL]\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "fadd z23.s, z23.s, z19.s\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z22.s, z22.s, z18.s\n"
+ "fadd z21.s, z21.s, z17.s\n"
+ "fadd z20.s, z20.s, z16.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ ".inst 0x658aaef1 // bfcvt z17.h, p3/M, z23.s\n"
+ ".inst 0x658aaed0 // bfcvt z16.h, p3/M, z22.s\n"
+ "st1h { z17.s }, p0, [x10, #2, MUL VL]\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z16.s }, p0, [x28, #2, MUL VL]\n"
+ "inch x28, ALL, MUL #3\n"
+ ".inst 0x658aaeb1 // bfcvt z17.h, p3/M, z21.s\n"
+ ".inst 0x658aae90 // bfcvt z16.h, p3/M, z20.s\n"
+ "st1h { z17.s }, p0, [x27, #2, MUL VL]\n"
+ "inch x27, ALL, MUL #3\n"
+ "st1h { z16.s }, p0, [x26, #2, MUL VL]\n"
+ "inch x26, ALL, MUL #3\n"
+ "bgt 43b\n"
+ "b 52f\n"
+ "44:" // Accumulate: Height 5
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "45:" // Accumulate: Height 5: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "ld1h { z20.s }, p2/Z, [x10]\n"
+ "ld1h { z19.s }, p2/Z, [x28]\n"
+ "ld1h { z18.s }, p2/Z, [x27]\n"
+ "ld1h { z17.s }, p2/Z, [x26]\n"
+ "ld1h { z16.s }, p2/Z, [x25]\n"
+ "ld1w { z1.s }, p2/Z, [%x[in_ptr]]\n"
+ "lsl z22.s, z20.s, #0x10\n"
+ "ld1w { z0.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "lsl z21.s, z19.s, #0x10\n"
+ "ld1w { z31.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "whilelt p1.s, x21, x11\n"
+ "lsl z19.s, z18.s, #0x10\n"
+ "ld1w { z20.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "lsl z18.s, z17.s, #0x10\n"
+ "ld1w { z17.s }, p2/Z, [x20, #-4, MUL VL]\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z1.s, z1.s, z22.s\n"
+ "incw x21\n"
+ "fadd z0.s, z0.s, z21.s\n"
+ "ld1w { z30.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "ld1w { z29.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "fadd z31.s, z31.s, z19.s\n"
+ "fadd z20.s, z20.s, z18.s\n"
+ "ld1w { z28.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "ld1w { z27.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "fadd z17.s, z17.s, z16.s\n"
+ "fmin z1.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z26.s }, p1/Z, [x20, #-3, MUL VL]\n"
+ "whilelt p0.s, x21, x11\n"
+ "fmin z0.s, p3/M, z0.s, z12.s\n"
+ "fmin z31.s, p3/M, z31.s, z12.s\n"
+ "fmin z20.s, p3/M, z20.s, z12.s\n"
+ "fmin z17.s, p3/M, z17.s, z12.s\n"
+ "fmax z1.s, p3/M, z1.s, z11.s\n"
+ "ld1w { z25.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z24.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ "fmax z0.s, p3/M, z0.s, z11.s\n"
+ "fmax z31.s, p3/M, z31.s, z11.s\n"
+ "ld1w { z23.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "ld1w { z22.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "fmax z20.s, p3/M, z20.s, z11.s\n"
+ "fmax z17.s, p3/M, z17.s, z11.s\n"
+ "ld1w { z21.s }, p0/Z, [x20, #-2, MUL VL]\n"
+ ".inst 0x658aac30 // bfcvt z16.h, p3/M, z1.s\n"
+ "cmp x11, XZR\n"
+ "incw x21\n"
+ ".inst 0x658aac13 // bfcvt z19.h, p3/M, z0.s\n"
+ ".inst 0x658aaff2 // bfcvt z18.h, p3/M, z31.s\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "st1h { z16.s }, p2, [x10]\n"
+ ".inst 0x658aae90 // bfcvt z16.h, p3/M, z20.s\n"
+ ".inst 0x658aae31 // bfcvt z17.h, p3/M, z17.s\n"
+ "st1h { z19.s }, p2, [x28]\n"
+ "st1h { z18.s }, p2, [x27]\n"
+ "st1h { z16.s }, p2, [x26]\n"
+ "ld1h { z16.s }, p1/Z, [x10, #1, MUL VL]\n"
+ "st1h { z17.s }, p2, [x25]\n"
+ "ld1h { z19.s }, p1/Z, [x28, #1, MUL VL]\n"
+ "ld1h { z18.s }, p1/Z, [x27, #1, MUL VL]\n"
+ "ld1h { z17.s }, p1/Z, [x26, #1, MUL VL]\n"
+ "lsl z20.s, z16.s, #0x10\n"
+ "ld1h { z16.s }, p1/Z, [x25, #1, MUL VL]\n"
+ "lsl z19.s, z19.s, #0x10\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z30.s, z30.s, z20.s\n"
+ "fadd z29.s, z29.s, z19.s\n"
+ "fadd z28.s, z28.s, z18.s\n"
+ "fadd z27.s, z27.s, z17.s\n"
+ "fadd z26.s, z26.s, z16.s\n"
+ "fmin z30.s, p3/M, z30.s, z12.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmax z30.s, p3/M, z30.s, z11.s\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ ".inst 0x658aafd2 // bfcvt z18.h, p3/M, z30.s\n"
+ ".inst 0x658aafb1 // bfcvt z17.h, p3/M, z29.s\n"
+ ".inst 0x658aaf90 // bfcvt z16.h, p3/M, z28.s\n"
+ "st1h { z18.s }, p1, [x10, #1, MUL VL]\n"
+ "st1h { z17.s }, p1, [x28, #1, MUL VL]\n"
+ ".inst 0x658aaf72 // bfcvt z18.h, p3/M, z27.s\n"
+ ".inst 0x658aaf51 // bfcvt z17.h, p3/M, z26.s\n"
+ "st1h { z16.s }, p1, [x27, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x10, #2, MUL VL]\n"
+ "st1h { z18.s }, p1, [x26, #1, MUL VL]\n"
+ "st1h { z17.s }, p1, [x25, #1, MUL VL]\n"
+ "ld1h { z19.s }, p0/Z, [x28, #2, MUL VL]\n"
+ "ld1h { z18.s }, p0/Z, [x27, #2, MUL VL]\n"
+ "lsl z20.s, z16.s, #0x10\n"
+ "ld1h { z17.s }, p0/Z, [x26, #2, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x25, #2, MUL VL]\n"
+ "lsl z19.s, z19.s, #0x10\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "fadd z25.s, z25.s, z20.s\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z24.s, z24.s, z19.s\n"
+ "fadd z23.s, z23.s, z18.s\n"
+ "fadd z22.s, z22.s, z17.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fadd z21.s, z21.s, z16.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ ".inst 0x658aaf31 // bfcvt z17.h, p3/M, z25.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ ".inst 0x658aaf10 // bfcvt z16.h, p3/M, z24.s\n"
+ "st1h { z17.s }, p0, [x10, #2, MUL VL]\n"
+ "inch x10, ALL, MUL #3\n"
+ ".inst 0x658aaef2 // bfcvt z18.h, p3/M, z23.s\n"
+ ".inst 0x658aaed1 // bfcvt z17.h, p3/M, z22.s\n"
+ "st1h { z16.s }, p0, [x28, #2, MUL VL]\n"
+ "inch x28, ALL, MUL #3\n"
+ ".inst 0x658aaeb0 // bfcvt z16.h, p3/M, z21.s\n"
+ "st1h { z18.s }, p0, [x27, #2, MUL VL]\n"
+ "inch x27, ALL, MUL #3\n"
+ "st1h { z17.s }, p0, [x26, #2, MUL VL]\n"
+ "inch x26, ALL, MUL #3\n"
+ "st1h { z16.s }, p0, [x25, #2, MUL VL]\n"
+ "inch x25, ALL, MUL #3\n"
+ "bgt 45b\n"
+ "b 52f\n"
+ "46:" // Accumulate: Height 6
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "47:" // Accumulate: Height 6: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "ld1h { z21.s }, p2/Z, [x10]\n"
+ "ld1h { z20.s }, p2/Z, [x28]\n"
+ "ld1h { z19.s }, p2/Z, [x27]\n"
+ "ld1h { z18.s }, p2/Z, [x26]\n"
+ "ld1h { z17.s }, p2/Z, [x25]\n"
+ "ld1h { z16.s }, p2/Z, [x24]\n"
+ "ld1w { z6.s }, p2/Z, [%x[in_ptr]]\n"
+ "lsl z22.s, z21.s, #0x10\n"
+ "ld1w { z5.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "lsl z21.s, z20.s, #0x10\n"
+ "ld1w { z4.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "lsl z20.s, z19.s, #0x10\n"
+ "ld1w { z3.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "lsl z19.s, z18.s, #0x10\n"
+ "ld1w { z2.s }, p2/Z, [x20, #-4, MUL VL]\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "ld1w { z18.s }, p2/Z, [x20, #-1, MUL VL]\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z6.s, z6.s, z22.s\n"
+ "fadd z5.s, z5.s, z21.s\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "fadd z4.s, z4.s, z20.s\n"
+ "fadd z3.s, z3.s, z19.s\n"
+ "fadd z2.s, z2.s, z17.s\n"
+ "fadd z18.s, z18.s, z16.s\n"
+ "fmin z6.s, p3/M, z6.s, z12.s\n"
+ "fmin z5.s, p3/M, z5.s, z12.s\n"
+ "ld1w { z1.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "ld1w { z0.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "whilelt p0.s, x21, x11\n"
+ "decw x11, ALL, MUL #3\n"
+ "fmin z4.s, p3/M, z4.s, z12.s\n"
+ "fmin z3.s, p3/M, z3.s, z12.s\n"
+ "ld1w { z31.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "ld1w { z30.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "fmin z2.s, p3/M, z2.s, z12.s\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "ld1w { z29.s }, p1/Z, [x20, #-3, MUL VL]\n"
+ "ld1w { z28.s }, p1/Z, [x20]\n"
+ "fmax z6.s, p3/M, z6.s, z11.s\n"
+ "fmax z5.s, p3/M, z5.s, z11.s\n"
+ "ld1w { z27.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z26.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "fmax z4.s, p3/M, z4.s, z11.s\n"
+ "fmax z3.s, p3/M, z3.s, z11.s\n"
+ "ld1w { z25.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "ld1w { z24.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "fmax z2.s, p3/M, z2.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ "ld1w { z23.s }, p0/Z, [x20, #-2, MUL VL]\n"
+ "ld1w { z22.s }, p0/Z, [x20, #1, MUL VL]\n"
+ ".inst 0x658aacd5 // bfcvt z21.h, p3/M, z6.s\n"
+ ".inst 0x658aacb4 // bfcvt z20.h, p3/M, z5.s\n"
+ "cmp x11, XZR\n"
+ "incw x21\n"
+ ".inst 0x658aac93 // bfcvt z19.h, p3/M, z4.s\n"
+ ".inst 0x658aac71 // bfcvt z17.h, p3/M, z3.s\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ ".inst 0x658aac50 // bfcvt z16.h, p3/M, z2.s\n"
+ ".inst 0x658aae52 // bfcvt z18.h, p3/M, z18.s\n"
+ "st1h { z21.s }, p2, [x10]\n"
+ "st1h { z20.s }, p2, [x28]\n"
+ "st1h { z19.s }, p2, [x27]\n"
+ "st1h { z17.s }, p2, [x26]\n"
+ "ld1h { z17.s }, p1/Z, [x10, #1, MUL VL]\n"
+ "st1h { z16.s }, p2, [x25]\n"
+ "ld1h { z16.s }, p1/Z, [x28, #1, MUL VL]\n"
+ "st1h { z18.s }, p2, [x24]\n"
+ "ld1h { z19.s }, p1/Z, [x27, #1, MUL VL]\n"
+ "ld1h { z18.s }, p1/Z, [x26, #1, MUL VL]\n"
+ "lsl z21.s, z17.s, #0x10\n"
+ "ld1h { z17.s }, p1/Z, [x25, #1, MUL VL]\n"
+ "lsl z20.s, z16.s, #0x10\n"
+ "ld1h { z16.s }, p1/Z, [x24, #1, MUL VL]\n"
+ "lsl z19.s, z19.s, #0x10\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "fadd z1.s, z1.s, z21.s\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z0.s, z0.s, z20.s\n"
+ "fadd z31.s, z31.s, z19.s\n"
+ "fadd z30.s, z30.s, z18.s\n"
+ "fmin z1.s, p3/M, z1.s, z12.s\n"
+ "fadd z29.s, z29.s, z17.s\n"
+ "fadd z28.s, z28.s, z16.s\n"
+ "fmin z0.s, p3/M, z0.s, z12.s\n"
+ "fmin z31.s, p3/M, z31.s, z12.s\n"
+ "fmin z30.s, p3/M, z30.s, z12.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "fmax z1.s, p3/M, z1.s, z11.s\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fmax z0.s, p3/M, z0.s, z11.s\n"
+ "fmax z31.s, p3/M, z31.s, z11.s\n"
+ "fmax z30.s, p3/M, z30.s, z11.s\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ ".inst 0x658aac34 // bfcvt z20.h, p3/M, z1.s\n"
+ ".inst 0x658aac12 // bfcvt z18.h, p3/M, z0.s\n"
+ ".inst 0x658aaff3 // bfcvt z19.h, p3/M, z31.s\n"
+ ".inst 0x658aafd1 // bfcvt z17.h, p3/M, z30.s\n"
+ ".inst 0x658aafb0 // bfcvt z16.h, p3/M, z29.s\n"
+ "st1h { z20.s }, p1, [x10, #1, MUL VL]\n"
+ "st1h { z18.s }, p1, [x28, #1, MUL VL]\n"
+ ".inst 0x658aaf92 // bfcvt z18.h, p3/M, z28.s\n"
+ "st1h { z19.s }, p1, [x27, #1, MUL VL]\n"
+ "st1h { z17.s }, p1, [x26, #1, MUL VL]\n"
+ "ld1h { z17.s }, p0/Z, [x10, #2, MUL VL]\n"
+ "st1h { z16.s }, p1, [x25, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x28, #2, MUL VL]\n"
+ "st1h { z18.s }, p1, [x24, #1, MUL VL]\n"
+ "ld1h { z19.s }, p0/Z, [x27, #2, MUL VL]\n"
+ "ld1h { z18.s }, p0/Z, [x26, #2, MUL VL]\n"
+ "lsl z21.s, z17.s, #0x10\n"
+ "ld1h { z17.s }, p0/Z, [x25, #2, MUL VL]\n"
+ "lsl z20.s, z16.s, #0x10\n"
+ "ld1h { z16.s }, p0/Z, [x24, #2, MUL VL]\n"
+ "lsl z19.s, z19.s, #0x10\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "fadd z27.s, z27.s, z21.s\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z26.s, z26.s, z20.s\n"
+ "fadd z25.s, z25.s, z19.s\n"
+ "fadd z24.s, z24.s, z18.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fadd z23.s, z23.s, z17.s\n"
+ "fadd z22.s, z22.s, z16.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ ".inst 0x658aaf74 // bfcvt z20.h, p3/M, z27.s\n"
+ ".inst 0x658aaf50 // bfcvt z16.h, p3/M, z26.s\n"
+ ".inst 0x658aaf33 // bfcvt z19.h, p3/M, z25.s\n"
+ ".inst 0x658aaf12 // bfcvt z18.h, p3/M, z24.s\n"
+ ".inst 0x658aaef1 // bfcvt z17.h, p3/M, z23.s\n"
+ "st1h { z20.s }, p0, [x10, #2, MUL VL]\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z16.s }, p0, [x28, #2, MUL VL]\n"
+ ".inst 0x658aaed0 // bfcvt z16.h, p3/M, z22.s\n"
+ "inch x28, ALL, MUL #3\n"
+ "st1h { z19.s }, p0, [x27, #2, MUL VL]\n"
+ "inch x27, ALL, MUL #3\n"
+ "st1h { z18.s }, p0, [x26, #2, MUL VL]\n"
+ "inch x26, ALL, MUL #3\n"
+ "st1h { z17.s }, p0, [x25, #2, MUL VL]\n"
+ "inch x25, ALL, MUL #3\n"
+ "st1h { z16.s }, p0, [x24, #2, MUL VL]\n"
+ "inch x24, ALL, MUL #3\n"
+ "bgt 47b\n"
+ "b 52f\n"
+ "48:" // Accumulate: Height 7
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "49:" // Accumulate: Height 7: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "ld1h { z22.s }, p2/Z, [x10]\n"
+ "ld1h { z21.s }, p2/Z, [x28]\n"
+ "ld1h { z20.s }, p2/Z, [x27]\n"
+ "ld1h { z19.s }, p2/Z, [x26]\n"
+ "ld1h { z18.s }, p2/Z, [x25]\n"
+ "ld1h { z17.s }, p2/Z, [x24]\n"
+ "ld1h { z16.s }, p2/Z, [x23]\n"
+ "ld1w { z8.s }, p2/Z, [%x[in_ptr]]\n"
+ "lsl z25.s, z22.s, #0x10\n"
+ "lsl z24.s, z21.s, #0x10\n"
+ "ld1w { z21.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "ld1w { z7.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "lsl z20.s, z20.s, #0x10\n"
+ "lsl z19.s, z19.s, #0x10\n"
+ "ld1w { z23.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "ld1w { z6.s }, p2/Z, [x20, #-4, MUL VL]\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "ld1w { z5.s }, p2/Z, [x20, #-1, MUL VL]\n"
+ "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z8.s, z8.s, z25.s\n"
+ "fadd z21.s, z21.s, z24.s\n"
+ "fadd z7.s, z7.s, z20.s\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "fadd z23.s, z23.s, z19.s\n"
+ "fadd z6.s, z6.s, z18.s\n"
+ "fadd z5.s, z5.s, z17.s\n"
+ "fadd z22.s, z22.s, z16.s\n"
+ "fmin z8.s, p3/M, z8.s, z12.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z7.s, p3/M, z7.s, z12.s\n"
+ "ld1w { z4.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "ld1w { z3.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "whilelt p0.s, x21, x11\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z6.s, p3/M, z6.s, z12.s\n"
+ "ld1w { z2.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "ld1w { z1.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "fmin z5.s, p3/M, z5.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "ld1w { z0.s }, p1/Z, [x20, #-3, MUL VL]\n"
+ "ld1w { z31.s }, p1/Z, [x20]\n"
+ "fmax z8.s, p3/M, z8.s, z11.s\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "ld1w { z30.s }, p1/Z, [x20, #3, MUL VL]\n"
+ "ld1w { z29.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "fmax z7.s, p3/M, z7.s, z11.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "ld1w { z28.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "ld1w { z27.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "fmax z6.s, p3/M, z6.s, z11.s\n"
+ "fmax z5.s, p3/M, z5.s, z11.s\n"
+ "ld1w { z26.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "ld1w { z25.s }, p0/Z, [x20, #-2, MUL VL]\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ ".inst 0x658aad13 // bfcvt z19.h, p3/M, z8.s\n"
+ ".inst 0x658aaeb5 // bfcvt z21.h, p3/M, z21.s\n"
+ "ld1w { z24.s }, p0/Z, [x20, #1, MUL VL]\n"
+ ".inst 0x658aacf4 // bfcvt z20.h, p3/M, z7.s\n"
+ ".inst 0x658aaef2 // bfcvt z18.h, p3/M, z23.s\n"
+ "ld1w { z23.s }, p0/Z, [x20, #4, MUL VL]\n"
+ "decw x11, ALL, MUL #3\n"
+ ".inst 0x658aacd1 // bfcvt z17.h, p3/M, z6.s\n"
+ ".inst 0x658aacb0 // bfcvt z16.h, p3/M, z5.s\n"
+ "incw x21\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "st1h { z19.s }, p2, [x10]\n"
+ ".inst 0x658aaed3 // bfcvt z19.h, p3/M, z22.s\n"
+ "st1h { z21.s }, p2, [x28]\n"
+ "cmp x11, XZR\n"
+ "st1h { z20.s }, p2, [x27]\n"
+ "st1h { z18.s }, p2, [x26]\n"
+ "ld1h { z18.s }, p1/Z, [x10, #1, MUL VL]\n"
+ "st1h { z17.s }, p2, [x25]\n"
+ "ld1h { z17.s }, p1/Z, [x28, #1, MUL VL]\n"
+ "st1h { z16.s }, p2, [x24]\n"
+ "ld1h { z16.s }, p1/Z, [x27, #1, MUL VL]\n"
+ "st1h { z19.s }, p2, [x23]\n"
+ "ld1h { z19.s }, p1/Z, [x26, #1, MUL VL]\n"
+ "lsl z22.s, z18.s, #0x10\n"
+ "ld1h { z18.s }, p1/Z, [x25, #1, MUL VL]\n"
+ "lsl z21.s, z17.s, #0x10\n"
+ "ld1h { z17.s }, p1/Z, [x24, #1, MUL VL]\n"
+ "lsl z20.s, z16.s, #0x10\n"
+ "ld1h { z16.s }, p1/Z, [x23, #1, MUL VL]\n"
+ "lsl z19.s, z19.s, #0x10\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "fadd z4.s, z4.s, z22.s\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z3.s, z3.s, z21.s\n"
+ "fadd z2.s, z2.s, z20.s\n"
+ "fadd z1.s, z1.s, z19.s\n"
+ "fadd z0.s, z0.s, z18.s\n"
+ "fadd z31.s, z31.s, z17.s\n"
+ "fmin z4.s, p3/M, z4.s, z12.s\n"
+ "fadd z30.s, z30.s, z16.s\n"
+ "fmin z3.s, p3/M, z3.s, z12.s\n"
+ "fmin z2.s, p3/M, z2.s, z12.s\n"
+ "fmin z1.s, p3/M, z1.s, z12.s\n"
+ "fmin z0.s, p3/M, z0.s, z12.s\n"
+ "fmin z31.s, p3/M, z31.s, z12.s\n"
+ "fmax z4.s, p3/M, z4.s, z11.s\n"
+ "fmin z30.s, p3/M, z30.s, z12.s\n"
+ "fmax z3.s, p3/M, z3.s, z11.s\n"
+ "fmax z2.s, p3/M, z2.s, z11.s\n"
+ "fmax z1.s, p3/M, z1.s, z11.s\n"
+ "fmax z0.s, p3/M, z0.s, z11.s\n"
+ "fmax z31.s, p3/M, z31.s, z11.s\n"
+ ".inst 0x658aac90 // bfcvt z16.h, p3/M, z4.s\n"
+ "fmax z30.s, p3/M, z30.s, z11.s\n"
+ ".inst 0x658aac74 // bfcvt z20.h, p3/M, z3.s\n"
+ ".inst 0x658aac53 // bfcvt z19.h, p3/M, z2.s\n"
+ ".inst 0x658aac32 // bfcvt z18.h, p3/M, z1.s\n"
+ "st1h { z16.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aac11 // bfcvt z17.h, p3/M, z0.s\n"
+ ".inst 0x658aaff0 // bfcvt z16.h, p3/M, z31.s\n"
+ "st1h { z20.s }, p1, [x28, #1, MUL VL]\n"
+ "st1h { z19.s }, p1, [x27, #1, MUL VL]\n"
+ ".inst 0x658aafd3 // bfcvt z19.h, p3/M, z30.s\n"
+ "st1h { z18.s }, p1, [x26, #1, MUL VL]\n"
+ "ld1h { z18.s }, p0/Z, [x10, #2, MUL VL]\n"
+ "st1h { z17.s }, p1, [x25, #1, MUL VL]\n"
+ "ld1h { z17.s }, p0/Z, [x28, #2, MUL VL]\n"
+ "st1h { z16.s }, p1, [x24, #1, MUL VL]\n"
+ "ld1h { z16.s }, p0/Z, [x27, #2, MUL VL]\n"
+ "st1h { z19.s }, p1, [x23, #1, MUL VL]\n"
+ "ld1h { z19.s }, p0/Z, [x26, #2, MUL VL]\n"
+ "lsl z22.s, z18.s, #0x10\n"
+ "ld1h { z18.s }, p0/Z, [x25, #2, MUL VL]\n"
+ "lsl z21.s, z17.s, #0x10\n"
+ "ld1h { z17.s }, p0/Z, [x24, #2, MUL VL]\n"
+ "lsl z20.s, z16.s, #0x10\n"
+ "ld1h { z16.s }, p0/Z, [x23, #2, MUL VL]\n"
+ "lsl z19.s, z19.s, #0x10\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "fadd z29.s, z29.s, z22.s\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z28.s, z28.s, z21.s\n"
+ "fadd z27.s, z27.s, z20.s\n"
+ "fadd z26.s, z26.s, z19.s\n"
+ "fadd z25.s, z25.s, z18.s\n"
+ "fadd z24.s, z24.s, z17.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "fadd z23.s, z23.s, z16.s\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ ".inst 0x658aafb1 // bfcvt z17.h, p3/M, z29.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ ".inst 0x658aaf94 // bfcvt z20.h, p3/M, z28.s\n"
+ ".inst 0x658aaf70 // bfcvt z16.h, p3/M, z27.s\n"
+ ".inst 0x658aaf53 // bfcvt z19.h, p3/M, z26.s\n"
+ "st1h { z17.s }, p0, [x10, #2, MUL VL]\n"
+ "inch x10, ALL, MUL #3\n"
+ ".inst 0x658aaf32 // bfcvt z18.h, p3/M, z25.s\n"
+ ".inst 0x658aaf11 // bfcvt z17.h, p3/M, z24.s\n"
+ "st1h { z20.s }, p0, [x28, #2, MUL VL]\n"
+ "inch x28, ALL, MUL #3\n"
+ "st1h { z16.s }, p0, [x27, #2, MUL VL]\n"
+ ".inst 0x658aaef0 // bfcvt z16.h, p3/M, z23.s\n"
+ "inch x27, ALL, MUL #3\n"
+ "st1h { z19.s }, p0, [x26, #2, MUL VL]\n"
+ "inch x26, ALL, MUL #3\n"
+ "st1h { z18.s }, p0, [x25, #2, MUL VL]\n"
+ "inch x25, ALL, MUL #3\n"
+ "st1h { z17.s }, p0, [x24, #2, MUL VL]\n"
+ "inch x24, ALL, MUL #3\n"
+ "st1h { z16.s }, p0, [x23, #2, MUL VL]\n"
+ "inch x23, ALL, MUL #3\n"
+ "bgt 49b\n"
+ "b 52f\n"
+ "50:" // Accumulate: Height 8
+ "mov x10, %x[out_ptr]\n"
+ "mov x11, %x[cols]\n"
+ "add x28, x10, %x[ldout], LSL #1\n"
+ "add x27, x28, %x[ldout], LSL #1\n"
+ "add x26, x27, %x[ldout], LSL #1\n"
+ "add x25, x26, %x[ldout], LSL #1\n"
+ "add x24, x25, %x[ldout], LSL #1\n"
+ "add x23, x24, %x[ldout], LSL #1\n"
+ "add x22, x23, %x[ldout], LSL #1\n"
+ "51:" // Accumulate: Height 8: Block loop
+ "mov x21, #0x0\n"
+ "addvl x20, %x[in_ptr], #16\n"
+ "whilelt p2.s, x21, x11\n"
+ "incw x21\n"
+ "ld1h { z23.s }, p2/Z, [x10]\n"
+ "ld1h { z22.s }, p2/Z, [x28]\n"
+ "ld1h { z21.s }, p2/Z, [x27]\n"
+ "ld1h { z20.s }, p2/Z, [x26]\n"
+ "ld1h { z19.s }, p2/Z, [x25]\n"
+ "ld1h { z18.s }, p2/Z, [x24]\n"
+ "ld1h { z17.s }, p2/Z, [x23]\n"
+ "ld1h { z16.s }, p2/Z, [x22]\n"
+ "lsl z31.s, z23.s, #0x10\n"
+ "lsl z30.s, z22.s, #0x10\n"
+ "ld1w { z29.s }, p2/Z, [%x[in_ptr]]\n"
+ "ld1w { z28.s }, p2/Z, [%x[in_ptr], #3, MUL VL]\n"
+ "lsl z27.s, z21.s, #0x10\n"
+ "lsl z26.s, z20.s, #0x10\n"
+ "ld1w { z21.s }, p2/Z, [%x[in_ptr], #6, MUL VL]\n"
+ "ld1w { z25.s }, p2/Z, [x20, #-7, MUL VL]\n"
+ "lsl z20.s, z19.s, #0x10\n"
+ "lsl z19.s, z18.s, #0x10\n"
+ "ld1w { z18.s }, p2/Z, [x20, #-4, MUL VL]\n"
+ "ld1w { z24.s }, p2/Z, [x20, #-1, MUL VL]\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "ld1w { z23.s }, p2/Z, [x20, #2, MUL VL]\n"
+ "ld1w { z22.s }, p2/Z, [x20, #5, MUL VL]\n"
+ "fadd z29.s, z29.s, z31.s\n"
+ "fadd z28.s, z28.s, z30.s\n"
+ "fadd z21.s, z21.s, z27.s\n"
+ "fadd z25.s, z25.s, z26.s\n"
+ "whilelt p1.s, x21, x11\n"
+ "incw x21\n"
+ "fadd z18.s, z18.s, z20.s\n"
+ "fadd z24.s, z24.s, z19.s\n"
+ "fadd z23.s, z23.s, z17.s\n"
+ "fadd z22.s, z22.s, z16.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fmin z21.s, p3/M, z21.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "ld1w { z6.s }, p1/Z, [%x[in_ptr], #1, MUL VL]\n"
+ "ld1w { z5.s }, p1/Z, [%x[in_ptr], #4, MUL VL]\n"
+ "fmin z18.s, p3/M, z18.s, z12.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "ld1w { z4.s }, p1/Z, [%x[in_ptr], #7, MUL VL]\n"
+ "ld1w { z3.s }, p1/Z, [x20, #-6, MUL VL]\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmin z22.s, p3/M, z22.s, z12.s\n"
+ "ld1w { z2.s }, p1/Z, [x20, #-3, MUL VL]\n"
+ "ld1w { z1.s }, p1/Z, [x20]\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
+ "ld1w { z31.s }, p1/Z, [x20, #6, MUL VL]\n"
+ "fmax z21.s, p3/M, z21.s, z11.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "fmax z22.s, p3/M, z22.s, z11.s\n"
+ ".inst 0x658aafb4 // bfcvt z20.h, p3/M, z29.s\n"
+ ".inst 0x658aaf93 // bfcvt z19.h, p3/M, z28.s\n"
+ ".inst 0x658aaeb5 // bfcvt z21.h, p3/M, z21.s\n"
+ ".inst 0x658aaf30 // bfcvt z16.h, p3/M, z25.s\n"
+ "whilelt p0.s, x21, x11\n"
+ "decw x11, ALL, MUL #3\n"
+ ".inst 0x658aae52 // bfcvt z18.h, p3/M, z18.s\n"
+ ".inst 0x658aaf11 // bfcvt z17.h, p3/M, z24.s\n"
+ "incw x21\n"
+ "st1h { z20.s }, p2, [x10]\n"
+ "st1h { z19.s }, p2, [x28]\n"
+ ".inst 0x658aaef4 // bfcvt z20.h, p3/M, z23.s\n"
+ ".inst 0x658aaed3 // bfcvt z19.h, p3/M, z22.s\n"
+ "st1h { z21.s }, p2, [x27]\n"
+ "ld1w { z30.s }, p0/Z, [%x[in_ptr], #2, MUL VL]\n"
+ "ld1w { z29.s }, p0/Z, [%x[in_ptr], #5, MUL VL]\n"
+ "cmp x11, XZR\n"
+ "st1h { z16.s }, p2, [x26]\n"
+ "ld1h { z16.s }, p1/Z, [x10, #1, MUL VL]\n"
+ "ld1w { z28.s }, p0/Z, [x20, #-8, MUL VL]\n"
+ "addvl %x[in_ptr], %x[in_ptr], #24\n"
+ "st1h { z18.s }, p2, [x25]\n"
+ "ld1h { z18.s }, p1/Z, [x28, #1, MUL VL]\n"
+ "ld1w { z27.s }, p0/Z, [x20, #-5, MUL VL]\n"
+ "st1h { z17.s }, p2, [x24]\n"
+ "ld1h { z17.s }, p1/Z, [x27, #1, MUL VL]\n"
+ "ld1w { z26.s }, p0/Z, [x20, #-2, MUL VL]\n"
+ "st1h { z20.s }, p2, [x23]\n"
+ "ld1h { z20.s }, p1/Z, [x26, #1, MUL VL]\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "ld1w { z25.s }, p0/Z, [x20, #1, MUL VL]\n"
+ "st1h { z19.s }, p2, [x22]\n"
+ "ld1h { z19.s }, p1/Z, [x25, #1, MUL VL]\n"
+ "lsl z22.s, z18.s, #0x10\n"
+ "ld1w { z24.s }, p0/Z, [x20, #4, MUL VL]\n"
+ "ld1h { z18.s }, p1/Z, [x24, #1, MUL VL]\n"
+ "lsl z21.s, z17.s, #0x10\n"
+ "ld1w { z23.s }, p0/Z, [x20, #7, MUL VL]\n"
+ "ld1h { z17.s }, p1/Z, [x23, #1, MUL VL]\n"
+ "lsl z20.s, z20.s, #0x10\n"
+ "fadd z6.s, z6.s, z16.s\n"
+ "ld1h { z16.s }, p1/Z, [x22, #1, MUL VL]\n"
+ "lsl z19.s, z19.s, #0x10\n"
+ "fadd z5.s, z5.s, z22.s\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "fadd z4.s, z4.s, z21.s\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fmin z6.s, p3/M, z6.s, z12.s\n"
+ "fadd z3.s, z3.s, z20.s\n"
+ "fadd z2.s, z2.s, z19.s\n"
+ "fmin z5.s, p3/M, z5.s, z12.s\n"
+ "fadd z1.s, z1.s, z18.s\n"
+ "fmin z4.s, p3/M, z4.s, z12.s\n"
+ "fadd z0.s, z0.s, z17.s\n"
+ "fadd z31.s, z31.s, z16.s\n"
+ "fmax z6.s, p3/M, z6.s, z11.s\n"
+ "fmin z3.s, p3/M, z3.s, z12.s\n"
+ "fmin z2.s, p3/M, z2.s, z12.s\n"
+ "fmax z5.s, p3/M, z5.s, z11.s\n"
+ "fmin z1.s, p3/M, z1.s, z12.s\n"
+ "fmin z0.s, p3/M, z0.s, z12.s\n"
+ "fmin z31.s, p3/M, z31.s, z12.s\n"
+ "fmax z4.s, p3/M, z4.s, z11.s\n"
+ ".inst 0x658aacd0 // bfcvt z16.h, p3/M, z6.s\n"
+ "fmax z3.s, p3/M, z3.s, z11.s\n"
+ "fmax z2.s, p3/M, z2.s, z11.s\n"
+ ".inst 0x658aacb1 // bfcvt z17.h, p3/M, z5.s\n"
+ "fmax z1.s, p3/M, z1.s, z11.s\n"
+ "fmax z0.s, p3/M, z0.s, z11.s\n"
+ "fmax z31.s, p3/M, z31.s, z11.s\n"
+ "st1h { z16.s }, p1, [x10, #1, MUL VL]\n"
+ ".inst 0x658aac90 // bfcvt z16.h, p3/M, z4.s\n"
+ "st1h { z17.s }, p1, [x28, #1, MUL VL]\n"
+ ".inst 0x658aac75 // bfcvt z21.h, p3/M, z3.s\n"
+ ".inst 0x658aac52 // bfcvt z18.h, p3/M, z2.s\n"
+ ".inst 0x658aac31 // bfcvt z17.h, p3/M, z1.s\n"
+ ".inst 0x658aac14 // bfcvt z20.h, p3/M, z0.s\n"
+ "st1h { z16.s }, p1, [x27, #1, MUL VL]\n"
+ ".inst 0x658aaff3 // bfcvt z19.h, p3/M, z31.s\n"
+ "ld1h { z16.s }, p0/Z, [x10, #2, MUL VL]\n"
+ "st1h { z21.s }, p1, [x26, #1, MUL VL]\n"
+ "st1h { z18.s }, p1, [x25, #1, MUL VL]\n"
+ "ld1h { z18.s }, p0/Z, [x28, #2, MUL VL]\n"
+ "st1h { z17.s }, p1, [x24, #1, MUL VL]\n"
+ "ld1h { z17.s }, p0/Z, [x27, #2, MUL VL]\n"
+ "st1h { z20.s }, p1, [x23, #1, MUL VL]\n"
+ "ld1h { z20.s }, p0/Z, [x26, #2, MUL VL]\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "st1h { z19.s }, p1, [x22, #1, MUL VL]\n"
+ "ld1h { z19.s }, p0/Z, [x25, #2, MUL VL]\n"
+ "lsl z22.s, z18.s, #0x10\n"
+ "ld1h { z18.s }, p0/Z, [x24, #2, MUL VL]\n"
+ "lsl z21.s, z17.s, #0x10\n"
+ "ld1h { z17.s }, p0/Z, [x23, #2, MUL VL]\n"
+ "lsl z20.s, z20.s, #0x10\n"
+ "fadd z30.s, z30.s, z16.s\n"
+ "ld1h { z16.s }, p0/Z, [x22, #2, MUL VL]\n"
+ "lsl z19.s, z19.s, #0x10\n"
+ "lsl z18.s, z18.s, #0x10\n"
+ "fadd z29.s, z29.s, z22.s\n"
+ "lsl z17.s, z17.s, #0x10\n"
+ "fadd z28.s, z28.s, z21.s\n"
+ "lsl z16.s, z16.s, #0x10\n"
+ "fadd z27.s, z27.s, z20.s\n"
+ "fmin z30.s, p3/M, z30.s, z12.s\n"
+ "fadd z26.s, z26.s, z19.s\n"
+ "fadd z25.s, z25.s, z18.s\n"
+ "fmin z29.s, p3/M, z29.s, z12.s\n"
+ "fadd z24.s, z24.s, z17.s\n"
+ "fmin z28.s, p3/M, z28.s, z12.s\n"
+ "fadd z23.s, z23.s, z16.s\n"
+ "fmin z27.s, p3/M, z27.s, z12.s\n"
+ "fmax z30.s, p3/M, z30.s, z11.s\n"
+ "fmin z26.s, p3/M, z26.s, z12.s\n"
+ "fmin z25.s, p3/M, z25.s, z12.s\n"
+ "fmax z29.s, p3/M, z29.s, z11.s\n"
+ "fmin z24.s, p3/M, z24.s, z12.s\n"
+ "fmin z23.s, p3/M, z23.s, z12.s\n"
+ "fmax z28.s, p3/M, z28.s, z11.s\n"
+ "fmax z27.s, p3/M, z27.s, z11.s\n"
+ ".inst 0x658aafd0 // bfcvt z16.h, p3/M, z30.s\n"
+ "fmax z26.s, p3/M, z26.s, z11.s\n"
+ "fmax z25.s, p3/M, z25.s, z11.s\n"
+ ".inst 0x658aafb1 // bfcvt z17.h, p3/M, z29.s\n"
+ "fmax z24.s, p3/M, z24.s, z11.s\n"
+ "fmax z23.s, p3/M, z23.s, z11.s\n"
+ "st1h { z16.s }, p0, [x10, #2, MUL VL]\n"
+ ".inst 0x658aaf90 // bfcvt z16.h, p3/M, z28.s\n"
+ ".inst 0x658aaf74 // bfcvt z20.h, p3/M, z27.s\n"
+ "inch x10, ALL, MUL #3\n"
+ "st1h { z17.s }, p0, [x28, #2, MUL VL]\n"
+ "inch x28, ALL, MUL #3\n"
+ ".inst 0x658aaf53 // bfcvt z19.h, p3/M, z26.s\n"
+ ".inst 0x658aaf32 // bfcvt z18.h, p3/M, z25.s\n"
+ "st1h { z16.s }, p0, [x27, #2, MUL VL]\n"
+ "inch x27, ALL, MUL #3\n"
+ ".inst 0x658aaf11 // bfcvt z17.h, p3/M, z24.s\n"
+ ".inst 0x658aaef0 // bfcvt z16.h, p3/M, z23.s\n"
+ "st1h { z20.s }, p0, [x26, #2, MUL VL]\n"
+ "inch x26, ALL, MUL #3\n"
+ "st1h { z19.s }, p0, [x25, #2, MUL VL]\n"
+ "inch x25, ALL, MUL #3\n"
+ "st1h { z18.s }, p0, [x24, #2, MUL VL]\n"
+ "inch x24, ALL, MUL #3\n"
+ "st1h { z17.s }, p0, [x23, #2, MUL VL]\n"
+ "inch x23, ALL, MUL #3\n"
+ "st1h { z16.s }, p0, [x22, #2, MUL VL]\n"
+ "inch x22, ALL, MUL #3\n"
+ "bgt 51b\n"
+ "subs %x[rows], %x[rows], #0x8\n"
+ "add %x[out_ptr], %x[out_ptr], x12\n"
+ "bgt 35b\n"
+ "52:" // Exit
+ : [in_ptr] "+&r" (in_ptr), [out_ptr] "+&r" (out_ptr), [rows] "+&r" (rows)
+ : [accumulate] "r" (accumulate), [bias] "r" (bias), [cols] "r" (cols), [ldout] "r" (ldout), [maxval] "r" (maxval), [minval] "r" (minval)
+ : "cc", "memory", "p0", "p1", "p2", "p3", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+ );
+}
+
+#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/cpu/operators/CpuMatMul.cpp b/src/cpu/operators/CpuMatMul.cpp
index 8908712..f68ae98 100644
--- a/src/cpu/operators/CpuMatMul.cpp
+++ b/src/cpu/operators/CpuMatMul.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2023 Arm Limited.
+ * Copyright (c) 2023-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -102,8 +102,8 @@
const ActivationLayerInfo &act_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, rhs, dst);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lhs, 1, DataType::F32, DataType::F16, DataType::QASYMM8,
- DataType::QASYMM8_SIGNED);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lhs, 1, DataType::F32, DataType::F16, DataType::BFLOAT16,
+ DataType::QASYMM8, DataType::QASYMM8_SIGNED);
ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs->are_values_constant(), "LHS Tensor must be dynamic.");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs->are_values_constant(), "RHS Tensor must be dynamic.");
ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(lhs);
@@ -120,6 +120,7 @@
auto gemm_info = AsmGemmInfo();
gemm_info.activation_info = act_info;
gemm_info.fast_mode = settings.fast_math();
+ gemm_info.fixed_format = settings.fixed_format();
// Validate and then permute a/b
if (adj_lhs)
@@ -157,6 +158,14 @@
gemm_info.activation_info, gemm_info.output_stage));
}
+ if (gemm_info.fixed_format)
+ {
+ gemm_info.weight_format = WeightFormat::ANY;
+ arm_compute::WeightFormat expected_weight_format = WeightFormat::ANY;
+ ARM_COMPUTE_RETURN_ON_ERROR(cpu::CpuGemmAssemblyDispatch::has_opt_impl(expected_weight_format, lhs_to_use,
+ rhs_to_use, nullptr, dst, gemm_info));
+ }
+
cpu::CpuGemmAssemblyDispatch::validate(lhs_to_use, rhs_to_use, nullptr, dst, gemm_info);
return Status{};
@@ -221,6 +230,7 @@
// Fill AsmGemmInfo class object before configuration
_gemm_info.activation_info = act_info;
_gemm_info.fast_mode = settings.fast_math();
+ _gemm_info.fixed_format = settings.fixed_format();
_gemm_info.negated_offsets = false;
lhs_to_use = (_adj_lhs) ? _lhs_transposed : lhs_to_use;
@@ -233,6 +243,18 @@
_gemm_info.output_stage);
}
+ if (_gemm_info.fixed_format)
+ {
+ _gemm_info.weight_format = WeightFormat::ANY;
+ arm_compute::WeightFormat expected_weight_format = WeightFormat::ANY;
+ ARM_COMPUTE_ERROR_THROW_ON(cpu::CpuGemmAssemblyDispatch::has_opt_impl(expected_weight_format, &lhs_to_use,
+ &rhs_to_use, nullptr, dst, _gemm_info));
+ // Set gemm weights info to the one returned by has_opt_impl
+ _gemm_info.weight_format = expected_weight_format;
+ // has_opt_impl may return a non fast math kernel, even if we requested one
+ _gemm_info.fast_mode = arm_compute::is_fixed_format_fast_math(expected_weight_format);
+ }
+
// Configure Asm Kernel
_asm_glue = std::make_unique<cpu::CpuGemmAssemblyDispatch>();
_asm_glue->configure(&lhs_to_use, &rhs_to_use, nullptr, &dst_to_use,
diff --git a/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp b/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp
index 58ee68f..efe2a7a 100644
--- a/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp
+++ b/src/cpu/operators/internal/CpuGemmAssemblyDispatch.cpp
@@ -581,7 +581,6 @@
// Fixed format kernels need no pretranspose.
ARM_COMPUTE_ERROR_ON(arm_compute::is_fixed_format(
assembly_utils::map_to_arm_compute_weight_format(_gemm_kernel_asm->get_config().weight_format)));
-
const int ldb = b_to_use->info()->strides_in_bytes().y() / b_to_use->info()->element_size();
const auto in1_ptr = reinterpret_cast<const TypeInput *>(b_to_use->buffer() +
b_to_use->info()->offset_first_element_in_bytes());
@@ -857,6 +856,7 @@
arm_gemm::WeightFormat arm_gemm_expected_wf = assembly_utils::map_to_arm_gemm_weight_format(expected_weight_format);
arm_gemm::GemmArgs args(&ci, p.M, p.N, p.K, p.sections, p.batches, p.multis, p.indirect, act, num_threads,
info.fixed_format, info.fast_mode, &cfg);
+
// TODO: Incorporate info.transpose_b COMPMID-6595
switch (a->data_type())
{
@@ -900,9 +900,18 @@
#if defined(ARM_COMPUTE_ENABLE_BF16)
case DataType::BFLOAT16:
{
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(
- !(arm_gemm::has_opt_gemm<bfloat16, float, arm_gemm::Nothing>(arm_gemm_expected_wf, args, {})),
- "We could not find an optimized kernel for BFLOAT16 input and F32 output");
+ if (d->data_type() == DataType::BFLOAT16)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+ !(arm_gemm::has_opt_gemm<bfloat16, bfloat16, arm_gemm::Nothing>(arm_gemm_expected_wf, args, {})),
+ "We could not find an optimized kernel for BFLOAT16 input and BFLOAT16 output");
+ }
+ else
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+ !(arm_gemm::has_opt_gemm<bfloat16, float, arm_gemm::Nothing>(arm_gemm_expected_wf, args, {})),
+ "We could not find an optimized kernel for BFLOAT16 input and F32 output");
+ }
break;
}
#endif /* defined(ARM_COMPUTE_ENABLE_BF16) */
@@ -958,8 +967,9 @@
"Only F32 output supported for F32 input");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::F16 && d->data_type() != DataType::F16,
"Only F16 output supported for F16 input");
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::BFLOAT16 && d->data_type() != DataType::F32,
- "Only F32 output supported for BFLOAT16 input");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::BFLOAT16 &&
+ (d->data_type() != DataType::F32 && d->data_type() != DataType::BFLOAT16),
+ "Only F32/BFLOAT16 output supported for BFLOAT16 input");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::U8 && d->data_type() != DataType::U32,
"Only U32 output supported for U8 input");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->data_type() == DataType::S8 && d->data_type() != DataType::S32,
@@ -1030,7 +1040,14 @@
#endif /* __aarch64__ */
#if defined(ARM_COMPUTE_ENABLE_BF16)
case DataType::BFLOAT16:
- create_arm_gemm<bfloat16, float>(_arm_gemm, a, b, c, d, act, info);
+ if (d->data_type() == DataType::BFLOAT16)
+ {
+ create_arm_gemm<bfloat16, bfloat16>(_arm_gemm, a, b, c, d, act, info);
+ }
+ else
+ {
+ create_arm_gemm<bfloat16, float>(_arm_gemm, a, b, c, d, act, info);
+ }
break;
#endif /* defined(ARM_COMPUTE_ENABLE_BF16) */
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC